def train(args):
print('Training:')
train_data = data_utils.load_dataset(args.train_dataset, args)
train_data_size = len(train_data)
if args.train_proportion < 1.0:
random.shuffle(train_data)
train_data_size = int(train_data_size * args.train_proportion)
train_data = train_data[:train_data_size]
eval_data = data_utils.load_dataset(args.val_dataset, args)
DataProcessor = data_utils.jspDataProcessor(args)
model_supervisor = create_model(args)
if args.resume:
resume_step = True
else:
resume_step = False
resume_idx = args.resume * args.batch_size
logger = model_utils.Logger(args)
if args.resume:
logs = pd.read_csv("../logs/" + args.log_name)
for index, log in logs.iterrows():
val_summary = {
'avg_reward': log['avg_reward'],
'global_step': log['global_step']
}
logger.write_summary(val_summary)
for epoch in range(resume_idx // train_data_size, args.num_epochs):
random.shuffle(train_data)
for batch_idx in range(0 + resume_step * resume_idx % train_data_size,
train_data_size, args.batch_size):
resume_step = False
print(epoch, batch_idx)
batch_data = DataProcessor.get_batch(train_data, args.batch_size,
batch_idx)
train_loss, train_reward = model_supervisor.train(batch_data)
print('train loss: %.4f train reward: %.4f' %
(train_loss, train_reward))
if (args.resume +
model_supervisor.global_step) % args.eval_every_n == 0:
eval_loss, eval_reward = model_supervisor.eval(
eval_data, args.output_trace_flag, args.max_eval_size)
val_summary = {
'avg_reward': eval_reward,
'global_step': args.resume + model_supervisor.global_step
}
logger.write_summary(val_summary)
model_supervisor.save_model()
if args.lr_decay_steps is not None and (
args.resume +
model_supervisor.global_step) % args.lr_decay_steps == 0:
model_supervisor.model.lr_decay(args.lr_decay_rate)
if model_supervisor.model.cont_prob > 0.01:
model_supervisor.model.cont_prob *= 0.5
def train(args):
print('Training:')
data_processor = data_utils.DataProcessor(args)
train_data = data_processor.load_data(args.train_dataset)
train_data, train_indices = data_processor.preprocess(train_data)
dev_data = data_processor.load_data(args.dev_dataset)
dev_data, dev_indices = data_processor.preprocess(dev_data)
train_data_size = len(train_data)
args.word_vocab_size = data_processor.word_vocab_size
args.code_vocab_size = data_processor.code_vocab_size
model_supervisor = create_model(args, data_processor.word_vocab,
data_processor.code_vocab)
logger = model_utils.Logger(args)
for epoch in range(args.num_epochs):
random.shuffle(train_data)
for batch_idx in range(0, train_data_size, args.batch_size):
print(epoch, batch_idx)
batch_input, batch_labels = data_processor.get_batch(
train_data, args.batch_size, batch_idx)
train_loss, train_acc = model_supervisor.train(
batch_input, batch_labels)
print('train loss: %.4f train acc: %.4f' % (train_loss, train_acc))
if model_supervisor.global_step % args.eval_every_n == 0:
model_supervisor.save_model()
eval_loss, eval_label_acc, eval_data_acc, eval_acc, pred_labels = model_supervisor.eval(
dev_data, args.data_order_invariant, args.max_eval_size)
val_summary = {
'train_loss': train_loss,
'train_acc': train_acc,
'eval_loss': eval_loss,
'eval_label_acc': eval_label_acc,
'eval_data_acc': eval_data_acc,
'eval_acc': eval_acc
}
val_summary['global_step'] = model_supervisor.global_step
logger.write_summary(val_summary)
if args.lr_decay_steps is not None and model_supervisor.global_step % args.lr_decay_steps == 0:
model_supervisor.model.lr_decay(args.lr_decay_rate)
def train(args):
print('Training:')
train_data = data_utils.load_dataset(args.train_dataset, args)
eval_data = data_utils.load_dataset(args.val_dataset, args)
DataProcessor = data_utils.HalideDataProcessor()
if args.train_proportion < 1.0:
random.shuffle(train_data)
train_data_size = int(train_data_size * args.train_proportion)
train_data = train_data[:train_data_size]
if args.train_max_len is not None:
train_data = DataProcessor.prune_dataset(train_data, max_len=args.train_max_len)
train_data_size = len(train_data)
term_vocab, term_vocab_list = DataProcessor.load_term_vocab()
op_vocab, op_vocab_list = DataProcessor.load_ops()
args.term_vocab_size = len(term_vocab)
args.op_vocab_size = len(op_vocab)
model_supervisor = create_model(args, term_vocab, term_vocab_list, op_vocab, op_vocab_list)
logger = model_utils.Logger(args)
for epoch in range(args.num_epochs):
random.shuffle(train_data)
for batch_idx in range(0, train_data_size, args.batch_size):
print(epoch, batch_idx)
batch_data = DataProcessor.get_batch(train_data, args.batch_size, batch_idx)
train_loss, train_reward = model_supervisor.train(batch_data)
print('train loss: %.4f train reward: %.4f' % (train_loss, train_reward))
if model_supervisor.global_step % args.eval_every_n == 0:
eval_loss, eval_reward = model_supervisor.eval(eval_data, args.output_trace_flag, args.output_trace_option, args.output_trace_file, args.max_eval_size)
val_summary = {'avg_reward': eval_reward}
model_supervisor.save_model()
val_summary['global_step'] = model_supervisor.global_step
logger.write_summary(val_summary)
if args.lr_decay_steps is not None and model_supervisor.global_step % args.lr_decay_steps == 0:
model_supervisor.model.lr_decay(args.lr_decay_rate)
if model_supervisor.model.cont_prob > 0.01:
model_supervisor.model.cont_prob *= 0.5
