keras Conv2d值错误：负维度大小

Question

我是 keras 新手，正在处理以 Tensorflow 作为后端的回归问题。

X1 = TrainingSet[:,0:603]
Y1 = TrainingSet[:,603:607]
###################################
#reshape Xtrain for CNN
X1 = X1.reshape(9999,3,201,1)

###################################
# create model
model = Sequential()

model.add(Conv2D(32, kernel_size=(3, 3), strides=(1, 1),
             activation='relu',
             input_shape=(3,201,1), data_format='channels_first'))
model.add(MaxPooling2D(pool_size=(2, 2), dim_ordering="tf",strides=(1, 1)))
model.add(Conv2D(32, (3, 3), activation='relu', data_format='channels_first'))
model.add(MaxPooling2D(pool_size=(2, 2), dim_ordering="tf",strides=(1, 1)))
model.add(Flatten())
model.add(Dense(1000, activation="tanh", kernel_initializer="uniform"))
model.add(Dense(4, activation="relu", kernel_initializer="uniform"))

# Compile model
model.compile(loss='mse', optimizer='adam', metrics=['mae'])

# Fit the model
history = model.fit(X1, Y1, validation_split=0.1, epochs=100, batch_size=100,  verbose=1)

# Calculate predictions
PredTestSet = model.predict(X1)

我得到了值错误

ValueError: Negative dimension size caused by subtracting 3 from 1 for 'conv2d_26/convolution' (op: 
'Conv2D') with input shapes: [?,201,1,3], [3,3,3,32].

我已经检查过类似的问题，例如 https://github.com/keras-team/keras/issues/7611

“Conv2D”的 1 减 3 导致的负维度大小

然而，他们的解决方案对我不起作用。

Answer 1

我能够重现您的错误

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, Activation, Flatten
from tensorflow.keras.layers import Conv2D, MaxPooling2D
import numpy as np

model = Sequential()

model.add(Conv2D(32, kernel_size=(3, 3), strides=(1, 1),
             activation='relu',
             input_shape=(3,201,1), data_format='channels_first'))
model.add(MaxPooling2D(pool_size=(2, 2), dim_ordering="tf",strides=(1, 1)))
model.add(Conv2D(32, (3, 3), activation='relu', data_format='channels_first'))
model.add(MaxPooling2D(pool_size=(2, 2), dim_ordering="tf",strides=(1, 1)))
model.add(Flatten())
model.add(Dense(1000, activation="tanh", kernel_initializer="uniform"))
model.add(Dense(4, activation="relu", kernel_initializer="uniform"))

model.summary()

输出：

---------------------------------------------------------------------------
InvalidArgumentError                      Traceback (most recent call last)
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/ops.py in _create_c_op(graph, node_def, inputs, control_inputs, op_def)
   1653   try:
-> 1654     c_op = pywrap_tf_session.TF_FinishOperation(op_desc)
   1655   except errors.InvalidArgumentError as e:

InvalidArgumentError: Negative dimension size caused by subtracting 3 from 1 for '{{node conv2d/Conv2D}} = Conv2D[T=DT_FLOAT, data_format="NCHW", dilations=[1, 1, 1, 1], explicit_paddings=[], padding="VALID", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true](conv2d_input, conv2d/Conv2D/ReadVariableOp)' with input shapes: [?,3,201,1], [3,3,3,32].

During handling of the above exception, another exception occurred:

ValueError                                Traceback (most recent call last)
14 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/ops.py in _create_c_op(graph, node_def, inputs, control_inputs, op_def)
   1655   except errors.InvalidArgumentError as e:
   1656     # Convert to ValueError for backwards compatibility.
-> 1657     raise ValueError(str(e))
   1658 
   1659   return c_op

ValueError: Negative dimension size caused by subtracting 3 from 1 for '{{node conv2d/Conv2D}} = Conv2D[T=DT_FLOAT, data_format="NCHW", dilations=[1, 1, 1, 1], explicit_paddings=[], padding="VALID", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true](conv2d_input, conv2d/Conv2D/ReadVariableOp)' with input shapes: [?,3,201,1], [3,3,3,32].

解决方案：

padding='same'这个问题可以通过包含到所有convolution层来解决，maxpooling通过这些层我们可以获得相同的结果dimensionality（在默认行为中，即padding = 'valid' 在卷积和最大池化期间会发生自动降维，并且将面临负维度问题）。

在解决上述问题时，遇到了另一个问题是dim_ordering在MaxPooling层中到期的。此选项在 Keras 2 中更改为“data_format”。

dim_ordering='tf'相当于data_format="channels_last"我相应地对网络进行了更改

model = Sequential()

model.add(Conv2D(32, kernel_size=(3, 3), strides=(1, 1),
             activation='relu',
             input_shape=(3,201,1), padding='same', data_format='channels_first'))
model.add(MaxPooling2D(pool_size=(2, 2), padding='same', data_format='channels_last',strides=(1, 1)))
model.add(Conv2D(32, (3, 3), activation='relu', padding='same', data_format='channels_first'))
model.add(MaxPooling2D(pool_size=(2, 2), padding='same', data_format='channels_last',strides=(1, 1)))
model.add(Flatten())
model.add(Dense(1000, activation="tanh", kernel_initializer="uniform"))
model.add(Dense(4, activation="relu", kernel_initializer="uniform"))

model.summary()

输出：

Model: "sequential_7"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_8 (Conv2D)            (None, 32, 201, 1)        896       
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 32, 201, 1)        0         
_________________________________________________________________
conv2d_9 (Conv2D)            (None, 32, 201, 1)        9248      
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 32, 201, 1)        0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 6432)              0         
_________________________________________________________________
dense_2 (Dense)              (None, 1000)              6433000   
_________________________________________________________________
dense_3 (Dense)              (None, 4)                 4004      
=================================================================
Total params: 6,447,148
Trainable params: 6,447,148
Non-trainable params: 0
_________________________________________________________________