13  Hyperparameter Tuning for PyTorch With spotPython

In this tutorial, we will show how spotPython can be integrated into the PyTorch training workflow. It is based on the tutorial “Hyperparameter Tuning with Ray Tune” from the PyTorch documentation (PyTorch 2023a), which is an extension of the tutorial “Training a Classifier” (PyTorch 2023b) for training a CIFAR10 image classifier.

This document refers to the following software versions:

spotPython can be installed via pip1.

!pip install spotPython

Results that refer to the Ray Tune package are taken from https://PyTorch.org/tutorials/beginner/hyperparameter_tuning_tutorial.html2.

13.1 Setup

Before we consider the detailed experimental setup, we select the parameters that affect run time, initial design size and the device that is used.

MAX_TIME = 60
INIT_SIZE = 20
DEVICE = "cpu" # "cuda:0"

13.2 Initialization of the fun_control Dictionary

spotPython uses a Python dictionary for storing the information required for the hyperparameter tuning process. This dictionary is called fun_control and is initialized with the function fun_control_init. The function fun_control_init returns a skeleton dictionary. The dictionary is filled with the required information for the hyperparameter tuning process. It stores the hyperparameter tuning settings, e.g., the deep learning network architecture that should be tuned, the classification (or regression) problem, and the data that is used for the tuning. The dictionary is used as an input for the SPOT function.

from spotPython.utils.init import fun_control_init
fun_control = fun_control_init(task="classification",
    tensorboard_path="runs/14_spot_ray_hpt_torch_cifar10")

13.3 Data Loading

The data loading process is implemented in the same manner as described in the Section “Data loaders” in PyTorch (2023a). The data loaders are wrapped into the function load_data_cifar10 which is identical to the function load_data in PyTorch (2023a). A global data directory is used, which allows sharing the data directory between different trials. The method load_data_cifar10 is part of the spotPython package and can be imported from spotPython.data.torchdata.

In the following step, the test and train data are added to the dictionary fun_control.

from spotPython.data.torchdata import load_data_cifar10
train, test = load_data_cifar10()
n_samples = len(train)
# add the dataset to the fun_control
fun_control.update({
    "train": train,
    "test": test,
    "n_samples": n_samples})

13.4 The Model (Algorithm) to be Tuned

13.4.1 Specification of the Preprocessing Model

After the training and test data are specified and added to the fun_control dictionary, spotPython allows the specification of a data preprocessing pipeline, e.g., for the scaling of the data or for the one-hot encoding of categorical variables. The preprocessing model is called prep_model (“preparation” or pre-processing) and includes steps that are not subject to the hyperparameter tuning process. The preprocessing model is specified in the fun_control dictionary. The preprocessing model can be implemented as a sklearn pipeline. The following code shows a typical preprocessing pipeline:

categorical_columns = ["cities", "colors"]
one_hot_encoder = OneHotEncoder(handle_unknown="ignore",
                                    sparse_output=False)
prep_model = ColumnTransformer(
        transformers=[
             ("categorical", one_hot_encoder, categorical_columns),
         ],
         remainder=StandardScaler(),
     )

Because the Ray Tune (ray[tune]) hyperparameter tuning as described in PyTorch (2023a) does not use a preprocessing model, the preprocessing model is set to None here.

prep_model = None
fun_control.update({"prep_model": prep_model})

13.4.2 Select algorithm and core_model_hyper_dict

The same neural network model as implemented in the section “Configurable neural network” of the PyTorch tutorial (PyTorch 2023a) is used here. We will show the implementation from PyTorch (2023a) in Section 13.4.2.1 first, before the extended implementation with spotPython is shown in Section 13.4.2.2.

13.4.2.1 Implementing a Configurable Neural Network With Ray Tune

We used the same hyperparameters that are implemented as configurable in the PyTorch tutorial. We specify the layer sizes, namely l1 and l2, of the fully connected layers:

class Net(nn.Module):
    def __init__(self, l1=120, l2=84):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, l1)
        self.fc2 = nn.Linear(l1, l2)
        self.fc3 = nn.Linear(l2, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = x.view(-1, 16 * 5 * 5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

The learning rate, i.e., lr, of the optimizer is made configurable, too:

optimizer = optim.SGD(net.parameters(), lr=config["lr"], momentum=0.9)

13.4.2.2 Implementing a Configurable Neural Network With spotPython

spotPython implements a class which is similar to the class described in the PyTorch tutorial. The class is called Net_CIFAR10 and is implemented in the file netcifar10.py.

from torch import nn
import torch.nn.functional as F
import spotPython.torch.netcore as netcore


class Net_CIFAR10(netcore.Net_Core):
    def __init__(self, l1, l2, lr_mult, batch_size, epochs, k_folds, patience,
    optimizer, sgd_momentum):
        super(Net_CIFAR10, self).__init__(
            lr_mult=lr_mult,
            batch_size=batch_size,
            epochs=epochs,
            k_folds=k_folds,
            patience=patience,
            optimizer=optimizer,
            sgd_momentum=sgd_momentum,
        )
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, l1)
        self.fc2 = nn.Linear(l1, l2)
        self.fc3 = nn.Linear(l2, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = x.view(-1, 16 * 5 * 5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

13.4.3 The Net_Core class

Net_CIFAR10 inherits from the class Net_Core which is implemented in the file netcore.py. It implements the additional attributes that are common to all neural network models. The Net_Core class is implemented in the file netcore.py. It implements hyperparameters as attributes, that are not used by the core_model, e.g.:

  • optimizer (optimizer),
  • learning rate (lr),
  • batch size (batch_size),
  • epochs (epochs),
  • k_folds (k_folds), and
  • early stopping criterion “patience” (patience).

Users can add further attributes to the class. The class Net_Core is shown below.

from torch import nn


class Net_Core(nn.Module):
    def __init__(self, lr_mult, batch_size, epochs, k_folds, patience,
        optimizer, sgd_momentum):
        super(Net_Core, self).__init__()
        self.lr_mult = lr_mult
        self.batch_size = batch_size
        self.epochs = epochs
        self.k_folds = k_folds
        self.patience = patience
        self.optimizer = optimizer
        self.sgd_momentum = sgd_momentum

13.4.4 Comparison of the Approach Described in the PyTorch Tutorial With spotPython

Comparing the class Net from the PyTorch tutorial and the class Net_CIFAR10 from spotPython, we see that the class Net_CIFAR10 has additional attributes and does not inherit from nn directly. It adds an additional class, Net_core, that takes care of additional attributes that are common to all neural network models, e.g., the learning rate multiplier lr_mult or the batch size batch_size.

spotPython’s core_model implements an instance of the Net_CIFAR10 class. In addition to the basic neural network model, the core_model can use these additional attributes. spotPython provides methods for handling these additional attributes to guarantee 100% compatibility with the PyTorch classes. The method add_core_model_to_fun_control adds the hyperparameters and additional attributes to the fun_control dictionary. The method is shown below.

from spotPython.torch.netcifar10 import Net_CIFAR10
from spotPython.data.torch_hyper_dict import TorchHyperDict
from spotPython.hyperparameters.values import add_core_model_to_fun_control
core_model = Net_CIFAR10
fun_control = add_core_model_to_fun_control(core_model=core_model,
                              fun_control=fun_control,
                              hyper_dict=TorchHyperDict,
                              filename=None)

13.5 The Search Space: Hyperparameters

In Section 13.5.1, we first describe how to configure the search space with ray[tune] (as shown in PyTorch (2023a)) and then how to configure the search space with spotPython in Section 13.5.2.

13.5.1 Configuring the Search Space With Ray Tune

Ray Tune’s search space can be configured as follows (PyTorch 2023a):

config = {
    "l1": tune.sample_from(lambda _: 2**np.random.randint(2, 9)),
    "l2": tune.sample_from(lambda _: 2**np.random.randint(2, 9)),
    "lr": tune.loguniform(1e-4, 1e-1),
    "batch_size": tune.choice([2, 4, 8, 16])
}

The tune.sample_from() function enables the user to define sample methods to obtain hyperparameters. In this example, the l1 and l2 parameters should be powers of 2 between 4 and 256, so either 4, 8, 16, 32, 64, 128, or 256. The lr (learning rate) should be uniformly sampled between 0.0001 and 0.1. Lastly, the batch size is a choice between 2, 4, 8, and 16.

At each trial, ray[tune] will randomly sample a combination of parameters from these search spaces. It will then train a number of models in parallel and find the best performing one among these. ray[tune] uses the ASHAScheduler which will terminate bad performing trials early.

13.5.2 Configuring the Search Space With spotPython

13.5.2.1 The hyper_dict Hyperparameters for the Selected Algorithm

spotPython uses JSON files for the specification of the hyperparameters. Users can specify their individual JSON files, or they can use the JSON files provided by spotPython. The JSON file for the core_model is called torch_hyper_dict.json.

In contrast to ray[tune], spotPython can handle numerical, boolean, and categorical hyperparameters. They can be specified in the JSON file in a similar way as the numerical hyperparameters as shown below. Each entry in the JSON file represents one hyperparameter with the following structure: type, default, transform, lower, and upper.

"factor_hyperparameter": {
    "levels": ["A", "B", "C"],
    "type": "factor",
    "default": "B",
    "transform": "None",
    "core_model_parameter_type": "str",
    "lower": 0,
    "upper": 2},

The corresponding entries for the Net_CIFAR10 class are shown below.

{"Net_CIFAR10":
    {
        "l1": {
            "type": "int",
            "default": 5,
            "transform": "transform_power_2_int",
            "lower": 2,
            "upper": 9},
        "l2": {
            "type": "int",
            "default": 5,
            "transform": "transform_power_2_int",
            "lower": 2,
            "upper": 9},
        "lr_mult": {
            "type": "float",
            "default": 1.0,
            "transform": "None",
            "lower": 0.1,
            "upper": 10},
        "batch_size": {
            "type": "int",
            "default": 4,
            "transform": "transform_power_2_int",
            "lower": 1,
            "upper": 4},
        "epochs": {
            "type": "int",
            "default": 3,
            "transform": "transform_power_2_int",
            "lower": 1,
            "upper": 4},
        "k_folds": {
            "type": "int",
            "default": 2,
            "transform": "None",
            "lower": 2,
            "upper": 3},
        "patience": {
            "type": "int",
            "default": 5,
            "transform": "None",
            "lower": 2,
            "upper": 10},
        "optimizer": {
            "levels": ["Adadelta",
                       "Adagrad",
                       "Adam",
                       "AdamW",
                       "SparseAdam",
                       "Adamax",
                       "ASGD",
                       "LBFGS",
                       "NAdam",
                       "RAdam",
                       "RMSprop",
                       "Rprop",
                       "SGD"],
            "type": "factor",
            "default": "SGD",
            "transform": "None",
            "class_name": "torch.optim",
            "core_model_parameter_type": "str",
            "lower": 0,
            "upper": 12},
        "sgd_momentum": {
            "type": "float",
            "default": 0.0,
            "transform": "None",
            "lower": 0.0,
            "upper": 1.0}
    }
}

13.5.3 Modifying the Hyperparameters

Ray tune (PyTorch 2023a) does not provide a way to change the specified hyperparameters without re-compilation. However, spotPython provides functions for modifying the hyperparameters, their bounds and factors as well as for activating and de-activating hyperparameters without re-compilation of the Python source code. These functions are described in the following.

13.5.3.1 Modify hyper_dict Hyperparameters for the Selected Algorithm aka core_model

After specifying the model, the corresponding hyperparameters, their types and bounds are loaded from the JSON file torch_hyper_dict.json. After loading, the user can modify the hyperparameters, e.g., the bounds. spotPython provides a simple rule for de-activating hyperparameters: If the lower and the upper bound are set to identical values, the hyperparameter is de-activated. This is useful for the hyperparameter tuning, because it allows to specify a hyperparameter in the JSON file, but to de-activate it in the fun_control dictionary. This is done in the next step.

13.5.3.2 Modify Hyperparameters of Type numeric and integer (boolean)

Since the hyperparameter k_folds is not used in the PyTorch tutorial, it is de-activated here by setting the lower and upper bound to the same value. Note, k_folds is of type “integer”.

from spotPython.hyperparameters.values import modify_hyper_parameter_bounds
fun_control = modify_hyper_parameter_bounds(fun_control, 
    "batch_size", bounds=[1, 5])
fun_control = modify_hyper_parameter_bounds(fun_control, 
    "k_folds", bounds=[0, 0])
fun_control = modify_hyper_parameter_bounds(fun_control, 
    "patience", bounds=[3, 3])

13.5.3.3 Modify Hyperparameter of Type factor

In a similar manner as for the numerical hyperparameters, the categorical hyperparameters can be modified. New configurations can be chosen by adding or deleting levels. For example, the hyperparameter optimizer can be re-configured as follows:

In the following setting, two optimizers ("SGD" and "Adam") will be compared during the spotPython hyperparameter tuning. The hyperparameter optimizer is active.

from spotPython.hyperparameters.values import modify_hyper_parameter_levels
fun_control = modify_hyper_parameter_levels(fun_control,
     "optimizer", ["SGD", "Adam"])

The hyperparameter optimizer can be de-activated by choosing only one value (level), here: "SGD".

fun_control = modify_hyper_parameter_levels(fun_control, "optimizer", ["SGD"])

As discussed in Section 13.6, there are some issues with the LBFGS optimizer. Therefore, the usage of the LBFGS optimizer is not deactivated in spotPython by default. However, the LBFGS optimizer can be activated by adding it to the list of optimizers. Rprop was removed, because it does perform very poorly (as some pre-tests have shown). However, it can also be activated by adding it to the list of optimizers. Since SparseAdam does not support dense gradients, Adam was used instead. Therefore, there are 10 default optimizers:

fun_control = modify_hyper_parameter_levels(fun_control, "optimizer",
    ["Adadelta", "Adagrad", "Adam", "AdamW", "Adamax", "ASGD", 
    "NAdam", "RAdam", "RMSprop", "SGD"])

13.6 Optimizers

Table 14.1 shows some of the optimizers available in PyTorch:

\(a\) denotes (0.9,0.999), \(b\) (0.5,1.2), and \(c\) (1e-6, 50), respectively. \(R\) denotes required, but unspecified. “m” denotes momentum, “w_d” weight_decay, “d” dampening, “n” nesterov, “r” rho, “l_s” learning rate for scaling delta, “l_d” lr_decay, “b” betas, “l” lambd, “a” alpha, “m_d” for momentum_decay, “e” etas, and “s_s” for step_sizes.

Table 13.1: Optimizers available in PyTorch (selection). The default values are shown in the table.
Optimizer lr m w_d d n r l_s l_d b l a m_d e s_s
Adadelta - - 0. - - 0.9 1. - - - - - - -
Adagrad 1e-2 - 0. - - - - 0. - - - - - -
Adam 1e-3 - 0. - - - - - \(a\) - - - - -
AdamW 1e-3 - 1e-2 - - - - - \(a\) - - - - -
SparseAdam 1e-3 - - - - - - - \(a\) - - - - -
Adamax 2e-3 - 0. - - - - - \(a\) - - - - -
ASGD 1e-2 .9 0. - F - - - - 1e-4 .75 - - -
LBFGS 1. - - - - - - - - - - - - -
NAdam 2e-3 - 0. - - - - - \(a\) - - 0 - -
RAdam 1e-3 - 0. - - - - - \(a\) - - - - -
RMSprop 1e-2 0. 0. - - - - - \(a\) - - - - -
Rprop 1e-2 - - - - - - - - - \(b\) \(c\) - -
SGD \(R\) 0. 0. 0. F - - - - - - - - -

spotPython implements an optimization handler that maps the optimizer names to the corresponding PyTorch optimizers.

A note on LBFGS

We recommend deactivating PyTorch’s LBFGS optimizer, because it does not perform very well. The PyTorch documentation, see https://pytorch.org/docs/stable/generated/torch.optim.LBFGS.html#torch.optim.LBFGS, states:

This is a very memory intensive optimizer (it requires additional param_bytes * (history_size + 1) bytes). If it doesn’t fit in memory try reducing the history size, or use a different algorithm.

Furthermore, the LBFGS optimizer is not compatible with the PyTorch tutorial. The reason is that the LBFGS optimizer requires the closure function, which is not implemented in the PyTorch tutorial. Therefore, the LBFGS optimizer is recommended here. Since there are ten optimizers in the portfolio, it is not recommended tuning the hyperparameters that effect one single optimizer only.

A note on the learning rate

spotPython provides a multiplier for the default learning rates, lr_mult, because optimizers use different learning rates. Using a multiplier for the learning rates might enable a simultaneous tuning of the learning rates for all optimizers. However, this is not recommended, because the learning rates are not comparable across optimizers. Therefore, we recommend fixing the learning rate for all optimizers if multiple optimizers are used. This can be done by setting the lower and upper bounds of the learning rate multiplier to the same value as shown below.

Thus, the learning rate, which affects the SGD optimizer, will be set to a fixed value. We choose the default value of 1e-3 for the learning rate, because it is used in other PyTorch examples (it is also the default value used by spotPython as defined in the optimizer_handler() method). We recommend tuning the learning rate later, when a reduced set of optimizers is fixed. Here, we will demonstrate how to select in a screening phase the optimizers that should be used for the hyperparameter tuning.

For the same reason, we will fix the sgd_momentum to 0.9.

fun_control = modify_hyper_parameter_bounds(fun_control,
    "lr_mult", bounds=[1.0, 1.0])
fun_control = modify_hyper_parameter_bounds(fun_control,
    "sgd_momentum", bounds=[0.9, 0.9])

13.7 Evaluation: Data Splitting

The evaluation procedure requires the specification of the way how the data is split into a train and a test set and the loss function (and a metric). As a default, spotPython provides a standard hold-out data split and cross validation.

13.7.1 Hold-out Data Split

If a hold-out data split is used, the data will be partitioned into a training, a validation, and a test data set. The split depends on the setting of the eval parameter. If eval is set to train_hold_out, one data set, usually the original training data set, is split into a new training and a validation data set. The training data set is used for training the model. The validation data set is used for the evaluation of the hyperparameter configuration and early stopping to prevent overfitting. In this case, the original test data set is not used.

spotPython returns the hyperparameters of the machine learning and deep learning models, e.g., number of layers, learning rate, or optimizer, but not the model weights. Therefore, after the SPOT run is finished, the corresponding model with the optimized architecture has to be trained again with the best hyperparameter configuration. The training is performed on the training data set. The test data set is used for the final evaluation of the model.

Summarizing, the following splits are performed in the hold-out setting:

  1. Run spotPython with eval set to train_hold_out to determine the best hyperparameter configuration.
  2. Train the model with the best hyperparameter configuration (“architecture”) on the training data set: train_tuned(model_spot, train, "model_spot.pt").
  3. Test the model on the test data: test_tuned(model_spot, test, "model_spot.pt")

These steps will be exemplified in the following sections. In addition to this hold-out setting, spotPython provides another hold-out setting, where an explicit test data is specified by the user that will be used as the validation set. To choose this option, the eval parameter is set to test_hold_out. In this case, the training data set is used for the model training. Then, the explicitly defined test data set is used for the evaluation of the hyperparameter configuration (the validation).

13.7.2 Cross-Validation

The cross validation setting is used by setting the eval parameter to train_cv or test_cv. In both cases, the data set is split into \(k\) folds. The model is trained on \(k-1\) folds and evaluated on the remaining fold. This is repeated \(k\) times, so that each fold is used exactly once for evaluation. The final evaluation is performed on the test data set. The cross validation setting is useful for small data sets, because it allows to use all data for training and evaluation. However, it is computationally expensive, because the model has to be trained \(k\) times.

Combinations of the above settings are possible, e.g., cross validation can be used for training and hold-out for evaluation or vice versa. Also, cross validation can be used for training and testing. Because cross validation is not used in the PyTorch tutorial (PyTorch 2023a), it is not considered further here.

13.7.3 Overview of the Evaluation Settings

13.7.3.1 Settings for the Hyperparameter Tuning

An overview of the training evaluations is shown in Table 14.2. "train_cv" and "test_cv" use sklearn.model_selection.KFold() internally. More details on the data splitting are provided in Section 15.14 (in the Appendix).

Table 13.2: Overview of the evaluation settings.
eval train test function comment
"train_hold_out" \(\checkmark\) train_one_epoch(), validate_one_epoch() for early stopping splits the train data set internally
"test_hold_out" \(\checkmark\) \(\checkmark\) train_one_epoch(), validate_one_epoch() for early stopping use the test data set for validate_one_epoch()
"train_cv" \(\checkmark\) evaluate_cv(net, train) CV using the train data set
"test_cv" \(\checkmark\) evaluate_cv(net, test) CV using the test data set . Identical to "train_cv", uses only test data.

13.7.3.2 Settings for the Final Evaluation of the Tuned Architecture

13.7.3.2.1 Training of the Tuned Architecture

train_tuned(model, train): train the model with the best hyperparameter configuration (or simply the default) on the training data set. It splits the traindata into new train and validation sets using create_train_val_data_loaders(), which calls torch.utils.data.random_split() internally. Currently, 60% of the data is used for training and 40% for validation. The train data is used for training the model with train_hold_out(). The validation data is used for early stopping using validate_fold_or_hold_out() on the validation data set.

13.7.3.2.2 Testing of the Tuned Architecture

test_tuned(model, test): test the model on the test data set. No data splitting is performed. The (trained) model is evaluated using the validate_fold_or_hold_out() function. Note: During training, shuffle is set to True, whereas during testing, shuffle is set to False.

Section 15.14.1.4 describes the final evaluation of the tuned architecture.

13.8 Evaluation: Loss Functions and Metrics

The key "loss_function" specifies the loss function which is used during the optimization. There are several different loss functions under PyTorch’s nn package. For example, a simple loss is MSELoss, which computes the mean-squared error between the output and the target. In this tutorial we will use CrossEntropyLoss, because it is also used in the PyTorch tutorial.

from torch.nn import CrossEntropyLoss
loss_function = CrossEntropyLoss()
fun_control.update({"loss_function": loss_function})

In addition to the loss functions, spotPython provides access to a large number of metrics.

  • The key "metric_sklearn" is used for metrics that follow the scikit-learn conventions.
  • The key "river_metric" is used for the river based evaluation (Montiel et al. 2021) via eval_oml_iter_progressive, and
  • the key "metric_torch" is used for the metrics from TorchMetrics.

TorchMetrics is a collection of more than 90 PyTorch metrics, see https://torchmetrics.readthedocs.io/en/latest/. Because the PyTorch tutorial uses the accuracy as metric, we use the same metric here. Currently, accuracy is computed in the tutorial’s example code. We will use TorchMetrics instead, because it offers more flexibilty, e.g., it can be used for regression and classification. Furthermore, TorchMetrics offers the following advantages:

* A standardized interface to increase reproducibility
* Reduces Boilerplate
* Distributed-training compatible
* Rigorously tested
* Automatic accumulation over batches
* Automatic synchronization between multiple devices

Therefore, we set

import torchmetrics
metric_torch = torchmetrics.Accuracy(task="multiclass", num_classes=10)
loss_function = CrossEntropyLoss()
weights = 1.0
metric_torch = torchmetrics.Accuracy(task="multiclass", num_classes=10)
shuffle = True
eval = "train_hold_out"
device = DEVICE
show_batch_interval = 100_000
path="torch_model.pt"

fun_control.update({
               "data_dir": None,
               "checkpoint_dir": None,
               "horizon": None,
               "oml_grace_period": None,
               "weights": weights,
               "step": None,
               "log_level": 50,
               "weight_coeff": None,
               "metric_torch": metric_torch,
               "metric_river": None,
               "metric_sklearn": None,
               "loss_function": loss_function,
               "shuffle": shuffle,
               "eval": eval,
               "device": device,
               "show_batch_interval": show_batch_interval,
               "path": path,
               })

13.9 Preparing the SPOT Call

The following code passes the information about the parameter ranges and bounds to spot.

Now, the dictionary fun_control contains all information needed for the hyperparameter tuning. Before the hyperparameter tuning is started, it is recommended to take a look at the experimental design. The method gen_design_table generates a design table as follows:

from spotPython.utils.eda import gen_design_table
print(gen_design_table(fun_control))

This allows to check if all information is available and if the information is correct. Table 14.3 shows the experimental design for the hyperparameter tuning. The table shows the hyperparameters, their types, default values, lower and upper bounds, and the transformation function. The transformation function is used to transform the hyperparameter values from the unit hypercube to the original domain. The transformation function is applied to the hyperparameter values before the evaluation of the objective function. Hyperparameter transformations are shown in the column “transform”, e.g., the l1 default is 5, which results in the value \(2^5 = 32\) for the network, because the transformation transform_power_2_int was selected in the JSON file. The default value of the batch_size is set to 4, which results in a batch size of \(2^4 = 16\).

Table 13.3: Experimental design for the hyperparameter tuning.
name type default lower upper transform
l1 int 5 2 9 transform_power_2_int
l2 int 5 2 9 transform_power_2_int
lr_mult float 1.0 1 1 None
batch_size int 4 1 5 transform_power_2_int
epochs int 3 3 4 transform_power_2_int
k_folds int 1 0 0 None
patience int 5 3 3 None
optimizer factor SGD 0 9 None
sgd_momentum float 0.0 0.9 0.9 None

13.10 The Objective Function fun_torch

The objective function fun_torch is selected next. It implements an interface from PyTorch’s training, validation, and testing methods to spotPython.

from spotPython.fun.hypertorch import HyperTorch
fun = HyperTorch().fun_torch

13.11 Using Default Hyperparameters or Results from Previous Runs

We add the default setting to the initial design:

from spotPython.hyperparameters.values import get_default_hyperparameters_as_array
hyper_dict=TorchHyperDict().load()
X_start = get_default_hyperparameters_as_array(fun_control, hyper_dict)

13.12 Starting the Hyperparameter Tuning

The spotPython hyperparameter tuning is started by calling the Spot function. Here, we will run the tuner for approximately 30 minutes (max_time). Note: the initial design is always evaluated in the spotPython run. As a consequence, the run may take longer than specified by max_time, because the evaluation time of initial design (here: init_size, 10 points) is performed independently of max_time.

from spotPython.spot import spot
from math import inf
import numpy as np
spot_tuner = spot.Spot(fun=fun,
                   lower = lower,
                   upper = upper,
                   fun_evals = inf,
                   fun_repeats = 1,
                   max_time = MAX_TIME,
                   noise = False,
                   tolerance_x = np.sqrt(np.spacing(1)),
                   var_type = var_type,
                   var_name = var_name,
                   infill_criterion = "y",
                   n_points = 1,
                   seed=123,
                   log_level = 50,
                   show_models= False,
                   show_progress= True,
                   fun_control = fun_control,
                   design_control={"init_size": INIT_SIZE,
                                   "repeats": 1},
                   surrogate_control={"noise": True,
                                      "cod_type": "norm",
                                      "min_theta": -4,
                                      "max_theta": 3,
                                      "n_theta": len(var_name),
                                      "model_fun_evals": 10_000,
                                      "log_level": 50
                                      })
spot_tuner.run(X_start=X_start)

During the run, the following output is shown:

config: {'l1': 128, 'l2': 8, 'lr_mult': 1.0, 'batch_size': 32,
    'epochs': 16, 'k_folds': 0, 'patience': 3,
    'optimizer': 'AdamW', 'sgd_momentum': 0.9}
Epoch: 1
Loss on hold-out set: 1.5143253986358642
Accuracy on hold-out set: 0.4447
MulticlassAccuracy value on hold-out data: 0.4447000026702881
Epoch: 2
...
Epoch: 15
Loss on hold-out set: 1.2061678514480592
Accuracy on hold-out set: 0.59505
MulticlassAccuracy value on hold-out data: 0.5950499773025513
Early stopping at epoch 14
Returned to Spot: Validation loss: 1.2061678514480592
----------------------------------------------

13.13 Tensorboard

The textual output shown in the console (or code cell) can be visualized with Tensorboard.

13.13.1 Tensorboard: Start Tensorboard

Start TensorBoard through the command line to visualize data you logged. Specify the root log directory as used in fun_control = fun_control_init(task="regression", tensorboard_path="runs/24_spot_torch_regression") as the tensorboard_path. The argument logdir points to directory where TensorBoard will look to find event files that it can display. TensorBoard will recursively walk the directory structure rooted at logdir, looking for .tfevents. files.

tensorboard --logdir=runs

Go to the URL it provides or to http://localhost:6006/. The following figures show some screenshots of Tensorboard.

Figure 13.1: Tensorboard

Figure 13.2: Tensorboard

13.13.2 Saving the State of the Notebook

The state of the notebook can be saved and reloaded as follows:

import pickle
SAVE = False
LOAD = False

if SAVE:
    result_file_name = "res_" + experiment_name + ".pkl"
    with open(result_file_name, 'wb') as f:
        pickle.dump(spot_tuner, f)

if LOAD:
    result_file_name = "add_the_name_of_the_result_file_here.pkl"
    with open(result_file_name, 'rb') as f:
        spot_tuner =  pickle.load(f)

13.14 Results

After the hyperparameter tuning run is finished, the progress of the hyperparameter tuning can be visualized. The following code generates the progress plot from Figure 14.4.

spot_tuner.plot_progress(log_y=False, 
    filename="./figures/" + experiment_name+"_progress.png")

Figure 13.3: Progress plot. Black dots denote results from the initial design. Red dots illustrate the improvement found by the surrogate model based optimization (surrogate model based optimization).

Figure 14.4 shows a typical behaviour that can be observed in many hyperparameter studies (Bartz et al. 2022): the largest improvement is obtained during the evaluation of the initial design. The surrogate model based optimization-optimization with the surrogate refines the results. Figure 14.4 also illustrates one major difference between ray[tune] as used in PyTorch (2023a) and spotPython: the ray[tune] uses a random search and will generate results similar to the black dots, whereas spotPython uses a surrogate model based optimization and presents results represented by red dots in Figure 14.4. The surrogate model based optimization is considered to be more efficient than a random search, because the surrogate model guides the search towards promising regions in the hyperparameter space.

In addition to the improved (“optimized”) hyperparameter values, spotPython allows a statistical analysis, e.g., a sensitivity analysis, of the results. We can print the results of the hyperparameter tuning, see Table 14.4. The table shows the hyperparameters, their types, default values, lower and upper bounds, and the transformation function. The column “tuned” shows the tuned values. The column “importance” shows the importance of the hyperparameters. The column “stars” shows the importance of the hyperparameters in stars. The importance is computed by the SPOT software.

print(gen_design_table(fun_control=fun_control, spot=spot_tuner))
Table 13.4: Results of the hyperparameter tuning.
name type default lower upper tuned transform importance stars
l1 int 5 2.0 9.0 7.0 pow_2_int 100.00 ***
l2 int 5 2.0 9.0 3.0 pow_2_int 96.29 ***
lr_mult float 1.0 0.1 10.0 0.1 None 0.00
batchsize int 4 1.0 5.0 4.0 pow_2_int 0.00
epochs int 3 3.0 4.0 4.0 pow_2_int 4.18 *
k_folds int 2 0.0 0.0 0.0 None 0.00
patience int 5 3.0 3.0 3.0 None 0.00
optimizer factor SGD 0.0 9.0 3.0 None 0.16 .

To visualize the most important hyperparameters, spotPython provides the function plot_importance. The following code generates the importance plot from Figure 14.5.

spot_tuner.plot_importance(threshold=0.025,
    filename="./figures/" + experiment_name+"_importance.png")

Figure 13.4: Variable importance

13.14.1 Get SPOT Results

The architecture of the spotPython model can be obtained by the following code:

from spotPython.hyperparameters.values import get_one_core_model_from_X
X = spot_tuner.to_all_dim(spot_tuner.min_X.reshape(1,-1))
model_spot = get_one_core_model_from_X(X, fun_control)
model_spot

First, the numerical representation of the hyperparameters are obtained, i.e., the numpy array X is generated. This array is then used to generate the model model_spot by the function get_one_core_model_from_X. The model model_spot has the following architecture:

Net_CIFAR10(
  (conv1): Conv2d(3, 6, kernel_size=(5, 5), stride=(1, 1))
  (pool): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1,
    ceil_mode=False)
  (conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
  (fc1): Linear(in_features=400, out_features=64, bias=True)
  (fc2): Linear(in_features=64, out_features=32, bias=True)
  (fc3): Linear(in_features=32, out_features=10, bias=True)
)

13.14.2 Get Default Hyperparameters

In a similar manner as in Section 13.14.1, the default hyperparameters can be obtained.

# fun_control was modified, we generate a new one with the original 
# default hyperparameters
from spotPython.hyperparameters.values import get_one_core_model_from_X
fc = fun_control
fc.update({"core_model_hyper_dict":
    hyper_dict[fun_control["core_model"].__name__]})
model_default = get_one_core_model_from_X(X_start, fun_control=fc)

The corresponding default model has the following architecture:

Net_CIFAR10(
  (conv1): Conv2d(3, 6, kernel_size=(5, 5), stride=(1, 1))
  (pool): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1,
    ceil_mode=False)
  (conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
  (fc1): Linear(in_features=400, out_features=32, bias=True)
  (fc2): Linear(in_features=32, out_features=32, bias=True)
  (fc3): Linear(in_features=32, out_features=10, bias=True)
)

13.14.3 Evaluation of the Default Architecture

The method train_tuned takes a model architecture without trained weights and trains this model with the train data. The train data is split into train and validation data. The validation data is used for early stopping. The trained model weights are saved as a dictionary.

This evaluation is similar to the final evaluation in PyTorch (2023a).

from spotPython.torch.traintest import (
    train_tuned,
    test_tuned,
    )
train_tuned(net=model_default, train_dataset=train, shuffle=True,
        loss_function=fun_control["loss_function"],
        metric=fun_control["metric_torch"],
        device = DEVICE, show_batch_interval=1_000_000,
        path=None,
        task=fun_control["task"],)

test_tuned(net=model_default, test_dataset=test, 
        loss_function=fun_control["loss_function"],
        metric=fun_control["metric_torch"],
        shuffle=False, 
        device = DEVICE,
        task=fun_control["task"],)        

13.14.4 Evaluation of the Tuned Architecture

The following code trains the model model_spot. If path is set to a filename, e.g., path = "model_spot_trained.pt", the weights of the trained model will be saved to this file.

train_tuned(net=model_spot, train_dataset=train,
        loss_function=fun_control["loss_function"],
        metric=fun_control["metric_torch"],
        shuffle=True,
        device = DEVICE,
        path=None,
        task=fun_control["task"],)
test_tuned(net=model_spot, test_dataset=test,
            shuffle=False,
            loss_function=fun_control["loss_function"],
            metric=fun_control["metric_torch"],
            device = DEVICE,
            task=fun_control["task"],)

These runs will generate output similar to the following:

Loss on hold-out set: 1.2267619131326675
Accuracy on hold-out set: 0.58955
Early stopping at epoch 13

13.14.5 Comparison with Default Hyperparameters and Ray Tune

Table 14.5 shows the loss and accuracy of the default model, the model with the hyperparameters from SPOT, and the model with the hyperparameters from ray[tune]. The table shows a comparison of the loss and accuracy of the default model, the model with the hyperparameters from SPOT, and the model with the hyperparameters from ray[tune]. ray[tune] only shows the validation loss, because training loss is not reported by ray[tune]

Table 13.5: Comparison.
Model Validation Loss Validation Accuracy Loss Accuracy
Default 2.1221 0.2452 2.1182 0.2425
spotPython 1.2268 0.5896 1.2426 0.5957
ray[tune] 1.1815 0.5836 - 0.5806

13.14.6 Detailed Hyperparameter Plots

The contour plots in this section visualize the interactions of the three most important hyperparameters, l1, l2, and epochs, and optimizer of the surrogate model used to optimize the hyperparameters. Since some of these hyperparameters take fatorial or integer values, sometimes step-like fitness landcapes (or response surfaces) are generated. SPOT draws the interactions of the main hyperparameters by default. It is also possible to visualize all interactions. For this, again refer to the notebook (Bartz-Beielstein 2023).

filename = "./figures/" + experiment_name
spot_tuner.plot_important_hyperparameter_contour(filename=filename)

Figure 13.5: Contour plot of the loss as a function of l1 and l2, i.e., the number of neurons in the layers.

Figure 13.6: Contour plot of the loss as a function of the number of epochs and the neurons in layer l1.

Figure 13.7: Contour plot of the loss as a function of the optimizer and the neurons in layer l1.

Figure 13.8: Contour plot of the loss as a function of the number of epochs and the neurons in layer l2.

Figure 13.9: Contour plot of the loss as a function of the optimizer and the neurons in layer l2.

Figure 13.10: Contour plot of the loss as a function of the optimizer and the number of epochs.

The figures (Figure 13.5 to Figure 13.10) show the contour plots of the loss as a function of the hyperparameters. These plots are very helpful for benchmark studies and for understanding neural networks. spotPython provides additional tools for a visual inspection of the results and give valuable insights into the hyperparameter tuning process. This is especially useful for model explainability, transparency, and trustworthiness. In addition to the contour plots, Figure 14.7 shows the parallel plot of the hyperparameters.

spot_tuner.parallel_plot()

Figure 13.11: Parallel plot

13.15 Summary and Outlook

This tutorial presents the hyperparameter tuning open source software spotPython for PyTorch. To show its basic features, a comparison with the “official” PyTorch hyperparameter tuning tutorial (PyTorch 2023a) is presented. Some of the advantages of spotPython are:

  • Numerical and categorical hyperparameters.
  • Powerful surrogate models.
  • Flexible approach and easy to use.
  • Simple JSON files for the specification of the hyperparameters.
  • Extension of default and user specified network classes.
  • Noise handling techniques.
  • Interaction with tensorboard.

Currently, only rudimentary parallel and distributed neural network training is possible, but these capabilities will be extended in the future. The next version of spotPython will also include a more detailed documentation and more examples.

Important

Important: This tutorial does not present a complete benchmarking study (Bartz-Beielstein et al. 2020). The results are only preliminary and highly dependent on the local configuration (hard- and software). Our goal is to provide a first impression of the performance of the hyperparameter tuning package spotPython. To demonstrate its capabilities, a quick comparison with ray[tune] was performed. ray[tune] was chosen, because it is presented as “an industry standard tool for distributed hyperparameter tuning.” The results should be interpreted with care.

13.16 Appendix

13.16.1 Sample Output From Ray Tune’s Run

The output from ray[tune] could look like this (PyTorch 2023b):

Number of trials: 10 (10 TERMINATED)
------+------+-------------+--------------+---------+------------+--------------------+
|   l1 |   l2 |          lr |   batch_size |    loss |   accuracy | training_iteration |
+------+------+-------------+--------------+---------+------------+--------------------|
|   64 |    4 | 0.00011629  |            2 | 1.87273 |     0.244  |                  2 |
|   32 |   64 | 0.000339763 |            8 | 1.23603 |     0.567  |                  8 |
|    8 |   16 | 0.00276249  |           16 | 1.1815  |     0.5836 |                 10 |
|    4 |   64 | 0.000648721 |            4 | 1.31131 |     0.5224 |                  8 |
|   32 |   16 | 0.000340753 |            8 | 1.26454 |     0.5444 |                  8 |
|    8 |    4 | 0.000699775 |            8 | 1.99594 |     0.1983 |                  2 |
|  256 |    8 | 0.0839654   |           16 | 2.3119  |     0.0993 |                  1 |
|   16 |  128 | 0.0758154   |           16 | 2.33575 |     0.1327 |                  1 |
|   16 |    8 | 0.0763312   |           16 | 2.31129 |     0.1042 |                  4 |
|  128 |   16 | 0.000124903 |            4 | 2.26917 |     0.1945 |                  1 |
+-----+------+------+-------------+--------------+---------+------------+--------------------+
Best trial config: {'l1': 8, 'l2': 16, 'lr': 0.00276249, 'batch_size': 16, 'data_dir': '...'}
Best trial final validation loss: 1.181501
Best trial final validation accuracy: 0.5836
Best trial test set accuracy: 0.5806

  1. Alternatively, the source code can be downloaded from gitHub: https://github.com/sequential-parameter-optimization/spotPython.↩︎

  2. We were not able to install Ray Tune on our system. Therefore, we used the results from the PyTorch tutorial.↩︎