from_logits=True and from_logits=False get different training result for tf.losses.CategoricalCrossentropy for UNet

Question

I am doing the image semantic segmentation job with unet, if I set the Softmax Activation for last layer like this:

...
conv9 = Conv2D(n_classes, (3,3), padding = 'same')(conv9)
conv10 = (Activation('softmax'))(conv9)
model = Model(inputs, conv10)
return model
...

and then using loss = tf.keras.losses.CategoricalCrossentropy(from_logits=False) The training will not converge even for only one training image.

But if I do not set the Softmax Activation for last layer like this:

...
conv9 = Conv2D(n_classes, (3,3), padding = 'same')(conv9)
model = Model(inputs, conv9)
return model
...

and then using loss = tf.keras.losses.CategoricalCrossentropy(from_logits=True) The training will converge for one training image.

My groundtruth dataset is generated like this:

X = []
Y = []
im = cv2.imread(impath)
X.append(im)
seg_labels = np.zeros((height, width, n_classes))
for spath in segpaths:
    mask = cv2.imread(spath, 0)
    seg_labels[:, :, c] += mask
Y.append(seg_labels.reshape(width*height, n_classes))

Why? Is there something wrong for my usage?

This is my experiment code of git: https://github.com/honeytidy/unet You can checkout and run (can run on cpu). You can change the Activation layer and from_logits of CategoricalCrossentropy and see what i said.

Answer 1

Pushing the "softmax" activation into the cross-entropy loss layer significantly simplifies the loss computation and makes it more numerically stable.
It might be the case that in your example the numerical issues are significant enough to render the training process ineffective for the from_logits=False option.

You can find a derivation of the cross entropy loss (a special case of "info gain" loss) in this post. This derivation illustrates the numerical issues that are averted when combining softmax with cross entropy loss.