def initialize(self, ctx):
self.manifest = ctx.manifest
properties = ctx.system_properties
model_dir = properties.get("model_dir")
serialized_file = self.manifest["model"]["serializedFile"]
model_pt_path = os.path.join(model_dir, serialized_file)
self.device = torch.device("cuda:" +
str(properties.get("gpu_id")) if torch.cuda.
is_available() else "cpu")
# Read model serialize/pt file
self.tokenizer = BertTokenizer.from_pretrained(
self.config.model_name_or_path)
self.processor = TRADEPreprocessor(self.slot_meta, self.tokenizer)
tokenized_slot_meta = []
for slot in self.slot_meta:
tokenized_slot_meta.append(
self.tokenizer.encode(slot.replace("-", " "),
add_special_tokens=False))
self.model = TRADE(self.config, tokenized_slot_meta)
ckpt = torch.load(model_pt_path, map_location="cpu")
self.model.load_state_dict(ckpt)
self.model.to(self.device)
print("Model is loaded")
self.initialized = True
# Data Loading
train_data_file = f"{args.data_dir}/train_dials.json"
slot_meta = json.load(open(f"{args.data_dir}/slot_meta.json"))
train_data, dev_data, dev_labels = load_dataset(train_data_file)
train_examples = get_examples_from_dialogues(
train_data, user_first=False, dialogue_level=False
)
dev_examples = get_examples_from_dialogues(
dev_data, user_first=False, dialogue_level=False
)
# Define Preprocessor
tokenizer = BertTokenizer.from_pretrained(args.model_name_or_path)
processor = TRADEPreprocessor(slot_meta, tokenizer)
args.vocab_size = len(tokenizer)
args.n_gate = len(processor.gating2id) # gating 갯수 none, dontcare, ptr
# Extracting Featrues
train_features = processor.convert_examples_to_features(train_examples)
dev_features = processor.convert_examples_to_features(dev_examples)
# Slot Meta tokenizing for the decoder initial inputs
tokenized_slot_meta = []
for slot in slot_meta:
tokenized_slot_meta.append(
tokenizer.encode(slot.replace("-", " "), add_special_tokens=False)
)
# Model 선언
parser.add_argument("--model_dir", type=str, default=None)
parser.add_argument("--output_dir", type=str, default=None)
parser.add_argument("--eval_batch_size", type=int, default=32)
args = parser.parse_args()
args.data_dir = os.environ['SM_CHANNEL_EVAL']
args.model_dir = os.environ['SM_CHANNEL_MODEL']
args.output_dir = os.environ['SM_OUTPUT_DATA_DIR']
model_dir_path = os.path.dirname(args.model_dir)
eval_data = json.load(open(f"{args.data_dir}/eval_dials.json", "r"))
config = json.load(open(f"{model_dir_path}/exp_config.json", "r"))
config = argparse.Namespace(**config)
slot_meta = json.load(open(f"{model_dir_path}/slot_meta.json", "r"))
tokenizer = BertTokenizer.from_pretrained(config.model_name_or_path)
processor = TRADEPreprocessor(slot_meta, tokenizer)
eval_examples = get_examples_from_dialogues(
eval_data, user_first=False, dialogue_level=False
)
# Extracting Featrues
eval_features = processor.convert_examples_to_features(eval_examples)
eval_data = WOSDataset(eval_features)
eval_sampler = SequentialSampler(eval_data)
eval_loader = DataLoader(
eval_data,
batch_size=args.eval_batch_size,
sampler=eval_sampler,
collate_fn=processor.collate_fn,
)
# args.data_dir = os.environ['SM_CHANNEL_TRAIN']
# args.model_dir = os.environ['SM_MODEL_DIR']
output_dir = increment_output_dir(args.model_dir)
# random seed 고정
set_seed(args.random_seed)
# Data Loading
train_data_file = f"{args.data_dir}/train_dials.json"
slot_meta = json.load(open(f"{args.data_dir}/slot_meta.json"))
# Define Preprocessor
tokenizer = BertTokenizer.from_pretrained(args.model_name_or_path)
processor = TRADEPreprocessor(
slot_meta, tokenizer,
word_drop=args.word_dropout) ## preprocessor에 word dropout 적용
_config = BertConfig.from_pretrained(args.model_name_or_path)
args.hidden_act = _config.hidden_act
args.layer_norm_eps = _config.layer_norm_eps
args.vocab_size = len(tokenizer)
args.n_gate = len(
processor.gating2id) # gating 갯수 none, dontcare, ptr, yes, no
feature_path = '/opt/ml/code/p3-dst-chatting-day/features/' + args.feature_model
if os.path.exists(feature_path):
# 저장된 feature 피클 가져오기
with open(feature_path + '/train_features.pickle', 'rb') as f:
train_features = pickle.load(f)
with open(feature_path + '/dev_features.pickle', 'rb') as f:
dev_features = pickle.load(f)
def train(args):
# Define Tokenizer
tokenizer_module = getattr(import_module("transformers"),
f"{args.model_name}Tokenizer")
tokenizer = tokenizer_module.from_pretrained(args.pretrained_name_or_path)
slot_meta, train_examples, dev_examples, dev_labels = train_data_loading(
args, isUserFirst=False, isDialogueLevel=False)
# Define Preprocessor
processor = TRADEPreprocessor(slot_meta,
tokenizer,
max_seq_length=args.max_seq_length,
use_n_gate=args.use_n_gate)
train_features = processor.convert_examples_to_features(train_examples)
dev_features = processor.convert_examples_to_features(dev_examples)
train_loader = get_data_loader(processor, train_features,
args.train_batch_size)
dev_loader = get_data_loader(processor, dev_features, args.eval_batch_size)
args.vocab_size = len(tokenizer)
args.n_gate = len(
processor.gating2id
) # gating 갯수 : (none, dontcare, ptr) or (none, yes, no, dontcare, ptr)
# Slot Meta tokenizing for the decoder initial inputs
tokenized_slot_meta = []
for slot in slot_meta:
tokenized_slot_meta.append(
tokenizer.encode(slot.replace("-", " "), add_special_tokens=False))
# Model 선언
model = TRADE(args, tokenized_slot_meta)
# model.set_subword_embedding(args) # Subword Embedding 초기화
print(f"Subword Embeddings is loaded from {args.pretrained_name_or_path}")
model.to(device)
print("Model is initialized")
# Optimizer 및 Scheduler 선언
n_epochs = args.epochs
t_total = len(train_loader) * n_epochs
# get_optimizer 부분에서 자동으로 warmup_steps를 계산할 수 있도록 바꿨음 (아래가 원래의 code)
# warmup_steps = int(t_total * args.warmup_ratio)
optimizer = get_optimizer(model,
args) # get optimizer (Adam, sgd, AdamP, ..)
scheduler = get_scheduler(
optimizer, t_total, args) # get scheduler (custom, linear, cosine, ..)
loss_fnc_1 = masked_cross_entropy_for_value # generation - # classes: vocab_size
loss_fnc_2 = nn.CrossEntropyLoss()
# loss_fnc_2 = LabelSmoothingLoss(classes=model.decoder.n_gate,smoothing=args.smoothing_factor)
json.dump(
vars(args),
open(f"{args.model_dir}/{args.model_fold}/exp_config.json", "w"),
indent=2,
ensure_ascii=False,
)
json.dump(
slot_meta,
open(f"{args.model_dir}/{args.model_fold}/slot_meta.json", "w"),
indent=2,
ensure_ascii=False,
)
best_score, best_checkpoint = 0, 0
for epoch in range(n_epochs):
model.train()
for step, batch in enumerate(train_loader):
input_ids, segment_ids, input_masks, gating_ids, target_ids, guids = [
b.to(device) if not isinstance(b, list) else b for b in batch
]
# teacher forcing
if (args.teacher_forcing_ratio > 0.0
and random.random() < args.teacher_forcing_ratio):
tf = target_ids
else:
tf = None
all_point_outputs, all_gate_outputs = model(
input_ids, segment_ids, input_masks, target_ids.size(-1), tf)
# generation loss
loss_1 = loss_fnc_1(
all_point_outputs.contiguous(),
target_ids.contiguous().view(-1),
tokenizer.pad_token_id,
)
# gating loss
loss_2 = loss_fnc_2(
all_gate_outputs.contiguous().view(-1, args.n_gate),
gating_ids.contiguous().view(-1),
)
loss = loss_1 + loss_2
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step()
for learning_rate in scheduler.get_lr():
wandb.log({"learning_rate": learning_rate})
optimizer.zero_grad()
if step % 100 == 0:
print(
f"[{epoch}/{n_epochs}] [{step}/{len(train_loader)}] loss: {loss.item()} gen: {loss_1.item()} gate: {loss_2.item()}"
)
wandb.log({
"epoch": epoch,
"Train epoch loss": loss.item(),
"Train epoch gen loss": loss_1.item(),
"Train epoch gate loss": loss_2.item(),
})
predictions = inference_TRADE(model, dev_loader, processor, device)
eval_result = _evaluation(predictions, dev_labels, slot_meta)
for k, v in eval_result.items():
if k in ("joint_goal_accuracy", 'turn_slot_accuracy',
'turn_slot_f1'):
print(f"{k}: {v}")
if best_score < eval_result["joint_goal_accuracy"]:
print("Update Best checkpoint!")
best_score = eval_result["joint_goal_accuracy"]
best_checkpoint = epoch
wandb.log({
"epoch":
epoch,
"Best joint goal accuracy":
best_score,
"Best turn slot accuracy":
eval_result['turn_slot_accuracy'],
"Best turn slot f1":
eval_result['turn_slot_f1']
})
if args.logging_accuracy_per_domain_slot:
wandb.log({
k: v
for k, v in eval_result.items()
if k not in ("joint_goal_accuracy", 'turn_slot_accuracy',
'turn_slot_f1')
})
torch.save(model.state_dict(),
f"{args.model_dir}/{args.model_fold}/model-{epoch}.bin")
print(f"Best checkpoint: {args.model_dir}/model-{best_checkpoint}.bin")
wandb.log(
{"Best checkpoint": f"{args.model_dir}/model-{best_checkpoint}.bin"})
def main_inference(args, config):
slot_meta = json.load(
open(f"{args.model_dir}/{args.model_fold}/slot_meta.json", "r"))
ontology = json.load(open(f"{CFG.TrainOntology}", "r"))
if config.replace_word_data:
slot_meta = [meta.replace('택시', '버스') for meta in slot_meta]
ontology = {
domain_slot_key.replace('택시', '버스'): domain_slot_value
for domain_slot_key, domain_slot_value in ontology.items()
}
# Define Tokenizer
tokenizer_module = getattr(import_module("transformers"),
f"{config.model_name}Tokenizer")
tokenizer = tokenizer_module.from_pretrained(
config.pretrained_name_or_path)
# Extracting Featrues
if config.dst == 'TRADE':
eval_examples = test_data_loading(args,
isUserFirst=False,
isDialogueLevel=False)
processor = TRADEPreprocessor(slot_meta, tokenizer)
tokenized_slot_meta = []
for slot in slot_meta:
tokenized_slot_meta.append(
tokenizer.encode(slot.replace("-", " "),
add_special_tokens=False))
# Model 선언
model = TRADE(config, tokenized_slot_meta)
model.set_subword_embedding(config) # Subword Embedding 초기화
elif config.dst == 'SUMBT':
eval_examples = test_data_loading(args,
isUserFirst=True,
isDialogueLevel=True)
max_turn = max([len(e) * 2 for e in eval_examples])
processor = SUMBTPreprocessor(
slot_meta,
tokenizer,
ontology=ontology, # predefined ontology
max_seq_length=config.max_seq_length, # 각 turn마다 최대 길이
max_turn_length=max_turn) # 각 dialogue의 최대 turn 길이
slot_type_ids, slot_values_ids = tokenize_ontology(
ontology, tokenizer, config.max_label_length)
# Model 선언
num_labels = [len(s)
for s in slot_values_ids] # 각 Slot 별 후보 Values의 갯수
model = SUMBT(config, num_labels, device)
model.initialize_slot_value_lookup(
slot_values_ids,
slot_type_ids) # Tokenized Ontology의 Pre-encoding using BERT_SV
eval_features = processor.convert_examples_to_features(eval_examples)
eval_loader = get_data_loader(processor, eval_features,
config.eval_batch_size)
print("# eval:", len(eval_loader))
ckpt = torch.load(
f'{args.model_dir}/{args.model_fold}/model-{args.chkpt_idx}.bin',
map_location="cpu")
model.load_state_dict(ckpt)
model.to(device)
print("Model is loaded")
inference_module = getattr(import_module("inference"),
f"inference_{config.dst}")
predictions = inference_module(model, eval_loader, processor, device)
os.makedirs(args.output_dir, exist_ok=True)
json.dump(
predictions,
open(f"{args.output_dir}/{args.model_fold}-predictions.csv", "w"),
indent=2,
ensure_ascii=False,
)
class TRADEHandler(BaseHandler, ABC):
"""
Transformers text classifier handler class. This handler takes a text (string) and
as input and returns the classification text based on the serialized transformers checkpoint.
"""
def __init__(self):
super(TRADEHandler, self).__init__()
self.initialized = False
self.config, self.slot_meta = self.load_json_data(
"./exp_config.json", "./slot_meta.json")
def load_json_data(self, exp_config_path, slot_meta_path):
config = json.load(open(exp_config_path, "r"))
config = argparse.Namespace(**config)
slot_meta = json.load(open(slot_meta_path, "r"))
return config, slot_meta
def initialize(self, ctx):
self.manifest = ctx.manifest
properties = ctx.system_properties
model_dir = properties.get("model_dir")
serialized_file = self.manifest["model"]["serializedFile"]
model_pt_path = os.path.join(model_dir, serialized_file)
self.device = torch.device("cuda:" +
str(properties.get("gpu_id")) if torch.cuda.
is_available() else "cpu")
# Read model serialize/pt file
self.tokenizer = BertTokenizer.from_pretrained(
self.config.model_name_or_path)
self.processor = TRADEPreprocessor(self.slot_meta, self.tokenizer)
tokenized_slot_meta = []
for slot in self.slot_meta:
tokenized_slot_meta.append(
self.tokenizer.encode(slot.replace("-", " "),
add_special_tokens=False))
self.model = TRADE(self.config, tokenized_slot_meta)
ckpt = torch.load(model_pt_path, map_location="cpu")
self.model.load_state_dict(ckpt)
self.model.to(self.device)
print("Model is loaded")
self.initialized = True
def preprocess(self, requests):
""" Very basic preprocessing code - only tokenizes.
Extend with your own preprocessing steps as needed.
"""
input_batch = []
for idx, data in enumerate(requests):
input_text = data.get("data")
if input_text is None:
input_text = data.get("body")
if isinstance(input_text, (bytes, bytearray)):
input_text = input_text.decode('utf-8')
input_text = json.loads(input_text)
input_batch.extend(input_text)
eval_examples = get_examples_from_dialogues(input_batch,
user_first=False,
dialogue_level=False)
eval_features = self.processor.convert_examples_to_features(
eval_examples)
eval_data = WOSDataset(eval_features)
eval_sampler = SequentialSampler(eval_data)
eval_loader = DataLoader(
eval_data,
batch_size=1,
sampler=eval_sampler,
collate_fn=self.processor.collate_fn,
)
return eval_loader
def postprocess_state(self, state):
for i, s in enumerate(state):
s = s.replace(" : ", ":")
state[i] = s.replace(" , ", ", ")
return state
def inference(self, inputs):
self.model.eval()
output_lst = []
predictions = {}
for batch in inputs:
input_ids, segment_ids, input_masks, gating_ids, target_ids, guids = [
b.to(self.device) if not isinstance(b, list) else b
for b in batch
]
with torch.no_grad():
o, g = self.model(input_ids, segment_ids, input_masks, 9)
_, generated_ids = o.max(-1)
_, gated_ids = g.max(-1)
for guid, gate, gen in zip(guids, gated_ids.tolist(),
generated_ids.tolist()):
prediction = self.processor.recover_state(gate, gen)
prediction = self.postprocess_state(prediction)
predictions[guid] = prediction
output_lst.append(predictions)
return output_lst
# def inference(self, inputs):
# """
# Predict the class of a text using a trained transformer model.
# """
# # NOTE: This makes the assumption that your model expects text to be tokenized
# # with "input_ids" and "token_type_ids" - which is true for some popular transformer models, e.g. bert.
# # If your transformer model expects different tokenization, adapt this code to suit
# # its expected input format.
# prediction = self.model(
# inputs['input_ids'].to(self.device),
# token_type_ids=inputs['token_type_ids'].to(self.device)
# )[0].argmax().item()
# logger.info("Model predicted: '%s'", prediction)
#
# if self.mapping:
# prediction = self.mapping[str(prediction)]
#
# return [prediction]
def postprocess(self, inference_output):
# TODO: Add any needed post-processing of the model predictions here
return inference_output