transformers/pytorch_pretrained_bert/model_utils.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."""

from __future__ import absolute_import, division, print_function, unicode_literals

import logging
import os
import json
import copy

import torch
from torch import nn
from torch.nn import CrossEntropyLoss, MSELoss

from .file_utils import cached_path

logger = logging.getLogger(__name__)

CONFIG_NAME = "config.json"
WEIGHTS_NAME = "pytorch_model.bin"


class PretrainedConfig(object):
    """ An abstract class to handle dowloading a model pretrained config.
    """
    pretrained_config_archive_map = {}

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
        """
        Instantiate a PretrainedConfig from a pre-trained model configuration.

        Params:
            pretrained_model_name_or_path: either:
                - a str with the name of a pre-trained model to load selected in the list of:
                    . `xlnet-large-cased`
                - a path or url to a pretrained model archive containing:
                    . `config.json` a configuration file for the model
            cache_dir: an optional path to a folder in which the pre-trained model configuration will be cached.
        """
        cache_dir = kwargs.get('cache_dir', None)
        kwargs.pop('cache_dir', None)

        if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
            config_file = cls.pretrained_config_archive_map[pretrained_model_name_or_path]
        else:
            config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
        # redirect to the cache, if necessary
        try:
            resolved_config_file = cached_path(config_file, cache_dir=cache_dir)
        except EnvironmentError:
            if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
                logger.error(
                    "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
                        config_file))
            else:
                logger.error(
                    "Model name '{}' was not found in model name list ({}). "
                    "We assumed '{}' was a path or url but couldn't find any file "
                    "associated to this path or url.".format(
                        pretrained_model_name_or_path,
                        ', '.join(cls.pretrained_config_archive_map.keys()),
                        config_file))
            return None
        if resolved_config_file == config_file:
            logger.info("loading configuration file {}".format(config_file))
        else:
            logger.info("loading configuration file {} from cache at {}".format(
                config_file, resolved_config_file))

        # Load config
        config = cls.from_json_file(resolved_config_file)

        # Update config with kwargs if needed
        to_remove = []
        for key, value in kwargs.items():
            if hasattr(config, key):
                setattr(config, key, value)
                to_remove.append(key)
        for key in to_remove:
            kwargs.pop(key, None)

        logger.info("Model config {}".format(config))
        return config

    @classmethod
    def from_dict(cls, json_object):
        """Constructs a `Config` from a Python dictionary of parameters."""
        config = cls(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"

    def to_json_file(self, json_file_path):
        """ Save this instance to a json file."""
        with open(json_file_path, "w", encoding='utf-8') as writer:
            writer.write(self.to_json_string())


def prune_linear_layer(layer, index, dim=0):
    """ Prune a linear layer (a model parameters) to keep only entries in index.
        Return the pruned layer as a new layer with requires_grad=True.
        Used to remove heads.
    """
    index = index.to(layer.weight.device)
    W = layer.weight.index_select(dim, index).clone().detach()
    if layer.bias is not None:
        if dim == 1:
            b = layer.bias.clone().detach()
        else:
            b = layer.bias[index].clone().detach()
    new_size = list(layer.weight.size())
    new_size[dim] = len(index)
    new_layer = nn.Linear(new_size[1], new_size[0], bias=layer.bias is not None).to(layer.weight.device)
    new_layer.weight.requires_grad = False
    new_layer.weight.copy_(W.contiguous())
    new_layer.weight.requires_grad = True
    if layer.bias is not None:
        new_layer.bias.requires_grad = False
        new_layer.bias.copy_(b.contiguous())
        new_layer.bias.requires_grad = True
    return new_layer


class Conv1D(nn.Module):
    """ Conv1D layer as defined by Alec Radford for GPT (and also used in GPT-2)
        Basically works like a Linear layer but the weights are transposed
    """
    def __init__(self, nf, nx):
        super(Conv1D, self).__init__()
        self.nf = nf
        w = torch.empty(nx, nf)
        nn.init.normal_(w, std=0.02)
        self.weight = nn.Parameter(w)
        self.bias = nn.Parameter(torch.zeros(nf))

    def forward(self, x):
        size_out = x.size()[:-1] + (self.nf,)
        x = torch.addmm(self.bias, x.view(-1, x.size(-1)), self.weight)
        x = x.view(*size_out)
        return x


def prune_conv1d_layer(layer, index, dim=1):
    """ Prune a Conv1D layer (a model parameters) to keep only entries in index.
        A Conv1D work as a Linear layer (see e.g. BERT) but the weights are transposed.
        Return the pruned layer as a new layer with requires_grad=True.
        Used to remove heads.
    """
    index = index.to(layer.weight.device)
    W = layer.weight.index_select(dim, index).clone().detach()
    if dim == 0:
        b = layer.bias.clone().detach()
    else:
        b = layer.bias[index].clone().detach()
    new_size = list(layer.weight.size())
    new_size[dim] = len(index)
    new_layer = Conv1D(new_size[1], new_size[0]).to(layer.weight.device)
    new_layer.weight.requires_grad = False
    new_layer.weight.copy_(W.contiguous())
    new_layer.weight.requires_grad = True
    new_layer.bias.requires_grad = False
    new_layer.bias.copy_(b.contiguous())
    new_layer.bias.requires_grad = True
    return new_layer