def cam_weighted_image(model, image_path, character_idx):
pixels = np.load(image_path)['pixels']
cam = visualize_cam(model,
layer_idx=-1,
filter_indices=[character_idx],
seed_input=pixels)
return np.uint8(pixels * np.dstack([cam] * 3))
def plot(predictions,
targets,
images,
original_images,
path_info,
prefix=''):
for i in range(len(targets)):
# if i < 140: continue
file_path = '../results/CAM/{}/{}{}'.format(
args['sherpa_trial'], prefix, i)
cam = visualize_cam(model,
len(model.layers) - 1, predictions[i],
images[i][np.newaxis])
plt.subplot(1, 2, 1)
plt.imshow(original_images[i])
plt.title('Original Image')
plt.tight_layout()
plt.axis('off')
plt.subplot(1, 2, 2)
plt.imshow(overlay(cam, (original_images[i]) * 255.))
plt.title('Heat Map')
plt.tight_layout()
plt.axis('off')
plt.suptitle(path_info.iloc[i]['path'].replace(
'/baldig/physicstest/ValleyFeverDogs/', ''))
plt.savefig(file_path)
def grad_cam(model, image):
model.layers[-1].activation = keras.activations.linear
return visualize_cam(
model,
layer_idx=len(model.layers) -
1, # modify this as the network changes, must be the last layer
filter_indices=[1], # index of importance in the layer, change
seed_input=image, # image of interest
penultimate_layer_idx=147, # layer prior to the GLOBAL max pooling layer
backprop_modifier=None, # no need to modify
grad_modifier=None) # no need to modify
weights = ['java_python.h5'
] #,'java_python_no_code.h5','java_python_pv_no_code.h5']
for weight in weights:
if weight == 'all_four.h5':
options = four_options
model = VGG(shape, 4)
else:
options = two_options
model = VGG(shape, 2)
# load weights file
model.load_weights('../../jp_Fold_0/' + weight)
# make directory for images
os.mkdir('jp_Images/' + weight.replace('.h5', '/'))
# predict classes for testing images
predicitions = model.predict(images)
# iterate over all the predictions to produce cam
for i in range(len(predicitions)):
# get class label from prediction
code = np.argmax(predicitions[i])
# label photo
name = '_Predicted ' + options[code]
location = 'jp_Images/' + weight.replace('.h5', '/') + str(i) + name
cam = visualize_cam(model,
len(model.layers) - 1, code,
images[i].reshape((1, ) + shape))
img = images[i]
plt.imshow(overlay(cam, img))
plt.savefig(location)
X_TRAIN, Y_TRAIN, X_TEST, Y_TEST)
# load weights file
model.load_weights('../../' + weights)
# make directory for images
os.mkdir('Images/' + weights.replace('.h5', '/'))
# predict classes for testing images
predicitions = model.predict(x_test)
# set options
if len(y_test[0]) > 2:
options = four_options
else:
options = two_options
# iterate over all the predictions to produce cam
for i in range(len(predicitions)):
# get class label from prediction
code = np.argmax(predicitions[i])
# label photo as correct/incorrect with label predicted
if np.argmax(y_test[i]) == code:
name = 'correct ' + options[code]
else:
name = 'Actual ' + options[np.argmax(
y_test[i])] + ' Predicted ' + options[code]
location = 'Images/' + weights.replace('.h5', '/') + str(i) + name
cam = visualize_cam(model,
len(model.layers) - 1, code,
x_test[i].reshape(1, 300, 300, 3))
img = x_test[i]
plt.imshow(overlay(cam, img))
plt.savefig(location)