update_resylts

config.json

@@ -1,11 +1,13 @@
 {
   "architectures": [
-    "StructRoberta"
+    "StructRoberta",
+    "StructRobertaForSequenceClassification"
   ],
   "attention_probs_dropout_prob": 0.1,
   "auto_map": {
     "AutoConfig": "modeling_structroberta.StructRobertaConfig",
-    "AutoModelForMaskedLM": "modeling_structroberta.StructRoberta"
+    "AutoModelForMaskedLM": "modeling_structroberta.StructRoberta",
+    "AutoModelForSequenceClassification": "modeling_structroberta.StructRobertaForSequenceClassification"
   },
   "bos_token_id": 0,
   "classifier_dropout": null,
@@ -36,4 +38,4 @@
   "use_cache": true,
   "vocab_size": 32000,
   "weight_act": "softmax"
-}
+}

finetune/boolq/all_results.json

@@ -0,0 +1,16 @@
+{
+    "epoch": 10.0,
+    "eval_accuracy": 0.6486860513687134,
+    "eval_f1": 0.7221006564551423,
+    "eval_loss": 1.0183159112930298,
+    "eval_mcc": 0.25076276821098453,
+    "eval_runtime": 1.5451,
+    "eval_samples": 723,
+    "eval_samples_per_second": 467.928,
+    "eval_steps_per_second": 58.896,
+    "train_loss": 0.3817570156521267,
+    "train_runtime": 95.4851,
+    "train_samples": 2072,
+    "train_samples_per_second": 216.997,
+    "train_steps_per_second": 1.885
+}

finetune/boolq/config.json

@@ -0,0 +1,50 @@
+{
+  "_name_or_path": "final_models/structroberta_sx2_final",
+  "architectures": [
+    "StructRobertaForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "auto_map": {
+    "AutoConfig": "modeling_structroberta.StructRobertaConfig",
+    "AutoModelForMaskedLM": "modeling_structroberta.StructRoberta",
+    "AutoModelForSequenceClassification": "modeling_structroberta.StructRobertaForSequenceClassification"
+  },
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "conv_size": 9,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": 0,
+    "1": 1
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "label2id": {
+    "0": 0,
+    "1": 1
+  },
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "roberta",
+  "n_cntxt_layers": 4,
+  "n_cntxt_layers_2": 0,
+  "n_parser_layers": 6,
+  "num_attention_heads": 12,
+  "num_hidden_layers": 8,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "problem_type": "single_label_classification",
+  "relations": [
+    "head",
+    "child"
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.1",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 32000,
+  "weight_act": "softmax"
+}

finetune/boolq/eval_results.json

@@ -0,0 +1,11 @@
+{
+    "epoch": 10.0,
+    "eval_accuracy": 0.6486860513687134,
+    "eval_f1": 0.7221006564551423,
+    "eval_loss": 1.0183159112930298,
+    "eval_mcc": 0.25076276821098453,
+    "eval_runtime": 1.5451,
+    "eval_samples": 723,
+    "eval_samples_per_second": 467.928,
+    "eval_steps_per_second": 58.896
+}

finetune/boolq/modeling_structroberta.py

@@ -0,0 +1,2146 @@
+import copy
+import math
+from typing import List, Optional, Tuple, Union
+import torch
+import torch.utils.checkpoint
+from packaging import version
+from torch import nn
+import torch.nn.functional as F
+from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
+from transformers.activations import ACT2FN, gelu
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    MaskedLMOutput,
+    SequenceClassifierOutput
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.utils import logging
+from transformers import RobertaConfig
+
+
+logger = logging.get_logger(__name__)
+ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "roberta-base",
+    "roberta-large",
+    "roberta-large-mnli",
+    "distilroberta-base",
+    "roberta-base-openai-detector",
+    "roberta-large-openai-detector",
+    # See all RoBERTa models at https://huggingface.co/models?filter=roberta
+]
+
+
+class StructRobertaConfig(RobertaConfig):
+    model_type = "roberta"
+
+    def __init__(
+        self,
+        n_parser_layers=4,
+        conv_size=9,
+        relations=("head", "child"),
+        weight_act="softmax",
+        n_cntxt_layers=3,
+        n_cntxt_layers_2=0,
+        **kwargs,):
+        
+        super().__init__(**kwargs)
+        self.n_cntxt_layers = n_cntxt_layers
+        self.n_parser_layers = n_parser_layers
+        self.n_cntxt_layers_2 = n_cntxt_layers_2
+        self.conv_size = conv_size
+        self.relations = relations
+        self.weight_act = weight_act
+
+class Conv1d(nn.Module):
+    """1D convolution layer."""
+
+    def __init__(self, hidden_size, kernel_size, dilation=1):
+        """Initialization.
+
+        Args:
+        hidden_size: dimension of input embeddings
+        kernel_size: convolution kernel size
+        dilation: the spacing between the kernel points
+        """
+        super(Conv1d, self).__init__()
+
+        if kernel_size % 2 == 0:
+            padding = (kernel_size // 2) * dilation
+            self.shift = True
+        else:
+            padding = ((kernel_size - 1) // 2) * dilation
+            self.shift = False
+        self.conv = nn.Conv1d(
+            hidden_size, hidden_size, kernel_size, padding=padding, dilation=dilation
+        )
+
+    def forward(self, x):
+        """Compute convolution.
+
+        Args:
+          x: input embeddings
+        Returns:
+          conv_output: convolution results
+        """
+
+        if self.shift:
+            return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
+        else:
+            return self.conv(x.transpose(1, 2)).transpose(1, 2)
+
+
+class RobertaEmbeddings(nn.Module):
+    """
+    Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
+    """
+
+    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(
+            config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id
+        )
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.hidden_size
+        )
+        self.token_type_embeddings = nn.Embedding(
+            config.type_vocab_size, config.hidden_size
+        )
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.position_embedding_type = getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1))
+        )
+        if version.parse(torch.__version__) > version.parse("1.6.0"):
+            self.register_buffer(
+                "token_type_ids",
+                torch.zeros(self.position_ids.size(), dtype=torch.long),
+                persistent=False,
+            )
+
+        # End copy
+        self.padding_idx = config.pad_token_id
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings,
+            config.hidden_size,
+            padding_idx=self.padding_idx,
+        )
+
+    def forward(
+        self,
+        input_ids=None,
+        token_type_ids=None,
+        position_ids=None,
+        inputs_embeds=None,
+        past_key_values_length=0,
+    ):
+        if position_ids is None:
+            if input_ids is not None:
+                # Create the position ids from the input token ids. Any padded tokens remain padded.
+                position_ids = create_position_ids_from_input_ids(
+                    input_ids, self.padding_idx, past_key_values_length
+                )
+            else:
+                position_ids = self.create_position_ids_from_inputs_embeds(
+                    inputs_embeds
+                )
+
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(
+                    input_shape[0], seq_length
+                )
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(
+                    input_shape, dtype=torch.long, device=self.position_ids.device
+                )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+        """
+        We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
+
+        Args:
+            inputs_embeds: torch.Tensor
+
+        Returns: torch.Tensor
+        """
+        input_shape = inputs_embeds.size()[:-1]
+        sequence_length = input_shape[1]
+
+        position_ids = torch.arange(
+            self.padding_idx + 1,
+            sequence_length + self.padding_idx + 1,
+            dtype=torch.long,
+            device=inputs_embeds.device,
+        )
+        return position_ids.unsqueeze(0).expand(input_shape)
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfAttention with Bert->Roberta
+class RobertaSelfAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(
+            config, "embedding_size"
+        ):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = position_embedding_type or getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(
+                2 * config.max_position_embeddings - 1, self.attention_head_size
+            )
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (
+            self.num_attention_heads,
+            self.attention_head_size,
+        )
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        parser_att_mask=None,
+    ) -> Tuple[torch.Tensor]:
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(-1, 1)
+            position_ids_r = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(
+                distance + self.max_position_embeddings - 1
+            )
+            positional_embedding = positional_embedding.to(
+                dtype=query_layer.dtype
+            )  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                relative_position_scores_key = torch.einsum(
+                    "bhrd,lrd->bhlr", key_layer, positional_embedding
+                )
+                attention_scores = (
+                    attention_scores
+                    + relative_position_scores_query
+                    + relative_position_scores_key
+                )
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in RobertaModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        if parser_att_mask is None:
+            # Normalize the attention scores to probabilities.
+            attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+        else:
+            attention_probs = torch.sigmoid(attention_scores) * parser_att_mask
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (
+            (context_layer, attention_probs) if output_attentions else (context_layer,)
+        )
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput
+class RobertaSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(
+        self, hidden_states: torch.Tensor, input_tensor: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->Roberta
+class RobertaAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        self.self = RobertaSelfAttention(
+            config, position_embedding_type=position_embedding_type
+        )
+        self.output = RobertaSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads,
+            self.self.num_attention_heads,
+            self.self.attention_head_size,
+            self.pruned_heads,
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = (
+            self.self.attention_head_size * self.self.num_attention_heads
+        )
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        parser_att_mask=None,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+            parser_att_mask=parser_att_mask,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[
+            1:
+        ]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate
+class RobertaIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOutput
+class RobertaOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(
+        self, hidden_states: torch.Tensor, input_tensor: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertLayer with Bert->Roberta
+class RobertaLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = RobertaAttention(config)
+        self.is_decoder = config.is_decoder
+        self.add_cross_attention = config.add_cross_attention
+        if self.add_cross_attention:
+            if not self.is_decoder:
+                raise ValueError(
+                    f"{self} should be used as a decoder model if cross attention is added"
+                )
+            self.crossattention = RobertaAttention(
+                config, position_embedding_type="absolute"
+            )
+        self.intermediate = RobertaIntermediate(config)
+        self.output = RobertaOutput(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        parser_att_mask=None,
+    ) -> Tuple[torch.Tensor]:
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = (
+            past_key_value[:2] if past_key_value is not None else None
+        )
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+            parser_att_mask=parser_att_mask,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[
+                1:
+            ]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            if not hasattr(self, "crossattention"):
+                raise ValueError(
+                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
+                )
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = (
+                past_key_value[-2:] if past_key_value is not None else None
+            )
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                cross_attn_past_key_value,
+                output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = (
+                outputs + cross_attention_outputs[1:-1]
+            )  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk,
+            self.chunk_size_feed_forward,
+            self.seq_len_dim,
+            attention_output,
+        )
+        outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+# Copied from transformers.models.bert.modeling_bert.BertEncoder with Bert->Roberta
+class RobertaEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList(
+            [RobertaLayer(config) for _ in range(config.num_hidden_layers)]
+        )
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+        parser_att_mask=None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = (
+            () if output_attentions and self.config.add_cross_attention else None
+        )
+
+        next_decoder_cache = () if use_cache else None
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                if parser_att_mask is not None:
+                    layer_outputs = layer_module(
+                        hidden_states,
+                        attention_mask,
+                        layer_head_mask,
+                        encoder_hidden_states,
+                        encoder_attention_mask,
+                        past_key_value,
+                        output_attentions,
+                        parser_att_mask=parser_att_mask[i])
+                else:
+                    layer_outputs = layer_module(
+                        hidden_states,
+                        attention_mask,
+                        layer_head_mask,
+                        encoder_hidden_states,
+                        encoder_attention_mask,
+                        past_key_value,
+                        output_attentions,
+                        parser_att_mask=None)
+                    
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertPooler
+class RobertaPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class RobertaPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = RobertaConfig
+    base_model_prefix = "roberta"
+    supports_gradient_checkpointing = True
+
+    # Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            if module.bias is not None:
+                module.bias.data.zero_()
+                module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, RobertaEncoder):
+            module.gradient_checkpointing = value
+
+    def update_keys_to_ignore(self, config, del_keys_to_ignore):
+        """Remove some keys from ignore list"""
+        if not config.tie_word_embeddings:
+            # must make a new list, or the class variable gets modified!
+            self._keys_to_ignore_on_save = [
+                k for k in self._keys_to_ignore_on_save if k not in del_keys_to_ignore
+            ]
+            self._keys_to_ignore_on_load_missing = [
+                k
+                for k in self._keys_to_ignore_on_load_missing
+                if k not in del_keys_to_ignore
+            ]
+
+
+ROBERTA_START_DOCSTRING = r"""
+
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`RobertaConfig`]): Model configuration class with all the parameters of the
+            model. Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+ROBERTA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`RobertaTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class RobertaModel(RobertaPreTrainedModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in *Attention is
+    all you need*_ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
+    Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set
+    to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and
+    `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass.
+
+    .. _*Attention is all you need*: https://arxiv.org/abs/1706.03762
+
+    """
+
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.__init__ with Bert->Roberta
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = RobertaEmbeddings(config)
+        self.encoder = RobertaEncoder(config)
+
+        self.pooler = RobertaPooler(config) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.forward
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        parser_att_mask=None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
+        r"""
+        encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        """
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = (
+            past_key_values[0][0].shape[2] if past_key_values is not None else 0
+        )
+
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                ((batch_size, seq_length + past_key_values_length)), device=device
+            )
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(
+                    batch_size, seq_length
+                )
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(
+                    input_shape, dtype=torch.long, device=device
+                )
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(
+            attention_mask, input_shape, device
+        )
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            (
+                encoder_batch_size,
+                encoder_sequence_length,
+                _,
+            ) = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(
+                encoder_attention_mask
+            )
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=parser_att_mask,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = (
+            self.pooler(sequence_output) if self.pooler is not None else None
+        )
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+class StructRoberta(RobertaPreTrainedModel):
+    _keys_to_ignore_on_save = [r"lm_head.decoder.weight", r"lm_head.decoder.bias"]
+    _keys_to_ignore_on_load_missing = [
+        r"position_ids",
+        r"lm_head.decoder.weight",
+        r"lm_head.decoder.bias",
+    ]
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        if config.is_decoder:
+            logger.warning(
+                "If you want to use `RobertaForMaskedLM` make sure `config.is_decoder=False` for "
+                "bi-directional self-attention."
+            )
+
+        
+        if config.n_cntxt_layers > 0:
+            config_cntxt = copy.deepcopy(config)
+            config_cntxt.num_hidden_layers = config.n_cntxt_layers
+            
+            self.cntxt_layers = RobertaModel(config_cntxt, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers_2 > 0:
+            self.parser_layers_1 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_1 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_1 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_1 = nn.Parameter(
+                torch.zeros((config.n_cntxt_layers_2, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_1.data.normal_(0, 0.1)
+
+            self._scaler_1 = nn.Parameter(torch.zeros(2))
+            
+            config_cntxt_2 = copy.deepcopy(config)
+            config_cntxt_2.num_hidden_layers = config.n_cntxt_layers_2
+            
+            self.cntxt_layers_2 = RobertaModel(config_cntxt_2, add_pooling_layer=False)
+            
+            
+            self.parser_layers_2 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_2 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_2 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_2 = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_2.data.normal_(0, 0.1)
+
+            self._scaler_2 = nn.Parameter(torch.zeros(2))
+            
+        else:
+            self.parser_layers = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(config.n_parser_layers)
+                ]
+            )
+
+            self.distance_ff = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight.data.normal_(0, 0.1)
+
+            self._scaler = nn.Parameter(torch.zeros(2))
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers > 0:
+            self.cntxt_layers.embeddings = self.roberta.embeddings
+        if config.n_cntxt_layers_2 > 0:
+            self.cntxt_layers_2.embeddings = self.roberta.embeddings
+        
+        self.lm_head = RobertaLMHead(config)
+
+        self.pad = config.pad_token_id
+
+        # The LM head weights require special treatment only when they are tied with the word embeddings
+        self.update_keys_to_ignore(config, ["lm_head.decoder.weight"])
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head.decoder = new_embeddings
+
+    @property
+    def scaler(self):
+        return self._scaler.exp()
+    
+    @property
+    def scaler_1(self):
+        return self._scaler_1.exp()
+    
+    @property
+    def scaler_2(self):
+        return self._scaler_2.exp()
+
+    @property
+    def rel_weight(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight, dim=-1)
+        
+    @property
+    def rel_weight_1(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_1)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_1, dim=-1)
+
+    
+    @property
+    def rel_weight_2(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_2)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_2, dim=-1)
+
+
+    def compute_block(self, distance, height, n_cntxt_layers=0):
+        """Compute constituents from distance and height."""
+
+        if n_cntxt_layers>0:
+            if n_cntxt_layers == 1:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_1[0]
+            elif n_cntxt_layers == 2:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_2[0]
+        else:
+            beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+        gamma = torch.sigmoid(-beta_logits)
+        ones = torch.ones_like(gamma)
+
+        block_mask_left = cummin(
+            gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1
+        )
+        block_mask_left = block_mask_left - F.pad(
+            block_mask_left[:, :, :-1], (1, 0), value=0
+        )
+        block_mask_left.tril_(0)
+
+        block_mask_right = cummin(
+            gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1
+        )
+        block_mask_right = block_mask_right - F.pad(
+            block_mask_right[:, :, 1:], (0, 1), value=0
+        )
+        block_mask_right.triu_(0)
+
+        block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+        block = cumsum(block_mask_left).tril(0) + cumsum(
+            block_mask_right, reverse=True
+        ).triu(1)
+
+        return block_p, block
+
+    def compute_head(self, height, n_cntxt_layers=0):
+        """Estimate head for each constituent."""
+
+        _, length = height.size()
+        if n_cntxt_layers>0:
+          if n_cntxt_layers == 1:
+              head_logits = height * self.scaler_1[1]
+          elif n_cntxt_layers == 2:
+              head_logits = height * self.scaler_2[1]  
+        else:
+            head_logits = height * self.scaler[1]
+        index = torch.arange(length, device=height.device)
+
+        mask = (index[:, None, None] <= index[None, None, :]) * (
+            index[None, None, :] <= index[None, :, None]
+        )
+        head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+        head_logits.masked_fill_(~mask[None, :, :, :], -1e9)
+
+        head_p = torch.softmax(head_logits, dim=-1)
+
+        return head_p
+
+    def parse(self, x, embs=None, n_cntxt_layers=0):
+        """Parse input sentence.
+
+        Args:
+        x: input tokens (required).
+        pos: position for each token (optional).
+        Returns:
+        distance: syntactic distance
+        height: syntactic height
+        """
+
+        mask = x != self.pad
+        mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+        if embs is None:
+            h = self.roberta.embeddings(x)
+        else:
+            h = embs
+        
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                parser_layers = self.parser_layers_1
+                height_ff = self.height_ff_1
+                distance_ff = self.distance_ff_1
+            elif n_cntxt_layers == 2:
+                parser_layers = self.parser_layers_2
+                height_ff = self.height_ff_2
+                distance_ff = self.distance_ff_2
+            for i in range(int(self.config.n_parser_layers/2)):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = parser_layers[i](h)
+
+            height = height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+        else:
+            for i in range(self.config.n_parser_layers):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = self.parser_layers[i](h)
+
+            height = self.height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = self.distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+
+        return distance, height
+
+    def generate_mask(self, x, distance, height, n_cntxt_layers=0):
+        """Compute head and cibling distribution for each token."""
+
+        bsz, length = x.size()
+
+        eye = torch.eye(length, device=x.device, dtype=torch.bool)
+        eye = eye[None, :, :].expand((bsz, -1, -1))
+
+        block_p, block = self.compute_block(distance, height, n_cntxt_layers=n_cntxt_layers)
+        head_p = self.compute_head(height, n_cntxt_layers=n_cntxt_layers)
+        head = torch.einsum("blij,bijh->blh", block_p, head_p)
+        head = head.masked_fill(eye, 0)
+        child = head.transpose(1, 2)
+        cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+        rel_list = []
+        if "head" in self.config.relations:
+            rel_list.append(head)
+        if "child" in self.config.relations:
+            rel_list.append(child)
+        if "cibling" in self.config.relations:
+            rel_list.append(cibling)
+
+        rel = torch.stack(rel_list, dim=1)
+
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                rel_weight = self.rel_weight_1
+            elif n_cntxt_layers == 2:
+                rel_weight = self.rel_weight_2
+        else:
+            rel_weight = self.rel_weight
+
+        dep = torch.einsum("lhr,brij->lbhij", rel_weight, rel)
+        
+        if n_cntxt_layers == 1:
+            num_layers = self.cntxt_layers_2.config.num_hidden_layers
+        else:
+            num_layers = self.roberta.config.num_hidden_layers
+            
+        att_mask = dep.reshape(
+            num_layers,
+            bsz,
+            self.config.num_attention_heads,
+            length,
+            length,
+        )
+
+        return att_mask, cibling, head, block
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, MaskedLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
+            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+        kwargs (`Dict[str, any]`, optional, defaults to *{}*):
+            Used to hide legacy arguments that have been deprecated.
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        
+        if self.config.n_cntxt_layers > 0:
+            cntxt_outputs = self.cntxt_layers(
+                input_ids,
+                attention_mask=attention_mask,
+                token_type_ids=token_type_ids,
+                position_ids=position_ids,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict)
+            
+        
+        if self.config.n_cntxt_layers_2 > 0:
+            distance_1, height_1 = self.parse(input_ids, cntxt_outputs[0], n_cntxt_layers=1)
+            att_mask_1, _, _, _ = self.generate_mask(input_ids, distance_1, height_1, n_cntxt_layers=1)
+            
+            cntxt_outputs_2 = self.cntxt_layers_2(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask_1)
+            
+            sequence_output = cntxt_outputs_2[0]
+            
+            distance_2, height_2 = self.parse(input_ids, sequence_output[0], n_cntxt_layers=2)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance_2, height_2, n_cntxt_layers=2)
+
+        elif self.config.n_cntxt_layers > 0:
+            distance, height = self.parse(input_ids, cntxt_outputs[0])
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+        else:
+            distance, height = self.parse(input_ids)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask,
+        )
+        sequence_output = outputs[0]
+        prediction_scores = self.lm_head(sequence_output)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(
+                prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return (
+                ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+            )
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class RobertaLMHead(nn.Module):
+    """Roberta Head for masked language modeling."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size)
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+        self.decoder.bias = self.bias
+
+    def forward(self, features, **kwargs):
+        x = self.dense(features)
+        x = gelu(x)
+        x = self.layer_norm(x)
+
+        # project back to size of vocabulary with bias
+        x = self.decoder(x)
+
+        return x
+
+    def _tie_weights(self):
+        # To tie those two weights if they get disconnected (on TPU or when the bias is resized)
+        self.bias = self.decoder.bias
+
+
+class StructRobertaForSequenceClassification(RobertaPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        if config.n_cntxt_layers > 0:
+            config_cntxt = copy.deepcopy(config)
+            config_cntxt.num_hidden_layers = config.n_cntxt_layers
+            
+            self.cntxt_layers = RobertaModel(config_cntxt, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers_2 > 0:
+            self.parser_layers_1 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_1 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_1 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_1 = nn.Parameter(
+                torch.zeros((config.n_cntxt_layers_2, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_1.data.normal_(0, 0.1)
+
+            self._scaler_1 = nn.Parameter(torch.zeros(2))
+            
+            config_cntxt_2 = copy.deepcopy(config)
+            config_cntxt_2.num_hidden_layers = config.n_cntxt_layers_2
+            
+            self.cntxt_layers_2 = RobertaModel(config_cntxt_2, add_pooling_layer=False)
+            
+            
+            self.parser_layers_2 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_2 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_2 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_2 = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_2.data.normal_(0, 0.1)
+
+            self._scaler_2 = nn.Parameter(torch.zeros(2))
+            
+        else:
+            self.parser_layers = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(config.n_parser_layers)
+                ]
+            )
+
+            self.distance_ff = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight.data.normal_(0, 0.1)
+
+            self._scaler = nn.Parameter(torch.zeros(2))
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers > 0:
+            self.cntxt_layers.embeddings = self.roberta.embeddings
+        if config.n_cntxt_layers_2 > 0:
+            self.cntxt_layers_2.embeddings = self.roberta.embeddings
+        
+        
+        self.pad = config.pad_token_id
+        self.classifier = RobertaClassificationHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    
+    @property
+    def scaler(self):
+        return self._scaler.exp()
+    
+    @property
+    def scaler_1(self):
+        return self._scaler_1.exp()
+    
+    @property
+    def scaler_2(self):
+        return self._scaler_2.exp()
+
+    @property
+    def rel_weight(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight, dim=-1)
+        
+    @property
+    def rel_weight_1(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_1)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_1, dim=-1)
+
+    
+    @property
+    def rel_weight_2(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_2)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_2, dim=-1)
+
+
+    def compute_block(self, distance, height, n_cntxt_layers=0):
+        """Compute constituents from distance and height."""
+
+        if n_cntxt_layers>0:
+            if n_cntxt_layers == 1:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_1[0]
+            elif n_cntxt_layers == 2:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_2[0]
+        else:
+            beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+        gamma = torch.sigmoid(-beta_logits)
+        ones = torch.ones_like(gamma)
+
+        block_mask_left = cummin(
+            gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1
+        )
+        block_mask_left = block_mask_left - F.pad(
+            block_mask_left[:, :, :-1], (1, 0), value=0
+        )
+        block_mask_left.tril_(0)
+
+        block_mask_right = cummin(
+            gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1
+        )
+        block_mask_right = block_mask_right - F.pad(
+            block_mask_right[:, :, 1:], (0, 1), value=0
+        )
+        block_mask_right.triu_(0)
+
+        block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+        block = cumsum(block_mask_left).tril(0) + cumsum(
+            block_mask_right, reverse=True
+        ).triu(1)
+
+        return block_p, block
+
+    def compute_head(self, height, n_cntxt_layers=0):
+        """Estimate head for each constituent."""
+
+        _, length = height.size()
+        if n_cntxt_layers>0:
+          if n_cntxt_layers == 1:
+              head_logits = height * self.scaler_1[1]
+          elif n_cntxt_layers == 2:
+              head_logits = height * self.scaler_2[1]  
+        else:
+            head_logits = height * self.scaler[1]
+        index = torch.arange(length, device=height.device)
+
+        mask = (index[:, None, None] <= index[None, None, :]) * (
+            index[None, None, :] <= index[None, :, None]
+        )
+        head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+        head_logits.masked_fill_(~mask[None, :, :, :], -1e9)
+
+        head_p = torch.softmax(head_logits, dim=-1)
+
+        return head_p
+
+    def parse(self, x, embs=None, n_cntxt_layers=0):
+        """Parse input sentence.
+
+        Args:
+        x: input tokens (required).
+        pos: position for each token (optional).
+        Returns:
+        distance: syntactic distance
+        height: syntactic height
+        """
+
+        mask = x != self.pad
+        mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+        if embs is None:
+            h = self.roberta.embeddings(x)
+        else:
+            h = embs
+        
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                parser_layers = self.parser_layers_1
+                height_ff = self.height_ff_1
+                distance_ff = self.distance_ff_1
+            elif n_cntxt_layers == 2:
+                parser_layers = self.parser_layers_2
+                height_ff = self.height_ff_2
+                distance_ff = self.distance_ff_2
+            for i in range(int(self.config.n_parser_layers/2)):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = parser_layers[i](h)
+
+            height = height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+        else:
+            for i in range(self.config.n_parser_layers):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = self.parser_layers[i](h)
+
+            height = self.height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = self.distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+
+        return distance, height
+
+    def generate_mask(self, x, distance, height, n_cntxt_layers=0):
+        """Compute head and cibling distribution for each token."""
+
+        bsz, length = x.size()
+
+        eye = torch.eye(length, device=x.device, dtype=torch.bool)
+        eye = eye[None, :, :].expand((bsz, -1, -1))
+
+        block_p, block = self.compute_block(distance, height, n_cntxt_layers=n_cntxt_layers)
+        head_p = self.compute_head(height, n_cntxt_layers=n_cntxt_layers)
+        head = torch.einsum("blij,bijh->blh", block_p, head_p)
+        head = head.masked_fill(eye, 0)
+        child = head.transpose(1, 2)
+        cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+        rel_list = []
+        if "head" in self.config.relations:
+            rel_list.append(head)
+        if "child" in self.config.relations:
+            rel_list.append(child)
+        if "cibling" in self.config.relations:
+            rel_list.append(cibling)
+
+        rel = torch.stack(rel_list, dim=1)
+
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                rel_weight = self.rel_weight_1
+            elif n_cntxt_layers == 2:
+                rel_weight = self.rel_weight_2
+        else:
+            rel_weight = self.rel_weight
+
+        dep = torch.einsum("lhr,brij->lbhij", rel_weight, rel)
+        
+        if n_cntxt_layers == 1:
+            num_layers = self.cntxt_layers_2.config.num_hidden_layers
+        else:
+            num_layers = self.roberta.config.num_hidden_layers
+            
+        att_mask = dep.reshape(
+            num_layers,
+            bsz,
+            self.config.num_attention_heads,
+            length,
+            length,
+        )
+
+        return att_mask, cibling, head, block
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        
+        if self.config.n_cntxt_layers > 0:
+            cntxt_outputs = self.cntxt_layers(
+                input_ids,
+                attention_mask=attention_mask,
+                token_type_ids=token_type_ids,
+                position_ids=position_ids,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict)
+            
+        
+        if self.config.n_cntxt_layers_2 > 0:
+            distance_1, height_1 = self.parse(input_ids, cntxt_outputs[0], n_cntxt_layers=1)
+            att_mask_1, _, _, _ = self.generate_mask(input_ids, distance_1, height_1, n_cntxt_layers=1)
+            
+            cntxt_outputs_2 = self.cntxt_layers_2(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask_1)
+            
+            sequence_output = cntxt_outputs_2[0]
+            
+            distance_2, height_2 = self.parse(input_ids, sequence_output[0], n_cntxt_layers=2)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance_2, height_2, n_cntxt_layers=2)
+
+        elif self.config.n_cntxt_layers > 0:
+            distance, height = self.parse(input_ids, cntxt_outputs[0])
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+        else:
+            distance, height = self.parse(input_ids)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask,
+        )
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class RobertaClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+def create_position_ids_from_input_ids(
+    input_ids, padding_idx, past_key_values_length=0
+):
+    """
+    Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
+    are ignored. This is modified from fairseq's `utils.make_positions`.
+
+    Args:
+        x: torch.Tensor x:
+
+    Returns: torch.Tensor
+    """
+    # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
+    mask = input_ids.ne(padding_idx).int()
+    incremental_indices = (
+        torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length
+    ) * mask
+    return incremental_indices.long() + padding_idx
+
+
+def cumprod(x, reverse=False, exclusive=False):
+    """cumulative product."""
+    if reverse:
+        x = x.flip([-1])
+
+    if exclusive:
+        x = F.pad(x[:, :, :-1], (1, 0), value=1)
+
+    cx = x.cumprod(-1)
+
+    if reverse:
+        cx = cx.flip([-1])
+    return cx
+
+
+def cumsum(x, reverse=False, exclusive=False):
+    """cumulative sum."""
+    bsz, _, length = x.size()
+    device = x.device
+    if reverse:
+        if exclusive:
+            w = torch.ones([bsz, length, length], device=device).tril(-1)
+        else:
+            w = torch.ones([bsz, length, length], device=device).tril(0)
+        cx = torch.bmm(x, w)
+    else:
+        if exclusive:
+            w = torch.ones([bsz, length, length], device=device).triu(1)
+        else:
+            w = torch.ones([bsz, length, length], device=device).triu(0)
+        cx = torch.bmm(x, w)
+    return cx
+
+
+def cummin(x, reverse=False, exclusive=False, max_value=1e9):
+    """cumulative min."""
+    if reverse:
+        if exclusive:
+            x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
+        x = x.flip([-1]).cummin(-1)[0].flip([-1])
+    else:
+        if exclusive:
+            x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
+        x = x.cummin(-1)[0]
+    return x

finetune/boolq/predict_results.txt

@@ -0,0 +1,724 @@
+index	prediction
+0	1
+1	1
+2	1
+3	0
+4	1
+5	1
+6	1
+7	0
+8	1
+9	0
+10	0
+11	1
+12	0
+13	1
+14	1
+15	1
+16	1
+17	0
+18	1
+19	1
+20	0
+21	1
+22	1
+23	1
+24	0
+25	1
+26	0
+27	1
+28	0
+29	1
+30	0
+31	1
+32	0
+33	1
+34	0
+35	1
+36	0
+37	1
+38	1
+39	1
+40	0
+41	0
+42	1
+43	1
+44	0
+45	1
+46	1
+47	1
+48	1
+49	1
+50	1
+51	1
+52	1
+53	0
+54	1
+55	1
+56	1
+57	0
+58	0
+59	0
+60	1
+61	1
+62	1
+63	0
+64	0
+65	1
+66	1
+67	1
+68	1
+69	1
+70	1
+71	0
+72	1
+73	1
+74	0
+75	0
+76	0
+77	0
+78	1
+79	1
+80	1
+81	1
+82	0
+83	1
+84	1
+85	1
+86	1
+87	1
+88	1
+89	0
+90	1
+91	1
+92	0
+93	1
+94	0
+95	1
+96	1
+97	1
+98	0
+99	1
+100	1
+101	0
+102	1
+103	0
+104	1
+105	1
+106	0
+107	0
+108	1
+109	1
+110	1
+111	1
+112	0
+113	1
+114	1
+115	1
+116	1
+117	1
+118	1
+119	1
+120	1
+121	1
+122	1
+123	0
+124	1
+125	1
+126	0
+127	0
+128	1
+129	0
+130	0
+131	1
+132	0
+133	0
+134	1
+135	1
+136	1
+137	0
+138	1
+139	0
+140	1
+141	1
+142	1
+143	1
+144	1
+145	0
+146	1
+147	0
+148	1
+149	0
+150	0
+151	1
+152	0
+153	1
+154	1
+155	0
+156	0
+157	1
+158	0
+159	1
+160	1
+161	1
+162	1
+163	1
+164	0
+165	1
+166	1
+167	1
+168	0
+169	0
+170	1
+171	1
+172	0
+173	0
+174	1
+175	1
+176	1
+177	1
+178	0
+179	1
+180	1
+181	0
+182	1
+183	1
+184	0
+185	1
+186	0
+187	0
+188	1
+189	1
+190	1
+191	0
+192	1
+193	0
+194	0
+195	0
+196	1
+197	0
+198	0
+199	1
+200	1
+201	1
+202	1
+203	0
+204	1
+205	1
+206	1
+207	0
+208	1
+209	1
+210	1
+211	1
+212	1
+213	0
+214	1
+215	0
+216	1
+217	1
+218	1
+219	1
+220	1
+221	1
+222	1
+223	0
+224	0
+225	0
+226	1
+227	1
+228	0
+229	1
+230	0
+231	0
+232	0
+233	0
+234	1
+235	1
+236	0
+237	0
+238	1
+239	1
+240	0
+241	1
+242	1
+243	1
+244	0
+245	1
+246	1
+247	1
+248	1
+249	1
+250	0
+251	0
+252	1
+253	1
+254	1
+255	1
+256	1
+257	1
+258	1
+259	0
+260	1
+261	0
+262	1
+263	1
+264	0
+265	1
+266	1
+267	0
+268	1
+269	1
+270	1
+271	1
+272	1
+273	1
+274	1
+275	1
+276	1
+277	1
+278	1
+279	1
+280	1
+281	1
+282	0
+283	0
+284	0
+285	0
+286	0
+287	0
+288	1
+289	0
+290	1
+291	0
+292	1
+293	1
+294	1
+295	1
+296	1
+297	0
+298	1
+299	1
+300	1
+301	0
+302	1
+303	0
+304	1
+305	0
+306	0
+307	0
+308	0
+309	1
+310	0
+311	1
+312	1
+313	0
+314	1
+315	1
+316	1
+317	1
+318	1
+319	0
+320	1
+321	1
+322	1
+323	1
+324	0
+325	0
+326	1
+327	0
+328	1
+329	1
+330	1
+331	1
+332	1
+333	1
+334	0
+335	0
+336	0
+337	1
+338	1
+339	1
+340	1
+341	0
+342	1
+343	0
+344	0
+345	1
+346	1
+347	0
+348	1
+349	1
+350	1
+351	0
+352	1
+353	1
+354	1
+355	0
+356	0
+357	1
+358	0
+359	1
+360	1
+361	1
+362	1
+363	1
+364	0
+365	1
+366	0
+367	1
+368	1
+369	1
+370	1
+371	1
+372	1
+373	1
+374	1
+375	1
+376	1
+377	1
+378	0
+379	0
+380	0
+381	1
+382	1
+383	1
+384	1
+385	0
+386	1
+387	1
+388	0
+389	1
+390	1
+391	0
+392	0
+393	0
+394	1
+395	1
+396	1
+397	1
+398	0
+399	1
+400	1
+401	0
+402	1
+403	1
+404	1
+405	1
+406	0
+407	0
+408	0
+409	1
+410	1
+411	1
+412	0
+413	1
+414	1
+415	0
+416	1
+417	1
+418	0
+419	1
+420	1
+421	1
+422	1
+423	0
+424	1
+425	0
+426	0
+427	1
+428	1
+429	1
+430	1
+431	0
+432	1
+433	1
+434	1
+435	1
+436	1
+437	1
+438	0
+439	1
+440	0
+441	0
+442	1
+443	1
+444	1
+445	1
+446	1
+447	1
+448	1
+449	1
+450	1
+451	1
+452	1
+453	0
+454	1
+455	1
+456	1
+457	1
+458	1
+459	0
+460	0
+461	1
+462	1
+463	0
+464	0
+465	0
+466	1
+467	0
+468	1
+469	0
+470	1
+471	1
+472	1
+473	1
+474	0
+475	1
+476	1
+477	0
+478	0
+479	1
+480	1
+481	0
+482	0
+483	1
+484	1
+485	1
+486	1
+487	1
+488	1
+489	0
+490	1
+491	1
+492	1
+493	0
+494	1
+495	1
+496	1
+497	0
+498	0
+499	1
+500	1
+501	0
+502	1
+503	1
+504	0
+505	1
+506	1
+507	1
+508	1
+509	1
+510	0
+511	0
+512	1
+513	1
+514	0
+515	0
+516	0
+517	1
+518	0
+519	1
+520	0
+521	1
+522	0
+523	1
+524	0
+525	1
+526	0
+527	1
+528	0
+529	0
+530	1
+531	1
+532	0
+533	1
+534	1
+535	1
+536	1
+537	0
+538	1
+539	0
+540	1
+541	1
+542	1
+543	1
+544	1
+545	1
+546	1
+547	1
+548	1
+549	1
+550	1
+551	0
+552	0
+553	0
+554	0
+555	0
+556	1
+557	0
+558	1
+559	1
+560	1
+561	1
+562	1
+563	0
+564	1
+565	0
+566	1
+567	0
+568	1
+569	1
+570	0
+571	1
+572	1
+573	1
+574	0
+575	1
+576	1
+577	1
+578	1
+579	1
+580	0
+581	0
+582	1
+583	0
+584	1
+585	1
+586	1
+587	1
+588	0
+589	0
+590	0
+591	1
+592	0
+593	1
+594	1
+595	1
+596	1
+597	0
+598	1
+599	1
+600	1
+601	0
+602	1
+603	1
+604	1
+605	0
+606	1
+607	1
+608	1
+609	1
+610	1
+611	1
+612	1
+613	1
+614	1
+615	0
+616	1
+617	1
+618	1
+619	1
+620	0
+621	1
+622	1
+623	1
+624	1
+625	0
+626	1
+627	0
+628	0
+629	0
+630	1
+631	1
+632	0
+633	1
+634	1
+635	0
+636	1
+637	1
+638	1
+639	1
+640	1
+641	0
+642	1
+643	0
+644	0
+645	1
+646	1
+647	0
+648	0
+649	1
+650	1
+651	0
+652	1
+653	0
+654	0
+655	1
+656	1
+657	0
+658	1
+659	1
+660	0
+661	0
+662	1
+663	1
+664	0
+665	1
+666	0
+667	1
+668	0
+669	1
+670	0
+671	1
+672	1
+673	1
+674	0
+675	1
+676	1
+677	1
+678	1
+679	1
+680	0
+681	1
+682	1
+683	1
+684	1
+685	1
+686	0
+687	1
+688	1
+689	0
+690	1
+691	0
+692	0
+693	1
+694	1
+695	1
+696	0
+697	1
+698	1
+699	1
+700	0
+701	1
+702	1
+703	1
+704	1
+705	1
+706	1
+707	1
+708	1
+709	1
+710	1
+711	0
+712	1
+713	0
+714	0
+715	1
+716	0
+717	0
+718	1
+719	0
+720	1
+721	1
+722	1

finetune/boolq/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51931eeadf8ad3ae73a1fc172ba1d9006633436fb37eb9357a7c35007e423a5e
+size 577069633

finetune/boolq/special_tokens_map.json

@@ -0,0 +1,15 @@
+{
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "mask_token": {
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "unk_token": "<unk>"
+}

finetune/boolq/tokenizer_config.json

@@ -0,0 +1,65 @@
+{
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "cls_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "mask_token": {
+    "__type": "AddedToken",
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "model_max_length": 512,
+  "name_or_path": "final_models/structroberta_sx2_final",
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "special_tokens_map_file": null,
+  "tokenizer_class": "RobertaTokenizer",
+  "trim_offsets": true,
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

finetune/boolq/train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 10.0,
+    "train_loss": 0.3817570156521267,
+    "train_runtime": 95.4851,
+    "train_samples": 2072,
+    "train_samples_per_second": 216.997,
+    "train_steps_per_second": 1.885
+}

finetune/boolq/trainer_state.json

@@ -0,0 +1,25 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 10.0,
+  "global_step": 180,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 10.0,
+      "step": 180,
+      "total_flos": 1898033803223040.0,
+      "train_loss": 0.3817570156521267,
+      "train_runtime": 95.4851,
+      "train_samples_per_second": 216.997,
+      "train_steps_per_second": 1.885
+    }
+  ],
+  "max_steps": 180,
+  "num_train_epochs": 10,
+  "total_flos": 1898033803223040.0,
+  "trial_name": null,
+  "trial_params": null
+}

finetune/boolq/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b8ff19c91b7744abf01e78ed8ae95d953473a8e341bcbf64b13d1376902090c
+size 3503

finetune/cola/all_results.json

@@ -0,0 +1,16 @@
+{
+    "epoch": 10.0,
+    "eval_accuracy": 0.7065750956535339,
+    "eval_f1": 0.7994634473507714,
+    "eval_loss": 0.7585899829864502,
+    "eval_mcc": 0.262551358979245,
+    "eval_runtime": 2.1721,
+    "eval_samples": 1019,
+    "eval_samples_per_second": 469.121,
+    "eval_steps_per_second": 58.928,
+    "train_loss": 0.3363707777382671,
+    "train_runtime": 371.6095,
+    "train_samples": 8164,
+    "train_samples_per_second": 219.693,
+    "train_steps_per_second": 1.857
+}

finetune/cola/checkpoint-400/config.json

@@ -0,0 +1,50 @@
+{
+  "_name_or_path": "final_models/structroberta_sx2_final",
+  "architectures": [
+    "StructRobertaForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "auto_map": {
+    "AutoConfig": "modeling_structroberta.StructRobertaConfig",
+    "AutoModelForMaskedLM": "modeling_structroberta.StructRoberta",
+    "AutoModelForSequenceClassification": "modeling_structroberta.StructRobertaForSequenceClassification"
+  },
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "conv_size": 9,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": 0,
+    "1": 1
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "label2id": {
+    "0": 0,
+    "1": 1
+  },
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "roberta",
+  "n_cntxt_layers": 4,
+  "n_cntxt_layers_2": 0,
+  "n_parser_layers": 6,
+  "num_attention_heads": 12,
+  "num_hidden_layers": 8,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "problem_type": "single_label_classification",
+  "relations": [
+    "head",
+    "child"
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.1",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 32000,
+  "weight_act": "softmax"
+}

finetune/cola/checkpoint-400/modeling_structroberta.py

@@ -0,0 +1,2146 @@
+import copy
+import math
+from typing import List, Optional, Tuple, Union
+import torch
+import torch.utils.checkpoint
+from packaging import version
+from torch import nn
+import torch.nn.functional as F
+from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
+from transformers.activations import ACT2FN, gelu
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    MaskedLMOutput,
+    SequenceClassifierOutput
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.utils import logging
+from transformers import RobertaConfig
+
+
+logger = logging.get_logger(__name__)
+ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "roberta-base",
+    "roberta-large",
+    "roberta-large-mnli",
+    "distilroberta-base",
+    "roberta-base-openai-detector",
+    "roberta-large-openai-detector",
+    # See all RoBERTa models at https://huggingface.co/models?filter=roberta
+]
+
+
+class StructRobertaConfig(RobertaConfig):
+    model_type = "roberta"
+
+    def __init__(
+        self,
+        n_parser_layers=4,
+        conv_size=9,
+        relations=("head", "child"),
+        weight_act="softmax",
+        n_cntxt_layers=3,
+        n_cntxt_layers_2=0,
+        **kwargs,):
+        
+        super().__init__(**kwargs)
+        self.n_cntxt_layers = n_cntxt_layers
+        self.n_parser_layers = n_parser_layers
+        self.n_cntxt_layers_2 = n_cntxt_layers_2
+        self.conv_size = conv_size
+        self.relations = relations
+        self.weight_act = weight_act
+
+class Conv1d(nn.Module):
+    """1D convolution layer."""
+
+    def __init__(self, hidden_size, kernel_size, dilation=1):
+        """Initialization.
+
+        Args:
+        hidden_size: dimension of input embeddings
+        kernel_size: convolution kernel size
+        dilation: the spacing between the kernel points
+        """
+        super(Conv1d, self).__init__()
+
+        if kernel_size % 2 == 0:
+            padding = (kernel_size // 2) * dilation
+            self.shift = True
+        else:
+            padding = ((kernel_size - 1) // 2) * dilation
+            self.shift = False
+        self.conv = nn.Conv1d(
+            hidden_size, hidden_size, kernel_size, padding=padding, dilation=dilation
+        )
+
+    def forward(self, x):
+        """Compute convolution.
+
+        Args:
+          x: input embeddings
+        Returns:
+          conv_output: convolution results
+        """
+
+        if self.shift:
+            return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
+        else:
+            return self.conv(x.transpose(1, 2)).transpose(1, 2)
+
+
+class RobertaEmbeddings(nn.Module):
+    """
+    Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
+    """
+
+    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(
+            config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id
+        )
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.hidden_size
+        )
+        self.token_type_embeddings = nn.Embedding(
+            config.type_vocab_size, config.hidden_size
+        )
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.position_embedding_type = getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1))
+        )
+        if version.parse(torch.__version__) > version.parse("1.6.0"):
+            self.register_buffer(
+                "token_type_ids",
+                torch.zeros(self.position_ids.size(), dtype=torch.long),
+                persistent=False,
+            )
+
+        # End copy
+        self.padding_idx = config.pad_token_id
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings,
+            config.hidden_size,
+            padding_idx=self.padding_idx,
+        )
+
+    def forward(
+        self,
+        input_ids=None,
+        token_type_ids=None,
+        position_ids=None,
+        inputs_embeds=None,
+        past_key_values_length=0,
+    ):
+        if position_ids is None:
+            if input_ids is not None:
+                # Create the position ids from the input token ids. Any padded tokens remain padded.
+                position_ids = create_position_ids_from_input_ids(
+                    input_ids, self.padding_idx, past_key_values_length
+                )
+            else:
+                position_ids = self.create_position_ids_from_inputs_embeds(
+                    inputs_embeds
+                )
+
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(
+                    input_shape[0], seq_length
+                )
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(
+                    input_shape, dtype=torch.long, device=self.position_ids.device
+                )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+        """
+        We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
+
+        Args:
+            inputs_embeds: torch.Tensor
+
+        Returns: torch.Tensor
+        """
+        input_shape = inputs_embeds.size()[:-1]
+        sequence_length = input_shape[1]
+
+        position_ids = torch.arange(
+            self.padding_idx + 1,
+            sequence_length + self.padding_idx + 1,
+            dtype=torch.long,
+            device=inputs_embeds.device,
+        )
+        return position_ids.unsqueeze(0).expand(input_shape)
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfAttention with Bert->Roberta
+class RobertaSelfAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(
+            config, "embedding_size"
+        ):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = position_embedding_type or getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(
+                2 * config.max_position_embeddings - 1, self.attention_head_size
+            )
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (
+            self.num_attention_heads,
+            self.attention_head_size,
+        )
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        parser_att_mask=None,
+    ) -> Tuple[torch.Tensor]:
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(-1, 1)
+            position_ids_r = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(
+                distance + self.max_position_embeddings - 1
+            )
+            positional_embedding = positional_embedding.to(
+                dtype=query_layer.dtype
+            )  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                relative_position_scores_key = torch.einsum(
+                    "bhrd,lrd->bhlr", key_layer, positional_embedding
+                )
+                attention_scores = (
+                    attention_scores
+                    + relative_position_scores_query
+                    + relative_position_scores_key
+                )
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in RobertaModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        if parser_att_mask is None:
+            # Normalize the attention scores to probabilities.
+            attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+        else:
+            attention_probs = torch.sigmoid(attention_scores) * parser_att_mask
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (
+            (context_layer, attention_probs) if output_attentions else (context_layer,)
+        )
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput
+class RobertaSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(
+        self, hidden_states: torch.Tensor, input_tensor: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->Roberta
+class RobertaAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        self.self = RobertaSelfAttention(
+            config, position_embedding_type=position_embedding_type
+        )
+        self.output = RobertaSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads,
+            self.self.num_attention_heads,
+            self.self.attention_head_size,
+            self.pruned_heads,
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = (
+            self.self.attention_head_size * self.self.num_attention_heads
+        )
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        parser_att_mask=None,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+            parser_att_mask=parser_att_mask,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[
+            1:
+        ]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate
+class RobertaIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOutput
+class RobertaOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(
+        self, hidden_states: torch.Tensor, input_tensor: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertLayer with Bert->Roberta
+class RobertaLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = RobertaAttention(config)
+        self.is_decoder = config.is_decoder
+        self.add_cross_attention = config.add_cross_attention
+        if self.add_cross_attention:
+            if not self.is_decoder:
+                raise ValueError(
+                    f"{self} should be used as a decoder model if cross attention is added"
+                )
+            self.crossattention = RobertaAttention(
+                config, position_embedding_type="absolute"
+            )
+        self.intermediate = RobertaIntermediate(config)
+        self.output = RobertaOutput(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        parser_att_mask=None,
+    ) -> Tuple[torch.Tensor]:
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = (
+            past_key_value[:2] if past_key_value is not None else None
+        )
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+            parser_att_mask=parser_att_mask,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[
+                1:
+            ]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            if not hasattr(self, "crossattention"):
+                raise ValueError(
+                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
+                )
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = (
+                past_key_value[-2:] if past_key_value is not None else None
+            )
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                cross_attn_past_key_value,
+                output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = (
+                outputs + cross_attention_outputs[1:-1]
+            )  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk,
+            self.chunk_size_feed_forward,
+            self.seq_len_dim,
+            attention_output,
+        )
+        outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+# Copied from transformers.models.bert.modeling_bert.BertEncoder with Bert->Roberta
+class RobertaEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList(
+            [RobertaLayer(config) for _ in range(config.num_hidden_layers)]
+        )
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+        parser_att_mask=None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = (
+            () if output_attentions and self.config.add_cross_attention else None
+        )
+
+        next_decoder_cache = () if use_cache else None
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                if parser_att_mask is not None:
+                    layer_outputs = layer_module(
+                        hidden_states,
+                        attention_mask,
+                        layer_head_mask,
+                        encoder_hidden_states,
+                        encoder_attention_mask,
+                        past_key_value,
+                        output_attentions,
+                        parser_att_mask=parser_att_mask[i])
+                else:
+                    layer_outputs = layer_module(
+                        hidden_states,
+                        attention_mask,
+                        layer_head_mask,
+                        encoder_hidden_states,
+                        encoder_attention_mask,
+                        past_key_value,
+                        output_attentions,
+                        parser_att_mask=None)
+                    
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertPooler
+class RobertaPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class RobertaPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = RobertaConfig
+    base_model_prefix = "roberta"
+    supports_gradient_checkpointing = True
+
+    # Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            if module.bias is not None:
+                module.bias.data.zero_()
+                module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, RobertaEncoder):
+            module.gradient_checkpointing = value
+
+    def update_keys_to_ignore(self, config, del_keys_to_ignore):
+        """Remove some keys from ignore list"""
+        if not config.tie_word_embeddings:
+            # must make a new list, or the class variable gets modified!
+            self._keys_to_ignore_on_save = [
+                k for k in self._keys_to_ignore_on_save if k not in del_keys_to_ignore
+            ]
+            self._keys_to_ignore_on_load_missing = [
+                k
+                for k in self._keys_to_ignore_on_load_missing
+                if k not in del_keys_to_ignore
+            ]
+
+
+ROBERTA_START_DOCSTRING = r"""
+
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`RobertaConfig`]): Model configuration class with all the parameters of the
+            model. Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+ROBERTA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`RobertaTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class RobertaModel(RobertaPreTrainedModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in *Attention is
+    all you need*_ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
+    Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set
+    to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and
+    `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass.
+
+    .. _*Attention is all you need*: https://arxiv.org/abs/1706.03762
+
+    """
+
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.__init__ with Bert->Roberta
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = RobertaEmbeddings(config)
+        self.encoder = RobertaEncoder(config)
+
+        self.pooler = RobertaPooler(config) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.forward
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        parser_att_mask=None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
+        r"""
+        encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        """
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = (
+            past_key_values[0][0].shape[2] if past_key_values is not None else 0
+        )
+
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                ((batch_size, seq_length + past_key_values_length)), device=device
+            )
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(
+                    batch_size, seq_length
+                )
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(
+                    input_shape, dtype=torch.long, device=device
+                )
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(
+            attention_mask, input_shape, device
+        )
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            (
+                encoder_batch_size,
+                encoder_sequence_length,
+                _,
+            ) = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(
+                encoder_attention_mask
+            )
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=parser_att_mask,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = (
+            self.pooler(sequence_output) if self.pooler is not None else None
+        )
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+class StructRoberta(RobertaPreTrainedModel):
+    _keys_to_ignore_on_save = [r"lm_head.decoder.weight", r"lm_head.decoder.bias"]
+    _keys_to_ignore_on_load_missing = [
+        r"position_ids",
+        r"lm_head.decoder.weight",
+        r"lm_head.decoder.bias",
+    ]
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        if config.is_decoder:
+            logger.warning(
+                "If you want to use `RobertaForMaskedLM` make sure `config.is_decoder=False` for "
+                "bi-directional self-attention."
+            )
+
+        
+        if config.n_cntxt_layers > 0:
+            config_cntxt = copy.deepcopy(config)
+            config_cntxt.num_hidden_layers = config.n_cntxt_layers
+            
+            self.cntxt_layers = RobertaModel(config_cntxt, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers_2 > 0:
+            self.parser_layers_1 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_1 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_1 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_1 = nn.Parameter(
+                torch.zeros((config.n_cntxt_layers_2, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_1.data.normal_(0, 0.1)
+
+            self._scaler_1 = nn.Parameter(torch.zeros(2))
+            
+            config_cntxt_2 = copy.deepcopy(config)
+            config_cntxt_2.num_hidden_layers = config.n_cntxt_layers_2
+            
+            self.cntxt_layers_2 = RobertaModel(config_cntxt_2, add_pooling_layer=False)
+            
+            
+            self.parser_layers_2 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_2 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_2 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_2 = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_2.data.normal_(0, 0.1)
+
+            self._scaler_2 = nn.Parameter(torch.zeros(2))
+            
+        else:
+            self.parser_layers = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(config.n_parser_layers)
+                ]
+            )
+
+            self.distance_ff = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight.data.normal_(0, 0.1)
+
+            self._scaler = nn.Parameter(torch.zeros(2))
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers > 0:
+            self.cntxt_layers.embeddings = self.roberta.embeddings
+        if config.n_cntxt_layers_2 > 0:
+            self.cntxt_layers_2.embeddings = self.roberta.embeddings
+        
+        self.lm_head = RobertaLMHead(config)
+
+        self.pad = config.pad_token_id
+
+        # The LM head weights require special treatment only when they are tied with the word embeddings
+        self.update_keys_to_ignore(config, ["lm_head.decoder.weight"])
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head.decoder = new_embeddings
+
+    @property
+    def scaler(self):
+        return self._scaler.exp()
+    
+    @property
+    def scaler_1(self):
+        return self._scaler_1.exp()
+    
+    @property
+    def scaler_2(self):
+        return self._scaler_2.exp()
+
+    @property
+    def rel_weight(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight, dim=-1)
+        
+    @property
+    def rel_weight_1(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_1)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_1, dim=-1)
+
+    
+    @property
+    def rel_weight_2(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_2)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_2, dim=-1)
+
+
+    def compute_block(self, distance, height, n_cntxt_layers=0):
+        """Compute constituents from distance and height."""
+
+        if n_cntxt_layers>0:
+            if n_cntxt_layers == 1:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_1[0]
+            elif n_cntxt_layers == 2:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_2[0]
+        else:
+            beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+        gamma = torch.sigmoid(-beta_logits)
+        ones = torch.ones_like(gamma)
+
+        block_mask_left = cummin(
+            gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1
+        )
+        block_mask_left = block_mask_left - F.pad(
+            block_mask_left[:, :, :-1], (1, 0), value=0
+        )
+        block_mask_left.tril_(0)
+
+        block_mask_right = cummin(
+            gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1
+        )
+        block_mask_right = block_mask_right - F.pad(
+            block_mask_right[:, :, 1:], (0, 1), value=0
+        )
+        block_mask_right.triu_(0)
+
+        block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+        block = cumsum(block_mask_left).tril(0) + cumsum(
+            block_mask_right, reverse=True
+        ).triu(1)
+
+        return block_p, block
+
+    def compute_head(self, height, n_cntxt_layers=0):
+        """Estimate head for each constituent."""
+
+        _, length = height.size()
+        if n_cntxt_layers>0:
+          if n_cntxt_layers == 1:
+              head_logits = height * self.scaler_1[1]
+          elif n_cntxt_layers == 2:
+              head_logits = height * self.scaler_2[1]  
+        else:
+            head_logits = height * self.scaler[1]
+        index = torch.arange(length, device=height.device)
+
+        mask = (index[:, None, None] <= index[None, None, :]) * (
+            index[None, None, :] <= index[None, :, None]
+        )
+        head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+        head_logits.masked_fill_(~mask[None, :, :, :], -1e9)
+
+        head_p = torch.softmax(head_logits, dim=-1)
+
+        return head_p
+
+    def parse(self, x, embs=None, n_cntxt_layers=0):
+        """Parse input sentence.
+
+        Args:
+        x: input tokens (required).
+        pos: position for each token (optional).
+        Returns:
+        distance: syntactic distance
+        height: syntactic height
+        """
+
+        mask = x != self.pad
+        mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+        if embs is None:
+            h = self.roberta.embeddings(x)
+        else:
+            h = embs
+        
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                parser_layers = self.parser_layers_1
+                height_ff = self.height_ff_1
+                distance_ff = self.distance_ff_1
+            elif n_cntxt_layers == 2:
+                parser_layers = self.parser_layers_2
+                height_ff = self.height_ff_2
+                distance_ff = self.distance_ff_2
+            for i in range(int(self.config.n_parser_layers/2)):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = parser_layers[i](h)
+
+            height = height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+        else:
+            for i in range(self.config.n_parser_layers):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = self.parser_layers[i](h)
+
+            height = self.height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = self.distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+
+        return distance, height
+
+    def generate_mask(self, x, distance, height, n_cntxt_layers=0):
+        """Compute head and cibling distribution for each token."""
+
+        bsz, length = x.size()
+
+        eye = torch.eye(length, device=x.device, dtype=torch.bool)
+        eye = eye[None, :, :].expand((bsz, -1, -1))
+
+        block_p, block = self.compute_block(distance, height, n_cntxt_layers=n_cntxt_layers)
+        head_p = self.compute_head(height, n_cntxt_layers=n_cntxt_layers)
+        head = torch.einsum("blij,bijh->blh", block_p, head_p)
+        head = head.masked_fill(eye, 0)
+        child = head.transpose(1, 2)
+        cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+        rel_list = []
+        if "head" in self.config.relations:
+            rel_list.append(head)
+        if "child" in self.config.relations:
+            rel_list.append(child)
+        if "cibling" in self.config.relations:
+            rel_list.append(cibling)
+
+        rel = torch.stack(rel_list, dim=1)
+
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                rel_weight = self.rel_weight_1
+            elif n_cntxt_layers == 2:
+                rel_weight = self.rel_weight_2
+        else:
+            rel_weight = self.rel_weight
+
+        dep = torch.einsum("lhr,brij->lbhij", rel_weight, rel)
+        
+        if n_cntxt_layers == 1:
+            num_layers = self.cntxt_layers_2.config.num_hidden_layers
+        else:
+            num_layers = self.roberta.config.num_hidden_layers
+            
+        att_mask = dep.reshape(
+            num_layers,
+            bsz,
+            self.config.num_attention_heads,
+            length,
+            length,
+        )
+
+        return att_mask, cibling, head, block
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, MaskedLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
+            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+        kwargs (`Dict[str, any]`, optional, defaults to *{}*):
+            Used to hide legacy arguments that have been deprecated.
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        
+        if self.config.n_cntxt_layers > 0:
+            cntxt_outputs = self.cntxt_layers(
+                input_ids,
+                attention_mask=attention_mask,
+                token_type_ids=token_type_ids,
+                position_ids=position_ids,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict)
+            
+        
+        if self.config.n_cntxt_layers_2 > 0:
+            distance_1, height_1 = self.parse(input_ids, cntxt_outputs[0], n_cntxt_layers=1)
+            att_mask_1, _, _, _ = self.generate_mask(input_ids, distance_1, height_1, n_cntxt_layers=1)
+            
+            cntxt_outputs_2 = self.cntxt_layers_2(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask_1)
+            
+            sequence_output = cntxt_outputs_2[0]
+            
+            distance_2, height_2 = self.parse(input_ids, sequence_output[0], n_cntxt_layers=2)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance_2, height_2, n_cntxt_layers=2)
+
+        elif self.config.n_cntxt_layers > 0:
+            distance, height = self.parse(input_ids, cntxt_outputs[0])
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+        else:
+            distance, height = self.parse(input_ids)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask,
+        )
+        sequence_output = outputs[0]
+        prediction_scores = self.lm_head(sequence_output)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(
+                prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return (
+                ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+            )
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class RobertaLMHead(nn.Module):
+    """Roberta Head for masked language modeling."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size)
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+        self.decoder.bias = self.bias
+
+    def forward(self, features, **kwargs):
+        x = self.dense(features)
+        x = gelu(x)
+        x = self.layer_norm(x)
+
+        # project back to size of vocabulary with bias
+        x = self.decoder(x)
+
+        return x
+
+    def _tie_weights(self):
+        # To tie those two weights if they get disconnected (on TPU or when the bias is resized)
+        self.bias = self.decoder.bias
+
+
+class StructRobertaForSequenceClassification(RobertaPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        if config.n_cntxt_layers > 0:
+            config_cntxt = copy.deepcopy(config)
+            config_cntxt.num_hidden_layers = config.n_cntxt_layers
+            
+            self.cntxt_layers = RobertaModel(config_cntxt, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers_2 > 0:
+            self.parser_layers_1 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_1 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_1 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_1 = nn.Parameter(
+                torch.zeros((config.n_cntxt_layers_2, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_1.data.normal_(0, 0.1)
+
+            self._scaler_1 = nn.Parameter(torch.zeros(2))
+            
+            config_cntxt_2 = copy.deepcopy(config)
+            config_cntxt_2.num_hidden_layers = config.n_cntxt_layers_2
+            
+            self.cntxt_layers_2 = RobertaModel(config_cntxt_2, add_pooling_layer=False)
+            
+            
+            self.parser_layers_2 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_2 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_2 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_2 = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_2.data.normal_(0, 0.1)
+
+            self._scaler_2 = nn.Parameter(torch.zeros(2))
+            
+        else:
+            self.parser_layers = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(config.n_parser_layers)
+                ]
+            )
+
+            self.distance_ff = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight.data.normal_(0, 0.1)
+
+            self._scaler = nn.Parameter(torch.zeros(2))
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers > 0:
+            self.cntxt_layers.embeddings = self.roberta.embeddings
+        if config.n_cntxt_layers_2 > 0:
+            self.cntxt_layers_2.embeddings = self.roberta.embeddings
+        
+        
+        self.pad = config.pad_token_id
+        self.classifier = RobertaClassificationHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    
+    @property
+    def scaler(self):
+        return self._scaler.exp()
+    
+    @property
+    def scaler_1(self):
+        return self._scaler_1.exp()
+    
+    @property
+    def scaler_2(self):
+        return self._scaler_2.exp()
+
+    @property
+    def rel_weight(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight, dim=-1)
+        
+    @property
+    def rel_weight_1(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_1)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_1, dim=-1)
+
+    
+    @property
+    def rel_weight_2(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_2)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_2, dim=-1)
+
+
+    def compute_block(self, distance, height, n_cntxt_layers=0):
+        """Compute constituents from distance and height."""
+
+        if n_cntxt_layers>0:
+            if n_cntxt_layers == 1:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_1[0]
+            elif n_cntxt_layers == 2:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_2[0]
+        else:
+            beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+        gamma = torch.sigmoid(-beta_logits)
+        ones = torch.ones_like(gamma)
+
+        block_mask_left = cummin(
+            gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1
+        )
+        block_mask_left = block_mask_left - F.pad(
+            block_mask_left[:, :, :-1], (1, 0), value=0
+        )
+        block_mask_left.tril_(0)
+
+        block_mask_right = cummin(
+            gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1
+        )
+        block_mask_right = block_mask_right - F.pad(
+            block_mask_right[:, :, 1:], (0, 1), value=0
+        )
+        block_mask_right.triu_(0)
+
+        block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+        block = cumsum(block_mask_left).tril(0) + cumsum(
+            block_mask_right, reverse=True
+        ).triu(1)
+
+        return block_p, block
+
+    def compute_head(self, height, n_cntxt_layers=0):
+        """Estimate head for each constituent."""
+
+        _, length = height.size()
+        if n_cntxt_layers>0:
+          if n_cntxt_layers == 1:
+              head_logits = height * self.scaler_1[1]
+          elif n_cntxt_layers == 2:
+              head_logits = height * self.scaler_2[1]  
+        else:
+            head_logits = height * self.scaler[1]
+        index = torch.arange(length, device=height.device)
+
+        mask = (index[:, None, None] <= index[None, None, :]) * (
+            index[None, None, :] <= index[None, :, None]
+        )
+        head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+        head_logits.masked_fill_(~mask[None, :, :, :], -1e9)
+
+        head_p = torch.softmax(head_logits, dim=-1)
+
+        return head_p
+
+    def parse(self, x, embs=None, n_cntxt_layers=0):
+        """Parse input sentence.
+
+        Args:
+        x: input tokens (required).
+        pos: position for each token (optional).
+        Returns:
+        distance: syntactic distance
+        height: syntactic height
+        """
+
+        mask = x != self.pad
+        mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+        if embs is None:
+            h = self.roberta.embeddings(x)
+        else:
+            h = embs
+        
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                parser_layers = self.parser_layers_1
+                height_ff = self.height_ff_1
+                distance_ff = self.distance_ff_1
+            elif n_cntxt_layers == 2:
+                parser_layers = self.parser_layers_2
+                height_ff = self.height_ff_2
+                distance_ff = self.distance_ff_2
+            for i in range(int(self.config.n_parser_layers/2)):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = parser_layers[i](h)
+
+            height = height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+        else:
+            for i in range(self.config.n_parser_layers):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = self.parser_layers[i](h)
+
+            height = self.height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = self.distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+
+        return distance, height
+
+    def generate_mask(self, x, distance, height, n_cntxt_layers=0):
+        """Compute head and cibling distribution for each token."""
+
+        bsz, length = x.size()
+
+        eye = torch.eye(length, device=x.device, dtype=torch.bool)
+        eye = eye[None, :, :].expand((bsz, -1, -1))
+
+        block_p, block = self.compute_block(distance, height, n_cntxt_layers=n_cntxt_layers)
+        head_p = self.compute_head(height, n_cntxt_layers=n_cntxt_layers)
+        head = torch.einsum("blij,bijh->blh", block_p, head_p)
+        head = head.masked_fill(eye, 0)
+        child = head.transpose(1, 2)
+        cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+        rel_list = []
+        if "head" in self.config.relations:
+            rel_list.append(head)
+        if "child" in self.config.relations:
+            rel_list.append(child)
+        if "cibling" in self.config.relations:
+            rel_list.append(cibling)
+
+        rel = torch.stack(rel_list, dim=1)
+
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                rel_weight = self.rel_weight_1
+            elif n_cntxt_layers == 2:
+                rel_weight = self.rel_weight_2
+        else:
+            rel_weight = self.rel_weight
+
+        dep = torch.einsum("lhr,brij->lbhij", rel_weight, rel)
+        
+        if n_cntxt_layers == 1:
+            num_layers = self.cntxt_layers_2.config.num_hidden_layers
+        else:
+            num_layers = self.roberta.config.num_hidden_layers
+            
+        att_mask = dep.reshape(
+            num_layers,
+            bsz,
+            self.config.num_attention_heads,
+            length,
+            length,
+        )
+
+        return att_mask, cibling, head, block
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        
+        if self.config.n_cntxt_layers > 0:
+            cntxt_outputs = self.cntxt_layers(
+                input_ids,
+                attention_mask=attention_mask,
+                token_type_ids=token_type_ids,
+                position_ids=position_ids,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict)
+            
+        
+        if self.config.n_cntxt_layers_2 > 0:
+            distance_1, height_1 = self.parse(input_ids, cntxt_outputs[0], n_cntxt_layers=1)
+            att_mask_1, _, _, _ = self.generate_mask(input_ids, distance_1, height_1, n_cntxt_layers=1)
+            
+            cntxt_outputs_2 = self.cntxt_layers_2(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask_1)
+            
+            sequence_output = cntxt_outputs_2[0]
+            
+            distance_2, height_2 = self.parse(input_ids, sequence_output[0], n_cntxt_layers=2)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance_2, height_2, n_cntxt_layers=2)
+
+        elif self.config.n_cntxt_layers > 0:
+            distance, height = self.parse(input_ids, cntxt_outputs[0])
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+        else:
+            distance, height = self.parse(input_ids)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask,
+        )
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class RobertaClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+def create_position_ids_from_input_ids(
+    input_ids, padding_idx, past_key_values_length=0
+):
+    """
+    Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
+    are ignored. This is modified from fairseq's `utils.make_positions`.
+
+    Args:
+        x: torch.Tensor x:
+
+    Returns: torch.Tensor
+    """
+    # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
+    mask = input_ids.ne(padding_idx).int()
+    incremental_indices = (
+        torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length
+    ) * mask
+    return incremental_indices.long() + padding_idx
+
+
+def cumprod(x, reverse=False, exclusive=False):
+    """cumulative product."""
+    if reverse:
+        x = x.flip([-1])
+
+    if exclusive:
+        x = F.pad(x[:, :, :-1], (1, 0), value=1)
+
+    cx = x.cumprod(-1)
+
+    if reverse:
+        cx = cx.flip([-1])
+    return cx
+
+
+def cumsum(x, reverse=False, exclusive=False):
+    """cumulative sum."""
+    bsz, _, length = x.size()
+    device = x.device
+    if reverse:
+        if exclusive:
+            w = torch.ones([bsz, length, length], device=device).tril(-1)
+        else:
+            w = torch.ones([bsz, length, length], device=device).tril(0)
+        cx = torch.bmm(x, w)
+    else:
+        if exclusive:
+            w = torch.ones([bsz, length, length], device=device).triu(1)
+        else:
+            w = torch.ones([bsz, length, length], device=device).triu(0)
+        cx = torch.bmm(x, w)
+    return cx
+
+
+def cummin(x, reverse=False, exclusive=False, max_value=1e9):
+    """cumulative min."""
+    if reverse:
+        if exclusive:
+            x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
+        x = x.flip([-1]).cummin(-1)[0].flip([-1])
+    else:
+        if exclusive:
+            x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
+        x = x.cummin(-1)[0]
+    return x

finetune/cola/checkpoint-400/optimizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:894d94568206a98ff7d62fb899d5439877e05449227f8a6c0e558525bfe3e464
+size 1154109317

finetune/cola/checkpoint-400/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec45d466ef3ab84ba3054a1d8517f7eb5329b61b2723f30fa84011c74b37b089
+size 577069633

finetune/cola/checkpoint-400/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f2b88a0416415a793212908cf2c110bbf133d3071d297a54461dc8554f1b61c
+size 14503

finetune/cola/checkpoint-400/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b09db5c9b9264697e44562c70af05a12a67180f480992d419fa287214c7be7a
+size 623

finetune/cola/checkpoint-400/special_tokens_map.json

@@ -0,0 +1,15 @@
+{
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "mask_token": {
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "unk_token": "<unk>"
+}

finetune/cola/checkpoint-400/tokenizer_config.json

@@ -0,0 +1,65 @@
+{
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "cls_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "mask_token": {
+    "__type": "AddedToken",
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "model_max_length": 512,
+  "name_or_path": "final_models/structroberta_sx2_final",
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "special_tokens_map_file": null,
+  "tokenizer_class": "RobertaTokenizer",
+  "trim_offsets": true,
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

finetune/cola/checkpoint-400/trainer_state.json

@@ -0,0 +1,27 @@
+{
+  "best_metric": 0.7994634473507714,
+  "best_model_checkpoint": "final_models/structroberta_sx2_final/finetune/cola/checkpoint-400",
+  "epoch": 5.797101449275362,
+  "global_step": 400,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 5.8,
+      "eval_accuracy": 0.7065750956535339,
+      "eval_f1": 0.7994634473507714,
+      "eval_loss": 0.7585899829864502,
+      "eval_mcc": 0.262551358979245,
+      "eval_runtime": 2.1393,
+      "eval_samples_per_second": 476.322,
+      "eval_steps_per_second": 59.832,
+      "step": 400
+    }
+  ],
+  "max_steps": 690,
+  "num_train_epochs": 10,
+  "total_flos": 4343859215677440.0,
+  "trial_name": null,
+  "trial_params": null
+}

finetune/cola/checkpoint-400/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:535606b0b8045695abdbfa17932190b45a57263b731642e3613fb33c791f7fa5
+size 3503

finetune/cola/config.json

@@ -0,0 +1,50 @@
+{
+  "_name_or_path": "final_models/structroberta_sx2_final",
+  "architectures": [
+    "StructRobertaForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "auto_map": {
+    "AutoConfig": "modeling_structroberta.StructRobertaConfig",
+    "AutoModelForMaskedLM": "modeling_structroberta.StructRoberta",
+    "AutoModelForSequenceClassification": "modeling_structroberta.StructRobertaForSequenceClassification"
+  },
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "conv_size": 9,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": 0,
+    "1": 1
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "label2id": {
+    "0": 0,
+    "1": 1
+  },
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "roberta",
+  "n_cntxt_layers": 4,
+  "n_cntxt_layers_2": 0,
+  "n_parser_layers": 6,
+  "num_attention_heads": 12,
+  "num_hidden_layers": 8,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "problem_type": "single_label_classification",
+  "relations": [
+    "head",
+    "child"
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.1",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 32000,
+  "weight_act": "softmax"
+}

finetune/cola/eval_results.json

@@ -0,0 +1,11 @@
+{
+    "epoch": 10.0,
+    "eval_accuracy": 0.7065750956535339,
+    "eval_f1": 0.7994634473507714,
+    "eval_loss": 0.7585899829864502,
+    "eval_mcc": 0.262551358979245,
+    "eval_runtime": 2.1721,
+    "eval_samples": 1019,
+    "eval_samples_per_second": 469.121,
+    "eval_steps_per_second": 58.928
+}

finetune/cola/modeling_structroberta.py

@@ -0,0 +1,2146 @@
+import copy
+import math
+from typing import List, Optional, Tuple, Union
+import torch
+import torch.utils.checkpoint
+from packaging import version
+from torch import nn
+import torch.nn.functional as F
+from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
+from transformers.activations import ACT2FN, gelu
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    MaskedLMOutput,
+    SequenceClassifierOutput
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.utils import logging
+from transformers import RobertaConfig
+
+
+logger = logging.get_logger(__name__)
+ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "roberta-base",
+    "roberta-large",
+    "roberta-large-mnli",
+    "distilroberta-base",
+    "roberta-base-openai-detector",
+    "roberta-large-openai-detector",
+    # See all RoBERTa models at https://huggingface.co/models?filter=roberta
+]
+
+
+class StructRobertaConfig(RobertaConfig):
+    model_type = "roberta"
+
+    def __init__(
+        self,
+        n_parser_layers=4,
+        conv_size=9,
+        relations=("head", "child"),
+        weight_act="softmax",
+        n_cntxt_layers=3,
+        n_cntxt_layers_2=0,
+        **kwargs,):
+        
+        super().__init__(**kwargs)
+        self.n_cntxt_layers = n_cntxt_layers
+        self.n_parser_layers = n_parser_layers
+        self.n_cntxt_layers_2 = n_cntxt_layers_2
+        self.conv_size = conv_size
+        self.relations = relations
+        self.weight_act = weight_act
+
+class Conv1d(nn.Module):
+    """1D convolution layer."""
+
+    def __init__(self, hidden_size, kernel_size, dilation=1):
+        """Initialization.
+
+        Args:
+        hidden_size: dimension of input embeddings
+        kernel_size: convolution kernel size
+        dilation: the spacing between the kernel points
+        """
+        super(Conv1d, self).__init__()
+
+        if kernel_size % 2 == 0:
+            padding = (kernel_size // 2) * dilation
+            self.shift = True
+        else:
+            padding = ((kernel_size - 1) // 2) * dilation
+            self.shift = False
+        self.conv = nn.Conv1d(
+            hidden_size, hidden_size, kernel_size, padding=padding, dilation=dilation
+        )
+
+    def forward(self, x):
+        """Compute convolution.
+
+        Args:
+          x: input embeddings
+        Returns:
+          conv_output: convolution results
+        """
+
+        if self.shift:
+            return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
+        else:
+            return self.conv(x.transpose(1, 2)).transpose(1, 2)
+
+
+class RobertaEmbeddings(nn.Module):
+    """
+    Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
+    """
+
+    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(
+            config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id
+        )
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.hidden_size
+        )
+        self.token_type_embeddings = nn.Embedding(
+            config.type_vocab_size, config.hidden_size
+        )
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.position_embedding_type = getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1))
+        )
+        if version.parse(torch.__version__) > version.parse("1.6.0"):
+            self.register_buffer(
+                "token_type_ids",
+                torch.zeros(self.position_ids.size(), dtype=torch.long),
+                persistent=False,
+            )
+
+        # End copy
+        self.padding_idx = config.pad_token_id
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings,
+            config.hidden_size,
+            padding_idx=self.padding_idx,
+        )
+
+    def forward(
+        self,
+        input_ids=None,
+        token_type_ids=None,
+        position_ids=None,
+        inputs_embeds=None,
+        past_key_values_length=0,
+    ):
+        if position_ids is None:
+            if input_ids is not None:
+                # Create the position ids from the input token ids. Any padded tokens remain padded.
+                position_ids = create_position_ids_from_input_ids(
+                    input_ids, self.padding_idx, past_key_values_length
+                )
+            else:
+                position_ids = self.create_position_ids_from_inputs_embeds(
+                    inputs_embeds
+                )
+
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(
+                    input_shape[0], seq_length
+                )
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(
+                    input_shape, dtype=torch.long, device=self.position_ids.device
+                )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+        """
+        We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
+
+        Args:
+            inputs_embeds: torch.Tensor
+
+        Returns: torch.Tensor
+        """
+        input_shape = inputs_embeds.size()[:-1]
+        sequence_length = input_shape[1]
+
+        position_ids = torch.arange(
+            self.padding_idx + 1,
+            sequence_length + self.padding_idx + 1,
+            dtype=torch.long,
+            device=inputs_embeds.device,
+        )
+        return position_ids.unsqueeze(0).expand(input_shape)
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfAttention with Bert->Roberta
+class RobertaSelfAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(
+            config, "embedding_size"
+        ):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = position_embedding_type or getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(
+                2 * config.max_position_embeddings - 1, self.attention_head_size
+            )
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (
+            self.num_attention_heads,
+            self.attention_head_size,
+        )
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        parser_att_mask=None,
+    ) -> Tuple[torch.Tensor]:
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(-1, 1)
+            position_ids_r = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(
+                distance + self.max_position_embeddings - 1
+            )
+            positional_embedding = positional_embedding.to(
+                dtype=query_layer.dtype
+            )  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                relative_position_scores_key = torch.einsum(
+                    "bhrd,lrd->bhlr", key_layer, positional_embedding
+                )
+                attention_scores = (
+                    attention_scores
+                    + relative_position_scores_query
+                    + relative_position_scores_key
+                )
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in RobertaModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        if parser_att_mask is None:
+            # Normalize the attention scores to probabilities.
+            attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+        else:
+            attention_probs = torch.sigmoid(attention_scores) * parser_att_mask
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (
+            (context_layer, attention_probs) if output_attentions else (context_layer,)
+        )
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput
+class RobertaSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(
+        self, hidden_states: torch.Tensor, input_tensor: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->Roberta
+class RobertaAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        self.self = RobertaSelfAttention(
+            config, position_embedding_type=position_embedding_type
+        )
+        self.output = RobertaSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads,
+            self.self.num_attention_heads,
+            self.self.attention_head_size,
+            self.pruned_heads,
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = (
+            self.self.attention_head_size * self.self.num_attention_heads
+        )
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        parser_att_mask=None,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+            parser_att_mask=parser_att_mask,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[
+            1:
+        ]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate
+class RobertaIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOutput
+class RobertaOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(
+        self, hidden_states: torch.Tensor, input_tensor: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertLayer with Bert->Roberta
+class RobertaLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = RobertaAttention(config)
+        self.is_decoder = config.is_decoder
+        self.add_cross_attention = config.add_cross_attention
+        if self.add_cross_attention:
+            if not self.is_decoder:
+                raise ValueError(
+                    f"{self} should be used as a decoder model if cross attention is added"
+                )
+            self.crossattention = RobertaAttention(
+                config, position_embedding_type="absolute"
+            )
+        self.intermediate = RobertaIntermediate(config)
+        self.output = RobertaOutput(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        parser_att_mask=None,
+    ) -> Tuple[torch.Tensor]:
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = (
+            past_key_value[:2] if past_key_value is not None else None
+        )
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+            parser_att_mask=parser_att_mask,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[
+                1:
+            ]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            if not hasattr(self, "crossattention"):
+                raise ValueError(
+                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
+                )
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = (
+                past_key_value[-2:] if past_key_value is not None else None
+            )
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                cross_attn_past_key_value,
+                output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = (
+                outputs + cross_attention_outputs[1:-1]
+            )  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk,
+            self.chunk_size_feed_forward,
+            self.seq_len_dim,
+            attention_output,
+        )
+        outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+# Copied from transformers.models.bert.modeling_bert.BertEncoder with Bert->Roberta
+class RobertaEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList(
+            [RobertaLayer(config) for _ in range(config.num_hidden_layers)]
+        )
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+        parser_att_mask=None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = (
+            () if output_attentions and self.config.add_cross_attention else None
+        )
+
+        next_decoder_cache = () if use_cache else None
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                if parser_att_mask is not None:
+                    layer_outputs = layer_module(
+                        hidden_states,
+                        attention_mask,
+                        layer_head_mask,
+                        encoder_hidden_states,
+                        encoder_attention_mask,
+                        past_key_value,
+                        output_attentions,
+                        parser_att_mask=parser_att_mask[i])
+                else:
+                    layer_outputs = layer_module(
+                        hidden_states,
+                        attention_mask,
+                        layer_head_mask,
+                        encoder_hidden_states,
+                        encoder_attention_mask,
+                        past_key_value,
+                        output_attentions,
+                        parser_att_mask=None)
+                    
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertPooler
+class RobertaPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class RobertaPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = RobertaConfig
+    base_model_prefix = "roberta"
+    supports_gradient_checkpointing = True
+
+    # Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            if module.bias is not None:
+                module.bias.data.zero_()
+                module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, RobertaEncoder):
+            module.gradient_checkpointing = value
+
+    def update_keys_to_ignore(self, config, del_keys_to_ignore):
+        """Remove some keys from ignore list"""
+        if not config.tie_word_embeddings:
+            # must make a new list, or the class variable gets modified!
+            self._keys_to_ignore_on_save = [
+                k for k in self._keys_to_ignore_on_save if k not in del_keys_to_ignore
+            ]
+            self._keys_to_ignore_on_load_missing = [
+                k
+                for k in self._keys_to_ignore_on_load_missing
+                if k not in del_keys_to_ignore
+            ]
+
+
+ROBERTA_START_DOCSTRING = r"""
+
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`RobertaConfig`]): Model configuration class with all the parameters of the
+            model. Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+ROBERTA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`RobertaTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class RobertaModel(RobertaPreTrainedModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in *Attention is
+    all you need*_ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
+    Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set
+    to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and
+    `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass.
+
+    .. _*Attention is all you need*: https://arxiv.org/abs/1706.03762
+
+    """
+
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.__init__ with Bert->Roberta
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = RobertaEmbeddings(config)
+        self.encoder = RobertaEncoder(config)
+
+        self.pooler = RobertaPooler(config) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.forward
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        parser_att_mask=None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
+        r"""
+        encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        """
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = (
+            past_key_values[0][0].shape[2] if past_key_values is not None else 0
+        )
+
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                ((batch_size, seq_length + past_key_values_length)), device=device
+            )
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(
+                    batch_size, seq_length
+                )
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(
+                    input_shape, dtype=torch.long, device=device
+                )
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(
+            attention_mask, input_shape, device
+        )
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            (
+                encoder_batch_size,
+                encoder_sequence_length,
+                _,
+            ) = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(
+                encoder_attention_mask
+            )
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=parser_att_mask,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = (
+            self.pooler(sequence_output) if self.pooler is not None else None
+        )
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+class StructRoberta(RobertaPreTrainedModel):
+    _keys_to_ignore_on_save = [r"lm_head.decoder.weight", r"lm_head.decoder.bias"]
+    _keys_to_ignore_on_load_missing = [
+        r"position_ids",
+        r"lm_head.decoder.weight",
+        r"lm_head.decoder.bias",
+    ]
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        if config.is_decoder:
+            logger.warning(
+                "If you want to use `RobertaForMaskedLM` make sure `config.is_decoder=False` for "
+                "bi-directional self-attention."
+            )
+
+        
+        if config.n_cntxt_layers > 0:
+            config_cntxt = copy.deepcopy(config)
+            config_cntxt.num_hidden_layers = config.n_cntxt_layers
+            
+            self.cntxt_layers = RobertaModel(config_cntxt, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers_2 > 0:
+            self.parser_layers_1 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_1 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_1 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_1 = nn.Parameter(
+                torch.zeros((config.n_cntxt_layers_2, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_1.data.normal_(0, 0.1)
+
+            self._scaler_1 = nn.Parameter(torch.zeros(2))
+            
+            config_cntxt_2 = copy.deepcopy(config)
+            config_cntxt_2.num_hidden_layers = config.n_cntxt_layers_2
+            
+            self.cntxt_layers_2 = RobertaModel(config_cntxt_2, add_pooling_layer=False)
+            
+            
+            self.parser_layers_2 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_2 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_2 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_2 = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_2.data.normal_(0, 0.1)
+
+            self._scaler_2 = nn.Parameter(torch.zeros(2))
+            
+        else:
+            self.parser_layers = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(config.n_parser_layers)
+                ]
+            )
+
+            self.distance_ff = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight.data.normal_(0, 0.1)
+
+            self._scaler = nn.Parameter(torch.zeros(2))
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers > 0:
+            self.cntxt_layers.embeddings = self.roberta.embeddings
+        if config.n_cntxt_layers_2 > 0:
+            self.cntxt_layers_2.embeddings = self.roberta.embeddings
+        
+        self.lm_head = RobertaLMHead(config)
+
+        self.pad = config.pad_token_id
+
+        # The LM head weights require special treatment only when they are tied with the word embeddings
+        self.update_keys_to_ignore(config, ["lm_head.decoder.weight"])
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head.decoder = new_embeddings
+
+    @property
+    def scaler(self):
+        return self._scaler.exp()
+    
+    @property
+    def scaler_1(self):
+        return self._scaler_1.exp()
+    
+    @property
+    def scaler_2(self):
+        return self._scaler_2.exp()
+
+    @property
+    def rel_weight(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight, dim=-1)
+        
+    @property
+    def rel_weight_1(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_1)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_1, dim=-1)
+
+    
+    @property
+    def rel_weight_2(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_2)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_2, dim=-1)
+
+
+    def compute_block(self, distance, height, n_cntxt_layers=0):
+        """Compute constituents from distance and height."""
+
+        if n_cntxt_layers>0:
+            if n_cntxt_layers == 1:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_1[0]
+            elif n_cntxt_layers == 2:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_2[0]
+        else:
+            beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+        gamma = torch.sigmoid(-beta_logits)
+        ones = torch.ones_like(gamma)
+
+        block_mask_left = cummin(
+            gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1
+        )
+        block_mask_left = block_mask_left - F.pad(
+            block_mask_left[:, :, :-1], (1, 0), value=0
+        )
+        block_mask_left.tril_(0)
+
+        block_mask_right = cummin(
+            gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1
+        )
+        block_mask_right = block_mask_right - F.pad(
+            block_mask_right[:, :, 1:], (0, 1), value=0
+        )
+        block_mask_right.triu_(0)
+
+        block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+        block = cumsum(block_mask_left).tril(0) + cumsum(
+            block_mask_right, reverse=True
+        ).triu(1)
+
+        return block_p, block
+
+    def compute_head(self, height, n_cntxt_layers=0):
+        """Estimate head for each constituent."""
+
+        _, length = height.size()
+        if n_cntxt_layers>0:
+          if n_cntxt_layers == 1:
+              head_logits = height * self.scaler_1[1]
+          elif n_cntxt_layers == 2:
+              head_logits = height * self.scaler_2[1]  
+        else:
+            head_logits = height * self.scaler[1]
+        index = torch.arange(length, device=height.device)
+
+        mask = (index[:, None, None] <= index[None, None, :]) * (
+            index[None, None, :] <= index[None, :, None]
+        )
+        head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+        head_logits.masked_fill_(~mask[None, :, :, :], -1e9)
+
+        head_p = torch.softmax(head_logits, dim=-1)
+
+        return head_p
+
+    def parse(self, x, embs=None, n_cntxt_layers=0):
+        """Parse input sentence.
+
+        Args:
+        x: input tokens (required).
+        pos: position for each token (optional).
+        Returns:
+        distance: syntactic distance
+        height: syntactic height
+        """
+
+        mask = x != self.pad
+        mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+        if embs is None:
+            h = self.roberta.embeddings(x)
+        else:
+            h = embs
+        
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                parser_layers = self.parser_layers_1
+                height_ff = self.height_ff_1
+                distance_ff = self.distance_ff_1
+            elif n_cntxt_layers == 2:
+                parser_layers = self.parser_layers_2
+                height_ff = self.height_ff_2
+                distance_ff = self.distance_ff_2
+            for i in range(int(self.config.n_parser_layers/2)):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = parser_layers[i](h)
+
+            height = height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+        else:
+            for i in range(self.config.n_parser_layers):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = self.parser_layers[i](h)
+
+            height = self.height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = self.distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+
+        return distance, height
+
+    def generate_mask(self, x, distance, height, n_cntxt_layers=0):
+        """Compute head and cibling distribution for each token."""
+
+        bsz, length = x.size()
+
+        eye = torch.eye(length, device=x.device, dtype=torch.bool)
+        eye = eye[None, :, :].expand((bsz, -1, -1))
+
+        block_p, block = self.compute_block(distance, height, n_cntxt_layers=n_cntxt_layers)
+        head_p = self.compute_head(height, n_cntxt_layers=n_cntxt_layers)
+        head = torch.einsum("blij,bijh->blh", block_p, head_p)
+        head = head.masked_fill(eye, 0)
+        child = head.transpose(1, 2)
+        cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+        rel_list = []
+        if "head" in self.config.relations:
+            rel_list.append(head)
+        if "child" in self.config.relations:
+            rel_list.append(child)
+        if "cibling" in self.config.relations:
+            rel_list.append(cibling)
+
+        rel = torch.stack(rel_list, dim=1)
+
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                rel_weight = self.rel_weight_1
+            elif n_cntxt_layers == 2:
+                rel_weight = self.rel_weight_2
+        else:
+            rel_weight = self.rel_weight
+
+        dep = torch.einsum("lhr,brij->lbhij", rel_weight, rel)
+        
+        if n_cntxt_layers == 1:
+            num_layers = self.cntxt_layers_2.config.num_hidden_layers
+        else:
+            num_layers = self.roberta.config.num_hidden_layers
+            
+        att_mask = dep.reshape(
+            num_layers,
+            bsz,
+            self.config.num_attention_heads,
+            length,
+            length,
+        )
+
+        return att_mask, cibling, head, block
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, MaskedLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
+            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+        kwargs (`Dict[str, any]`, optional, defaults to *{}*):
+            Used to hide legacy arguments that have been deprecated.
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        
+        if self.config.n_cntxt_layers > 0:
+            cntxt_outputs = self.cntxt_layers(
+                input_ids,
+                attention_mask=attention_mask,
+                token_type_ids=token_type_ids,
+                position_ids=position_ids,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict)
+            
+        
+        if self.config.n_cntxt_layers_2 > 0:
+            distance_1, height_1 = self.parse(input_ids, cntxt_outputs[0], n_cntxt_layers=1)
+            att_mask_1, _, _, _ = self.generate_mask(input_ids, distance_1, height_1, n_cntxt_layers=1)
+            
+            cntxt_outputs_2 = self.cntxt_layers_2(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask_1)
+            
+            sequence_output = cntxt_outputs_2[0]
+            
+            distance_2, height_2 = self.parse(input_ids, sequence_output[0], n_cntxt_layers=2)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance_2, height_2, n_cntxt_layers=2)
+
+        elif self.config.n_cntxt_layers > 0:
+            distance, height = self.parse(input_ids, cntxt_outputs[0])
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+        else:
+            distance, height = self.parse(input_ids)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask,
+        )
+        sequence_output = outputs[0]
+        prediction_scores = self.lm_head(sequence_output)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(
+                prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return (
+                ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+            )
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class RobertaLMHead(nn.Module):
+    """Roberta Head for masked language modeling."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size)
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+        self.decoder.bias = self.bias
+
+    def forward(self, features, **kwargs):
+        x = self.dense(features)
+        x = gelu(x)
+        x = self.layer_norm(x)
+
+        # project back to size of vocabulary with bias
+        x = self.decoder(x)
+
+        return x
+
+    def _tie_weights(self):
+        # To tie those two weights if they get disconnected (on TPU or when the bias is resized)
+        self.bias = self.decoder.bias
+
+
+class StructRobertaForSequenceClassification(RobertaPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        if config.n_cntxt_layers > 0:
+            config_cntxt = copy.deepcopy(config)
+            config_cntxt.num_hidden_layers = config.n_cntxt_layers
+            
+            self.cntxt_layers = RobertaModel(config_cntxt, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers_2 > 0:
+            self.parser_layers_1 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_1 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_1 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_1 = nn.Parameter(
+                torch.zeros((config.n_cntxt_layers_2, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_1.data.normal_(0, 0.1)
+
+            self._scaler_1 = nn.Parameter(torch.zeros(2))
+            
+            config_cntxt_2 = copy.deepcopy(config)
+            config_cntxt_2.num_hidden_layers = config.n_cntxt_layers_2
+            
+            self.cntxt_layers_2 = RobertaModel(config_cntxt_2, add_pooling_layer=False)
+            
+            
+            self.parser_layers_2 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_2 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_2 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_2 = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_2.data.normal_(0, 0.1)
+
+            self._scaler_2 = nn.Parameter(torch.zeros(2))
+            
+        else:
+            self.parser_layers = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(config.n_parser_layers)
+                ]
+            )
+
+            self.distance_ff = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight.data.normal_(0, 0.1)
+
+            self._scaler = nn.Parameter(torch.zeros(2))
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers > 0:
+            self.cntxt_layers.embeddings = self.roberta.embeddings
+        if config.n_cntxt_layers_2 > 0:
+            self.cntxt_layers_2.embeddings = self.roberta.embeddings
+        
+        
+        self.pad = config.pad_token_id
+        self.classifier = RobertaClassificationHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    
+    @property
+    def scaler(self):
+        return self._scaler.exp()
+    
+    @property
+    def scaler_1(self):
+        return self._scaler_1.exp()
+    
+    @property
+    def scaler_2(self):
+        return self._scaler_2.exp()
+
+    @property
+    def rel_weight(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight, dim=-1)
+        
+    @property
+    def rel_weight_1(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_1)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_1, dim=-1)
+
+    
+    @property
+    def rel_weight_2(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_2)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_2, dim=-1)
+
+
+    def compute_block(self, distance, height, n_cntxt_layers=0):
+        """Compute constituents from distance and height."""
+
+        if n_cntxt_layers>0:
+            if n_cntxt_layers == 1:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_1[0]
+            elif n_cntxt_layers == 2:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_2[0]
+        else:
+            beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+        gamma = torch.sigmoid(-beta_logits)
+        ones = torch.ones_like(gamma)
+
+        block_mask_left = cummin(
+            gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1
+        )
+        block_mask_left = block_mask_left - F.pad(
+            block_mask_left[:, :, :-1], (1, 0), value=0
+        )
+        block_mask_left.tril_(0)
+
+        block_mask_right = cummin(
+            gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1
+        )
+        block_mask_right = block_mask_right - F.pad(
+            block_mask_right[:, :, 1:], (0, 1), value=0
+        )
+        block_mask_right.triu_(0)
+
+        block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+        block = cumsum(block_mask_left).tril(0) + cumsum(
+            block_mask_right, reverse=True
+        ).triu(1)
+
+        return block_p, block
+
+    def compute_head(self, height, n_cntxt_layers=0):
+        """Estimate head for each constituent."""
+
+        _, length = height.size()
+        if n_cntxt_layers>0:
+          if n_cntxt_layers == 1:
+              head_logits = height * self.scaler_1[1]
+          elif n_cntxt_layers == 2:
+              head_logits = height * self.scaler_2[1]  
+        else:
+            head_logits = height * self.scaler[1]
+        index = torch.arange(length, device=height.device)
+
+        mask = (index[:, None, None] <= index[None, None, :]) * (
+            index[None, None, :] <= index[None, :, None]
+        )
+        head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+        head_logits.masked_fill_(~mask[None, :, :, :], -1e9)
+
+        head_p = torch.softmax(head_logits, dim=-1)
+
+        return head_p
+
+    def parse(self, x, embs=None, n_cntxt_layers=0):
+        """Parse input sentence.
+
+        Args:
+        x: input tokens (required).
+        pos: position for each token (optional).
+        Returns:
+        distance: syntactic distance
+        height: syntactic height
+        """
+
+        mask = x != self.pad
+        mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+        if embs is None:
+            h = self.roberta.embeddings(x)
+        else:
+            h = embs
+        
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                parser_layers = self.parser_layers_1
+                height_ff = self.height_ff_1
+                distance_ff = self.distance_ff_1
+            elif n_cntxt_layers == 2:
+                parser_layers = self.parser_layers_2
+                height_ff = self.height_ff_2
+                distance_ff = self.distance_ff_2
+            for i in range(int(self.config.n_parser_layers/2)):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = parser_layers[i](h)
+
+            height = height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+        else:
+            for i in range(self.config.n_parser_layers):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = self.parser_layers[i](h)
+
+            height = self.height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = self.distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+
+        return distance, height
+
+    def generate_mask(self, x, distance, height, n_cntxt_layers=0):
+        """Compute head and cibling distribution for each token."""
+
+        bsz, length = x.size()
+
+        eye = torch.eye(length, device=x.device, dtype=torch.bool)
+        eye = eye[None, :, :].expand((bsz, -1, -1))
+
+        block_p, block = self.compute_block(distance, height, n_cntxt_layers=n_cntxt_layers)
+        head_p = self.compute_head(height, n_cntxt_layers=n_cntxt_layers)
+        head = torch.einsum("blij,bijh->blh", block_p, head_p)
+        head = head.masked_fill(eye, 0)
+        child = head.transpose(1, 2)
+        cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+        rel_list = []
+        if "head" in self.config.relations:
+            rel_list.append(head)
+        if "child" in self.config.relations:
+            rel_list.append(child)
+        if "cibling" in self.config.relations:
+            rel_list.append(cibling)
+
+        rel = torch.stack(rel_list, dim=1)
+
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                rel_weight = self.rel_weight_1
+            elif n_cntxt_layers == 2:
+                rel_weight = self.rel_weight_2
+        else:
+            rel_weight = self.rel_weight
+
+        dep = torch.einsum("lhr,brij->lbhij", rel_weight, rel)
+        
+        if n_cntxt_layers == 1:
+            num_layers = self.cntxt_layers_2.config.num_hidden_layers
+        else:
+            num_layers = self.roberta.config.num_hidden_layers
+            
+        att_mask = dep.reshape(
+            num_layers,
+            bsz,
+            self.config.num_attention_heads,
+            length,
+            length,
+        )
+
+        return att_mask, cibling, head, block
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        
+        if self.config.n_cntxt_layers > 0:
+            cntxt_outputs = self.cntxt_layers(
+                input_ids,
+                attention_mask=attention_mask,
+                token_type_ids=token_type_ids,
+                position_ids=position_ids,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict)
+            
+        
+        if self.config.n_cntxt_layers_2 > 0:
+            distance_1, height_1 = self.parse(input_ids, cntxt_outputs[0], n_cntxt_layers=1)
+            att_mask_1, _, _, _ = self.generate_mask(input_ids, distance_1, height_1, n_cntxt_layers=1)
+            
+            cntxt_outputs_2 = self.cntxt_layers_2(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask_1)
+            
+            sequence_output = cntxt_outputs_2[0]
+            
+            distance_2, height_2 = self.parse(input_ids, sequence_output[0], n_cntxt_layers=2)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance_2, height_2, n_cntxt_layers=2)
+
+        elif self.config.n_cntxt_layers > 0:
+            distance, height = self.parse(input_ids, cntxt_outputs[0])
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+        else:
+            distance, height = self.parse(input_ids)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask,
+        )
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class RobertaClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+def create_position_ids_from_input_ids(
+    input_ids, padding_idx, past_key_values_length=0
+):
+    """
+    Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
+    are ignored. This is modified from fairseq's `utils.make_positions`.
+
+    Args:
+        x: torch.Tensor x:
+
+    Returns: torch.Tensor
+    """
+    # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
+    mask = input_ids.ne(padding_idx).int()
+    incremental_indices = (
+        torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length
+    ) * mask
+    return incremental_indices.long() + padding_idx
+
+
+def cumprod(x, reverse=False, exclusive=False):
+    """cumulative product."""
+    if reverse:
+        x = x.flip([-1])
+
+    if exclusive:
+        x = F.pad(x[:, :, :-1], (1, 0), value=1)
+
+    cx = x.cumprod(-1)
+
+    if reverse:
+        cx = cx.flip([-1])
+    return cx
+
+
+def cumsum(x, reverse=False, exclusive=False):
+    """cumulative sum."""
+    bsz, _, length = x.size()
+    device = x.device
+    if reverse:
+        if exclusive:
+            w = torch.ones([bsz, length, length], device=device).tril(-1)
+        else:
+            w = torch.ones([bsz, length, length], device=device).tril(0)
+        cx = torch.bmm(x, w)
+    else:
+        if exclusive:
+            w = torch.ones([bsz, length, length], device=device).triu(1)
+        else:
+            w = torch.ones([bsz, length, length], device=device).triu(0)
+        cx = torch.bmm(x, w)
+    return cx
+
+
+def cummin(x, reverse=False, exclusive=False, max_value=1e9):
+    """cumulative min."""
+    if reverse:
+        if exclusive:
+            x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
+        x = x.flip([-1]).cummin(-1)[0].flip([-1])
+    else:
+        if exclusive:
+            x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
+        x = x.cummin(-1)[0]
+    return x

finetune/cola/predict_results.txt

@@ -0,0 +1,1020 @@
+index	prediction
+0	1
+1	1
+2	0
+3	1
+4	0
+5	0
+6	0
+7	1
+8	1
+9	1
+10	0
+11	1
+12	1
+13	0
+14	1
+15	1
+16	1
+17	1
+18	1
+19	1
+20	1
+21	1
+22	1
+23	1
+24	1
+25	1
+26	1
+27	1
+28	1
+29	1
+30	1
+31	0
+32	1
+33	1
+34	1
+35	1
+36	1
+37	1
+38	1
+39	0
+40	1
+41	0
+42	1
+43	1
+44	0
+45	0
+46	1
+47	0
+48	1
+49	1
+50	1
+51	1
+52	1
+53	0
+54	1
+55	1
+56	1
+57	0
+58	1
+59	0
+60	0
+61	1
+62	1
+63	1
+64	1
+65	0
+66	1
+67	0
+68	0
+69	1
+70	1
+71	1
+72	1
+73	0
+74	1
+75	0
+76	1
+77	1
+78	1
+79	1
+80	0
+81	1
+82	1
+83	0
+84	1
+85	0
+86	0
+87	1
+88	1
+89	1
+90	1
+91	1
+92	1
+93	1
+94	1
+95	1
+96	0
+97	1
+98	1
+99	1
+100	1
+101	0
+102	1
+103	1
+104	1
+105	0
+106	1
+107	1
+108	0
+109	0
+110	1
+111	0
+112	1
+113	1
+114	0
+115	1
+116	0
+117	1
+118	0
+119	1
+120	1
+121	1
+122	1
+123	1
+124	1
+125	1
+126	1
+127	0
+128	0
+129	1
+130	1
+131	1
+132	1
+133	1
+134	1
+135	1
+136	1
+137	1
+138	1
+139	1
+140	1
+141	1
+142	1
+143	1
+144	1
+145	1
+146	1
+147	1
+148	1
+149	1
+150	0
+151	0
+152	0
+153	1
+154	1
+155	1
+156	1
+157	1
+158	0
+159	0
+160	1
+161	1
+162	1
+163	1
+164	1
+165	1
+166	0
+167	1
+168	0
+169	1
+170	0
+171	0
+172	1
+173	0
+174	0
+175	1
+176	1
+177	1
+178	1
+179	1
+180	0
+181	1
+182	1
+183	1
+184	1
+185	1
+186	1
+187	0
+188	0
+189	1
+190	1
+191	1
+192	0
+193	1
+194	1
+195	1
+196	1
+197	0
+198	1
+199	1
+200	1
+201	1
+202	1
+203	1
+204	1
+205	1
+206	1
+207	1
+208	1
+209	1
+210	1
+211	1
+212	1
+213	0
+214	1
+215	1
+216	0
+217	1
+218	1
+219	1
+220	0
+221	0
+222	1
+223	1
+224	1
+225	0
+226	1
+227	1
+228	1
+229	1
+230	1
+231	1
+232	0
+233	1
+234	1
+235	1
+236	1
+237	1
+238	1
+239	1
+240	0
+241	1
+242	1
+243	0
+244	1
+245	1
+246	1
+247	1
+248	0
+249	1
+250	1
+251	1
+252	1
+253	1
+254	0
+255	0
+256	1
+257	0
+258	1
+259	1
+260	1
+261	1
+262	0
+263	0
+264	0
+265	0
+266	0
+267	1
+268	0
+269	0
+270	1
+271	1
+272	1
+273	1
+274	1
+275	1
+276	0
+277	1
+278	1
+279	0
+280	1
+281	1
+282	1
+283	0
+284	1
+285	1
+286	0
+287	1
+288	1
+289	1
+290	1
+291	1
+292	1
+293	0
+294	1
+295	1
+296	0
+297	1
+298	1
+299	1
+300	0
+301	1
+302	1
+303	1
+304	0
+305	1
+306	1
+307	1
+308	1
+309	1
+310	1
+311	1
+312	1
+313	1
+314	1
+315	1
+316	1
+317	1
+318	1
+319	1
+320	0
+321	1
+322	0
+323	0
+324	0
+325	1
+326	1
+327	1
+328	0
+329	1
+330	1
+331	1
+332	1
+333	1
+334	1
+335	1
+336	1
+337	1
+338	0
+339	0
+340	1
+341	1
+342	1
+343	1
+344	1
+345	0
+346	1
+347	1
+348	1
+349	1
+350	1
+351	1
+352	1
+353	1
+354	1
+355	1
+356	0
+357	1
+358	1
+359	1
+360	1
+361	1
+362	1
+363	1
+364	1
+365	0
+366	1
+367	1
+368	1
+369	1
+370	1
+371	1
+372	1
+373	1
+374	1
+375	1
+376	1
+377	1
+378	0
+379	1
+380	1
+381	1
+382	1
+383	0
+384	1
+385	1
+386	1
+387	1
+388	1
+389	1
+390	1
+391	1
+392	0
+393	1
+394	1
+395	1
+396	1
+397	0
+398	1
+399	1
+400	1
+401	1
+402	0
+403	1
+404	1
+405	1
+406	1
+407	1
+408	1
+409	0
+410	1
+411	1
+412	1
+413	1
+414	1
+415	1
+416	1
+417	1
+418	1
+419	1
+420	1
+421	1
+422	1
+423	1
+424	0
+425	1
+426	0
+427	1
+428	0
+429	1
+430	0
+431	0
+432	1
+433	1
+434	1
+435	1
+436	1
+437	1
+438	0
+439	1
+440	1
+441	1
+442	1
+443	1
+444	1
+445	0
+446	1
+447	0
+448	0
+449	1
+450	0
+451	0
+452	0
+453	0
+454	1
+455	1
+456	1
+457	1
+458	1
+459	0
+460	1
+461	1
+462	0
+463	1
+464	1
+465	0
+466	1
+467	0
+468	1
+469	1
+470	1
+471	1
+472	0
+473	1
+474	0
+475	1
+476	1
+477	1
+478	0
+479	1
+480	0
+481	0
+482	1
+483	0
+484	0
+485	1
+486	0
+487	1
+488	0
+489	1
+490	1
+491	0
+492	1
+493	0
+494	0
+495	0
+496	0
+497	0
+498	1
+499	1
+500	1
+501	1
+502	1
+503	1
+504	0
+505	0
+506	0
+507	0
+508	0
+509	1
+510	1
+511	1
+512	1
+513	0
+514	1
+515	1
+516	0
+517	0
+518	1
+519	1
+520	1
+521	1
+522	1
+523	1
+524	1
+525	1
+526	0
+527	0
+528	0
+529	0
+530	0
+531	0
+532	1
+533	0
+534	1
+535	1
+536	0
+537	0
+538	0
+539	1
+540	0
+541	1
+542	1
+543	1
+544	1
+545	0
+546	1
+547	1
+548	1
+549	1
+550	1
+551	0
+552	1
+553	1
+554	1
+555	1
+556	1
+557	1
+558	1
+559	1
+560	1
+561	1
+562	1
+563	1
+564	0
+565	1
+566	1
+567	1
+568	1
+569	1
+570	1
+571	1
+572	1
+573	1
+574	1
+575	1
+576	1
+577	1
+578	1
+579	1
+580	1
+581	1
+582	1
+583	1
+584	1
+585	1
+586	1
+587	1
+588	1
+589	1
+590	1
+591	1
+592	1
+593	1
+594	1
+595	1
+596	1
+597	1
+598	1
+599	1
+600	0
+601	1
+602	1
+603	1
+604	1
+605	1
+606	1
+607	1
+608	1
+609	1
+610	1
+611	1
+612	1
+613	1
+614	1
+615	1
+616	0
+617	1
+618	1
+619	1
+620	1
+621	0
+622	1
+623	0
+624	1
+625	1
+626	1
+627	0
+628	1
+629	1
+630	1
+631	1
+632	1
+633	1
+634	1
+635	1
+636	1
+637	1
+638	0
+639	1
+640	1
+641	1
+642	0
+643	1
+644	1
+645	1
+646	0
+647	1
+648	0
+649	0
+650	0
+651	0
+652	0
+653	1
+654	0
+655	0
+656	1
+657	0
+658	1
+659	1
+660	1
+661	1
+662	1
+663	0
+664	1
+665	0
+666	1
+667	1
+668	1
+669	1
+670	1
+671	0
+672	1
+673	0
+674	1
+675	1
+676	1
+677	1
+678	0
+679	0
+680	1
+681	1
+682	0
+683	1
+684	1
+685	1
+686	1
+687	1
+688	1
+689	1
+690	1
+691	1
+692	0
+693	1
+694	1
+695	0
+696	0
+697	0
+698	0
+699	1
+700	1
+701	0
+702	0
+703	0
+704	0
+705	1
+706	0
+707	0
+708	0
+709	0
+710	1
+711	1
+712	1
+713	1
+714	0
+715	0
+716	1
+717	1
+718	0
+719	1
+720	0
+721	1
+722	1
+723	0
+724	0
+725	0
+726	1
+727	1
+728	1
+729	0
+730	1
+731	1
+732	1
+733	1
+734	1
+735	0
+736	1
+737	0
+738	1
+739	0
+740	1
+741	0
+742	1
+743	1
+744	1
+745	1
+746	1
+747	1
+748	1
+749	1
+750	1
+751	1
+752	1
+753	1
+754	1
+755	1
+756	0
+757	1
+758	1
+759	0
+760	1
+761	0
+762	1
+763	0
+764	1
+765	0
+766	0
+767	1
+768	1
+769	1
+770	1
+771	1
+772	1
+773	1
+774	1
+775	1
+776	1
+777	1
+778	1
+779	1
+780	1
+781	1
+782	1
+783	1
+784	1
+785	1
+786	0
+787	0
+788	1
+789	1
+790	1
+791	1
+792	1
+793	1
+794	0
+795	0
+796	0
+797	1
+798	1
+799	1
+800	1
+801	1
+802	1
+803	1
+804	1
+805	1
+806	1
+807	1
+808	1
+809	1
+810	1
+811	1
+812	1
+813	0
+814	0
+815	0
+816	1
+817	0
+818	0
+819	1
+820	1
+821	1
+822	1
+823	1
+824	1
+825	1
+826	1
+827	1
+828	0
+829	1
+830	1
+831	1
+832	1
+833	1
+834	0
+835	1
+836	1
+837	1
+838	1
+839	1
+840	1
+841	1
+842	1
+843	1
+844	1
+845	0
+846	1
+847	1
+848	1
+849	1
+850	1
+851	1
+852	1
+853	1
+854	1
+855	1
+856	1
+857	1
+858	1
+859	1
+860	1
+861	1
+862	1
+863	1
+864	0
+865	1
+866	1
+867	0
+868	1
+869	1
+870	1
+871	1
+872	1
+873	1
+874	1
+875	1
+876	1
+877	1
+878	0
+879	1
+880	0
+881	1
+882	1
+883	1
+884	1
+885	1
+886	1
+887	1
+888	1
+889	1
+890	1
+891	1
+892	1
+893	1
+894	1
+895	1
+896	1
+897	1
+898	0
+899	1
+900	1
+901	1
+902	1
+903	1
+904	1
+905	1
+906	1
+907	1
+908	1
+909	1
+910	1
+911	1
+912	1
+913	1
+914	0
+915	0
+916	1
+917	1
+918	1
+919	1
+920	1
+921	1
+922	1
+923	1
+924	1
+925	1
+926	1
+927	1
+928	1
+929	1
+930	1
+931	1
+932	1
+933	1
+934	1
+935	1
+936	1
+937	1
+938	1
+939	1
+940	1
+941	1
+942	1
+943	0
+944	1
+945	1
+946	1
+947	1
+948	1
+949	0
+950	0
+951	1
+952	1
+953	1
+954	1
+955	1
+956	1
+957	1
+958	1
+959	1
+960	0
+961	0
+962	1
+963	1
+964	1
+965	1
+966	1
+967	1
+968	1
+969	1
+970	1
+971	0
+972	1
+973	1
+974	1
+975	1
+976	1
+977	1
+978	0
+979	1
+980	1
+981	1
+982	1
+983	1
+984	1
+985	1
+986	1
+987	1
+988	1
+989	1
+990	1
+991	1
+992	0
+993	1
+994	1
+995	0
+996	1
+997	1
+998	1
+999	1
+1000	1
+1001	1
+1002	1
+1003	0
+1004	1
+1005	0
+1006	1
+1007	1
+1008	0
+1009	0
+1010	1
+1011	1
+1012	1
+1013	1
+1014	1
+1015	1
+1016	1
+1017	1
+1018	1

finetune/cola/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec45d466ef3ab84ba3054a1d8517f7eb5329b61b2723f30fa84011c74b37b089
+size 577069633

finetune/cola/special_tokens_map.json

@@ -0,0 +1,15 @@
+{
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "mask_token": {
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "unk_token": "<unk>"
+}

finetune/cola/tokenizer_config.json

@@ -0,0 +1,65 @@
+{
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "cls_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "mask_token": {
+    "__type": "AddedToken",
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "model_max_length": 512,
+  "name_or_path": "final_models/structroberta_sx2_final",
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "special_tokens_map_file": null,
+  "tokenizer_class": "RobertaTokenizer",
+  "trim_offsets": true,
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

finetune/cola/train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 10.0,
+    "train_loss": 0.3363707777382671,
+    "train_runtime": 371.6095,
+    "train_samples": 8164,
+    "train_samples_per_second": 219.693,
+    "train_steps_per_second": 1.857
+}

finetune/cola/trainer_state.json

@@ -0,0 +1,42 @@
+{
+  "best_metric": 0.7994634473507714,
+  "best_model_checkpoint": "final_models/structroberta_sx2_final/finetune/cola/checkpoint-400",
+  "epoch": 10.0,
+  "global_step": 690,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 5.8,
+      "eval_accuracy": 0.7065750956535339,
+      "eval_f1": 0.7994634473507714,
+      "eval_loss": 0.7585899829864502,
+      "eval_mcc": 0.262551358979245,
+      "eval_runtime": 2.1393,
+      "eval_samples_per_second": 476.322,
+      "eval_steps_per_second": 59.832,
+      "step": 400
+    },
+    {
+      "epoch": 7.25,
+      "learning_rate": 1.3768115942028985e-05,
+      "loss": 0.4074,
+      "step": 500
+    },
+    {
+      "epoch": 10.0,
+      "step": 690,
+      "total_flos": 7478546317332480.0,
+      "train_loss": 0.3363707777382671,
+      "train_runtime": 371.6095,
+      "train_samples_per_second": 219.693,
+      "train_steps_per_second": 1.857
+    }
+  ],
+  "max_steps": 690,
+  "num_train_epochs": 10,
+  "total_flos": 7478546317332480.0,
+  "trial_name": null,
+  "trial_params": null
+}

finetune/cola/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:535606b0b8045695abdbfa17932190b45a57263b731642e3613fb33c791f7fa5
+size 3503

finetune/control_raising_control/all_results.json

@@ -0,0 +1,16 @@
+{
+    "epoch": 10.0,
+    "eval_accuracy": 0.8752802014350891,
+    "eval_f1": 0.8866091446429785,
+    "eval_loss": 0.801540732383728,
+    "eval_mcc": 0.7701394030930233,
+    "eval_runtime": 28.0068,
+    "eval_samples": 13382,
+    "eval_samples_per_second": 477.812,
+    "eval_steps_per_second": 59.735,
+    "train_loss": 0.058391484198245136,
+    "train_runtime": 328.0846,
+    "train_samples": 6570,
+    "train_samples_per_second": 200.253,
+    "train_steps_per_second": 1.676
+}

finetune/control_raising_control/checkpoint-400/config.json

@@ -0,0 +1,50 @@
+{
+  "_name_or_path": "final_models/structroberta_sx2_final",
+  "architectures": [
+    "StructRobertaForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "auto_map": {
+    "AutoConfig": "modeling_structroberta.StructRobertaConfig",
+    "AutoModelForMaskedLM": "modeling_structroberta.StructRoberta",
+    "AutoModelForSequenceClassification": "modeling_structroberta.StructRobertaForSequenceClassification"
+  },
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "conv_size": 9,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": 0,
+    "1": 1
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "label2id": {
+    "0": 0,
+    "1": 1
+  },
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "roberta",
+  "n_cntxt_layers": 4,
+  "n_cntxt_layers_2": 0,
+  "n_parser_layers": 6,
+  "num_attention_heads": 12,
+  "num_hidden_layers": 8,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "problem_type": "single_label_classification",
+  "relations": [
+    "head",
+    "child"
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.1",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 32000,
+  "weight_act": "softmax"
+}

finetune/control_raising_control/checkpoint-400/modeling_structroberta.py

@@ -0,0 +1,2146 @@
+import copy
+import math
+from typing import List, Optional, Tuple, Union
+import torch
+import torch.utils.checkpoint
+from packaging import version
+from torch import nn
+import torch.nn.functional as F
+from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
+from transformers.activations import ACT2FN, gelu
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    MaskedLMOutput,
+    SequenceClassifierOutput
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.utils import logging
+from transformers import RobertaConfig
+
+
+logger = logging.get_logger(__name__)
+ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "roberta-base",
+    "roberta-large",
+    "roberta-large-mnli",
+    "distilroberta-base",
+    "roberta-base-openai-detector",
+    "roberta-large-openai-detector",
+    # See all RoBERTa models at https://huggingface.co/models?filter=roberta
+]
+
+
+class StructRobertaConfig(RobertaConfig):
+    model_type = "roberta"
+
+    def __init__(
+        self,
+        n_parser_layers=4,
+        conv_size=9,
+        relations=("head", "child"),
+        weight_act="softmax",
+        n_cntxt_layers=3,
+        n_cntxt_layers_2=0,
+        **kwargs,):
+        
+        super().__init__(**kwargs)
+        self.n_cntxt_layers = n_cntxt_layers
+        self.n_parser_layers = n_parser_layers
+        self.n_cntxt_layers_2 = n_cntxt_layers_2
+        self.conv_size = conv_size
+        self.relations = relations
+        self.weight_act = weight_act
+
+class Conv1d(nn.Module):
+    """1D convolution layer."""
+
+    def __init__(self, hidden_size, kernel_size, dilation=1):
+        """Initialization.
+
+        Args:
+        hidden_size: dimension of input embeddings
+        kernel_size: convolution kernel size
+        dilation: the spacing between the kernel points
+        """
+        super(Conv1d, self).__init__()
+
+        if kernel_size % 2 == 0:
+            padding = (kernel_size // 2) * dilation
+            self.shift = True
+        else:
+            padding = ((kernel_size - 1) // 2) * dilation
+            self.shift = False
+        self.conv = nn.Conv1d(
+            hidden_size, hidden_size, kernel_size, padding=padding, dilation=dilation
+        )
+
+    def forward(self, x):
+        """Compute convolution.
+
+        Args:
+          x: input embeddings
+        Returns:
+          conv_output: convolution results
+        """
+
+        if self.shift:
+            return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
+        else:
+            return self.conv(x.transpose(1, 2)).transpose(1, 2)
+
+
+class RobertaEmbeddings(nn.Module):
+    """
+    Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
+    """
+
+    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(
+            config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id
+        )
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.hidden_size
+        )
+        self.token_type_embeddings = nn.Embedding(
+            config.type_vocab_size, config.hidden_size
+        )
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.position_embedding_type = getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1))
+        )
+        if version.parse(torch.__version__) > version.parse("1.6.0"):
+            self.register_buffer(
+                "token_type_ids",
+                torch.zeros(self.position_ids.size(), dtype=torch.long),
+                persistent=False,
+            )
+
+        # End copy
+        self.padding_idx = config.pad_token_id
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings,
+            config.hidden_size,
+            padding_idx=self.padding_idx,
+        )
+
+    def forward(
+        self,
+        input_ids=None,
+        token_type_ids=None,
+        position_ids=None,
+        inputs_embeds=None,
+        past_key_values_length=0,
+    ):
+        if position_ids is None:
+            if input_ids is not None:
+                # Create the position ids from the input token ids. Any padded tokens remain padded.
+                position_ids = create_position_ids_from_input_ids(
+                    input_ids, self.padding_idx, past_key_values_length
+                )
+            else:
+                position_ids = self.create_position_ids_from_inputs_embeds(
+                    inputs_embeds
+                )
+
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(
+                    input_shape[0], seq_length
+                )
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(
+                    input_shape, dtype=torch.long, device=self.position_ids.device
+                )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+        """
+        We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
+
+        Args:
+            inputs_embeds: torch.Tensor
+
+        Returns: torch.Tensor
+        """
+        input_shape = inputs_embeds.size()[:-1]
+        sequence_length = input_shape[1]
+
+        position_ids = torch.arange(
+            self.padding_idx + 1,
+            sequence_length + self.padding_idx + 1,
+            dtype=torch.long,
+            device=inputs_embeds.device,
+        )
+        return position_ids.unsqueeze(0).expand(input_shape)
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfAttention with Bert->Roberta
+class RobertaSelfAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(
+            config, "embedding_size"
+        ):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = position_embedding_type or getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(
+                2 * config.max_position_embeddings - 1, self.attention_head_size
+            )
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (
+            self.num_attention_heads,
+            self.attention_head_size,
+        )
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        parser_att_mask=None,
+    ) -> Tuple[torch.Tensor]:
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(-1, 1)
+            position_ids_r = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(
+                distance + self.max_position_embeddings - 1
+            )
+            positional_embedding = positional_embedding.to(
+                dtype=query_layer.dtype
+            )  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                relative_position_scores_key = torch.einsum(
+                    "bhrd,lrd->bhlr", key_layer, positional_embedding
+                )
+                attention_scores = (
+                    attention_scores
+                    + relative_position_scores_query
+                    + relative_position_scores_key
+                )
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in RobertaModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        if parser_att_mask is None:
+            # Normalize the attention scores to probabilities.
+            attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+        else:
+            attention_probs = torch.sigmoid(attention_scores) * parser_att_mask
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (
+            (context_layer, attention_probs) if output_attentions else (context_layer,)
+        )
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput
+class RobertaSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(
+        self, hidden_states: torch.Tensor, input_tensor: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->Roberta
+class RobertaAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        self.self = RobertaSelfAttention(
+            config, position_embedding_type=position_embedding_type
+        )
+        self.output = RobertaSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads,
+            self.self.num_attention_heads,
+            self.self.attention_head_size,
+            self.pruned_heads,
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = (
+            self.self.attention_head_size * self.self.num_attention_heads
+        )
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        parser_att_mask=None,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+            parser_att_mask=parser_att_mask,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[
+            1:
+        ]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate
+class RobertaIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOutput
+class RobertaOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(
+        self, hidden_states: torch.Tensor, input_tensor: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertLayer with Bert->Roberta
+class RobertaLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = RobertaAttention(config)
+        self.is_decoder = config.is_decoder
+        self.add_cross_attention = config.add_cross_attention
+        if self.add_cross_attention:
+            if not self.is_decoder:
+                raise ValueError(
+                    f"{self} should be used as a decoder model if cross attention is added"
+                )
+            self.crossattention = RobertaAttention(
+                config, position_embedding_type="absolute"
+            )
+        self.intermediate = RobertaIntermediate(config)
+        self.output = RobertaOutput(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+        parser_att_mask=None,
+    ) -> Tuple[torch.Tensor]:
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = (
+            past_key_value[:2] if past_key_value is not None else None
+        )
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+            parser_att_mask=parser_att_mask,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[
+                1:
+            ]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            if not hasattr(self, "crossattention"):
+                raise ValueError(
+                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
+                )
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = (
+                past_key_value[-2:] if past_key_value is not None else None
+            )
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                cross_attn_past_key_value,
+                output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = (
+                outputs + cross_attention_outputs[1:-1]
+            )  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk,
+            self.chunk_size_feed_forward,
+            self.seq_len_dim,
+            attention_output,
+        )
+        outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+# Copied from transformers.models.bert.modeling_bert.BertEncoder with Bert->Roberta
+class RobertaEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList(
+            [RobertaLayer(config) for _ in range(config.num_hidden_layers)]
+        )
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+        parser_att_mask=None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = (
+            () if output_attentions and self.config.add_cross_attention else None
+        )
+
+        next_decoder_cache = () if use_cache else None
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                if parser_att_mask is not None:
+                    layer_outputs = layer_module(
+                        hidden_states,
+                        attention_mask,
+                        layer_head_mask,
+                        encoder_hidden_states,
+                        encoder_attention_mask,
+                        past_key_value,
+                        output_attentions,
+                        parser_att_mask=parser_att_mask[i])
+                else:
+                    layer_outputs = layer_module(
+                        hidden_states,
+                        attention_mask,
+                        layer_head_mask,
+                        encoder_hidden_states,
+                        encoder_attention_mask,
+                        past_key_value,
+                        output_attentions,
+                        parser_att_mask=None)
+                    
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertPooler
+class RobertaPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class RobertaPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = RobertaConfig
+    base_model_prefix = "roberta"
+    supports_gradient_checkpointing = True
+
+    # Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            if module.bias is not None:
+                module.bias.data.zero_()
+                module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, RobertaEncoder):
+            module.gradient_checkpointing = value
+
+    def update_keys_to_ignore(self, config, del_keys_to_ignore):
+        """Remove some keys from ignore list"""
+        if not config.tie_word_embeddings:
+            # must make a new list, or the class variable gets modified!
+            self._keys_to_ignore_on_save = [
+                k for k in self._keys_to_ignore_on_save if k not in del_keys_to_ignore
+            ]
+            self._keys_to_ignore_on_load_missing = [
+                k
+                for k in self._keys_to_ignore_on_load_missing
+                if k not in del_keys_to_ignore
+            ]
+
+
+ROBERTA_START_DOCSTRING = r"""
+
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`RobertaConfig`]): Model configuration class with all the parameters of the
+            model. Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+ROBERTA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`RobertaTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class RobertaModel(RobertaPreTrainedModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in *Attention is
+    all you need*_ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
+    Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set
+    to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and
+    `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass.
+
+    .. _*Attention is all you need*: https://arxiv.org/abs/1706.03762
+
+    """
+
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.__init__ with Bert->Roberta
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = RobertaEmbeddings(config)
+        self.encoder = RobertaEncoder(config)
+
+        self.pooler = RobertaPooler(config) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.forward
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        parser_att_mask=None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
+        r"""
+        encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        """
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = (
+            past_key_values[0][0].shape[2] if past_key_values is not None else 0
+        )
+
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                ((batch_size, seq_length + past_key_values_length)), device=device
+            )
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(
+                    batch_size, seq_length
+                )
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(
+                    input_shape, dtype=torch.long, device=device
+                )
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(
+            attention_mask, input_shape, device
+        )
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            (
+                encoder_batch_size,
+                encoder_sequence_length,
+                _,
+            ) = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(
+                encoder_attention_mask
+            )
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=parser_att_mask,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = (
+            self.pooler(sequence_output) if self.pooler is not None else None
+        )
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+class StructRoberta(RobertaPreTrainedModel):
+    _keys_to_ignore_on_save = [r"lm_head.decoder.weight", r"lm_head.decoder.bias"]
+    _keys_to_ignore_on_load_missing = [
+        r"position_ids",
+        r"lm_head.decoder.weight",
+        r"lm_head.decoder.bias",
+    ]
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        if config.is_decoder:
+            logger.warning(
+                "If you want to use `RobertaForMaskedLM` make sure `config.is_decoder=False` for "
+                "bi-directional self-attention."
+            )
+
+        
+        if config.n_cntxt_layers > 0:
+            config_cntxt = copy.deepcopy(config)
+            config_cntxt.num_hidden_layers = config.n_cntxt_layers
+            
+            self.cntxt_layers = RobertaModel(config_cntxt, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers_2 > 0:
+            self.parser_layers_1 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_1 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_1 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_1 = nn.Parameter(
+                torch.zeros((config.n_cntxt_layers_2, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_1.data.normal_(0, 0.1)
+
+            self._scaler_1 = nn.Parameter(torch.zeros(2))
+            
+            config_cntxt_2 = copy.deepcopy(config)
+            config_cntxt_2.num_hidden_layers = config.n_cntxt_layers_2
+            
+            self.cntxt_layers_2 = RobertaModel(config_cntxt_2, add_pooling_layer=False)
+            
+            
+            self.parser_layers_2 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_2 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_2 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_2 = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_2.data.normal_(0, 0.1)
+
+            self._scaler_2 = nn.Parameter(torch.zeros(2))
+            
+        else:
+            self.parser_layers = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(config.n_parser_layers)
+                ]
+            )
+
+            self.distance_ff = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight.data.normal_(0, 0.1)
+
+            self._scaler = nn.Parameter(torch.zeros(2))
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers > 0:
+            self.cntxt_layers.embeddings = self.roberta.embeddings
+        if config.n_cntxt_layers_2 > 0:
+            self.cntxt_layers_2.embeddings = self.roberta.embeddings
+        
+        self.lm_head = RobertaLMHead(config)
+
+        self.pad = config.pad_token_id
+
+        # The LM head weights require special treatment only when they are tied with the word embeddings
+        self.update_keys_to_ignore(config, ["lm_head.decoder.weight"])
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head.decoder = new_embeddings
+
+    @property
+    def scaler(self):
+        return self._scaler.exp()
+    
+    @property
+    def scaler_1(self):
+        return self._scaler_1.exp()
+    
+    @property
+    def scaler_2(self):
+        return self._scaler_2.exp()
+
+    @property
+    def rel_weight(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight, dim=-1)
+        
+    @property
+    def rel_weight_1(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_1)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_1, dim=-1)
+
+    
+    @property
+    def rel_weight_2(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_2)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_2, dim=-1)
+
+
+    def compute_block(self, distance, height, n_cntxt_layers=0):
+        """Compute constituents from distance and height."""
+
+        if n_cntxt_layers>0:
+            if n_cntxt_layers == 1:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_1[0]
+            elif n_cntxt_layers == 2:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_2[0]
+        else:
+            beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+        gamma = torch.sigmoid(-beta_logits)
+        ones = torch.ones_like(gamma)
+
+        block_mask_left = cummin(
+            gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1
+        )
+        block_mask_left = block_mask_left - F.pad(
+            block_mask_left[:, :, :-1], (1, 0), value=0
+        )
+        block_mask_left.tril_(0)
+
+        block_mask_right = cummin(
+            gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1
+        )
+        block_mask_right = block_mask_right - F.pad(
+            block_mask_right[:, :, 1:], (0, 1), value=0
+        )
+        block_mask_right.triu_(0)
+
+        block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+        block = cumsum(block_mask_left).tril(0) + cumsum(
+            block_mask_right, reverse=True
+        ).triu(1)
+
+        return block_p, block
+
+    def compute_head(self, height, n_cntxt_layers=0):
+        """Estimate head for each constituent."""
+
+        _, length = height.size()
+        if n_cntxt_layers>0:
+          if n_cntxt_layers == 1:
+              head_logits = height * self.scaler_1[1]
+          elif n_cntxt_layers == 2:
+              head_logits = height * self.scaler_2[1]  
+        else:
+            head_logits = height * self.scaler[1]
+        index = torch.arange(length, device=height.device)
+
+        mask = (index[:, None, None] <= index[None, None, :]) * (
+            index[None, None, :] <= index[None, :, None]
+        )
+        head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+        head_logits.masked_fill_(~mask[None, :, :, :], -1e9)
+
+        head_p = torch.softmax(head_logits, dim=-1)
+
+        return head_p
+
+    def parse(self, x, embs=None, n_cntxt_layers=0):
+        """Parse input sentence.
+
+        Args:
+        x: input tokens (required).
+        pos: position for each token (optional).
+        Returns:
+        distance: syntactic distance
+        height: syntactic height
+        """
+
+        mask = x != self.pad
+        mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+        if embs is None:
+            h = self.roberta.embeddings(x)
+        else:
+            h = embs
+        
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                parser_layers = self.parser_layers_1
+                height_ff = self.height_ff_1
+                distance_ff = self.distance_ff_1
+            elif n_cntxt_layers == 2:
+                parser_layers = self.parser_layers_2
+                height_ff = self.height_ff_2
+                distance_ff = self.distance_ff_2
+            for i in range(int(self.config.n_parser_layers/2)):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = parser_layers[i](h)
+
+            height = height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+        else:
+            for i in range(self.config.n_parser_layers):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = self.parser_layers[i](h)
+
+            height = self.height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = self.distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+
+        return distance, height
+
+    def generate_mask(self, x, distance, height, n_cntxt_layers=0):
+        """Compute head and cibling distribution for each token."""
+
+        bsz, length = x.size()
+
+        eye = torch.eye(length, device=x.device, dtype=torch.bool)
+        eye = eye[None, :, :].expand((bsz, -1, -1))
+
+        block_p, block = self.compute_block(distance, height, n_cntxt_layers=n_cntxt_layers)
+        head_p = self.compute_head(height, n_cntxt_layers=n_cntxt_layers)
+        head = torch.einsum("blij,bijh->blh", block_p, head_p)
+        head = head.masked_fill(eye, 0)
+        child = head.transpose(1, 2)
+        cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+        rel_list = []
+        if "head" in self.config.relations:
+            rel_list.append(head)
+        if "child" in self.config.relations:
+            rel_list.append(child)
+        if "cibling" in self.config.relations:
+            rel_list.append(cibling)
+
+        rel = torch.stack(rel_list, dim=1)
+
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                rel_weight = self.rel_weight_1
+            elif n_cntxt_layers == 2:
+                rel_weight = self.rel_weight_2
+        else:
+            rel_weight = self.rel_weight
+
+        dep = torch.einsum("lhr,brij->lbhij", rel_weight, rel)
+        
+        if n_cntxt_layers == 1:
+            num_layers = self.cntxt_layers_2.config.num_hidden_layers
+        else:
+            num_layers = self.roberta.config.num_hidden_layers
+            
+        att_mask = dep.reshape(
+            num_layers,
+            bsz,
+            self.config.num_attention_heads,
+            length,
+            length,
+        )
+
+        return att_mask, cibling, head, block
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, MaskedLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
+            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+        kwargs (`Dict[str, any]`, optional, defaults to *{}*):
+            Used to hide legacy arguments that have been deprecated.
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        
+        if self.config.n_cntxt_layers > 0:
+            cntxt_outputs = self.cntxt_layers(
+                input_ids,
+                attention_mask=attention_mask,
+                token_type_ids=token_type_ids,
+                position_ids=position_ids,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict)
+            
+        
+        if self.config.n_cntxt_layers_2 > 0:
+            distance_1, height_1 = self.parse(input_ids, cntxt_outputs[0], n_cntxt_layers=1)
+            att_mask_1, _, _, _ = self.generate_mask(input_ids, distance_1, height_1, n_cntxt_layers=1)
+            
+            cntxt_outputs_2 = self.cntxt_layers_2(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask_1)
+            
+            sequence_output = cntxt_outputs_2[0]
+            
+            distance_2, height_2 = self.parse(input_ids, sequence_output[0], n_cntxt_layers=2)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance_2, height_2, n_cntxt_layers=2)
+
+        elif self.config.n_cntxt_layers > 0:
+            distance, height = self.parse(input_ids, cntxt_outputs[0])
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+        else:
+            distance, height = self.parse(input_ids)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask,
+        )
+        sequence_output = outputs[0]
+        prediction_scores = self.lm_head(sequence_output)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(
+                prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return (
+                ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+            )
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class RobertaLMHead(nn.Module):
+    """Roberta Head for masked language modeling."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size)
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+        self.decoder.bias = self.bias
+
+    def forward(self, features, **kwargs):
+        x = self.dense(features)
+        x = gelu(x)
+        x = self.layer_norm(x)
+
+        # project back to size of vocabulary with bias
+        x = self.decoder(x)
+
+        return x
+
+    def _tie_weights(self):
+        # To tie those two weights if they get disconnected (on TPU or when the bias is resized)
+        self.bias = self.decoder.bias
+
+
+class StructRobertaForSequenceClassification(RobertaPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        if config.n_cntxt_layers > 0:
+            config_cntxt = copy.deepcopy(config)
+            config_cntxt.num_hidden_layers = config.n_cntxt_layers
+            
+            self.cntxt_layers = RobertaModel(config_cntxt, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers_2 > 0:
+            self.parser_layers_1 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_1 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_1 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_1 = nn.Parameter(
+                torch.zeros((config.n_cntxt_layers_2, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_1.data.normal_(0, 0.1)
+
+            self._scaler_1 = nn.Parameter(torch.zeros(2))
+            
+            config_cntxt_2 = copy.deepcopy(config)
+            config_cntxt_2.num_hidden_layers = config.n_cntxt_layers_2
+            
+            self.cntxt_layers_2 = RobertaModel(config_cntxt_2, add_pooling_layer=False)
+            
+            
+            self.parser_layers_2 = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(int(config.n_parser_layers/2))
+                ]
+            )
+
+            self.distance_ff_2 = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff_2 = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight_2 = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight_2.data.normal_(0, 0.1)
+
+            self._scaler_2 = nn.Parameter(torch.zeros(2))
+            
+        else:
+            self.parser_layers = nn.ModuleList(
+                [
+                    nn.Sequential(
+                        Conv1d(config.hidden_size, config.conv_size),
+                        nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                        nn.Tanh(),
+                    )
+                    for i in range(config.n_parser_layers)
+                ]
+            )
+
+            self.distance_ff = nn.Sequential(
+                Conv1d(config.hidden_size, 2),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            self.height_ff = nn.Sequential(
+                nn.Linear(config.hidden_size, config.hidden_size),
+                nn.LayerNorm(config.hidden_size, elementwise_affine=False),
+                nn.Tanh(),
+                nn.Linear(config.hidden_size, 1),
+            )
+
+            n_rel = len(config.relations)
+            self._rel_weight = nn.Parameter(
+                torch.zeros((config.num_hidden_layers, config.num_attention_heads, n_rel))
+            )
+            self._rel_weight.data.normal_(0, 0.1)
+
+            self._scaler = nn.Parameter(torch.zeros(2))
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        
+        if config.n_cntxt_layers > 0:
+            self.cntxt_layers.embeddings = self.roberta.embeddings
+        if config.n_cntxt_layers_2 > 0:
+            self.cntxt_layers_2.embeddings = self.roberta.embeddings
+        
+        
+        self.pad = config.pad_token_id
+        self.classifier = RobertaClassificationHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    
+    @property
+    def scaler(self):
+        return self._scaler.exp()
+    
+    @property
+    def scaler_1(self):
+        return self._scaler_1.exp()
+    
+    @property
+    def scaler_2(self):
+        return self._scaler_2.exp()
+
+    @property
+    def rel_weight(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight, dim=-1)
+        
+    @property
+    def rel_weight_1(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_1)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_1, dim=-1)
+
+    
+    @property
+    def rel_weight_2(self):
+        if self.config.weight_act == "sigmoid":
+            return torch.sigmoid(self._rel_weight_2)
+        elif self.config.weight_act == "softmax":
+            return torch.softmax(self._rel_weight_2, dim=-1)
+
+
+    def compute_block(self, distance, height, n_cntxt_layers=0):
+        """Compute constituents from distance and height."""
+
+        if n_cntxt_layers>0:
+            if n_cntxt_layers == 1:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_1[0]
+            elif n_cntxt_layers == 2:
+              beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler_2[0]
+        else:
+            beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+        gamma = torch.sigmoid(-beta_logits)
+        ones = torch.ones_like(gamma)
+
+        block_mask_left = cummin(
+            gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1
+        )
+        block_mask_left = block_mask_left - F.pad(
+            block_mask_left[:, :, :-1], (1, 0), value=0
+        )
+        block_mask_left.tril_(0)
+
+        block_mask_right = cummin(
+            gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1
+        )
+        block_mask_right = block_mask_right - F.pad(
+            block_mask_right[:, :, 1:], (0, 1), value=0
+        )
+        block_mask_right.triu_(0)
+
+        block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+        block = cumsum(block_mask_left).tril(0) + cumsum(
+            block_mask_right, reverse=True
+        ).triu(1)
+
+        return block_p, block
+
+    def compute_head(self, height, n_cntxt_layers=0):
+        """Estimate head for each constituent."""
+
+        _, length = height.size()
+        if n_cntxt_layers>0:
+          if n_cntxt_layers == 1:
+              head_logits = height * self.scaler_1[1]
+          elif n_cntxt_layers == 2:
+              head_logits = height * self.scaler_2[1]  
+        else:
+            head_logits = height * self.scaler[1]
+        index = torch.arange(length, device=height.device)
+
+        mask = (index[:, None, None] <= index[None, None, :]) * (
+            index[None, None, :] <= index[None, :, None]
+        )
+        head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+        head_logits.masked_fill_(~mask[None, :, :, :], -1e9)
+
+        head_p = torch.softmax(head_logits, dim=-1)
+
+        return head_p
+
+    def parse(self, x, embs=None, n_cntxt_layers=0):
+        """Parse input sentence.
+
+        Args:
+        x: input tokens (required).
+        pos: position for each token (optional).
+        Returns:
+        distance: syntactic distance
+        height: syntactic height
+        """
+
+        mask = x != self.pad
+        mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+        if embs is None:
+            h = self.roberta.embeddings(x)
+        else:
+            h = embs
+        
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                parser_layers = self.parser_layers_1
+                height_ff = self.height_ff_1
+                distance_ff = self.distance_ff_1
+            elif n_cntxt_layers == 2:
+                parser_layers = self.parser_layers_2
+                height_ff = self.height_ff_2
+                distance_ff = self.distance_ff_2
+            for i in range(int(self.config.n_parser_layers/2)):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = parser_layers[i](h)
+
+            height = height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+        else:
+            for i in range(self.config.n_parser_layers):
+                h = h.masked_fill(~mask[:, :, None], 0)
+                h = self.parser_layers[i](h)
+
+            height = self.height_ff(h).squeeze(-1)
+            height.masked_fill_(~mask, -1e9)
+
+            distance = self.distance_ff(h).squeeze(-1)
+            distance.masked_fill_(~mask_shifted, 1e9)
+
+            # Calbrating the distance and height to the same level
+            length = distance.size(1)
+            height_max = height[:, None, :].expand(-1, length, -1)
+            height_max = torch.cummax(
+                height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9, dim=-1
+            )[0].triu(0)
+
+            margin_left = torch.relu(
+                F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max
+            )
+            margin_right = torch.relu(distance[:, None, :] - height_max)
+            margin = torch.where(
+                margin_left > margin_right, margin_right, margin_left
+            ).triu(0)
+
+            margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+            margin.masked_fill_(~margin_mask, 0)
+            margin = margin.max()
+
+            distance = distance - margin
+
+        return distance, height
+
+    def generate_mask(self, x, distance, height, n_cntxt_layers=0):
+        """Compute head and cibling distribution for each token."""
+
+        bsz, length = x.size()
+
+        eye = torch.eye(length, device=x.device, dtype=torch.bool)
+        eye = eye[None, :, :].expand((bsz, -1, -1))
+
+        block_p, block = self.compute_block(distance, height, n_cntxt_layers=n_cntxt_layers)
+        head_p = self.compute_head(height, n_cntxt_layers=n_cntxt_layers)
+        head = torch.einsum("blij,bijh->blh", block_p, head_p)
+        head = head.masked_fill(eye, 0)
+        child = head.transpose(1, 2)
+        cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+        rel_list = []
+        if "head" in self.config.relations:
+            rel_list.append(head)
+        if "child" in self.config.relations:
+            rel_list.append(child)
+        if "cibling" in self.config.relations:
+            rel_list.append(cibling)
+
+        rel = torch.stack(rel_list, dim=1)
+
+        if n_cntxt_layers > 0:
+            if n_cntxt_layers == 1:
+                rel_weight = self.rel_weight_1
+            elif n_cntxt_layers == 2:
+                rel_weight = self.rel_weight_2
+        else:
+            rel_weight = self.rel_weight
+
+        dep = torch.einsum("lhr,brij->lbhij", rel_weight, rel)
+        
+        if n_cntxt_layers == 1:
+            num_layers = self.cntxt_layers_2.config.num_hidden_layers
+        else:
+            num_layers = self.roberta.config.num_hidden_layers
+            
+        att_mask = dep.reshape(
+            num_layers,
+            bsz,
+            self.config.num_attention_heads,
+            length,
+            length,
+        )
+
+        return att_mask, cibling, head, block
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        
+        if self.config.n_cntxt_layers > 0:
+            cntxt_outputs = self.cntxt_layers(
+                input_ids,
+                attention_mask=attention_mask,
+                token_type_ids=token_type_ids,
+                position_ids=position_ids,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict)
+            
+        
+        if self.config.n_cntxt_layers_2 > 0:
+            distance_1, height_1 = self.parse(input_ids, cntxt_outputs[0], n_cntxt_layers=1)
+            att_mask_1, _, _, _ = self.generate_mask(input_ids, distance_1, height_1, n_cntxt_layers=1)
+            
+            cntxt_outputs_2 = self.cntxt_layers_2(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask_1)
+            
+            sequence_output = cntxt_outputs_2[0]
+            
+            distance_2, height_2 = self.parse(input_ids, sequence_output[0], n_cntxt_layers=2)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance_2, height_2, n_cntxt_layers=2)
+
+        elif self.config.n_cntxt_layers > 0:
+            distance, height = self.parse(input_ids, cntxt_outputs[0])
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+        else:
+            distance, height = self.parse(input_ids)
+            att_mask, _, _, _ = self.generate_mask(input_ids, distance, height)
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            parser_att_mask=att_mask,
+        )
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class RobertaClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+def create_position_ids_from_input_ids(
+    input_ids, padding_idx, past_key_values_length=0
+):
+    """
+    Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
+    are ignored. This is modified from fairseq's `utils.make_positions`.
+
+    Args:
+        x: torch.Tensor x:
+
+    Returns: torch.Tensor
+    """
+    # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
+    mask = input_ids.ne(padding_idx).int()
+    incremental_indices = (
+        torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length
+    ) * mask
+    return incremental_indices.long() + padding_idx
+
+
+def cumprod(x, reverse=False, exclusive=False):
+    """cumulative product."""
+    if reverse:
+        x = x.flip([-1])
+
+    if exclusive:
+        x = F.pad(x[:, :, :-1], (1, 0), value=1)
+
+    cx = x.cumprod(-1)
+
+    if reverse:
+        cx = cx.flip([-1])
+    return cx
+
+
+def cumsum(x, reverse=False, exclusive=False):
+    """cumulative sum."""
+    bsz, _, length = x.size()
+    device = x.device
+    if reverse:
+        if exclusive:
+            w = torch.ones([bsz, length, length], device=device).tril(-1)
+        else:
+            w = torch.ones([bsz, length, length], device=device).tril(0)
+        cx = torch.bmm(x, w)
+    else:
+        if exclusive:
+            w = torch.ones([bsz, length, length], device=device).triu(1)
+        else:
+            w = torch.ones([bsz, length, length], device=device).triu(0)
+        cx = torch.bmm(x, w)
+    return cx
+
+
+def cummin(x, reverse=False, exclusive=False, max_value=1e9):
+    """cumulative min."""
+    if reverse:
+        if exclusive:
+            x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
+        x = x.flip([-1]).cummin(-1)[0].flip([-1])
+    else:
+        if exclusive:
+            x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
+        x = x.cummin(-1)[0]
+    return x

finetune/control_raising_control/checkpoint-400/optimizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:276b4418ede949bd6f8dd84d85070e2d14eda7ab99bb967a5fb4a15d83387a47
+size 1154109317

finetune/control_raising_control/checkpoint-400/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1da9b8cae6ed0f931ee814194e2fb0860595439291dbf891b54455f0e5edbcb9
+size 577069633

finetune/control_raising_control/checkpoint-400/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a5299232550db0e37485cb678803cf54c345c834976d36e81571c3d16c29cf7
+size 14503

finetune/control_raising_control/checkpoint-400/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c2403f14641b3caeb1b4d17bf70ec776358494ec9059cbe53a4c9c5a18c4c15
+size 623

finetune/control_raising_control/checkpoint-400/special_tokens_map.json

@@ -0,0 +1,15 @@
+{
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "mask_token": {
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "unk_token": "<unk>"
+}

finetune/control_raising_control/checkpoint-400/tokenizer_config.json

@@ -0,0 +1,65 @@
+{
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "cls_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "mask_token": {
+    "__type": "AddedToken",
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "model_max_length": 512,
+  "name_or_path": "final_models/structroberta_sx2_final",
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "special_tokens_map_file": null,
+  "tokenizer_class": "RobertaTokenizer",
+  "trim_offsets": true,
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

finetune/control_raising_control/checkpoint-400/trainer_state.json

@@ -0,0 +1,27 @@
+{
+  "best_metric": 0.8866091446429785,
+  "best_model_checkpoint": "final_models/structroberta_sx2_final/finetune/control_raising_control/checkpoint-400",
+  "epoch": 7.2727272727272725,
+  "global_step": 400,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 7.27,
+      "eval_accuracy": 0.8752802014350891,
+      "eval_f1": 0.8866091446429785,
+      "eval_loss": 0.801540732383728,
+      "eval_mcc": 0.7701394030930233,
+      "eval_runtime": 28.1112,
+      "eval_samples_per_second": 476.038,
+      "eval_steps_per_second": 59.514,
+      "step": 400
+    }
+  ],
+  "max_steps": 550,
+  "num_train_epochs": 10,
+  "total_flos": 4377752676449280.0,
+  "trial_name": null,
+  "trial_params": null
+}