如何在多输出 model 训练期间加载数据，而无需在 Keras 中进行迭代？

Question

I have a Keras model with 1 input and 2 outputs in TensorFlow 2. When calling model.fit I want to pass dataset as x=train_dataset and call model.fit once. The train_dataset is made with tf.data.Dataset.from_generator which yields: x1, y1, y2.

The only way I can run training is the following:

for x1, y1,y2 in train_dataset:
    model.fit(x=x1, y=[y1,y2],...)

How to tell TensorFlow to unpack variables and train without the explicit for loop? Using the for loop makes many things less practical, as well as usage of train_on_batch .

If I want to run model.fit(train_dataset, ...) the function doesn't understand what x and y are, even the model is defined like:

model = Model(name ='Joined_Model',inputs=self.x, outputs=[self.network.y1, self.network.y2])

It throws an error that it is expecting 2 targets while getting 1, even the dataset has 3 variables, which can be iterated trough in the loop.

The dataset and mini-batch are generated as:

def dataset_joined(self, n_epochs, buffer_size=32):
    dataset = tf.data.Dataset.from_generator(
        self.mbatch_gen_joined,
        (tf.float32, tf.float32,tf.int32),
        (tf.TensorShape([None, None, self.n_feat]),
            tf.TensorShape([None, None, self.n_feat]),
            tf.TensorShape([None, None])),
        [tf.constant(n_epochs)]
        )
    dataset = dataset.prefetch(buffer_size)
    return dataset

    def mbatch_gen_joined(self, n_epochs):
    for _ in range(n_epochs):
        random.shuffle(self.train_s_list)
        start_idx, end_idx = 0, self.mbatch_size
        for _ in range(self.n_iter):
            s_mbatch_list = self.train_s_list[start_idx:end_idx]
            d_mbatch_list = random.sample(self.train_d_list, end_idx-start_idx)
            s_mbatch, d_mbatch, s_mbatch_len, d_mbatch_len, snr_mbatch, label_mbatch, _ = \
                self.wav_batch(s_mbatch_list, d_mbatch_list)
            x_STMS_mbatch, xi_bar_mbatch, _ = \
                self.training_example(s_mbatch, d_mbatch, s_mbatch_len,
                d_mbatch_len, snr_mbatch)
            #seq_mask_mbatch = tf.cast(tf.sequence_mask(n_frames_mbatch), tf.float32)
            start_idx += self.mbatch_size; end_idx += self.mbatch_size
            if end_idx > self.n_examples: end_idx = self.n_examples

            yield x_STMS_mbatch, xi_bar_mbatch, label_mbatch

Answer 1

Keras models expect the Python generators or tf.data.Dataset objects provide the input data as a tuple with the format of (input_data, target_data) (or (input_data, target_data, sample_weights) ). Each of input_data or target_data could and should be a list/tuple if the model has multiple input/output layers. Therefore, in your code, the generated data should also be compatible with this expected format:

yield x_STMS_mbatch, (xi_bar_mbatch, label_mbatch)  # <- the second element is a tuple itself

Also, this should be considered in the arguments passed to the from_generator method as well:

dataset = tf.data.Dataset.from_generator(
    self.mbatch_gen_joined,
    output_types=(
        tf.float32,
        (tf.float32, tf.int32)
    ),
    output_shapes=(
        tf.TensorShape([None, None, self.n_feat]),
        (
            tf.TensorShape([None, None, self.n_feat]),
            tf.TensorShape([None, None])
        )
    ),
    args=(tf.constant(n_epochs),)
)

Answer 2

Use yield(x1, [y1,y2]) so model.fit will understand your generator output.