output_model_file = "./datas/mybert.bin"
lr = 2e-5
batch_size = 32
accumulation_steps = 2
np.random.seed(SEED)
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
torch.backends.cudnn.deterministic = True
lr_layers = [1 for i in range(12)]
for i in range(len(lr_layers)):
lr_layers[i] = lr_layers[i - 1] * 0.95 if i > 0 else lr
lr_layers = lr_layers[::-1]
model.zero_grad()
model = model.to(device)
param_optimizer = list(model.named_parameters())
import re
pattern = re.compile('[0-9]+')
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if n.find(".0.") != -1 and not any(nd in n for nd in no_decay)
],
'lr':
lr_layers[0],
def main():
train_df = pd.read_csv(TRAIN_PATH)
train_df['male'] = np.load(
"../input/identity-column-data/male_labeled.npy")
train_df['female'] = np.load(
"../input/identity-column-data/female_labeled.npy")
train_df['homosexual_gay_or_lesbian'] = np.load(
"../input/identity-column-data/homosexual_gay_or_lesbian_labeled.npy")
train_df['christian'] = np.load(
"../input/identity-column-data/christian_labeled.npy")
train_df['jewish'] = np.load(
"../input/identity-column-data/jewish_labeled.npy")
train_df['muslim'] = np.load(
"../input/identity-column-data/muslim_labeled.npy")
train_df['black'] = np.load(
"../input/identity-column-data/black_labeled.npy")
train_df['white'] = np.load(
"../input/identity-column-data/white_labeled.npy")
train_df['psychiatric_or_mental_illness'] = np.load(
"../input/identity-column-data/psychiatric_or_mental_illness_labeled.npy"
)
fold_df = pd.read_csv(FOLD_PATH)
# y = np.where(train_df['target'] >= 0.5, 1, 0)
y = train_df['target'].values
y_aux = train_df[AUX_COLUMNS].values
identity_columns_new = []
for column in identity_columns + ['target']:
train_df[column + "_bin"] = np.where(train_df[column] >= 0.5, True,
False)
if column != "target":
identity_columns_new.append(column + "_bin")
# Overall
weights = np.ones((len(train_df), )) / 4
# Subgroup
weights += (train_df[identity_columns].fillna(0).values >= 0.5).sum(
axis=1).astype(bool).astype(np.int) / 4
# Background Positive, Subgroup Negative
weights += (
((train_df["target"].values >= 0.5).astype(bool).astype(np.int) +
(1 - (train_df[identity_columns].fillna(0).values >= 0.5).sum(
axis=1).astype(bool).astype(np.int))) > 1).astype(bool).astype(
np.int) / 4
# Background Negative, Subgroup Positive
weights += (
((train_df["target"].values < 0.5).astype(bool).astype(np.int) +
(train_df[identity_columns].fillna(0).values >= 0.5).sum(
axis=1).astype(bool).astype(np.int)) > 1).astype(bool).astype(
np.int) / 4
loss_weight = 0.5
with timer('preprocessing text'):
# df["comment_text"] = [analyzer_embed(text) for text in df["comment_text"]]
train_df['comment_text'] = train_df['comment_text'].astype(str)
train_df = train_df.fillna(0)
with timer('load embedding'):
tokenizer = BertTokenizer.from_pretrained(BERT_MODEL_PATH,
cache_dir=None,
do_lower_case=True)
X_text = convert_lines_head_tail(
train_df["comment_text"].fillna("DUMMY_VALUE"), max_len, head_len,
tokenizer)
del tokenizer
gc.collect()
LOGGER.info(f"X_text {X_text.shape}")
with timer('train'):
train_index = fold_df.fold_id != fold_id
valid_index = fold_df.fold_id == fold_id
X_train, y_train, y_aux_train, w_train = X_text[train_index].astype(
"int32"), y[train_index], y_aux[train_index], weights[train_index]
X_val, y_val, y_aux_val, w_val = X_text[valid_index].astype("int32"), y[valid_index], y_aux[valid_index], \
weights[
valid_index]
test_df = train_df[valid_index]
del X_text, y, y_aux, weights, train_index, valid_index, train_df
gc.collect()
model = BertForSequenceClassification(bert_config, num_labels=n_labels)
model.load_state_dict(torch.load(model_path))
model.zero_grad()
model = model.to(device)
y_train = np.concatenate(
(y_train.reshape(-1, 1), w_train.reshape(-1, 1), y_aux_train),
axis=1).astype("float32")
y_val = np.concatenate(
(y_val.reshape(-1, 1), w_val.reshape(-1, 1), y_aux_val),
axis=1).astype("float32")
train_dataset = torch.utils.data.TensorDataset(
torch.tensor(X_train, dtype=torch.long),
torch.tensor(y_train, dtype=torch.float32))
valid = torch.utils.data.TensorDataset(
torch.tensor(X_val, dtype=torch.long),
torch.tensor(y_val, dtype=torch.float32))
ran_sampler = torch.utils.data.RandomSampler(train_dataset)
len_sampler = LenMatchBatchSampler(ran_sampler,
batch_size=batch_size,
drop_last=False)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_sampler=len_sampler)
valid_loader = torch.utils.data.DataLoader(valid,
batch_size=batch_size * 2,
shuffle=False)
LOGGER.info(f"done data loader setup")
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
num_train_optimization_steps = int(epochs * len(X_train) / batch_size /
accumulation_steps)
total_step = int(epochs * len(X_train) / batch_size)
optimizer = BertAdam(optimizer_grouped_parameters,
lr=base_lr,
warmup=0.005,
t_total=num_train_optimization_steps)
LOGGER.info(f"done optimizer loader setup")
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O1",
verbosity=0)
# criterion = torch.nn.BCEWithLogitsLoss().to(device)
criterion = CustomLoss(loss_weight).to(device)
LOGGER.info(f"done amp setup")
for epoch in range(1, epochs + 1):
LOGGER.info(f"Starting {epoch} epoch...")
LOGGER.info(f"length {len(X_train)} train {len(X_val)} train...")
if epoch == 1:
for param_group in optimizer.param_groups:
param_group['lr'] = base_lr * gammas[1]
tr_loss, train_losses = train_one_epoch(model,
train_loader,
criterion,
optimizer,
device,
accumulation_steps,
total_step,
n_labels,
base_lr,
gamma=gammas[2 * epoch])
LOGGER.info(f'Mean train loss: {round(tr_loss,5)}')
torch.save(model.state_dict(),
'{}_epoch{}_fold{}.pth'.format(exp, epoch, fold_id))
valid_loss, oof_pred = validate(model, valid_loader, criterion,
device, n_labels)
LOGGER.info(f'Mean valid loss: {round(valid_loss,5)}')
if epochs > 1:
test_df_cp = test_df.copy()
test_df_cp["pred"] = oof_pred[:, 0]
test_df_cp = convert_dataframe_to_bool(test_df_cp)
bias_metrics_df = compute_bias_metrics_for_model(
test_df_cp, identity_columns)
LOGGER.info(bias_metrics_df)
score = get_final_metric(bias_metrics_df,
calculate_overall_auc(test_df_cp))
LOGGER.info(f'score is {score}')
del model
gc.collect()
torch.cuda.empty_cache()
test_df["pred"] = oof_pred[:, 0]
test_df = convert_dataframe_to_bool(test_df)
bias_metrics_df = compute_bias_metrics_for_model(test_df, identity_columns)
LOGGER.info(bias_metrics_df)
score = get_final_metric(bias_metrics_df, calculate_overall_auc(test_df))
LOGGER.info(f'final score is {score}')
test_df.to_csv("oof.csv", index=False)
xs = list(range(1, len(train_losses) + 1))
plt.plot(xs, train_losses, label='Train loss')
plt.legend()
plt.xticks(xs)
plt.xlabel('Iter')
plt.savefig("loss.png")
def main():
# train_df = pd.read_csv(TRAIN_PATH).sample(frac=1.0, random_state=seed)
# train_size = int(len(train_df) * 0.9)
train_df = pd.read_csv(TRAIN_PATH).sample(train_size + valid_size, random_state=seed)
LOGGER.info(f'data_size is {len(train_df)}')
LOGGER.info(f'train_size is {train_size}')
y = np.where(train_df['target'] >= 0.5, 1, 0)
y_aux = train_df[AUX_COLUMNS].values
identity_columns_new = []
for column in identity_columns + ['target']:
train_df[column + "_bin"] = np.where(train_df[column] >= 0.5, True, False)
if column != "target":
identity_columns_new.append(column + "_bin")
sample_weights = np.ones(len(train_df), dtype=np.float32)
sample_weights += train_df[identity_columns_new].sum(axis=1)
sample_weights += train_df['target_bin'] * (~train_df[identity_columns_new]).sum(axis=1)
sample_weights += (~train_df['target_bin']) * train_df[identity_columns_new].sum(axis=1) * 5
sample_weights /= sample_weights.mean()
with timer('preprocessing text'):
# df["comment_text"] = [analyzer_embed(text) for text in df["comment_text"]]
train_df['comment_text'] = train_df['comment_text'].astype(str)
train_df = train_df.fillna(0)
with timer('load embedding'):
tokenizer = BertTokenizer.from_pretrained(BERT_MODEL_PATH, cache_dir=None, do_lower_case=True)
X_text = convert_lines(train_df["comment_text"].fillna("DUMMY_VALUE"), max_len, tokenizer)
test_df = train_df[train_size:]
with timer('train'):
X_train, y_train, y_aux_train, w_train = X_text[:train_size], y[:train_size], y_aux[
:train_size], sample_weights[
:train_size]
X_val, y_val, y_aux_val, w_val = X_text[train_size:], y[train_size:], y_aux[train_size:], sample_weights[
train_size:]
model = BertForSequenceClassification(bert_config, num_labels=n_labels)
model.load_state_dict(torch.load(model_path))
model.zero_grad()
model = model.to(device)
train_dataset = torch.utils.data.TensorDataset(torch.tensor(X_train, dtype=torch.long),
torch.tensor(y_train, dtype=torch.float))
valid = torch.utils.data.TensorDataset(torch.tensor(X_val, dtype=torch.long),
torch.tensor(y_val, dtype=torch.float))
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
valid_loader = torch.utils.data.DataLoader(valid, batch_size=batch_size * 2, shuffle=False)
sample_weight_train = [w_train.values, np.ones_like(w_train)]
sample_weight_val = [w_val.values, np.ones_like(w_val)]
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
num_train_optimization_steps = int(epochs * train_size / batch_size / accumulation_steps)
total_step = int(epochs * train_size / batch_size)
optimizer = BertAdam(optimizer_grouped_parameters,
lr=2e-5*gamma,
warmup=0.05,
t_total=num_train_optimization_steps)
model, optimizer = amp.initialize(model, optimizer, opt_level="O1", verbosity=0)
criterion = torch.nn.BCEWithLogitsLoss().to(device)
LOGGER.info(f"Starting 1 epoch...")
tr_loss, train_losses = train_one_epoch(model, train_loader, criterion, optimizer, device,
accumulation_steps, total_step, n_labels)
LOGGER.info(f'Mean train loss: {round(tr_loss,5)}')
torch.save(model.state_dict(), '{}_dic'.format(exp))
valid_loss, oof_pred = validate(model, valid_loader, criterion, device, n_labels)
del model
gc.collect()
torch.cuda.empty_cache()
test_df["pred"] = oof_pred.reshape(-1)
test_df = convert_dataframe_to_bool(test_df)
bias_metrics_df = compute_bias_metrics_for_model(test_df, identity_columns)
LOGGER.info(bias_metrics_df)
score = get_final_metric(bias_metrics_df, calculate_overall_auc(test_df))
LOGGER.info(f'final score is {score}')
test_df.to_csv("oof.csv", index=False)
xs = list(range(1, len(train_losses) + 1))
plt.plot(xs, train_losses, label='Train loss');
plt.legend();
plt.xticks(xs);
plt.xlabel('Iter')
plt.savefig("loss.png")
