def quantize_model(output_graph_file, test_data, input_holder, output):
"""Quantize tf model by using amct tool."""
batch = load_image(test_data, input_holder.shape)
input_name = input_holder.name
output_name = output.name
with tf.io.gfile.GFile(output_graph_file, mode='rb') as model:
graph_def = tf.compat.v1.GraphDef()
graph_def.ParseFromString(model.read())
graph = tf.Graph()
with graph.as_default():
tf.import_graph_def(graph_def, name='')
input_tensor = graph.get_tensor_by_name(input_name)
output_tensor = graph.get_tensor_by_name(output_name)
base_dir = os.path.dirname(output_graph_file)
config_path = os.path.join(base_dir, 'config.json')
amct.create_quant_config(config_file=config_path,
graph=graph,
skip_layers=[],
batch_num=1)
record_path = os.path.join(base_dir, 'record.txt')
amct.quantize_model(graph=graph,
config_file=config_path,
record_file=record_path)
with tf.compat.v1.Session(graph=graph) as sess:
sess.run(tf.compat.v1.global_variables_initializer())
sess.run(output_tensor, feed_dict={input_tensor: batch})
save_path = os.path.join(base_dir, 'tf_model')
amct.save_model(pb_model=output_graph_file,
outputs=[output_name[:-2]],
record_file=record_path,
save_path=save_path)
os.system('cp {}_quantized.pb {}'.format(save_path, output_graph_file))
logging.info('amct quantinize successfully.')
def main():
"""main process"""
args_check(ARGS)
outputs = [PREDICTIONS]
record_file = os.path.join(OUTPUTS, 'record.txt')
config_file = os.path.join(OUTPUTS, 'config.json')
with tf.io.gfile.GFile(ARGS.model, mode='rb') as model:
graph_def = tf.compat.v1.GraphDef()
graph_def.ParseFromString(model.read())
tf.import_graph_def(graph_def, name='')
graph = tf.compat.v1.get_default_graph()
amct.create_quant_config(config_file, graph)
save_dir = os.path.join(OUTPUTS, 'mobilenet_v2')
evaluator = MobileNetV2Evaluator(ARGS.dataset, ARGS.keyword, ARGS.num_parallel_reads, ARGS.batch_size)
amct.accuracy_based_auto_calibration(ARGS.model, outputs, record_file, config_file, save_dir, evaluator)
def quantize_all_layers(model_path, input_tensor_name, output_tensor_name):
"""Quantify all layers of the whole network."""
# Step one, load original model.
graph, input_tensor, output_tensor = load_model(model_path,
input_tensor_name,
output_tensor_name)
# Step two, generate the quantization config file.
config_path = os.path.join(OUTPUTS, 'config.json')
if ARGS.cfg_define is not None:
amct.create_quant_config(config_file=config_path,
graph=graph,
skip_layers=[],
batch_num=1,
config_defination=ARGS.cfg_define)
else:
amct.create_quant_config(config_file=config_path,
graph=graph,
skip_layers=[],
batch_num=1)
# Step three, quantize the model.
record_path = os.path.join(OUTPUTS, 'record.txt')
amct.quantize_model(graph=graph,
config_file=config_path,
record_file=record_path)
# Step four, calibrate and save the quantized model.
calibration_path = os.path.join(PATH, 'data/calibration')
batch = load_image(calibration_path)
with tf.compat.v1.Session() as session:
session.run(tf.compat.v1.global_variables_initializer())
session.run(output_tensor, feed_dict={input_tensor: batch})
amct.save_model(pb_model=model_path,
outputs=['MobilenetV2/Predictions/Reshape_1'],
record_file=record_path,
save_path=os.path.join(OUTPUTS, 'mobilenet_v2_all_layers'))
def main(): # pylint: disable=too-many-statements, too-many-locals, not-context-manager
"""
Before run this script, please check whether the following files
exist in the same directory.
calibration.jpg
COCO_labels.txt,
detection.jpg,
yolov3_tensorflow_1.5.pb,
:return: None
"""
# Step one, load the trained model.
# This sample will use YOLOv3 which is trained with COCO dataset
# and save as 'yolov3_tensorflow_1.5.pb'. Therefore, loading COCO
# labels and test image to do inference.
model_path = os.path.join(PATH, 'model/yolov3_tensorflow_1.5.pb')
with tf.io.gfile.GFile(model_path, mode='rb') as model:
graph_def = tf.compat.v1.GraphDef()
graph_def.ParseFromString(model.read())
tf.import_graph_def(graph_def, name='')
graph = tf.compat.v1.get_default_graph()
input_tensor = graph.get_tensor_by_name(INPUT_NAME + ':0')
xy_tensor = graph.get_tensor_by_name(XY_NAME + ':0')
confidence_tensor = graph.get_tensor_by_name(CONF_NAME + ':0')
probability_tensor = graph.get_tensor_by_name(PROB_NAME + ':0')
labels = []
label_path = os.path.join(PATH, 'data/COCO_labels.txt')
with open(label_path) as label_file:
for line in label_file:
labels.append(line[:-1])
image_path = os.path.join(PATH, 'data/detection.jpg')
image_input = preprocessing(image_path)
image_input = image_input.reshape([1, SIDE, SIDE, 3])
with tf.compat.v1.Session() as session:
xy, confidence, probability = session.run(
[xy_tensor, confidence_tensor, probability_tensor],
feed_dict={input_tensor: image_input})
xy = xy.reshape([-1, 4])
confidence = confidence.reshape([-1, 1])
probability = probability.reshape([-1, 80])
bbox_origin = np.concatenate((xy, confidence, probability), axis=-1)
# Step two, generate the quantization config file.
# The function 'create_quant_config' will generate a config file
# that describe how to quantize the model in graph. The config file
# is saved as JSON format, and you can edit the file to configurate
# your quantization parameters of each layers(support FC, CONV and
# DW) easily.
config_path = os.path.join(OUTPUTS, 'config.json')
cfg_define = os.path.join(PATH, 'src/yolo_quant.cfg')
amct.create_quant_config(config_file=config_path,
graph=graph,
config_defination=cfg_define)
# Step three, quantize the model.
# The function 'quantize_model' will quantize your model in graph
# according to config file.
record_path = os.path.join(OUTPUTS, 'record.txt')
amct.quantize_model(graph=graph,
config_file=config_path,
record_file=record_path)
# Step four, calibrate and save the quantized model.
# The quantization parameters require one or more batch data to do
# inference and find the optimal values. This process is referred to
# calibration. For example, we use 1 pictures for calibration. Be
# sure to initialize the quantization parameters first. If your
# model have variables, you must reload the checkpoint before
# inference. After calibration, you can save the quantized model from
# original model and record_file. The quantized model can be used for
# simulating test on CPU/GPU and evaluate the accuracy of quantized
# model. and it can also be used for ATC to generate the model
# running on Ascend AI Processor.
calibration_path = os.path.join(PATH, 'data/calibration.jpg')
image_calibration = preprocessing(calibration_path)
image_calibration = image_calibration.reshape([1, SIDE, SIDE, 3])
with tf.compat.v1.Session() as session:
session.run(tf.compat.v1.global_variables_initializer())
session.run([xy_tensor, confidence_tensor, probability_tensor],
feed_dict={input_tensor: image_calibration})
amct.save_model(pb_model=model_path,
outputs=[XY_NAME, CONF_NAME, PROB_NAME],
record_file=record_path,
save_path=os.path.join(OUTPUTS, 'yolo_v3'))
# Step five, reload and test the quantized model for 'Fakequant'.
model_path = os.path.join(OUTPUTS, 'yolo_v3_quantized.pb')
with tf.io.gfile.GFile(name=model_path, mode='rb') as model:
graph_def_reload = tf.compat.v1.GraphDef()
graph_def_reload.ParseFromString(model.read())
graph_reload = tf.Graph()
with graph_reload.as_default():
tf.import_graph_def(graph_def=graph_def_reload, name='')
input_tensor = graph_reload.get_tensor_by_name(INPUT_NAME + ':0')
xy_tensor = graph_reload.get_tensor_by_name(XY_NAME + ':0')
confidence_tensor = graph_reload.get_tensor_by_name(CONF_NAME + ':0')
probability_tensor = graph_reload.get_tensor_by_name(PROB_NAME + ':0')
with tf.compat.v1.Session(graph=graph_reload) as session:
xy, confidence, probability = session.run(
[xy_tensor, confidence_tensor, probability_tensor],
feed_dict={input_tensor: image_input})
xy = xy.reshape([-1, 4])
confidence = confidence.reshape([-1, 1])
probability = probability.reshape([-1, 80])
bbox_fakequant = np.concatenate((xy, confidence, probability), axis=-1)
bounds, classes, scores = postprocessing(bbox_origin, image_path)
origin_image = annotate(image_path, bounds, classes, scores, labels)
origin_image.save('origin.png', 'png')
origin_image.show('origin')
print('origin.png save successfully!')
bounds, classes, scores = postprocessing(bbox_fakequant, image_path)
quantize_image = annotate(image_path, bounds, classes, scores, labels)
quantize_image.save('quantize.png', 'png')
quantize_image.show('quantize')
print('quantize.png save successfully!')
def main(): # pylint: disable=too-many-locals, not-context-manager
"""
Before run this script, please check whether the following files
exist in the same directory.
classification.jpg,
folder 'calibration' contains 32 images
:return: None
"""
# Step one, inference the original model to obtain the original
# precision before quantization.
model_path = os.path.realpath('./model/mobilenetv2_tf.pb')
with tf.io.gfile.GFile(model_path, mode='rb') as model:
graph_def = tf.compat.v1.GraphDef()
graph_def.ParseFromString(model.read())
tf.import_graph_def(graph_def, name='')
graph = tf.compat.v1.get_default_graph()
input_tensor = graph.get_tensor_by_name(INPUT_NAME + ':0')
output_tensor = graph.get_tensor_by_name(OUTPUT_NAME + ':0')
image_path = os.path.realpath('./data/classification.jpg')
image_test = Image.open(image_path).resize([SIDE, SIDE])
image_test = np.array(image_test).astype(np.float32) / 128 - 1
image_test = image_test.reshape([1, SIDE, SIDE, 3])
with tf.compat.v1.Session() as session:
origin_prediction = session.run(output_tensor,
feed_dict={input_tensor: image_test})
# Step two, generate the quantization config file.
# The function 'create_quant_config' will generate a config file
# that describe how to quantize the model in graph. The config file
# is saved as JSON format, and you can edit the file to configurate
# your quantization parameters of each layers(support FC, CONV and
# DW) easily.
config_path = os.path.join(OUTPUTS, 'config.json')
amct.create_quant_config(config_file=config_path,
graph=graph,
skip_layers=[],
batch_num=1)
# Step three, quantize the model.
# The function 'quantize_model' will quantize your model in graph
# according to config file.
record_path = os.path.join(OUTPUTS, 'record.txt')
amct.quantize_model(graph=graph,
config_file=config_path,
record_file=record_path)
# Step four, calibrate and save the quantized model.
# The quantization parameters require one or more batch data to do
# inference and find the optimal values. This process is referred
# to calibration. For example, we use 32 pictures for calibration.
# Be sure to initialize the quantization parameters first. If your
# model have variables, you must reload the checkpoint before
# inference. After calibration, you can save the quantized model
# from original model and record_file. The quantized model can be
# used for simulating test on CPU/GPU and evaluate the accuracy of
# quantized model, and it can also be used for ATC to generate the
# model running on Ascend AI Processor.
calibration_path = os.path.realpath('./data/calibration')
batch = load_image(calibration_path)
with tf.compat.v1.Session() as session:
session.run(tf.compat.v1.global_variables_initializer())
session.run(output_tensor, feed_dict={input_tensor: batch})
amct.save_model(pb_model=model_path,
outputs=[OUTPUT_NAME],
record_file=record_path,
save_path=os.path.join(OUTPUTS, 'mobilenet_v2'))
# Step five, reload and test the quantized model for 'Fakequant'.
model_path = os.path.join(OUTPUTS, 'mobilenet_v2_quantized.pb')
with tf.io.gfile.GFile(name=model_path, mode='rb') as model:
graph_def_reload = tf.compat.v1.GraphDef()
graph_def_reload.ParseFromString(model.read())
graph_reload = tf.Graph()
with graph_reload.as_default():
tf.import_graph_def(graph_def=graph_def_reload, name='')
input_tensor = graph_reload.get_tensor_by_name(INPUT_NAME + ':0')
output_tensor = graph_reload.get_tensor_by_name(OUTPUT_NAME + ':0')
with tf.compat.v1.Session(graph=graph_reload) as session:
fakequant_prediction = session.run(
output_tensor, feed_dict={input_tensor: image_test})
print('Origin Model Prediction:\n',
'\tcategory index: %d\n' % origin_prediction.argmax(),
'\tcategory prob: %.3f\n' % round(origin_prediction.max(), 3),
end='')
print('Quantized Model Prediction:\n',
'\tcategory index: %d\n' % fakequant_prediction.argmax(),
'\tcategory prob: %.3f\n' % round(fakequant_prediction.max(), 3),
end='')
def main(): # pylint: disable=R0914
"""
Before run this script, please check whether the following files
exist in the same directory.
classification.jpg
folder 'calibration' contains 32 images
:return: None
"""
model_file = os.path.join(PATH, 'model/resnet_v1_50.pb')
load_graph(model_file)
graph = tf.get_default_graph()
input_tensor = graph.get_tensor_by_name('input:0')
output_tensor = graph.get_tensor_by_name('Reshape_1:0')
image_path = os.path.join(PATH, 'data/classification.jpg')
image_test = Image.open(image_path).resize([SIDE, SIDE])
image_test = np.array(image_test).astype(np.float32) / 128 - 1
image_test = image_test.reshape([1, SIDE, SIDE, 3])
print('inference with origin pb********************')
with tf.Session() as session:
origin_prediction = session.run(output_tensor,
feed_dict={input_tensor: image_test})
config_file = os.path.join(OUTPUTS, 'config.json')
record_file = os.path.join(OUTPUTS, 'record.txt')
amct.create_quant_config(config_file,
graph,
batch_num=1,
config_defination='./src/nuq_conf/nuq_quant.cfg')
amct.quantize_model(graph, config_file, record_file)
calibration_path = os.path.join(PATH, 'data/calibration')
batch = load_image(calibration_path)
with tf.Session() as session:
session.run(tf.global_variables_initializer())
session.run(output_tensor, feed_dict={input_tensor: batch})
amct.save_model(model_file, ['Reshape_1'], record_file,
os.path.join(OUTPUTS, 'resnet-50_v1'))
# reload and test the quantized model for 'Fakequant'.
model_file = os.path.join(OUTPUTS, 'resnet-50_v1_quantized.pb')
with tf.io.gfile.GFile(model_file, mode='rb') as model:
graph_def_reload = tf.GraphDef()
graph_def_reload.ParseFromString(model.read())
graph_reload = tf.Graph()
with graph_reload.as_default():
tf.import_graph_def(graph_def_reload, name='')
print('inference with quantized pb====================')
with tf.Session(graph=graph_reload) as session:
fakequant_prediction = session.run('Reshape_1:0',
feed_dict={'input:0': image_test})
print('Origin Model Prediction:\n',
'\tcategory index: %d\n' % origin_prediction.argmax(),
'\tcategory prob: %.3f\n' % round(origin_prediction.max(), 3),
end='')
print('Quantized Model Prediction:\n',
'\tcategory index: %d\n' % fakequant_prediction.argmax(),
'\tcategory prob: %.3f\n' % round(fakequant_prediction.max(), 3),
end='')
