Upload PureBertForSequenceClassification

config.json

@@ -1,12 +1,13 @@
 {
-  "_name_or_path": "/root/.cache/torch/sentence_transformers/BAAI_bge-large-en/",
+  "_name_or_path": "BAAI/bge-large-en-v1.5",
   "alpha": 1,
   "architectures": [
-    "BertModel"
+    "PureBertForSequenceClassification"
   ],
   "attention_probs_dropout_prob": 0.1,
   "auto_map": {
-    "AutoConfig": "configuration_pure_bert.PureBertConfig"
+    "AutoConfig": "configuration_pure_bert.PureBertConfig",
+    "AutoModelForSequenceClassification": "modeling_pure_bert.PureBertForSequenceClassification"
   },
   "center": false,
   "classifier_dropout": null,

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:88d12b0f3b00bfde7a5ee7d3e054714e68ee6c358de80add011040ac7a009e26
+size 1336420068

modeling_pure_bert.py

@@ -0,0 +1,864 @@
+import torch
+import torch.nn as nn
+import numpy as np
+from torch.autograd import Function
+from transformers import PreTrainedModel
+from transformers.models.bert.modeling_bert import (
+    BertEmbeddings, BertEncoder, BertPooler
+)
+from typing import Union, Tuple, Optional, List
+from transformers.modeling_outputs import (
+    SequenceClassifierOutput, 
+    MultipleChoiceModelOutput, 
+    QuestionAnsweringModelOutput, 
+    BaseModelOutputWithPoolingAndCrossAttentions
+)
+from transformers.modeling_attn_mask_utils import (
+    _prepare_4d_attention_mask_for_sdpa,
+    _prepare_4d_causal_attention_mask_for_sdpa,
+)
+from transformers.utils import ModelOutput
+
+from .configuration_pure_bert import PureBertConfig
+
+
+class CovarianceFunction(Function):
+
+    @staticmethod
+    def forward(ctx, inputs):
+        x = inputs
+        b, c, h, w = x.data.shape
+        m = h * w
+        x = x.view(b, c, m)
+        I_hat = (-1.0 / m / m) * torch.ones(m, m, device=x.device) + (
+            1.0 / m
+        ) * torch.eye(m, m, device=x.device)
+        I_hat = I_hat.view(1, m, m).repeat(b, 1, 1).type(x.dtype)
+        y = x @ I_hat @ x.transpose(-1, -2)
+        ctx.save_for_backward(inputs, I_hat)
+        return y
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        inputs, I_hat = ctx.saved_tensors
+        x = inputs
+        b, c, h, w = x.data.shape
+        m = h * w
+        x = x.view(b, c, m)
+        grad_input = grad_output + grad_output.transpose(1, 2)
+        grad_input = grad_input @ x @ I_hat
+        grad_input = grad_input.reshape(b, c, h, w)
+        return grad_input
+
+
+class Covariance(nn.Module):
+
+    def __init__(self):
+        super(Covariance, self).__init__()
+        
+    def _covariance(self, x):
+        return CovarianceFunction.apply(x)
+
+    def forward(self, x):
+        # x should be [batch_size, seq_len, embed_dim]
+        if x.dim() == 2:
+            x = x.transpose(-1, -2)
+            C = self._covariance(x[None, :, :, None])
+        C = C.squeeze(dim=0)
+        return C
+
+
+class PFSA(torch.nn.Module):
+    """
+    https://openreview.net/pdf?id=isodM5jTA7h
+    """
+    def __init__(self, input_dim, alpha=1):
+        super(PFSA, self).__init__()
+        self.input_dim = input_dim
+        self.alpha = alpha
+
+    def forward_one_sample(self, x):
+        x = x.transpose(1, 2)[..., None]
+        k = torch.mean(x, dim=[-1, -2], keepdim=True)
+        kd = torch.sqrt((k - k.mean(dim=1, keepdim=True)).pow(2).sum(dim=1, keepdim=True)) # [B, 1, 1, 1]
+        qd = torch.sqrt((x - x.mean(dim=1, keepdim=True)).pow(2).sum(dim=1, keepdim=True)) # [B, 1, T, 1]
+        C_qk = (((x - x.mean(dim=1, keepdim=True)) * (k - k.mean(dim=1, keepdim=True))).sum(dim=1, keepdim=True)) / (qd * kd)
+        A = (1 - torch.sigmoid(C_qk)) ** self.alpha
+        out = x * A
+        out = out.squeeze(dim=-1).transpose(1, 2)
+        return out
+
+    def forward(self, input_values, attention_mask=None):
+        """
+        x: [B, T, F]
+        """
+        out = []
+        b, t, f = input_values.shape
+        for x, mask in zip(input_values, attention_mask):
+            x = x.view(1, t, f)
+            # x_in = x[:, :sum(mask), :]
+            x_in = x[:, :int(mask.sum().item()), :]
+            x_out = self.forward_one_sample(x_in)
+            x_expanded = torch.zeros_like(x, device=x.device)
+            x_expanded[:, :x_out.shape[-2], :x_out.shape[-1]] = x_out
+            out.append(x_expanded)
+        out = torch.vstack(out)
+        out = out.view(b, t, f)
+        return out
+
+
+class PURE(torch.nn.Module):
+
+    def __init__(
+        self, 
+        in_dim, 
+        svd_rank=16, 
+        num_pc_to_remove=1, 
+        center=False, 
+        num_iters=2, 
+        alpha=1, 
+        disable_pcr=False, 
+        disable_pfsa=False, 
+        disable_covariance=True, 
+        *args, **kwargs
+    ):
+        super().__init__()
+        self.in_dim = in_dim
+        self.svd_rank = svd_rank
+        self.num_pc_to_remove = num_pc_to_remove
+        self.center = center
+        self.num_iters = num_iters
+        self.do_pcr = not disable_pcr
+        self.do_pfsa = not disable_pfsa
+        self.do_covariance = not disable_covariance
+        self.attention = PFSA(in_dim, alpha=alpha)
+    
+    def _compute_pc(self, X, attention_mask):
+        """
+        x: (B, T, F)
+        """
+        pcs = []
+        bs, seqlen, dim = X.shape
+        for x, mask in zip(X, attention_mask):
+            rank = int(mask.sum().item())
+            x = x[:rank, :]
+            if self.do_covariance:
+                x = Covariance()(x)
+                q = self.svd_rank
+            else:
+                q = min(self.svd_rank, rank)
+            _, _, V = torch.pca_lowrank(x, q=q, center=self.center, niter=self.num_iters)
+            # _, _, Vh = torch.linalg.svd(x_, full_matrices=False)
+            # V = Vh.mH
+            pc = V.transpose(0, 1)[:self.num_pc_to_remove, :] # pc: [K, F]
+            pcs.append(pc)
+        # pcs = torch.vstack(pcs)
+        # pcs = pcs.view(bs, self.num_pc_to_remove, dim)
+        return pcs
+
+    def _remove_pc(self, X, pcs):
+        """
+        [B, T, F], [B, ..., F]
+        """
+        b, t, f = X.shape
+        out = []
+        for i, (x, pc) in enumerate(zip(X, pcs)):
+            # v = []
+            # for j, t in enumerate(x):
+            #     t_ = t
+            #     for c_ in c:
+            #         t_ = t_.view(f, 1) - c_.view(f, 1) @ c_.view(1, f) @ t.view(f, 1)
+            #     v.append(t_.transpose(-1, -2))
+            # v = torch.vstack(v)
+            v = x - x @ pc.transpose(0, 1) @ pc
+            out.append(v[None, ...])
+        out = torch.vstack(out)
+        return out
+
+    def forward(self, input_values, attention_mask=None, *args, **kwargs):
+        """
+        PCR -> Attention
+        x: (B, T, F)
+        """
+        x = input_values
+        if self.do_pcr:
+            pc = self._compute_pc(x, attention_mask) # pc: [B, K, F]
+            xx = self._remove_pc(x, pc)
+            # xx = xt - xt @ pc.transpose(1, 2) @ pc # [B, T, F] * [B, F, K] * [B, K, F] = [B, T, F]
+        else:
+            xx = x
+        if self.do_pfsa:
+            xx = self.attention(xx, attention_mask)
+        return xx
+
+
+class StatisticsPooling(torch.nn.Module):
+
+    def __init__(self, return_mean=True, return_std=True):
+        super().__init__()
+
+        # Small value for GaussNoise
+        self.eps = 1e-5
+        self.return_mean = return_mean
+        self.return_std = return_std
+        if not (self.return_mean or self.return_std):
+            raise ValueError(
+                "both of statistics are equal to False \n"
+                "consider enabling mean and/or std statistic pooling"
+            )
+
+    def forward(self, input_values, attention_mask=None):
+        """Calculates mean and std for a batch (input tensor).
+
+        Arguments
+        ---------
+        x : torch.Tensor
+            It represents a tensor for a mini-batch.
+        """
+        x = input_values
+        if attention_mask is None:
+            if self.return_mean:
+                mean = x.mean(dim=1)
+            if self.return_std:
+                std = x.std(dim=1)
+        else:
+            mean = []
+            std = []
+            for snt_id in range(x.shape[0]):
+                # Avoiding padded time steps
+                lengths = torch.sum(attention_mask, dim=1)
+                relative_lengths = lengths / torch.max(lengths)
+                actual_size = torch.round(relative_lengths[snt_id] * x.shape[1]).int()
+                # actual_size = int(torch.round(lengths[snt_id] * x.shape[1]))
+
+                # computing statistics
+                if self.return_mean:
+                    mean.append(
+                        torch.mean(x[snt_id, 0:actual_size, ...], dim=0)
+                    )
+                if self.return_std:
+                    std.append(torch.std(x[snt_id, 0:actual_size, ...], dim=0))
+            if self.return_mean:
+                mean = torch.stack(mean)
+            if self.return_std:
+                std = torch.stack(std)
+
+        if self.return_mean:
+            gnoise = self._get_gauss_noise(mean.size(), device=mean.device)
+            gnoise = gnoise
+            mean += gnoise
+        if self.return_std:
+            std = std + self.eps
+
+        # Append mean and std of the batch
+        if self.return_mean and self.return_std:
+            pooled_stats = torch.cat((mean, std), dim=1)
+            pooled_stats = pooled_stats.unsqueeze(1)
+        elif self.return_mean:
+            pooled_stats = mean.unsqueeze(1)
+        elif self.return_std:
+            pooled_stats = std.unsqueeze(1)
+
+        return pooled_stats
+
+    def _get_gauss_noise(self, shape_of_tensor, device="cpu"):
+        """Returns a tensor of epsilon Gaussian noise.
+
+        Arguments
+        ---------
+        shape_of_tensor : tensor
+            It represents the size of tensor for generating Gaussian noise.
+        """
+        gnoise = torch.randn(shape_of_tensor, device=device)
+        gnoise -= torch.min(gnoise)
+        gnoise /= torch.max(gnoise)
+        gnoise = self.eps * ((1 - 9) * gnoise + 9)
+
+        return gnoise
+
+
+class PureBertPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = PureBertConfig
+    base_model_prefix = "bert"
+    supports_gradient_checkpointing = True
+    _supports_sdpa = True
+
+    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)
+
+
+class PureBertModel(PureBertPreTrainedModel):
+    """
+
+    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](https://arxiv.org/abs/1706.03762) 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.
+    """
+
+    _no_split_modules = ["BertEmbeddings", "BertLayer"]
+
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BertEmbeddings(config)
+        self.encoder = BertEncoder(config)
+
+        self.pooler = BertPooler(config) if add_pooling_layer else None
+
+        self.attn_implementation = config._attn_implementation
+        self.position_embedding_type = config.position_embedding_type
+
+        # 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)
+
+    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)` or `(batch_size, sequence_length, target_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:
+            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
+            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 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)
+
+        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,
+        )
+
+        if attention_mask is None:
+            attention_mask = torch.ones((batch_size, seq_length + past_key_values_length), device=device)
+
+        use_sdpa_attention_masks = (
+            self.attn_implementation == "sdpa"
+            and self.position_embedding_type == "absolute"
+            and head_mask is None
+            and not output_attentions
+        )
+
+        # Expand the attention mask
+        if use_sdpa_attention_masks and attention_mask.dim() == 2:
+            # Expand the attention mask for SDPA.
+            # [bsz, seq_len] -> [bsz, 1, seq_len, seq_len]
+            if self.config.is_decoder:
+                extended_attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
+                    attention_mask,
+                    input_shape,
+                    embedding_output,
+                    past_key_values_length,
+                )
+            else:
+                extended_attention_mask = _prepare_4d_attention_mask_for_sdpa(
+                    attention_mask, embedding_output.dtype, tgt_len=seq_length
+                )
+        else:
+            # 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 = 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)
+
+            if use_sdpa_attention_masks and encoder_attention_mask.dim() == 2:
+                # Expand the attention mask for SDPA.
+                # [bsz, seq_len] -> [bsz, 1, seq_len, seq_len]
+                encoder_extended_attention_mask = _prepare_4d_attention_mask_for_sdpa(
+                    encoder_attention_mask, embedding_output.dtype, tgt_len=seq_length
+                )
+            else:
+                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)
+
+        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,
+        )
+
+
+class PureBertForSequenceClassification(PureBertPreTrainedModel):
+
+    def __init__(
+        self, 
+        config, 
+        label_smoothing=0.0, 
+    ):
+        super().__init__(config)
+        self.label_smoothing = label_smoothing
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.bert = PureBertModel(config, add_pooling_layer=False)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.pure = PURE(
+            in_dim=config.hidden_size, 
+            svd_rank=config.svd_rank, 
+            num_pc_to_remove=config.num_pc_to_remove, 
+            center=config.center, 
+            num_iters=config.num_iters, 
+            alpha=config.alpha, 
+            disable_pcr=config.disable_pcr, 
+            disable_pfsa=config.disable_pfsa, 
+            disable_covariance=config.disable_covariance
+        )
+        self.mean = StatisticsPooling(return_mean=True, return_std=False)
+        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()
+
+    def forward_pure_embeddings(
+        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,
+        labels: Optional[torch.Tensor] = 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.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,
+        )
+
+        token_embeddings = outputs.last_hidden_state
+        token_embeddings = self.pure(token_embeddings, attention_mask)
+
+        return ModelOutput(
+            last_hidden_state=token_embeddings,
+        )
+
+    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,
+        labels: Optional[torch.Tensor] = 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.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,
+        )
+
+        token_embeddings = outputs.last_hidden_state
+        token_embeddings = self.pure(token_embeddings, attention_mask)
+        pooled_output = self.mean(token_embeddings).squeeze(1)
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_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 = nn.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 = nn.CrossEntropyLoss(label_smoothing=self.label_smoothing)
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = nn.BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class PureBertForMultipleChoice(PureBertPreTrainedModel):
+
+    def __init__(
+        self, 
+        config, 
+        label_smoothing=0.0, 
+    ):
+        super().__init__(config)
+        self.label_smoothing = label_smoothing
+
+        self.bert = PureBertModel(config)
+        self.pure = PURE(
+            in_dim=config.hidden_size, 
+            svd_rank=config.svd_rank, 
+            num_pc_to_remove=config.num_pc_to_remove, 
+            center=config.center, 
+            num_iters=config.num_iters, 
+            alpha=config.alpha, 
+            disable_pcr=config.disable_pcr, 
+            disable_pfsa=config.disable_pfsa, 
+            disable_covariance=config.disable_covariance
+        )
+        self.mean = StatisticsPooling(return_mean=True, return_std=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, 1)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    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,
+        labels: Optional[torch.Tensor] = 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]
+
+        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+
+        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,
+        )
+
+        token_embeddings = outputs.last_hidden_state
+        token_embeddings = self.pure(token_embeddings, attention_mask)
+        pooled_output = self.mean(token_embeddings).squeeze(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 = nn.CrossEntropyLoss(label_smoothing=self.label_smoothing)
+            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,
+        )
+
+
+class PureBertForQuestionAnswering(PureBertPreTrainedModel):
+
+    def __init__(
+        self, 
+        config, 
+        label_smoothing=0.0, 
+    ):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.label_smoothing = label_smoothing
+
+        self.bert = PureBertModel(config, add_pooling_layer=False)
+        self.pure = PURE(
+            in_dim=config.hidden_size, 
+            svd_rank=config.svd_rank, 
+            num_pc_to_remove=config.num_pc_to_remove, 
+            center=config.center, 
+            num_iters=config.num_iters, 
+            alpha=config.alpha, 
+            disable_pcr=config.disable_pcr, 
+            disable_pfsa=config.disable_pfsa, 
+            disable_covariance=config.disable_covariance
+        )
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    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,
+        start_positions: Optional[torch.Tensor] = None,
+        end_positions: Optional[torch.Tensor] = 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.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,
+        )
+
+        token_embeddings = outputs.last_hidden_state
+        sequence_output = self.pure(token_embeddings, attention_mask)
+
+        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 = nn.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,
+        )