Keras中添加Softmax激活层报错，求正确实现方法

Hey there! Let's work through that Softmax issue you're hitting with your Keras CNN. I've got a couple of straightforward fixes for you, plus an explanation of why those errors popped up in the first place.

First, let's break down the problems: #

1. The TypeError: softmax() got an unexpected keyword argument 'axis' error happens because when you pass activation='softmax' to a Dense layer, Keras uses the underlying keras.activations.softmax function. In older versions of standalone Keras (not the TensorFlow-integrated tf.keras), this function doesn't support the axis parameter, but the code tries to pass it anyway—hence the error.
2. The warning you saw when using Softmax(axis=-1) as an activation argument is a key hint: advanced activation layers (like the Softmax layer class) shouldn't be passed as the activation parameter of another layer. Instead, they need to be added as separate layers in your model, just like your Conv2D or MaxPool2D layers.

Fix 1: Use the Softmax layer as a standalone layer (Recommended) #

This is the cleanest approach, and it follows Keras's best practices for advanced activations. Just define your final Dense layer without an activation, then add the Softmax layer right after it:

# Replace your original output layer code with this:
# First, create a Dense layer that outputs raw logits (no activation)
output_dense = Dense(units=n_classes, kernel_initializer='uniform')
cnn.add(output_dense)
# Add the Softmax layer as a separate step
from keras.layers import Softmax
cnn.add(Softmax(axis=-1))

Setting axis=-1 tells Softmax to compute probabilities across the last dimension (which is exactly what we want for multi-class classification).

Fix 2: Use tf.keras's softmax activation function (if using TensorFlow backend) #

If you prefer to keep the activation tied directly to the Dense layer, switch to using tf.keras.activations.softmax (this requires you to be using the TensorFlow-integrated version of Keras):

import tensorflow as tf
# Define your output layer with the tf.keras activation
output_layer = Dense(
    units=n_classes,
    activation=tf.keras.activations.softmax(axis=-1),
    kernel_initializer='uniform'
)
cnn.add(output_layer)

That said, Fix 1 is more readable and aligns better with how Keras expects you to use advanced activation layers.

Quick note on your original code #

I spotted a small mistake in your CNN setup: you're adding the Flatten layer twice! You have:

flat_layer_0 = Flatten()
cnn.add(flat_layer_0)
cnn.add(Flatten()) # This is redundant

Make sure to remove one of those Flatten() calls—having two will mess up your tensor dimensions.

Full modified code example #

Here's what your corrected CNN should look like (with Fix 1 implemented):

from keras.models import Sequential
from keras.layers import Conv2D, MaxPool2D, Dropout, Flatten, Dense, Softmax

cnn = Sequential()
kernelSize = (3, 3)
ip_activation = 'relu'
im_shape = (64, 64, 3) # Replace with your actual input shape
n_classes = 10 # Replace with your number of classes

ip_conv_0 = Conv2D(filters=32, kernel_size=kernelSize, input_shape=im_shape, activation=ip_activation)
cnn.add(ip_conv_0)
ip_conv_0_1 = Conv2D(filters=64, kernel_size=kernelSize, activation=ip_activation)
cnn.add(ip_conv_0_1)
pool_0 = MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding="same")
cnn.add(pool_0)
drop_layer_0 = Dropout(0.2)
cnn.add(drop_layer_0)

# Only one Flatten layer needed!
flat_layer_0 = Flatten()
cnn.add(flat_layer_0)

h_dense_0 = Dense(units=128, activation=ip_activation, kernel_initializer='uniform')
cnn.add(h_dense_0)
h_dense_1 = Dense(units=64, activation=ip_activation, kernel_initializer='uniform')
cnn.add(h_dense_1)

# Corrected output layers
output_dense = Dense(units=n_classes, kernel_initializer='uniform')
cnn.add(output_dense)
cnn.add(Softmax(axis=-1))

内容的提问来源于stack exchange，提问作者John