Upload 21 files

app.py

@@ -0,0 +1,91 @@
+import argparse
+import logging
+import math
+import os
+import random
+
+import datasets
+from datasets import load_dataset, load_metric
+from torch.utils.data import DataLoader
+from tqdm.auto import tqdm
+import gradio as gr
+
+import transformers
+from accelerate import Accelerator # huggingface package
+from transformers import (
+    AdamW,
+    AutoConfig,
+    AutoModelForSequenceClassification,
+    AutoTokenizer,
+    DataCollatorWithPadding,
+    PretrainedConfig,
+    SchedulerType,
+    default_data_collator,
+    get_scheduler,
+    set_seed,
+    BertTokenizer,
+)
+from transformers.utils.versions import require_version
+
+import torch
+from test_module.modeling_transkimer import BertForSequenceClassification as TranskimerForSequenceClassification
+from test_module.modeling_transkimer_roberta import RobertaForSequenceClassification as TranskimerRobertaForSequenceClassification
+from test_module.modeling_utils import convert_softmax_mask_to_digit
+
+from blackbox_utils.my_attack import CharacterAttack
+from transformers import glue_processors as processors
+
+
+task_to_keys = {
+    "cola": ("sentence", None),
+    "mnli": ("premise", "hypothesis"),
+    "mrpc": ("sentence1", "sentence2"),
+    "qnli": ("question", "sentence"),
+    "qqp": ("question1", "question2"),
+    "rte": ("sentence1", "sentence2"),
+    "sst2": ("sentence", None),
+    "stsb": ("sentence1", "sentence2"),
+    "wnli": ("sentence1", "sentence2"),
+    "imdb": ("text", None),
+}
+
+model_path_dict = {
+    "transkimer_sst2_not_pad":'./not_pad_0.5',
+}
+
+
+datasets.utils.logging.set_verbosity_error()
+transformers.utils.logging.set_verbosity_error()
+
+
+task_name = 'sst2'
+model_type = 'transkimer'
+
+# Load pretrained model and tokenizer
+model_path_key =  f'{model_type}_{task_name}_not_pad'
+model_path = model_path_dict[model_path_key]
+config = AutoConfig.from_pretrained(model_path, num_labels=num_labels, finetuning_task=task_name)
+tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased', use_fast=True)
+model = TranskimerForSequenceClassification.from_pretrained(model_path,from_tf=bool(".ckpt" in model_path),config=config,)
+
+# Preprocessing the datasets
+sentence1_key, sentence2_key = task_to_keys[task_name]
+    
+processor = processors[task_name]()
+label_list = processor.get_labels()
+    
+label_to_id = {v: i for i, v in enumerate(label_list)}
+
+padding = False
+
+attack = CharacterAttack(f'{model_type}_{task_name}',model,tokenizer,device='cpu',max_per=10,padding=padding,max_length=128,label_to_id=label_to_id,sentence1_key=sentence1_key,sentence2_key=sentence2_key)
+
+
+def greet(text):
+    text_input = [(text,None)]
+    outputs,time = attack.get_prob(text_input)
+    _,token_remained,_ = attack.output_analysis(outputs)
+    return time,token_remained.item()
+
+iface = gr.Interface(fn=greet, inputs=["text","text"], outputs=["number","number"])
+iface.launch()

blackbox_utils/Attack_base.py

@@ -0,0 +1,238 @@
+import torch
+import torch.nn as nn
+# import jieba
+import string
+import numpy as np
+from copy import deepcopy
+from tqdm import tqdm
+import time
+from datetime import datetime
+import os
+from sklearn.linear_model import LinearRegression
+from torch.multiprocessing import Process,Pool
+
+from transformers import BertTokenizer
+
+os.environ['TOKENIZERS_PARALLELISM']='True'
+# torch.autograd.set_detect_anomaly(True)
+
+class BaseAttack:
+    def __init__(self, name, model, tokenizer, device, max_per, padding,max_length,label_to_id,sentence1_key,sentence2_key):
+        self.name = name
+        self.model = model
+        self.tokenizer = tokenizer
+        self.device = device
+        self.model = self.model.to(self.device)
+        self.model.eval()
+        self.padding = padding
+        self.max_length = max_length
+        self.label_to_id = label_to_id
+        self.sentence1_key = sentence1_key
+        self.sentence2_key = sentence2_key
+        # 修改token个数的最大值
+        self.max_per = max_per
+        # linear regression model initialization
+        self.linear_regression()
+        self.random_tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+
+    def run_attack(self, x):
+        pass
+
+    def compute_loss(self, x):
+        pass
+
+    def preprocess_function(self,examples,to_device=True):
+        # Tokenize the texts
+        texts = ((examples[0],) if self.sentence2_key is None else (examples[0], examples[1]))
+        result = self.tokenizer(*texts, padding=self.padding, max_length=self.max_length, truncation=True)
+        new_result = {}
+        for key,item in result.items():
+            if to_device:
+                new_result[key] = torch.tensor(item).unsqueeze(0).to(self.device)
+            else:
+                new_result[key] = torch.tensor(item).unsqueeze(0)
+        return new_result
+    
+    def get_pred(self,input_):
+        return self.get_prob(input_).logits.argmax(dim=-1)
+
+    def get_prob(self,input_):
+        toc = datetime.now()
+        batch = self.preprocess_function(input_)
+        # batch['gumbel_softmax']=gradient
+        # print(batch)
+        outputs = self.model(**batch) # get all logits
+        tic = datetime.now()
+        running_time = (tic-toc).total_seconds()
+        return outputs,running_time
+    
+    def output_analysis(self,outputs):
+        # print(outputs)
+
+        all_skim_loss, all_tokens_remained = list(), list()
+        all_layer_tokens_remained = [[] for _ in range(len(outputs.layer_tokens_remained))]
+        
+        all_skim_loss.append(outputs.skim_loss)
+        all_tokens_remained.append(outputs.tokens_remained)
+        for layer_idx,mac in enumerate(outputs.layer_tokens_remained):
+            all_layer_tokens_remained[layer_idx].append(mac)
+
+        skim_loss = torch.mean(torch.stack(all_skim_loss))
+        tokens_remained = torch.mean(torch.stack(all_tokens_remained))
+        layers_result = [torch.mean(torch.stack(macs)) for i,macs in enumerate(all_layer_tokens_remained)]
+
+        return skim_loss,tokens_remained,layers_result
+    
+    def load_data(self,model_path_key,mode='train'):
+        path = f'flops_count/{model_path_key}/{mode}'
+        if os.path.exists(f'{path}/process_data.pth'):
+            print(f'loading data from {path}')
+            data = torch.load(f'{path}/process_data.pth')
+        else:
+            time_list = torch.load(f'{path}/time_list.pth')
+            ratio_list = torch.load(f'{path}/ratio_list.pth')
+            token_num_list = torch.load(f'{path}/text_len_list_tokenizer.pth')
+
+            ratio_list_ = []
+            for ratio in ratio_list:
+                ratio_list_.append(ratio.item())
+            y = np.expand_dims(np.array(ratio_list_),axis=1)   
+            # print(x.shape)
+
+            time_list_ = []
+            for time,token_num in zip(time_list,token_num_list):
+                time_list_.append((time/(token_num*(10**8))))
+            x = np.expand_dims(np.array(time_list_),axis=1)
+            # print(y.shape)
+            
+            data = dict()
+            data['x']=x
+            data['y']=y
+            torch.save(data,f'{path}/process_data.pth')
+
+        return data
+    
+    def predict(self,x):
+        return self.w*x+self.b
+    
+    def linear_regression(self):
+        print("="*20)
+        print('Linear Regression Generation')
+        data_train = self.load_data(self.name,mode='train')
+        data_test = self.load_data(self.name,mode='test')
+        # print(data_train,data_test)
+
+        reg = LinearRegression().fit(data_train['x'],data_train['y'])
+        train_score = reg.score(data_train['x'],data_train['y'])
+        test_score = reg.score(data_test['x'],data_test['y'])
+        print(f'train set score: {train_score}')
+        print(f'test set score: {test_score}')
+
+        self.w = reg.coef_[0][0]
+        self.b = reg.intercept_[0]
+        print("w:",self.w)
+        print("b:",self.b)
+
+        print(self.predict(0.8))
+
+
+class MyAttack(BaseAttack):
+    def __init__(self, name, model, tokenizer, device, max_per, padding, max_length, label_to_id, sentence1_key, sentence2_key):
+        super(MyAttack, self).__init__(name, model, tokenizer, device, max_per, padding, max_length, label_to_id, sentence1_key, sentence2_key)
+        # self.insert_character = string.punctuation
+        self.insert_character = string.digits
+        self.insert_character += string.ascii_letters
+        # self.insert_character -= """"'/\\"""
+        # print(self.insert_character)
+
+        self.origin_ratio = []
+        self.attack_ratio = []
+        self.layer_result = []
+        self.origin_layer_result = []
+
+    # @torch.no_grad()
+    # def select_best(self, new_strings):
+    #     best_string = None
+    #     best_loss = 0
+    #     for new_string in new_strings:
+    #         new_predicted_loss = self.compute_loss(new_string)
+    #         if new_predicted_loss>best_loss:
+    #             best_loss = new_predicted_loss
+    #             best_string = new_string
+
+    #     assert best_string is not None
+    #     return best_string,best_loss
+
+    @torch.no_grad()
+    def select_best(self, new_strings):
+        # self.model.to('cpu')
+        best_string = None
+        best_loss = 0
+        with Pool(processes=4) as pool:
+            loss_list = pool.map(self.compute_loss,new_strings)
+        idx = np.argmax(np.array(loss_list))
+        best_loss = loss_list[idx]
+        best_string = new_strings[idx]
+        # self.model.to(self.device)
+        # for new_string in new_strings:
+        #     new_predicted_loss = self.compute_loss(new_string)
+        #     if new_predicted_loss>best_loss:
+        #         best_loss = new_predicted_loss
+        #         best_string = new_string
+
+        assert best_string is not None
+        # self.model.to(self.device)
+        return best_string,best_loss
+
+    def compute_loss(self, xxx):
+        raise NotImplementedError
+
+    def mutation(self, current_adv_text, grad, modify_pos):
+        raise NotImplementedError
+
+    def run_attack(self, text):
+        # assert len(text) == 1
+        # print(text)
+        text[0] = text[0].strip(" .")
+        text[1] = text[1].strip(" .")
+        print(f'Origin Text: {text}')
+        current_adv_text = deepcopy(text)
+        # max_per 最多扰动单词的个数
+        # pbar = tqdm(range(self.max_per))
+
+        best_loss = 0
+        best_tokens_remained = 0
+        best_layer_result = None
+
+        output,_ = self.get_prob(current_adv_text)
+        origin_skim_loss,origin_ratio_,origin_layer_result_ = self.output_analysis(output)
+        print(origin_skim_loss,origin_ratio_)
+        self.origin_ratio.append(origin_ratio_.item())
+        self.origin_layer_result.append(origin_layer_result_)
+
+
+        # for it in pbar:
+        for _ in range(self.max_per):
+            # 得到每个修改的位置
+            new_strings = self.mutation(current_adv_text)
+            #print(new_strings)
+            current_adv_text,current_loss = self.select_best(new_strings)
+            # print(new_strings)
+            # print(current_adv_text,current_loss,current_tokens_remained)
+            if current_loss > best_loss:
+                best_adv_text = deepcopy(current_adv_text)
+                best_loss = current_loss
+            print(best_adv_text)
+
+        output,_ = self.get_prob(best_adv_text)
+        _,best_tokens_remained,best_layer_result = self.output_analysis(output)
+
+        self.attack_ratio.append(best_tokens_remained.item())
+        self.layer_result.append(best_layer_result)
+        print(f'Malicious Text: {best_adv_text}')
+        print(f'Origin Ratio: {self.origin_ratio[-1]} Attack Ratio: {self.attack_ratio[-1]}')
+        print(f'Layer Result: {self.layer_result[-1]}')
+
+        return best_adv_text,best_loss,best_tokens_remained,best_layer_result
+
+

blackbox_utils/my_attack.py

@@ -0,0 +1,99 @@
+import copy
+import time
+import torch
+import torch.nn as nn
+import numpy as np
+from tqdm import tqdm
+import nltk
+import string
+from copy import deepcopy
+from torchprofile import profile_macs
+from datetime import datetime
+
+from transformers import BertTokenizer, BertModel, BertForMaskedLM
+from nltk.tokenize.treebank import TreebankWordTokenizer, TreebankWordDetokenizer
+from blackbox_utils.Attack_base import MyAttack
+
+
+class CharacterAttack(MyAttack):
+    # TODO: 存储一个list每次只修改不同的token位置
+    def __init__(self, name, model, tokenizer, device, max_per, padding, max_length, label_to_id, sentence1_key, sentence2_key):
+        super(CharacterAttack, self).__init__(name, model, tokenizer, device, max_per, padding, max_length, label_to_id, sentence1_key, sentence2_key)
+
+    def compute_importance(self, text):
+        current_tensor = self.preprocess_function(text)["input_ids"][0]
+        # print(current_tensor)
+        word_losses = {}
+        for idx in range(1,len(current_tensor)-1):
+            # print(current_tensor[:idx])
+            # print(current_tensor[idx+1:])
+            sentence_tokens_without = torch.cat([current_tensor[:idx],current_tensor[idx + 1:]])
+            sentence_without = self.tokenizer.decode(sentence_tokens_without)
+            sentence_without = [sentence_without,text[1]]
+            word_losses[int(current_tensor[idx])] = self.compute_loss(sentence_without)
+        word_losses = [k for k, _ in sorted(word_losses.items(), key=lambda item: item[1], reverse=True)] 
+        return word_losses 
+
+    def compute_loss(self, text):
+        inputs = self.preprocess_function(text)
+        shift_inputs = (inputs['input_ids'],inputs['attention_mask'],inputs['token_type_ids'])
+        # toc = datetime.now()
+        macs = profile_macs(self.model, shift_inputs)
+        # tic = datetime.now()
+        # print((tic-toc).total_seconds())
+        result = self.random_tokenizer(*inputs, padding=self.padding, max_length=self.max_length, truncation=True)
+        token_length = len(result["input_ids"])
+        macs_per_token = macs/(token_length*10**8)
+
+        return self.predict(macs_per_token)
+
+    def mutation(self, current_adv_text):
+        current_tensor = self.preprocess_function(current_adv_text)
+        # print(current_tensor)
+        current_tensor = current_tensor["input_ids"][0]
+        # print(current_tensor)
+        new_strings = self.character_replace_mutation(current_adv_text, current_tensor)
+        return new_strings
+
+    @staticmethod
+    def transfer(c: str):
+        if c in string.ascii_lowercase:
+            return c.upper()
+        elif c in string.ascii_uppercase:
+            return c.lower()
+        return c
+
+    def character_replace_mutation(self, current_text, current_tensor):
+        important_tensor = self.compute_importance(current_text)
+        # current_string = [self.tokenizer.decoder[int(t)] for t in current_tensor]
+        new_strings = [current_text]
+        # 遍历每个vocabulary，查找文本有的第一个token
+        # print(current_tensor)
+        for t in important_tensor:
+            if int(t) not in current_tensor:
+                continue
+            ori_decode_token = self.tokenizer.decode([int(t)])
+            # print(ori_decode_token)
+            # if self.space_token in ori_decode_token:
+            #     ori_token = ori_decode_token.replace(self.space_token, '')
+            # else:
+            ori_token = ori_decode_token
+            # 如果只有一个长度
+            if len(ori_token) == 1 or ori_token not in current_text[0]:  #todo
+                continue
+            # 随意插入一个字符
+            candidate = [ori_token[:i] + insert + ori_token[i:] for i in range(len(ori_token)) for insert in self.insert_character]
+            # 随意更换一个大小写
+            candidate += [ori_token[:i - 1] + self.transfer(ori_token[i - 1]) + ori_token[i:] for i in range(1, len(ori_token))]
+            # print(candidate)
+            # 最多只替换一次
+            new_strings += [[current_text[0].replace(ori_token, c, 1),current_text[1]] for c in candidate]
+
+            # ori_tensor_pos = current_tensor.eq(int(t)).nonzero()
+            #
+            # for p in ori_tensor_pos:
+            #     new_strings += [current_string[:p] + c + current_string[p + 1:] for c in candidate]
+            # 存在一个有效的改动就返回
+            if len(new_strings) > 1:
+                return new_strings
+        return new_strings

not_pad_0.5/config.json

@@ -0,0 +1,37 @@
+{
+  "_name_or_path": "bert-base-uncased",
+  "architectures": [
+    "BertForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "classifier_dropout": null,
+  "finetuning_task": "sst2",
+  "gradient_checkpointing": false,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": "negative",
+    "1": "positive"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "label2id": {
+    "negative": 0,
+    "positive": 1
+  },
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "position_embedding_type": "absolute",
+  "problem_type": "single_label_classification",
+  "skim_coefficient": 0.5,
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.1",
+  "type_vocab_size": 2,
+  "use_cache": true,
+  "vocab_size": 30522
+}

not_pad_0.5/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b49ee701f19e2ad22efc89e07452b0aed5c3075800004d2a10b63308a29f4363
+size 466610485

requirements.txt

@@ -0,0 +1,18 @@
+argparse
+logging
+math
+os
+random
+tqdm
+gradio
+transformers
+accelerate
+torch
+datetime
+time
+copy
+numpy
+string
+sklearn
+nltk
+torchprofile

test_module/modeling_skim_predictor.py

@@ -0,0 +1,51 @@
+from torch import nn
+import torch
+
+def init_skim_predictor(module_list, mean_bias=5.0):
+    for module in module_list:
+        if not isinstance(module, torch.nn.Linear):
+            raise ValueError("only support initialization of linear skim predictor")
+
+        # module.bias.data[1].fill_(5.0)
+        # module.bias.data[0].fill_(-5.0)
+        # module.weight.data.zero_()
+        module.bias.data[1].normal_(mean=mean_bias, std=0.02)
+        module.bias.data[0].normal_(mean=-mean_bias, std=0.02)
+        module.weight.data.normal_(mean=0.0, std=0.02)
+
+        module._skim_initialized = True
+
+class SkimPredictor(nn.Module):
+    def __init__(self, input_size, output_size, hidden_size=None):
+        super().__init__()
+        
+        self.hidden_size = hidden_size if hidden_size else input_size
+
+        self.predictor = nn.Sequential(
+            nn.LayerNorm(input_size),
+            nn.Linear(input_size, self.hidden_size),
+            # nn.GELU(),
+            # nn.Linear(self.hidden_size, self.hidden_size),
+            nn.LayerNorm(self.hidden_size),
+            nn.GELU(),
+            nn.Linear(self.hidden_size, output_size),
+        )
+
+        init_skim_predictor([self.predictor[-1]])
+
+    def forward(self, hidden_states):
+        return self.predictor(hidden_states)
+
+def test_init_skim_predictor():
+    num_layers = 12
+
+    skim_predictors = torch.nn.ModuleList([torch.nn.Linear(768,2) for _ in range(num_layers)])
+    init_skim_predictor(skim_predictors)
+
+    print(skim_predictors[0].weight, skim_predictors[0].bias)
+
+    rand_input = torch.rand((4, 16, 768))
+    print(skim_predictors[0](rand_input))
+
+if __name__ == "__main__":
+    test_init_skim_predictor()

test_module/modeling_transkimer.py

@@ -0,0 +1,2002 @@
+# 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 BERT model. """
+
+
+import math
+import os
+import warnings
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+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
+from transformers.file_utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    NextSentencePredictorOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.utils import logging
+from transformers.models.bert.configuration_bert import BertConfig
+
+from .modeling_skim_predictor import SkimPredictor
+from .modeling_utils import BaseModelOutputWithPastAndCrossAttentionsSkim, BaseModelOutputWithPoolingAndCrossAttentionsSkim, MaskedLMOutputSkim, QuestionAnsweringModelOutputSkim, SequenceClassifierOutputSkim, convert_softmax_mask_to_digit, trunc_with_mask_batched, masked_softmax
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "bert-base-uncased"
+_CONFIG_FOR_DOC = "BertConfig"
+_TOKENIZER_FOR_DOC = "BertTokenizer"
+
+BERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "bert-base-uncased",
+    "bert-large-uncased",
+    "bert-base-cased",
+    "bert-large-cased",
+    "bert-base-multilingual-uncased",
+    "bert-base-multilingual-cased",
+    "bert-base-chinese",
+    "bert-base-german-cased",
+    "bert-large-uncased-whole-word-masking",
+    "bert-large-cased-whole-word-masking",
+    "bert-large-uncased-whole-word-masking-finetuned-squad",
+    "bert-large-cased-whole-word-masking-finetuned-squad",
+    "bert-base-cased-finetuned-mrpc",
+    "bert-base-german-dbmdz-cased",
+    "bert-base-german-dbmdz-uncased",
+    "cl-tohoku/bert-base-japanese",
+    "cl-tohoku/bert-base-japanese-whole-word-masking",
+    "cl-tohoku/bert-base-japanese-char",
+    "cl-tohoku/bert-base-japanese-char-whole-word-masking",
+    "TurkuNLP/bert-base-finnish-cased-v1",
+    "TurkuNLP/bert-base-finnish-uncased-v1",
+    "wietsedv/bert-base-dutch-cased",
+    # See all BERT models at https://huggingface.co/models?filter=bert
+]
+
+
+def load_tf_weights_in_bert(model, config, tf_checkpoint_path):
+    """Load tf checkpoints in a pytorch model."""
+    try:
+        import re
+
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info(f"Loading TF weight {name} with shape {shape}")
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        name = name.split("/")
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(
+            n in ["adam_v", "adam_m", "AdamWeightDecayOptimizer", "AdamWeightDecayOptimizer_1", "global_step"]
+            for n in name
+        ):
+            logger.info(f"Skipping {'/'.join(name)}")
+            continue
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
+                scope_names = re.split(r"_(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+            if scope_names[0] == "kernel" or scope_names[0] == "gamma":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
+                pointer = getattr(pointer, "bias")
+            elif scope_names[0] == "output_weights":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "squad":
+                pointer = getattr(pointer, "classifier")
+            else:
+                try:
+                    pointer = getattr(pointer, scope_names[0])
+                except AttributeError:
+                    logger.info(f"Skipping {'/'.join(name)}")
+                    continue
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+        if m_name[-11:] == "_embeddings":
+            pointer = getattr(pointer, "weight")
+        elif m_name == "kernel":
+            array = np.transpose(array)
+        try:
+            assert (
+                pointer.shape == array.shape
+            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info(f"Initialize PyTorch weight {name}")
+        pointer.data = torch.from_numpy(array)
+    return model
+
+
+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 = 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, device=self.position_ids.device),
+                persistent=False,
+            )
+
+    def forward(
+        self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
+    ):
+        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[:, past_key_values_length : seq_length + past_key_values_length]
+
+        # 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
+
+
+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(
+                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 = getattr(config, "position_embedding_type", "absolute")
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        skim_mask=None,
+    ):
+        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 BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+        # attention_probs = masked_softmax(attention_scores, skim_mask, dim=3)
+
+        # mask attention probs during training for skimming
+        attention_probs = attention_probs * skim_mask[:, None, None, :]
+
+        # 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
+
+
+class BertSelfOutput(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, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self = BertSelfAttention(config)
+        self.output = BertSelfOutput(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,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        skim_mask=None,
+    ):
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+            skim_mask,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+class BertIntermediate(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):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class BertOutput(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, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertLayer(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 = BertAttention(config)
+        self.is_decoder = config.is_decoder
+        self.add_cross_attention = config.add_cross_attention
+        if self.add_cross_attention:
+            assert self.is_decoder, f"{self} should be used as a decoder model if cross attention is added"
+            self.crossattention = BertAttention(config)
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        skim_mask=None,
+    ):
+        # 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,
+            skim_mask=skim_mask,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            assert hasattr(
+                self, "crossattention"
+            ), 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
+
+
+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)])
+
+        # skim predictors for each layer
+        self.skim_predictors = nn.ModuleList([SkimPredictor(config.hidden_size, 2) for _ in range(config.num_hidden_layers)])
+        # init_skim_predictor(self.skim_predictors)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=False,
+        output_hidden_states=False,
+        return_dict=True,
+    ):
+        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
+        all_skim_mask = ()
+
+        next_decoder_cache = () if use_cache else None
+
+        forward_hidden_states = hidden_states.clone()
+        forward_skim_mask = None
+
+
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+            
+            # if gumbel_softmax:
+            #     # print('gradient')
+            #     skim_mask = nn.functional.gumbel_softmax(self.skim_predictors[i](hidden_states[:,1:,:]), hard=True, tau=1)
+            # else:
+                # print('not gradient')
+            logits = nn.functional.softmax(self.skim_predictors[i](hidden_states[:,1:,:]),dim=-1) 
+                # print(logits)
+            index = logits.max(dim=-1, keepdim=True)[1]
+                # print(index)
+            skim_mask = torch.zeros_like(logits, memory_format=torch.legacy_contiguous_format).scatter_(-1, index, 1.0)
+                # print(skim_mask)
+                
+            skim_mask = skim_mask[:,:,1]
+            skim_mask_with_cls = torch.ones(skim_mask.shape[0], skim_mask.shape[1]+1, device=skim_mask.device)
+            skim_mask_with_cls[:,1:] = skim_mask
+            skim_mask = skim_mask_with_cls
+            # multiple current layer skim mask with last layer skim mask
+            # to gurantee skimmed tokens are never recovered
+            if all_skim_mask and hidden_states.shape[0] != 1:
+                skim_mask = skim_mask * all_skim_mask[-1]
+            all_skim_mask += (skim_mask, )
+
+            # 最大的一个不同之处：就是这个trunc掉了
+            if hidden_states.shape[0] == 1:
+                bool_skim_mask = skim_mask.to(dtype=torch.bool)
+                hidden_states = trunc_with_mask_batched(hidden_states, bool_skim_mask, 1)
+                attention_mask = trunc_with_mask_batched(attention_mask, bool_skim_mask, 3)
+                skim_mask = trunc_with_mask_batched(skim_mask, bool_skim_mask, 1)
+                if forward_skim_mask is None:
+                    forward_skim_mask = torch.ones_like(bool_skim_mask).to(dtype=torch.bool)
+
+            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 getattr(self.config, "gradient_checkpointing", False) and self.training:
+
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. 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,
+                    skim_mask,
+                )
+
+            hidden_states = layer_outputs[0]
+            # print(hidden_states.shape)
+            if hidden_states.shape[0] == 1:
+                forward_skim_mask[forward_skim_mask.clone()] = bool_skim_mask
+                forward_hidden_states[forward_skim_mask] = hidden_states
+            else:
+                forward_hidden_states = forward_hidden_states * (1-skim_mask.view(*skim_mask.shape,1)) + hidden_states * skim_mask.view(*skim_mask.shape,1)
+
+            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 [
+                    forward_hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentionsSkim(
+            last_hidden_state=forward_hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+            attention_mask=attention_mask,
+            skim_mask=all_skim_mask,
+        )
+
+
+class BertPooler(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):
+        # 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 BertPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class BertLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = BertPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+class BertOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = BertLMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+class BertOnlyNSPHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.seq_relationship = nn.Linear(config.hidden_size, 2)
+
+    def forward(self, pooled_output):
+        seq_relationship_score = self.seq_relationship(pooled_output)
+        return seq_relationship_score
+
+
+class BertPreTrainingHeads(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = BertLMPredictionHead(config)
+        self.seq_relationship = nn.Linear(config.hidden_size, 2)
+
+    def forward(self, sequence_output, pooled_output):
+        prediction_scores = self.predictions(sequence_output)
+        seq_relationship_score = self.seq_relationship(pooled_output)
+        return prediction_scores, seq_relationship_score
+
+
+class BertPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = BertConfig
+    load_tf_weights = load_tf_weights_in_bert
+    base_model_prefix = "bert"
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if hasattr(module, '_skim_initialized') and module._skim_initialized:
+            return
+        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)
+
+
+@dataclass
+class BertForPreTrainingOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.BertForPreTraining`.
+
+    Args:
+        loss (`optional`, returned when ``labels`` is provided, ``torch.FloatTensor`` of shape :obj:`(1,)`):
+            Total loss as the sum of the masked language modeling loss and the next sequence prediction
+            (classification) loss.
+        prediction_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        seq_relationship_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
+            before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    prediction_logits: torch.FloatTensor = None
+    seq_relationship_logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+BERT_START_DOCSTRING = r"""
+
+    This model inherits from :class:`~transformers.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 (:class:`~transformers.BertConfig`): 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 :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
+            weights.
+"""
+
+BERT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.BertTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`({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.html#attention-mask>`__
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`({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.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`({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.html#position-ids>`_
+        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(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 (:obj:`torch.FloatTensor` of shape :obj:`({0}, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
+        output_attentions (:obj:`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 (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare Bert Model transformer outputting raw hidden-states without any specific head on top.",
+    BERT_START_DOCSTRING,
+)
+class BertModel(BertPreTrainedModel):
+    """
+
+    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 :obj:`is_decoder` argument of the configuration
+    set to :obj:`True`. To be used in a Seq2Seq model, the model needs to initialized with both :obj:`is_decoder`
+    argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an
+    input to the forward pass.
+    """
+
+    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.init_weights()
+
+    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(BERT_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,
+    )
+    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,
+        past_key_values=None,
+        use_cache=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 tokens that are **masked**.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(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 :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        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 BaseModelOutputWithPoolingAndCrossAttentionsSkim(
+            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,
+            attention_mask=encoder_outputs.attention_mask,
+            skim_mask=encoder_outputs.skim_mask,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bert Model with two heads on top as done during the pretraining: a `masked language modeling` head and a `next
+    sentence prediction (classification)` head.
+    """,
+    BERT_START_DOCSTRING,
+)
+class BertForPreTraining(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config)
+        self.cls = BertPreTrainingHeads(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=BertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        next_sentence_label=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`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]``
+        next_sentence_label (``torch.LongTensor`` of shape ``(batch_size,)``, `optional`):
+            Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair
+            (see :obj:`input_ids` docstring) Indices should be in ``[0, 1]``:
+
+            - 0 indicates sequence B is a continuation of sequence A,
+            - 1 indicates sequence B is a random sequence.
+        kwargs (:obj:`Dict[str, any]`, optional, defaults to `{}`):
+            Used to hide legacy arguments that have been deprecated.
+
+        Returns:
+
+        Example::
+
+            >>> from transformers import BertTokenizer, BertForPreTraining
+            >>> import torch
+
+            >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+            >>> model = BertForPreTraining.from_pretrained('bert-base-uncased')
+
+            >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+            >>> outputs = model(**inputs)
+
+            >>> prediction_logits = outputs.prediction_logits
+            >>> seq_relationship_logits = outputs.seq_relationship_logits
+        """
+        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,
+        )
+
+        sequence_output, pooled_output = outputs[:2]
+        prediction_scores, seq_relationship_score = self.cls(sequence_output, pooled_output)
+
+        total_loss = None
+        if labels is not None and next_sentence_label is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
+            total_loss = masked_lm_loss + next_sentence_loss
+
+        if not return_dict:
+            output = (prediction_scores, seq_relationship_score) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return BertForPreTrainingOutput(
+            loss=total_loss,
+            prediction_logits=prediction_scores,
+            seq_relationship_logits=seq_relationship_score,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """Bert Model with a `language modeling` head on top for CLM fine-tuning. """, BERT_START_DOCSTRING
+)
+class BertLMHeadModel(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        if not config.is_decoder:
+            logger.warning("If you want to use `BertLMHeadModel` as a standalone, add `is_decoder=True.`")
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
+    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,
+        labels=None,
+        past_key_values=None,
+        use_cache=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 tokens that are **masked**.
+        labels (:obj:`torch.LongTensor` of shape :obj:`(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 n ``[0, ..., config.vocab_size]``
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(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 :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+
+        Returns:
+
+        Example::
+
+            >>> from transformers import BertTokenizer, BertLMHeadModel, BertConfig
+            >>> import torch
+
+            >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
+            >>> config = BertConfig.from_pretrained("bert-base-cased")
+            >>> config.is_decoder = True
+            >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', 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.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,
+            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.cls(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("""Bert Model with a `language modeling` head on top. """, BERT_START_DOCSTRING)
+class BertForMaskedLM(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        if config.is_decoder:
+            logger.warning(
+                "If you want to use `BertForMaskedLM` make sure `config.is_decoder=False` for "
+                "bi-directional self-attention."
+            )
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(BERT_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,
+    )
+    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,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(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]``
+        """
+
+        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,
+            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.cls(sequence_output)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
+            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,
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **model_kwargs):
+        input_shape = input_ids.shape
+        effective_batch_size = input_shape[0]
+
+        #  add a dummy token
+        assert self.config.pad_token_id is not None, "The PAD token should be defined for generation"
+        attention_mask = torch.cat([attention_mask, attention_mask.new_zeros((attention_mask.shape[0], 1))], dim=-1)
+        dummy_token = torch.full(
+            (effective_batch_size, 1), self.config.pad_token_id, dtype=torch.long, device=input_ids.device
+        )
+        input_ids = torch.cat([input_ids, dummy_token], dim=1)
+
+        return {"input_ids": input_ids, "attention_mask": attention_mask}
+
+
+@add_start_docstrings(
+    """Bert Model with a `next sentence prediction (classification)` head on top. """,
+    BERT_START_DOCSTRING,
+)
+class BertForNextSentencePrediction(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config)
+        self.cls = BertOnlyNSPHead(config)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=NextSentencePredictorOutput, config_class=_CONFIG_FOR_DOC)
+    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,
+        **kwargs,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair
+            (see ``input_ids`` docstring). Indices should be in ``[0, 1]``:
+
+            - 0 indicates sequence B is a continuation of sequence A,
+            - 1 indicates sequence B is a random sequence.
+
+        Returns:
+
+        Example::
+
+            >>> from transformers import BertTokenizer, BertForNextSentencePrediction
+            >>> import torch
+
+            >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+            >>> model = BertForNextSentencePrediction.from_pretrained('bert-base-uncased')
+
+            >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
+            >>> next_sentence = "The sky is blue due to the shorter wavelength of blue light."
+            >>> encoding = tokenizer(prompt, next_sentence, return_tensors='pt')
+
+            >>> outputs = model(**encoding, labels=torch.LongTensor([1]))
+            >>> logits = outputs.logits
+            >>> assert logits[0, 0] < logits[0, 1] # next sentence was random
+        """
+
+        if "next_sentence_label" in kwargs:
+            warnings.warn(
+                "The `next_sentence_label` argument is deprecated and will be removed in a future version, use `labels` instead.",
+                FutureWarning,
+            )
+            labels = kwargs.pop("next_sentence_label")
+
+        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]
+
+        seq_relationship_scores = self.cls(pooled_output)
+
+        next_sentence_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            next_sentence_loss = loss_fct(seq_relationship_scores.view(-1, 2), labels.view(-1))
+
+        if not return_dict:
+            output = (seq_relationship_scores,) + outputs[2:]
+            return ((next_sentence_loss,) + output) if next_sentence_loss is not None else output
+
+        return NextSentencePredictorOutput(
+            loss=next_sentence_loss,
+            logits=seq_relationship_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
+    output) e.g. for GLUE tasks.
+    """,
+    BERT_START_DOCSTRING,
+)
+class BertForSequenceClassification(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.bert = BertModel(config)
+        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)
+
+        self.skim_coefficient = config.skim_coefficient if hasattr(config, 'skim_coefficient') else 1
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    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).
+        """
+        # assert gumbel_softmax is not None
+        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.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
+
+        skim_loss, neat_mac = 0.0, 0.0
+        layer_neat_mac = list()
+        all_tokens_length = torch.mean(torch.sum(attention_mask.to(torch.float32),dim=-1))
+        for mask in outputs.skim_mask:
+            accumulated_skim_mask = torch.mean(torch.sum(mask,dim=1))
+            skim_loss += accumulated_skim_mask/mask.shape[1]
+            layer_neat_mac.append(accumulated_skim_mask/all_tokens_length)
+            neat_mac += accumulated_skim_mask/all_tokens_length
+        skim_loss /= self.config.num_hidden_layers
+        neat_mac /= self.config.num_hidden_layers
+        classification_loss = loss
+        # print(skim_loss, neat_mac, loss)
+        # loss = skim_loss
+        if labels is not None:
+            loss = self.skim_coefficient * skim_loss + loss
+
+        return SequenceClassifierOutputSkim(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            attention_mask=outputs.attention_mask,
+            skim_mask=outputs.skim_mask,
+            skim_loss=skim_loss,
+            classification_loss=classification_loss,
+            tokens_remained=neat_mac,
+            layer_tokens_remained=layer_neat_mac,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bert 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.
+    """,
+    BERT_START_DOCSTRING,
+)
+class BertForMultipleChoice(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config)
+        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)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BERT_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=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 multiple choice classification loss. Indices should be in ``[0, ...,
+            num_choices-1]`` where :obj:`num_choices` is the size of the second dimension of the input tensors. (See
+            :obj:`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,
+        )
+
+        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(
+    """
+    Bert 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.
+    """,
+    BERT_START_DOCSTRING,
+)
+class BertForTokenClassification(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(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)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    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, 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.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,
+        )
+
+        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()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)
+                active_labels = torch.where(
+                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
+                )
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                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,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bert 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`).
+    """,
+    BERT_START_DOCSTRING,
+)
+class BertForQuestionAnswering(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.skim_coefficient = config.skim_coefficient if hasattr(config, 'skim_coefficient') else 1
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        start_positions=None,
+        end_positions=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        start_positions (:obj:`torch.LongTensor` of shape :obj:`(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 (:obj:`sequence_length`). Position outside of the
+            sequence are not taken into account for computing the loss.
+        end_positions (:obj:`torch.LongTensor` of shape :obj:`(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 (:obj:`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,
+        )
+
+        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
+
+        skim_loss, neat_mac = 0.0, 0.0
+        layer_neat_mac = list()
+        all_tokens_length = torch.mean(torch.sum(attention_mask.to(torch.float32),dim=-1))
+        for mask in outputs.skim_mask:
+            accumulated_skim_mask = torch.mean(torch.sum(mask,dim=1))
+            skim_loss += accumulated_skim_mask/mask.shape[1]
+            layer_neat_mac.append(accumulated_skim_mask/all_tokens_length)
+            neat_mac += accumulated_skim_mask/all_tokens_length
+        skim_loss /= self.config.num_hidden_layers
+        neat_mac /= self.config.num_hidden_layers
+        qa_loss = total_loss
+        if start_positions is not None and end_positions is not None:
+            # print(skim_loss, neat_mac, loss)
+            # loss = skim_loss
+            total_loss = self.skim_coefficient * skim_loss + qa_loss
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+
+        return QuestionAnsweringModelOutputSkim(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            attention_mask=outputs.attention_mask,
+            skim_mask=outputs.skim_mask,
+            skim_loss=skim_loss,
+            classification_loss=qa_loss,
+            tokens_remained=neat_mac,
+            layer_tokens_remained=layer_neat_mac,
+        )
+
+
+def test_BertEncoder():
+    import transformers
+
+    logging.debug(f'Start unit test for BertEncoder')
+
+    config = transformers.BertConfig.from_pretrained('bert-base-uncased')
+    # config.output_attentions = False
+    encoder = BertEncoder(config)
+
+    rand_hidden_states = torch.rand((1,8,768))
+    # rand_hidden_states = torch.rand((4,128,768))
+
+    encoder_outputs = encoder(rand_hidden_states)
+
+    logging.debug(f'output attention: {config.output_attentions}, {encoder_outputs[-1][0].shape}')
+
+if __name__ == "__main__":
+    import logging
+
+    logging.basicConfig(level=logging.DEBUG)
+
+    test_BertEncoder()

test_module/modeling_transkimer_roberta.py

@@ -0,0 +1,1624 @@
+# 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
+
+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.file_utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.utils import logging
+from transformers.models.roberta.configuration_roberta import RobertaConfig
+
+from module.modeling_skim_predictor import SkimPredictor
+from module.modeling_utils import BaseModelOutputWithPastAndCrossAttentionsSkim, BaseModelOutputWithPoolingAndCrossAttentionsSkim, SequenceClassifierOutputSkim
+
+
+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, device=self.position_ids.device),
+                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):
+        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 = getattr(config, "position_embedding_type", "absolute")
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        skim_mask=None,
+    ):
+        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.Softmax(dim=-1)(attention_scores)
+
+        # mask attention probs during training for skimming
+        attention_probs = attention_probs * skim_mask[:, None, None, :]
+
+        # 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, input_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):
+        super().__init__()
+        self.self = RobertaSelfAttention(config)
+        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,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        skim_mask=None,
+    ):
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+            skim_mask,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate
+class RobertaIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states):
+        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, input_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:
+            assert self.is_decoder, f"{self} should be used as a decoder model if cross attention is added"
+            self.crossattention = RobertaAttention(config)
+        self.intermediate = RobertaIntermediate(config)
+        self.output = RobertaOutput(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        skim_mask=None,
+    ):
+        # 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,
+            skim_mask=skim_mask,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            assert hasattr(
+                self, "crossattention"
+            ), 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)])
+
+        # skim predictors for each layer
+        self.skim_predictors = nn.ModuleList([SkimPredictor(config.hidden_size, 2) 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,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=False,
+        output_hidden_states=False,
+        return_dict=True,
+    ):
+        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
+        all_skim_mask = ()
+        forward_hidden_states = hidden_states.clone()
+
+        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,)
+
+            skim_mask = nn.functional.gumbel_softmax(self.skim_predictors[i](hidden_states[:,1:,:]), hard=True, tau=1)
+            skim_mask = skim_mask[:,:,1]
+            skim_mask_with_cls = torch.ones(skim_mask.shape[0], skim_mask.shape[1]+1, device=skim_mask.device)
+            skim_mask_with_cls[:,1:] = skim_mask
+            skim_mask = skim_mask_with_cls
+            # multiple current layer skim mask with last layer skim mask
+            # to gurantee skimmed tokens are never recovered
+            if all_skim_mask:
+                skim_mask = skim_mask * all_skim_mask[-1]
+            all_skim_mask += (skim_mask, )
+
+            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 getattr(self.config, "gradient_checkpointing", False) and self.training:
+
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. 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,
+                    skim_mask,
+                )
+
+            hidden_states = layer_outputs[0]
+            forward_hidden_states = forward_hidden_states * (1-skim_mask.view(*skim_mask.shape,1)) + hidden_states * skim_mask.view(*skim_mask.shape,1)
+            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 [
+                    forward_hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentionsSkim(
+            last_hidden_state=forward_hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+            attention_mask=attention_mask,
+            skim_mask=all_skim_mask,
+        )
+
+
+# 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):
+        # 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"
+
+    # Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if hasattr(module, '_skim_initialized') and module._skim_initialized:
+            return
+        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 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 :class:`~transformers.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 (:class:`~transformers.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 :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
+            weights.
+"""
+
+ROBERTA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.RobertaTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`({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.html#attention-mask>`__
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`({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.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`({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.html#position-ids>`_
+        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(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 (:obj:`torch.FloatTensor` of shape :obj:`({0}, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
+        output_attentions (:obj:`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 (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_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 :obj:`is_decoder` argument of the configuration
+    set to :obj:`True`. To be used in a Seq2Seq model, the model needs to initialized with both :obj:`is_decoder`
+    argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`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
+
+        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 _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=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,
+        past_key_values=None,
+        use_cache=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 tokens that are **masked**.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(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 :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        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 BaseModelOutputWithPoolingAndCrossAttentionsSkim(
+            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,
+            attention_mask=encoder_outputs.attention_mask,
+            skim_mask=encoder_outputs.skim_mask,
+        )
+
+
+@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"])
+
+        self.init_weights()
+
+    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=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,
+        labels=None,
+        past_key_values=None,
+        use_cache=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 tokens that are **masked**.
+
+        labels (:obj:`torch.LongTensor` of shape :obj:`(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 (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(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 :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+
+        Returns:
+
+        Example::
+
+            >>> 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"])
+
+        self.init_weights()
+
+    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>",
+    )
+    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,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(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 (:obj:`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)
+
+        self.skim_coefficient = config.skim_coefficient if hasattr(config, 'skim_coefficient') else 1
+
+        self.init_weights()
+
+    @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=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    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.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
+
+        skim_loss, neat_mac = 0.0, 0.0
+        layer_neat_mac = list()
+        all_tokens_length = torch.mean(torch.sum(attention_mask.to(torch.float32),dim=-1))
+        for mask in outputs.skim_mask:
+            accumulated_skim_mask = torch.mean(torch.sum(mask,dim=1))
+            skim_loss += accumulated_skim_mask/mask.shape[1]
+            layer_neat_mac.append(accumulated_skim_mask/all_tokens_length)
+            neat_mac += accumulated_skim_mask/all_tokens_length
+        skim_loss /= self.config.num_hidden_layers
+        neat_mac /= self.config.num_hidden_layers
+        classification_loss = loss
+        # print(skim_loss, neat_mac, loss)
+        # loss = skim_loss
+        loss = self.skim_coefficient * skim_loss + loss
+
+        return SequenceClassifierOutputSkim(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            attention_mask=outputs.attention_mask,
+            skim_mask=outputs.skim_mask,
+            skim_loss=skim_loss,
+            classification_loss=classification_loss,
+            tokens_remained=neat_mac,
+            layer_tokens_remained=layer_neat_mac,            
+        )
+
+
+@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)
+
+        self.init_weights()
+
+    @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=None,
+        token_type_ids=None,
+        attention_mask=None,
+        labels=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=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 multiple choice classification loss. Indices should be in ``[0, ...,
+            num_choices-1]`` where :obj:`num_choices` is the size of the second dimension of the input tensors. (See
+            :obj:`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)
+
+        self.init_weights()
+
+    @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=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    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, 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()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)
+                active_labels = torch.where(
+                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
+                )
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                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)
+
+        self.init_weights()
+
+    @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=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        start_positions=None,
+        end_positions=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        start_positions (:obj:`torch.LongTensor` of shape :obj:`(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 (:obj:`sequence_length`). Position outside of the
+            sequence are not taken into account for computing the loss.
+        end_positions (:obj:`torch.LongTensor` of shape :obj:`(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 (:obj:`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

test_module/modeling_utils.py

@@ -0,0 +1,92 @@
+from dataclasses import dataclass
+import torch
+from transformers.file_utils import ModelOutput
+from typing import Optional, Tuple
+
+from transformers.modeling_outputs import MaskedLMOutput, QuestionAnsweringModelOutput
+
+@dataclass
+class BaseModelOutputWithPastAndCrossAttentionsSkim(ModelOutput):
+    last_hidden_state: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    attention_mask: Optional[torch.FloatTensor] = None
+    skim_mask: Optional[torch.FloatTensor] = None
+
+@dataclass
+class BaseModelOutputWithPoolingAndCrossAttentionsSkim(ModelOutput):
+    last_hidden_state: torch.FloatTensor = None
+    pooler_output: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    attention_mask: Optional[torch.FloatTensor] = None
+    skim_mask: Optional[torch.FloatTensor] = None
+
+
+@dataclass
+class SequenceClassifierOutputSkim(ModelOutput):
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    attention_mask: Optional[torch.FloatTensor] = None
+    skim_mask: Optional[torch.FloatTensor] = None
+    skim_loss: Optional[torch.FloatTensor] = None
+    classification_loss: Optional[torch.FloatTensor] = None
+    tokens_remained: Optional[torch.FloatTensor] = None
+    layer_tokens_remained: Optional[Tuple[torch.FloatTensor]] = None
+
+@dataclass
+class QuestionAnsweringModelOutputSkim(QuestionAnsweringModelOutput):
+    attention_mask: Optional[torch.FloatTensor] = None
+    skim_mask: Optional[torch.FloatTensor] = None
+    skim_loss: Optional[torch.FloatTensor] = None
+    classification_loss: Optional[torch.FloatTensor] = None
+    tokens_remained: Optional[torch.FloatTensor] = None
+    layer_tokens_remained: Optional[Tuple[torch.FloatTensor]] = None
+
+@dataclass
+class MaskedLMOutputSkim(MaskedLMOutput):
+    attention_mask: Optional[torch.FloatTensor] = None
+    skim_mask: Optional[torch.FloatTensor] = None
+    skim_loss: Optional[torch.FloatTensor] = None
+    classification_loss: Optional[torch.FloatTensor] = None
+    tokens_remained: Optional[torch.FloatTensor] = None
+    layer_tokens_remained: Optional[Tuple[torch.FloatTensor]] = None
+
+def masked_softmax(vec, mask, dim=1, eps=1e-6):
+    mask = mask[:,None,None,:]
+    exps = torch.exp(vec)
+    masked_exps = exps * mask.float() + eps
+    masked_sums = masked_exps.sum(dim, keepdim=True) 
+    return (masked_exps/masked_sums)
+
+def convert_softmax_mask_to_digit(skim_mask):
+    # skim_mask [batch, from, to, seq_len]
+    return (skim_mask == 0).to(dtype=torch.int64).unsqueeze(1).unsqueeze(1)
+
+def trunc_with_mask_batched(input, mask, dim):
+    """
+    trunc a batched input at dim
+        e.g. hidden_states ([batch, seq_len, hidden_size])
+            attention_mask ([batch, layer, head, seq_len])
+    mask: [batch, seq_len]
+    """
+    assert input.shape[dim]==mask.shape[1]
+
+    if dim != 1:
+        input = input.transpose(1, dim)
+
+    transpose_shape = list(input.shape)
+    transpose_shape[1] = -1
+
+    trunc_input = input[mask].view(transpose_shape)
+
+    if dim != 1:
+        trunc_input = trunc_input.transpose(1, dim)
+
+    return trunc_input

utils/extend_auto_mapping.py

@@ -0,0 +1,13 @@
+def extend_lazy_auto_mapping(mapping, key, config_value, value):
+    prev_config_mapping = mapping._config_mapping
+    prev_model_mapping = mapping._model_mapping
+
+    prev_config_mapping[key] = config_value
+    prev_model_mapping[key] = value
+
+def test_extending():
+    import transformers
+    extend_lazy_auto_mapping(transformers.models.auto.MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING, 'test', transformers.models.bert.configuration_bert, transformers.models.bert.BertForSequenceClassification)
+
+if __name__ == "__main__":
+    test_extending()

utils/utils_qa.py

@@ -0,0 +1,427 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Team 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.
+"""
+Post-processing utilities for question answering.
+"""
+import collections
+import json
+import logging
+import os
+from typing import Optional, Tuple
+
+import numpy as np
+from tqdm.auto import tqdm
+
+
+logger = logging.getLogger(__name__)
+
+
+def postprocess_qa_predictions(
+    examples,
+    features,
+    predictions: Tuple[np.ndarray, np.ndarray],
+    version_2_with_negative: bool = False,
+    n_best_size: int = 20,
+    max_answer_length: int = 30,
+    null_score_diff_threshold: float = 0.0,
+    output_dir: Optional[str] = None,
+    prefix: Optional[str] = None,
+    log_level: Optional[int] = logging.WARNING,
+):
+    """
+    Post-processes the predictions of a question-answering model to convert them to answers that are substrings of the
+    original contexts. This is the base postprocessing functions for models that only return start and end logits.
+
+    Args:
+        examples: The non-preprocessed dataset (see the main script for more information).
+        features: The processed dataset (see the main script for more information).
+        predictions (:obj:`Tuple[np.ndarray, np.ndarray]`):
+            The predictions of the model: two arrays containing the start logits and the end logits respectively. Its
+            first dimension must match the number of elements of :obj:`features`.
+        version_2_with_negative (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not the underlying dataset contains examples with no answers.
+        n_best_size (:obj:`int`, `optional`, defaults to 20):
+            The total number of n-best predictions to generate when looking for an answer.
+        max_answer_length (:obj:`int`, `optional`, defaults to 30):
+            The maximum length of an answer that can be generated. This is needed because the start and end predictions
+            are not conditioned on one another.
+        null_score_diff_threshold (:obj:`float`, `optional`, defaults to 0):
+            The threshold used to select the null answer: if the best answer has a score that is less than the score of
+            the null answer minus this threshold, the null answer is selected for this example (note that the score of
+            the null answer for an example giving several features is the minimum of the scores for the null answer on
+            each feature: all features must be aligned on the fact they `want` to predict a null answer).
+
+            Only useful when :obj:`version_2_with_negative` is :obj:`True`.
+        output_dir (:obj:`str`, `optional`):
+            If provided, the dictionaries of predictions, n_best predictions (with their scores and logits) and, if
+            :obj:`version_2_with_negative=True`, the dictionary of the scores differences between best and null
+            answers, are saved in `output_dir`.
+        prefix (:obj:`str`, `optional`):
+            If provided, the dictionaries mentioned above are saved with `prefix` added to their names.
+        log_level (:obj:`int`, `optional`, defaults to ``logging.WARNING``):
+            ``logging`` log level (e.g., ``logging.WARNING``)
+    """
+    assert len(predictions) == 2, "`predictions` should be a tuple with two elements (start_logits, end_logits)."
+    all_start_logits, all_end_logits = predictions
+
+    assert len(predictions[0]) == len(features), f"Got {len(predictions[0])} predictions and {len(features)} features."
+
+    # Build a map example to its corresponding features.
+    example_id_to_index = {k: i for i, k in enumerate(examples["id"])}
+    features_per_example = collections.defaultdict(list)
+    for i, feature in enumerate(features):
+        features_per_example[example_id_to_index[feature["example_id"]]].append(i)
+
+    # The dictionaries we have to fill.
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    if version_2_with_negative:
+        scores_diff_json = collections.OrderedDict()
+
+    # Logging.
+    logger.setLevel(log_level)
+    logger.info(f"Post-processing {len(examples)} example predictions split into {len(features)} features.")
+
+    # Let's loop over all the examples!
+    for example_index, example in enumerate(tqdm(examples)):
+        # Those are the indices of the features associated to the current example.
+        feature_indices = features_per_example[example_index]
+
+        min_null_prediction = None
+        prelim_predictions = []
+
+        # Looping through all the features associated to the current example.
+        for feature_index in feature_indices:
+            # We grab the predictions of the model for this feature.
+            start_logits = all_start_logits[feature_index]
+            end_logits = all_end_logits[feature_index]
+            # This is what will allow us to map some the positions in our logits to span of texts in the original
+            # context.
+            offset_mapping = features[feature_index]["offset_mapping"]
+            # Optional `token_is_max_context`, if provided we will remove answers that do not have the maximum context
+            # available in the current feature.
+            token_is_max_context = features[feature_index].get("token_is_max_context", None)
+
+            # Update minimum null prediction.
+            feature_null_score = start_logits[0] + end_logits[0]
+            if min_null_prediction is None or min_null_prediction["score"] > feature_null_score:
+                min_null_prediction = {
+                    "offsets": (0, 0),
+                    "score": feature_null_score,
+                    "start_logit": start_logits[0],
+                    "end_logit": end_logits[0],
+                }
+
+            # Go through all possibilities for the `n_best_size` greater start and end logits.
+            start_indexes = np.argsort(start_logits)[-1 : -n_best_size - 1 : -1].tolist()
+            end_indexes = np.argsort(end_logits)[-1 : -n_best_size - 1 : -1].tolist()
+            for start_index in start_indexes:
+                for end_index in end_indexes:
+                    # Don't consider out-of-scope answers, either because the indices are out of bounds or correspond
+                    # to part of the input_ids that are not in the context.
+                    if (
+                        start_index >= len(offset_mapping)
+                        or end_index >= len(offset_mapping)
+                        or offset_mapping[start_index] is None
+                        or offset_mapping[end_index] is None
+                    ):
+                        continue
+                    # Don't consider answers with a length that is either < 0 or > max_answer_length.
+                    if end_index < start_index or end_index - start_index + 1 > max_answer_length:
+                        continue
+                    # Don't consider answer that don't have the maximum context available (if such information is
+                    # provided).
+                    if token_is_max_context is not None and not token_is_max_context.get(str(start_index), False):
+                        continue
+                    prelim_predictions.append(
+                        {
+                            "offsets": (offset_mapping[start_index][0], offset_mapping[end_index][1]),
+                            "score": start_logits[start_index] + end_logits[end_index],
+                            "start_logit": start_logits[start_index],
+                            "end_logit": end_logits[end_index],
+                        }
+                    )
+        if version_2_with_negative:
+            # Add the minimum null prediction
+            prelim_predictions.append(min_null_prediction)
+            null_score = min_null_prediction["score"]
+
+        # Only keep the best `n_best_size` predictions.
+        predictions = sorted(prelim_predictions, key=lambda x: x["score"], reverse=True)[:n_best_size]
+
+        # Add back the minimum null prediction if it was removed because of its low score.
+        if version_2_with_negative and not any(p["offsets"] == (0, 0) for p in predictions):
+            predictions.append(min_null_prediction)
+
+        # Use the offsets to gather the answer text in the original context.
+        context = example["context"]
+        for pred in predictions:
+            offsets = pred.pop("offsets")
+            pred["text"] = context[offsets[0] : offsets[1]]
+
+        # In the very rare edge case we have not a single non-null prediction, we create a fake prediction to avoid
+        # failure.
+        if len(predictions) == 0 or (len(predictions) == 1 and predictions[0]["text"] == ""):
+            predictions.insert(0, {"text": "empty", "start_logit": 0.0, "end_logit": 0.0, "score": 0.0})
+
+        # Compute the softmax of all scores (we do it with numpy to stay independent from torch/tf in this file, using
+        # the LogSumExp trick).
+        scores = np.array([pred.pop("score") for pred in predictions])
+        exp_scores = np.exp(scores - np.max(scores))
+        probs = exp_scores / exp_scores.sum()
+
+        # Include the probabilities in our predictions.
+        for prob, pred in zip(probs, predictions):
+            pred["probability"] = prob
+
+        # Pick the best prediction. If the null answer is not possible, this is easy.
+        if not version_2_with_negative:
+            all_predictions[example["id"]] = predictions[0]["text"]
+        else:
+            # Otherwise we first need to find the best non-empty prediction.
+            i = 0
+            while predictions[i]["text"] == "":
+                i += 1
+            best_non_null_pred = predictions[i]
+
+            # Then we compare to the null prediction using the threshold.
+            score_diff = null_score - best_non_null_pred["start_logit"] - best_non_null_pred["end_logit"]
+            scores_diff_json[example["id"]] = float(score_diff)  # To be JSON-serializable.
+            if score_diff > null_score_diff_threshold:
+                all_predictions[example["id"]] = ""
+            else:
+                all_predictions[example["id"]] = best_non_null_pred["text"]
+
+        # Make `predictions` JSON-serializable by casting np.float back to float.
+        all_nbest_json[example["id"]] = [
+            {k: (float(v) if isinstance(v, (np.float16, np.float32, np.float64)) else v) for k, v in pred.items()}
+            for pred in predictions
+        ]
+
+    # If we have an output_dir, let's save all those dicts.
+    if output_dir is not None:
+        assert os.path.isdir(output_dir), f"{output_dir} is not a directory."
+
+        prediction_file = os.path.join(
+            output_dir, "predictions.json" if prefix is None else f"{prefix}_predictions.json"
+        )
+        nbest_file = os.path.join(
+            output_dir, "nbest_predictions.json" if prefix is None else f"{prefix}_nbest_predictions.json"
+        )
+        if version_2_with_negative:
+            null_odds_file = os.path.join(
+                output_dir, "null_odds.json" if prefix is None else f"{prefix}_null_odds.json"
+            )
+
+        logger.info(f"Saving predictions to {prediction_file}.")
+        with open(prediction_file, "w") as writer:
+            writer.write(json.dumps(all_predictions, indent=4) + "\n")
+        logger.info(f"Saving nbest_preds to {nbest_file}.")
+        with open(nbest_file, "w") as writer:
+            writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+        if version_2_with_negative:
+            logger.info(f"Saving null_odds to {null_odds_file}.")
+            with open(null_odds_file, "w") as writer:
+                writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
+
+    return all_predictions
+
+
+def postprocess_qa_predictions_with_beam_search(
+    examples,
+    features,
+    predictions: Tuple[np.ndarray, np.ndarray],
+    version_2_with_negative: bool = False,
+    n_best_size: int = 20,
+    max_answer_length: int = 30,
+    start_n_top: int = 5,
+    end_n_top: int = 5,
+    output_dir: Optional[str] = None,
+    prefix: Optional[str] = None,
+    log_level: Optional[int] = logging.WARNING,
+):
+    """
+    Post-processes the predictions of a question-answering model with beam search to convert them to answers that are substrings of the
+    original contexts. This is the postprocessing functions for models that return start and end logits, indices, as well as
+    cls token predictions.
+
+    Args:
+        examples: The non-preprocessed dataset (see the main script for more information).
+        features: The processed dataset (see the main script for more information).
+        predictions (:obj:`Tuple[np.ndarray, np.ndarray]`):
+            The predictions of the model: two arrays containing the start logits and the end logits respectively. Its
+            first dimension must match the number of elements of :obj:`features`.
+        version_2_with_negative (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not the underlying dataset contains examples with no answers.
+        n_best_size (:obj:`int`, `optional`, defaults to 20):
+            The total number of n-best predictions to generate when looking for an answer.
+        max_answer_length (:obj:`int`, `optional`, defaults to 30):
+            The maximum length of an answer that can be generated. This is needed because the start and end predictions
+            are not conditioned on one another.
+        start_n_top (:obj:`int`, `optional`, defaults to 5):
+            The number of top start logits too keep when searching for the :obj:`n_best_size` predictions.
+        end_n_top (:obj:`int`, `optional`, defaults to 5):
+            The number of top end logits too keep when searching for the :obj:`n_best_size` predictions.
+        output_dir (:obj:`str`, `optional`):
+            If provided, the dictionaries of predictions, n_best predictions (with their scores and logits) and, if
+            :obj:`version_2_with_negative=True`, the dictionary of the scores differences between best and null
+            answers, are saved in `output_dir`.
+        prefix (:obj:`str`, `optional`):
+            If provided, the dictionaries mentioned above are saved with `prefix` added to their names.
+        log_level (:obj:`int`, `optional`, defaults to ``logging.WARNING``):
+            ``logging`` log level (e.g., ``logging.WARNING``)
+    """
+    assert len(predictions) == 5, "`predictions` should be a tuple with five elements."
+    start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits = predictions
+
+    assert len(predictions[0]) == len(
+        features
+    ), f"Got {len(predictions[0])} predicitions and {len(features)} features."
+
+    # Build a map example to its corresponding features.
+    example_id_to_index = {k: i for i, k in enumerate(examples["id"])}
+    features_per_example = collections.defaultdict(list)
+    for i, feature in enumerate(features):
+        features_per_example[example_id_to_index[feature["example_id"]]].append(i)
+
+    # The dictionaries we have to fill.
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    scores_diff_json = collections.OrderedDict() if version_2_with_negative else None
+
+    # Logging.
+    logger.setLevel(log_level)
+    logger.info(f"Post-processing {len(examples)} example predictions split into {len(features)} features.")
+
+    # Let's loop over all the examples!
+    for example_index, example in enumerate(tqdm(examples)):
+        # Those are the indices of the features associated to the current example.
+        feature_indices = features_per_example[example_index]
+
+        min_null_score = None
+        prelim_predictions = []
+
+        # Looping through all the features associated to the current example.
+        for feature_index in feature_indices:
+            # We grab the predictions of the model for this feature.
+            start_log_prob = start_top_log_probs[feature_index]
+            start_indexes = start_top_index[feature_index]
+            end_log_prob = end_top_log_probs[feature_index]
+            end_indexes = end_top_index[feature_index]
+            feature_null_score = cls_logits[feature_index]
+            # This is what will allow us to map some the positions in our logits to span of texts in the original
+            # context.
+            offset_mapping = features[feature_index]["offset_mapping"]
+            # Optional `token_is_max_context`, if provided we will remove answers that do not have the maximum context
+            # available in the current feature.
+            token_is_max_context = features[feature_index].get("token_is_max_context", None)
+
+            # Update minimum null prediction
+            if min_null_score is None or feature_null_score < min_null_score:
+                min_null_score = feature_null_score
+
+            # Go through all possibilities for the `n_start_top`/`n_end_top` greater start and end logits.
+            for i in range(start_n_top):
+                for j in range(end_n_top):
+                    start_index = int(start_indexes[i])
+                    j_index = i * end_n_top + j
+                    end_index = int(end_indexes[j_index])
+                    # Don't consider out-of-scope answers (last part of the test should be unnecessary because of the
+                    # p_mask but let's not take any risk)
+                    if (
+                        start_index >= len(offset_mapping)
+                        or end_index >= len(offset_mapping)
+                        or offset_mapping[start_index] is None
+                        or offset_mapping[end_index] is None
+                    ):
+                        continue
+                    # Don't consider answers with a length negative or > max_answer_length.
+                    if end_index < start_index or end_index - start_index + 1 > max_answer_length:
+                        continue
+                    # Don't consider answer that don't have the maximum context available (if such information is
+                    # provided).
+                    if token_is_max_context is not None and not token_is_max_context.get(str(start_index), False):
+                        continue
+                    prelim_predictions.append(
+                        {
+                            "offsets": (offset_mapping[start_index][0], offset_mapping[end_index][1]),
+                            "score": start_log_prob[i] + end_log_prob[j_index],
+                            "start_log_prob": start_log_prob[i],
+                            "end_log_prob": end_log_prob[j_index],
+                        }
+                    )
+
+        # Only keep the best `n_best_size` predictions.
+        predictions = sorted(prelim_predictions, key=lambda x: x["score"], reverse=True)[:n_best_size]
+
+        # Use the offsets to gather the answer text in the original context.
+        context = example["context"]
+        for pred in predictions:
+            offsets = pred.pop("offsets")
+            pred["text"] = context[offsets[0] : offsets[1]]
+
+        # In the very rare edge case we have not a single non-null prediction, we create a fake prediction to avoid
+        # failure.
+        if len(predictions) == 0:
+            predictions.insert(0, {"text": "", "start_logit": -1e-6, "end_logit": -1e-6, "score": -2e-6})
+
+        # Compute the softmax of all scores (we do it with numpy to stay independent from torch/tf in this file, using
+        # the LogSumExp trick).
+        scores = np.array([pred.pop("score") for pred in predictions])
+        exp_scores = np.exp(scores - np.max(scores))
+        probs = exp_scores / exp_scores.sum()
+
+        # Include the probabilities in our predictions.
+        for prob, pred in zip(probs, predictions):
+            pred["probability"] = prob
+
+        # Pick the best prediction and set the probability for the null answer.
+        all_predictions[example["id"]] = predictions[0]["text"]
+        if version_2_with_negative:
+            scores_diff_json[example["id"]] = float(min_null_score)
+
+        # Make `predictions` JSON-serializable by casting np.float back to float.
+        all_nbest_json[example["id"]] = [
+            {k: (float(v) if isinstance(v, (np.float16, np.float32, np.float64)) else v) for k, v in pred.items()}
+            for pred in predictions
+        ]
+
+    # If we have an output_dir, let's save all those dicts.
+    if output_dir is not None:
+        assert os.path.isdir(output_dir), f"{output_dir} is not a directory."
+
+        prediction_file = os.path.join(
+            output_dir, "predictions.json" if prefix is None else f"{prefix}_predictions.json"
+        )
+        nbest_file = os.path.join(
+            output_dir, "nbest_predictions.json" if prefix is None else f"{prefix}_nbest_predictions.json"
+        )
+        if version_2_with_negative:
+            null_odds_file = os.path.join(
+                output_dir, "null_odds.json" if prefix is None else f"{prefix}_null_odds.json"
+            )
+
+        logger.info(f"Saving predictions to {prediction_file}.")
+        with open(prediction_file, "w") as writer:
+            writer.write(json.dumps(all_predictions, indent=4) + "\n")
+        logger.info(f"Saving nbest_preds to {nbest_file}.")
+        with open(nbest_file, "w") as writer:
+            writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+        if version_2_with_negative:
+            logger.info(f"Saving null_odds to {null_odds_file}.")
+            with open(null_odds_file, "w") as writer:
+                writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
+
+    return all_predictions, scores_diff_json