Validation loss is the loss computed from a model’s predictions on data withheld from training. Keeping training and validation data separate reveals whether learning transfers to unseen examples rather than memorizing training pairs. section 3.6.7
model <- keras_model_sequential() |> layer_dense(units = 1) model |> compile( optimizer = optimizer_rmsprop(learning_rate = 0.1), loss = loss_mean_squared_error(), metrics = list(metric_binary_accuracy()) ) indices_permutation <- sample.int(nrow(inputs)) #1 shuffled_inputs <- inputs[indices_permutation, , drop = FALSE] shuffled_targets <- targets[indices_permutation, , drop = FALSE] num_validation_samples <- as.integer(0.3 * nrow(inputs)) #2 val_inputs <- shuffled_inputs[1:num_validation_samples, ] val_targets <- shuffled_targets[1:num_validation_samples, ] training_inputs <- shuffled_inputs[-(1:num_validation_samples), ] training_targets <- shuffled_targets[-(1:num_validation_samples), ] model |> fit( training_inputs, training_targets, #3 epochs = 5, batch_size = 16, validation_data = list(val_inputs, val_targets) #4 )
Split shuffled inputs and targets, reserve 30% for validation, then pass the untouched subset through fit() with validation_data. section 3.6.7
Use validation loss to:
- Monitor generalization alongside training loss and validation metrics; worsening validation performance signals that the model has begun to overfit. section 6.2.4
- Evaluate after training with evaluate(), which returns validation loss first, followed by configured metrics. section 3.6.7
Listing 8.15 Fitting the model using a Dataset
callbacks <- list(
callback_model_checkpoint(
filepath = "convnet_from_scratch.keras",
save_best_only = TRUE,
monitor = "val_loss"
)
)
history <- model |> fit(
train_dataset,
epochs = 50,
validation_data = validation_dataset,
callbacks = callbacks
)
Fit a model on training and validation datasets while configuring ModelCheckpoint to save only the epoch with the lowest val_loss. section 8.2.4
