def main(_):
# Build the bridge with intel model zoo configuration for inference run
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "../config/models.json")
model = ModelZooBridge(args.model, args.in_graph, args.data_location,
args.models_zoo_location, args.models_source_dir,
config_file_path)
# pick up the quantization calibration parameters from models.json
inference_cmd_gen_minmax_log = model.inference_calib_cmd
graph_converter_params = model.quantize_params_dict
inputs = []
outputs = []
excluded_ops = []
excluded_nodes = []
per_channel = False
if ModelZooBridge.INPUT_NODE_LIST in graph_converter_params.keys():
inputs = graph_converter_params[ModelZooBridge.INPUT_NODE_LIST]
if ModelZooBridge.OUTPUT_NODE_LIST in graph_converter_params.keys():
outputs = graph_converter_params[ModelZooBridge.OUTPUT_NODE_LIST]
if ModelZooBridge.EXCLUDED_OPS_LIST in graph_converter_params.keys():
excluded_ops = graph_converter_params[ModelZooBridge.EXCLUDED_OPS_LIST]
if ModelZooBridge.EXCLUDED_NODE_LIST in graph_converter_params.keys():
excluded_nodes = graph_converter_params[
ModelZooBridge.EXCLUDED_NODE_LIST]
if ModelZooBridge.PER_CHANNEL_FLAG in graph_converter_params.keys():
per_channel = graph_converter_params[ModelZooBridge.PER_CHANNEL_FLAG]
# Call the GraphConverter to do the FP32 calibration, INT8 quantization, and INT8 calibration
qt = converter.GraphConverter(args.in_graph, args.out_graph, inputs,
outputs, excluded_ops, excluded_nodes,
per_channel)
qt.debug = args.debug
qt.gen_calib_data_cmds = inference_cmd_gen_minmax_log
qt.convert()
def main(_):
c = None
per_channel_value = False
output_shape = args.model + "/predictions/Reshape_1"
image_size=224
if args.model == 'inception_v1':
output_shape = 'InceptionV1/Logits/Predictions/Reshape_1'
elif args.model == 'inception_v2':
output_shape = 'InceptionV2/Predictions/Reshape_1'
elif args.model == 'inception_v4':
output_shape = 'InceptionV4/Logits/Predictions'
image_size=299
elif args.model == 'mobilenet_v1':
per_channel_value = True
output_shape = 'MobilenetV1/Predictions/Reshape_1'
elif args.model == 'mobilenet_v2':
per_channel_value = True
output_shape = 'MobilenetV2/Predictions/Reshape_1'
elif args.model == 'vgg_16':
output_shape = 'vgg_16/fc8/squeezed'
elif args.model == 'vgg_19':
output_shape = 'vgg_19/fc8/squeezed'
elif args.model == 'nasnet_large' or args.model == 'pnasnet_large':
output_shape = 'final_layer/predictions'
image_size=331
if per_channel_value:
c = converter.GraphConverter(args.model_location, args.out_graph, ['input'], [output_shape],
per_channel=True)
else:
c = converter.GraphConverter(args.model_location, args.out_graph, ['input'], [output_shape])
c.debug = True
c.gen_calib_data_cmds = model_callback_cmds(args.data_location,output_shape,image_size)
c.convert()
def main(_):
print(args.inputs.split(','), args.outputs.split(','), args.output_graph)
if not os.path.exists(args.input_graph):
print("{} doesn't exist!".format(args.input_graph))
sys.exit(-1)
if args.inputs:
inputs = args.inputs.split(',')
else:
inputs = []
if args.outputs:
outputs = args.outputs.split(',')
else:
outputs = []
if args.excluded_ops:
excluded_ops = args.excluded_ops.split(',')
else:
excluded_ops = []
if args.excluded_nodes:
excluded_nodes = args.excluded_nodes.split(',')
else:
excluded_nodes = []
qt = converter.GraphConverter(args.input_graph, args.output_graph, inputs,
outputs, excluded_ops, excluded_nodes,
args.per_channel)
qt.debug = args.debug
if 'input_graph=' in args.callback:
prefix = args.callback.split('input_graph=')[0]
postfix = ' '.join(
args.callback.split('input_graph=')[-1].split(' ')[1:])
callback_cmd = prefix + 'input_graph={} ' + postfix
else:
callback_cmd = args.callback
qt.gen_calib_data_cmds = callback_cmd
qt.convert()
def quantization(input_graph, model_path, tf_records):
minigo_converter = converter.GraphConverter(
input_graph, model_path + '.pb', ["pos_tensor"],
["policy_output", "value_output"])
minigo_converter.gen_calib_data_cmds = minigo_callback_cmds(tf_records)
minigo_converter.quantize()