PyTorch - Deep Neural Network - Natural Language Processing

Version 1.1 by KzXuan

Contains CNN, RNN and Transformer layers and models implemented by pytorch for classification and sequence labeling tasks in NLP.

• Newly designed modules.
• Reduce usage complexity.
• Use mask as the sequence length identifier.
• Multi-GPU parallel for grid search.

Dependecies

python 3.5+ & pytorch 1.2.0+

Install

> pip install dnnnlp

API Document (In Chinese)

• Layer - layer.py
• Model - model.py
• Execution - exec.py
• Utility - utils.py

Hyperparameters

Usage

# import our modules
from dnnnlp.model import RNNModel
from dnnnlp.exec import default_args, Classify

# load the embedding matrix
emb_mat = np.array((-1, 300))
# load the train data
train_x = np.array((800, 50))
train_y = np.array((800,))
train_mask = np.array((800, 50))
# load the test data
test_x = np.array((200, 50))
test_y = np.array((200,))
test_mask = np.array((200, 50))

# get the default arguments
args = default_args()
# modify part of the arguments
args.space_turbo = False
args.n_hidden = 100
args.batch_size = 32

• Classification

# initilize a model
model = RNNModel(args, emb_mat, bi_direction=False, rnn_type='GRU', use_attention=True)
# initilize a classifier
nn = Classify(model, args, train_x, train_y, train_mask, test_x, test_y, test_mask)
# do training and testing
evals = nn.train_test(device_id=0)

• Run several times and get the average score.

# initilize a model
model = CNNModel(args, emb_mat, kernel_widths=[2, 3, 4])
# initilize a classifier
nn = Classify(model, args, train_x, train_y, train_mask)
# run the model several times
avg_evals = average_several_run(nn.train_test, args, n_times=8, n_paral=4, fold=5)

• Parameters' grid search.

# initilize a model
model = TransformerModel(args, n_layer=12, n_head=8)
# initilize a classifier
nn = Classify(model, args, train_x, train_y, train_mask, test_x, test_y, test_mask)
# set searching params
params_search = {'learning_rate': [0.1, 0.01], 'n_hidden': [50, 100]}
# run grid search
max_evals = grid_search(nn, nn.train_test, args, params_search)

• Sequence labeling

from dnnnlp.model import RNNCRFModel
from dnnnlp.exec import default_args, SequenceLabeling

# load the train data
train_x = np.array((1000, 50))
train_y = np.array((1000, 50))
train_mask = np.array((1000, 50))

# initilize a model
model = RNNCRFModel(args)
# initilize a labeler
nn = SequenceLabeling(model, args, train_x, train_y, train_mask)
# do cross validation
nn.cross_validation(fold=5)

History

version 1.1

• Add CRFLayer: packaging CRF for both training and testing.
• Add RNNCRFModel: a integrated RNN-CRF sequence labeling model.
• Add SequenceLabeling: a sequence labeling execution module that inherits from Classify.
• Fix errors in judging whether a tensor is None.

version 1.0

• Rename project dnn to dnnnlp.
• Remove file base, add file utils.
• Optimize and rename SoftmaxLayer and SoftAttentionLayer.
• Rewrite and rename EmbeddingLayer, CNNLayer and RNNLayer.
• Rewrite MultiheadAttentionLayer: a packaging attention layer based on nn.MultiheadAttention.
• Rewrite TransformerLayer: support new MultiheadAttentionLayer.
• Optimize and rename CNNModel, RNNModel and TransformerModel.
• Optimize and rename Classify: a highly applicable classification execution module.
• Rewrite average_several_run and grid_search: support multi-GPU parallel.
• Support pytorch 1.2.0.

version 0.12

• Update RNN_layer: fully support for tanh, LSTM and GRU.
• Fix errors in some mask operations.
• Support pytorch 1.1.0.

Old version 0.12.3.

version 0.11

• Provides an acceleration method by using more GPU memories.
• Fix the problem of memory consumption caused by abnormal data reading.
• Add multi_head_attention_layer: packaging multi-head attention for Transformer.
• Add Transformer_layer and Transformer_model: packaging Transformer layer and model written by ourself.
• Support data disruption for training.

version 0.10

• Split the code into four files: base, layer, model, exec.
• Add CNN_layer and CNN_model: packaging CNN layer and model.
• Support multi-GPU parallel for each model.

version 0.9

• Fix the problem of output format.
• Fix the statistical errors in cross-validation part of LSTM_classify.
• Rename: LSTM_model to RNN_layer, self_attention to self_attention_layer.
• Add softmax_layer: a packaging fully-connected layer.

version 0.8

• Adjust the applicability of functions in LSTM_classify to avoid rewriting in LSTM_sequence.
• Optimize the way of parameter transfer.
• A more complete evaluation mechanism.

version 0.7

• Add LSTM_sequence: a sequence labeling module for LSTM_model.
• Fix the nan-value problem in hierarchical classification.
• Support pytorch 1.0.0.

version 0.6

• Update LSTM_classify: support hierarchical classification.
• The GRU_model is merged into the LSTM_model.
• Adapt to CPU operation.

version 0.5

• Split the running part of LSTM_classify to reduce the rewrite of custom models.
• Add control for visual output.
• Create function average_several_run: support to get the average score after several training and testing.
• Create function grid_search: support parameters' grid search.

version 0.4

• Add GRU_model: a packaging GRU model based on nn.GRU.
• Support L2 regular.

version 0.3

• Add self_attention: provides attention mechanism support.
• Update LSTM_classify: adapts to complex custom models.

version 0.2

• Support mode selection of embedding.
• Default usage of nn.Dropout.
• Create function default_args to provide default hyperparameters.

version 0.1

• Initilization of project dnn: based on pytorch 0.4.1.
• Add LSTM_model: a packaging LSTM model based on nn.LSTM.
• Add LSTM_classify: a classification module for LSTM model, which supports train-test and corss-validation.