def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
seed = 1111
set_seed(seed)
#### get data
data_obj = _DATA()
train_data, valid_data, vocab_obj = data_obj.f_load_data_yelp(args)
# train_data, valid_data = data()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("device", device)
if args.train:
now_time = datetime.datetime.now()
time_name = str(now_time.month)+"_"+str(now_time.day)+"_"+str(now_time.hour)+"_"+str(now_time.minute)
model_file = os.path.join(args.model_path, args.data_name+"_"+args.model_name)
if not os.path.isdir(model_file):
print("create a directory ", model_file)
os.mkdir(model_file)
args.model_file = model_file+"/model_best_"+time_name+".pt"
print("model_file", model_file)
print("vocab_size", vocab_obj.vocab_size)
print("user num", vocab_obj.user_size)
### get model
network = _NETWORK(vocab_obj, args, device=device)
### add count parameters
total_param_num = 0
for name, param in network.named_parameters():
if param.requires_grad:
param_num = param.numel()
total_param_num += param_num
print(name, "\t", param_num)
print("total parameters num", total_param_num)
if args.train:
logger_obj = _LOGGER()
logger_obj.f_add_writer(args)
# if torch.cuda.device_count() > 1:
# print("... let us use", torch.cuda.device_count(), "GPUs!")
# network = nn.DataParallel(network)
# print("=="*20)
# print("device", network.cuda())
# en_parameters = list(network.module.m_embedding.parameters()) + list(network.module.m_user_item_encoder.parameters()) + list(network.module.m_output2vocab.parameters())
# en_optimizer = _OPTIM(en_parameters, args)
# de_parameters = network.module.m_generator.parameters()
# de_optimizer = _OPTIM(de_parameters, args)
en_parameters = list(network.m_embedding.parameters()) + list(network.m_user_item_encoder.parameters()) + list(network.m_output2vocab.parameters())
en_optimizer = _OPTIM(en_parameters, args)
de_parameters = network.m_generator.parameters()
de_optimizer = _OPTIM(de_parameters, args)
trainer = _TRAINER(vocab_obj, args, device)
trainer.f_train(train_data, valid_data, network, en_optimizer, de_optimizer, logger_obj)
logger_obj.f_close_writer()
if args.test:
print("="*10, "test", "="*10)
infer_obj = _INFER(vocab_obj, args, device)
infer_obj.f_init_infer(network, args.model_file, reload_model=True)
infer_obj.f_inference(train_data, valid_data)
if args.eval:
print("="*10, "eval", "="*10)
eval_obj = _EVAL(vocab_obj, args, device)
eval_obj.f_init_eval(network, args.model_file, reload_model=True)
eval_obj.f_eval(train_data, valid_data)
de_optimizer = _OPTIM(de_parameters, args)
# print("=="*20)
# print("generator parameter")
# for name, p in network.m_generator.named_parameters():
# print(name)
# print("=="*20)
# exit()
trainer = _TRAINER(vocab_obj, args, device)
trainer.f_train(train_data, valid_data, pretrain_encoder, pretrain_optimizer, network, en_optimizer, de_optimizer, logger_obj)
logger_obj.f_close_writer()
if args.test:
print("="*10, "test", "="*10)
infer_obj = _INFER(vocab_obj, args, device)
infer_obj.f_init_infer(network, args.model_file, reload_model=True)
infer_obj.f_inference(train_data, valid_data)
if args.eval:
print("="*10, "eval", "="*10)
eval_obj = _EVAL(vocab_obj, args, device)
eval_obj.f_init_eval(network, args.model_file, reload_model=True)
eval_obj.f_eval(train_data, valid_data)
if __name__ == "__main__":
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
seed = 1111
set_seed(seed)
#### get data
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("device", device)
if args.parallel:
local_rank = args.local_rank
torch.distributed.init_process_group(backend="nccl")
device = torch.device('cuda:{}'.format(local_rank))
data_obj = _DATA()
train_data, valid_data, vocab_obj = data_obj.f_load_data_yelp(args)
# train_data, valid_data = data()
if args.train:
now_time = datetime.datetime.now()
time_name = str(now_time.month) + "_" + str(now_time.day) + "_" + str(
now_time.hour) + "_" + str(now_time.minute)
model_file = os.path.join(args.model_path,
args.data_name + "_" + args.model_name)
if not os.path.isdir(model_file):
print("create a directory ", model_file)
os.mkdir(model_file)
args.model_file = model_file + "/model_best_" + time_name + ".pt"
print("model_file", model_file)
print("vocab_size", vocab_obj.vocab_size)
print("user num", vocab_obj.user_size)
### get model
network = _GEN_NETWORK(vocab_obj, args)
### add count parameters
total_param_num = 0
for name, param in network.named_parameters():
if param.requires_grad:
param_num = param.numel()
total_param_num += param_num
print(name, "\t", param_num)
print("total parameters num", total_param_num)
if args.train:
logger_obj = _LOGGER()
logger_obj.f_add_writer(args)
E_network = _ENC_NETWORK(vocab_obj, args)
E_network = E_network.to(device)
# E_network = torch.nn.parallel.DistributedDataParallel(E_network, device_ids=[local_rank], output_device=local_rank, find_unused_parameters=True)
# torch.distributed.barrier()
# map_location = {'cuda:%d'%0:'cuda:%d'%local_rank}
model_path = args.model_path
E_model_file = args.E_model_file
E_model_abs_file = os.path.join(model_path, E_model_file)
print("E_model_abs_file", E_model_abs_file)
check_point = torch.load(E_model_abs_file)
# check_point = torch.load(E_model_abs_file, map_location=map_location)
E_network.load_state_dict(check_point['model'])
# torch.distributed.barrier()
network = network.to(device)
if args.parallel:
network = torch.nn.parallel.DistributedDataParallel(
network,
device_ids=[local_rank],
output_device=local_rank,
find_unused_parameters=True)
de_parameters = network.parameters()
de_optimizer = _OPTIM(de_parameters, args)
trainer = _TRAINER(vocab_obj, args, device)
trainer.f_train_M(train_data, valid_data, E_network, network,
de_optimizer, logger_obj, local_rank)
logger_obj.f_close_writer()
if args.test:
print("=" * 10, "test", "=" * 10)
infer_obj = _INFER(vocab_obj, args, device)
infer_obj.f_init_infer(network, args.model_file, reload_model=True)
infer_obj.f_inference(train_data, valid_data)
if args.eval:
print("=" * 10, "eval", "=" * 10)
eval_obj = _EVAL(vocab_obj, args, device)
eval_obj.f_init_eval(network, args.model_file, reload_model=True)
eval_obj.f_eval(train_data, valid_data)
def main(args):
ts = time.strftime('%Y-%b-%d-%H:%M:%S', time.gmtime())
seed = 1111
set_seed(seed)
#### get data
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("device", device)
args.decoder_model_type = args.decoder_model_type.lower()
global_step = args.global_step_eval
print("checkpoint dir", args.checkpoint_dir)
output_decoder_dir = os.path.join(args.checkpoint_dir, "checkpoint-decoder-{}".format(global_step))
output_full_dir = os.path.join(args.checkpoint_dir, "checkpoint-full-{}".format(global_step))
print("output_decoder_dir: ", output_decoder_dir)
print("output_full_dir: ", output_full_dir)
checkpoints = [[output_decoder_dir]]
MODEL_CLASSES = {'gpt2': (GPT2Config, GPT2ForLatentConnector, GPT2Tokenizer)}
decoder_config_class, decoder_model_class, decoder_tokenizer_class = MODEL_CLASSES[args.decoder_model_type]
model_decoder = decoder_model_class.from_pretrained(output_decoder_dir, latent_size=args.latent_size)
tokenizer_decoder = decoder_tokenizer_class.from_pretrained(args.decoder_tokenizer_name if args.decoder_tokenizer_name else args.decoder_model_name_or_path, do_lower_case=args.do_lower_case)
print("decoder_tokenizer_name ", args.decoder_tokenizer_name)
print("decoder_model_name_or_path ", args.decoder_model_name_or_path)
model_decoder.to(device)
if args.block_size <= 0:
args.block_size = tokenizer_decoder.max_len_single_sentence
print("max_len_single_sentence: ", tokenizer_decoder.max_len_single_sentence)
args.block_size = min(args.block_size, tokenizer_decoder.max_len_single_sentence)
print("block size: ", args.block_size)
checkpoint = torch.load(os.path.join(output_full_dir, "training.bin"))
special_tokens_dict = {'pad_token': '<PAD>', 'bos_token': '<BOS>', 'eos_token': '<EOS>'}
num_added_toks = tokenizer_decoder.add_special_tokens(special_tokens_dict)
print('We have added', num_added_toks, 'tokens to GPT2')
model_decoder.resize_token_embeddings(len(tokenizer_decoder))
assert tokenizer_decoder.pad_token == '<PAD>'
data_obj = _DATA()
train_data, valid_data, vocab_obj = data_obj.f_load_data_yelp_GPT(tokenizer_decoder, args)
# train_data, valid_data = data()
if args.train:
now_time = datetime.datetime.now()
time_name = str(now_time.month)+"_"+str(now_time.day)+"_"+str(now_time.hour)+"_"+str(now_time.minute)
model_file = os.path.join(args.model_path, args.data_name+"_"+args.model_name)
if not os.path.isdir(model_file):
print("create a directory ", model_file)
os.mkdir(model_file)
args.model_file = model_file+"/model_best_"+time_name+".pt"
print("model_file", model_file)
network = _GEN_NETWORK(vocab_obj, args)
### add count parameters
total_param_num = 0
for name, param in network.named_parameters():
if param.requires_grad:
param_num = param.numel()
total_param_num += param_num
print(name, "\t", param_num)
print("total parameters num", total_param_num)
if args.train:
logger_obj = _LOGGER()
logger_obj.f_add_writer(args)
E_network = _ENC_NETWORK(vocab_obj, args)
E_network = E_network.to(device)
# E_network = torch.nn.parallel.DistributedDataParallel(E_network, device_ids=[local_rank], output_device=local_rank, find_unused_parameters=True)
# torch.distributed.barrier()
# map_location = {'cuda:%d'%0:'cuda:%d'%local_rank}
model_path = args.model_path
# E_model_file = args.E_model_file
# E_model_abs_file = os.path.join(model_path, E_model_file)
# print("E_model_abs_file", E_model_abs_file)
# check_point = torch.load(E_model_abs_file)
# check_point = torch.load(E_model_abs_file, map_location=map_location)
# E_network.load_state_dict(check_point['model'])
# if args.user_pretrained_model:
# pre_model = check_point['model_state_dict']
# model_dict = network.state_dict()
# pre_dict = {k:v for k, v in pre_model.items() if k in model_dict}
# model_dict.update(pre_dict)
# network.load_state_dict(model_dict)
network.init_tokenizer_decoder(tokenizer_decoder, model_decoder)
network = network.to(device)
no_decay = ['bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in network.named_parameters() if not any(nd in n for nd in no_decay)],'weight_decay':args.weight_decay },
{'params': [p for n, p in network.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay':0.0}
]
t_total = len(train_data) // args.gradient_accumulation_steps * args.num_train_epochs
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
local_rank = 0
trainer = _TRAINER(vocab_obj, args, device)
trainer.f_train_M(train_data, valid_data, E_network, network, optimizer, scheduler, logger_obj, local_rank, args)
logger_obj.f_close_writer()
if args.test:
print("="*10, "test", "="*10)
infer_obj = _INFER(vocab_obj, args, device)
infer_obj.f_init_infer(network, args.model_file, reload_model=True)
infer_obj.f_inference(train_data, valid_data)
if args.eval:
print("="*10, "eval", "="*10)
eval_obj = _EVAL(vocab_obj, args, device)
eval_obj.f_init_eval(network, args.model_file, reload_model=True)
eval_obj.f_eval(train_data, valid_data)