modelingpreln.py

# Copyright (c) Microsoft Corporation.
# SPDX-License-Identifier: Apache-2.0

# DeepSpeed Team

from __future__ import absolute_import, division, print_function, unicode_literals
# Copyright The Microsoft DeepSpeed Team
# DeepSpeed note, code taken from commit 3d59216cec89a363649b4fe3d15295ba936ced0f
# https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/LanguageModeling/BERT/modeling.py

# 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 copy
import json
import logging
import math
import os
import shutil
import tarfile
import tempfile
from io import open

import torch
from torch import nn
from torch.nn import CrossEntropyLoss
from torch.utils import checkpoint
import deepspeed.comm as dist

from torch.nn import Module
import torch.nn.functional as F
import torch.nn.init as init
from deepspeed.accelerator import get_accelerator

#from numba import cuda

#from deepspeed_cuda import DeepSpeedSoftmaxConfig, DeepSpeedSoftmax

logger = logging.getLogger(__name__)

PRETRAINED_MODEL_ARCHIVE_MAP = {
    'bert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased.tar.gz",
    'bert-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased.tar.gz",
    'bert-base-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased.tar.gz",
    'bert-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased.tar.gz",
    'bert-base-multilingual-uncased':
    "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased.tar.gz",
    'bert-base-multilingual-cased':
    "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased.tar.gz",
    'bert-base-chinese': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese.tar.gz",
}
CONFIG_NAME = 'bert_config.json'
WEIGHTS_NAME = 'pytorch_model.bin'
TF_WEIGHTS_NAME = 'model.ckpt'


def load_tf_weights_in_bert(model, tf_checkpoint_path):
    """ Load tf checkpoints in a pytorch model
    """
    try:
        import re
        import numpy as np
        import tensorflow as tf
    except ImportError:
        print("Loading a TensorFlow models 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)
    print("Converting TensorFlow checkpoint from {}".format(tf_path))
    # Load weights from TF model
    init_vars = tf.train.list_variables(tf_path)
    names = []
    arrays = []
    for name, shape in init_vars:
        print("Loading TF weight {} with shape {}".format(name, 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"] for n in name):
            print("Skipping {}".format("/".join(name)))
            continue
        pointer = model
        for m_name in name:
            if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
                l = re.split(r'_(\d+)', m_name)
            else:
                l = [m_name]
            if l[0] == 'kernel' or l[0] == 'gamma':
                pointer = getattr(pointer, 'weight')
            elif l[0] == 'output_bias' or l[0] == 'beta':
                pointer = getattr(pointer, 'bias')
            elif l[0] == 'output_weights':
                pointer = getattr(pointer, 'weight')
            else:
                pointer = getattr(pointer, l[0])
            if len(l) >= 2:
                num = int(l[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
        except AssertionError as e:
            e.args += (pointer.shape, array.shape)
            raise
        print("Initialize PyTorch weight {}".format(name))
        pointer.data = torch.from_numpy(array)
    return model


"""
@torch.jit.script
def f_gelu(x):
    return x * 0.5 * (1.0 + torch.erf(x / 1.41421))
@torch.jit.script
def bias_gelu(bias, y):
    x = bias + y
    return x * 0.5 * (1.0 + torch.erf(x / 1.41421))
@torch.jit.script
def bias_tanh(bias, y):
    x = bias + y
    return torch.tanh(x)
 """


def f_gelu(x):
    x_type = x.dtype
    x = x.float()
    x = x * 0.5 * (1.0 + torch.erf(x / 1.41421))
    return x.to(x_type)


def bias_gelu(bias, y):
    y_type = y.dtype
    x = bias.float() + y.float()
    x = x * 0.5 * (1.0 + torch.erf(x / 1.41421))
    return x.to(y_type)


def bias_tanh(bias, y):
    y_type = y.dtype
    x = bias.float() + y.float()
    x = torch.tanh(x)
    return x.to(y_type)


def gelu(x):
    """Implementation of the gelu activation function.
        For information: OpenAI GPT's gelu is slightly different (and gives slightly different results):
        0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
        Also see https://arxiv.org/abs/1606.08415
    """
    return f_gelu(x)


def swish(x):
    return x * torch.sigmoid(x)


ACT2FN = {"gelu": gelu, "relu": torch.nn.functional.relu, "swish": swish}


class GPUTimer:

    def __init__(self):
        super().__init__()
        self.start = get_accelerator().Event()  # noqa: F821
        self.stop = get_accelerator().Event()  # noqa: F821

    def record(self):
        self.start.record()

    def elapsed(self):
        self.stop.record()
        self.stop.synchronize()
        return self.start.elapsed_time(self.stop) / 1000.0


class LinearActivation(Module):
    r"""Fused Linear and activation Module.
    """
    __constants__ = ['bias']

    def __init__(self, in_features, out_features, weights, biases, act='gelu', bias=True):
        super(LinearActivation, self).__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.fused_gelu = False
        self.fused_tanh = False
        if isinstance(act, str):
            if bias and act == 'gelu':
                self.fused_gelu = True
            elif bias and act == 'tanh':
                self.fused_tanh = True
            else:
                self.act_fn = ACT2FN[act]
        else:
            self.act_fn = act
        #self.weight = Parameter(torch.Tensor(out_features, in_features))
        self.weight = weights[5]
        self.bias = biases[5]
        #if bias:
        #    self.bias = Parameter(torch.Tensor(out_features))
        #else:
        #    self.register_parameter('bias', None)
        #self.reset_parameters()

    def reset_parameters(self):
        init.kaiming_uniform_(self.weight, a=math.sqrt(5))
        if self.bias is not None:
            fan_in, _ = init._calculate_fan_in_and_fan_out(self.weight)
            bound = 1 / math.sqrt(fan_in)
            init.uniform_(self.bias, -bound, bound)

    def forward(self, input):
        if self.fused_gelu:
            #timing = []
            #t1 = GPUTimer()
            #t1.record()
            y = F.linear(input, self.weight, None)
            #timing.append(t1.elapsed())
            #t1.record()
            bg = bias_gelu(self.bias, y)
            #timing.append(t1.elapsed())
            return bg
        elif self.fused_tanh:
            return bias_tanh(self.bias, F.linear(input, self.weight, None))
        else:
            return self.act_fn(F.linear(input, self.weight, self.bias))

    def extra_repr(self):
        return 'in_features={}, out_features={}, bias={}'.format(self.in_features, self.out_features, self.bias
                                                                 is not None)


class BertConfig(object):
    """Configuration class to store the configuration of a `BertModel`.
    """

    def __init__(self,
                 vocab_size_or_config_json_file,
                 hidden_size=768,
                 num_hidden_layers=12,
                 num_attention_heads=12,
                 intermediate_size=3072,
                 batch_size=8,
                 hidden_act="gelu",
                 hidden_dropout_prob=0.1,
                 attention_probs_dropout_prob=0.1,
                 max_position_embeddings=512,
                 type_vocab_size=2,
                 initializer_range=0.02,
                 fp16=False):
        """Constructs BertConfig.

        Args:
            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `BertModel`.
            hidden_size: Size of the encoder layers and the pooler layer.
            num_hidden_layers: Number of hidden layers in the Transformer encoder.
            num_attention_heads: Number of attention heads for each attention layer in
                the Transformer encoder.
            intermediate_size: The size of the "intermediate" (i.e., feed-forward)
                layer in the Transformer encoder.
            hidden_act: The non-linear activation function (function or string) in the
                encoder and pooler. If string, "gelu", "relu" and "swish" are supported.
            hidden_dropout_prob: The dropout probability for all fully connected
                layers in the embeddings, encoder, and pooler.
            attention_probs_dropout_prob: The dropout ratio for the attention
                probabilities.
            max_position_embeddings: The maximum sequence length that this model might
                ever be used with. Typically set this to something large just in case
                (e.g., 512 or 1024 or 2048).
            type_vocab_size: The vocabulary size of the `token_type_ids` passed into
                `BertModel`.
            initializer_range: The sttdev of the truncated_normal_initializer for
                initializing all weight matrices.
        """
        if isinstance(vocab_size_or_config_json_file, str):
            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
                json_config = json.loads(reader.read())
            for key, value in json_config.items():
                self.__dict__[key] = value
        elif isinstance(vocab_size_or_config_json_file, int):
            self.vocab_size = vocab_size_or_config_json_file
            self.hidden_size = hidden_size
            self.num_hidden_layers = num_hidden_layers
            self.num_attention_heads = num_attention_heads
            self.batch_size = batch_size
            self.hidden_act = hidden_act
            self.intermediate_size = intermediate_size
            self.hidden_dropout_prob = hidden_dropout_prob
            self.attention_probs_dropout_prob = attention_probs_dropout_prob
            self.max_position_embeddings = max_position_embeddings
            self.type_vocab_size = type_vocab_size
            self.initializer_range = initializer_range
            self.fp16 = fp16
        else:
            raise ValueError("First argument must be either a vocabulary size (int)"
                             "or the path to a pretrained model config file (str)")

    @classmethod
    def from_dict(cls, json_object):
        """Constructs a `BertConfig` from a Python dictionary of parameters."""
        config = BertConfig(vocab_size_or_config_json_file=-1)
        for key, value in json_object.items():
            config.__dict__[key] = value
        return config

    @classmethod
    def from_json_file(cls, json_file):
        """Constructs a `BertConfig` from a json file of parameters."""
        with open(json_file, "r", encoding='utf-8') as reader:
            text = reader.read()
        return cls.from_dict(json.loads(text))

    def __repr__(self):
        return str(self.to_json_string())

    def to_dict(self):
        """Serializes this instance to a Python dictionary."""
        output = copy.deepcopy(self.__dict__)
        return output

    def to_json_string(self):
        """Serializes this instance to a JSON string."""
        return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"


try:
    import apex
    #apex.amp.register_half_function(apex.normalization.fused_layer_norm, 'FusedLayerNorm')
    import apex.normalization
    #apex.amp.register_float_function(apex.normalization.FusedLayerNorm, 'forward')
    BertLayerNorm = apex.normalization.FusedLayerNorm
except ImportError:
    print("Better speed can be achieved with apex installed from https://www.github.com/nvidia/apex.")

    class BertLayerNorm(nn.Module):

        def __init__(self, hidden_size, eps=1e-12):
            """Construct a layernorm module in the TF style (epsilon inside the square root).
            """
            super(BertLayerNorm, self).__init__()
            self.weight = nn.Parameter(torch.ones(hidden_size))
            self.bias = nn.Parameter(torch.zeros(hidden_size))
            self.variance_epsilon = eps

        def forward(self, x):
            pdtype = x.dtype
            x = x.float()
            u = x.mean(-1, keepdim=True)
            s = (x - u).pow(2).mean(-1, keepdim=True)
            x = (x - u) / torch.sqrt(s + self.variance_epsilon)
            return self.weight * x.to(pdtype) + self.bias

        #def forward(self, x):
        #    u = x.mean(-1, keepdim=True)
        #    s = (x - u).pow(2).mean(-1, keepdim=True)
        #    x = (x - u) / torch.sqrt(s + self.variance_epsilon)
        #    return self.weight * x + self.bias


class BertEmbeddings(nn.Module):
    """Construct the embeddings from word, position and token_type embeddings.
    """

    def __init__(self, config):
        super(BertEmbeddings, self).__init__()
        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
        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 = BertLayerNorm(config.hidden_size, eps=1e-12)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, input_ids, token_type_ids=None):
        seq_length = input_ids.size(1)
        position_ids = torch.arange(seq_length, dtype=torch.long, device=input_ids.device)
        position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
        if token_type_ids is None:
            token_type_ids = torch.zeros_like(input_ids)

        words_embeddings = self.word_embeddings(input_ids)
        position_embeddings = self.position_embeddings(position_ids)
        token_type_embeddings = self.token_type_embeddings(token_type_ids)

        embeddings = words_embeddings + position_embeddings + token_type_embeddings
        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings


class BertSelfAttention(nn.Module):

    def __init__(self, i, config, weights, biases):
        super(BertSelfAttention, self).__init__()
        if config.hidden_size % config.num_attention_heads != 0:
            raise ValueError("The hidden size (%d) is not a multiple of the number of attention "
                             "heads (%d)" % (config.hidden_size, config.num_attention_heads))
        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(config.hidden_size, self.all_head_size)
        self.query.weight = weights[0]
        self.query.bias = biases[0]
        self.key = nn.Linear(config.hidden_size, self.all_head_size)
        self.key.weight = weights[1]
        self.key.bias = biases[1]
        self.value = nn.Linear(config.hidden_size, self.all_head_size)
        self.value.weight = weights[2]
        self.value.bias = biases[2]

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
        self.softmax = nn.Softmax(dim=-1)
        #self.softmax_config = DeepSpeedSoftmaxConfig()
        #self.softmax_config.batch_size = config.batch_size
        #self.softmax_config.max_seq_length = config.max_position_embeddings
        #self.softmax_config.hidden_size = config.hidden_size
        #self.softmax_config.heads = config.num_attention_heads
        #self.softmax_config.softmax_id = i
        #self.softmax_config.fp16 = config.fp16
        #self.softmax_config.prob_drop_out = 0.0
        #self.softmax = DeepSpeedSoftmax(i, self.softmax_config)

    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
        x = x.view(*new_x_shape)
        return x.permute(0, 2, 1, 3)

    def transpose_key_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, 3, 1)

    def forward(self, hidden_states, attention_mask, grads=None):
        #timing = []
        #t1 = GPUTimer()
        #t1.record()
        mixed_query_layer = self.query(hidden_states)

        #timing.append(t1.elapsed())
        #print("Query elapsed: %s" % (time.clock() - start))
        #t1.record()
        mixed_key_layer = self.key(hidden_states)

        #timing.append(t1.elapsed())
        #print("Key elapsed: %s" % (time.clock() - start))
        #t1.record()
        mixed_value_layer = self.value(hidden_states)
        #timing.append(t1.elapsed())
        #print("Value elapsed: %s" % (time.clock() - start))

        #t1.record()
        query_layer = self.transpose_for_scores(mixed_query_layer)
        # print(query_layer)
        #timing.append(t1.elapsed())
        #print("Query-Transform elapsed: %s" % (time.clock() - start))
        #t1.record()
        key_layer = self.transpose_key_for_scores(mixed_key_layer)
        # print(key_layer)
        #timing.append(t1.elapsed())
        #print("Key-Transform elapsed: %s" % (time.clock() - start))
        #t1.record()
        value_layer = self.transpose_for_scores(mixed_value_layer)
        #print(value_layer)
        #timing.append(t1.elapsed())
        #print("Value-Transform elapsed: %s" % (time.clock() - start))

        # Take the dot product between "query" and "key" to get the raw attention scores.
        #t1.record()
        #print(query_layer.shape)
        #print(key_layer.shape)
        attention_scores = torch.matmul(query_layer, key_layer)
        #print(attention_scores.shape)
        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        #print("Pytorch: ", attention_scores)
        #timing.append(t1.elapsed())
        #print("Attention-Score elapsed: %s" % (time.clock() - start))
        # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
        #t1.record()

        # context_layer = self.softmax(query_layer, key_layer, value_layer, attention_mask)
        #print("context shape is :", context_layer.shape)
        #print("Cuda-ext:, ", attention_scores1)
        # Normalize the attention scores to probabilities.
        ####attention_probs = self.softmax(attention_scores)
        #timing.append(t1.elapsed())
        #print("Softmax elapsed: %s" % (time.clock() - start))
        #t1 = GPUTimer()
        #t1.record()
        attention_scores = attention_scores + attention_mask
        attention_probs = self.softmax(attention_scores)
        #attention_scores = self.softmax(attention_scores, attention_mask)
        #print("Softmax elapse {0:8.2f} ms", t1.elapsed() * 1000)

        # 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)

        #t1.record()
        context_layer = torch.matmul(attention_probs, value_layer)
        #timing.append(t1.elapsed())
        #print("Context elapsed: %s" % (time.clock() - start))
        #t1.record()
        #context_layer1 = context_layer.permute(
        #                0, 1, 3, 2, 4).contiguous()
        #if grads is not None:
        # context_layer.register_hook(lambda x, self = self : grads.append([x, "Context"]))
        context_layer1 = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer1.size()[:-2] + (self.all_head_size, )
        context_layer1 = context_layer1.view(*new_context_layer_shape)
        #timing.append(t1.elapsed())
        #print("Context-Transform elapsed: %s" % (time.clock() - start))

        if grads is not None:
            query_layer.register_hook(lambda x, self=self: grads.append([x, "Query"]))
            key_layer.register_hook(lambda x, self=self: grads.append([x, "Key"]))
            value_layer.register_hook(lambda x, self=self: grads.append([x, "Value"]))

        return context_layer1


class BertSelfOutput(nn.Module):

    def __init__(self, config, weights, biases):
        super(BertSelfOutput, self).__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.dense.weight = weights[3]
        self.dense.bias = biases[3]
        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=1e-12)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):
        #timing = []
        #t1 = GPUTimer()
        #t1.record()
        hidden_states = self.dense(hidden_states)
        #timing.append(t1.elapsed())
        #print("Attention Output elapsed: %s" % (time.clock() - start))
        hidden_states = self.dropout(hidden_states)
        #t1.record()
        #hidden_states = self.LayerNorm(hidden_states + input_tensor)
        #timing.append(t1.elapsed())
        #print("LayerNorm elapsed: %s" % (time.clock() - start))
        return hidden_states

    def get_w(self):
        return self.dense.weight


class BertAttention(nn.Module):

    def __init__(self, i, config, weights, biases):
        super(BertAttention, self).__init__()
        self.self = BertSelfAttention(i, config, weights, biases)
        self.output = BertSelfOutput(config, weights, biases)

    def forward(self, input_tensor, attention_mask):
        self_output = self.self(input_tensor, attention_mask)
        attention_output = self.output(self_output, input_tensor)
        return attention_output

    def get_w(self):
        return self.output.get_w()


class BertIntermediate(nn.Module):

    def __init__(self, config, weights, biases):
        super(BertIntermediate, self).__init__()
        self.dense_act = LinearActivation(config.hidden_size,
                                          config.intermediate_size,
                                          weights,
                                          biases,
                                          act=config.hidden_act)

    def forward(self, hidden_states):
        hidden_states = self.dense_act(hidden_states)
        return hidden_states


class BertOutput(nn.Module):

    def __init__(self, config, weights, biases):
        super(BertOutput, self).__init__()
        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
        self.dense.weight = weights[6]
        self.dense.bias = biases[6]
        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=1e-12)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):
        #timing = []
        #t1 = GPUTimer()
        #t1.record()
        #print (hidden_states)
        #print (self.dense.weight)
        hidden_states = self.dense(hidden_states)
        #timing.append(t1.elapsed())
        #print("FF2 elapsed: %s" % (time.clock() - start))
        hidden_states = self.dropout(hidden_states)
        #t1.record()
        #hidden_states = self.LayerNorm(hidden_states + input_tensor)
        #timing.append(t1.elapsed())
        #print("LayerNorm elapsed: %s" % (time.clock() - start))
        return hidden_states


class BertLayer(nn.Module):

    def __init__(self, i, config, weights, biases):
        super(BertLayer, self).__init__()
        self.attention = BertAttention(i, config, weights, biases)
        self.PreAttentionLayerNorm = BertLayerNorm(config.hidden_size, eps=1e-12)
        self.PostAttentionLayerNorm = BertLayerNorm(config.hidden_size, eps=1e-12)
        self.intermediate = BertIntermediate(config, weights, biases)
        self.output = BertOutput(config, weights, biases)
        self.weight = weights
        self.biases = biases

    def forward(self, hidden_states, attention_mask, grads, collect_all_grads=False):
        input_layer_norm = self.PreAttentionLayerNorm(hidden_states)
        attention_output = self.attention(input_layer_norm, attention_mask)
        #print ("hidden shape is :", hidden_states.shape)
        intermediate_input = hidden_states + attention_output

        intermediate_layer_norm = self.PostAttentionLayerNorm(intermediate_input)
        intermediate_output = self.intermediate(intermediate_layer_norm)
        layer_output = self.output(intermediate_output, attention_output)

        #attention_output = self.attention(hidden_states, attention_mask)
        #intermediate_output = self.intermediate(attention_output)
        #layer_output = self.output(intermediate_output, attention_output)

        if collect_all_grads:
            # self.weight[0].register_hook(lambda x, self=self: grads.append([x,"Q_W"]))
            # self.biases[0].register_hook(lambda x, self=self: grads.append([x,"Q_B"]))
            # self.weight[1].register_hook(lambda x, self=self: grads.append([x,"K_W"]))
            # self.biases[1].register_hook(lambda x, self=self: grads.append([x,"K_B"]))
            self.weight[2].register_hook(lambda x, self=self: grads.append([x, "V_W"]))
            self.biases[2].register_hook(lambda x, self=self: grads.append([x, "V_B"]))
            self.weight[3].register_hook(lambda x, self=self: grads.append([x, "O_W"]))
            self.biases[3].register_hook(lambda x, self=self: grads.append([x, "O_B"]))
            self.PostAttentionLayerNorm.weight.register_hook(lambda x, self=self: grads.append([x, "N2_W"]))
            self.PostAttentionLayerNorm.bias.register_hook(lambda x, self=self: grads.append([x, "N2_B"]))
            self.weight[5].register_hook(lambda x, self=self: grads.append([x, "int_W"]))
            self.biases[5].register_hook(lambda x, self=self: grads.append([x, "int_B"]))
            self.weight[6].register_hook(lambda x, self=self: grads.append([x, "out_W"]))
            self.biases[6].register_hook(lambda x, self=self: grads.append([x, "out_B"]))
            self.PreAttentionLayerNorm.weight.register_hook(lambda x, self=self: grads.append([x, "norm_W"]))
            self.PreAttentionLayerNorm.bias.register_hook(lambda x, self=self: grads.append([x, "norm_B"]))

        return layer_output + intermediate_input

    def get_w(self):
        return self.attention.get_w()


class BertEncoder(nn.Module):

    def __init__(self, config, weights, biases):
        super(BertEncoder, self).__init__()
        #layer = BertLayer(config, weights, biases)
        self.FinalLayerNorm = BertLayerNorm(config.hidden_size, eps=1e-12)

        self.layer = nn.ModuleList(
            [copy.deepcopy(BertLayer(i, config, weights, biases)) for i in range(config.num_hidden_layers)])
        self.grads = []
        self.graph = []

    def get_grads(self):
        return self.grads

    # def forward(self, hidden_states, attention_mask, output_all_encoded_layers=True):
    #     all_encoder_layers = []
    #     for layer_module in self.layer:
    #         hidden_states = layer_module(hidden_states, attention_mask)
    #         if output_all_encoded_layers:
    #             all_encoder_layers.append(hidden_states)
    #     if not output_all_encoded_layers:
    #         all_encoder_layers.append(hidden_states)
    #     return all_encoder_layers

    def get_modules(self, big_node, input):
        for mdl in big_node.named_children():
            self.graph.append(mdl)
            self.get_modules(self, mdl, input)

    def forward(self, hidden_states, attention_mask, output_all_encoded_layers=True, checkpoint_activations=False):
        all_encoder_layers = []

        def custom(start, end):

            def custom_forward(*inputs):
                layers = self.layer[start:end]
                x_ = inputs[0]
                for layer in layers:
                    x_ = layer(x_, inputs[1])
                return x_

            return custom_forward

        if checkpoint_activations:
            l = 0
            num_layers = len(self.layer)
            chunk_length = math.ceil(math.sqrt(num_layers))
            while l < num_layers:
                hidden_states = checkpoint.checkpoint(custom(l, l + chunk_length), hidden_states, attention_mask * 1)
                l += chunk_length
            # decoder layers
        else:
            for i, layer_module in enumerate(self.layer):
                hidden_states = layer_module(hidden_states, attention_mask, self.grads, collect_all_grads=True)
                hidden_states.register_hook(lambda x, i=i, self=self: self.grads.append([x, "hidden_state"]))
                #print("pytorch weight is: ", layer_module.get_w())

                if output_all_encoded_layers:
                    all_encoder_layers.append((hidden_states))

        if not output_all_encoded_layers or checkpoint_activations:
            hidden_states = self.FinalLayerNorm(hidden_states)
            all_encoder_layers.append((hidden_states))
        return all_encoder_layers


#class BertEncoder(nn.Module):
#    def __init__(self, config):
#        super(BertEncoder, self).__init__()
#        layer = BertLayer(config)
#        self.layer = nn.ModuleList([copy.deepcopy(layer) for _ in range(config.num_hidden_layers)])
#
#    def forward(self, hidden_states, attention_mask, output_all_encoded_layers=True):
#        all_encoder_layers = []
#        for layer_module in self.layer:
#            hidden_states = layer_module(hidden_states, attention_mask)
#            if output_all_encoded_layers:
#                all_encoder_layers.append(hidden_states)
#        if not output_all_encoded_layers:
#            all_encoder_layers.append(hidden_states)
#        return all_encoder_layers


class BertPooler(nn.Module):

    def __init__(self, config):
        super(BertPooler, self).__init__()
        self.dense_act = LinearActivation(config.hidden_size, config.hidden_size, act="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_act(first_token_tensor)
        return pooled_output


class BertPredictionHeadTransform(nn.Module):

    def __init__(self, config):
        super(BertPredictionHeadTransform, self).__init__()
        self.dense_act = LinearActivation(config.hidden_size, config.hidden_size, act=config.hidden_act)
        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=1e-12)

    def forward(self, hidden_states):
        hidden_states = self.dense_act(hidden_states)
        hidden_states = self.LayerNorm(hidden_states)
        return hidden_states


class BertLMPredictionHead(nn.Module):

    def __init__(self, config, bert_model_embedding_weights):
        super(BertLMPredictionHead, self).__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(bert_model_embedding_weights.size(1),
                                 bert_model_embedding_weights.size(0),
                                 bias=False)
        self.decoder.weight = bert_model_embedding_weights
        self.bias = nn.Parameter(torch.zeros(bert_model_embedding_weights.size(0)))

    def forward(self, hidden_states):
        hidden_states = self.transform(hidden_states)
        get_accelerator().range_push("decoder input.size() = {}, weight.size() = {}".format(
            hidden_states.size(), self.decoder.weight.size()))
        hidden_states = self.decoder(hidden_states) + self.bias
        get_accelerator().range_pop()
        return hidden_states


class BertOnlyMLMHead(nn.Module):

    def __init__(self, config, bert_model_embedding_weights):
        super(BertOnlyMLMHead, self).__init__()
        self.predictions = BertLMPredictionHead(config, bert_model_embedding_weights)

    def forward(self, sequence_output):
        prediction_scores = self.predictions(sequence_output)
        return prediction_scores


class BertOnlyNSPHead(nn.Module):

    def __init__(self, config):
        super(BertOnlyNSPHead, self).__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, bert_model_embedding_weights):
        super(BertPreTrainingHeads, self).__init__()
        self.predictions = BertLMPredictionHead(config, bert_model_embedding_weights)
        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(nn.Module):
    """ An abstract class to handle weights initialization and
        a simple interface for downloading and loading pretrained models.
    """

    def __init__(self, config, *inputs, **kwargs):
        super(BertPreTrainedModel, self).__init__()
        if not isinstance(config, BertConfig):
            raise ValueError("Parameter config in `{}(config)` should be an instance of class `BertConfig`. "
                             "To create a model from a Google pretrained model use "
                             "`model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`".format(
                                 self.__class__.__name__, self.__class__.__name__))
        self.config = config

    def init_bert_weights(self, module):
        """ Initialize the weights.
        """
        if isinstance(module, (nn.Linear, nn.Embedding)):
            # 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)
        elif isinstance(module, BertLayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)
        if isinstance(module, nn.Linear) and module.bias is not None:
            module.bias.data.zero_()

    @classmethod
    def from_pretrained(cls,
                        pretrained_model_name_or_path,
                        state_dict=None,
                        cache_dir=None,
                        from_tf=False,
                        *inputs,
                        **kwargs):
        """
        Instantiate a BertPreTrainedModel from a pre-trained model file or a pytorch state dict.
        Download and cache the pre-trained model file if needed.

        Params:
            pretrained_model_name_or_path: either:
                - a str with the name of a pre-trained model to load selected in the list of:
                    . `bert-base-uncased`
                    . `bert-large-uncased`
                    . `bert-base-cased`
                    . `bert-large-cased`
                    . `bert-base-multilingual-uncased`
                    . `bert-base-multilingual-cased`
                    . `bert-base-chinese`
                - a path or url to a pretrained model archive containing:
                    . `bert_config.json` a configuration file for the model
                    . `pytorch_model.bin` a PyTorch dump of a BertForPreTraining instance
                - a path or url to a pretrained model archive containing:
                    . `bert_config.json` a configuration file for the model
                    . `model.chkpt` a TensorFlow checkpoint
            from_tf: should we load the weights from a locally saved TensorFlow checkpoint
            cache_dir: an optional path to a folder in which the pre-trained models will be cached.
            state_dict: an optional state dictionary (collections.OrderedDict object) to use instead of Google pre-trained models
            *inputs, **kwargs: additional input for the specific Bert class
                (ex: num_labels for BertForSequenceClassification)
        """
        if pretrained_model_name_or_path in PRETRAINED_MODEL_ARCHIVE_MAP:
            archive_file = PRETRAINED_MODEL_ARCHIVE_MAP[pretrained_model_name_or_path]
        else:
            archive_file = pretrained_model_name_or_path
        if resolved_archive_file == archive_file:  # noqa: F821
            logger.info("loading archive file {}".format(archive_file))
        else:
            logger.info("loading archive file {} from cache at {}".format(archive_file,
                                                                          resolved_archive_file))  # noqa: F821
        tempdir = None
        if os.path.isdir(resolved_archive_file) or from_tf:  # noqa: F821
            serialization_dir = resolved_archive_file  # noqa: F821
        else:
            # Extract archive to temp dir
            tempdir = tempfile.mkdtemp()
            logger.info("extracting archive file {} to temp dir {}".format(
                resolved_archive_file,  # noqa: F821
                tempdir))
            with tarfile.open(resolved_archive_file, 'r:gz') as archive:  # noqa: F821
                archive.extractall(tempdir)
            serialization_dir = tempdir
        # Load config
        config_file = os.path.join(serialization_dir, CONFIG_NAME)
        config = BertConfig.from_json_file(config_file)
        logger.info("Model config {}".format(config))
        # Instantiate model.
        model = cls(config, *inputs, **kwargs)
        if state_dict is None and not from_tf:
            weights_path = os.path.join(serialization_dir, WEIGHTS_NAME)
            state_dict = torch.load(weights_path, map_location='cpu' if not get_accelerator().is_available() else None)
        if tempdir:
            # Clean up temp dir
            shutil.rmtree(tempdir)
        if from_tf:
            # Directly load from a TensorFlow checkpoint
            weights_path = os.path.join(serialization_dir, TF_WEIGHTS_NAME)
            return load_tf_weights_in_bert(model, weights_path)
        # Load from a PyTorch state_dict
        old_keys = []
        new_keys = []
        for key in state_dict.keys():
            new_key = None
            if 'gamma' in key:
                new_key = key.replace('gamma', 'weight')
            if 'beta' in key:
                new_key = key.replace('beta', 'bias')
            if new_key:
                old_keys.append(key)
                new_keys.append(new_key)
        for old_key, new_key in zip(old_keys, new_keys):
            state_dict[new_key] = state_dict.pop(old_key)

        missing_keys = []
        unexpected_keys = []
        error_msgs = []
        # copy state_dict so _load_from_state_dict can modify it
        metadata = getattr(state_dict, '_metadata', None)
        state_dict = state_dict.copy()
        if metadata is not None:
            state_dict._metadata = metadata

        def load(module, prefix=''):
            local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
            module._load_from_state_dict(state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys,
                                         error_msgs)
            for name, child in module._modules.items():
                if child is not None:
                    load(child, prefix + name + '.')

        start_prefix = ''
        if not hasattr(model, 'bert') and any(s.startswith('bert.') for s in state_dict.keys()):
            start_prefix = 'bert.'
        load(model, prefix=start_prefix)
        if len(missing_keys) > 0:
            logger.info("Weights of {} not initialized from pretrained model: {}".format(
                model.__class__.__name__, missing_keys))
        if len(unexpected_keys) > 0:
            logger.info("Weights from pretrained model not used in {}: {}".format(model.__class__.__name__,
                                                                                  unexpected_keys))
        if len(error_msgs) > 0:
            raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
                model.__class__.__name__, "\n\t".join(error_msgs)))
        return model


class BertModel(BertPreTrainedModel):
    """BERT model ("Bidirectional Embedding Representations from a Transformer").

    Params:
        config: a BertConfig class instance with the configuration to build a new model

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary(see the tokens preprocessing logic in the scripts
            `extract_features.py`, `run_classifier.py` and `run_squad.py`)
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `output_all_encoded_layers`: boolean which controls the content of the `encoded_layers` output as described below. Default: `True`.

    Outputs: Tuple of (encoded_layers, pooled_output)
        `encoded_layers`: controlled by `output_all_encoded_layers` argument:
            - `output_all_encoded_layers=True`: outputs a list of the full sequences of encoded-hidden-states at the end
                of each attention block (i.e. 12 full sequences for BERT-base, 24 for BERT-large), each
                encoded-hidden-state is a torch.FloatTensor of size [batch_size, sequence_length, hidden_size],
            - `output_all_encoded_layers=False`: outputs only the full sequence of hidden-states corresponding
                to the last attention block of shape [batch_size, sequence_length, hidden_size],
        `pooled_output`: a torch.FloatTensor of size [batch_size, hidden_size] which is the output of a
            classifier pretrained on top of the hidden state associated to the first character of the
            input (`CLS`) to train on the Next-Sentence task (see BERT's paper).

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
    input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
    token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])

    config = modeling.BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)

    model = modeling.BertModel(config=config)
    all_encoder_layers, pooled_output = model(input_ids, token_type_ids, input_mask)
    ```
    """

    def __init__(self, config):
        super(BertModel, self).__init__(config)
        self.embeddings = BertEmbeddings(config)
        self.encoder = BertEncoder(config)
        self.pooler = BertPooler(config)
        self.apply(self.init_bert_weights)

    def forward(self,
                input_ids,
                token_type_ids=None,
                attention_mask=None,
                output_all_encoded_layers=True,
                checkpoint_activations=False):
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids)
        if token_type_ids is None:
            token_type_ids = torch.zeros_like(input_ids)

        # We create a 3D attention mask from a 2D tensor mask.
        # Sizes are [batch_size, 1, 1, to_seq_length]
        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
        # this attention mask is more simple than the triangular masking of causal attention
        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)

        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
        # masked positions, this operation will create a tensor which is 0.0 for
        # positions we want to attend and -10000.0 for masked positions.
        # Since we are adding it to the raw scores before the softmax, this is
        # effectively the same as removing these entirely.
        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0

        embedding_output = self.embeddings(input_ids, token_type_ids)
        encoded_layers = self.encoder(embedding_output,
                                      extended_attention_mask,
                                      output_all_encoded_layers=output_all_encoded_layers,
                                      checkpoint_activations=checkpoint_activations)
        sequence_output = encoded_layers[-1]
        pooled_output = self.pooler(sequence_output)
        if not output_all_encoded_layers:
            encoded_layers = encoded_layers[-1]
        return encoded_layers, pooled_output


class BertForPreTraining(BertPreTrainedModel):
    """BERT model with pre-training heads.
    This module comprises the BERT model followed by the two pre-training heads:
        - the masked language modeling head, and
        - the next sentence classification head.

    Params:
        config: a BertConfig class instance with the configuration to build a new model.

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary(see the tokens preprocessing logic in the scripts
            `extract_features.py`, `run_classifier.py` and `run_squad.py`)
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `masked_lm_labels`: optional masked language modeling labels: torch.LongTensor of shape [batch_size, sequence_length]
            with indices selected in [-1, 0, ..., vocab_size]. All labels set to -1 are ignored (masked), the loss
            is only computed for the labels set in [0, ..., vocab_size]
        `next_sentence_label`: optional next sentence classification loss: torch.LongTensor of shape [batch_size]
            with indices selected in [0, 1].
            0 => next sentence is the continuation, 1 => next sentence is a random sentence.

    Outputs:
        if `masked_lm_labels` and `next_sentence_label` are not `None`:
            Outputs the total_loss which is the sum of the masked language modeling loss and the next
            sentence classification loss.
        if `masked_lm_labels` or `next_sentence_label` is `None`:
            Outputs a tuple comprising
            - the masked language modeling logits of shape [batch_size, sequence_length, vocab_size], and
            - the next sentence classification logits of shape [batch_size, 2].

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
    input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
    token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])

    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)

    model = BertForPreTraining(config)
    masked_lm_logits_scores, seq_relationship_logits = model(input_ids, token_type_ids, input_mask)
    ```
    """

    def __init__(self, config, args):
        super(BertForPreTraining, self).__init__(config)
        self.summary_writer = None
        if dist.get_rank() == 0:
            self.summary_writer = args.summary_writer
        self.samples_per_step = dist.get_world_size() * args.train_batch_size
        self.sample_count = self.samples_per_step
        self.bert = BertModel(config)
        self.cls = BertPreTrainingHeads(config, self.bert.embeddings.word_embeddings.weight)
        self.apply(self.init_bert_weights)

    def log_summary_writer(self, logs: dict, base='Train'):
        if dist.get_rank() == 0:
            module_name = "Samples"  #self._batch_module_name.get(batch_type, self._get_batch_type_error(batch_type))
            for key, log in logs.items():
                self.summary_writer.add_scalar(f'{base}/{module_name}/{key}', log, self.sample_count)
            self.sample_count += self.samples_per_step

    def forward(self, batch, log=True):
        #input_ids, token_type_ids=None, attention_mask=None, masked_lm_labels=None, next_sentence_label=None, checkpoint_activations=False):
        input_ids = batch[1]
        token_type_ids = batch[3]
        attention_mask = batch[2]
        masked_lm_labels = batch[5]
        next_sentence_label = batch[4]
        checkpoint_activations = False

        sequence_output, pooled_output = self.bert(input_ids,
                                                   token_type_ids,
                                                   attention_mask,
                                                   output_all_encoded_layers=False,
                                                   checkpoint_activations=checkpoint_activations)
        prediction_scores, seq_relationship_score = self.cls(sequence_output, pooled_output)

        if masked_lm_labels is not None and next_sentence_label is not None:
            loss_fct = CrossEntropyLoss(ignore_index=-1)
            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
            next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
            #print("loss is {} {}".format(masked_lm_loss, next_sentence_loss))
            total_loss = masked_lm_loss + next_sentence_loss
            #            if log:
            #                self.log_summary_writer(logs={'train_loss': total_loss.item()})
            return total_loss
        else:
            return prediction_scores, seq_relationship_score


class BertForMaskedLM(BertPreTrainedModel):
    """BERT model with the masked language modeling head.
    This module comprises the BERT model followed by the masked language modeling head.

    Params:
        config: a BertConfig class instance with the configuration to build a new model.

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary(see the tokens preprocessing logic in the scripts
            `extract_features.py`, `run_classifier.py` and `run_squad.py`)
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `masked_lm_labels`: masked language modeling labels: torch.LongTensor of shape [batch_size, sequence_length]
            with indices selected in [-1, 0, ..., vocab_size]. All labels set to -1 are ignored (masked), the loss
            is only computed for the labels set in [0, ..., vocab_size]

    Outputs:
        if `masked_lm_labels` is  not `None`:
            Outputs the masked language modeling loss.
        if `masked_lm_labels` is `None`:
            Outputs the masked language modeling logits of shape [batch_size, sequence_length, vocab_size].

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
    input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
    token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])

    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)

    model = BertForMaskedLM(config)
    masked_lm_logits_scores = model(input_ids, token_type_ids, input_mask)
    ```
    """

    def __init__(self, config):
        super(BertForMaskedLM, self).__init__(config)
        self.bert = BertModel(config)
        self.cls = BertOnlyMLMHead(config, self.bert.embeddings.word_embeddings.weight)
        self.apply(self.init_bert_weights)

    def forward(self,
                input_ids,
                token_type_ids=None,
                attention_mask=None,
                masked_lm_labels=None,
                checkpoint_activations=False):
        sequence_output, _ = self.bert(input_ids, token_type_ids, attention_mask, output_all_encoded_layers=False)
        prediction_scores = self.cls(sequence_output)

        if masked_lm_labels is not None:
            loss_fct = CrossEntropyLoss(ignore_index=-1)
            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
            return masked_lm_loss
        else:
            return prediction_scores


class BertForNextSentencePrediction(BertPreTrainedModel):
    """BERT model with next sentence prediction head.
    This module comprises the BERT model followed by the next sentence classification head.

    Params:
        config: a BertConfig class instance with the configuration to build a new model.

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary(see the tokens preprocessing logic in the scripts
            `extract_features.py`, `run_classifier.py` and `run_squad.py`)
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `next_sentence_label`: next sentence classification loss: torch.LongTensor of shape [batch_size]
            with indices selected in [0, 1].
            0 => next sentence is the continuation, 1 => next sentence is a random sentence.

    Outputs:
        if `next_sentence_label` is not `None`:
            Outputs the total_loss which is the sum of the masked language modeling loss and the next
            sentence classification loss.
        if `next_sentence_label` is `None`:
            Outputs the next sentence classification logits of shape [batch_size, 2].

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
    input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
    token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])

    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)

    model = BertForNextSentencePrediction(config)
    seq_relationship_logits = model(input_ids, token_type_ids, input_mask)
    ```
    """

    def __init__(self, config):
        super(BertForNextSentencePrediction, self).__init__(config)
        self.bert = BertModel(config)
        self.cls = BertOnlyNSPHead(config)
        self.apply(self.init_bert_weights)

    def forward(self,
                input_ids,
                token_type_ids=None,
                attention_mask=None,
                next_sentence_label=None,
                checkpoint_activations=False):
        _, pooled_output = self.bert(input_ids, token_type_ids, attention_mask, output_all_encoded_layers=False)
        seq_relationship_score = self.cls(pooled_output)

        if next_sentence_label is not None:
            loss_fct = CrossEntropyLoss(ignore_index=-1)
            next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
            return next_sentence_loss
        else:
            return seq_relationship_score


class BertForSequenceClassification(BertPreTrainedModel):
    """BERT model for classification.
    This module is composed of the BERT model with a linear layer on top of
    the pooled output.

    Params:
        `config`: a BertConfig class instance with the configuration to build a new model.
        `num_labels`: the number of classes for the classifier. Default = 2.

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary(see the tokens preprocessing logic in the scripts
            `extract_features.py`, `run_classifier.py` and `run_squad.py`)
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `labels`: labels for the classification output: torch.LongTensor of shape [batch_size]
            with indices selected in [0, ..., num_labels].

    Outputs:
        if `labels` is not `None`:
            Outputs the CrossEntropy classification loss of the output with the labels.
        if `labels` is `None`:
            Outputs the classification logits of shape [batch_size, num_labels].

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
    input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
    token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])

    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)

    num_labels = 2

    model = BertForSequenceClassification(config, num_labels)
    logits = model(input_ids, token_type_ids, input_mask)
    ```
    """

    def __init__(self, config, num_labels):
        super(BertForSequenceClassification, self).__init__(config)
        self.num_labels = num_labels
        self.bert = BertModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, num_labels)
        self.apply(self.init_bert_weights)

    def forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None, checkpoint_activations=False):
        _, pooled_output = self.bert(input_ids, token_type_ids, attention_mask, output_all_encoded_layers=False)
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)

        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
            return loss
        else:
            return logits


class BertForMultipleChoice(BertPreTrainedModel):
    """BERT model for multiple choice tasks.
    This module is composed of the BERT model with a linear layer on top of
    the pooled output.

    Params:
        `config`: a BertConfig class instance with the configuration to build a new model.
        `num_choices`: the number of classes for the classifier. Default = 2.

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, num_choices, sequence_length]
            with the word token indices in the vocabulary(see the tokens preprocessing logic in the scripts
            `extract_features.py`, `run_classifier.py` and `run_squad.py`)
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, num_choices, sequence_length]
            with the token types indices selected in [0, 1]. Type 0 corresponds to a `sentence A`
            and type 1 corresponds to a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, num_choices, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `labels`: labels for the classification output: torch.LongTensor of shape [batch_size]
            with indices selected in [0, ..., num_choices].

    Outputs:
        if `labels` is not `None`:
            Outputs the CrossEntropy classification loss of the output with the labels.
        if `labels` is `None`:
            Outputs the classification logits of shape [batch_size, num_labels].

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[[31, 51, 99], [15, 5, 0]], [[12, 16, 42], [14, 28, 57]]])
    input_mask = torch.LongTensor([[[1, 1, 1], [1, 1, 0]],[[1,1,0], [1, 0, 0]]])
    token_type_ids = torch.LongTensor([[[0, 0, 1], [0, 1, 0]],[[0, 1, 1], [0, 0, 1]]])
    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)

    num_choices = 2

    model = BertForMultipleChoice(config, num_choices)
    logits = model(input_ids, token_type_ids, input_mask)
    ```
    """

    def __init__(self, config, num_choices):
        super(BertForMultipleChoice, self).__init__(config)
        self.num_choices = num_choices
        self.bert = BertModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, 1)
        self.apply(self.init_bert_weights)

    def forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None, checkpoint_activations=False):
        flat_input_ids = input_ids.view(-1, input_ids.size(-1))
        flat_token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1))
        flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1))
        _, pooled_output = self.bert(flat_input_ids,
                                     flat_token_type_ids,
                                     flat_attention_mask,
                                     output_all_encoded_layers=False)
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        reshaped_logits = logits.view(-1, self.num_choices)

        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(reshaped_logits, labels)
            return loss
        else:
            return reshaped_logits


class BertForTokenClassification(BertPreTrainedModel):
    """BERT model for token-level classification.
    This module is composed of the BERT model with a linear layer on top of
    the full hidden state of the last layer.

    Params:
        `config`: a BertConfig class instance with the configuration to build a new model.
        `num_labels`: the number of classes for the classifier. Default = 2.

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary(see the tokens preprocessing logic in the scripts
            `extract_features.py`, `run_classifier.py` and `run_squad.py`)
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `labels`: labels for the classification output: torch.LongTensor of shape [batch_size, sequence_length]
            with indices selected in [0, ..., num_labels].

    Outputs:
        if `labels` is not `None`:
            Outputs the CrossEntropy classification loss of the output with the labels.
        if `labels` is `None`:
            Outputs the classification logits of shape [batch_size, sequence_length, num_labels].

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
    input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
    token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])

    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)

    num_labels = 2

    model = BertForTokenClassification(config, num_labels)
    logits = model(input_ids, token_type_ids, input_mask)
    ```
    """

    def __init__(self, config, num_labels):
        super(BertForTokenClassification, self).__init__(config)
        self.num_labels = num_labels
        self.bert = BertModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, num_labels)
        self.apply(self.init_bert_weights)

    def forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None, checkpoint_activations=False):
        sequence_output, _ = self.bert(input_ids, token_type_ids, attention_mask, output_all_encoded_layers=False)
        sequence_output = self.dropout(sequence_output)
        logits = self.classifier(sequence_output)

        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_loss]
                active_labels = labels.view(-1)[active_loss]
                loss = loss_fct(active_logits, active_labels)
            else:
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
            return loss
        else:
            return logits


class BertForQuestionAnswering(BertPreTrainedModel):
    """BERT model for Question Answering (span extraction).
    This module is composed of the BERT model with a linear layer on top of
    the sequence output that computes start_logits and end_logits

    Params:
        `config`: a BertConfig class instance with the configuration to build a new model.

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary(see the tokens preprocessing logic in the scripts
            `extract_features.py`, `run_classifier.py` and `run_squad.py`)
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see BERT paper for more details).
        `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `start_positions`: position of the first token for the labeled span: torch.LongTensor of shape [batch_size].
            Positions are clamped to the length of the sequence and position outside of the sequence are not taken
            into account for computing the loss.
        `end_positions`: position of the last token for the labeled span: torch.LongTensor of shape [batch_size].
            Positions are clamped to the length of the sequence and position outside of the sequence are not taken
            into account for computing the loss.

    Outputs:
        if `start_positions` and `end_positions` are not `None`:
            Outputs the total_loss which is the sum of the CrossEntropy loss for the start and end token positions.
        if `start_positions` or `end_positions` is `None`:
            Outputs a tuple of start_logits, end_logits which are the logits respectively for the start and end
            position tokens of shape [batch_size, sequence_length].

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
    input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
    token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])

    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)

    model = BertForQuestionAnswering(config)
    start_logits, end_logits = model(input_ids, token_type_ids, input_mask)
    ```
    """

    def __init__(self, config):
        super(BertForQuestionAnswering, self).__init__(config)
        self.bert = BertModel(config)
        # TODO check with Google if it's normal there is no dropout on the token classifier of SQuAD in the TF version
        # self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.qa_outputs = nn.Linear(config.hidden_size, 2)
        self.apply(self.init_bert_weights)

    def forward(self,
                input_ids,
                token_type_ids=None,
                attention_mask=None,
                start_positions=None,
                end_positions=None,
                checkpoint_activations=False):
        sequence_output, _ = self.bert(input_ids, token_type_ids, attention_mask, output_all_encoded_layers=False)
        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1)
        end_logits = end_logits.squeeze(-1)

        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.clamp_(0, ignored_index)
            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
            return total_loss
        else:
            return start_logits, end_logits