python - 第二个 MaxPool2D 之后的宽度、高度比下采样因子小 1

Question

我对 ML 和使用 Keras 和 TF 创建 NN 还很陌生，所以我正在按照这里的教程创建一个验证码 ocr 模型，并为我自己的验证码图像稍微调整代码。

我正在使用以下函数来构建网络：

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

# image dimensions
image_width = 120
image_height = 40

# batch size for training and validation
batch_size = 16

def build_model():
  # inputs
  input_image = layers.Input(
      shape=(image_width, image_height, 1), name='image', dtype='float32'
  )
  labels = layers.Input(shape = (None,), name='label', dtype='float32')

  # first conv block
  x = layers.Conv2D(
      32,
      (3, 3),
      activation='relu',
      kernel_initializer='he_normal',
      padding='same',
      name='Conv1'
  )(input_image)
  x = layers.MaxPool2D((2, 2), name='Pool1')(x)
  print()

  # second conv block
  x = layers.Conv2D(
      64,
      (3, 3),
      activation='relu',
      kernel_initializer='he_normal',
      name='Conv2'
  )(x)
  x = layers.MaxPool2D((2, 2), name='Pool2')(x)

  # reshape
  new_shape = ((image_width // 4), (image_height // 4) * 64)
  x = layers.Reshape(target_shape=new_shape, name='Reshape')(x)
  x = layers.Dense(64, activation='relu', name='Dense1')(x)
  x = layers.Dropout(0.2)(x)

  # recurrent block
  x = layers.Bidirectional(layers.LSTM(128, return_sequences=True, dropout=0.25))(x)
  x = layers.Bidirectional(layers.LSTM(64, return_sequences=True, dropout=0.25))(x)

  # output
  x = layers.Dense(11, activation='softmax', name='Dense2')(x)
  output = CTCLayer(name='CTCLoss')(labels, x)

  # define model
  model = keras.models.Model(
      inputs=[input_image, labels], outputs=output, name='CaptchaOCRv1'
  )
  model.compile(optimizer=keras.optimizers.Adam())
  return model

给定输入图像尺寸 (120, 40, 1) 和两个 (2,2) 池化层，我希望重塑层输入的形状为 (30, 10, 64)。但是我收到以下错误：

ValueError: total size of new array must be unchanged, input_shape = [29, 9, 64], output_shape = [30, 640]

我不明白为什么宽度和高度比因子 4 下采样小 1。谁能指出我的错误？

标签： pythontensorflowkeras