try to make it work quick and dirty

BERT_explainability/BERT.py

@@ -0,0 +1,671 @@
+from __future__ import absolute_import
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+import math
+from transformers import BertConfig
+from transformers.modeling_outputs import BaseModelOutputWithPooling, BaseModelOutput
+from BERT_explainability.modules.layers_ours import *
+from transformers import (
+    BertPreTrainedModel,
+    PreTrainedModel,
+)
+
+ACT2FN = {
+    "relu": ReLU,
+    "tanh": Tanh,
+    "gelu": GELU,
+}
+
+
+def get_activation(activation_string):
+    if activation_string in ACT2FN:
+        return ACT2FN[activation_string]
+    else:
+        raise KeyError("function {} not found in ACT2FN mapping {}".format(activation_string, list(ACT2FN.keys())))
+
+def compute_rollout_attention(all_layer_matrices, start_layer=0):
+    # adding residual consideration
+    num_tokens = all_layer_matrices[0].shape[1]
+    batch_size = all_layer_matrices[0].shape[0]
+    eye = torch.eye(num_tokens).expand(batch_size, num_tokens, num_tokens).to(all_layer_matrices[0].device)
+    all_layer_matrices = [all_layer_matrices[i] + eye for i in range(len(all_layer_matrices))]
+    all_layer_matrices = [all_layer_matrices[i] / all_layer_matrices[i].sum(dim=-1, keepdim=True)
+                          for i in range(len(all_layer_matrices))]
+    joint_attention = all_layer_matrices[start_layer]
+    for i in range(start_layer+1, len(all_layer_matrices)):
+        joint_attention = all_layer_matrices[i].bmm(joint_attention)
+    return joint_attention
+
+class BertEmbeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    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 = LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+
+        self.add1 = Add()
+        self.add2 = Add()
+
+    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, :seq_length]
+
+        if token_type_ids is None:
+            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)
+        position_embeddings = self.position_embeddings(position_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        # embeddings = inputs_embeds + position_embeddings + token_type_embeddings
+        embeddings = self.add1([token_type_embeddings, position_embeddings])
+        embeddings = self.add2([embeddings, inputs_embeds])
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+    def relprop(self, cam, **kwargs):
+        cam = self.dropout.relprop(cam, **kwargs)
+        cam = self.LayerNorm.relprop(cam, **kwargs)
+
+        # [inputs_embeds, position_embeddings, token_type_embeddings]
+        (cam) = self.add2.relprop(cam, **kwargs)
+
+        return cam
+
+class BertEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
+
+    def forward(
+            self,
+            hidden_states,
+            attention_mask=None,
+            head_mask=None,
+            encoder_hidden_states=None,
+            encoder_attention_mask=None,
+            output_attentions=False,
+            output_hidden_states=False,
+            return_dict=False,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions 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
+
+            if getattr(self.config, "gradient_checkpointing", False):
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    output_attentions,
+                )
+            hidden_states = layer_outputs[0]
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+    def relprop(self, cam, **kwargs):
+        # assuming output_hidden_states is False
+        for layer_module in reversed(self.layer):
+            cam = layer_module.relprop(cam, **kwargs)
+        return cam
+
+# not adding relprop since this is only pooling at the end of the network, does not impact tokens importance
+class BertPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = Linear(config.hidden_size, config.hidden_size)
+        self.activation = Tanh()
+        self.pool = IndexSelect()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        self._seq_size = hidden_states.shape[1]
+
+        # first_token_tensor = hidden_states[:, 0]
+        first_token_tensor = self.pool(hidden_states, 1, torch.tensor(0, device=hidden_states.device))
+        first_token_tensor = first_token_tensor.squeeze(1)
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+    def relprop(self, cam, **kwargs):
+        cam = self.activation.relprop(cam, **kwargs)
+        #print(cam.sum())
+        cam = self.dense.relprop(cam, **kwargs)
+        #print(cam.sum())
+        cam = cam.unsqueeze(1)
+        cam = self.pool.relprop(cam, **kwargs)
+        #print(cam.sum())
+
+        return cam
+
+class BertAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self = BertSelfAttention(config)
+        self.output = BertSelfOutput(config)
+        self.pruned_heads = set()
+        self.clone = Clone()
+
+    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,
+            attention_mask=None,
+            head_mask=None,
+            encoder_hidden_states=None,
+            encoder_attention_mask=None,
+            output_attentions=False,
+    ):
+        h1, h2 = self.clone(hidden_states, 2)
+        self_outputs = self.self(
+            h1,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], h2)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+    def relprop(self, cam, **kwargs):
+        # assuming that we don't ouput the attentions (outputs = (attention_output,)), self_outputs=(context_layer,)
+        (cam1, cam2) = self.output.relprop(cam, **kwargs)
+        #print(cam1.sum(), cam2.sum(), (cam1 + cam2).sum())
+        cam1 = self.self.relprop(cam1, **kwargs)
+        #print(cam1.sum(), cam2.sum(), (cam1 + cam2).sum())
+
+        return self.clone.relprop((cam1, cam2), **kwargs)
+
+class BertSelfAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, 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 = Linear(config.hidden_size, self.all_head_size)
+        self.key = Linear(config.hidden_size, self.all_head_size)
+        self.value = Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = Dropout(config.attention_probs_dropout_prob)
+
+        self.matmul1 = MatMul()
+        self.matmul2 = MatMul()
+        self.softmax = Softmax(dim=-1)
+        self.add = Add()
+        self.mul = Mul()
+        self.head_mask = None
+        self.attention_mask = None
+        self.clone = Clone()
+
+        self.attn_cam = None
+        self.attn = None
+        self.attn_gradients = None
+
+    def get_attn(self):
+        return self.attn
+
+    def save_attn(self, attn):
+        self.attn = attn
+
+    def save_attn_cam(self, cam):
+        self.attn_cam = cam
+
+    def get_attn_cam(self):
+        return self.attn_cam
+
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+
+    def get_attn_gradients(self):
+        return self.attn_gradients
+
+    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 transpose_for_scores_relprop(self, x):
+        return x.permute(0, 2, 1, 3).flatten(2)
+
+    def forward(
+            self,
+            hidden_states,
+            attention_mask=None,
+            head_mask=None,
+            encoder_hidden_states=None,
+            encoder_attention_mask=None,
+            output_attentions=False,
+    ):
+        self.head_mask = head_mask
+        self.attention_mask = attention_mask
+
+        h1, h2, h3 = self.clone(hidden_states, 3)
+        mixed_query_layer = self.query(h1)
+
+        # 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.
+        if encoder_hidden_states is not None:
+            mixed_key_layer = self.key(encoder_hidden_states)
+            mixed_value_layer = self.value(encoder_hidden_states)
+            attention_mask = encoder_attention_mask
+        else:
+            mixed_key_layer = self.key(h2)
+            mixed_value_layer = self.value(h3)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+        key_layer = self.transpose_for_scores(mixed_key_layer)
+        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = self.matmul1([query_layer, key_layer.transpose(-1, -2)])
+        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 BertModel forward() function)
+            attention_scores = self.add([attention_scores, attention_mask])
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = self.softmax(attention_scores)
+
+        self.save_attn(attention_probs)
+        attention_probs.register_hook(self.save_attn_gradients)
+
+        # 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 = self.matmul2([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,)
+        return outputs
+
+    def relprop(self, cam, **kwargs):
+        # Assume output_attentions == False
+        cam = self.transpose_for_scores(cam)
+
+        # [attention_probs, value_layer]
+        (cam1, cam2) = self.matmul2.relprop(cam, **kwargs)
+        cam1 /= 2
+        cam2 /= 2
+        if self.head_mask is not None:
+            # [attention_probs, head_mask]
+            (cam1, _)= self.mul.relprop(cam1, **kwargs)
+
+
+        self.save_attn_cam(cam1)
+
+        cam1 = self.dropout.relprop(cam1, **kwargs)
+
+        cam1 = self.softmax.relprop(cam1, **kwargs)
+
+        if self.attention_mask is not None:
+            # [attention_scores, attention_mask]
+            (cam1, _) = self.add.relprop(cam1, **kwargs)
+
+        # [query_layer, key_layer.transpose(-1, -2)]
+        (cam1_1, cam1_2) = self.matmul1.relprop(cam1, **kwargs)
+        cam1_1 /= 2
+        cam1_2 /= 2
+
+        # query
+        cam1_1 = self.transpose_for_scores_relprop(cam1_1)
+        cam1_1 = self.query.relprop(cam1_1, **kwargs)
+
+        # key
+        cam1_2 = self.transpose_for_scores_relprop(cam1_2.transpose(-1, -2))
+        cam1_2 = self.key.relprop(cam1_2, **kwargs)
+
+        # value
+        cam2 = self.transpose_for_scores_relprop(cam2)
+        cam2 = self.value.relprop(cam2, **kwargs)
+
+        cam = self.clone.relprop((cam1_1, cam1_2, cam2), **kwargs)
+
+        return cam
+
+
+class BertSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = Dropout(config.hidden_dropout_prob)
+        self.add = Add()
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        add = self.add([hidden_states, input_tensor])
+        hidden_states = self.LayerNorm(add)
+        return hidden_states
+
+    def relprop(self, cam, **kwargs):
+        cam = self.LayerNorm.relprop(cam, **kwargs)
+        # [hidden_states, input_tensor]
+        (cam1, cam2) = self.add.relprop(cam, **kwargs)
+        cam1 = self.dropout.relprop(cam1, **kwargs)
+        cam1 = self.dense.relprop(cam1, **kwargs)
+
+        return (cam1, cam2)
+
+
+class BertIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = 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):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+    def relprop(self, cam, **kwargs):
+        cam = self.intermediate_act_fn.relprop(cam, **kwargs)  # FIXME only ReLU
+        #print(cam.sum())
+        cam = self.dense.relprop(cam, **kwargs)
+        #print(cam.sum())
+        return cam
+
+
+class BertOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = Dropout(config.hidden_dropout_prob)
+        self.add = Add()
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        add = self.add([hidden_states, input_tensor])
+        hidden_states = self.LayerNorm(add)
+        return hidden_states
+
+    def relprop(self, cam, **kwargs):
+        # print("in", cam.sum())
+        cam = self.LayerNorm.relprop(cam, **kwargs)
+        #print(cam.sum())
+        # [hidden_states, input_tensor]
+        (cam1, cam2)= self.add.relprop(cam, **kwargs)
+        # print("add", cam1.sum(), cam2.sum(), cam1.sum() + cam2.sum())
+        cam1 = self.dropout.relprop(cam1, **kwargs)
+        #print(cam1.sum())
+        cam1 = self.dense.relprop(cam1, **kwargs)
+        # print("dense", cam1.sum())
+
+        # print("out", cam1.sum() + cam2.sum(), cam1.sum(), cam2.sum())
+        return (cam1, cam2)
+
+
+class BertLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.attention = BertAttention(config)
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config)
+        self.clone = Clone()
+
+    def forward(
+            self,
+            hidden_states,
+            attention_mask=None,
+            head_mask=None,
+            output_attentions=False,
+    ):
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        ao1, ao2 = self.clone(attention_output, 2)
+        intermediate_output = self.intermediate(ao1)
+        layer_output = self.output(intermediate_output, ao2)
+
+        outputs = (layer_output,) + outputs
+        return outputs
+
+    def relprop(self, cam, **kwargs):
+        (cam1, cam2) = self.output.relprop(cam, **kwargs)
+        # print("output", cam1.sum(), cam2.sum(), cam1.sum() + cam2.sum())
+        cam1 = self.intermediate.relprop(cam1, **kwargs)
+        # print("intermediate", cam1.sum())
+        cam = self.clone.relprop((cam1, cam2), **kwargs)
+        # print("clone", cam.sum())
+        cam = self.attention.relprop(cam, **kwargs)
+        # print("attention", cam.sum())
+        return cam
+
+
+class BertModel(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BertEmbeddings(config)
+        self.encoder = BertEncoder(config)
+        self.pooler = BertPooler(config)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def forward(
+            self,
+            input_ids=None,
+            attention_mask=None,
+            token_type_ids=None,
+            position_ids=None,
+            head_mask=None,
+            inputs_embeds=None,
+            encoder_hidden_states=None,
+            encoder_attention_mask=None,
+            output_attentions=None,
+            output_hidden_states=None,
+            return_dict=None,
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(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 (:obj:`torch.FloatTensor` of shape :obj:`(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 MASKED tokens.
+        """
+        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 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")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            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
+        )
+
+        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,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output)
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+    def relprop(self, cam, **kwargs):
+        cam = self.pooler.relprop(cam, **kwargs)
+        # print("111111111111",cam.sum())
+        cam = self.encoder.relprop(cam, **kwargs)
+        # print("222222222222222", cam.sum())
+        # print("conservation: ", cam.sum())
+        return cam
+
+
+if __name__ == '__main__':
+    class Config:
+      def __init__(self, hidden_size, num_attention_heads, attention_probs_dropout_prob):
+        self.hidden_size = hidden_size
+        self.num_attention_heads = num_attention_heads
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+
+    model = BertSelfAttention(Config(1024, 4, 0.1))
+    x = torch.rand(2, 20, 1024)
+    x.requires_grad_()
+
+    model.eval()
+
+    y = model.forward(x)
+
+    relprop = model.relprop(torch.rand(2, 20, 1024), (torch.rand(2, 20, 1024),))
+
+    print(relprop[1][0].shape)

BERT_explainability/BERT_cls_lrp.py

@@ -0,0 +1,202 @@
+from transformers import BertPreTrainedModel
+from transformers.utils import logging
+from BERT_explainability.modules.layers_lrp import *
+from BERT_explainability.modules.BERT.BERT_orig_lrp import BertModel
+from torch.nn import CrossEntropyLoss, MSELoss
+import torch.nn as nn
+from typing import List, Any
+import torch
+from BERT_rationale_benchmark.models.model_utils import PaddedSequence
+
+
+class BertForSequenceClassification(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config)
+        self.dropout = Dropout(config.hidden_dropout_prob)
+        self.classifier = Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(
+            self,
+            input_ids=None,
+            attention_mask=None,
+            token_type_ids=None,
+            position_ids=None,
+            head_mask=None,
+            inputs_embeds=None,
+            labels=None,
+            output_attentions=None,
+            output_hidden_states=None,
+            return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
+            If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`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
+
+        outputs = self.bert(
+            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,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        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,
+        )
+
+    def relprop(self, cam=None, **kwargs):
+        cam = self.classifier.relprop(cam, **kwargs)
+        cam = self.dropout.relprop(cam, **kwargs)
+        cam = self.bert.relprop(cam, **kwargs)
+        return cam
+
+
+# this is the actual classifier we will be using
+class BertClassifier(nn.Module):
+    """Thin wrapper around BertForSequenceClassification"""
+
+    def __init__(self,
+                 bert_dir: str,
+                 pad_token_id: int,
+                 cls_token_id: int,
+                 sep_token_id: int,
+                 num_labels: int,
+                 max_length: int = 512,
+                 use_half_precision=True):
+        super(BertClassifier, self).__init__()
+        bert = BertForSequenceClassification.from_pretrained(bert_dir, num_labels=num_labels)
+        if use_half_precision:
+            import apex
+            bert = bert.half()
+        self.bert = bert
+        self.pad_token_id = pad_token_id
+        self.cls_token_id = cls_token_id
+        self.sep_token_id = sep_token_id
+        self.max_length = max_length
+
+    def forward(self,
+                query: List[torch.tensor],
+                docids: List[Any],
+                document_batch: List[torch.tensor]):
+        assert len(query) == len(document_batch)
+        print(query)
+        # note about device management:
+        # since distributed training is enabled, the inputs to this module can be on *any* device (preferably cpu, since we wrap and unwrap the module)
+        # we want to keep these params on the input device (assuming CPU) for as long as possible for cheap memory access
+        target_device = next(self.parameters()).device
+        cls_token = torch.tensor([self.cls_token_id]).to(device=document_batch[0].device)
+        sep_token = torch.tensor([self.sep_token_id]).to(device=document_batch[0].device)
+        input_tensors = []
+        position_ids = []
+        for q, d in zip(query, document_batch):
+            if len(q) + len(d) + 2 > self.max_length:
+                d = d[:(self.max_length - len(q) - 2)]
+            input_tensors.append(torch.cat([cls_token, q, sep_token, d]))
+            position_ids.append(torch.tensor(list(range(0, len(q) + 1)) + list(range(0, len(d) + 1))))
+        bert_input = PaddedSequence.autopad(input_tensors, batch_first=True, padding_value=self.pad_token_id,
+                                            device=target_device)
+        positions = PaddedSequence.autopad(position_ids, batch_first=True, padding_value=0, device=target_device)
+        (classes,) = self.bert(bert_input.data,
+                               attention_mask=bert_input.mask(on=0.0, off=float('-inf'), device=target_device),
+                               position_ids=positions.data)
+        assert torch.all(classes == classes)  # for nans
+
+        print(input_tensors[0])
+        print(self.relprop()[0])
+
+        return classes
+
+    def relprop(self, cam=None, **kwargs):
+        return self.bert.relprop(cam, **kwargs)
+
+
+if __name__ == '__main__':
+    from transformers import BertTokenizer
+    import os
+
+    class Config:
+        def __init__(self, hidden_size, num_attention_heads, attention_probs_dropout_prob, num_labels,
+                     hidden_dropout_prob):
+            self.hidden_size = hidden_size
+            self.num_attention_heads = num_attention_heads
+            self.attention_probs_dropout_prob = attention_probs_dropout_prob
+            self.num_labels = num_labels
+            self.hidden_dropout_prob = hidden_dropout_prob
+
+
+    tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+    x = tokenizer.encode_plus("In this movie the acting is great. The movie is perfect! [sep]",
+                         add_special_tokens=True,
+                         max_length=512,
+                         return_token_type_ids=False,
+                         return_attention_mask=True,
+                         pad_to_max_length=True,
+                         return_tensors='pt',
+                         truncation=True)
+
+    print(x['input_ids'])
+
+    model = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)
+    model_save_file = os.path.join('./BERT_explainability/output_bert/movies/classifier/', 'classifier.pt')
+    model.load_state_dict(torch.load(model_save_file))
+
+    # x = torch.randint(100, (2, 20))
+    # x = torch.tensor([[101, 2054, 2003, 1996, 15792, 1997, 2023, 3319, 1029, 102,
+    #                    101, 4079, 102, 101, 6732, 102, 101, 2643, 102, 101,
+    #                    2038, 102, 101, 1037, 102, 101, 2933, 102, 101, 2005,
+    #                    102, 101, 2032, 102, 101, 1010, 102, 101, 1037, 102,
+    #                    101, 3800, 102, 101, 2005, 102, 101, 2010, 102, 101,
+    #                    2166, 102, 101, 1010, 102, 101, 1998, 102, 101, 2010,
+    #                    102, 101, 4650, 102, 101, 1010, 102, 101, 2002, 102,
+    #                    101, 2074, 102, 101, 2515, 102, 101, 1050, 102, 101,
+    #                    1005, 102, 101, 1056, 102, 101, 2113, 102, 101, 2054,
+    #                    102, 101, 1012, 102]])
+    # x.requires_grad_()
+
+    model.eval()
+
+    y = model(x['input_ids'], x['attention_mask'])
+    print(y)
+
+    cam, _ = model.relprop()
+
+    #print(cam.shape)
+
+    cam = cam.sum(-1)
+    #print(cam)

BERT_explainability/BERT_orig_lrp.py

@@ -0,0 +1,671 @@
+from __future__ import absolute_import
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+import math
+from transformers import BertConfig
+from transformers.modeling_outputs import BaseModelOutputWithPooling, BaseModelOutput
+from BERT_explainability.modules.layers_lrp import *
+from transformers import (
+    BertPreTrainedModel,
+    PreTrainedModel,
+)
+
+ACT2FN = {
+    "relu": ReLU,
+    "tanh": Tanh,
+    "gelu": GELU,
+}
+
+
+def get_activation(activation_string):
+    if activation_string in ACT2FN:
+        return ACT2FN[activation_string]
+    else:
+        raise KeyError("function {} not found in ACT2FN mapping {}".format(activation_string, list(ACT2FN.keys())))
+
+def compute_rollout_attention(all_layer_matrices, start_layer=0):
+    # adding residual consideration
+    num_tokens = all_layer_matrices[0].shape[1]
+    batch_size = all_layer_matrices[0].shape[0]
+    eye = torch.eye(num_tokens).expand(batch_size, num_tokens, num_tokens).to(all_layer_matrices[0].device)
+    all_layer_matrices = [all_layer_matrices[i] + eye for i in range(len(all_layer_matrices))]
+    all_layer_matrices = [all_layer_matrices[i] / all_layer_matrices[i].sum(dim=-1, keepdim=True)
+                          for i in range(len(all_layer_matrices))]
+    joint_attention = all_layer_matrices[start_layer]
+    for i in range(start_layer+1, len(all_layer_matrices)):
+        joint_attention = all_layer_matrices[i].bmm(joint_attention)
+    return joint_attention
+
+class BertEmbeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    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 = LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+
+        self.add1 = Add()
+        self.add2 = Add()
+
+    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, :seq_length]
+
+        if token_type_ids is None:
+            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)
+        position_embeddings = self.position_embeddings(position_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        # embeddings = inputs_embeds + position_embeddings + token_type_embeddings
+        embeddings = self.add1([token_type_embeddings, position_embeddings])
+        embeddings = self.add2([embeddings, inputs_embeds])
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+    def relprop(self, cam, **kwargs):
+        cam = self.dropout.relprop(cam, **kwargs)
+        cam = self.LayerNorm.relprop(cam, **kwargs)
+
+        # [inputs_embeds, position_embeddings, token_type_embeddings]
+        (cam) = self.add2.relprop(cam, **kwargs)
+
+        return cam
+
+class BertEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
+
+    def forward(
+            self,
+            hidden_states,
+            attention_mask=None,
+            head_mask=None,
+            encoder_hidden_states=None,
+            encoder_attention_mask=None,
+            output_attentions=False,
+            output_hidden_states=False,
+            return_dict=False,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions 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
+
+            if getattr(self.config, "gradient_checkpointing", False):
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    output_attentions,
+                )
+            hidden_states = layer_outputs[0]
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+    def relprop(self, cam, **kwargs):
+        # assuming output_hidden_states is False
+        for layer_module in reversed(self.layer):
+            cam = layer_module.relprop(cam, **kwargs)
+        return cam
+
+# not adding relprop since this is only pooling at the end of the network, does not impact tokens importance
+class BertPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = Linear(config.hidden_size, config.hidden_size)
+        self.activation = Tanh()
+        self.pool = IndexSelect()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        self._seq_size = hidden_states.shape[1]
+
+        # first_token_tensor = hidden_states[:, 0]
+        first_token_tensor = self.pool(hidden_states, 1, torch.tensor(0, device=hidden_states.device))
+        first_token_tensor = first_token_tensor.squeeze(1)
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+    def relprop(self, cam, **kwargs):
+        cam = self.activation.relprop(cam, **kwargs)
+        #print(cam.sum())
+        cam = self.dense.relprop(cam, **kwargs)
+        #print(cam.sum())
+        cam = cam.unsqueeze(1)
+        cam = self.pool.relprop(cam, **kwargs)
+        #print(cam.sum())
+
+        return cam
+
+class BertAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self = BertSelfAttention(config)
+        self.output = BertSelfOutput(config)
+        self.pruned_heads = set()
+        self.clone = Clone()
+
+    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,
+            attention_mask=None,
+            head_mask=None,
+            encoder_hidden_states=None,
+            encoder_attention_mask=None,
+            output_attentions=False,
+    ):
+        h1, h2 = self.clone(hidden_states, 2)
+        self_outputs = self.self(
+            h1,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], h2)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+    def relprop(self, cam, **kwargs):
+        # assuming that we don't ouput the attentions (outputs = (attention_output,)), self_outputs=(context_layer,)
+        (cam1, cam2) = self.output.relprop(cam, **kwargs)
+        #print(cam1.sum(), cam2.sum(), (cam1 + cam2).sum())
+        cam1 = self.self.relprop(cam1, **kwargs)
+        #print(cam1.sum(), cam2.sum(), (cam1 + cam2).sum())
+
+        return self.clone.relprop((cam1, cam2), **kwargs)
+
+class BertSelfAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, 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 = Linear(config.hidden_size, self.all_head_size)
+        self.key = Linear(config.hidden_size, self.all_head_size)
+        self.value = Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = Dropout(config.attention_probs_dropout_prob)
+
+        self.matmul1 = MatMul()
+        self.matmul2 = MatMul()
+        self.softmax = Softmax(dim=-1)
+        self.add = Add()
+        self.mul = Mul()
+        self.head_mask = None
+        self.attention_mask = None
+        self.clone = Clone()
+
+        self.attn_cam = None
+        self.attn = None
+        self.attn_gradients = None
+
+    def get_attn(self):
+        return self.attn
+
+    def save_attn(self, attn):
+        self.attn = attn
+
+    def save_attn_cam(self, cam):
+        self.attn_cam = cam
+
+    def get_attn_cam(self):
+        return self.attn_cam
+
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+
+    def get_attn_gradients(self):
+        return self.attn_gradients
+
+    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 transpose_for_scores_relprop(self, x):
+        return x.permute(0, 2, 1, 3).flatten(2)
+
+    def forward(
+            self,
+            hidden_states,
+            attention_mask=None,
+            head_mask=None,
+            encoder_hidden_states=None,
+            encoder_attention_mask=None,
+            output_attentions=False,
+    ):
+        self.head_mask = head_mask
+        self.attention_mask = attention_mask
+
+        h1, h2, h3 = self.clone(hidden_states, 3)
+        mixed_query_layer = self.query(h1)
+
+        # 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.
+        if encoder_hidden_states is not None:
+            mixed_key_layer = self.key(encoder_hidden_states)
+            mixed_value_layer = self.value(encoder_hidden_states)
+            attention_mask = encoder_attention_mask
+        else:
+            mixed_key_layer = self.key(h2)
+            mixed_value_layer = self.value(h3)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+        key_layer = self.transpose_for_scores(mixed_key_layer)
+        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = self.matmul1([query_layer, key_layer.transpose(-1, -2)])
+        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 BertModel forward() function)
+            attention_scores = self.add([attention_scores, attention_mask])
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = self.softmax(attention_scores)
+
+        self.save_attn(attention_probs)
+        attention_probs.register_hook(self.save_attn_gradients)
+
+        # 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 = self.matmul2([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,)
+        return outputs
+
+    def relprop(self, cam, **kwargs):
+        # Assume output_attentions == False
+        cam = self.transpose_for_scores(cam)
+
+        # [attention_probs, value_layer]
+        (cam1, cam2) = self.matmul2.relprop(cam, **kwargs)
+        cam1 /= 2
+        cam2 /= 2
+        if self.head_mask is not None:
+            # [attention_probs, head_mask]
+            (cam1, _)= self.mul.relprop(cam1, **kwargs)
+
+
+        self.save_attn_cam(cam1)
+
+        cam1 = self.dropout.relprop(cam1, **kwargs)
+
+        cam1 = self.softmax.relprop(cam1, **kwargs)
+
+        if self.attention_mask is not None:
+            # [attention_scores, attention_mask]
+            (cam1, _) = self.add.relprop(cam1, **kwargs)
+
+        # [query_layer, key_layer.transpose(-1, -2)]
+        (cam1_1, cam1_2) = self.matmul1.relprop(cam1, **kwargs)
+        cam1_1 /= 2
+        cam1_2 /= 2
+
+        # query
+        cam1_1 = self.transpose_for_scores_relprop(cam1_1)
+        cam1_1 = self.query.relprop(cam1_1, **kwargs)
+
+        # key
+        cam1_2 = self.transpose_for_scores_relprop(cam1_2.transpose(-1, -2))
+        cam1_2 = self.key.relprop(cam1_2, **kwargs)
+
+        # value
+        cam2 = self.transpose_for_scores_relprop(cam2)
+        cam2 = self.value.relprop(cam2, **kwargs)
+
+        cam = self.clone.relprop((cam1_1, cam1_2, cam2), **kwargs)
+
+        return cam
+
+
+class BertSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = Dropout(config.hidden_dropout_prob)
+        self.add = Add()
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        add = self.add([hidden_states, input_tensor])
+        hidden_states = self.LayerNorm(add)
+        return hidden_states
+
+    def relprop(self, cam, **kwargs):
+        cam = self.LayerNorm.relprop(cam, **kwargs)
+        # [hidden_states, input_tensor]
+        (cam1, cam2) = self.add.relprop(cam, **kwargs)
+        cam1 = self.dropout.relprop(cam1, **kwargs)
+        cam1 = self.dense.relprop(cam1, **kwargs)
+
+        return (cam1, cam2)
+
+
+class BertIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = 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):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+    def relprop(self, cam, **kwargs):
+        cam = self.intermediate_act_fn.relprop(cam, **kwargs)  # FIXME only ReLU
+        #print(cam.sum())
+        cam = self.dense.relprop(cam, **kwargs)
+        #print(cam.sum())
+        return cam
+
+
+class BertOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = Dropout(config.hidden_dropout_prob)
+        self.add = Add()
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        add = self.add([hidden_states, input_tensor])
+        hidden_states = self.LayerNorm(add)
+        return hidden_states
+
+    def relprop(self, cam, **kwargs):
+        # print("in", cam.sum())
+        cam = self.LayerNorm.relprop(cam, **kwargs)
+        #print(cam.sum())
+        # [hidden_states, input_tensor]
+        (cam1, cam2)= self.add.relprop(cam, **kwargs)
+        # print("add", cam1.sum(), cam2.sum(), cam1.sum() + cam2.sum())
+        cam1 = self.dropout.relprop(cam1, **kwargs)
+        #print(cam1.sum())
+        cam1 = self.dense.relprop(cam1, **kwargs)
+        # print("dense", cam1.sum())
+
+        # print("out", cam1.sum() + cam2.sum(), cam1.sum(), cam2.sum())
+        return (cam1, cam2)
+
+
+class BertLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.attention = BertAttention(config)
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config)
+        self.clone = Clone()
+
+    def forward(
+            self,
+            hidden_states,
+            attention_mask=None,
+            head_mask=None,
+            output_attentions=False,
+    ):
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        ao1, ao2 = self.clone(attention_output, 2)
+        intermediate_output = self.intermediate(ao1)
+        layer_output = self.output(intermediate_output, ao2)
+
+        outputs = (layer_output,) + outputs
+        return outputs
+
+    def relprop(self, cam, **kwargs):
+        (cam1, cam2) = self.output.relprop(cam, **kwargs)
+        # print("output", cam1.sum(), cam2.sum(), cam1.sum() + cam2.sum())
+        cam1 = self.intermediate.relprop(cam1, **kwargs)
+        # print("intermediate", cam1.sum())
+        cam = self.clone.relprop((cam1, cam2), **kwargs)
+        # print("clone", cam.sum())
+        cam = self.attention.relprop(cam, **kwargs)
+        # print("attention", cam.sum())
+        return cam
+
+
+class BertModel(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BertEmbeddings(config)
+        self.encoder = BertEncoder(config)
+        self.pooler = BertPooler(config)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def forward(
+            self,
+            input_ids=None,
+            attention_mask=None,
+            token_type_ids=None,
+            position_ids=None,
+            head_mask=None,
+            inputs_embeds=None,
+            encoder_hidden_states=None,
+            encoder_attention_mask=None,
+            output_attentions=None,
+            output_hidden_states=None,
+            return_dict=None,
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(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 (:obj:`torch.FloatTensor` of shape :obj:`(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 MASKED tokens.
+        """
+        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 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")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            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
+        )
+
+        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,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output)
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+    def relprop(self, cam, **kwargs):
+        cam = self.pooler.relprop(cam, **kwargs)
+        # print("111111111111",cam.sum())
+        cam = self.encoder.relprop(cam, **kwargs)
+        # print("222222222222222", cam.sum())
+        # print("conservation: ", cam.sum())
+        return cam
+
+
+if __name__ == '__main__':
+    class Config:
+      def __init__(self, hidden_size, num_attention_heads, attention_probs_dropout_prob):
+        self.hidden_size = hidden_size
+        self.num_attention_heads = num_attention_heads
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+
+    model = BertSelfAttention(Config(1024, 4, 0.1))
+    x = torch.rand(2, 20, 1024)
+    x.requires_grad_()
+
+    model.eval()
+
+    y = model.forward(x)
+
+    relprop = model.relprop(torch.rand(2, 20, 1024), (torch.rand(2, 20, 1024),))
+
+    print(relprop[1][0].shape)

BERT_explainability/BERTalt.py

@@ -0,0 +1,551 @@
+from __future__ import absolute_import
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+import math
+from BERT_explainability.modules.layers_ours import *
+
+import transformers
+
+from transformers import BertConfig
+from transformers.modeling_outputs import BaseModelOutputWithPooling, BaseModelOutput
+from transformers import (
+    BertPreTrainedModel,
+    PreTrainedModel,
+)
+
+
+ACT2FN = {
+    "relu": ReLU,
+    "tanh": Tanh,
+    "gelu": GELU,
+}
+
+
+def get_activation(activation_string):
+    if activation_string in ACT2FN:
+        return ACT2FN[activation_string]
+    else:
+        raise KeyError("function {} not found in ACT2FN mapping {}".format(activation_string, list(ACT2FN.keys())))
+
+def compute_rollout_attention(all_layer_matrices, start_layer=0):
+    # adding residual consideration
+    num_tokens = all_layer_matrices[0].shape[1]
+    batch_size = all_layer_matrices[0].shape[0]
+    eye = torch.eye(num_tokens).expand(batch_size, num_tokens, num_tokens).to(all_layer_matrices[0].device)
+    all_layer_matrices = [all_layer_matrices[i] + eye for i in range(len(all_layer_matrices))]
+    all_layer_matrices = [all_layer_matrices[i] / all_layer_matrices[i].sum(dim=-1, keepdim=True)
+                          for i in range(len(all_layer_matrices))]
+    joint_attention = all_layer_matrices[start_layer]
+    for i in range(start_layer+1, len(all_layer_matrices)):
+        joint_attention = all_layer_matrices[i].bmm(joint_attention)
+    return joint_attention
+
+class RPBertEmbeddings(BertEmbeddings):
+    def __init__(self, config):
+        super().__init__()
+
+        self.add1 = Add()
+        self.add2 = Add()
+
+    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, :seq_length]
+
+        if token_type_ids is None:
+            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)
+        position_embeddings = self.position_embeddings(position_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        # embeddings = inputs_embeds + position_embeddings + token_type_embeddings
+        embeddings = self.add1([token_type_embeddings, position_embeddings])
+        embeddings = self.add2([embeddings, inputs_embeds])
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+    def relprop(self, cam, **kwargs):
+        cam = self.dropout.relprop(cam, **kwargs)
+        cam = self.LayerNorm.relprop(cam, **kwargs)
+
+        # [inputs_embeds, position_embeddings, token_type_embeddings]
+        (cam) = self.add2.relprop(cam, **kwargs)
+
+        return cam
+
+class RPBertEncoder(transformers.modeling_bert.BertEncoder):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
+
+    def relprop(self, cam, **kwargs):
+        # assuming output_hidden_states is False
+        for layer_module in reversed(self.layer):
+            cam = layer_module.relprop(cam, **kwargs)
+        return cam
+
+
+# not adding relprop since this is only pooling at the end of the network, does not impact tokens importance
+class RPBertPooler(transformers.modeling_bert.BertPooler):
+    def __init__(self, config):
+        super().__init__()
+        self.pool = IndexSelect()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        self._seq_size = hidden_states.shape[1]
+
+        # first_token_tensor = hidden_states[:, 0]
+        first_token_tensor = self.pool(hidden_states, 1, torch.tensor(0, device=hidden_states.device))
+        first_token_tensor = first_token_tensor.squeeze(1)
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+    def relprop(self, cam, **kwargs):
+        cam = self.activation.relprop(cam, **kwargs)
+        #print(cam.sum())
+        cam = self.dense.relprop(cam, **kwargs)
+        #print(cam.sum())
+        cam = cam.unsqueeze(1)
+        cam = self.pool.relprop(cam, **kwargs)
+        #print(cam.sum())
+
+        return cam
+
+class BertAttention(transformers.modeling_bert.BertAttention):
+    def __init__(self, config):
+        super().__init__()
+        self.clone = Clone()
+
+    def forward(
+            self,
+            hidden_states,
+            attention_mask=None,
+            head_mask=None,
+            encoder_hidden_states=None,
+            encoder_attention_mask=None,
+            output_attentions=False,
+    ):
+        h1, h2 = self.clone(hidden_states, 2)
+        self_outputs = self.self(
+            h1,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], h2)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+    def relprop(self, cam, **kwargs):
+        # assuming that we don't ouput the attentions (outputs = (attention_output,)), self_outputs=(context_layer,)
+        (cam1, cam2) = self.output.relprop(cam, **kwargs)
+        #print(cam1.sum(), cam2.sum(), (cam1 + cam2).sum())
+        cam1 = self.self.relprop(cam1, **kwargs)
+        #print(cam1.sum(), cam2.sum(), (cam1 + cam2).sum())
+
+        return self.clone.relprop((cam1, cam2), **kwargs)
+
+class BertSelfAttention(transformers.modeling_bert.BertSelfAttention):
+    def __init__(self, config):
+        super().__init__()
+
+        self.matmul1 = MatMul()
+        self.matmul2 = MatMul()
+        self.softmax = Softmax(dim=-1)
+        self.add = Add()
+        self.mul = Mul()
+        self.head_mask = None
+        self.attention_mask = None
+        self.clone = Clone()
+
+        self.attn_cam = None
+        self.attn = None
+        self.attn_gradients = None
+
+    def get_attn(self):
+        return self.attn
+
+    def save_attn(self, attn):
+        self.attn = attn
+
+    def save_attn_cam(self, cam):
+        self.attn_cam = cam
+
+    def get_attn_cam(self):
+        return self.attn_cam
+
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+
+    def get_attn_gradients(self):
+        return self.attn_gradients
+
+    def transpose_for_scores_relprop(self, x):
+        return x.permute(0, 2, 1, 3).flatten(2)
+
+    def forward(
+            self,
+            hidden_states,
+            attention_mask=None,
+            head_mask=None,
+            encoder_hidden_states=None,
+            encoder_attention_mask=None,
+            output_attentions=False,
+    ):
+        self.head_mask = head_mask
+        self.attention_mask = attention_mask
+
+        h1, h2, h3 = self.clone(hidden_states, 3)
+        mixed_query_layer = self.query(h1)
+
+        # 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.
+        if encoder_hidden_states is not None:
+            mixed_key_layer = self.key(encoder_hidden_states)
+            mixed_value_layer = self.value(encoder_hidden_states)
+            attention_mask = encoder_attention_mask
+        else:
+            mixed_key_layer = self.key(h2)
+            mixed_value_layer = self.value(h3)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+        key_layer = self.transpose_for_scores(mixed_key_layer)
+        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = self.matmul1([query_layer, key_layer.transpose(-1, -2)])
+        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 BertModel forward() function)
+            attention_scores = self.add([attention_scores, attention_mask])
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = self.softmax(attention_scores)
+
+        self.save_attn(attention_probs)
+        attention_probs.register_hook(self.save_attn_gradients)
+
+        # 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 = self.matmul2([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,)
+        return outputs
+
+    def relprop(self, cam, **kwargs):
+        # Assume output_attentions == False
+        cam = self.transpose_for_scores(cam)
+
+        # [attention_probs, value_layer]
+        (cam1, cam2) = self.matmul2.relprop(cam, **kwargs)
+        cam1 /= 2
+        cam2 /= 2
+        if self.head_mask is not None:
+            # [attention_probs, head_mask]
+            (cam1, _)= self.mul.relprop(cam1, **kwargs)
+
+
+        self.save_attn_cam(cam1)
+
+        cam1 = self.dropout.relprop(cam1, **kwargs)
+
+        cam1 = self.softmax.relprop(cam1, **kwargs)
+
+        if self.attention_mask is not None:
+            # [attention_scores, attention_mask]
+            (cam1, _) = self.add.relprop(cam1, **kwargs)
+
+        # [query_layer, key_layer.transpose(-1, -2)]
+        (cam1_1, cam1_2) = self.matmul1.relprop(cam1, **kwargs)
+        cam1_1 /= 2
+        cam1_2 /= 2
+
+        # query
+        cam1_1 = self.transpose_for_scores_relprop(cam1_1)
+        cam1_1 = self.query.relprop(cam1_1, **kwargs)
+
+        # key
+        cam1_2 = self.transpose_for_scores_relprop(cam1_2.transpose(-1, -2))
+        cam1_2 = self.key.relprop(cam1_2, **kwargs)
+
+        # value
+        cam2 = self.transpose_for_scores_relprop(cam2)
+        cam2 = self.value.relprop(cam2, **kwargs)
+
+        cam = self.clone.relprop((cam1_1, cam1_2, cam2), **kwargs)
+
+        return cam
+
+
+class BertSelfOutput(transformers.modeling_bert.BertSelfOutput):
+    def __init__(self, config):
+        super().__init__()
+        self.add = Add()
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        add = self.add([hidden_states, input_tensor])
+        hidden_states = self.LayerNorm(add)
+        return hidden_states
+
+    def relprop(self, cam, **kwargs):
+        cam = self.LayerNorm.relprop(cam, **kwargs)
+        # [hidden_states, input_tensor]
+        (cam1, cam2) = self.add.relprop(cam, **kwargs)
+        cam1 = self.dropout.relprop(cam1, **kwargs)
+        cam1 = self.dense.relprop(cam1, **kwargs)
+
+        return (cam1, cam2)
+
+
+class BertIntermediate(transformers.modeling_bert.BertIntermediate):
+    def relprop(self, cam, **kwargs):
+        cam = self.intermediate_act_fn.relprop(cam, **kwargs)  # FIXME only ReLU
+        #print(cam.sum())
+        cam = self.dense.relprop(cam, **kwargs)
+        #print(cam.sum())
+        return cam
+
+
+class BertOutput(transformers.modeling_bert.BertOutput):
+    def __init__(self, config):
+        super().__init__()
+        self.add = Add()
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        add = self.add([hidden_states, input_tensor])
+        hidden_states = self.LayerNorm(add)
+        return hidden_states
+
+    def relprop(self, cam, **kwargs):
+        # print("in", cam.sum())
+        cam = self.LayerNorm.relprop(cam, **kwargs)
+        #print(cam.sum())
+        # [hidden_states, input_tensor]
+        (cam1, cam2)= self.add.relprop(cam, **kwargs)
+        # print("add", cam1.sum(), cam2.sum(), cam1.sum() + cam2.sum())
+        cam1 = self.dropout.relprop(cam1, **kwargs)
+        #print(cam1.sum())
+        cam1 = self.dense.relprop(cam1, **kwargs)
+        # print("dense", cam1.sum())
+
+        # print("out", cam1.sum() + cam2.sum(), cam1.sum(), cam2.sum())
+        return (cam1, cam2)
+
+
+class RPBertLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.attention = BertAttention(config)
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config)
+        self.clone = Clone()
+
+    def forward(
+            self,
+            hidden_states,
+            attention_mask=None,
+            head_mask=None,
+            output_attentions=False,
+    ):
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        ao1, ao2 = self.clone(attention_output, 2)
+        intermediate_output = self.intermediate(ao1)
+        layer_output = self.output(intermediate_output, ao2)
+
+        outputs = (layer_output,) + outputs
+        return outputs
+
+    def relprop(self, cam, **kwargs):
+        (cam1, cam2) = self.output.relprop(cam, **kwargs)
+        # print("output", cam1.sum(), cam2.sum(), cam1.sum() + cam2.sum())
+        cam1 = self.intermediate.relprop(cam1, **kwargs)
+        # print("intermediate", cam1.sum())
+        cam = self.clone.relprop((cam1, cam2), **kwargs)
+        # print("clone", cam.sum())
+        cam = self.attention.relprop(cam, **kwargs)
+        # print("attention", cam.sum())
+        return cam
+
+
+class BertModel(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BertEmbeddings(config)
+        self.encoder = BertEncoder(config)
+        self.pooler = BertPooler(config)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def forward(
+            self,
+            input_ids=None,
+            attention_mask=None,
+            token_type_ids=None,
+            position_ids=None,
+            head_mask=None,
+            inputs_embeds=None,
+            encoder_hidden_states=None,
+            encoder_attention_mask=None,
+            output_attentions=None,
+            output_hidden_states=None,
+            return_dict=None,
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(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 (:obj:`torch.FloatTensor` of shape :obj:`(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 MASKED tokens.
+        """
+        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 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")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            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
+        )
+
+        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,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output)
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+    def relprop(self, cam, **kwargs):
+        cam = self.pooler.relprop(cam, **kwargs)
+        # print("111111111111",cam.sum())
+        cam = self.encoder.relprop(cam, **kwargs)
+        # print("222222222222222", cam.sum())
+        # print("conservation: ", cam.sum())
+        return cam
+
+
+transformers.modeling_bert.BertEmbeddings = RPBertEmbeddings
+transformers.modeling_bert.BertEncoder = RPBertEncoder
+
+if __name__ == '__main__':
+    class Config:
+      def __init__(self, hidden_size, num_attention_heads, attention_probs_dropout_prob):
+        self.hidden_size = hidden_size
+        self.num_attention_heads = num_attention_heads
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+
+    model = BertSelfAttention(Config(1024, 4, 0.1))
+    x = torch.rand(2, 20, 1024)
+    x.requires_grad_()
+
+    model.eval()
+
+    y = model.forward(x)
+
+    relprop = model.relprop(torch.rand(2, 20, 1024), (torch.rand(2, 20, 1024),))
+
+    print(relprop[1][0].shape)

BERT_explainability/BertForSequenceClassification.py

@@ -0,0 +1,204 @@
+from transformers import BertPreTrainedModel
+from transformers.modeling_outputs import SequenceClassifierOutput
+from transformers.utils import logging
+from BERT_explainability.modules.layers_ours import *
+from BERT_explainability.modules.BERT.BERT import BertModel
+from torch.nn import CrossEntropyLoss, MSELoss
+import torch.nn as nn
+from typing import List, Any
+import torch
+from BERT_rationale_benchmark.models.model_utils import PaddedSequence
+
+
+class BertForSequenceClassification(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config)
+        self.dropout = Dropout(config.hidden_dropout_prob)
+        self.classifier = Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(
+            self,
+            input_ids=None,
+            attention_mask=None,
+            token_type_ids=None,
+            position_ids=None,
+            head_mask=None,
+            inputs_embeds=None,
+            labels=None,
+            output_attentions=None,
+            output_hidden_states=None,
+            return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
+            If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`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
+
+        outputs = self.bert(
+            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,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        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,
+        )
+
+    def relprop(self, cam=None, **kwargs):
+        cam = self.classifier.relprop(cam, **kwargs)
+        cam = self.dropout.relprop(cam, **kwargs)
+        cam = self.bert.relprop(cam, **kwargs)
+        # print("conservation: ", cam.sum())
+        return cam
+
+
+# this is the actual classifier we will be using
+class BertClassifier(nn.Module):
+    """Thin wrapper around BertForSequenceClassification"""
+
+    def __init__(self,
+                 bert_dir: str,
+                 pad_token_id: int,
+                 cls_token_id: int,
+                 sep_token_id: int,
+                 num_labels: int,
+                 max_length: int = 512,
+                 use_half_precision=True):
+        super(BertClassifier, self).__init__()
+        bert = BertForSequenceClassification.from_pretrained(bert_dir, num_labels=num_labels)
+        if use_half_precision:
+            import apex
+            bert = bert.half()
+        self.bert = bert
+        self.pad_token_id = pad_token_id
+        self.cls_token_id = cls_token_id
+        self.sep_token_id = sep_token_id
+        self.max_length = max_length
+
+    def forward(self,
+                query: List[torch.tensor],
+                docids: List[Any],
+                document_batch: List[torch.tensor]):
+        assert len(query) == len(document_batch)
+        print(query)
+        # note about device management:
+        # since distributed training is enabled, the inputs to this module can be on *any* device (preferably cpu, since we wrap and unwrap the module)
+        # we want to keep these params on the input device (assuming CPU) for as long as possible for cheap memory access
+        target_device = next(self.parameters()).device
+        cls_token = torch.tensor([self.cls_token_id]).to(device=document_batch[0].device)
+        sep_token = torch.tensor([self.sep_token_id]).to(device=document_batch[0].device)
+        input_tensors = []
+        position_ids = []
+        for q, d in zip(query, document_batch):
+            if len(q) + len(d) + 2 > self.max_length:
+                d = d[:(self.max_length - len(q) - 2)]
+            input_tensors.append(torch.cat([cls_token, q, sep_token, d]))
+            position_ids.append(torch.tensor(list(range(0, len(q) + 1)) + list(range(0, len(d) + 1))))
+        bert_input = PaddedSequence.autopad(input_tensors, batch_first=True, padding_value=self.pad_token_id,
+                                            device=target_device)
+        positions = PaddedSequence.autopad(position_ids, batch_first=True, padding_value=0, device=target_device)
+        (classes,) = self.bert(bert_input.data,
+                               attention_mask=bert_input.mask(on=0.0, off=float('-inf'), device=target_device),
+                               position_ids=positions.data)
+        assert torch.all(classes == classes)  # for nans
+
+        print(input_tensors[0])
+        print(self.relprop()[0])
+
+        return classes
+
+    def relprop(self, cam=None, **kwargs):
+        return self.bert.relprop(cam, **kwargs)
+
+
+if __name__ == '__main__':
+    from transformers import BertTokenizer
+    import os
+
+    class Config:
+        def __init__(self, hidden_size, num_attention_heads, attention_probs_dropout_prob, num_labels,
+                     hidden_dropout_prob):
+            self.hidden_size = hidden_size
+            self.num_attention_heads = num_attention_heads
+            self.attention_probs_dropout_prob = attention_probs_dropout_prob
+            self.num_labels = num_labels
+            self.hidden_dropout_prob = hidden_dropout_prob
+
+
+    tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+    x = tokenizer.encode_plus("In this movie the acting is great. The movie is perfect! [sep]",
+                         add_special_tokens=True,
+                         max_length=512,
+                         return_token_type_ids=False,
+                         return_attention_mask=True,
+                         pad_to_max_length=True,
+                         return_tensors='pt',
+                         truncation=True)
+
+    print(x['input_ids'])
+
+    model = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)
+    model_save_file = os.path.join('./BERT_explainability/output_bert/movies/classifier/', 'classifier.pt')
+    model.load_state_dict(torch.load(model_save_file))
+
+    # x = torch.randint(100, (2, 20))
+    # x = torch.tensor([[101, 2054, 2003, 1996, 15792, 1997, 2023, 3319, 1029, 102,
+    #                    101, 4079, 102, 101, 6732, 102, 101, 2643, 102, 101,
+    #                    2038, 102, 101, 1037, 102, 101, 2933, 102, 101, 2005,
+    #                    102, 101, 2032, 102, 101, 1010, 102, 101, 1037, 102,
+    #                    101, 3800, 102, 101, 2005, 102, 101, 2010, 102, 101,
+    #                    2166, 102, 101, 1010, 102, 101, 1998, 102, 101, 2010,
+    #                    102, 101, 4650, 102, 101, 1010, 102, 101, 2002, 102,
+    #                    101, 2074, 102, 101, 2515, 102, 101, 1050, 102, 101,
+    #                    1005, 102, 101, 1056, 102, 101, 2113, 102, 101, 2054,
+    #                    102, 101, 1012, 102]])
+    # x.requires_grad_()
+
+    model.eval()
+
+    y = model(x['input_ids'], x['attention_mask'])
+    print(y)
+
+    cam, _ = model.relprop()
+
+    #print(cam.shape)
+
+    cam = cam.sum(-1)
+    #print(cam)

BERT_explainability/ExplanationGenerator.py

@@ -0,0 +1,165 @@
+import argparse
+import numpy as np
+import torch
+import glob
+
+from captum._utils.common import _get_module_from_name
+
+# compute rollout between attention layers
+def compute_rollout_attention(all_layer_matrices, start_layer=0):
+    # adding residual consideration- code adapted from https://github.com/samiraabnar/attention_flow
+    num_tokens = all_layer_matrices[0].shape[1]
+    batch_size = all_layer_matrices[0].shape[0]
+    eye = torch.eye(num_tokens).expand(batch_size, num_tokens, num_tokens).to(all_layer_matrices[0].device)
+    all_layer_matrices = [all_layer_matrices[i] + eye for i in range(len(all_layer_matrices))]
+    matrices_aug = [all_layer_matrices[i] / all_layer_matrices[i].sum(dim=-1, keepdim=True)
+                          for i in range(len(all_layer_matrices))]
+    joint_attention = matrices_aug[start_layer]
+    for i in range(start_layer+1, len(matrices_aug)):
+        joint_attention = matrices_aug[i].bmm(joint_attention)
+    return joint_attention
+
+class Generator:
+    def __init__(self, model, key="bert.encoder.layer"):
+        self.model = model
+        self.key = key
+        self.model.eval()
+
+    def forward(self, input_ids, attention_mask):
+        return self.model(input_ids, attention_mask)
+
+    def _calculate_gradients(self, output, index, do_relprop=True):
+        if index == None:
+            index = np.argmax(output.cpu().data.numpy(), axis=-1)
+
+        one_hot_vector = (torch.nn.functional
+                .one_hot(
+                        # one_hot requires ints
+                        torch.tensor(index, dtype=torch.int64),
+                        num_classes=output.size(-1)
+                    )
+                # but requires_grad_ needs floats
+                .to(torch.float)
+            ).to(output.device)
+
+        hot_output = torch.sum(one_hot_vector.clone().requires_grad_(True) * output)
+        self.model.zero_grad()
+        hot_output.backward(retain_graph=True)
+
+        if do_relprop:
+            return self.model.relprop(one_hot_vector, alpha=1)
+
+    def generate_LRP(self, input_ids, attention_mask,
+                     index=None, start_layer=11):
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+
+        if index == None:
+            index = np.argmax(output.cpu().data.numpy(), axis=-1)
+
+        self._calculate_gradients(output, index)
+
+        cams = []
+        blocks = _get_module_from_name(self.model, self.key)
+        for blk in blocks:
+            grad = blk.attention.self.get_attn_gradients()
+            cam = blk.attention.self.get_attn_cam()
+            cam = cam[0].reshape(-1, cam.shape[-1], cam.shape[-1])
+            grad = grad[0].reshape(-1, grad.shape[-1], grad.shape[-1])
+            cam = grad * cam
+            cam = cam.clamp(min=0).mean(dim=0)
+            cams.append(cam.unsqueeze(0))
+        rollout = compute_rollout_attention(cams, start_layer=start_layer)
+        rollout[:, 0, 0] = rollout[:, 0].min()
+        return rollout[:, 0]
+
+
+    def generate_LRP_last_layer(self, input_ids, attention_mask,
+                     index=None):
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+        if index == None:
+            index = np.argmax(output.cpu().data.numpy(), axis=-1)
+
+        self._calculate_gradients(output, index)
+
+        cam = _get_module_from_name(self.model, self.key)[-1].attention.self.get_attn_cam()[0]
+        cam = cam.clamp(min=0).mean(dim=0).unsqueeze(0)
+        cam[:, 0, 0] = 0
+        return cam[:, 0]
+
+    def generate_full_lrp(self, input_ids, attention_mask,
+                     index=None):
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+
+        if index == None:
+            index = np.argmax(output.cpu().data.numpy(), axis=-1)
+
+        cam = self._calculate_gradients(output, index)
+        cam = cam.sum(dim=2)
+        cam[:, 0] = 0
+        return cam
+
+    def generate_attn_last_layer(self, input_ids, attention_mask,
+                     index=None):
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+        cam = _get_module_from_name(self.model, self.key)[-1].attention.self.get_attn()[0]
+        cam = cam.mean(dim=0).unsqueeze(0)
+        cam[:, 0, 0] = 0
+        return cam[:, 0]
+
+    def generate_rollout(self, input_ids, attention_mask, start_layer=0, index=None):
+        self.model.zero_grad()
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+        blocks = _get_module_from_name(self.model, self.key)
+        all_layer_attentions = []
+        for blk in blocks:
+            attn_heads = blk.attention.self.get_attn()
+            avg_heads = (attn_heads.sum(dim=1) / attn_heads.shape[1]).detach()
+            all_layer_attentions.append(avg_heads)
+        rollout = compute_rollout_attention(all_layer_attentions, start_layer=start_layer)
+        rollout[:, 0, 0] = 0
+        return rollout[:, 0]
+
+    def generate_attn_gradcam(self, input_ids, attention_mask, index=None):
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+
+        if index == None:
+            index = np.argmax(output.cpu().data.numpy(), axis=-1)
+
+        self._calculate_gradients(output, index)
+
+        cam = _get_module_from_name(self.model, self.key)[-1].attention.self.get_attn()
+        grad = _get_module_from_name(self.model, self.key)[-1].attention.self.get_attn_gradients()
+
+        cam = cam[0].reshape(-1, cam.shape[-1], cam.shape[-1])
+        grad = grad[0].reshape(-1, grad.shape[-1], grad.shape[-1])
+        grad = grad.mean(dim=[1, 2], keepdim=True)
+        cam = (cam * grad).mean(0).clamp(min=0).unsqueeze(0)
+        cam = (cam - cam.min()) / (cam.max() - cam.min())
+        cam[:, 0, 0] = 0
+        return cam[:, 0]
+
+    def generate_rollout_attn_gradcam(self, input_ids, attention_mask, index=None, start_layer=0):
+        # rule 5 from paper
+        def avg_heads(cam, grad):
+            return (grad * cam).clamp(min=0).mean(dim=-3)
+
+        # rule 6 from paper
+        def apply_self_attention_rules(R_ss, cam_ss):
+            return torch.matmul(cam_ss, R_ss)
+
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+        blocks = _get_module_from_name(self.model, self.key)
+
+        num_tokens = input_ids.size(-1)
+        R = torch.eye(num_tokens).expand(output.size(0), -1, -1).clone().to(output.device)
+
+        for i, blk in enumerate(model.roberta.encoder.layer):
+            if i < start_layer:
+                continue
+            grad = blk.attention.self.get_attn_gradients().detach()
+            cam = blk.attention.self.get_attn().detach()
+            cam = avg_heads(cam, grad)
+            joint = apply_self_attention_rules(R, cam)
+            R += joint
+        return R[:, 0, 1:-1]
+

BERT_explainability/NewExplanationGenerator.py

@@ -0,0 +1,145 @@
+import argparse
+import numpy as np
+import torch
+import glob
+
+from captum._utils.common import _get_module_from_name
+
+# compute rollout between attention layers
+def compute_rollout_attention(all_layer_matrices, start_layer=0):
+    # adding residual consideration- code adapted from https://github.com/samiraabnar/attention_flow
+    num_tokens = all_layer_matrices[0].shape[1]
+    batch_size = all_layer_matrices[0].shape[0]
+    eye = torch.eye(num_tokens).expand(batch_size, num_tokens, num_tokens).to(all_layer_matrices[0].device)
+    all_layer_matrices = [all_layer_matrices[i] + eye for i in range(len(all_layer_matrices))]
+    matrices_aug = [all_layer_matrices[i] / all_layer_matrices[i].sum(dim=-1, keepdim=True)
+                          for i in range(len(all_layer_matrices))]
+    joint_attention = matrices_aug[start_layer]
+    for i in range(start_layer+1, len(matrices_aug)):
+        joint_attention = matrices_aug[i].bmm(joint_attention)
+    return joint_attention
+
+class Generator:
+    def __init__(self, model, key="bert.encoder.layer"):
+        self.model = model
+        self.key = key
+        self.model.eval()
+
+    def forward(self, input_ids, attention_mask):
+        return self.model(input_ids, attention_mask)
+
+    def _build_one_hot(self, output, index):
+        if index == None:
+            index = np.argmax(output.cpu().data.numpy(), axis=-1)
+
+        one_hot = np.zeros((1, output.size()[-1]), dtype=np.float32)
+        one_hot[0, index] = 1
+        one_hot_vector = one_hot
+        one_hot = torch.from_numpy(one_hot).requires_grad_(True).to(output.device)
+        one_hot = torch.sum(one_hot * output)
+
+        return one_hot, one_hot_vector
+
+    def generate_LRP(self, input_ids, attention_mask,
+                     index=None, start_layer=11):
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+        kwargs = {"alpha": 1}
+
+        one_hot, one_hot_vector = self._build_one_hot(output, index)
+        self.model.zero_grad()
+        one_hot.backward(retain_graph=True)
+
+        self.model.relprop(torch.tensor(one_hot_vector).to(input_ids.device), **kwargs)
+
+        cams = []
+        blocks = _get_module_from_name(self.model, self.key)
+        for blk in blocks:
+            grad = blk.attention.self.get_attn_gradients()
+            cam = blk.attention.self.get_attn_cam()
+            cam = cam[0].reshape(-1, cam.shape[-1], cam.shape[-1])
+            grad = grad[0].reshape(-1, grad.shape[-1], grad.shape[-1])
+            cam = grad * cam
+            cam = cam.clamp(min=0).mean(dim=0)
+            cams.append(cam.unsqueeze(0))
+        rollout = compute_rollout_attention(cams, start_layer=start_layer)
+        rollout[:, 0, 0] = rollout[:, 0].min()
+        return rollout[:, 0]
+
+    def generate_LRP_last_layer(self, input_ids, attention_mask,
+                     index=None):
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+        kwargs = {"alpha": 1}
+
+        one_hot, one_hot_vector = self._build_one_hot(output, index)
+
+        self.model.zero_grad()
+        one_hot.backward(retain_graph=True)
+
+        self.model.relprop(torch.tensor(one_hot_vector).to(input_ids.device), **kwargs)
+
+        cam = _get_module_from_name(self.model, self.key)[-1].attention.self.get_attn_cam()[0]
+        cam = cam.clamp(min=0).mean(dim=0).unsqueeze(0)
+        cam[:, 0, 0] = 0
+        return cam[:, 0]
+
+    def generate_full_lrp(self, input_ids, attention_mask,
+                     index=None):
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+        kwargs = {"alpha": 1}
+
+        one_hot, one_hot_vector = self._build_one_hot(output, index)
+
+        self.model.zero_grad()
+        one_hot.backward(retain_graph=True)
+
+        cam = self.model.relprop(torch.tensor(one_hot_vector).to(input_ids.device), **kwargs)
+        cam = cam.sum(dim=2)
+        cam[:, 0] = 0
+        return cam
+
+    def generate_attn_last_layer(self, input_ids, attention_mask,
+                     index=None):
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+        cam = _get_module_from_name(self.model, self.key)[-1].attention.self.get_attn()[0]
+        cam = cam.mean(dim=0).unsqueeze(0)
+        cam[:, 0, 0] = 0
+        return cam[:, 0]
+
+    def generate_rollout(self, input_ids, attention_mask, start_layer=0, index=None):
+        self.model.zero_grad()
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+        blocks = _get_module_from_name(self.model, self.key)
+        all_layer_attentions = []
+        for blk in blocks:
+            attn_heads = blk.attention.self.get_attn()
+            avg_heads = (attn_heads.sum(dim=1) / attn_heads.shape[1]).detach()
+            all_layer_attentions.append(avg_heads)
+        rollout = compute_rollout_attention(all_layer_attentions, start_layer=start_layer)
+        rollout[:, 0, 0] = 0
+        return rollout[:, 0]
+
+    def generate_attn_gradcam(self, input_ids, attention_mask, index=None):
+        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
+        kwargs = {"alpha": 1}
+
+        if index == None:
+            index = np.argmax(output.cpu().data.numpy(), axis=-1)
+
+        one_hot, one_hot_vector = self._build_one_hot(output, index)
+
+        self.model.zero_grad()
+        one_hot.backward(retain_graph=True)
+
+        self.model.relprop(torch.tensor(one_hot_vector).to(input_ids.device), **kwargs)
+
+        cam = _get_module_from_name(self.model, self.key)[-1].attention.self.get_attn()
+        grad = _get_module_from_name(self.model, self.key)[-1].attention.self.get_attn_gradients()
+
+        cam = cam[0].reshape(-1, cam.shape[-1], cam.shape[-1])
+        grad = grad[0].reshape(-1, grad.shape[-1], grad.shape[-1])
+        grad = grad.mean(dim=[1, 2], keepdim=True)
+        cam = (cam * grad).mean(0).clamp(min=0).unsqueeze(0)
+        cam = (cam - cam.min()) / (cam.max() - cam.min())
+        cam[:, 0, 0] = 0
+        return cam[:, 0]
+

BERT_explainability/RobertaForSequenceClassification.py

@@ -0,0 +1,204 @@
+from transformers import BertPreTrainedModel
+from transformers.modeling_outputs import SequenceClassifierOutput
+from transformers.utils import logging
+from BERT_explainability.modules.layers_ours import *
+from BERT_explainability.modules.BERT.BERT import BertModel
+from torch.nn import CrossEntropyLoss, MSELoss
+import torch.nn as nn
+from typing import List, Any
+import torch
+from BERT_rationale_benchmark.models.model_utils import PaddedSequence
+
+
+class BertForSequenceClassification(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config)
+        self.dropout = Dropout(config.hidden_dropout_prob)
+        self.classifier = Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(
+            self,
+            input_ids=None,
+            attention_mask=None,
+            token_type_ids=None,
+            position_ids=None,
+            head_mask=None,
+            inputs_embeds=None,
+            labels=None,
+            output_attentions=None,
+            output_hidden_states=None,
+            return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
+            If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`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
+
+        outputs = self.bert(
+            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,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        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,
+        )
+
+    def relprop(self, cam=None, **kwargs):
+        cam = self.classifier.relprop(cam, **kwargs)
+        cam = self.dropout.relprop(cam, **kwargs)
+        cam = self.bert.relprop(cam, **kwargs)
+        # print("conservation: ", cam.sum())
+        return cam
+
+
+# this is the actual classifier we will be using
+class BertClassifier(nn.Module):
+    """Thin wrapper around BertForSequenceClassification"""
+
+    def __init__(self,
+                 bert_dir: str,
+                 pad_token_id: int,
+                 cls_token_id: int,
+                 sep_token_id: int,
+                 num_labels: int,
+                 max_length: int = 512,
+                 use_half_precision=True):
+        super(BertClassifier, self).__init__()
+        bert = BertForSequenceClassification.from_pretrained(bert_dir, num_labels=num_labels)
+        if use_half_precision:
+            import apex
+            bert = bert.half()
+        self.bert = bert
+        self.pad_token_id = pad_token_id
+        self.cls_token_id = cls_token_id
+        self.sep_token_id = sep_token_id
+        self.max_length = max_length
+
+    def forward(self,
+                query: List[torch.tensor],
+                docids: List[Any],
+                document_batch: List[torch.tensor]):
+        assert len(query) == len(document_batch)
+        print(query)
+        # note about device management:
+        # since distributed training is enabled, the inputs to this module can be on *any* device (preferably cpu, since we wrap and unwrap the module)
+        # we want to keep these params on the input device (assuming CPU) for as long as possible for cheap memory access
+        target_device = next(self.parameters()).device
+        cls_token = torch.tensor([self.cls_token_id]).to(device=document_batch[0].device)
+        sep_token = torch.tensor([self.sep_token_id]).to(device=document_batch[0].device)
+        input_tensors = []
+        position_ids = []
+        for q, d in zip(query, document_batch):
+            if len(q) + len(d) + 2 > self.max_length:
+                d = d[:(self.max_length - len(q) - 2)]
+            input_tensors.append(torch.cat([cls_token, q, sep_token, d]))
+            position_ids.append(torch.tensor(list(range(0, len(q) + 1)) + list(range(0, len(d) + 1))))
+        bert_input = PaddedSequence.autopad(input_tensors, batch_first=True, padding_value=self.pad_token_id,
+                                            device=target_device)
+        positions = PaddedSequence.autopad(position_ids, batch_first=True, padding_value=0, device=target_device)
+        (classes,) = self.bert(bert_input.data,
+                               attention_mask=bert_input.mask(on=0.0, off=float('-inf'), device=target_device),
+                               position_ids=positions.data)
+        assert torch.all(classes == classes)  # for nans
+
+        print(input_tensors[0])
+        print(self.relprop()[0])
+
+        return classes
+
+    def relprop(self, cam=None, **kwargs):
+        return self.bert.relprop(cam, **kwargs)
+
+
+if __name__ == '__main__':
+    from transformers import BertTokenizer
+    import os
+
+    class Config:
+        def __init__(self, hidden_size, num_attention_heads, attention_probs_dropout_prob, num_labels,
+                     hidden_dropout_prob):
+            self.hidden_size = hidden_size
+            self.num_attention_heads = num_attention_heads
+            self.attention_probs_dropout_prob = attention_probs_dropout_prob
+            self.num_labels = num_labels
+            self.hidden_dropout_prob = hidden_dropout_prob
+
+
+    tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+    x = tokenizer.encode_plus("In this movie the acting is great. The movie is perfect! [sep]",
+                         add_special_tokens=True,
+                         max_length=512,
+                         return_token_type_ids=False,
+                         return_attention_mask=True,
+                         pad_to_max_length=True,
+                         return_tensors='pt',
+                         truncation=True)
+
+    print(x['input_ids'])
+
+    model = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)
+    model_save_file = os.path.join('./BERT_explainability/output_bert/movies/classifier/', 'classifier.pt')
+    model.load_state_dict(torch.load(model_save_file))
+
+    # x = torch.randint(100, (2, 20))
+    # x = torch.tensor([[101, 2054, 2003, 1996, 15792, 1997, 2023, 3319, 1029, 102,
+    #                    101, 4079, 102, 101, 6732, 102, 101, 2643, 102, 101,
+    #                    2038, 102, 101, 1037, 102, 101, 2933, 102, 101, 2005,
+    #                    102, 101, 2032, 102, 101, 1010, 102, 101, 1037, 102,
+    #                    101, 3800, 102, 101, 2005, 102, 101, 2010, 102, 101,
+    #                    2166, 102, 101, 1010, 102, 101, 1998, 102, 101, 2010,
+    #                    102, 101, 4650, 102, 101, 1010, 102, 101, 2002, 102,
+    #                    101, 2074, 102, 101, 2515, 102, 101, 1050, 102, 101,
+    #                    1005, 102, 101, 1056, 102, 101, 2113, 102, 101, 2054,
+    #                    102, 101, 1012, 102]])
+    # x.requires_grad_()
+
+    model.eval()
+
+    y = model(x['input_ids'], x['attention_mask'])
+    print(y)
+
+    cam, _ = model.relprop()
+
+    #print(cam.shape)
+
+    cam = cam.sum(-1)
+    #print(cam)

BERT_explainability/roberta2.py

@@ -0,0 +1,1596 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch RoBERTa model."""
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from packaging import version
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN, gelu
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from transformers.utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.models.roberta.configuration_roberta import RobertaConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "roberta-base"
+_CONFIG_FOR_DOC = "RobertaConfig"
+_TOKENIZER_FOR_DOC = "RobertaTokenizer"
+
+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 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 get_attn(self):
+        return self.attn
+
+    def save_attn(self, attn):
+        self.attn = attn
+
+    def save_attn_cam(self, cam):
+        self.attn_cam = cam
+
+    def get_attn_cam(self):
+        return self.attn_cam
+
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+
+    def get_attn_gradients(self):
+        return self.attn_gradients
+
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        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,
+    ) -> 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
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        self.save_attn(attention_probs)
+        attention_probs.register_hook(self.save_attn_gradients)
+
+        # 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,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        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,
+    ) -> 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,
+        )
+        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,
+    ) -> 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:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                )
+
+            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):
+            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.
+"""
+
+
+@add_start_docstrings(
+    "The bare RoBERTa Model transformer outputting raw hidden-states without any specific head on top.",
+    ROBERTA_START_DOCSTRING,
+)
+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)
+
+    @add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPoolingAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    # 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,
+    ) -> 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)
+
+        # 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,
+        )
+        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,
+        )
+
+
+@add_start_docstrings(
+    """RoBERTa Model with a `language modeling` head on top for CLM fine-tuning.""", ROBERTA_START_DOCSTRING
+)
+class RobertaForCausalLM(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 not config.is_decoder:
+            logger.warning("If you want to use `RobertaLMHeadModel` as a standalone, add `is_decoder=True.`")
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        self.lm_head = RobertaLMHead(config)
+
+        # 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
+
+    @add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
+    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,
+        past_key_values: Tuple[Tuple[torch.FloatTensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
+        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**.
+
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the left-to-right language modeling loss (next word prediction). 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]`
+        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`).
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import RobertaTokenizer, RobertaForCausalLM, RobertaConfig
+        >>> import torch
+
+        >>> tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
+        >>> config = RobertaConfig.from_pretrained("roberta-base")
+        >>> config.is_decoder = True
+        >>> model = RobertaForCausalLM.from_pretrained("roberta-base", config=config)
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> outputs = model(**inputs)
+
+        >>> prediction_logits = outputs.logits
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        if labels is not None:
+            use_cache = False
+
+        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,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+        prediction_scores = self.lm_head(sequence_output)
+
+        lm_loss = None
+        if labels is not None:
+            # we are doing next-token prediction; shift prediction scores and input ids by one
+            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
+            labels = labels[:, 1:].contiguous()
+            loss_fct = CrossEntropyLoss()
+            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((lm_loss,) + output) if lm_loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=lm_loss,
+            logits=prediction_scores,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, **model_kwargs):
+        input_shape = input_ids.shape
+        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
+        if attention_mask is None:
+            attention_mask = input_ids.new_ones(input_shape)
+
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            input_ids = input_ids[:, -1:]
+
+        return {"input_ids": input_ids, "attention_mask": attention_mask, "past_key_values": past}
+
+    def _reorder_cache(self, past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past
+
+
+@add_start_docstrings("""RoBERTa Model with a `language modeling` head on top.""", ROBERTA_START_DOCSTRING)
+class RobertaForMaskedLM(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."
+            )
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        self.lm_head = RobertaLMHead(config)
+
+        # 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
+
+    @add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+        mask="<mask>",
+        expected_output="' Paris'",
+        expected_loss=0.1,
+    )
+    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[torch.Tensor], 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
+
+        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,
+        )
+        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
+
+
+@add_start_docstrings(
+    """
+    RoBERTa Model transformer with a sequence classification/regression head on top (a linear layer on top of the
+    pooled output) e.g. for GLUE tasks.
+    """,
+    ROBERTA_START_DOCSTRING,
+)
+class RobertaForSequenceClassification(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
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        self.classifier = RobertaClassificationHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint="cardiffnlp/twitter-roberta-base-emotion",
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output="'optimism'",
+        expected_loss=0.08,
+    )
+    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[torch.Tensor], 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
+
+        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,
+        )
+        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,
+        )
+
+
+@add_start_docstrings(
+    """
+    Roberta Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
+    softmax) e.g. for RocStories/SWAG tasks.
+    """,
+    ROBERTA_START_DOCSTRING,
+)
+class RobertaForMultipleChoice(RobertaPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.roberta = RobertaModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], MultipleChoiceModelOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
+            num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
+            `input_ids` above)
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
+
+        flat_input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        flat_position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        flat_token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        flat_inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+
+        outputs = self.roberta(
+            flat_input_ids,
+            position_ids=flat_position_ids,
+            token_type_ids=flat_token_type_ids,
+            attention_mask=flat_attention_mask,
+            head_mask=head_mask,
+            inputs_embeds=flat_inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return MultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Roberta Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    ROBERTA_START_DOCSTRING,
+)
+class RobertaForTokenClassification(RobertaPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint="Jean-Baptiste/roberta-large-ner-english",
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output="['O', 'ORG', 'ORG', 'O', 'O', 'O', 'O', 'O', 'LOC', 'O', 'LOC', 'LOC']",
+        expected_loss=0.01,
+    )
+    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[torch.Tensor], TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        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,
+        )
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            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
+
+
+@add_start_docstrings(
+    """
+    Roberta Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    ROBERTA_START_DOCSTRING,
+)
+class RobertaForQuestionAnswering(RobertaPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.roberta = RobertaModel(config, add_pooling_layer=False)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint="deepset/roberta-base-squad2",
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output="' puppet'",
+        expected_loss=0.86,
+    )
+    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,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], QuestionAnsweringModelOutput]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        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,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+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

BERT_explainability/roberta2.py.rej

@@ -0,0 +1,63 @@
+--- modeling_roberta.py	2022-06-28 11:59:19.974278244 +0200
++++ roberta2.py	2022-06-28 14:13:05.765050058 +0200
+@@ -23,14 +23,14 @@
+ from torch import nn
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+ 
+-from ...activations import ACT2FN, gelu
+-from ...file_utils import (
++from transformers.activations import ACT2FN, gelu
++from transformers.file_utils import (
+     add_code_sample_docstrings,
+     add_start_docstrings,
+     add_start_docstrings_to_model_forward,
+     replace_return_docstrings,
+ )
+-from ...modeling_outputs import (
++from transformers.modeling_outputs import (
+     BaseModelOutputWithPastAndCrossAttentions,
+     BaseModelOutputWithPoolingAndCrossAttentions,
+     CausalLMOutputWithCrossAttentions,
+@@ -40,14 +40,14 @@
+     SequenceClassifierOutput,
+     TokenClassifierOutput,
+ )
+-from ...modeling_utils import (
++from transformers.modeling_utils import (
+     PreTrainedModel,
+     apply_chunking_to_forward,
+     find_pruneable_heads_and_indices,
+     prune_linear_layer,
+ )
+-from ...utils import logging
+-from .configuration_roberta import RobertaConfig
++from transformers.utils import logging
++from transformers.models.roberta.configuration_roberta import RobertaConfig
+ 
+ 
+ logger = logging.get_logger(__name__)
+@@ -183,6 +183,24 @@
+ 
+         self.is_decoder = config.is_decoder
+ 
++    def get_attn(self):
++        return self.attn
++
++    def save_attn(self, attn):
++        self.attn = attn
++
++    def save_attn_cam(self, cam):
++        self.attn_cam = cam
++
++    def get_attn_cam(self):
++        return self.attn_cam
++
++    def save_attn_gradients(self, attn_gradients):
++        self.attn_gradients = attn_gradients
++
++    def get_attn_gradients(self):
++        return self.attn_gradients
++
+     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)

app.py

@@ -1,7 +1,194 @@
+import sys
 import gradio
 
-def greet(name):
-    return f"Hello {name}. Check back soon for real content"
+sys.path.append("BERT_explainability")
 
-iface = gradio.Interface(fn=greet, inputs="text", outputs="text")
+import torch
+
+from BERT_explainability.ExplanationGenerator import Generator
+from BERT_explainability.roberta2 import RobertaForSequenceClassification
+from transformers import AutoTokenizer
+
+from captum.attr import (
+        visualization
+    )
+import torch
+
+# from https://discuss.pytorch.org/t/using-scikit-learns-scalers-for-torchvision/53455
+class PyTMinMaxScalerVectorized(object):
+    """
+    Transforms each channel to the range [0, 1].
+    """
+    def __init__(self, dimension=-1):
+        self.d = dimension
+    def __call__(self, tensor):
+        d = self.d
+        scale = 1.0 / (tensor.max(dim=d, keepdim=True)[0] - tensor.min(dim=d, keepdim=True)[0])
+        tensor.mul_(scale).sub_(tensor.min(dim=d, keepdim=True)[0])
+        return tensor
+
+
+if torch.cuda.is_available():
+    device = torch.device("cuda")
+else:
+    device = torch.device("cpu")
+
+model = RobertaForSequenceClassification.from_pretrained("textattack/roberta-base-SST-2").to(device)
+model.eval()
+tokenizer = AutoTokenizer.from_pretrained("textattack/roberta-base-SST-2")
+# initialize the explanations generator
+explanations = Generator(model, "roberta")
+
+classifications = ["NEGATIVE", "POSITIVE"]
+
+# rule 5 from paper
+def avg_heads(cam, grad):
+    cam = (
+            (grad * cam)
+            .clamp(min=0)
+            .mean(dim=-3)
+        )
+    # set negative values to 0, then average
+#    cam = cam.clamp(min=0).mean(dim=0)
+    return cam
+
+# rule 6 from paper
+def apply_self_attention_rules(R_ss, cam_ss):
+    R_ss_addition = torch.matmul(cam_ss, R_ss)
+    return R_ss_addition
+
+def generate_relevance(model, input_ids, attention_mask, index=None, start_layer=0):
+    output = model(input_ids=input_ids, attention_mask=attention_mask)[0]
+    if index == None:
+        #index = np.expand_dims(np.arange(input_ids.shape[1])
+        # by default explain the class with the highest score
+        index = output.argmax(axis=-1).detach().cpu().numpy()
+
+    # create a one-hot vector selecting class we want explanations for
+    one_hot = (torch.nn.functional
+            .one_hot(torch.tensor(index, dtype=torch.int64), num_classes=output.size(-1))
+            .to(torch.float)
+            .requires_grad_(True)
+        ).to(device)
+    print("ONE_HOT", one_hot.size(), one_hot)
+    one_hot = torch.sum(one_hot * output)
+    model.zero_grad()
+    # create the gradients for the class we're interested in
+    one_hot.backward(retain_graph=True)
+
+    num_tokens = model.roberta.encoder.layer[0].attention.self.get_attn().shape[-1]
+    print(input_ids.size(-1), num_tokens)
+    R = torch.eye(num_tokens).expand(output.size(0), -1, -1).clone().to(device)
+
+    for i, blk in enumerate(model.roberta.encoder.layer):
+        if i < start_layer:
+            continue
+        grad = blk.attention.self.get_attn_gradients()
+        cam = blk.attention.self.get_attn()
+        cam = avg_heads(cam, grad)
+        joint = apply_self_attention_rules(R, cam)
+        R += joint
+    return output, R[:, 0, 1:-1]
+
+def visualize_text(datarecords, legend=True):
+    dom = ["<table width: 100%>"]
+    rows = [
+        "<tr><th>True Label</th>"
+        "<th>Predicted Label</th>"
+        "<th>Attribution Label</th>"
+        "<th>Attribution Score</th>"
+        "<th>Word Importance</th>"
+    ]
+    for datarecord in datarecords:
+        rows.append(
+            "".join(
+                [
+                    "<tr>",
+                    format_classname(datarecord.true_class),
+                    format_classname(
+                        "{0} ({1:.2f})".format(
+                            datarecord.pred_class, datarecord.pred_prob
+                        )
+                    ),
+                    format_classname(datarecord.attr_class),
+                    format_classname("{0:.2f}".format(datarecord.attr_score)),
+                    format_word_importances(
+                        datarecord.raw_input_ids, datarecord.word_attributions
+                    ),
+                    "<tr>",
+                ]
+            )
+        )
+
+    if legend:
+        dom.append(
+            '<div style="border-top: 1px solid; margin-top: 5px; \
+            padding-top: 5px; display: inline-block">'
+        )
+        dom.append("<b>Legend: </b>")
+
+        for value, label in zip([-1, 0, 1], ["Negative", "Neutral", "Positive"]):
+            dom.append(
+                '<span style="display: inline-block; width: 10px; height: 10px; \
+                border: 1px solid; background-color: \
+                {value}"></span> {label}  '.format(
+                    value=_get_color(value), label=label
+                )
+            )
+        dom.append("</div>")
+
+    dom.append("".join(rows))
+    dom.append("</table>")
+    html = "".join(dom)
+
+    return html
+
+def show_explanation(model, input_ids, attention_mask, index=None, start_layer=0):
+    # generate an explanation for the input
+    output, expl = generate_relevance(model, input_ids, attention_mask, index=index, start_layer=start_layer)
+    print(output.shape, expl.shape)
+    # normalize scores
+    scaler = PyTMinMaxScalerVectorized()
+
+    norm = scaler(expl)
+    # get the model classification
+    output = torch.nn.functional.softmax(output, dim=-1)
+
+
+    vis_data_records = []
+    for record in range(input_ids.size(0)):
+        classification = output[record].argmax(dim=-1).item()
+        class_name = classifications[classification]
+        nrm = norm[record]
+
+        # if the classification is negative, higher explanation scores are more negative
+        # flip for visualization
+        if class_name == "NEGATIVE":
+            nrm *= (-1)
+        tokens = tokenizer.convert_ids_to_tokens(input_ids[record].flatten())[1:0 - ((attention_mask[record] == 0).sum().item() + 1)]
+        print([(tokens[i], nrm[i].item()) for i in range(len(tokens))])
+        vis_data_records.append(visualization.VisualizationDataRecord(
+                                    nrm,
+                                    output[record][classification],
+                                    classification,
+                                    classification,
+                                    index,
+                                    1,
+                                    tokens,
+                                    1))
+    return visualize_text(vis_data_records)
+
+def run(input_text):
+    text_batch = [input_text]
+    encoding = tokenizer(text_batch, return_tensors='pt')
+    input_ids = encoding['input_ids'].to(device)
+    attention_mask = encoding['attention_mask'].to(device)
+
+    # true class is positive - 1
+    true_class = 1
+
+    html = show_explanation(model, input_ids, attention_mask)
+    return html
+
+iface = gradio.Interface(fn=greet, inputs="text", outputs="html", examples=[["This movie was the best movie I have ever seen! some scenes were ridiculous, but acting was great"], ["I really didn't like this movie. Some of the actors were good, but overall the movie was boring"]])
 iface.launch()

requirements.txt

@@ -0,0 +1,4 @@
+pytorch
+transformers==4.21.2
+captum
+