def dream(self):
# Process image and return variable
self.processed_image = preprocess_image(self.created_image, True)
# Define optimizer for the image
# Earlier layers need higher learning rates to visualize whereas layer layers need less
optimizer = SGD([self.processed_image], lr=12, weight_decay=1e-4)
for i in range(1, 251):
optimizer.zero_grad()
# Assign create image to a variable to move forward in the model
x = self.processed_image
for index, layer in enumerate(self.model):
# Forward
x = layer(x)
# Only need to forward until we the selected layer is reached
if index == self.selected_layer:
break
# Loss function is the mean of the output of the selected layer/filter
# We try to minimize the mean of the output of that specific filter
loss = -torch.mean(self.conv_output)
print('Iteration:', str(i), 'Loss:',
"{0:.2f}".format(loss.data.numpy()))
# Backward
loss.backward()
# Update image
optimizer.step()
# Recreate image
self.created_image = recreate_image(self.processed_image)
# Save image every 20 iteration
if i % 10 == 0:
print(self.created_image.shape)
im_path = '../generated/ddream_l' + str(self.selected_layer) + \
'_f' + str(self.selected_filter) + '_iter' + str(i) + '.jpg'
save_image(self.created_image, im_path)
def generate(self):
initial_learning_rate = 6
for i in range(1, 150):
# Process image and return variable
self.processed_image = preprocess_image(self.created_image)
# Define optimizer for the image
optimizer = SGD([self.processed_image], lr=initial_learning_rate)
# Forward
output = self.model(self.processed_image)
# Target specific class
class_loss = -output[0, self.target_class]
print('Iteration:', str(i), 'Loss',
"{0:.2f}".format(class_loss.data.numpy()[0]))
# Zero grads
self.model.zero_grad()
# Backward
class_loss.backward()
# Update image
optimizer.step()
# Recreate image
self.created_image = recreate_image(self.processed_image)
# Save image
cv2.imwrite('../generated/c_specific_iteration_' + str(i) + '.jpg',
self.created_image)
return self.processed_image
def generate_inverted_image_specific_layer(self,
input_image,
img_size,
target_layer=3):
# Generate a random image which we will optimize
opt_img = Variable(1e-1 * torch.randn(1, 3, img_size, img_size),
requires_grad=True)
# Define optimizer for previously created image
optimizer = SGD([opt_img], lr=1e4, momentum=0.9)
# Get the output from the model after a forward pass until target_layer
# with the input image (real image, NOT the randomly generated one)
input_image_layer_output = \
self.get_output_from_specific_layer(input_image, target_layer)
# Alpha regularization parametrs
# Parameter alpha, which is actually sixth norm
alpha_reg_alpha = 6
# The multiplier, lambda alpha
alpha_reg_lambda = 1e-7
# Total variation regularization parameters
# Parameter beta, which is actually second norm
tv_reg_beta = 2
# The multiplier, lambda beta
tv_reg_lambda = 1e-8
for i in range(201):
optimizer.zero_grad()
# Get the output from the model after a forward pass until target_layer
# with the generated image (randomly generated one, NOT the real image)
output = self.get_output_from_specific_layer(opt_img, target_layer)
# Calculate euclidian loss
euc_loss = 1e-1 * self.euclidian_loss(
input_image_layer_output.detach(), output)
# Calculate alpha regularization
reg_alpha = alpha_reg_lambda * self.alpha_norm(
opt_img, alpha_reg_alpha)
# Calculate total variation regularization
reg_total_variation = tv_reg_lambda * self.total_variation_norm(
opt_img, tv_reg_beta)
# Sum all to optimize
loss = euc_loss + reg_alpha + reg_total_variation
# Step
loss.backward()
optimizer.step()
# Generate image every 5 iterations
if i % 5 == 0:
print('Iteration:', str(i), 'Loss:', loss.data.numpy())
recreated_im = recreate_image(opt_img)
im_path = '../generated/Inv_Image_Layer_' + str(target_layer) + \
'_Iteration_' + str(i) + '.jpg'
save_image(recreated_im, im_path)
# Reduce learning rate every 40 iterations
if i % 40 == 0:
for param_group in optimizer.param_groups:
param_group['lr'] *= 1 / 10
def visualise_layer_with_hooks(self):
# Hook the selected layer
self.hook_layer() # 여기서 선택된 필터
# Generate a random image
random_image = np.uint8(np.random.uniform(150, 180, (224, 224, 3)))
# Process image and return variable
processed_image = preprocess_image(
random_image, False) # 이미지 제로민 노말라이즈, torch tensor 형태로 만듬
# Define optimizer for the image
optimizer = Adam([processed_image], lr=0.1, weight_decay=1e-6)
for i in range(1, 31):
optimizer.zero_grad()
# Assign create image to a variable to move forward in the model
x = processed_image
for index, layer in enumerate(self.model):
# Forward pass layer by layer
# x is not used after this point because it is only needed to trigger
# the forward hook function
x = layer(x)
# print(">>>",x.size()) => 여기서 forward 해서 계속 feature map이 나온다.
# Only need to forward until the selected layer is reached
if index == self.selected_layer:
# (forward hook function triggered) => for 문나갈 떄 hook trigger 된다.
break
# Loss function is the mean of the output of the selected layer/filter
# We try to minimize the mean of the output of that specific filter
# print(self.conv_output.size())
loss = -torch.mean(
self.conv_output) # 특정 filter 의 output의 mean을 backprob으로 미니마이즈
exit()
print('Iteration:', str(i), 'Loss:',
"{0:.2f}".format(loss.data.numpy()))
# Backward
loss.backward()
# Update image
optimizer.step()
# Recreate image
self.created_image = recreate_image(
processed_image) # 다시 이미지 형태로 복원
# Save image
if i % 5 == 0:
im_path = '../generated/layer_vis_l' + str(self.selected_layer) + \
'_f' + str(self.selected_filter) + '_iter' + str(i) + '.jpg'
save_image(self.created_image, im_path)
def visualise_layer_without_hooks(self):
# Process image and return variable
# Generate a random image
random_image = np.uint8(np.random.uniform(150, 180, (224, 224, 3)))
# Process image and return variable
processed_image = preprocess_image(random_image, False)
# Define optimizer for the image
optimizer = Adam([processed_image], lr=0.1, weight_decay=1e-6)
for i in range(1, 31):
optimizer.zero_grad()
# Assign create image to a variable to move forward in the model
x = processed_image
for index, layer in enumerate(self.model):
# Forward pass layer by layer
x = layer(x)
if index == self.selected_layer:
# Only need to forward until the selected layer is reached
# Now, x is the output of the selected layer
break
# Here, we get the specific filter from the output of the convolution operation
# x is a tensor of shape 1x512x28x28.(For layer 17)
# So there are 512 unique filter outputs
# Following line selects a filter from 512 filters so self.conv_output will become
# a tensor of shape 28x28
self.conv_output = x[0, self.selected_filter]
# Loss function is the mean of the output of the selected layer/filter
# We try to minimize the mean of the output of that specific filter
loss = -torch.mean(self.conv_output)
print('Iteration:', str(i), 'Loss:',
"{0:.2f}".format(loss.data.numpy()))
# Backward
loss.backward()
# Update image
optimizer.step()
# Recreate image
self.created_image = recreate_image(processed_image)
# Save image
if i % 5 == 0:
im_path = '../generated/layer_vis_l' + str(self.selected_layer) + \
'_f' + str(self.selected_filter) + '_iter' + str(i) + '.jpg'
save_image(self.created_image, im_path)
def visualise_layer_without_hooks(self):
# Process image and return variable
self.processed_image = preprocess_image(self.created_image)
# Define optimizer for the image
# Earlier layers need higher learning rates to visualize whereas later layers need less
optimizer = SGD([self.processed_image], lr=5, weight_decay=1e-6)
for i in range(1, 51):
optimizer.zero_grad()
# Assign create image to a variable to move forward in the model
x = self.processed_image
for index, layer in enumerate(self.model):
# Forward pass layer by layer
x = layer(x)
if index == self.selected_layer:
# Only need to forward until the selected layer is reached
# Now, x is the output of the selected layer
break
# Here, we get the specific filter from the output of the convolution operation
# x is a tensor of shape 1x512x28x28.(For layer 17)
# So there are 512 unique filter outputs
# Following line selects a filter from 512 filters so self.conv_output will become
# a tensor of shape 28x28
self.conv_output = x[0, self.selected_filter]
# Loss function is the mean of the output of the selected layer/filter
# We try to minimize the mean of the output of that specific filter
loss = torch.mean(self.conv_output)
print('Iteration:', str(i), 'Loss:',
"{0:.2f}".format(loss.data.numpy()[0]))
# Backward
loss.backward()
# Update image
optimizer.step()
# Recreate image
self.created_image = recreate_image(self.processed_image)
# Save image
if i % 5 == 0:
cv2.imwrite(
'generated/layer_vis_l' + str(self.selected_layer) + '_f' +
str(self.selected_filter) + '_iter' + str(i) + '.jpg',
self.created_image)
def visualise_layer_with_hooks(self):
# Hook the selected layer
self.hook_layer()
# Process image and return variable
self.processed_image = preprocess_image(self.created_image)
# Define optimizer for the image
# Earlier layers need higher learning rates to visualize whereas later layers need less
optimizer = SGD([self.processed_image], lr=5, weight_decay=1e-6)
for i in range(1, 51):
optimizer.zero_grad()
# Assign create image to a variable to move forward in the model
x = self.processed_image
for index, layer in enumerate(self.model):
# Forward pass layer by layer
# x is not used after this point because it is only needed to trigger
# the forward hook function
x = layer(x)
# Only need to forward until the selected layer is reached
if index == self.selected_layer:
# (forward hook function triggered)
break
# Loss function is the mean of the output of the selected layer/filter
# We try to minimize the mean of the output of that specific filter
loss = torch.mean(self.conv_output)
print('Iteration:', str(i), 'Loss:',
"{0:.2f}".format(loss.data.numpy()[0]))
# Backward
loss.backward()
# Update image
optimizer.step()
# Recreate image
self.created_image = recreate_image(self.processed_image)
# Save image
if i % 5 == 0:
cv2.imwrite(
'generated/layer_vis_l' + str(self.selected_layer) + '_f' +
str(self.selected_filter) + '_iter' + str(i) + '.jpg',
self.created_image)
tolerance = vis_args.RESULTS.POINTING_GAME.tolerance
pg = PointingGame(vis_args.MODEL.n_classes, tolerance=tolerance)
for f, path in tqdm.tqdm(zip(files, paths)):
img = open_image(path)
prep_img = preprocess_image(img, h=h, w=w)
cam, pred = gcv2.generate_cam(prep_img, vis_args.DATASET.target_class)
guided_grads = GBP.generate_gradients(prep_img,
vis_args.DATASET.target_class)
cam_gb = guided_grad_cam(cam, guided_grads)
bw_cam_gb = convert_to_grayscale(cam_gb)
if vis_args.RESULTS.POINTING_GAME.state:
boxes = get_boxes(vis_args.RESULTS.DRAW_GT_BBOX.gt_src, f, img)
hit = pg.evaluate(boxes, bw_cam_gb.squeeze())
_ = pg.accumulate(hit[0], 1)
prep_img = recreate_image(prep_img)
r = alpha * bw_cam_gb + (1 - alpha) * prep_img
r = ((r - r.min()) / (r.max() - r.min()) * 255).astype(np.float32)
custom_save_gradient_images(bw_cam_gb,
vis_args.RESULTS.dir,
f,
obj="bw_ggradcam")
if vis_args.RESULTS.POINTING_GAME.state:
print(pg.print_stats())
print('Guided grad cam completed')