def test(args):
test_x, test_y, n_items = load_test(args.max_len)
args.n_items = n_items
test_batches = len(test_x) // args.batch_size
HR, NDCG, MRR = 0.0, 0.0, 0.0
test_loss = 0.0
model = RCNN(args)
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
with tf.Session(config=gpu_config) as sess:
saver = tf.train.Saver(tf.global_variables())
ckpt = tf.train.get_checkpoint_state(args.checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print('Restore model from {} successfully!'.format(
args.checkpoint_dir))
else:
print('Restore model from {} failed!'.format(args.checkpoint_dir))
return
for i in range(test_batches):
x = test_x[i * args.batch_size:(i + 1) * args.batch_size]
y = test_y[i * args.batch_size:(i + 1) * args.batch_size]
fetches = [model.sum_loss, model.top_k_index, model.y_labels]
feed_dict = {model.X: x, model.Y: y}
loss, top_k_index, labels = sess.run(fetches, feed_dict)
test_loss += loss
hr, ndcg, mrr = cal_eval(top_k_index, labels)
HR += hr
NDCG += ndcg
MRR += mrr
print('loss:{:6f}\tHR@{}:{:.6f}\tNDCG@{}:{:.6f}\tMRR@{}:{:.6f}'.format(
test_loss, args.top_k, HR, args.top_k, NDCG, args.top_k, MRR))
def main(args):
model = RCNN(vocab_size=args.vocab_size,
embedding_dim=args.embedding_dim,
hidden_size=args.hidden_size,
hidden_size_linear=args.hidden_size_linear,
class_num=args.class_num,
dropout=args.dropout).to(args.device)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model, dim=0)
train_texts, train_labels = read_file(args.train_file_path)
word2idx = build_dictionary(train_texts, vocab_size=args.vocab_size)
logger.info('Dictionary Finished!')
full_dataset = CustomTextDataset(train_texts, train_labels, word2idx)
num_train_data = len(full_dataset) - args.num_val_data
train_dataset, val_dataset = random_split(
full_dataset, [num_train_data, args.num_val_data])
train_dataloader = DataLoader(dataset=train_dataset,
collate_fn=lambda x: collate_fn(x, args),
batch_size=args.batch_size,
shuffle=True)
valid_dataloader = DataLoader(dataset=val_dataset,
collate_fn=lambda x: collate_fn(x, args),
batch_size=args.batch_size,
shuffle=True)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
train(model, optimizer, train_dataloader, valid_dataloader, args)
logger.info('******************** Train Finished ********************')
# Test
if args.test_set:
test_texts, test_labels = read_file(args.test_file_path)
test_dataset = CustomTextDataset(test_texts, test_labels, word2idx)
test_dataloader = DataLoader(dataset=test_dataset,
collate_fn=lambda x: collate_fn(x, args),
batch_size=args.batch_size,
shuffle=True)
model.load_state_dict(
torch.load(os.path.join(args.model_save_path, "best.pt")))
_, accuracy, precision, recall, f1, cm = evaluate(
model, test_dataloader, args)
logger.info('-' * 50)
logger.info(
f'|* TEST SET *| |ACC| {accuracy:>.4f} |PRECISION| {precision:>.4f} |RECALL| {recall:>.4f} |F1| {f1:>.4f}'
)
logger.info('-' * 50)
logger.info('---------------- CONFUSION MATRIX ----------------')
for i in range(len(cm)):
logger.info(cm[i])
logger.info('--------------------------------------------------')
def train(config):
train_data = pickle.load(open(os.path.join(config.data_path, config.train_name), "rb"))
dev_data = pickle.load(open(os.path.join(config.data_path, config.dev_name), "rb"))
test_data = pickle.load(open(os.path.join(config.data_path, config.test_name), "rb"))
vocabulary = pickle.load(open(os.path.join(config.data_path, config.vocabulary_name), "rb"))
# load w2v data
weight = pickle.load(open(os.path.join(config.data_path, config.weight_name), "rb"))
if config.task_name == "lstm":
text_model = LSTM(vocab_size=len(vocabulary), embed_dim=config.embed_dim,
output_dim=config.class_num, hidden_dim=config.hidden_dim,
num_layers=config.num_layers, dropout=config.dropout)
elif config.task_name == "lstm_maxpool":
text_model = LSTM_maxpool(vocab_size=len(vocabulary), embed_dim=config.embed_dim,
output_dim=config.class_num, hidden_dim=config.hidden_dim,
num_layers=config.num_layers, dropout=config.dropout)
elif config.task_name == "rnn":
text_model = RNN(vocab_size=len(vocabulary), embed_dim=config.embed_dim,
output_dim=config.class_num, hidden_dim=config.hidden_dim,
num_layers=config.num_layers, dropout=config.dropout)
elif config.task_name == "cnn":
text_model = CNN(vocab_size=len(vocabulary), embed_dim=config.embed_dim,
class_num=config.class_num, kernel_num=config.kernel_num,
kernel_sizes=config.kernel_sizes, dropout=config.dropout,
static=config.static, in_channels=config.in_channels)
elif config.task_name == "cnn_w2v":
text_model = CNN_w2v(vocab_size=len(vocabulary), embed_dim=config.embed_dim,
class_num=config.class_num, kernel_num=config.kernel_num,
kernel_sizes=config.kernel_sizes, dropout=config.dropout,
static=config.static, in_channels=config.in_channels,
weight=weight)
elif config.task_name == "rcnn":
text_model = RCNN(vocab_size=len(vocabulary), embed_dim=config.embed_dim,
output_dim=config.class_num, hidden_dim=config.hidden_dim,
num_layers=config.num_layers, dropout=config.dropout)
optimizer = Adam(lr=config.lr, weight_decay=config.weight_decay)
timing = TimingCallback()
early_stop = EarlyStopCallback(config.patience)
accuracy = AccuracyMetric(pred='output', target='target')
trainer = Trainer(train_data=train_data, model=text_model, loss=CrossEntropyLoss(),
batch_size=config.batch_size, check_code_level=0,
metrics=accuracy, n_epochs=config.epoch,
dev_data=dev_data, save_path=config.save_path,
print_every=config.print_every, validate_every=config.validate_every,
optimizer=optimizer, use_tqdm=False,
device=config.device, callbacks=[timing, early_stop])
trainer.train()
# test result
tester = Tester(test_data, text_model, metrics=accuracy)
tester.test()
def start_training(train_arguments, folder_index):
rcnn = RCNN(train_arguments.pos_loss_method, train_arguments.loss_weight_lambda,
train_arguments.prevent_overfitting_method).cuda()
rcnn.train() # train mode could use dropout.
npz_path = train_arguments.get_train_data_path(folder_index)
npz = np.load(npz_path)
print("\n\n\nload from: ", npz_path)
train_arguments.train_sentences = npz['train_sentences']
train_arguments.train_sentence_info = npz['train_sentence_info']
train_arguments.train_roi = npz['train_roi']
train_arguments.train_cls = npz['train_cls']
if train_arguments.normalize:
if train_arguments.dx_compute_method == "left_boundary":
train_arguments.train_tbbox = npz["train_norm_lb_tbbox"]
else:
train_arguments.train_tbbox = npz["train_norm_tbbox"]
else:
train_arguments.train_tbbox = npz['train_tbbox']
train_arguments.train_sentences = t.Tensor(train_arguments.train_sentences)
train_arguments.train_set = np.random.permutation(train_arguments.train_sentences.size(0)) # like shuffle
if train_arguments.prevent_overfitting_method.lower() == "l2 regu":
if train_arguments.partial_l2_penalty:
optimizer = optim.Adam([
{"params": rcnn.conv1.parameters(), "weight_decay": 0},
{"params": rcnn.cls_fc1.parameters(), "weight_decay": train_arguments.l2_beta},
{"params": rcnn.cls_score.parameters(), "weight_decay": train_arguments.l2_beta},
{"params": rcnn.bbox_fc1.parameters(), "weight_decay": train_arguments.l2_beta},
{"params": rcnn.bbox.parameters(), "weight_decay": train_arguments.l2_beta}
], lr=train_arguments.learning_rate)
else:
optimizer = optim.Adam(rcnn.parameters(), lr=train_arguments.learning_rate,
weight_decay=train_arguments.l2_beta)
else: # dropout optimizer
optimizer = optim.Adam(rcnn.parameters(), lr=train_arguments.learning_rate)
rcnn.optimizer = optimizer
for epoch_time in range(train_arguments.max_iter_epoch):
print('===========================================')
print('[Training Epoch {}]'.format(epoch_time + 1))
train_epoch(train_arguments, rcnn)
if epoch_time >= train_arguments.start_save_epoch:
save_directory = train_arguments.get_save_directory(folder_index)
save_path = save_directory + "model_epoch" + str(epoch_time + 1) + ".pth"
t.save(rcnn.state_dict(), save_path)
print("Model save in ", save_path)
def main(args):
model = RCNN(vocab_size=args.vocab_size,
embedding_dim=args.embedding_dim,
hidden_size=args.hidden_size,
hidden_size_linear=args.hidden_size_linear,
class_num=args.class_num,
dropout=args.dropout)
train_texts, train_labels = read_file(args.train_file_path)
test_texts, test_labels = read_file(args.test_file_path)
word2idx = build_dictionary(train_texts, vocab_size=args.vocab_size)
logger.info('Dictionary Finished!')
x_train, y_train = CustomTextDataset(train_texts, train_labels, word2idx)
x_test, y_test = CustomTextDataset(test_texts, test_labels, word2idx)
num_train_data = len(x_train)
optimizer = tf.keras.optimizers.Adam(learning_rate=args.lr)
hist = train(model, optimizer, x_train, x_test, y_train, y_test, args)
logger.info('******************** Train Finished ********************')
tf.saved_model.save(model, "/tmp/module_no_signatures")
def train_initialization(domain, classifier_name, all_data, data_type):
train_data, test_data, Final_test, Final_test_original, Final_test_gt, unique_vocab_dict, unique_vocab_list = all_data
output_size = 2
batch_size = 32
pre_train = True
embedding_tune = True
if data_type == 'train':
epoch_num = 10 if domain == 'captions' else 4
else: # 'dev'
epoch_num = 3 # sample test
embedding_length = 300 if domain != 'captions' else 50
hidden_size = 256 if domain != 'captions' else 32
learning_rate = collections.defaultdict(dict)
learning_rate['amazon'] = {'LSTM': 0.001, 'LSTMAtten': 0.0002, 'RNN': 0.001, 'RCNN': 0.001, 'SelfAttention': 0.001, 'CNN': 0.001}
learning_rate['yelp'] = {'LSTM': 0.002, 'LSTMAtten': 0.0002, 'RNN': 0.0001, 'RCNN': 0.001, 'SelfAttention': 0.0001, 'CNN': 0.001}
learning_rate['captions'] = {'LSTM': 0.005, 'LSTMAtten': 0.005, 'RNN': 0.01, 'RCNN': 0.01, 'SelfAttention': 0.005, 'CNN': 0.001}
TEXT, vocab_size, word_embeddings, train_iter, test_iter, Final_test_iter, Final_test_original_iter, Final_test_gt_iter = load_dataset(train_data, test_data, Final_test, Final_test_original, Final_test_gt, embedding_length, batch_size)
if classifier_name == 'LSTM':
model = LSTM(batch_size, output_size, hidden_size, vocab_size, embedding_length, word_embeddings, pre_train, embedding_tune)
elif classifier_name == 'LSTMAtten':
model = LSTM_AttentionModel(batch_size, output_size, hidden_size, vocab_size, embedding_length, word_embeddings, pre_train, embedding_tune)
elif classifier_name == 'RNN':
model = RNN(batch_size, output_size, hidden_size, vocab_size, embedding_length, word_embeddings, pre_train, embedding_tune)
elif classifier_name == 'RCNN':
model = RCNN(batch_size, output_size, hidden_size, vocab_size, embedding_length, word_embeddings, pre_train, embedding_tune)
elif classifier_name == 'SelfAttention':
model = SelfAttention(batch_size, output_size, hidden_size, vocab_size, embedding_length, word_embeddings, pre_train, embedding_tune)
elif classifier_name == 'CNN':
model = CNN(batch_size, output_size, 1, 32, [2,4,6], 1, 0, 0.6, vocab_size, embedding_length, word_embeddings, pre_train, embedding_tune)
else:
raise ValueError('Not a valid classifier_name!!!')
loss_fn = F.cross_entropy
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=learning_rate[domain][classifier_name])
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 2, gamma=0.1)
return train_iter, test_iter, Final_test_iter, Final_test_original_iter, Final_test_gt_iter, epoch_num, model, loss_fn, optimizer, scheduler
def main(args):
acc_list = []
f1_score_list = []
prec_list = []
recall_list = []
for i in range(10):
setup_data()
model = RCNN(vocab_size=args.vocab_size,
embedding_dim=args.embedding_dim,
hidden_size=args.hidden_size,
hidden_size_linear=args.hidden_size_linear,
class_num=args.class_num,
dropout=args.dropout).to(args.device)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model, dim=0)
train_texts, train_labels = read_file(args.train_file_path)
word2idx, embedding = build_dictionary(train_texts, args.vocab_size,
args.lexical, args.syntactic,
args.semantic)
logger.info('Dictionary Finished!')
full_dataset = CustomTextDataset(train_texts, train_labels, word2idx,
args)
num_train_data = len(full_dataset) - args.num_val_data
train_dataset, val_dataset = random_split(
full_dataset, [num_train_data, args.num_val_data])
train_dataloader = DataLoader(dataset=train_dataset,
collate_fn=lambda x: collate_fn(x, args),
batch_size=args.batch_size,
shuffle=True)
valid_dataloader = DataLoader(dataset=val_dataset,
collate_fn=lambda x: collate_fn(x, args),
batch_size=args.batch_size,
shuffle=True)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
train(model, optimizer, train_dataloader, valid_dataloader, embedding,
args)
logger.info('******************** Train Finished ********************')
# Test
if args.test_set:
test_texts, test_labels = read_file(args.test_file_path)
test_dataset = CustomTextDataset(test_texts, test_labels, word2idx,
args)
test_dataloader = DataLoader(
dataset=test_dataset,
collate_fn=lambda x: collate_fn(x, args),
batch_size=args.batch_size,
shuffle=True)
model.load_state_dict(
torch.load(os.path.join(args.model_save_path, "best.pt")))
_, accuracy, precision, recall, f1, cm = evaluate(
model, test_dataloader, embedding, args)
logger.info('-' * 50)
logger.info(
f'|* TEST SET *| |ACC| {accuracy:>.4f} |PRECISION| {precision:>.4f} |RECALL| {recall:>.4f} |F1| {f1:>.4f}'
)
logger.info('-' * 50)
logger.info('---------------- CONFUSION MATRIX ----------------')
for i in range(len(cm)):
logger.info(cm[i])
logger.info('--------------------------------------------------')
acc_list.append(accuracy / 100)
prec_list.append(precision)
recall_list.append(recall)
f1_score_list.append(f1)
avg_acc = sum(acc_list) / len(acc_list)
avg_prec = sum(prec_list) / len(prec_list)
avg_recall = sum(recall_list) / len(recall_list)
avg_f1_score = sum(f1_score_list) / len(f1_score_list)
logger.info('--------------------------------------------------')
logger.info(
f'|* TEST SET *| |Avg ACC| {avg_acc:>.4f} |Avg PRECISION| {avg_prec:>.4f} |Avg RECALL| {avg_recall:>.4f} |Avg F1| {avg_f1_score:>.4f}'
)
logger.info('--------------------------------------------------')
plot_df = pd.DataFrame({
'x_values': range(10),
'avg_acc': acc_list,
'avg_prec': prec_list,
'avg_recall': recall_list,
'avg_f1_score': f1_score_list
})
plt.plot('x_values',
'avg_acc',
data=plot_df,
marker='o',
markerfacecolor='blue',
markersize=12,
color='skyblue',
linewidth=4)
plt.plot('x_values',
'avg_prec',
data=plot_df,
marker='',
color='olive',
linewidth=2)
plt.plot('x_values',
'avg_recall',
data=plot_df,
marker='',
color='olive',
linewidth=2,
linestyle='dashed')
plt.plot('x_values',
'avg_f1_score',
data=plot_df,
marker='',
color='olive',
linewidth=2,
linestyle='dashed')
plt.legend()
fname = 'lexical-semantic-syntactic.png' if args.lexical and args.semantic and args.syntactic \
else 'semantic-syntactic.png' if args.semantic and args.syntactic \
else 'lexical-semantic.png' if args.lexical and args.semantic \
else 'lexical-syntactic.png'if args.lexical and args.syntactic \
else 'lexical.png' if args.lexical \
else 'syntactic.png' if args.syntactic \
else 'semantic.png' if args.semantic \
else 'plain.png'
if not (path.exists('./images')):
mkdir('./images')
plt.savefig(path.join('./images', fname))
iou_threshold=iou_threshold,
max_samples=max_samples,
verbose=verbose)
# Save it for later.
with open(imagesdata_pickle_path, 'wb') as fi:
pickle.dump(imagesdata, fi)
print('NB CLASSES : ' + str(imagesdata.get_num_classes())) # check the classes
tf.debugging.set_log_device_placement(True)
strategy = tf.distribute.MirroredStrategy() # run on multiple GPUs
with strategy.scope():
arch = RCNN(imagesdata, loss=loss, opt=opt, lr=lr, verbose=verbose)
arch.train(epochs=epochs,
batch_size=batch_size,
split_size=split_size,
checkpoint_path=checkpoint_path,
early_stopping=early_stopping,
verbose=verbose)
arch.model.save(filepath='../data/out_test/test2/model.h5',
save_format='h5')
loss = arch.history()['loss']
val_loss = arch.history()['val_loss']
accuracy = arch.history()['accuracy']
val_accuracy = arch.history()['val_accuracy']
model.train()
return score
if __name__ == "__main__":
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
train_data = pickle.load(open(os.path.join(data_path, train_name), "rb"))
dev_data = pickle.load(open(os.path.join(data_path, dev_name), "rb"))
vocabulary = pickle.load(open(os.path.join(data_path, vocabulary_name), "rb"))
print('dataset', len(train_data), len(dev_data))
# load w2v data
weight = pickle.load(open(os.path.join(data_path, weight_name), "rb"))
# model
train_device = torch.device(device if torch.cuda.is_available() else "cpu")
model = RCNN(vocab_size=len(vocabulary), embed_dim=embed_dim,
output_dim=class_num, hidden_dim=hidden_dim,
num_layers=num_layers, dropout=dropout, weight=weight)
model.to(train_device)
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate, weight_decay=weight_decay)
# train
writer = SummaryWriter(log_dir=log_path)
train()
writer.close()
def train(config, task_name):
train_data = pickle.load(
open(os.path.join(config.data_path, config.train_name), "rb"))
# debug
if config.debug:
train_data = train_data[0:30]
dev_data = pickle.load(
open(os.path.join(config.data_path, config.dev_name), "rb"))
# test_data = pickle.load(open(os.path.join(config.data_path, config.test_name), "rb"))
vocabulary = pickle.load(
open(os.path.join(config.data_path, config.vocabulary_name), "rb"))
# load w2v data
# weight = pickle.load(open(os.path.join(config.data_path, config.weight_name), "rb"))
if task_name == "lstm":
text_model = LSTM(vocab_size=len(vocabulary),
embed_dim=config.embed_dim,
output_dim=config.class_num,
hidden_dim=config.hidden_dim,
num_layers=config.num_layers,
dropout=config.dropout)
elif task_name == "lstm_maxpool":
text_model = LSTM_maxpool(vocab_size=len(vocabulary),
embed_dim=config.embed_dim,
output_dim=config.class_num,
hidden_dim=config.hidden_dim,
num_layers=config.num_layers,
dropout=config.dropout)
elif task_name == "cnn":
text_model = CNN(vocab_size=len(vocabulary),
embed_dim=config.embed_dim,
class_num=config.class_num,
kernel_num=config.kernel_num,
kernel_sizes=config.kernel_sizes,
dropout=config.dropout,
static=config.static,
in_channels=config.in_channels)
elif task_name == "rnn":
text_model = RNN(vocab_size=len(vocabulary),
embed_dim=config.embed_dim,
output_dim=config.class_num,
hidden_dim=config.hidden_dim,
num_layers=config.num_layers,
dropout=config.dropout)
# elif task_name == "cnn_w2v":
# text_model = CNN_w2v(vocab_size=len(vocabulary), embed_dim=config.embed_dim,
# class_num=config.class_num, kernel_num=config.kernel_num,
# kernel_sizes=config.kernel_sizes, dropout=config.dropout,
# static=config.static, in_channels=config.in_channels,
# weight=weight)
elif task_name == "rcnn":
text_model = RCNN(vocab_size=len(vocabulary),
embed_dim=config.embed_dim,
output_dim=config.class_num,
hidden_dim=config.hidden_dim,
num_layers=config.num_layers,
dropout=config.dropout)
#elif task_name == "bert":
# text_model = BertModel.from_pretrained(config.bert_path)
optimizer = Adam(lr=config.lr, weight_decay=config.weight_decay)
timing = TimingCallback()
early_stop = EarlyStopCallback(config.patience)
logs = FitlogCallback(dev_data)
f1 = F1_score(pred='output', target='target')
trainer = Trainer(train_data=train_data,
model=text_model,
loss=BCEWithLogitsLoss(),
batch_size=config.batch_size,
check_code_level=-1,
metrics=f1,
metric_key='f1',
n_epochs=config.epoch,
dev_data=dev_data,
save_path=config.save_path,
print_every=config.print_every,
validate_every=config.validate_every,
optimizer=optimizer,
use_tqdm=False,
device=config.device,
callbacks=[timing, early_stop, logs])
trainer.train()
# test result
tester = Tester(
dev_data,
text_model,
metrics=f1,
device=config.device,
batch_size=config.batch_size,
)
tester.test()
y_va_age = to_categorical(y_va_age)
x_train_current = x_train_age
x_train_left = np.hstack([np.expand_dims(x_train_age[:, 0], axis=1), x_train_age[:, 0:-1]])
x_train_right = np.hstack([x_train_age[:, 1:], np.expand_dims(x_train_age[:, -1], axis=1)])
print('x_train_current 维度:', x_train_current.shape)
print('x_train_left 维度:', x_train_left.shape)
print('x_train_right 维度:', x_train_right.shape)
x_val_current = x_va_age
x_val_left = np.hstack([np.expand_dims(x_va_age[:, 0], axis=1), x_va_age[:, 0:-1]])
x_val_right = np.hstack([x_va_age[:, 1:], np.expand_dims(x_va_age[:, -1], axis=1)])
print('开始RCNN建模......')
max_features = len(word2index) + 1 # 词表的大小
model = RCNN(maxlen, max_features, embedding_dims, 7, 'softmax').get_model()
# 指定optimizer、loss、评估标准
model.compile('adam', 'categorical_crossentropy', metrics=['accuracy'])
print('训练...')
my_callbacks = [
ModelCheckpoint(model_path + 'rcnn_model_age.h5', verbose=1),
EarlyStopping(monitor='val_accuracy', patience=2, mode='max')
]
# fit拟合数据
history = model.fit([x_train_current, x_train_left, x_train_right], y_train_age,
batch_size=batch_size,
epochs=epochs,
callbacks=my_callbacks,
validation_data=([x_val_current, x_val_left, x_val_right], y_va_age))
def train(x_train, y_train, x_test, y_test,params,
train_summary_dir_path,test_summary_dir_path,
checkpoint_dir_path,checkpoint_prefix_path):
# Training
# ==================================================
with tf.Graph().as_default():
sess = tf.Session()
with sess.as_default():
# Define Training procedure
if params['model']=='TextCNN':
model = TextCNN(
sequence_length=params['sequence_length'],
num_classes=params['num_classes'],
vocab_size=params['vocab_size'],
embedding_size=params['embedding_size'],
filter_sizes=list(map(int, params['filter_sizes'].split(","))),
num_filters=params['num_filters'],
l2_reg_lambda=params['l2_reg_lambda'])
elif params['model']=='BiLSTM':
print('train BiLSTM model')
model = BiLSTM(
sequence_length=params['sequence_length'],
num_classes=params['num_classes'],
vocab_size=params['vocab_size'],
embedding_size=params['embedding_size'],
hidden_size=params['hidden_size'],
batch_size=params['batch_size'],
l2_reg_lambda=params['l2_reg_lambda'])
elif params['model']=='Fasttext':
model = Fasttext(sequence_length=params['sequence_length'],
num_classes=params['num_classes'],
vocab_size=params['vocab_size'],
embedding_size=params['embedding_size'],
l2_reg_lambda=params['l2_reg_lambda'])
elif params['model']=='RCNN':
model = RCNN(sequence_length=params['sequence_length'],
num_classes=params['num_classes'],
vocab_size=params['vocab_size'],
embedding_size=params['embedding_size'],
hidden_size=params['hidden_size'],
output_size=params['output_size'],
l2_reg_lambda=params['l2_reg_lambda'])
global_step = tf.Variable(0, name="global_step", trainable=False)
optimizer = tf.train.AdamOptimizer(1e-3)
grads_and_vars = optimizer.compute_gradients(model.loss)
train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step) # 每迭代一个batch,global_step+1
# Keep track of gradient values and sparsity (optional)
grad_summaries = []
for g, v in grads_and_vars:
if g is not None:
grad_hist_summary = tf.summary.histogram("{}/grad/hist".format(v.name), g)
sparsity_summary = tf.summary.scalar("{}/grad/sparsity".format(v.name), tf.nn.zero_fraction(g))
grad_summaries.append(grad_hist_summary)
grad_summaries.append(sparsity_summary)
grad_summaries_merged = tf.summary.merge(grad_summaries)
# Output directory for models and summaries
out_dir = out_dir_path
print("Writing to {}\n".format(out_dir))
# Summaries for loss and accuracy
loss_summary = tf.summary.scalar("loss", model.loss)
acc_summary = tf.summary.scalar("accuracy", model.accuracy)
# Train Summaries
train_summary_op = tf.summary.merge([loss_summary, acc_summary, grad_summaries_merged])
train_summary_dir = train_summary_dir_path
if not os.path.exists(train_summary_dir):
os.makedirs(train_summary_dir)
train_summary_writer = tf.summary.FileWriter(train_summary_dir, sess.graph)
# Test summaries
test_summary_op = tf.summary.merge([loss_summary, acc_summary])
test_summary_dir = test_summary_dir_path
if not os.path.exists(test_summary_dir):
os.makedirs(test_summary_dir)
test_summary_writer = tf.summary.FileWriter(test_summary_dir, sess.graph)
# Checkpoint directory. Tensorflow assumes this directory already exists so we need to create it
checkpoint_dir = checkpoint_dir_path
checkpoint_prefix = checkpoint_prefix_path
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
saver = tf.train.Saver(tf.global_variables(), max_to_keep=params['num_checkpoints'])
# Write vocabulary
# vocab_processor.save(os.path.join(out_dir, "vocab"))
# Initialize all variables
sess.run(tf.global_variables_initializer())
def train_step(x_batch, y_batch):
"""
A single training step
"""
feed_dict = {
model.input_x: x_batch,
model.input_y: y_batch,
model.dropout_keep_prob: params['dropout_keep_prob']
}
_, step, summaries, loss, accuracy = sess.run(
[train_op, global_step, train_summary_op, model.loss, model.accuracy],
feed_dict)
time_str = datetime.datetime.now().isoformat()
res['step'].append(step)
res['loss'].append(loss)
res['acc'].append(accuracy)
print("{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy))
train_summary_writer.add_summary(summaries, step)
def test_step(x_batch, y_batch, writer=None):
"""
Evaluates model on a dev set
"""
feed_dict = {
model.input_x: x_batch,
model.input_y: y_batch,
model.dropout_keep_prob: 1.0
}
step, summaries, loss, accuracy = sess.run(
[global_step, test_summary_op, model.loss, model.accuracy],
feed_dict)
time_str = datetime.datetime.now().isoformat()
print("{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy))
if writer:
writer.add_summary(summaries, step)
# Generate batches
batches = batch_iter(
list(zip(x_train, y_train)), params['batch_size'], params['num_epochs'])
# Training loop. For each batch...
for batch in batches:
x_batch, y_batch = zip(*batch)
train_step(x_batch, y_batch)
current_step = tf.train.global_step(sess, global_step)
# if current_step % params['evaluate_every'] == 0:
# print("\nEvaluation:")
# test_step(x_test, y_test, writer=test_summary_writer)
# print("")
if current_step % params['checkpoint_every'] == 0:
path = saver.save(sess, checkpoint_prefix, global_step=current_step)
print("Saved model checkpoint to {}\n".format(path))
if current_step==10000:
return
import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from model import RCNN from dataloader_mnist import dataloader,batch_size,test_dataset_len,train_dataset_len n_classes = 10 net = RCNN(n_classes=n_classes) learning_rate = 1e-3 epoch = 30 criterion = nn.CrossEntropyLoss() # optimizer = optim.SGD(net.parameters(), lr=learning_rate, momentum=0.9) optimizer = optim.Adam(net.parameters(), lr=learning_rate) # scheduler = optim.lr_scheduler.ReduceLROnPlateau( # optimizer, 'min' , # factor=0.1 , # patience=(train_dataset_len/batch_size)*3, # verbose=True) use_gpu = torch.cuda.is_available() if use_gpu: net = net.cuda()
def train_RCNN(args):
train_x, train_y, n_items = load_train(args.max_len)
args.n_items = n_items
data = list(zip(train_x, train_y))
random.shuffle(data)
train_x, train_y = zip(*data)
num_batches = len(train_x) // args.batch_size
global valid_x
global valid_y
valid_x, valid_y, _ = load_valid(args.max_len)
print('#Items: {}'.format(n_items))
print('#Training Nums: {}'.format(len(train_x)))
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
with tf.Session(config=gpu_config) as sess:
model = RCNN(args)
if args.is_store:
saver = tf.train.Saver(tf.global_variables())
ckpt = tf.train.get_checkpoint_state(args.checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print('Restore model from {} successfully!'.format(
ckpt.model_checkpoint_path))
else:
print('Restore model from {} failed!'.format(
args.checkpoint_dir))
return
else:
sess.run(tf.global_variables_initializer())
best_epoch = -1
best_step = -1
best_loss = np.inf
best_HR = np.inf
max_stay, stay_cnt = 20, 0
losses = 0.0
for epoch in range(args.epochs):
for i in range(num_batches):
x = train_x[i * args.batch_size:(i + 1) * args.batch_size]
y = train_y[i * args.batch_size:(i + 1) * args.batch_size]
fetches = [
model.sum_loss, model.global_step, model.lr, model.train_op
]
feed_dict = {model.X: x, model.Y: y}
loss, step, lr, _ = sess.run(fetches, feed_dict)
losses += loss
if step % 50 == 0:
print('Epoch-{}\tstep-{}\tlr:{:.6f}\tloss: {:.6f}'.format(
epoch + 1, step, lr, losses / 50))
losses = 0.0
if step % 1000 == 0:
valid_loss, HR, NDCG, MRR = eval_validation(
model, sess, args.batch_size)
print(
'step-{}\teval_validation\tloss:{:6f}\tHR@{}:{:.6f}\tNDCG@{}:{:.6f}\tMRR@{}:{:.6f}'
.format(step, valid_loss, args.top_k, HR, args.top_k,
NDCG, args.top_k, MRR))
if HR > best_HR or (valid_loss < best_loss and HR > 0.0):
best_HR = HR
best_loss = valid_loss
best_epoch = epoch + 1
best_step = step
stay_cnt = 0
ckpt_path = args.checkpoint_dir + 'model.ckpt'
model.saver.save(sess, ckpt_path, global_step=step)
print("model saved to {}".format(ckpt_path))
else:
stay_cnt += 1
if stay_cnt >= max_stay:
break
if stay_cnt >= max_stay:
break
print("best model at:epoch-{}\tstep-{}\tloss:{:.6f}\tHR@{}:{:.6f}".
format(best_epoch, best_step, best_loss, args.top_k, best_HR))
def main():
#
print('\nRunnig fold: ' + sys.argv[1])
kfold_ = int(sys.argv[1]) # only train for one cross-validation fold at a time (this way we can train all folds in parallel)
print(type(kfold_))
# load data
tr_fact = 1 # 1 is 100% data for training
out_dir = './results/rcnn_merge_time_coch_cval10_brain_hfb/' + \
'n_back_6_cnnT_300h_cnnF_100h_rnn_300h_alt_alt2_concattest_train' + str(int(tr_fact * 100)) + '/'
x1_file = './data/M3_audio_mono_down.wav'
x2_file = './data/minoes_wav_freq_125Hz_abs.npy'
t_file = './data/minoes_hfb_6subjs.npy'
xtr1 = librosa.load(x1_file, sr=8000)[0]
xtr2 = np.load(x2_file).astype(np.float32)
ttr = np.load(t_file).astype(np.float32)
print('Train data: ' + str(int(tr_fact * 100)) + '%')
# resample brain and spectrogram data to 50 Hz
xtr2 = resample(xtr2, sr1=50, sr2=125)
ttr = resample(ttr, sr1=50, sr2=125)
# take a sample in sec
global sr1, sr2, sr3, n_back
sr1 = 8000
sr2 = 50
sr3 = 50
nsec = ttr.shape[0] / float(sr2)
nsamp = nsec * 1
n2 = int(nsamp * sr2)
n3 = int(nsamp * sr3)
xtr2 = xtr2[:n2]
ttr = ttr[:n3]
# cut raw audio to match brain data (ttr) length in sec
n1 = int(nsamp * sr1)
xtr1 = xtr1[:n1]
xtr1 = xtr1[:, None]
# set up cross-validation for performance accuracy: set-up the same way for all folds when folds are trained separately
kfolds = 10
nparts = 7 # test set is not a continuous chunk but is a concatenation of nparts fragments for better performance
ind1 = np.arange(xtr1.shape[0])
ind2 = np.arange(ttr.shape[0])
ind3 = np.arange(ttr.shape[0])
TestI_, TestI = [], []
kf = KFold(n_splits=kfolds * nparts)
for (_, ix1_test), (_, ix2_test), (_, it_test) in zip(kf.split(xtr1), kf.split(xtr2), kf.split(ttr)):
TestI_.append([ix1_test, ix2_test, it_test])
for kfold in range(kfolds):
TestI.append([np.array(
[item for sublist in [TestI_[i][j] for i in range(0 + kfold, kfolds * nparts + kfold, kfolds)] for item in
sublist])
for j in range(len(TestI_[0]))])
if (out_dir is not None) & (not os.path.exists(out_dir)): os.makedirs(out_dir)
process = psutil.Process(os.getpid())
print(process.memory_info().rss / 1024 / 1024 / 1024)
# standard sklearn preprocessing of data
scaler = Scaler()
kfold = kfold_
ktrain, ktest, _ = scaler([xtr1[np.setdiff1d(ind1, TestI[kfold][0])], xtr2[np.setdiff1d(ind2, TestI[kfold][1])], ttr[np.setdiff1d(ind3, TestI[kfold][2])]],
[xtr1[TestI[kfold][0]], xtr2[TestI[kfold][1]], ttr[TestI[kfold][2]]], None)
nsec_tr = ktrain[-1].shape[0] / float(sr2)
nsamp_tr = nsec_tr * tr_fact
ktrain = map(lambda x, n: x.copy()[:n], ktrain, [int(nsamp_tr *i) for i in [sr1, sr2, sr3]])
print(map(len, ktrain))
print(map(len, ktest))
# model parameters
dur = 1 # sec units
batch_size = 16
n_back = 6 * dur # in dur units, temporal window of input data (how much data the model sees at once)
nepochs = 30
n_out = ttr.shape[-1]
alpha = 5e-04
h_cnn_t = 300 # number of hidden units on top layer of CNN time
h_cnn_f = 100 # number of hidden units on top layer of CNN freq/spectra
h_rnn = 300 # number of hidden units of RNN
print('batch size: ' + str(batch_size) + ', nepochs: ' + str(nepochs) + ', lr: ' + str(alpha) +
', h_cnn_t: ' + str(h_cnn_t) + ', h_cnn_f: ' + str(h_cnn_f) + ', h_rnn: ' + str(h_rnn))
print('outdir: ' + out_dir)
# set up model
rcnn = RCNN(h_cnn_t, h_cnn_f, h_rnn, n_out)
opt = chainer.optimizers.Adam(alpha)
opt.setup(rcnn)
with open(out_dir + 'fold' + str(kfold) + '_run.log', 'wb'): pass # running epoch and best performance are saved to txt file for bookkeeping
with open(out_dir + 'fold' + str(kfold) + '_epoch.txt', 'wb'): pass
# train loop
best_acc = -1
for epoch in range(nepochs):
print('Epoch ' + str(epoch))
with open(out_dir + 'fold' + str(kfold) + '_run.log', 'a') as fid0:
fid0.write('epoch' + str(epoch) + '\n')
rcnn.reset_state()
x1, x2, t = roll_data(ktrain, [.14 * epoch * sr for sr in [sr1, sr2, sr3]])
x1, x2, t = prepare_input([x1, x2, t], [sr1, sr2, sr3], n_back)
xbs1, xbs2, tbs = get_batches([x1, x2, t], batch_size)
print(process.memory_info().rss / 1024 / 1024 / 1024)
for ib, (xb1, xb2, tb) in enumerate(zip(xbs1, xbs2, tbs)):
with chainer.using_config('train', True):
y = rcnn([np.expand_dims(xb1, 1), np.expand_dims(xb2, 1)], n_back)
loss = 0
for ni in range(y.shape[1]):
loss += F.mean_squared_error(tb[:, ni, :], y[:, ni, :])
r = acc_pass(tb.reshape((-1, n_out)), y.data.reshape((-1, n_out)))
print('\t\tbatch ' + str(ib) + ', train loss: ' + str(loss.data / tb.shape[1]) + ', max acc: ' + str(np.max(r)))
rcnn.cleargrads()
loss.backward()
loss.unchain_backward()
opt.update()
xb1_, xb2_, tb_ = prepare_input(ktest, [sr1, sr2, sr3], n_back)
rcnn.reset_state()
with chainer.using_config('train', False):
y_ = rcnn([np.expand_dims(xb1_, 1), np.expand_dims(xb2_, 1)], n_back)
loss_ = 0
for ni in range(y_.shape[1]):
loss_ += F.mean_squared_error(tb_[:, ni, :], y_[:, ni, :])
r = acc_pass(tb_.reshape((-1, n_out)), y_.data.reshape((-1, n_out)))
print('\t\ttest loss: ' + str(np.round(loss_.data / tb_.shape[1], 3)) + ', max acc: ' + str(
np.round(np.sort(r)[::-1][:10], 4)))
run_acc = np.mean(np.sort(r)[::-1][:10])
if run_acc > best_acc: # only if performance of current model is superior, save it to file
print('Current model is best: ' + str(np.round(run_acc, 4)) + ' > ' + str(
np.round(best_acc, 4)) + ': saving update to disk')
best_acc = run_acc.copy()
serializers.save_npz(out_dir + '/model' + str(kfold) + '.npz', rcnn)
with open(out_dir + 'fold' + str(kfold) + '_epoch.txt', 'a') as fid:
fid.write(str(epoch) + '\n')
fid.write(str(np.sort(r)[::-1][:10]) + '\n')
np.save(out_dir + '/predictions_fold' + str(kfold), y_.data.reshape((-1, n_out)))
np.save(out_dir + '/targets_fold' + str(kfold), tb_.reshape((-1, n_out)))
def load_model(test_arguments):
rcnn = RCNN(test_arguments.pos_loss_method, test_arguments.loss_weight_lambda).cuda()
rcnn.load_state_dict(t.load(test_arguments.model_path))
rcnn.eval() # dropout rate = 0
return rcnn
################################################################################
#command: main python main.py train/inference --weights=coco/last --image=link
#tensorboard --logdir=log_dir
################################################################################
parser = argparse.ArgumentParser()
parser.add_argument("command")
parser.add_argument("--weights", required=True)
parser.add_argument("--image", required=False)
parser.add_argument("--video", required=False)
args = parser.parse_args()
config = Config()
if args.command == "train":
model = RCNN(mode="train", config=config)
else:
model = RCNN(mode="inference", config=config)
#load resnet101 pretrained model
isCoco = 0
if args.weights == "coco":
weight_path = COCO_WEIGHTS
isCoco = 1
print(weight_path)
elif args.weights == "last":
weight_path = model.find_last()
print(weight_path)
else:
weight_path = args.weights
def load_models(config):
# train_data = pickle.load(open(os.path.join(config.data_path, config.train_name), "rb"))
# debug
# if config.debug:
# train_data = train_data[0:30]
# dev_data = pickle.load(open(os.path.join(config.data_path, config.dev_name), "rb"))
# test_data = pickle.load(open(os.path.join(config.data_path, config.test_name), "rb"))
vocabulary = pickle.load(
open(os.path.join(config.data_path, config.vocabulary_name), "rb"))
# load w2v data
# weight = pickle.load(open(os.path.join(config.data_path, config.weight_name), "rb"))
cnn = CNN(vocab_size=len(vocabulary),
embed_dim=config.embed_dim,
class_num=config.class_num,
kernel_num=config.kernel_num,
kernel_sizes=config.kernel_sizes,
dropout=config.dropout,
static=config.static,
in_channels=config.in_channels)
state_dict = torch.load(
os.path.join(config.save_path,
config.ensemble_models[0])).state_dict()
cnn.load_state_dict(state_dict)
lstm = LSTM(vocab_size=len(vocabulary),
embed_dim=config.embed_dim,
output_dim=config.class_num,
hidden_dim=config.hidden_dim,
num_layers=config.num_layers,
dropout=config.dropout)
state_dict = torch.load(
os.path.join(config.save_path,
config.ensemble_models[1])).state_dict()
lstm.load_state_dict(state_dict)
lstm_mxp = LSTM_maxpool(vocab_size=len(vocabulary),
embed_dim=config.embed_dim,
output_dim=config.class_num,
hidden_dim=config.hidden_dim,
num_layers=config.num_layers,
dropout=config.dropout)
state_dict = torch.load(
os.path.join(config.save_path,
config.ensemble_models[2])).state_dict()
lstm_mxp.load_state_dict(state_dict)
rcnn = RCNN(vocab_size=len(vocabulary),
embed_dim=config.embed_dim,
output_dim=config.class_num,
hidden_dim=config.hidden_dim,
num_layers=config.num_layers,
dropout=config.dropout)
state_dict = torch.load(
os.path.join(config.save_path,
config.ensemble_models[3])).state_dict()
rcnn.load_state_dict(state_dict)
schemas = get_schemas(config.source_path)
state_dict = torch.load(
os.path.join(config.save_path,
config.ensemble_models[4])).state_dict()
bert = BertForMultiLabelSequenceClassification.from_pretrained(
config.bert_folder, state_dict=state_dict, num_labels=len(schemas))
bert.load_state_dict(state_dict)
return cnn, lstm, lstm_mxp, rcnn, bert