zero mean intensity of gradient for some cases

[HRKpython]

I am using Keras with tensorflow backend and I have fine-tuned the last Conv layer and FC layer of my network based on VGG weights. Now I am using grad-CAM technique to visualize which parts of my image triggered the prediction and I get all zeros for mean intensity of the gradient over a specific feature map channel.

I have 4 classes, for my test sample these are the prediction:

preds_sample = model.predict(x)
output>> array([[1., 0., 0., 0.]], dtype=float32)

# This is the "sample image" entry in the prediction vector    
image_0 = model.output[:, 0]

last_conv_layer = model.get_layer('conv2d_13')
grads = K.gradients(toilet_w, last_conv_layer.output)[0]
grads.shape
output>> TensorShape([Dimension(None), Dimension(512), Dimension(14), Dimension(14)])

Since I am using theano image ordering - when I calculate the mean of grads my axis is (0,2,3)

from keras import backend as K
K.set_image_dim_ordering('th')

pooled_grads = K.mean(grads, axis=(0,2,3))
pooled_grads.shape
output>> TensorShape([Dimension(512)])

iterate = K.function([model.input], [pooled_grads, last_conv_layer.output[0]])
pooled_grads_value, conv_layer_output_value = iterate([x])
pooled_grads_value.shape, conv_layer_output_value.shape
output>> ((512,), (512, 14, 14))

pooled_grads_value is all zero

Reference: https://github.com/fchollet/deep-learning-with-python-notebooks/blob/master/5.4-visualizing-what-convnets-learn.ipynb

I tested the algorithm with more images and found out it works for some of the images. Then I noticed that I have a dropout layer after my last conv layer. I did more research #2 and modified the code as:

last_conv_layer = model.get_layer('conv2d_13')
grads = K.gradients(sample_output, last_conv_layer.output)[0]

# normalization trick: we normalize the gradient
#grads = normalize_grad(grads)

pooled_grads = K.mean(grads, axis=(0, 2, 3))
iterate = K.function([model.input, K.learning_phase()], [pooled_grads, last_conv_layer.output[0]])

pooled_grads_value, conv_layer_output_value = iterate([x, 0])

But Still for some of the images all pooled_grads are zeros.

[hequn]

I got the same problem and haven't solved it yet (¦3」∠)

[hequn]

I did solve my error for the case that i did not choose the right layer output as the target.

[janphhe]

What layer did you take? Or what is the correct layer?

[hequn]

@janphhe In my project, I used Xception. The target layer will be the global_avg_pooling layer, we should get a vector which seemed to be flattened. The grad visualization was made on it.

[surfaniac]

Somebody has a solution for this. I get the same issue with a VGG19 network :/ I appreciate every help! Thanks a lot!

[Ada-Nick]

I'm having the same problem, i'm sure I have selected the correct conv layer as it works for some images and not others. I'm not using a pre-trained model.

EDIT: Change to using Leaky-relu activations to prevent the vanishing gradient problem has solved it for me, all though this probably isn't easy with pre-trained networks

[andreimargeloiu]

@Ada-Nick, can you please give more details about your solution/intuition?

[Ada-Nick]

@margiki Leaky-relu allows the gradient of negative values to be non-zero, preventing pooled_grads from being a null-matrix

[andreimargeloiu]

In my case, using LeakyReLU didn't solve the issue.

I delved deeper and found that the gradients computed by GradCAM were actually negative. Then, GradCAM applies a ReLU and the result was an empty map.