Models API

keras4torch.Model wraps a torch.nn.Module to integrate training and inference features.

Configs

.compile(optimizer, loss, metrics, ...)

Configure the model for training.

• optimizer: String (name of optimizer) or optimizer instance.
• loss: String (name of objective function), objective function or loss instance.
• metrics: List of metrics to be evaluated by the model during training. You can also use dict to specify the abbreviation of each metric.
• epoch_metrics: List of non-linear metrics(e.g. ROC_AUC) that need to be evaluated on epoch end.
• device: Device of the model and its trainer, if None 'cuda' will be used when torch.cuda.is_available() otherwise 'cpu'.
• loop_config: Optional TrainerLoopConfig object to customize training and validation loop. See Training API for details.

NumPy workflow

.fit(x, y, epochs, batch_size, ...)

Train the model for a fixed number of epochs (iterations on a dataset).

• x (ndarray or torch.Tensor or Dataset): Input data
• y (ndarray or torch.Tensor): Target data
• epochs (int, default=10): Number of epochs to train the model
• batch_size (int, default=32): Number of samples per gradient update
• validation_batch_size (int, default=None): Number of samples for each step on validation loop, if None will use batch_size
• validation_split (float between 0 and 1): Fraction of the training data to be used as validation data
• shuffle_val_split (bool, default=True): Whether to do shuffling when validation_split is provided
• validation_data (tuple of x and y or Dataset): Data on which to evaluate the loss and any model metrics at the end of each epoch
• callbacks (list of keras4torch.callbacks.Callback): List of callbacks to apply during training
• verbose (int, default=1): 0, 1, or 2. Verbosity mode. 0 = silent, 1 = normal, 2 = brief
• shuffle (bool, default=True): Whether to shuffle the training data before each epoch
• sample_weight (list of floats): Optional weights for the training samples. If provided will enable WeightedRandomSampler
• num_workers (int, default=0): Workers of DataLoader. If -1 will use cpu_count() - 1 for multiprocessing
• use_amp (bool, default=False): Whether to use automatic mixed precision
• accum_grad_steps (int, default=1): How many steps to update the model parameters
• **dl_kwargs: Extra keyword arguments for DataLoader

.evaluate(x, y, batch_size, ...)

Return the loss value & metrics values for the model in test mode.

• x (ndarray or torch.Tensor or Dataset): Input data
• y (ndarray or torch.Tensor): Target data
• batch_size (int, default=32): Number of samples per batch
• num_workers (int, default=0): Workers of DataLoader. If -1 will use cpu_count() - 1 for multiprocessing
• use_amp (bool, default=False): Whether to use automatic mixed precision
• **dl_kwargs: Extra keyword arguments for DataLoader

.predict(x, batch_size, ...)

Generate output predictions for the input samples.

• x (ndarray or torch.Tensor or Dataset): Input data
• batch_size (int, default=32): Number of samples per batch
• device (default=None): Device to do inference
• output_numpy (bool, default=True): If True, the output will move to CPU and convert to NumPy array
• activation (Callable or str, default=None): Extra activation applied to the output tensor
• num_workers (int, default=0): Workers of DataLoader. If -1 will use cpu_count() - 1 for multiprocessing
• use_amp (bool, default=False): Whether to use automatic mixed precision
• **dl_kwargs: Extra keyword arguments for DataLoader

DataLoader workflow

.fit_dl(train_loader, val_loader, epochs, ...)

.evaluate_dl(data_loader, ...)

.predict_dl(data_loader, ...)

Saving & Serialization

.save_weights(filepath)

Equal to torch.save(model.state_dict(), filepath).

.load_weights(filepath)

Equal to model.load_state_dict(torch.load(filepath)).

Utilities

.summary(depth, ...)

Print a string summary of the network.

• depth (default=3): Summary details level

.count_params()

Count the total number of scalars composing the weights.

.trainable_params()

Return all trainable parameters of the model.