def compute_heatmap(gradcam, image_path, target_class):
# Obtain normalized image
normalized_image = get_normalized_image(image_path)
# Obtain the image mask by applying gradcam
mask = gradcam(normalized_image, target_index=target_class)
# Obtain the result of merging the mask with the torch image
heatmap = get_heatmap(mask)
return heatmap
def main():
# Retrieve the model
print(f'Loading weights for model: {model_name}...')
model, layer_name = load_fine_trained(model_name, model_weights)
print(f'Weights loaded!')
# Instantiate the gradCAM class
gradcam = GradCAM(model=model, \
target_layer=layer_name
)
# load pickle
data = pd.read_csv(experiment_data_fpath)
images = data['images'].tolist()
labels = data['labels'].apply(lambda x: x[1:-1].split(',')).tolist()
categories = data['categories'].apply(
lambda x: x[1:-1].split(',')).tolist()
main_category_index = data['main_category_index'].tolist()
for i, im_path in enumerate(images):
label_set = labels[i]
category_set = categories[i]
# n = random.randint(0,1)
n = int(main_category_index[i])
label = int(label_set[n])
category = category_set[n]
second_category = category_set[1 - n]
img_name = f'im_{i}_{category}_other_category_{second_category}'
# get guided grad cam map and guided backprop
""" Guided Back Propagation """
# Instantiate guided backpropagation to retrieve the mask
guided_backpropagation = GuidedBackPropagation(model=model)
# Obtain the initial image
input_image = get_normalized_image(im_path)
# Obtain the image mask by applying gradcam
guided_backpropagation_mask = guided_backpropagation(
input_image, label)
# Normalize the guided back propagation mask
guided_backpropagation_image = normalize(guided_backpropagation_mask)
# Save the image
save_image(
guided_backpropagation_image,
f'../results/experiment_5_1/{model_name}/{img_name}_guided_backpropagation.png'
)
""" Guided GradCAM"""
# Reset the guided_backpropagation_mask of the input, so that the input_image tensor can be used again
input_image.grad.zero_()
# Obtain the gradcam mask
mask = gradcam(input_image, target_index=label)
# Merge the gradcam with the guided back propagation to get guided gradcam results
guided_gradCAM = guided_backpropagation_mask * mask[..., np.newaxis]
# Normalize the guided gradCAM results to the [1,255] range
guided_gradCAM = normalize(guided_gradCAM)
# save image
save_image(
guided_gradCAM,
f'../results/experiment_5_1/{model_name}/{img_name}_guided_gradcam.png'
)
print(f"....Skipping unusual layer...{module_name}...")
continue
# Print information!
print(f"Investigating {model_name}'s {module_name} module")
# Instantiate the gradCAM class
gradcam = GradCAM( model = model, \
target_layer = module
)
# Obtain pre-normalized image
image = get_image('../../datasets/test_images/cat_dog.png')
# Obtain normalized image
normalized_image = get_normalized_image(
'../../datasets/test_images/cat_dog.png')
# Obtain the image mask by applying gradcam
mask = gradcam(normalized_image)
# Obtain the result of merging the mask with the torch image
heatmap = get_heatmap(mask)
# Save the heatmap!
save_image(
heatmap,
f'../../results/linear_probing/heatmap/cat_dog_heatmap_{model_name}_{module_name}.png'
)
# Merge the heatmap with
cam = heatmap + np.float32(image)
# Read the ground truth file to get list of images to process
with open(ground_truth_bounding_boxes_path,'rb') as opened_file:
ground_truth_bounding_boxes = pickle.load(opened_file)
# Get the list of images to iterate through
images = list(ground_truth_bounding_boxes.keys())
# Iterate through every image and compute bounding boxes
for image_identifier in tqdm(images):
# Strip the .xml from the identifier
image_identifier = image_identifier.strip('.xml')
# Define the image path
image_path = VOC2007_images_directory + image_identifier + '.jpg'
# Obtain normalized image
normalized_image = get_normalized_image(image_path)
# Obtain the image mask by applying gradcam
mask = gradcam(normalized_image)
# Obtain the result of merging the mask with the torch image
heatmap = get_heatmap(mask)
# Save the heatmap!
save_image(heatmap, f'../../../results/experiment_4_1/heatmaps/{image_identifier}.png')
def main():
# Retrieve the model
print(f'Loading weights for model: {model_name}...')
model, layer_name = load_fine_trained(model_name, model_weights)
print(f'Weights loaded!')
# Instantiate the gradCAM class
gradcam = GradCAM(model=model, \
target_layer=layer_name
)
# load pickle
data = pd.read_csv(experiment_data_fpath)
images = data['images'].tolist()
labels = data['labels'].tolist()
categories = data['categories'].tolist()
for i, im_path in enumerate(images):
label = int(labels[i])
category = categories[i]
img_name = f'im_{i}_{category}'
n = random.randint(0, 1) # if 0, guided back prop is model A and g-grad cam is model B, else guided grad cam is model A and guided back prop is model B
# get guided grad cam map and guided backprop
""" Guided Back Propagation """
# Instantiate guided backpropagation to retrieve the mask
guided_backpropagation = GuidedBackPropagation(model=model)
# Obtain the initial image
input_image = get_normalized_image(im_path)
# Obtain the image mask by applying gradcam
guided_backpropagation_mask = guided_backpropagation(input_image, label)
# Normalize the guided back propagation mask
guided_backpropagation_image = normalize(guided_backpropagation_mask)
# Save the image
img_dir = f'../results/experiment_5_2/{img_name}/guided_back_prop'
Path(img_dir).mkdir(parents=True, exist_ok=True)
save_image(guided_backpropagation_image, f'{img_dir}/{model_name}.png')
""" Guided GradCAM"""
# Reset the guided_backpropagation_mask of the input, so that the input_image tensor can be used again
input_image.grad.zero_()
# Obtain the gradcam mask
mask = gradcam(input_image, target_index=label)
# Merge the gradcam with the guided back propagation to get guided gradcam results
guided_gradCAM = guided_backpropagation_mask * mask[..., np.newaxis]
# Normalize the guided gradCAM results to the [1,255] range
guided_gradCAM = normalize(guided_gradCAM)
# save image
img_dir = f'../results/experiment_5_2/{img_name}/guided_gradcam'
Path(img_dir).mkdir(parents=True, exist_ok=True)
save_image(guided_gradCAM, f'{img_dir}/{model_name}.png')