def run():
sentences, pos, tag, enc_pos, enc_tag = utils.process_data(config.DATA_FILE)
meta_data = {
"enc_pos": enc_pos,
"enc_tag": enc_tag
}
joblib.dump(meta_data, "meta.bin")
num_pos = len(list(enc_pos.classes_))
num_tag = len(list(enc_tag.classes_))
(
train_sentences,
test_sentences,
train_pos,
test_pos,
train_tag,
test_tag
) = model_selection.train_test_split(sentences, pos, tag, random_state=42, test_size=0.1)
train_dataset = dataset.EntityDataset(
texts = train_sentences, pos=train_pos, tags=train_tag
)
test_dataset = dataset.EntityDataset(
texts = test_sentences, pos=test_pos, tags=test_tag
)
train_data_loader = torch.utils.data.DataLoader(
train_dataset, batch_size = config.TRAIN_BATCH_SIZE, num_workers=4
)
test_data_loader = torch.utils.data.DataLoader(
test_dataset, batch_size=config.VALID_BATCH_SIZE, num_workers=1
)
device = torch.device("cuda")
model = EntityModel(num_tag=num_tag, num_pos=num_pos)
model.to(device)
param_optimizer = list(model.named_parameters())
no_decay = ["bias", "LayerNorm.bias", "LayerNorm.weight"]
optimizer_param = [
{
"params" : [
p for n, p in param_optimizer if not any(nd in n for nd in no_decay)
],
"weight_decay": 0.001,
},
{
"params" : [
p for n, p in param_optimizer if any(nd in n for nd in no_decay)
],
"weight_decay": 0.0,
},
]
num_train_steps = int(len(train_sentences) / config.TRAIN_BATCH_SIZE * config.EPOCHS )
optimizer = AdamW(optimizer_param, lr=3e-5)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=0, num_training_steps=num_train_steps
)
best_loss = np.inf
for epoch in range(config.EPOCHS):
train_loss = engine.train_fn(train_data_loader, model, optimizer, device, scheduler)
test_loss = engine.eval_fn(test_data_loader, model, device)
print(f"Train Loss = {train_loss} Valod Loss = {test_loss}")
if test_loss < best_loss:
torch.save(model.state_dict(), config.MODEL_SAVE_PATH)
best_loss = test_loss
no_decay = ["bias", "LayerNorm.bias", "LayerNorm.weight"]
optimizer_parameters = [
{
"params": [
p for n, p in param_optimizer if not any(nd in n for nd in no_decay)
],
"weight_decay": 0.001,
},
{
"params": [
p for n, p in param_optimizer if any(nd in n for nd in no_decay)
],
"weight_decay": 0.0,
},
]
num_train_steps = int(len(train_sentences) / config.TRAIN_BATCH_SIZE * config.EPOCHS)
optimizer = AdamW(optimizer_parameters, lr=3e-5)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=0, num_training_steps=num_train_steps
)
best_loss = np.inf
for epoch in range(config.EPOCHS):
train_loss = engine.train_fn(train_data_loader, model, optimizer, device, scheduler)
test_loss = engine.eval_fn(valid_data_loader, model, device)
print(f"Train Loss = {train_loss} Valid Loss = {test_loss}")
if test_loss < best_loss:
torch.save(model.state_dict(), config.MODEL_PATH)
best_loss = test_loss
num_train_steps = int(len(train_sentences) / config.params["TRAIN_BATCH_SIZE"] * config.params["EPOCHS"])
optimizer = AdamW(optimizer_parameters, lr=3e-5)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=0, num_training_steps=num_train_steps
)
best_loss = np.inf
for epoch in range(config.params["EPOCHS"]):
train_loss = engine.train_fn(train_data_loader, model, optimizer, device, scheduler)
test_loss, metrics, classification_metrics = engine.eval_with_metrics_combined(valid_data_loader, model, device, enc_tag, classify_enc_tag)
df_metrics = pd.DataFrame.from_dict(metrics)
df_metrics = df_metrics.transpose()
weighted_average = df_metrics["f1-score"][-1]
micro_average = df_metrics["f1-score"][-3]
table = wandb.Table(dataframe=df_metrics.transpose())
wandb.log({"Train loss":train_loss, "Valid loss": test_loss})
wandb.log({"Validation Metric details": table})
wandb.log({"Weighted average": weighted_average, "Micro average": micro_average})
print(f"Train Loss = {train_loss} Valid Loss = {test_loss}, \nMetrics = {df_metrics}, \nClassify Metrics = {pd.DataFrame.from_dict(classification_metrics)} ")
if test_loss < best_loss:
torch.save(model.state_dict(), config.params["MODEL_PATH"])
model.save_pretrained_model(config.params["BASE_MODEL_PATH"]+"_finetuned_"+config.params["language"])
config.TOKENIZER.save_pretrained(config.params["BASE_MODEL_PATH"]+"_finetuned_"+config.params["language"])
best_loss = test_loss
wandb.run.summary["best_valloss"] = best_loss