Remove softmax classification layer and compute gradient

[Jamesswiz]

Hi,
I want to remove the softmax layer and then compute gradients using guided backpropogation. I used pop() for this:

model=load_model('nnet.h5', custom_objects={'keras':keras})
model.layers.pop()

However, I am getting the following error while using:
from guided_backprop import GuidedBackprop
guided_bprop = GuidedBackprop(model)

  Error: The name 'activation_5_1/Softmax:0' refers to a Tensor which does not exist.

I guess pop() does not changes the TF network graph. Can you please help how do I handle this?

Q: does your code ensures that the gradients are with respect to the output node with maximum activation?

[experiencor]

Hi @Jamesswiz! I remember that pop does not cleanly remove the last layer. You need to do something like model = Model(model.layers[0], model.layers[-1]). I know that it's hacky but I don't know of other ways to fix it.

[Jamesswiz]

HI, @experiencor thanks for help! Well the correct way is:

inp=model.input
outp=model.layers[-2].output
model1=keras.models.Model(inp,outp)

but again it doesn't works with your code and I get the error: The name 'dense_2_13/BiasAdd:0' refers to a Tensor which does not exist.

Now the code is trying to find the bias adding in TF graph! any idea how to deal with this?

[experiencor]

@Jamesswiz Sorry I cannot help you anymore. I'm quite slow with tensorflow bugs.

[Jamesswiz]

@experiencor No problem, I think I have figured it out. Going to test it on various examples to check if things are right.

Anyway, It will be good if you can update your code and mention this case also: since we should remove softmax layer to see gradients well, as output is depending on previous all nodes which we don't want. I am not sure what is usual convention in image processing as I work in speech. I guess its better to consider the previous layer and then propagate error backwards?

Again great stuff! thanks for sharing the code and helping me resolve the bugs
Cheers!

[experiencor]

@Jamesswiz Yes, sure. Can you share the way that you resolve it? Or better make a pull request 😁.

[Jamesswiz]

@experiencor Hi, well I am new to git so not sure how pull request works. Actually the error was due to the saved model files in /tmp/ directory because of which you need to reset the jupyter-nootbook kernel each time you make changes in the TF graph with the current session on. The code is working fine and results are interesting. Haven't tested on images though.

---

Remove softmax layer from keras model:

  model  = load_model('nnet.h5', custom_objects={'keras':keras})   # Load model
  input    = model.input                            # define input layer
  output  = model.layers[-2].output         # define output layer, [-1] is last layer
  
  # define new model with modified TF graph
  model_new = keras.models.Model(input,output)                            

---

Tip: one can compute gradients with respect to any layer's input and output by defining new model in this way.

[Jamesswiz]

Hi @experiencor ,
I need clarification for the operation of getting the output "self.imported_y" for computing gradients in guided_backprop.py i.e.,

self.imported_y = self.guided_graph.get_tensor_by_name(model.output.name)[0][output_index]

By default output_index is set to 0.
Does this means we are computing the gradients with respect to first output node or class?

waiting for a reply!

Best