python - Keras fit_generator 引发 ValueError 因为多个输入不正确

Question

我有一个工作模型，可以在小数据集上使用拟合时定期训练该模型。但是，由于数据集总量很大，我想切换到生成器。

我的模型是一个变分自动编码器，有 3 个输入层。首先，我使用常规拟合方法使模型正常工作。当我按如下方式运行模型时，它可以工作:

history = bvae.fit([data_class, data_regress, n_particles],
                        epochs = epochs,
                        verbose = 1,
                        batch_size = batch_size,
                        validation_split = 0.2,
                        callbacks = callback_list)

当我开始工作时，我想切换到生成器，以便我可以使用更多训练样本。我尝试按如下方式实现我的生成器:

class DataGenerator(keras.utils.Sequence):
    'Generates data for Keras'
    def __init__(self, filename, sample_size, batch_size):
        self.filename = filename
        self.sample_size = sample_size
        self.batch_size = batch_size          # batch size
        self.channel = h5py.File(self.filename, 'r')
        self.list_class_arg = range(3, 67, 4)
        self.list_regress_arg = [x for x in range(1, 67) if x not in self.list_class_arg]

    def __len__(self):
        'Denotes the number of batches per epoch'
        return int(np.floor(self.sample_size / self.batch_size))

    def __getitem__(self, index):
        'Generate one batch of data'
        # selects data
        events = self.channel['SM'][index*self.batch_size:(index+1)*self.batch_size]
        # splits events into class and regress
        events_class = self.class_norm_to_int(events[:,self.list_class_arg])
        events_regress = events[:,self.list_regress_arg]
        # count number of particles in event
        n_particles = self.count_particles(events)

        print(type(events_class))
        print(type(events_regress))
        print(type(n_particles))

        return events_class, events_regress, n_particles

    def count_particles(self, data):
        'Counts the number of particles per events and returns that number in an array'
        n_particles = np.zeros(len(data), dtype='int32')
        for i in range(len(data)):
            for j in range(3,67,4):
                if data[i,j] != 0:
                    n_particles[i] += 1
                else:
                    break
        return n_particles

    def class_norm_to_int(self, data_class):
        'Transform the particle type to an interger number'
        class_shape = np.shape(data_class)
        data_class = np.ndarray.flatten(data_class)
        data_class = [(x*2.1659782700803696)+5.2390088017131005 if x!=0 else x for x in data_class]
        data_class = np.reshape(data_class, newshape=class_shape)
        return data_class

但是我收到了这个错误:

Using TensorFlow backend.
Epoch 1/1
<class 'numpy.ndarray'>
<class 'numpy.ndarray'>
<class 'numpy.ndarray'>
Traceback (most recent call last):
  File "bvae_train_generator.py", line 96, in <module>
    callbacks = callback_list)
  File "/home/usr/venv/lib/python3.6/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)
  File "/home/usr/venv/lib/python3.6/site-packages/keras/engine/training.py", line 1658, in fit_generator
    initial_epoch=initial_epoch)
  File "/home/usr/venv/lib/python3.6/site-packages/keras/engine/training_generator.py", line 215, in fit_generator
    class_weight=class_weight)
  File "/home/usr/venv/lib/python3.6/site-packages/keras/engine/training.py", line 1443, in train_on_batch
    class_weight=class_weight)
  File "/home/usr/venv/lib/python3.6/site-packages/keras/engine/training.py", line 757, in _standardize_user_data
    exception_prefix='input')
  File "/home/usr/venv/lib/python3.6/site-packages/keras/engine/training_utils.py", line 105, in standardize_input_data
    str(len(data)) + ' arrays: ' + str(data)[:200] + '...')
ValueError: Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 3 array(s), but instead got the following list of 1 arrays: [array([[5., 5., 6., ..., 0., 0., 0.],
       [5., 0., 0., ..., 0., 0., 0.],
       [5., 7., 7., ..., 0., 0., 0.],
       ...,
       [5., 5., 0., ..., 0., 0., 0.],
       [5., 5., 0., ..., 0., 0., 0....

似乎认为我只返回一个包含一个数组的列表，但正如您在 print 语句的结果中看到的，所有三个给定变量都是 numpy.ndarray 类型

我尝试将 __getitem__ 中的返回更改为:

return [events_class, events_regress, n_particles]

或

return (events_class, events_regress, n_particles)

但是，这似乎并不能解决问题。欢迎任何帮助!

编辑:

我的模型具有以下形式:

def initialize_bvae(beta, n_particles, latent, activation_func, extra_layer = False):

    'Network parameters'
    input_shape_class = (n_particles, )
    input_shape_regress = (n_particles*3 + 2, )
    hidden_layer_dim = 512
    latent_dim = n_particles*latent


    'Build encoder'
    inputs_class = Input(shape=input_shape_class, name='input_1') # Input for classification
    inputs_class_t0 = Lambda(lambda i: i)(inputs_class)
    inputs_class_t1 = Lambda(lambda x: K.one_hot(K.cast(x, 'int32'), 8))(inputs_class_t0) # Onehot encode classification input
    inputs_class_t2 = Reshape((n_particles*8,))(inputs_class_t1)

    inputs_regress = Input(shape=input_shape_regress, name='input_2') # Input for regression

    inputs = Concatenate()([inputs_class_t2, inputs_regress]) # Combine inputs

    x = Dense(hidden_layer_dim, activation=activation_func)(inputs)
    if extra_layer:
        x = Dense(int(hidden_layer_dim), activation=activation_func)(x)
        x = Dense(int(hidden_layer_dim/2), activation=activation_func)(x)

    x = Dense(int(hidden_layer_dim/2), activation=activation_func)(x)
    x = Dense(int(hidden_layer_dim/4), activation=activation_func)(x)

    'Latent space'
    z_mean = Dense(latent_dim, activation='linear', name='z_mean')(x)
    z_log_var = Dense(latent_dim, activation='linear', name='z_log_var')(x)
    z = Lambda(sampling, output_shape=(latent_dim,), name='z')([z_mean, z_log_var])

    encoder = Model([inputs_class, inputs_regress], [z_mean, z_log_var, z], name='encoder')

    'Build decoder'
    inputs_latent = Input(shape=(latent_dim,), name='z_sampling')
    inputs_particles = Input(shape=(1,), name='input_3') # Input for particles

    # Implement mask
    mask_fn = lambda n: K.concatenate((K.ones(shape=latent*n, dtype='float32'), K.zeros(shape=latent_dim-(latent*n), dtype='float32')))
    mask_layer = Lambda(lambda x: K.map_fn(mask_fn, K.cast(x, 'int32'), dtype='float32'))(inputs_particles)
    masked_latent = Lambda(lambda x: x[0]*x[1])([inputs_latent, mask_layer])

    x = Dense(int(hidden_layer_dim/4), activation=activation_func)(masked_latent)
    x = Dense(int(hidden_layer_dim/2), activation=activation_func)(x)
    if extra_layer:
        x = Dense(int(hidden_layer_dim/2), activation=activation_func)(x)
        x = Dense(int(hidden_layer_dim), activation=activation_func)(x)
    x = Dense(hidden_layer_dim, activation=activation_func)(x)

    outputs_class_t = Dense(n_particles*8, activation='elu')(x)
    outputs_class_t2 = Reshape((n_particles, 8))(outputs_class_t)
    outputs_class = Conv1D(8, 1, activation='softmax')(outputs_class_t2) # Classification output
    outputs_regress = Dense(n_particles*3+2, activation='linear')(x) #Regression output

    decoder = Model([inputs_latent, inputs_particles], [outputs_class, outputs_regress], name='decoder')

    'Build autoencoder'
    outputs = decoder([encoder([inputs_class, inputs_regress])[2], inputs_particles])
    bvae = Model([inputs_class, inputs_regress, inputs_particles], outputs, name='bvae')

    'Calculate loss'
    classification_loss = categorical_crossentropy(inputs_class_t1, outputs[0])
    regression_loss = mse(inputs_regress, outputs[1])
    reconstruction_loss = 20*K.mean(classification_loss) + K.mean(regression_loss)
    kl_loss = -0.5 * K.sum((1+z_log_var-K.square(z_mean)-K.exp(z_log_var)), axis=-1)
    bvae_loss = K.mean((1-beta) * reconstruction_loss + beta * kl_loss)
    bvae.add_loss(bvae_loss)

    bvae.compile(optimizer = keras.optimizers.Adam(lr=1e-3, beta_1=0.9, beta_2=0.999))
    return bvae, encoder, decoder

编辑2:

这是我如何调用fit_generator:

history = bvae.fit_generator(load_data.DataGenerator(file_training, 1e6, 50000),
                        epochs = epochs,
                        verbose = 1,
                        workers=0,
                        validation_data = load_data.DataGenerator(file_validation, 1e5, 50000),
                        callbacks = callback_list)

Answer 1

您正在模型中定义多个输入 input_regress、input_class、input_latent 和 input_articles。所有这些都应该在拟合模型时输入数据，而您只是从 DataGenerator 返回 (None, n) 维数组。如果您提供所有输入，由于尺寸不匹配而导致的错误将会消失。