def inference(self, image_list):
"""Do an inference on the model with a set of inputs.
# Arguments:
image_list: The input image list
Return the result of the inference.
"""
image = image_list[0]
image = self.linear_to_srgb(image)
image = (image * 255).astype(np.uint8)
kernel = self.kernel_size * 2 + 1
blur = cv2.GaussianBlur(image, (kernel, kernel), 0)
blur = blur.astype(np.float32) / 255.
blur = self.srgb_to_linear(blur)
# If make_blur button is pressed in Nuke
if self.make_blur:
script_msg = message_pb2.FieldValuePairAttrib()
script_msg.name = "PythonScript"
# Create a Python script message to run in Nuke
python_script = self.blur_script(blur)
script_msg_val = script_msg.values.add()
script_msg_str = script_msg_val.string_attributes.add()
script_msg_str.values.extend([python_script])
return [blur, script_msg]
return [blur]
def inference(self, image_list):
"""Do an inference on the model with a set of inputs.
# Arguments:
image_list: The input image list
Return the result of the inference.
"""
image = image_list[0]
image = self.linear_to_srgb(image).copy()
image = (image * 255).astype(np.uint8)
if not hasattr(self, 'model'):
# Initialise tensorflow graph
tf.compat.v1.reset_default_graph()
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth=True
self.sess=tf.compat.v1.Session(config=config)
# Load most recent trained model
self.model = self.load_model()
self.class_labels = (self.checkpoint_name.split('.')[0]).split('_')
# If checkpoint name has changed, load new checkpoint
if self.prev_ckpt_name != self.checkpoint_name or self.checkpoint_name == '':
self.model = self.load_model()
self.class_labels = (self.checkpoint_name.split('.')[0]).split('_')
# If checkpoint correctly loaded, update previous checkpoint name
self.prev_ckpt_name = self.checkpoint_name
image = cv2.resize(image, dsize=(224, 224), interpolation=cv2.INTER_NEAREST)
# Predict on new data
image_batch = np.expand_dims(image, 0)
# Preprocess a numpy array encoding a batch of images (RGB values within [0, 255])
image_batch = tf.keras.applications.mobilenet.preprocess_input(image_batch)
start = time.time()
y_prob = self.model.predict(image_batch)
y_class = y_prob.argmax(axis=-1)[0]
duration = time.time() - start
# Print results on server side
print('Inference duration: {:4.3f}s'.format(duration))
class_scores = str(["{0:0.4f}".format(i) for i in y_prob[0]]).replace("'", "")
print("Class scores: {} --> Label: {}".format(class_scores, self.class_labels[y_class]))
# If get_label button is pressed in Nuke
if self.get_label:
# Send back which class was detected
script_msg = message_pb2.FieldValuePairAttrib()
script_msg.name = "PythonScript"
# Create a Python script message to run in Nuke
nuke_msg = "Class scores: {}\\nLabel: {}".format(class_scores, self.class_labels[y_class])
python_script = "nuke.message('{}')\n".format(nuke_msg)
script_msg_val = script_msg.values.add()
script_msg_str = script_msg_val.string_attributes.add()
script_msg_str.values.extend([python_script])
return [image_list[0], script_msg]
return [image_list[0]]
def inference(self, image_list):
"""Do an inference on the model with a set of inputs.
# Arguments:
image_list: The input image list
Return the result of the inference.
"""
image = image_list[0]
image = linear_to_srgb(image).copy()
if not hasattr(self, 'sess'):
# Initialise tensorflow graph
tf.compat.v1.reset_default_graph()
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
self.sess = tf.compat.v1.Session(config=config)
# Input is stacked histograms of original and gamma-graded images.
input_shape = [1, 2, 100]
# Initialise input placeholder size
self.input = tf.compat.v1.placeholder(tf.float32,
shape=input_shape)
self.model = baseline_model(
input_shape=input_shape[1:],
output_param_number=self.output_param_number)
self.infer_op = self.model(self.input)
# Load latest model checkpoint
self.saver = tf.compat.v1.train.Saver()
self.load(self.model)
self.prev_ckpt_name = self.checkpoint_name
# If checkpoint name has changed, load new checkpoint
if self.prev_ckpt_name != self.checkpoint_name or self.checkpoint_name == '':
self.load(self.model)
# If checkpoint correctly loaded, update previous checkpoint name
self.prev_ckpt_name = self.checkpoint_name
# Preprocess image same way we preprocessed it for training
# Here for gamma correction compute histograms
def histogram(x, value_range=[0.0, 1.0], nbins=100):
"""Return histogram of tensor x"""
h, w, c = x.shape
hist = tf.histogram_fixed_width(x, value_range, nbins=nbins)
hist = tf.divide(hist, h * w * c)
return hist
with tf.compat.v1.Session() as sess:
# Convert to grayscale
img_gray = tf.image.rgb_to_grayscale(image)
img_gray = tf.image.resize(img_gray,
[self.patch_size, self.patch_size])
# Apply gamma correction
img_gray_grade = tf.math.pow(img_gray, self.gamma_to_predict)
img_grade = tf.math.pow(image, self.gamma_to_predict)
# Compute histograms
img_hist = histogram(img_gray)
img_grade_hist = histogram(img_gray_grade)
hists_op = tf.stack([img_hist, img_grade_hist], axis=0)
hists, img_grade = sess.run([hists_op, img_grade])
res_img = srgb_to_linear(img_grade)
hists_batch = np.expand_dims(hists, 0)
start = time.time()
# Run model inference
inference = self.sess.run(self.infer_op,
feed_dict={self.input: hists_batch})
duration = time.time() - start
print('Inference duration: {:4.3f}s'.format(duration))
res = inference[-1]
print("Predicted gamma: {}".format(res))
# If predict button is pressed in Nuke
if self.predict:
script_msg = message_pb2.FieldValuePairAttrib()
script_msg.name = "PythonScript"
# Create a Python script message to run in Nuke
python_script = self.nuke_script(res)
script_msg_val = script_msg.values.add()
script_msg_str = script_msg_val.string_attributes.add()
script_msg_str.values.extend([python_script])
return [res_img, script_msg]
return [res_img]