Hi @getjaidev,
Firstly, I do agree with you here that the output of the grader is not at all helpful in this case. I should see what I can do about it.
With that being said, your code was somehow passing the test case, but was incorrect. Your code was unnecessarily complex and long, and was referencing variables in not the sequential order as proposed in the assignment. In some instances, you were using the global variable model instead of using the function parameter of the model provided.
I have fixed your code in your notebook, you can take a look at it for more detail. It now passes all the tests. Please use the file C3_W3_Assignment.ipynb.
As a reference, this is the exercise code and you can see how the order was different in your notebook.
# UNQ_C1 (UNIQUE CELL IDENTIFIER, DO NOT EDIT)
def grad_cam(input_model, image, category_index, layer_name):
"""
GradCAM method for visualizing input saliency.
Args:
input_model (Keras.model): model to compute cam for
image (tensor): input to model, shape (1, H, W, 3), where H (int) is height W (int) is width
category_index (int): class to compute cam with respect to
layer_name (str): relevant layer in model
Return:
cam ()
"""
cam = None
### START CODE HERE (REPLACE INSTANCES OF 'None' with your code) ###
# 1. Get placeholders for class output and last layer
# Get the model's output
output_with_batch_dim = None
# Remove the batch dimension
output_all_categories = None
# Retrieve only the disease category at the given category index
y_c = None
# Get the input model's layer specified by layer_name, and retrive the layer's output tensor
spatial_map_layer = None
# 2. Get gradients of last layer with respect to output
# get the gradients of y_c with respect to the spatial map layer (it's a list of length 1)
grads_l = None
# Get the gradient at index 0 of the list
grads = None
# 3. Get hook for the selected layer and its gradient, based on given model's input
# Hint: Use the variables produced by the previous two lines of code
spatial_map_and_gradient_function = None
# Put in the image to calculate the values of the spatial_maps (selected layer) and values of the gradients
spatial_map_all_dims, grads_val_all_dims = None
# Reshape activations and gradient to remove the batch dimension
# Shape goes from (B, H, W, C) to (H, W, C)
# B: Batch. H: Height. W: Width. C: Channel
# Reshape spatial map output to remove the batch dimension
spatial_map_val = None
# Reshape gradients to remove the batch dimension
grads_val = None
# 4. Compute weights using global average pooling on gradient
# grads_val has shape (Height, Width, Channels) (H,W,C)
# Take the mean across the height and also width, for each channel
# Make sure weights have shape (C)
weights = None
# 5. Compute dot product of spatial map values with the weights
cam = None
### END CODE HERE ###
# We'll take care of the postprocessing.
H, W = image.shape[1], image.shape[2]
cam = np.maximum(cam, 0) # ReLU so we only get positive importance
cam = cv2.resize(cam, (W, H), cv2.INTER_NEAREST)
cam = cam / cam.max()
return cam
Congratulations on completing the specialisation!
All the best,
Mubsi
