Блог

Я пытаюсь сравнить fit функцию с. GradientTape В то время как я получаю 80% процентов при использовании функции подгонки, я получаю 27% GradientTape . Я уже проверил возможные проблемы здесь, где используется дополнительный декоратор tf.function , однако, без каких-либо улучшений. Здесь вы можете найти мой код:

 epochs = 50
batch_size = 128

# First we're going to fit to task A and retain a copy of parameters trained on Task A
#model.fit([X_amp_tr_A, X_phase_tr_A, weight_tr_A, temp_tr_A, humidity_tr_A, loc_tr_A], shelf_life_tr_A,  
#                  validation_data = ([X_amp_test_A, X_phase_test_A, weight_test_A, 
#                                      temp_test_A, humidity_test_A, loc_test_A], np.array(shelf_life_test_A)))
theta_A = {n: p.value() for n, p in enumerate(model.trainable_variables.copy())}

# Metrics for the custom training loop
accuracy = tf.keras.metrics.CategoricalAccuracy('accuracy')
metric = tf.keras.metrics.CategoricalCrossentropy('categorical_crossentropy')
loss = tf.keras.losses.CategoricalCrossentropy(from_logits=True)

optimizer = tf.keras.optimizers.SGD()


dataset = tf.data.Dataset.from_tensor_slices(({'amp': X_amp_tr_A, 'phase': X_phase_tr_A, 
                                               'weight': weight_tr_A, 'temp': temp_tr_A, 
                                               'moist': humidity_tr_A, 'zone': loc_tr_A}, shelf_life_tr_A)).batch(batch_size)#

@tf.function
def train_step(images, labels, metric):
    loss = tf.keras.losses.CategoricalCrossentropy(from_logits=True)
    
    with tf.GradientTape() as tape:
        predictions = model(images, training=True)
        total_loss = loss(labels, predictions)  l2_penalty(model, theta_A) 

    grads = tape.gradient(total_loss, model.trainable_variables)
    optimizer.apply_gradients(zip(grads, model.trainable_variables))
    
    metric.update_state(labels, predictions)
    accuracy.update_state(labels, predictions)
    #metric.update_state(labels, preds)
    #return total_loss, accuracy


for epoch in range(epochs):
    #accuracy.reset_states()
    #metric.reset_states()
    cumulative_loss = 0
    for i, (batch) in enumerate(dataset):
        imgs = batch[0]
        labels = batch[1]
        #with tf.GradientTape() as tape:
            #preds = model(imgs)
            #total_loss = loss(labels, preds)  l2_penalty(model, theta_A) 
            #print(total_loss)
        train_step(imgs, labels, metric)
        #cumulative_loss = cumulative_loss   loss
        #grads = tape.gradient(cumulative_loss, model.trainable_variables)
        #optimizer.apply_gradients(zip(grads, model.trainable_variables))

        print("rEpoch: {}, Batch: {}, Loss: {:.3f}, Accuracy: {:.3f}".format(
        epoch 1, i 1, metric.result().numpy(), accuracy.result().numpy()), flush=True, end='')
    print("")

 

Есть ли у вас какие-либо