def run_inference(model, test_loader):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model.eval()
model.to(device)
preds = []
with torch.no_grad():
for x_ids, x_mask in tqdm(test_loader):
pred = model((x_ids.to(device), x_mask.to(device)))
preds.append(pred.cpu())
preds = torch.cat(preds)
_, preds = torch.topk(preds, 1, dim=1)
preds = [map_label(item, True) for item in preds.numpy().flatten()]
return preds
df_train, df_val = train_test_split(
df, test_size=test_size, random_state=123, stratify=df[label_col])
tokenizer = RobertaTokenizerFast.from_pretrained(
tokenizer_path, max_len=max_length)
train_loader, val_loader = get_roberta_dataloaders(df_train, df_val, list_procs,
text_col, label_col, tokenizer, max_length,
train_batch_size, val_batch_size, num_workers, pin_memory)
# prepare class_weigths
classes_weights = None
if balanced_classes:
val_counts = df_train[label_col].value_counts()
tmp_idx, tmp_vals = np.array(val_counts.index), val_counts.values
classes_weights = np.sum(tmp_vals)/tmp_vals
tmp_idx = np.array([map_label(i, False) for i in tmp_idx])
classes_weights = torch.FloatTensor(classes_weights[np.argsort(tmp_idx)])
# load model
dict_model = {}
with open(model_path) as f:
exec(f.read(), dict_model)
model = dict_model["model"]
optimizer = dict_model["optimizer"]
scheduler = dict_model.get("scheduler")
shutil.copy(model_path, os.path.join(output_path, "model_script.py"))
train_config["model_script"] = os.path.join(output_path, "model_script.py")
_pretty_print("Training start")
model, train_metrics, val_metrics = run_training(model, optimizer, scheduler,
output_path, train_loader,
tokenizer,
max_length,
train_batch_size,
val_batch_size,
num_workers,
pin_memory,
)
# prepare class_weigths
classes_weights = None
if balanced_classes:
val_counts = df_train[label_col].value_counts()
tmp_idx, tmp_vals = np.array(val_counts.index), val_counts.values
classes_weights = np.sum(tmp_vals) / tmp_vals
tmp_idx = np.array(
[map_label(i, False, dict_label) for i in tmp_idx])
classes_weights = torch.FloatTensor(
classes_weights[np.argsort(tmp_idx)])
# load model
dict_model = {}
with open(model_path) as f:
exec(f.read(), dict_model)
model = dict_model["model"]
optimizer = dict_model["optimizer"]
scheduler = dict_model.get("scheduler")
_pretty_print("Training start")
output_path_k = os.path.join(output_path, f"fold_{k}")
model, train_metrics, val_metrics = run_training(
model,
dict_label,
tokenizer,
max_length,
train_batch_size,
val_batch_size,
num_workers,
pin_memory,
)
# prepare class_weigths
classes_weights = None
if balanced_classes:
val_counts = df_train[label_col].value_counts()
tmp_idx, tmp_vals = np.array(val_counts.index), val_counts.values
classes_weights = np.sum(tmp_vals) / tmp_vals
tmp_idx = np.array([map_label(i, False, dict_label) for i in tmp_idx])
classes_weights = torch.FloatTensor(classes_weights[np.argsort(tmp_idx)])
# load model
dict_model = {}
with open(model_path) as f:
exec(f.read(), dict_model)
model = dict_model["model"]
optimizer = dict_model["optimizer"]
scheduler = dict_model.get("scheduler")
if k == 0:
shutil.copy(model_path, os.path.join(output_path, "model_script.py"))
train_config["model_script"] = os.path.join(output_path, "model_script.py")
_pretty_print("Training start")
output_path_k = os.path.join(output_path, f"fold_{k}")
tokenizer,
max_length,
train_batch_size,
val_batch_size,
num_workers,
pin_memory,
)
# prepare class_weigths
classes_weights = None
if balanced_classes:
val_counts = df_train[label_col].value_counts()
tmp_idx, tmp_vals = np.array(val_counts.index), val_counts.values
classes_weights = np.sum(tmp_vals) / tmp_vals
tmp_idx = np.array(
[map_label(i, False, dict_label) for i in tmp_idx])
classes_weights = torch.FloatTensor(
classes_weights[np.argsort(tmp_idx)])
# load model
dict_model = {}
with open(model_path) as f:
exec(f.read(), dict_model)
model = dict_model["model"]
optimizer = dict_model["optimizer"]
scheduler = dict_model.get("scheduler")
if k == 0:
shutil.copy(model_path, os.path.join(output_path,
"model_script.py"))
train_config["model_script"] = os.path.join(
output_path, "model_script.py")