Skip to content

Corrected hyperparameter optimization problem (changed df_train k-folding by normal validation). #2

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
victor-fdz merged 1 commit into from
May 27, 2026
Merged

Corrected hyperparameter optimization problem (changed df_train k-folding by normal validation). #2

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
161 changes: 75 additions & 86 deletions lib/model/train_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -304,44 +304,35 @@ def train_model(
# -----------------------------
def optimize_hyperparameters(
df_train: pd.DataFrame,
df_val: pd.DataFrame,
batch_size: int = DEFAULT_BATCH_SIZE,
n_trials: int = DEFAULT_OPTUNA_TRIALS,
n_jobs: int = DEFAULT_OPTUNA_JOBS,
n_folds: int = DEFAULT_OPTUNA_FOLDS,
trial_epochs: int = DEFAULT_OPTUNA_EPOCHS,
fixed_margin: float = DEFAULT_MARGIN,
attention_heads: int = DEFAULT_ATTENTION_HEADS,
embedding_dim: int = DEFAULT_EMBEDDING_DIM,
optuna_output_dir: Path | None = None,
) -> dict[str, Any]:
"""Run Optuna hyperparameter search using stratified cross-validation."""
"""Run Optuna hyperparameter search using the external validation set."""

if n_trials <= 0:
raise ValueError("n_trials must be a positive integer.")

if n_jobs <= 0:
raise ValueError("n_jobs must be a positive integer.")

if n_folds < 2:
raise ValueError("n_folds must be at least 2.")

if trial_epochs <= 0:
raise ValueError("trial_epochs must be a positive integer.")

labels_for_stratification = _prepare_binary_labels_for_stratification(df_train)
label_counts = labels_for_stratification.value_counts()

if label_counts.min() < n_folds:
raise ValueError(
"n_folds cannot be larger than the smallest label class. "
f"Label counts: {dict(label_counts)} | n_folds={n_folds}"
)
_prepare_binary_labels_for_stratification(df_train)
_prepare_binary_labels_for_stratification(df_val)

logger.info(
"Running Optuna optimization: trials=%d | jobs=%d | folds=%d | trial_epochs=%d.",
"Running Optuna optimization with external validation: "
"trials=%d | jobs=%d | trial_epochs=%d.",
n_trials,
n_jobs,
n_folds,
trial_epochs,
)

Expand All @@ -353,12 +344,6 @@ def optimize_hyperparameters(
logger.warning("Could not reset PyTorch interop threads. Continuing.")
logger.info("Parallel Optuna enabled. Forced PyTorch threads per trial to 1.")

stratified_splitter = StratifiedKFold(
n_splits=n_folds,
shuffle=True,
random_state=DEFAULT_RANDOM_STATE,
)

def objective(trial: optuna.trial.Trial) -> float:
learning_rate = trial.suggest_float(
"learning_rate",
Expand Down Expand Up @@ -393,87 +378,89 @@ def objective(trial: optuna.trial.Trial) -> float:
[18, 20, 24, 30],
)

fold_losses = []
trial_seed = DEFAULT_RANDOM_STATE + trial.number * 1000

for fold_index, (train_idx, val_idx) in enumerate(
stratified_splitter.split(df_train, labels_for_stratification)
):
trial_seed = DEFAULT_RANDOM_STATE + trial.number * 1000 + fold_index
train_loader_generator = torch.Generator()
train_loader_generator.manual_seed(trial_seed)

data_loader_generator = torch.Generator()
data_loader_generator.manual_seed(trial_seed)
train_loader = DataLoader(
DNAPairDatasetWithMeta(df_train),
batch_size=batch_size,
shuffle=True,
generator=train_loader_generator,
num_workers=0,
)

fold_train = df_train.iloc[train_idx].copy()
fold_val = df_train.iloc[val_idx].copy()
val_loader = DataLoader(
DNAPairDatasetWithMeta(df_val),
batch_size=batch_size,
shuffle=False,
num_workers=0,
)

train_loader = DataLoader(
DNAPairDatasetWithMeta(fold_train),
batch_size=batch_size,
shuffle=True,
generator=data_loader_generator,
num_workers=0,
)
_validate_loader_not_empty(train_loader, "Optuna training")
_validate_loader_not_empty(val_loader, "Optuna validation")

with torch.random.fork_rng(devices=[]):
torch.manual_seed(trial_seed)

val_loader = DataLoader(
DNAPairDatasetWithMeta(fold_val),
batch_size=batch_size,
shuffle=False,
num_workers=0,
model = SiameseCNN(
dropout_rate=dropout_rate,
small_kernel_size=small_kernel_size,
medium_kernel_size=medium_kernel_size,
large_kernel_size=large_kernel_size,
attention_heads=attention_heads,
embedding_dim=embedding_dim,
)

with torch.random.fork_rng(devices=[]):
torch.manual_seed(trial_seed)
model = SiameseCNN(
dropout_rate=dropout_rate,
small_kernel_size=small_kernel_size,
medium_kernel_size=medium_kernel_size,
large_kernel_size=large_kernel_size,
attention_heads=attention_heads,
embedding_dim=embedding_dim,
)
criterion = ContrastiveLossCosine(
margin=fixed_margin,
)

criterion = ContrastiveLossCosine(margin=fixed_margin)
optimizer = torch.optim.AdamW(
model.parameters(),
lr=learning_rate,
weight_decay=weight_decay,
)

optimizer = torch.optim.AdamW(
model.parameters(),
lr=learning_rate,
weight_decay=weight_decay,
)
for epoch in range(1, trial_epochs + 1):
model.train()

for _ in range(trial_epochs):
model.train()
for seq1, seq2, labels, *_ in train_loader:
optimizer.zero_grad()

for seq1, seq2, labels, *_ in train_loader:
optimizer.zero_grad()
out1, out2 = model(seq1, seq2)
out1, out2 = model(seq1, seq2)

loss = criterion(
out1,
out2,
labels,
)

loss = criterion(
out1,
out2,
labels,
)
loss.backward()
optimizer.step()

loss.backward()
optimizer.step()
model.eval()
validation_loss = 0.0

model.eval()
val_loss = 0.0
with torch.no_grad():
for seq1, seq2, labels, *_ in val_loader:
out1, out2 = model(seq1, seq2)

with torch.no_grad():
for seq1, seq2, labels, *_ in val_loader:
out1, out2 = model(seq1, seq2)
validation_loss += criterion(
out1,
out2,
labels,
).item()

val_loss += criterion(
out1,
out2,
labels,
).item()
mean_validation_loss = validation_loss / len(val_loader)

fold_losses.append(val_loss / len(val_loader))
return float(mean_validation_loss)

return float(np.mean(fold_losses))
study = optuna.create_study(
direction="minimize",
)

study = optuna.create_study(direction="minimize")
study.optimize(
objective,
n_trials=n_trials,
Expand All @@ -486,7 +473,10 @@ def objective(trial: optuna.trial.Trial) -> float:
if optuna_output_dir is not None:
optuna_output_dir.mkdir(parents=True, exist_ok=True)

with (optuna_output_dir / "optuna_best_params.json").open("w", encoding="utf-8") as handle:
with (optuna_output_dir / "optuna_best_params.json").open(
"w",
encoding="utf-8",
) as handle:
json.dump(study.best_params, handle, indent=2)

study.trials_dataframe().to_csv(
Expand All @@ -497,7 +487,6 @@ def objective(trial: optuna.trial.Trial) -> float:

return study.best_params


# -----------------------------
# Plots
# -----------------------------
Expand Down Expand Up @@ -763,10 +752,10 @@ def run_training(
if optimize_hparams:
best = optimize_hyperparameters(
df_train=df_train,
df_val=df_val,
batch_size=batch_size,
n_trials=optuna_trials,
n_jobs=optuna_jobs,
n_folds=optuna_folds,
trial_epochs=optuna_epochs,
fixed_margin=margin,
attention_heads=attention_heads,
Expand Down
Loading