Skip to content

Optimizing Hyperparameters

View all lessons
Optimizing a subset of hyperparameters to achieve an objective.
Goku Mohandas
· ·
Repository · Notebook

📬  Receive new lessons straight to your inbox (once a month) and join 30K+ developers in learning how to responsibly deliver value with ML.

Intuition

Optimization is the process of fine-tuning the hyperparameters in our experiment to optimize towards a particular objective. It can be a computationally involved process depending on the number of parameters, search space and model architectures. Hyperparameters don't just include the model's parameters but they also include parameters (choices) from preprocessing, splitting, etc. When we look at all the different parameters that can be tuned, it quickly becomes a very large search space. However, just because something is a hyperparameter doesn't mean we need to tune it.

  • It's absolutely alright to fix some hyperparameters (ex. lower=True during preprocessing) and remove them from the current tuning subset. Just be sure to note which parameters you are fixing and your reasoning for doing so.
  • You can initially just tune a small, yet influential, subset of hyperparameters that you believe will yield best results.

We want to optimize our hyperparameters so that we can understand how each of them affects our objective. By running many trials across a reasonable search space, we can determine near ideal values for our different parameters. It's also a great opportunity to determine if a smaller parameters yield similar performances as larger ones (efficiency).

Tools

There are many options for hyperparameter tuning (Optuna, Ray tune, Hyperopt, etc.). We'll be using Optuna for it's simplicity, popularity and efficiency though they are all equally so. It really comes down to familiarity and whether a library has a specific implementation readily tested and available.

Application

There are many factors to consider when performing hyperparameter optimization and luckily Optuna allows us to implement them with ease. We'll be conducting a small study where we'll tune a set of arguments (we'll do a much more thorough study of the parameter space when we move our code to Python scripts). Here's the process for the study:

  1. Define an objective (metric) and identifying the direction to optimize.
  2. [OPTIONAL] Choose a sampler for determining parameters for subsequent trials. (default is a tree based sampler).
  3. [OPTIONAL] Choose a pruner to end unpromising trials early.
  4. Define the parameters to tune in each trial and the distribution of values to sample.

There are many more options (multiple objectives, storage options, etc.) to explore but this basic set up will allow us to optimize quite well.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
from argparse import Namespace
from numpyencoder import NumpyEncoder

# Specify arguments
params = Namespace(
    char_level=True,
    filter_sizes=list(range(1, 11)),
    batch_size=64,
    embedding_dim=128,
    num_filters=128,
    hidden_dim=128,
    dropout_p=0.5,
    lr=2e-4,
    num_epochs=200,
    patience=10,
)

We're going to modify our Trainer object to be able to prune unpromising trials based on the trial's validation loss. 

1
2
3
4
5
6
7
8
9
class Trainer(object):
    ...
    def train(self, ...):
        ...
        # Pruning based on the intermediate value
        self.trial.report(val_loss, epoch)
        if self.trial.should_prune():
            raise optuna.TrialPruned()
        ...

Code for complete Trainer class
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
# Trainer (modified for experiment tracking)
class Trainer(object):
    def __init__(self, model, device, loss_fn=None,
                optimizer=None, scheduler=None, trial=None):

        # Set params
        self.model = model
        self.device = device
        self.loss_fn = loss_fn
        self.optimizer = optimizer
        self.scheduler = scheduler
        self.trial = trial

    def train_step(self, dataloader):
        """Train step."""
        # Set model to train mode
        self.model.train()
        loss = 0.0

        # Iterate over train batches
        for i, batch in enumerate(dataloader):
            # Step
            batch = [item.to(self.device) for item in batch]
            inputs, targets = batch[:-1], batch[-1]
            self.optimizer.zero_grad()  # Reset gradients
            z = self.model(inputs)  # Forward pass
            J = self.loss_fn(z, targets)  # Define loss
            J.backward()  # Backward pass
            self.optimizer.step()  # Update weights

            # Cumulative Metrics
            loss += (J.detach().item() - loss) / (i + 1)

        return loss

    def eval_step(self, dataloader):
        """Validation or test step."""
        # Set model to eval mode
        self.model.eval()
        loss = 0.0
        y_trues, y_probs = [], []

        # Iterate over val batches
        with torch.inference_mode():
            for i, batch in enumerate(dataloader):

                # Step
                batch = [item.to(self.device) for item in batch]  # Set device
                inputs, y_true = batch[:-1], batch[-1]
                z = self.model(inputs)  # Forward pass
                J = self.loss_fn(z, y_true).item()

                # Cumulative Metrics
                loss += (J - loss) / (i + 1)

                # Store outputs
                y_prob = torch.sigmoid(z).cpu().numpy()
                y_probs.extend(y_prob)
                y_trues.extend(y_true.cpu().numpy())

        return loss, np.vstack(y_trues), np.vstack(y_probs)

    def predict_step(self, dataloader):
        """Prediction step."""
        # Set model to eval mode
        self.model.eval()
        y_probs = []

        # Iterate over val batches
        with torch.inference_mode():
            for i, batch in enumerate(dataloader):

                # Forward pass w/ inputs
                inputs, targets = batch[:-1], batch[-1]
                z = self.model(inputs)

                # Store outputs
                y_prob = torch.sigmoid(z).cpu().numpy()
                y_probs.extend(y_prob)

        return np.vstack(y_probs)

    def train(self, num_epochs, patience, train_dataloader, val_dataloader):
        best_val_loss = np.inf
        for epoch in range(num_epochs):
            # Steps
            train_loss = self.train_step(dataloader=train_dataloader)
            val_loss, _, _ = self.eval_step(dataloader=val_dataloader)
            self.scheduler.step(val_loss)

            # Early stopping
            if val_loss < best_val_loss:
                best_val_loss = val_loss
                best_model = self.model
                _patience = patience  # reset _patience
            else:
                _patience -= 1
            if not _patience:  # 0
                print("Stopping early!")
                break

            # Logging
            print(
                f"Epoch: {epoch+1} | "
                f"train_loss: {train_loss:.5f}, "
                f"val_loss: {val_loss:.5f}, "
                f"lr: {self.optimizer.param_groups[0]['lr']:.2E}, "
                f"_patience: {_patience}"
            )

            # Pruning based on the intermediate value
            self.trial.report(val_loss, epoch)
            if self.trial.should_prune():
                raise optuna.TrialPruned()

        return best_model, best_val_loss

We'll also modify our train_cnn function to include information about the trial.

1
2
3
4
5
6
7
def train_cnn(params, df, trial=None):
    ...
    # Trainer module
    trainer = Trainer(
        model=model, device=device, loss_fn=loss_fn,
        optimizer=optimizer, scheduler=scheduler, trial=trial)
    ...
Code for complete train_cnn function
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
def train_cnn(params, df, trial=None):
    """Train a CNN using specific arguments."""

    # Set seeds
    set_seeds()

    # Get data splits
    preprocessed_df = df.copy()
    preprocessed_df.text = preprocessed_df.text.apply(preprocess, lower=True)
    X_train, X_val, X_test, y_train, y_val, y_test, label_encoder = get_data_splits(preprocessed_df)
    X_test_raw = X_test
    num_classes = len(label_encoder)

    # Set device
    cuda = True
    device = torch.device("cuda" if (
        torch.cuda.is_available() and cuda) else "cpu")
    torch.set_default_tensor_type("torch.FloatTensor")
    if device.type == "cuda":
        torch.set_default_tensor_type("torch.cuda.FloatTensor")

    # Tokenize
    tokenizer = Tokenizer(char_level=params.char_level)
    tokenizer.fit_on_texts(texts=X_train)
    vocab_size = len(tokenizer)

    # Convert texts to sequences of indices
    X_train = np.array(tokenizer.texts_to_sequences(X_train))
    X_val = np.array(tokenizer.texts_to_sequences(X_val))
    X_test = np.array(tokenizer.texts_to_sequences(X_test))

    # Class weights
    counts = np.bincount([label_encoder.class_to_index[class_] for class_ in all_tags])
    class_weights = {i: 1.0/count for i, count in enumerate(counts)}

    # Create datasets
    train_dataset = CNNTextDataset(
        X=X_train, y=y_train, max_filter_size=max(params.filter_sizes))
    val_dataset = CNNTextDataset(
        X=X_val, y=y_val, max_filter_size=max(params.filter_sizes))
    test_dataset = CNNTextDataset(
        X=X_test, y=y_test, max_filter_size=max(params.filter_sizes))

    # Create dataloaders
    train_dataloader = train_dataset.create_dataloader(
        batch_size=params.batch_size)
    val_dataloader = val_dataset.create_dataloader(
        batch_size=params.batch_size)
    test_dataloader = test_dataset.create_dataloader(
        batch_size=params.batch_size)

    # Initialize model
    model = CNN(
        embedding_dim=params.embedding_dim, vocab_size=vocab_size,
        num_filters=params.num_filters, filter_sizes=params.filter_sizes,
        hidden_dim=params.hidden_dim, dropout_p=params.dropout_p,
        num_classes=num_classes)
    model = model.to(device)

    # Define loss
    class_weights_tensor = torch.Tensor(np.array(list(class_weights.values())))
    loss_fn = nn.BCEWithLogitsLoss(weight=class_weights_tensor)

    # Define optimizer & scheduler
    optimizer = torch.optim.Adam(model.parameters(), lr=params.lr)
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
        optimizer, mode="min", factor=0.1, patience=5)

    # Trainer module
    trainer = Trainer(
        model=model, device=device, loss_fn=loss_fn,
        optimizer=optimizer, scheduler=scheduler, trial=trial)

    # Train
    best_model, best_val_loss = trainer.train(
        params.num_epochs, params.patience, train_dataloader, val_dataloader)

    # Best threshold for f1
    train_loss, y_true, y_prob = trainer.eval_step(dataloader=train_dataloader)
    precisions, recalls, thresholds = precision_recall_curve(y_true.ravel(), y_prob.ravel())
    threshold = find_best_threshold(y_true.ravel(), y_prob.ravel())

    # Determine predictions using threshold
    test_loss, y_true, y_prob = trainer.eval_step(dataloader=test_dataloader)
    y_pred = np.array([np.where(prob >= threshold, 1, 0) for prob in y_prob])

    # Evaluate
    performance = get_metrics(
        y_true=y_test, y_pred=y_pred, classes=label_encoder.classes)

    return {
        "params": params,
        "tokenizer": tokenizer,
        "label_encoder": label_encoder,
        "model": best_model,
        "performance": performance,
        "best_val_loss": best_val_loss,
        "threshold": threshold,
    }

Objective

We need to define an Objective function that will consume a trial and a set of arguments and produce the metric to optimize on (f1 in our case). 

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
def objective(trial, params):
    """Objective function for optimization trials."""

    # Paramters (to tune)
    params.embedding_dim = trial.suggest_int("embedding_dim", 128, 512)
    params.num_filters = trial.suggest_int("num_filters", 128, 512)
    params.hidden_dim = trial.suggest_int("hidden_dim", 128, 512)
    params.dropout_p = trial.suggest_uniform("dropout_p", 0.3, 0.8)
    params.lr = trial.suggest_loguniform("lr", 5e-5, 5e-4)

    # Train & evaluate
    artifacts = train_cnn(params=params, df=df, trial=trial)

    # Set additional attributes
    trial.set_user_attr("precision", artifacts["performance"]["precision"])
    trial.set_user_attr("recall", artifacts["performance"]["recall"])
    trial.set_user_attr("f1", artifacts["performance"]["f1"])
    trial.set_user_attr("threshold", artifacts["threshold"])

    return artifacts["performance"]["f1"]

Study

We're ready to kick off our study with our MLFlowCallback so we can track all of the different trials. 

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
from optuna.integration.mlflow import MLflowCallback

# Optimize
NUM_TRIALS = 50 # small sample for now
pruner = optuna.pruners.MedianPruner(n_startup_trials=5, n_warmup_steps=5)
study = optuna.create_study(study_name="optimization", direction="maximize", pruner=pruner)
mlflow_callback = MLflowCallback(
    tracking_uri=mlflow.get_tracking_uri(), metric_name="f1")
study.optimize(lambda trial: objective(trial, params),
               n_trials=NUM_TRIALS,
               callbacks=[mlflow_callback])


A new study created in memory with name: optimization
Epoch: 1 | train_loss: 0.00645, val_loss: 0.00314, lr: 3.48E-04, _patience: 10
...
Epoch: 23 | train_loss: 0.00029, val_loss: 0.00175, lr: 3.48E-05, _patience: 1
Stopping early!
Trial 0 finished with value: 0.5999225606985846 and parameters: {'embedding_dim': 508, 'num_filters': 359, 'hidden_dim': 262, 'dropout_p': 0.6008497926241321, 'lr': 0.0003484755175747328}. Best is trial 0 with value: 0.5999225606985846.
INFO: 'optimization' does not exist. Creating a new experiment

...

Trial 10 pruned.

...

Epoch: 25 | train_loss: 0.00029, val_loss: 0.00156, lr: 2.73E-05, _patience: 2
Epoch: 26 | train_loss: 0.00028, val_loss: 0.00152, lr: 2.73E-05, _patience: 1
Stopping early!
Trial 49 finished with value: 0.6220047640997922 and parameters: {'embedding_dim': 485, 'num_filters': 420, 'hidden_dim': 477, 'dropout_p': 0.7984462152799114, 'lr': 0.0002619841505205434}. Best is trial 46 with value: 0.63900047716579.

1
2
3
4
5
6
# MLFlow dashboard
get_ipython().system_raw("mlflow server -h 0.0.0.0 -p 5000 --backend-store-uri $PWD/experiments/ &")
ngrok.kill()
ngrok.set_auth_token("")
ngrok_tunnel = ngrok.connect(addr="5000", proto="http", bind_tls=True)
print("MLflow Tracking UI:", ngrok_tunnel.public_url)

MLflow Tracking UI: https://d19689b7ba4e.ngrok.io

You can compare all (or a subset) of the trials in our experiment.

compare experiments

We can then view the results through various lens (contours, parallel coordinates, etc.)

contour plot
parallel coordinates 
1
2
3
4
# All trials
trials_df = study.trials_dataframe()
trials_df = trials_df.sort_values(["value"], ascending=False)  # sort by metric
trials_df.head()
number value datetime_start datetime_complete duration params_dropout_p params_embedding_dim params_hidden_dim params_lr params_num_filters user_attrs_f1 user_attrs_precision user_attrs_recall user_attrs_threshold state
46 46 0.639000 2021-01-26 21:29:09.435991 2021-01-26 21:30:20.637867 0 days 00:01:11.201876 0.670784 335 458 0.000298 477 0.639000 0.852947 0.540094 0.221352 COMPLETE
32 32 0.638382 2021-01-26 21:08:27.456865 2021-01-26 21:09:54.151386 0 days 00:01:26.694521 0.485060 322 329 0.000143 458 0.638382 0.860706 0.535624 0.285308 COMPLETE
33 33 0.638135 2021-01-26 21:09:54.182560 2021-01-26 21:11:14.038009 0 days 00:01:19.855449 0.567419 323 405 0.000163 482 0.638135 0.872309 0.537566 0.298093 COMPLETE
39 39 0.637652 2021-01-26 21:18:37.735567 2021-01-26 21:20:01.271413 0 days 00:01:23.535846 0.689044 391 401 0.000496 512 0.637652 0.852757 0.536279 0.258009 COMPLETE
34 34 0.634339 2021-01-26 21:11:14.068099 2021-01-26 21:12:33.645090 0 days 00:01:19.576991 0.592627 371 379 0.000213 486 0.634339 0.863092 0.531822 0.263524 COMPLETE 
1
2
3
# Best trial
print (f"Best value (val loss): {study.best_trial.value}")
print (f"Best hyperparameters: {study.best_trial.params}")

Best value (f1): 0.6953118802537894
Best hyperparameters: {'embedding_dim': 234, 'num_filters': 383, 'hidden_dim': 265, 'dropout_p': 0.4174131267446717, 'lr': 0.0004392663090337615}

Don't forget to save learned parameters (ex. decision threshold) during training which you'll need later for inference.

1
2
3
4
# Save best parameters
params = {**params.__dict__, **study.best_trial.params}
params["threshold"] = study.best_trial.user_attrs["threshold"]
print (json.dumps(params, indent=2, cls=NumpyEncoder))

{
  "char_level": true,
  "filter_sizes": [
    1,
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9,
    10
  ],
  "batch_size": 64,
  "embedding_dim": 234,
  "num_filters": 383,
  "hidden_dim": 265,
  "dropout_p": 0.4174131267446717,
  "lr": 0.0004392663090337615,
  "num_epochs": 100,
  "patience": 10,
  "threshold": 0.3115537464618683
}

... and now we're finally ready to move from working in Jupyter notebooks to Python scripts. We'll be revisiting everything we did so far, but this time with proper software engineering prinicples such as object oriented programming (OOPs), styling, testing, etc. → https://madewithml.com/#mlops

You'll most likely be using the CLI application to optimize and train your models. If you don't have access to GPUs (personal machine, AWS, GCP, etc.), check out the optimize.ipynb notebook for how to train on Google Colab and transfer the entire MLFlow experiment to your local machine. We essentially run optimization, then train the best model to download and transfer it's artifacts.

To cite this lesson, please use:

1
2
3
4
5
6
@article{madewithml,
    author       = {Goku Mohandas},
    title        = { Optimization - Made With ML },
    howpublished = {\url{https://madewithml.com/}},
    year         = {2021}
}