tf.keras 卷积架构不起作用

Question

以下问题是我在设计基本自动编码器架构时遇到的实际问题的缩减。以下示例足以准确重现我遇到的错误。我已经尝试了大约两天，但我找不到任何解决办法。

import tensorflow as tf
import random
import os

RES = [256, 256]
def generator_data(n):
    for i in range(n):
        for j in range(6):
            yield tf.zeros((1, 256, 256, 3)), tf.zeros((1, 256, 256, 3))

def mymodel():
    model = tf.keras.Sequential()
    model.add(tf.keras.layers.Conv2D(8, (3, 3), activation='relu', padding='same'))
    # 256 x 256 x 8
    model.add(tf.keras.layers.MaxPooling2D((2, 2), padding='same'))
    # 128 x 128 x 8
    model.add(tf.keras.layers.Conv2D(16, (3, 3), activation='relu', padding='same'))
    # 128 x 128 x 16
    model.add(tf.keras.layers.MaxPooling2D((2, 2), padding='same'))
    # 64 x 64 x 16
    model.add(tf.keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same'))
    # 64 x 64 x 32
    model.add(tf.keras.layers.MaxPooling2D((2, 2), padding='same'))
    # 32 x 32 x 32

    # 32 x 32 x 32
    model.add(tf.keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same'))
    # 32 x 32 x 32
    model.add(tf.keras.layers.UpSampling2D((2, 2)))
    # 64 x 64 x 32
    model.add(tf.keras.layers.Conv2D(16, (3, 3), activation='relu', padding='same'))
    # 64 x 64 x 16
    model.add(tf.keras.layers.UpSampling2D((2, 2)))
    # 128 x 128 x 16
    model.add(tf.keras.layers.Conv2D(8, (3, 3), activation='relu', padding='same'))
    # 128 x 128 x 8
    model.add(tf.keras.layers.UpSampling2D((2, 2)))
    # 256 x 256 x 8
    model.add(tf.keras.layers.Conv2D(1, (3, 3), activation='sigmoid', padding='same'))
    return model


if __name__ == "__main__":
    # import some data to play with
    x_val, y_val = zip(*generator_data(20))

    model = mymodel()
    optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)
    model.compile(optimizer=optimizer, loss=tf.keras.losses.MeanSquaredError())
    model(tf.zeros((1, 256, 256, 3)))
    model.summary()

    # generator_data(train_list)
    model.fit(x=generator_data(1000),
        validation_data=(list(x_val), list(y_val)),
        verbose=1, epochs=1000)

首先，我有一个 model.summary() 的奇怪行为，其中包含：

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
conv2d (Conv2D)              multiple                  224
_________________________________________________________________
max_pooling2d (MaxPooling2D) multiple                  0
_________________________________________________________________
conv2d_1 (Conv2D)            multiple                  1168
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 multiple                  0
_________________________________________________________________
conv2d_2 (Conv2D)            multiple                  4640
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 multiple                  0
_________________________________________________________________
conv2d_3 (Conv2D)            multiple                  9248
_________________________________________________________________
up_sampling2d (UpSampling2D) multiple                  0
_________________________________________________________________
conv2d_4 (Conv2D)            multiple                  4624
_________________________________________________________________
up_sampling2d_1 (UpSampling2 multiple                  0
_________________________________________________________________
conv2d_5 (Conv2D)            multiple                  1160
_________________________________________________________________
up_sampling2d_2 (UpSampling2 multiple                  0
_________________________________________________________________
conv2d_6 (Conv2D)            multiple                  73
=================================================================
Total params: 21,137
Trainable params: 21,137
Non-trainable params: 0

输出形状只有多个。我已经查过了 here，但解决方法似乎不起作用。但其次也是更重要的是我得到了错误：

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 1 array(s), for inputs ['input_1'] but instead got the following list of 120 arrays: [<tf.Tensor: shape=(1, 256, 256, 3), dtype=float32, numpy=
array([[[[0., 0., 0.],
         [0., 0., 0.],
         [0., 0., 0.],
         ...,
         [0., 0., 0.],
         [0., 0., 0.],
         [0....

这对我来说毫无意义。我的生成器 returns [batch, x-dim, y-dim, channel] （我也尝试过 [batch, channel, x-dim, y-dim] 但也没有成功）。在这种情况下，批次等于 1 而不是 120。正如我所说，无论如何我都无法 tackle/debug 这些问题，因此非常感谢您的帮助。我对 DL 很陌生，但在 python 中不是，我在 python-3.7

中使用 Tensorflow-2.1.0

非常感谢。

Answer 1

这是工作代码。

import tensorflow as tf
import random
import os
import numpy as np

RES = [256, 256]
def generator_data(n):
    for i in range(n):
        for j in range(1):
            yield tf.zeros((1, 256, 256, 3)), tf.zeros((1, 256, 256, 3))

def mymodel():
    model = tf.keras.Sequential()
    model.add(tf.keras.layers.Conv2D(8, (3, 3), activation='relu', padding='same'))
    # 256 x 256 x 8
    model.add(tf.keras.layers.MaxPooling2D((2, 2), padding='same'))
    # 128 x 128 x 8
    model.add(tf.keras.layers.Conv2D(16, (3, 3), activation='relu', padding='same'))
    # 128 x 128 x 16
    model.add(tf.keras.layers.MaxPooling2D((2, 2), padding='same'))
    # 64 x 64 x 16
    model.add(tf.keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same'))
    # 64 x 64 x 32
    model.add(tf.keras.layers.MaxPooling2D((2, 2), padding='same'))
    # 32 x 32 x 32

    # 32 x 32 x 32
    model.add(tf.keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same'))
    # 32 x 32 x 32
    model.add(tf.keras.layers.UpSampling2D((2, 2)))
    # 64 x 64 x 32
    model.add(tf.keras.layers.Conv2D(16, (3, 3), activation='relu', padding='same'))
    # 64 x 64 x 16
    model.add(tf.keras.layers.UpSampling2D((2, 2)))
    # 128 x 128 x 16
    model.add(tf.keras.layers.Conv2D(8, (3, 3), activation='relu', padding='same'))
    # 128 x 128 x 8
    model.add(tf.keras.layers.UpSampling2D((2, 2)))
    # 256 x 256 x 8
    model.add(tf.keras.layers.Conv2D(1, (3, 3), activation='sigmoid', padding='same'))
    return model


if __name__ == "__main__":
    # import some data to play with
    z = list(zip(*generator_data(2)))

    x_val = z[0][0]
    y_val = z[0][1]

    model = mymodel()
    optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)
    model.compile(optimizer=optimizer, loss=tf.keras.losses.MeanSquaredError())
    model(tf.zeros((1, 256, 256, 3)))
    model.summary()


    print(x_val.numpy().shape)
    print(y_val.numpy().shape)
    model.fit(x=generator_data(10),
        validation_data=(x_val, y_val),
        verbose=1, epochs=1)

您以错误的方式使用生成器解压缩。我将输出类型转换为一个列表，以便它可以订阅。其中一个有用的技巧是在每个步骤中打印 X、y 的形状和长度，以找出错误所在。

更新：

是的，完全正确，但是你需要传递一个形状为 [batch, 256, 256, 3] 的张量。但是如果 a 是一个列表，并且 a[0] 的形状为 [1, 256, 256, 3]，那么您需要将 a[0] 传递给模型，这就是我所做的。但是，你正在通过一个。但是 a 是一个列表，而不是一个 numpy array/tensor，即使我们将它转换为一个 numpy 数组，我们也会得到 shape = (1, 1, 256, 256, 3) - 这是无效的。

此外，在您的 generator_data 为什么要使用不必要的第二个循环？

def generator_data(n):
    for i in range(n):
        for j in range(1): # ??????? Why?
            yield tf.zeros((1, 256, 256, 3)), tf.zeros((1, 256, 256, 3))

Answer 2

我通过每晚更新到 tf-2.2 并使用 tf.data 模块解决了这些问题。

如果你遇到同样的问题，请看这里：

import tensorflow as tf
import random
import os
from functools import partial

RES = [256, 256]
def generator_data(n):
    for i in range(n):
        for j in range(6):
            yield tf.zeros((1, 256, 256, 3)), tf.zeros((1, 256, 256, 3))

def generator_data_val(n):
    for i in range(n):
        for j in range(6):
            yield tf.zeros((256, 256, 3)), tf.zeros((256, 256, 3))


def model():
    model = tf.keras.Sequential()
    model.add(tf.keras.layers.Conv2D(8, (3, 3), activation='relu', padding='same', input_shape=(256, 256, 3)))
    # 256 x 256 x 8
    model.add(tf.keras.layers.MaxPooling2D((2, 2), padding='same'))
    # 128 x 128 x 8
    model.add(tf.keras.layers.Conv2D(16, (3, 3), activation='relu', padding='same'))
    # 128 x 128 x 16
    model.add(tf.keras.layers.MaxPooling2D((2, 2), padding='same'))
    # 64 x 64 x 16
    model.add(tf.keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same'))
    # 64 x 64 x 32
    model.add(tf.keras.layers.MaxPooling2D((2, 2), padding='same'))
    # 32 x 32 x 32

    # 32 x 32 x 32
    model.add(tf.keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same'))
    # 32 x 32 x 32
    model.add(tf.keras.layers.UpSampling2D((2, 2)))
    # 64 x 64 x 32
    model.add(tf.keras.layers.Conv2D(16, (3, 3), activation='relu', padding='same'))
    # 64 x 64 x 16
    model.add(tf.keras.layers.UpSampling2D((2, 2)))
    # 128 x 128 x 16
    model.add(tf.keras.layers.Conv2D(8, (3, 3), activation='relu', padding='same'))
    # 128 x 128 x 8
    model.add(tf.keras.layers.UpSampling2D((2, 2)))
    # 256 x 256 x 8
    model.add(tf.keras.layers.Conv2D(1, (3, 3), activation='sigmoid', padding='same'))
    return model


if __name__ == "__main__":
    # import some data to play with
    x_val, y_val = zip(*generator_data_val(5))
    x_val, y_val = list(x_val), list(y_val)


    model = model()
    optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)
    model.compile(optimizer=optimizer, loss=tf.keras.losses.MeanSquaredError())
    model(tf.zeros((1, 256, 256, 3)))
    model.summary()
    train_dataset = generator_data(5)

    gen = partial(generator_data, n=5)
    train_dataset = tf.data.Dataset.from_generator(
        gen, output_types=(tf.float32, tf.float32),
        output_shapes=(tf.TensorShape([1, 256, 256, 3]), tf.TensorShape([1, 256, 256, 3]))).repeat()
    val_dataset = tf.data.Dataset.from_tensor_slices((x_val, y_val)).batch(2)

    # generator_data(train_list)
    model.fit(x=train_dataset,
        steps_per_epoch=40,
        validation_data=val_dataset,
        verbose=1, epochs=1000)

tf.keras 卷积架构不起作用

tf.keras convolutional architecture isn't working

python

autoencoder

keras

tensorflow

tensorflow2.0