def test_train_use_pretrained_embedding(self):
""" 模型训练,使用预训练embed.
"""
print('{} test_train_use_pretrained_embedding {}'.format(
'-' * 15, '-' * 15))
# 数据集加载
test_data_pickle = './data/aclImdb/test_data.pkl'
test_data = load_serialize_obj(test_data_pickle)
test_data = test_data[:1000] # 数据量比较大，cpu电脑跑不动，取一部分进行训练
test_data_tokenized = get_tokenized_imdb(imdb_data=test_data)
test_data_vocab = get_tokenized_vocab(test_data_tokenized)
vocab_size = len(test_data_vocab)
print('vocab len:{}'.format(vocab_size)) # vocab len:4345
test_iter = get_imdb_data_iter(test_data,
test_data_vocab,
batch_size=8,
shuffle=True)
print('test_iter len:{}'.format(len(test_iter))) # test_iter len:125
# 构造模型
net = TextCNN(vocab_size=vocab_size, labels_size=2)
print('参数量:{}'.format(get_parameter_number(
net))) # total:263.152 Thousand, trainable:263.152 Thousand
# 使用预训练embed初始化
glove_embedding = torchtext.vocab.GloVe(name='6B',
dim=50,
cache='./data/torchtext')
print("glove_embedding 一共包含%d个词。" %
len(glove_embedding.stoi)) # 一共包含400000个词。
words = test_data_vocab.itos
embed = load_pretrained_embedding(
words=words,
pretrained_vocab=glove_embedding) # There are 1004 oov words.
net.embedding.weight.data.copy_(embed)
net.embedding.weight.requires_grad = False # 直接加载预训练好的, 所以不需要更新它
print('参数量:{}'.format(get_parameter_number(
net))) # total:263.152 Thousand, trainable:45.902 Thousand
print(net)
optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, net.parameters()))
loss_func = nn.CrossEntropyLoss()
# 训练
train_net(net,
train_iter=test_iter,
dev_iter=test_iter,
max_epoch=2,
optimizer=optimizer,
loss_func=loss_func)
init_file_path('./data/save/text_classify/textcnn/')
# 保存模型
torch.save(net, f='./data/save/text_classify/textcnn/model.pkl')
# 保存vocabulary
save_vocab_words(
test_data_vocab,
file_name='./data/save/text_classify/textcnn/vocab_words.txt')
save_serialize_obj(
test_data_vocab,
filename='./data/save/text_classify/textcnn/vocab.pkl')
def test_train(self):
""" 模型训练.
"""
print('{} test_train {}'.format('-' * 15, '-' * 15))
# 数据集加载
data_file = './data/nlp/fr-en-small.txt'
in_vocab, out_vocab, dataset = read_data(data_file, max_seq_len=7)
print('in_vocab len:{}, out_vocab len:{}'.format(
len(in_vocab), len(out_vocab))) # n_vocab len:46, out_vocab len:38
# 构造模型
encoder_net = Encoder(vocab_size=len(in_vocab),
embed_size=50,
num_hiddens=64,
num_layers=2,
drop_prob=0.5)
print('encoder_net 参数量:{}'.format(get_parameter_number(
encoder_net))) # total:49.532 Thousand, trainable:49.532 Thousand
decoder_net = Decoder(vocab_size=len(out_vocab),
embed_size=50,
num_hiddens=64,
num_layers=2,
attention_size=10,
drop_prob=0.5)
print('decoder_net 参数量:{}'.format(get_parameter_number(
decoder_net))) # total:65.18 Thousand, trainable:65.18 Thousand
# 训练
train_machine_translation_net(encoder=encoder_net,
decoder=decoder_net,
dataset=dataset,
out_vocab=out_vocab,
max_epoch=100)
# 2020-04-12 16:48:01 I [train.py:74] epoch 100, loss 0.008
# 保存模型
file_path = './data/save/machine_translation/attention'
init_file_path(file_path)
torch.save(encoder_net, f=os.path.join(file_path, 'encoder_net.pkl'))
torch.save(decoder_net, f=os.path.join(file_path, 'decoder_net.pkl'))
# 保存vocabulary
save_vocab_words(in_vocab,
file_name=os.path.join(file_path,
'in_vocab_words.txt'))
save_serialize_obj(in_vocab,
filename=os.path.join(file_path, 'in_vocab.pkl'))
save_vocab_words(out_vocab,
file_name=os.path.join(file_path,
'out_vocab_words.txt'))
save_serialize_obj(out_vocab,
filename=os.path.join(file_path, 'out_vocab.pkl'))
def test_train(self):
""" 模型训练.
"""
print('{} test_train {}'.format('-' * 15, '-' * 15))
# 数据集加载
test_data_pickle = './data/aclImdb/test_data.pkl'
test_data = load_serialize_obj(test_data_pickle)
test_data = test_data[:100] # 数据量比较大，cpu电脑跑不动，取一部分进行训练
test_data_tokenized = get_tokenized_imdb(imdb_data=test_data)
test_data_vocab = get_tokenized_vocab(test_data_tokenized)
vocab_size = len(test_data_vocab)
print('vocab len:{}'.format(vocab_size)) # vocab len:45098
test_iter = get_imdb_data_iter(test_data,
test_data_vocab,
batch_size=8,
shuffle=True)
print('test_iter len:{}'.format(len(test_iter))) # test_iter len:3218
# 构造模型
net = BiLSTM(vocab_size=vocab_size, labels_size=2)
print('参数量:{}'.format(get_parameter_number(
net))) # total:436.002 Thousand, trainable:436.002 Thousand
print(net)
optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, net.parameters()))
loss_func = nn.CrossEntropyLoss()
# 训练
train_net(net,
train_iter=test_iter,
dev_iter=test_iter,
max_epoch=5,
optimizer=optimizer,
loss_func=loss_func)
def test_get_parameter_number(self):
""" 统计神经网络参数个数.
"""
print('{} test_get_parameter_number {}'.format('-'*15, '-'*15))
net = LeNet()
print(net)
print(get_parameter_number(net)) # {'total': 44426, 'trainable': 44426}
def test_get_parameter_number(self):
""" 统计神经网络参数个数.
"""
print('{} test_get_parameter_number {}'.format('-' * 15, '-' * 15))
net = BiLSTM(vocab_size=100000, labels_size=2)
print(net)
print(get_parameter_number(
net)) # {'total': 454802, 'trainable': 454802}
def test_get_parameter_number(self):
""" 统计神经网络参数个数.
"""
print('{} test_get_parameter_number {}'.format('-' * 15, '-' * 15))
net = ResNet18()
print(net)
print(get_parameter_number(
net)) # {'total': 11178378, 'trainable': 11178378}
def test_get_parameter_number(self):
""" 统计神经网络参数个数.
"""
print('{} test_get_parameter_number {}'.format('-' * 15, '-' * 15))
net = VGG11()
print(net)
print(get_parameter_number(
net)) # {'total': 128806154, 'trainable': 128806154}
def test_get_parameter_number(self):
""" 统计神经网络参数个数.
"""
print('{} test_get_parameter_number {}'.format('-' * 15, '-' * 15))
encoder_net = Encoder(vocab_size=100,
embed_size=50,
num_hiddens=64,
num_layers=2,
drop_prob=0.5)
print(encoder_net)
print(get_parameter_number(
encoder_net)) # total:52.232 Thousand, trainable:52.232 Thousand
decoder_net = Decoder(vocab_size=100,
embed_size=50,
num_hiddens=64,
num_layers=2,
attention_size=10,
drop_prob=0.5)
print(decoder_net)
print(get_parameter_number(
decoder_net)) # total:72.31 Thousand, trainable:72.31 Thousand
