def __init__(self, corpus, **opts):
self.corpus = corpus
self.opts = opts
self.global_step = get_or_create_global_step()
self.increment_global_step_op = tf.assign(self.global_step, self.global_step + 1, name="increment_global_step")
self.corpus_size = get_corpus_size(self.corpus["train"])
self.corpus_size_valid = get_corpus_size(self.corpus["valid"])
self.word2idx, self.idx2word = build_vocab(self.corpus["train"])
self.vocab_size = len(self.word2idx)
self.generator_template = tf.make_template(GENERATOR_PREFIX, generator)
self.discriminator_template = tf.make_template(DISCRIMINATOR_PREFIX, discriminator)
self.enqueue_data, _, source, target, sequence_length = \
prepare_data(self.corpus["train"], self.word2idx, num_threads=7, **self.opts)
# TODO: option to either do pretrain or just generate?
self.g_tensors_pretrain = self.generator_template(
source, target, sequence_length, self.vocab_size, **self.opts)
self.enqueue_data_valid, self.input_ph, source_valid, target_valid, sequence_length_valid = \
prepare_data(self.corpus["valid"], self.word2idx, num_threads=1, **self.opts)
self.g_tensors_pretrain_valid = self.generator_template(
source_valid, target_valid, sequence_length_valid, self.vocab_size, **self.opts)
self.decoder_fn = prepare_custom_decoder(
sequence_length, self.g_tensors_pretrain.embedding_matrix, self.g_tensors_pretrain.output_projections)
self.g_tensors_fake = self.generator_template(
source, target, sequence_length, self.vocab_size, decoder_fn=self.decoder_fn, **self.opts)
self.g_tensors_fake_valid = self.generator_template(
source_valid, target_valid, sequence_length_valid, self.vocab_size, decoder_fn=self.decoder_fn, **self.opts)
# TODO: using the rnn outputs from pretraining as "real" instead of target embeddings (aka professor forcing)
self.d_tensors_real = self.discriminator_template(
self.g_tensors_pretrain.rnn_outputs, sequence_length, is_real=True, **self.opts)
# TODO: check to see if sequence_length is correct
self.d_tensors_fake = self.discriminator_template(
self.g_tensors_fake.rnn_outputs, None, is_real=False, **self.opts)
self.g_tvars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=GENERATOR_PREFIX)
self.d_tvars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=DISCRIMINATOR_PREFIX)
def __init__(self, corpus, **opts):
self.corpus = corpus
self.opts = opts
self.global_step = get_or_create_global_step()
self.increment_global_step_op = tf.assign(self.global_step, self.global_step + 1, name="increment_global_step")
self.corpus_size = get_corpus_size(self.corpus["train"])
self.corpus_size_valid = get_corpus_size(self.corpus["valid"])
self.word2idx, self.idx2word = build_vocab(self.corpus["train"])
self.vocab_size = len(self.word2idx)
self.generator_template = tf.make_template(GENERATOR_PREFIX, generator)
self.discriminator_template = tf.make_template(DISCRIMINATOR_PREFIX, discriminator)
self.enqueue_data, _, source, target, sequence_length = \
prepare_data(self.corpus["train"], self.word2idx, num_threads=7, **self.opts)
# TODO: option to either do pretrain or just generate?
self.g_tensors_pretrain = self.generator_template(
source, target, sequence_length, self.vocab_size, **self.opts)
self.enqueue_data_valid, self.input_ph, source_valid, target_valid, sequence_length_valid = \
prepare_data(self.corpus["valid"], self.word2idx, num_threads=1, **self.opts)
self.g_tensors_pretrain_valid = self.generator_template(
source_valid, target_valid, sequence_length_valid, self.vocab_size, **self.opts)
self.decoder_fn = prepare_custom_decoder(sequence_length)
self.g_tensors_fake = self.generator_template(
source, target, sequence_length, self.vocab_size, decoder_fn=self.decoder_fn, **self.opts)
# TODO: using the rnn outputs from pretraining as "real" instead of target embeddings (aka professor forcing)
self.d_tensors_real = self.discriminator_template(
self.g_tensors_pretrain.rnn_outputs, sequence_length, is_real=True, **self.opts)
# TODO: check to see if sequence_length is correct
self.d_tensors_fake = self.discriminator_template(
self.g_tensors_fake.rnn_outputs, None, is_real=False, **self.opts)
self.g_tvars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=GENERATOR_PREFIX)
self.d_tvars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=DISCRIMINATOR_PREFIX)
def main(args):
# Create model directory
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Image preprocessing
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.033, 0.032, 0.033), (0.027, 0.027, 0.027))
])
# Build vocab
vocab = build_vocab(args.root_path, threshold=0)
vocab_path = args.vocab_path
with open(vocab_path, 'wb') as f:
pickle.dump(vocab, f)
len_vocab = vocab.idx
print(vocab.idx2word)
# Build data loader
data_loader = get_loader(args.root_path,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
# Build the models
encoder = ResNet(ResidualBlock, [3, 3, 3], args.embed_size)
decoder = DecoderRNN(args.embed_size, args.hidden_size, len(vocab),
args.num_layers)
#Build atten models
if torch.cuda.is_available():
encoder.cuda(1)
decoder.cuda(1)
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(decoder.parameters()) + list(encoder.parameters())
optimizer = torch.optim.Adam(params, lr=args.learning_rate)
# Train the Models
total_step = len(data_loader)
for epoch in range(args.num_epochs):
for i, (images, captions, lengths) in enumerate(data_loader):
# make one hot
# cap_ = torch.unsqueeze(captions,2)
# one_hot_ = torch.FloatTensor(captions.size(0),captions.size(1),len_vocab).zero_()
# one_hot_caption = one_hot_.scatter_(2, cap_, 1)
# Set mini-batch dataset
images = to_var(images)
captions = to_var(captions)
#captions_ = to_var(one_hot_caption)
targets = pack_padded_sequence(captions, lengths,
batch_first=True)[0]
# Forward, Backward and Optimize
optimizer.zero_grad()
features = encoder(images)
outputs = decoder(features, captions, lengths)
captions = captions.view(-1)
outputs = outputs.view(-1, len_vocab)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
#print(targets)
#print(outputs)
# Print log info
if i % args.log_step == 0:
print(
'Epoch [%d/%d], Step [%d/%d], Loss: %.4f, Perplexity: %5.4f'
% (epoch, args.num_epochs, i, total_step, loss.data[0],
np.exp(loss.data[0])))
#test set accuracy
#print(outputs.max(1)[1])
outputs_np = outputs.max(1)[1].cpu().data.numpy()
targets_np = targets.cpu().data.numpy()
print(outputs_np)
print(targets_np)
location_match = 0
size_match = 0
shape_match = 0
exact_match = 0
for i in range(len(targets_np)):
if outputs_np[i] == targets_np[i]:
exact_match += 1
if i >= args.batch_size and i < args.batch_size * 2 and outputs_np[
i] == targets_np[i]:
shape_match += 1
elif i >= args.batch_size * 2 and i < args.batch_size * 3 and outputs_np[
i] == targets_np[i]:
location_match += 1
elif i >= args.batch_size * 3 and i < args.batch_size * 4 and outputs_np[
i] == targets_np[i]:
size_match += 1
print(
'location match : %.4f, shape match : %.4f, exact_match: %.4f'
% (location_match / (args.batch_size), shape_match /
args.batch_size, exact_match / len(targets_np)))
# Save the models
if (i + 1) % args.save_step == 0:
torch.save(
decoder.state_dict(),
os.path.join(args.model_path,
'decoder-%d-%d.pkl' % (epoch + 1, i + 1)))
torch.save(
encoder.state_dict(),
os.path.join(args.model_path,
'encoder-%d-%d.pkl' % (epoch + 1, i + 1)))
def __init__(self, dataset, opts, use_pretrained_embeddings=True):
# TODO: Add Dropout layer later.
self.dropout_keep_prob = tf.placeholder(tf.float32,
name="dropout_keep_prob")
if use_pretrained_embeddings:
word2vec = get_word2vec_model(WORD2VEC_PATH)
word2idx, idx2word, label2idx, idx2label = build_vocab(
dataset.training_files,
dataset.vocab_file,
word2vec,
min_counts=opts['min_counts'])
embedding_weights = get_embedding_weights(word2idx, word2vec)
embedding_length = embedding_weights.shape[1]
# TODO: embedding might be trainable.
self.embeddings = tf.Variable(embedding_weights,
dtype=tf.float32,
trainable=False)
else:
word2idx, idx2word, label2idx, idx2label = build_vocab(
dataset.training_files,
dataset.vocab_file,
min_counts=opts['min_counts'])
embedding_length = opts['embedding_length']
self.embeddings = tf.Variable(tf.random_uniform(
[len(word2idx), embedding_length], -1.0, 1.0),
dtype=tf.float32)
self.sess = tf.Session()
self.enqueue_data, self.source, self.target_word, self.label, \
self.sequence_length = prepare_data(self.sess, dataset.training_files, word2idx, label2idx, **opts)
self.target_words_embedded = tf.nn.embedding_lookup(
self.embeddings, self.target_word)
self.sentences_embedded = tf.nn.embedding_lookup(
self.embeddings, self.source)
hidden_unit_size = opts['hidden_unit_size']
num_senses = len(label2idx)
encoder_cell = LSTMCell(hidden_unit_size)
(encoder_fw_outputs, encoder_bw_outputs), (encoder_fw_final_state, encoder_bw_final_state) = \
tf.nn.bidirectional_dynamic_rnn(cell_fw=encoder_cell, cell_bw=encoder_cell, inputs=self.sentences_embedded,
sequence_length=self.sequence_length, dtype=tf.float32, time_major=True)
encoder_final_state_c = tf.concat(
(encoder_fw_final_state.c, encoder_bw_final_state.c), 1)
encoder_final_state_h = tf.concat(
(encoder_fw_final_state.h, encoder_bw_final_state.h), 1)
encoder_final_state = LSTMStateTuple(c=encoder_final_state_c,
h=encoder_final_state_h)
# self.encoder_target_embedding = encoder_final_state.c
self.encoder_target_embedding = tf.concat(
(encoder_final_state.c, self.target_words_embedded), 1)
with tf.name_scope("output"):
W = tf.Variable(tf.truncated_normal(
[hidden_unit_size * 2 + embedding_length, num_senses],
stddev=0.1),
name="W")
b = tf.Variable(tf.constant(0.1, shape=[num_senses]), name="b")
self.scores = tf.matmul(self.encoder_target_embedding, W) + b
self.predictions = tf.argmax(self.scores, 1, name="predictions")
with tf.name_scope('cross_entropy'):
labels = tf.one_hot(self.label, num_senses)
self.diff = tf.nn.softmax_cross_entropy_with_logits(
labels=labels, logits=self.scores)
with tf.name_scope('loss'):
self.loss = tf.reduce_mean(self.diff)
with tf.name_scope('train'):
self.train_step = tf.train.AdamOptimizer(
opts['learning_rate']).minimize(self.loss)
with tf.name_scope('accuracy'):
with tf.name_scope('correct_prediction'):
correct_prediction = tf.equal(self.predictions,
tf.argmax(labels, 1))
with tf.name_scope('accuracy'):
self.accuracy = tf.reduce_mean(
tf.cast(correct_prediction, tf.float32))
self.sess.run(tf.global_variables_initializer())
def main():
training_dataset = [[
"안녕하세요, 제 이름은 윤주성입니다", "hello, my name is joosung yoon"
], ["저는 텐서플로우를 좋아합니다", "i like tensorflow"]]
X_y_split = list(zip(*training_dataset))
X_train_str = list(
X_y_split[0]) # ['안녕하세요, 제 이름은 윤주성입니다', '저는 텐서플로우를 좋아합니다']
y_train_str = list(
X_y_split[1]
) # ['Hello, my name is joosung Yoon', 'I like TensorFlow']
print(X_train_str)
print(y_train_str)
corpus = []
corpus.extend(X_train_str)
corpus.extend(
y_train_str
) # ['안녕하세요, 제 이름은 윤주성입니다', '저는 텐서플로우를 좋아합니다', 'Hello, my name is joosung Yoon', 'I like TensorFlow']
vocab = build_vocab(corpus)
print(vocab.idx2word)
max_sequence_len = 13
X_train, _, _ = word_to_pad_word_ids(text_batch=X_train_str,
vocab=vocab,
maxlen=max_sequence_len,
add_start_end_token=True)
_, tar_inp, tar_real = word_to_pad_word_ids(
text_batch=y_train_str,
vocab=vocab,
maxlen=max_sequence_len,
add_start_end_token=True) # add +1 maxlen for start, end token
print(
X_train
) # [[ 5 6 7 8 9 10 11 12 13 14 0 0 0 0 0], [15 16 17 18 19 0 0 0 0 0 0 0 0 0 0]]
print(
tar_inp
) # [[20 8 21 22 23 24 25 0 0 0 0 0 0 0 0], [26 27 28 0 0 0 0 0 0 0 0 0 0 0 0]]
print(tar_real)
print(decode_word_ids(X_train, vocab))
# [['안녕/NNG', '하/XSV', '세요/EP+EF', ',/SC', '제/MM', '이름/NNG', '은/JX', '윤주/NNG', '성/XSN', '입니다/VCP+EC', '<pad>', '<pad>', '<pad>', '<pad>', '<pad>'],
# ['저/NP', '는/JX', '텐서플로우/NNP', '를/JKO', '좋아합니다/VV+EC', '<pad>', '<pad>', '<pad>', '<pad>', '<pad>', '<pad>', '<pad>', '<pad>', '<pad>', '<pad>']]
config = {}
config['vocab_size'] = len(vocab.idx2word)
config['maxlen'] = max_sequence_len
config['embed_dim'] = 100
config['head_num'] = 5
config['split_embed_dim'] = 20
config['layer_num'] = 2
config['feed_forward_dim'] = 100
# define model
model = Transformer(config=config)
loss_object = tf.keras.losses.SparseCategoricalCrossentropy(
from_logits=True, reduction='none') # input label == index of class
optimizer = tf.keras.optimizers.Adam()
train_loss = tf.keras.metrics.Mean(name='train_loss')
train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
name='train_accuracy')
test_loss = tf.keras.metrics.Mean(name='test_loss')
test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
name='test_accuracy')
def loss_function(real, pred):
mask = tf.math.logical_not(tf.math.equal(real, 0)) # padding 아닌건 1
loss_ = loss_object(real, pred)
mask = tf.cast(mask, dtype=loss_.dtype)
loss_ *= mask # 패딩이 아닌 1인 값은 살리고, 패딩인 값인 0인 값은 없앰
return tf.reduce_mean(loss_)
def create_padding_mask(seq):
seq = tf.cast(tf.math.equal(seq, 0), tf.float32)
# add extra dimensions so that we can add the padding
# to the attention logits.
return seq[:, tf.newaxis, tf.newaxis, :] # (batch_size, 1, 1, seq_len)
def create_look_ahead_mask(step_size):
"""
- decoder에서 각 상태에 대한 self-attention이 inference step에 맞게 future token을 보지 못하게 해야됨
- 각 step이 소유하고 있는 attention은 step개수 만큼임
- future token보지 못하게 하려면 각 step에서 future step에 대해서 마스킹 해야함
- 1 step에서는 나머지 n-1개 masking, 2번째 스텝에서는 앞에 두개 빼고 나머지 n-2개 마스킹
- 이렇게 하면 역삼각형 모양의 마스킹 매트릭스가 나옴
- step * step 을 대각선으로 나눈 모양임
example)
x = tf.random.uniform((1, 3))
temp = create_look_ahead_mask(x.shape[1])
temp:
<tf.Tensor: id=311521, shape=(3, 3), dtype=float32, numpy=
array([[ 0., 1., 1.],
[ 0., 0., 1.],
[ 0., 0., 0.]], dtype=float32)>
Special usecase:
tf.matrix_band_part(input, 0, -1) ==> Upper triangular part.
tf.matrix_band_part(input, -1, 0) ==> Lower triangular part.
tf.matrix_band_part(input, 0, 0) ==> Diagonal.
:param step_size:
:return:
"""
mask = 1 - tf.linalg.band_part(tf.ones((step_size, step_size)), -1, 0)
return mask # (seq_len, seq_len)
def create_masks(inp, tar):
# Encoder padding mask
enc_padding_mask = create_padding_mask(inp)
# Used in the 2nd attention block in the decoder.
# This padding mask is used to mask the encoder outputs.
dec_padding_mask = create_padding_mask(inp)
# Used in the 1st attention block in the decoder.
# It is used to pad and mask future tokens in the input received by
# the decoder.
look_ahead_mask = create_look_ahead_mask(tf.shape(tar)[1])
dec_target_padding_mask = create_padding_mask(tar)
combined_mask = tf.maximum(dec_target_padding_mask, look_ahead_mask)
return enc_padding_mask, combined_mask, dec_padding_mask
# 세션 대신 tf.function() decorator로 파이썬 함수를 감싸면, 이 함수를 하나의 그래프로 실행하기 위해 JIT 컴파일함
# tf.function()을 쓰면 eager mode -> graph mode 되는 것임
# @tf.function
def train_step(enc_input, tar_inp, tar_real):
# tar_inp = label[:, :-1] # remove </s>
# tar_real = label[:, 1:] # remove <s>
enc_padding_mask, combined_mask, dec_padding_mask = create_masks(
enc_input, tar_inp)
with tf.GradientTape() as tape:
predictions, attention_weights = model(enc_input, tar_inp, True,
enc_padding_mask,
combined_mask,
dec_padding_mask)
loss = loss_function(tar_real,
predictions) # masking losses for padding
predicted_id = tf.cast(tf.argmax(predictions, axis=-1),
tf.int32).numpy()
print("X_train: ", decode_word_ids(enc_input.numpy(), vocab))
print("tar_inp: ", decode_word_ids(tar_inp.numpy(), vocab))
print("tar_real: ", decode_word_ids(tar_real.numpy(), vocab))
print("result: ", decode_word_ids(predicted_id, vocab))
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
train_loss(loss)
train_accuracy(tar_real, predictions)
# @tf.function
# def test_step(Y_test, label):
# predictions = model(Y_test)
# t_loss = loss_object(label, predictions)
#
# test_loss(t_loss)
# test_accuracy(label, predictions)
def plot_attention_weights(attention, sentence, result, layer):
import matplotlib.pyplot as plt
from matplotlib import font_manager, rc
# print("font_list: ", font_manager.get_fontconfig_fonts())
font_name = font_manager.FontProperties(
fname='/Library/Fonts/NanumSquareBold.ttf').get_name()
rc('font', family=font_name)
fig = plt.figure(figsize=(16, 8))
sentence, _, _ = word_to_pad_word_ids(
text_batch=[sentence],
vocab=vocab,
maxlen=max_sequence_len,
add_start_end_token=True) #tokenizer_pt.encode(sentence)
attention = tf.squeeze(attention[layer], axis=0)
for head in range(attention.shape[0]):
ax = fig.add_subplot(2, 4, head + 1)
# plot the attention weights
im = ax.matshow(
attention[head][:, :],
cmap='viridis') # viridis #plt.cm.Reds # plt.cm.Blues
fontdict = {'fontsize': 10}
ax.set_xticks(range(len(decode_word_ids(sentence, vocab)[0])))
ax.set_yticks(range(len(decode_word_ids(result, vocab)[0])))
from mpl_toolkits.axes_grid1 import make_axes_locatable
divider = make_axes_locatable(ax)
cax = divider.append_axes("right", size="5%", pad=0.05)
plt.colorbar(im, cax=cax)
ax.set_xticklabels(decode_word_ids(sentence, vocab)[0],
fontdict=fontdict,
rotation=90)
ax.set_yticklabels(decode_word_ids(result, vocab)[0],
fontdict=fontdict)
ax.set_xlabel('Head {}'.format(head + 1))
plt.tight_layout()
plt.show()
def evaluate(inp_sentence, vocab, max_sequence_len):
# inference 일때는 굳이 length를 +1 하지 않아도됨
encoder_input, _, _ = word_to_pad_word_ids(text_batch=[inp_sentence],
vocab=vocab,
maxlen=max_sequence_len,
add_start_end_token=True)
print("encoder_input: ", encoder_input)
decoder_input = ['<s>']
decoder_input = [vocab.word2idx[_] for _ in decoder_input]
output = tf.expand_dims(decoder_input, 0)
print("output: ", decode_word_ids(output.numpy(), vocab))
for i in range(max_sequence_len):
enc_padding_mask, combined_mask, dec_padding_mask = create_masks(
encoder_input, output)
# predictions.shape == (batch_size, seq_len, vocab_size)
predictions, attention_weights = model(encoder_input, output,
False, enc_padding_mask,
combined_mask,
dec_padding_mask)
# select the last word from the seq_len dimension
print("predicted_id: ",
tf.cast(tf.argmax(predictions, axis=-1), tf.int32))
predictions = predictions[:, -1:, :] # (batch_size, 1, vocab_size)
predicted_id = tf.cast(tf.argmax(predictions, axis=-1), tf.int32)
# return the result if the predicted_id is equal to the end token
if tf.equal(predicted_id, vocab.word2idx['</s>']):
return tf.squeeze(output, axis=0), attention_weights
# concatentate the predicted_id to the output which is given to the decoder
# as its input.
output = tf.concat([output, predicted_id], axis=-1)
print("output: ", decode_word_ids(output.numpy(), vocab))
return tf.squeeze(output, axis=0), attention_weights
def translate(sentence, vocab, max_sequence_len, plot=''):
result, attention_weights = evaluate(sentence, vocab, max_sequence_len)
result = [result.numpy()]
predicted_sentence = decode_word_ids(result, vocab)
print('Input: {}'.format(sentence))
print('Predicted translation: {}'.format(predicted_sentence))
if plot:
plot_attention_weights(attention_weights, sentence, result, plot)
### Training
EPOCHS = 4000
BATCH_SIZE = 45
train_ds = tf.data.Dataset.from_tensor_slices((X_train, tar_inp, tar_real))
train_ds = train_ds.repeat(EPOCHS).shuffle(1024).batch(BATCH_SIZE)
train_ds = train_ds.prefetch(tf.data.experimental.AUTOTUNE)
for step, (X_train_batch, tar_inp, tar_real) in enumerate(train_ds):
train_step(X_train_batch, tar_inp, tar_real)
template = 'Step {}, Loss: {}, Accuracy: {}, Test Loss: {}, Test Accuracy: {}'
print(
template.format(step + 1, train_loss.result(),
train_accuracy.result() * 100, test_loss.result(),
test_accuracy.result() * 100))
translate("안녕하세요, 제 이름은 윤주성입니다",
vocab,
max_sequence_len,
plot='decoder_layer2_block2')
model.summary()
eval_data_dir = sys.argv[2]
base_dir = sys.argv[3]
window_size = sys.argv[4]
train_ratio = sys.argv[5]
vocab_dir = os.path.join(base_dir, 'vocab.txt')
save_dir = os.path.join(base_dir, train_ratio + '/checkpoints/textrnn')
save_path = os.path.join(save_dir, 'best_validation') # 最佳验证结果保存路径
window_size = int(window_size)
train_ratio = float(train_ratio)
print('Configuring RNN model...')
print('Building vocab if not exists.')
start_time_vocab = time.time()
config = TRNNConfig()
if not os.path.exists(vocab_dir): # 如果不存在词汇表，重建
build_vocab(train_data_dir, vocab_dir)
categories, cat_to_id = read_category()
words, word_to_id = read_vocab(vocab_dir)
config.vocab_size = len(words)
model = TextRNN(config)
time_dif_vocab = get_time_dif(start_time_vocab)
print("Time usage:", time_dif_vocab)
#读取原始数据并转换成三个集合
print("Processing and loading training and validation data...")
start_time = time.time()
x_train, x_val, x_test, y_train, y_val, y_test = process_all_file(
train_data_dir, eval_data_dir, train_ratio, word_to_id, cat_to_id,
config.seq_length, window_size)
time_dif = get_time_dif(start_time)
print("Time usage:", time_dif)
save_dir_bak = 'checkpoints/textcnn_bak'
save_path_bak = os.path.join(save_dir_bak, 'best_validation') # 最佳验证结果保存路径
save_dir = 'checkpoints/textcnn'
save_path = os.path.join(save_dir, 'best_validation')
print('Configuring CNN model...')
normal_num = [690] * 3
max_acc = 0
greatest_normal_num = 0
for i in normal_num:
get_train_data(i)
config = TCNNConfig()
if not os.path.exists(vocab_dir): # 如果不存在词汇表，重建
build_vocab(train_dir, vocab_dir, config.vocab_size)
categories, cat_to_id = read_category(train_dir)
words, word_to_id = read_vocab(vocab_dir)
config.vocab_size = len(words)
config.num_classes = len(categories)
model = TextCNN(config)
# 训练模型并保存到bak
train()
print("Loading test data...")
start_time = time.time()
x_test, y_test = process_file(test_dir, word_to_id, cat_to_id, config.seq_length)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.1)
session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
def main(args):
# Image preprocessing
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.033, 0.032, 0.033), (0.027, 0.027, 0.027))
])
vocab = build_vocab(args.root_path, threshold=0)
num_class = 9
# Build data loader
data_loader = get_loader(args.root_path,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
# Build the models
cnn = ResNet(ResidualBlock, [3, 3, 3], num_class)
if torch.cuda.is_available():
cnn.cuda(1)
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
params = list(cnn.parameters())
optimizer = torch.optim.Adam(params, lr=args.learning_rate)
# Train the Models
total_step = len(data_loader)
for epoch in range(args.num_epochs):
for i, (images, captions, lengths) in enumerate(data_loader):
#if i > 1 :
# break;
idx_arr = []
for element in captions[:, 1]:
idx_arr.append(int(vocab.idx2word[element]) - 1)
temp_arr = np.array(idx_arr)
trg_arr = torch.from_numpy(temp_arr)
target = to_var(trg_arr)
images = to_var(images)
optimizer.zero_grad()
features = cnn(images)
loss = criterion(features, target)
loss.backward()
optimizer.step()
# Print log info
if i % args.log_step == 0:
print(
'Epoch [%d/%d], Step [%d/%d], Loss: %.4f, Perplexity: %5.4f'
% (epoch, args.num_epochs, i, total_step, loss.data[0],
np.exp(loss.data[0])))
#print(features)
#print(target)
##test set accuracy
#rearrange tensor to batch_size * caption_size
re_target = rearrange_tensor(target, captions.size(0), 1)
re_out_max = rearrange_tensor(
features.max(1)[1], captions.size(0), 1)
#convert to numpy
outputs_np = re_out_max.cpu().data.numpy()
targets_np = re_target.cpu().data.numpy()
location_match = 0
for i in range(len(targets_np)):
if (outputs_np[i] == targets_np[i]):
location_match += 1
print('location match accuracy: %.4f' %
(location_match / len(targets_np)))
#test model
print('---------------------------------')
cnn.eval()
test_loader = get_loader(args.test_path,
vocab,
transform,
args.batch_size,
shuffle=True,
num_workers=args.num_workers)
for images, captions, lengths in test_loader:
idx_arr = []
for element in captions[:, 1]:
idx_arr.append(int(vocab.idx2word[element]) - 1)
temp_arr = np.array(idx_arr)
trg_arr = torch.from_numpy(temp_arr)
target = to_var(trg_arr)
images = to_var(images)
features = cnn(images)
re_target = rearrange_tensor(target, captions.size(0), 1)
re_out_max = rearrange_tensor(features.max(1)[1], captions.size(0), 1)
#convert to numpy
outputs_np = re_out_max.cpu().data.numpy()
targets_np = re_target.cpu().data.numpy()
location_match = 0
for i in range(len(targets_np)):
if (outputs_np[i] == targets_np[i]):
location_match += 1
print('location match accuracy: %.4f' %
(location_match / len(targets_np)))
test_txt_dirs = txt_dirs[int(len(txt_dirs) * 0.9):]
# print "载入测试样本..."
# test_txt_dirs = list()
# # test_data_dir = '/home/abc/ssd/pzw/nlp/data/0523/word_sep_test/'
# test_data_dir = '/home/abc/ssd/pzw/nlp/data/test_data3/'
# # test_data_dir = '/home/zhwpeng/abc/nlp/data/0324/word_sep_test/'
# # test_data_dir = '/home/zhwpeng/abc/text_classify/data/0412/raw/test_data3/'
# for fold in glob(test_data_dir + '*'):
# test_txt_dirs = test_txt_dirs + glob(fold + '/*.txt')
# # test_txt_dirs = test_txt_dirs + glob(fold + '/*.txt')[:10]
# # print "测试集样本总数是{}".format(len(test_txt_dirs))
# np.random.shuffle(test_txt_dirs)
print "配置CNN模型..."
config = TCNNConfig()
seq_length = config.seq_length
if not os.path.exists(vocab_dir): # 如果不存在词汇表，重建
build_vocab(txt_dirs, seq_length, vocab_dir, config.vocab_size)
categories, cat_to_id = read_category(types)
words, word_to_id = read_vocab(vocab_dir)
config.vocab_size = len(words)
model = TextCNN(config)
if sys.argv[1] == 'train':
print "开始训练..."
train()
else:
print "开始测试..."
model_tes_t(test_txt_dirs, train_flag=False)
