def generate(self, filename):
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/' + filename + 'c_specific_iteration_' + str(i) +
'.jpg', self.created_image)
return self.processed_image
def dream(self):
# Process image and return variable
self.processed_image = preprocess_image(self.created_image, False)
# 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()[0]))
# Backward
loss.backward()
# Update image
optimizer.step()
# Recreate image
self.created_image = recreate_image(self.processed_image)
# Save image every 20 iteration
if i % 20 == 0:
cv2.imwrite('generated/ddream_l' + str(self.selected_layer) +
'_f' + str(self.selected_filter) + '_iter' + str(
i) + '.jpg', self.created_image)
def generate_inverted_image_specific_layer(self,
input_image,
img_size,
filename,
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()[0])
x = recreate_image(opt_img)
cv2.imwrite(
'generated/' + filename + '_Inv_Image_Layer_' +
str(target_layer) + '_Iteration_' + str(i) + '.jpg', x)
# 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_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)
def dream(self,
device,
affine_size=4,
is_3d=False,
update_count=250,
is_depth_first=False):
# Process image and return variable
if type(self.created_image) is not np.ndarray:
self.created_image = np.asarray(self.created_image)
#if not is_3d:
# self.processed_image = preprocess_image(self.created_image, True)
else:
self.processed_image = torch.from_numpy(self.created_image).float()
self.processed_image /= 255
self.processed_image -= 0.485
self.processed_image /= 0.229
self.processed_image.requires_grad = 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.to(device)
for index, layer in (self.model._modules.items()):
# Forward
if (index == 'fc'): x = x.flatten(1)
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.cpu().numpy()))
# Backward
loss.backward()
# Update image
optimizer.step()
# Recreate image
if not is_3d and self.method is None:
self.created_image = recreate_image(self.processed_image)
else:
self.created_image = self.processed_image
# Save image every 20 iteration
if i % update_count == 0:
#print(self.created_image.shape)
if is_3d == False:
im_path = self.im_path +"CNN_Vis_DeepDream_" + str(self.selected_layer) + \
'_f' + str(self.selected_filter) + '_iter' + str(i) + '.png'
save_image(self.created_image, im_path)
else:
im_path = self.im_path +"CNN_Vis_DeepDream_" + str(self.selected_layer) + \
'_f' + str(self.selected_filter) + '_iter' + str(i) + '.nii.gz'
grads = self.created_image.squeeze().cpu().detach().numpy()
if is_depth_first:
grads = np.moveaxis(grads, 0, -1)
img = nb.Nifti1Image(grads, np.eye(affine_size))
nb.save(img, im_path)
def visualise_layer_with_hooks(self,
device,
opPath,
is_3d=False,
size=(224, 224, 3),
affine_size=4,
isDepthFirst=False):
# Hook the selected layer
self.hook_layer()
# Generate a random image
random_image = np.uint8(np.random.uniform(150, 180, size))
# Process image and return variable
if not is_3d:
processed_image = preprocess_image(random_image, False)
else:
processed_image = random_image
processed_image = torch.from_numpy(processed_image).float()
# 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.to(device)
for index, layer in (self.model._modules.items()):
if (index == 'fc'): x = x.flatten(1)
# 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
# print(self.conv_output)
loss = -torch.mean(self.conv_output)
#print('Iteration:', str(i), 'Loss:', "{0:.2f}".format(loss.data.cpu().numpy()))
# Backward
loss.backward()
# Update image
optimizer.step()
# Recreate image
if not is_3d: self.created_image = recreate_image(processed_image)
else: self.created_image = processed_image
# Save image
if i % 30 == 0:
if is_3d == False:
im_path = opPath + "Layer_Visualization_"+str(self.selected_layer) + \
'_f' + str(self.selected_filter) + '_iter' + str(i) + '.png'
save_image(self.created_image, im_path)
else:
im_path = opPath + "Layer_Visualization_"+str(self.selected_layer) + \
'_f' + str(self.selected_filter) + '_iter' + str(i) + '.nii.gz'
grads = self.created_image.squeeze().cpu().detach().numpy()
if isDepthFirst:
grads = np.moveaxis(grads, 0, -1)
img = nb.Nifti1Image(grads, np.eye(affine_size))
nb.save(img, im_path)
def visualise_layer_without_hooks(self,
device,
opPath,
is_3d=False,
size=(224, 224, 3),
affine_size=4,
isDepthFirst=False):
# Process image and return variable
# Generate a random image
random_image = np.uint8(np.random.uniform(150, 180, size))
# Process image and return variable
if not is_3d:
processed_image = preprocess_image(random_image, False)
else:
processed_image = random_image
processed_image = torch.from_numpy(processed_image).float()
# 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.to(device)
for index, layer in (self.model._modules.items()):
if (index == 'fc'): x = x.flatten(1)
# 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.cpu().numpy()))
# Backward
loss.backward()
# Update image
optimizer.step()
# Recreate image
if not is_3d:
self.created_image = recreate_image(processed_image)
else:
self.created_image = processed_image
# Save image
if i % 30 == 0:
if is_3d == False:
im_path = opPath + "Layer_Visualization_no_hook_" + str(self.selected_layer) + \
'_f' + str(self.selected_filter) + '_iter' + str(i) + '.png'
save_image(self.created_image, im_path)
else:
im_path = opPath + "Layer_Visualization_no_hook_" + str(self.selected_layer) + \
'_f' + str(self.selected_filter) + '_iter' + str(i) + '.nii.gz'
grads = self.created_image.squeeze().cpu().detach().numpy()
if isDepthFirst:
grads = np.moveaxis(grads, 0, -1)
img = nb.Nifti1Image(grads, np.eye(affine_size))
nb.save(img, im_path)