Closed Eduardo-Assuncao-hub closed 1 month ago
zhouyuchong
commented
1 month ago @Eduardo-Assuncao-hub there are several version arcface trained on different backbones and datasets. it may not wrong just different datasets. it doesn't matter so much as soon as you make sure the features are generated by the same model.
Eduardo-Assuncao-hub
commented
1 month ago @zhouyuchong thanks for your reply, Below are the cosine similarity of your features and my feartures in the knkow_faces image samples . Comparing Monica with Ross, Phoebe, Chandler, and Rachel: zhouyuchong _similarities: [[-0.0767541 -0.10237912 -0.03609521 0.1494083 ]] my_similarities: [[0.97455305 0.9796156 0.9683327 0.9882584 ]] In my results, the cosine similatity is very height for all sample. This means that there is no discrimination between faces. Because of that I think my model is not ok.
Note: Im using the sklearn.metrics.pairwise cosine_similarity.
zhouyuchong
commented
1 month ago these scores seem weird indeed. you can try these onnx file. they are from the official insightface, no differences. as for onnx2trt, you can use the scripts below:
'''
Author: zhouyuchong
Date: 2024-05-30 11:15:58
Description:
LastEditors: zhouyuchong
LastEditTime: 2024-09-14 11:23:59
'''
import os
import argparse
from loguru import logger
import tensorrt as trt
import math
TRT_LOGGER = trt.Logger()
def build_engine(onnx_file_path, inputname, engine_file_path="", set_input_shape=None):
"""Takes an ONNX file and creates a TensorRT engine to run inference with"""
network_creation_flag = 0
network_creation_flag = 1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
network_creation_flag
) as network, builder.create_builder_config() as config, trt.OnnxParser(
network, TRT_LOGGER
) as parser, trt.Runtime(
TRT_LOGGER
) as runtime:
config.set_memory_pool_limit(
trt.MemoryPoolType.WORKSPACE, 1 << 28
) # 256MiB
# Parse model file
if not os.path.exists(onnx_file_path):
print(
"ONNX file {} not found, please run yolov3_to_onnx.py first to generate it.".format(
onnx_file_path
)
)
exit(0)
print("Loading ONNX file from path {}...".format(onnx_file_path))
with open(onnx_file_path, "rb") as model:
print("Beginning ONNX file parsing")
if not parser.parse(model.read()):
print("ERROR: Failed to parse the ONNX file.")
for error in range(parser.num_errors):
print(parser.get_error(error))
return None
# set input shape
profile = builder.create_optimization_profile()
logger.debug("total input layer: {}".format(network.num_inputs))
logger.debug(network.num_outputs)
output = network.get_output(0)
logger.debug(output.shape)
for i in range(network.num_inputs):
input = network.get_input(i)
# assert input.shape[0] == -1
logger.debug("input layer-{}: {}".format(i, input.name))
profile.set_shape(inputname, set_input_shape[0], set_input_shape[1], set_input_shape[2])
config.add_optimization_profile(profile)
logger.debug("build, may take a while...")
plan = builder.build_serialized_network(network, config)
engine = runtime.deserialize_cuda_engine(plan)
print("Completed creating Engine")
with open(engine_file_path, "wb") as f:
f.write(plan)
return engine
def main(args):
trt_file_name = args.onnx.replace('.onnx', '_bs{}.trt'.format(args.batch))
input_shape = [(1, 3, args.size, args.size), (math.ceil(args.batch/2), 3, args.size, args.size), (args.batch, 3, args.size, args.size)]
logger.debug("set input shape: {}".format(input_shape))
build_engine(args.onnx, args.name, trt_file_name, input_shape)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--onnx', type=str, default='', help='onnx path', required=True)
parser.add_argument('-s', '--size', type=int, default=640, help='input shape')
parser.add_argument('-b', '--batch', type=int, default=1, help='max batch size')
parser.add_argument('-n', '--name', type=str, default='images', help='input name')
args = parser.parse_args()
main(args=args)
take a try!
Eduardo-Assuncao-hub
commented
1 month ago these scores seem weird indeed. you can try these onnx file. they are from the official insightface, no differences. as for onnx2trt, you can use the scripts below:
''' Author: zhouyuchong Date: 2024-05-30 11:15:58 Description: LastEditors: zhouyuchong LastEditTime: 2024-09-14 11:23:59 ''' import os import argparse from loguru import logger import tensorrt as trt import math TRT_LOGGER = trt.Logger() def build_engine(onnx_file_path, inputname, engine_file_path="", set_input_shape=None): """Takes an ONNX file and creates a TensorRT engine to run inference with""" network_creation_flag = 0 network_creation_flag = 1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH) with trt.Builder(TRT_LOGGER) as builder, builder.create_network( network_creation_flag ) as network, builder.create_builder_config() as config, trt.OnnxParser( network, TRT_LOGGER ) as parser, trt.Runtime( TRT_LOGGER ) as runtime: config.set_memory_pool_limit( trt.MemoryPoolType.WORKSPACE, 1 << 28 ) # 256MiB # Parse model file if not os.path.exists(onnx_file_path): print( "ONNX file {} not found, please run yolov3_to_onnx.py first to generate it.".format( onnx_file_path ) ) exit(0) print("Loading ONNX file from path {}...".format(onnx_file_path)) with open(onnx_file_path, "rb") as model: print("Beginning ONNX file parsing") if not parser.parse(model.read()): print("ERROR: Failed to parse the ONNX file.") for error in range(parser.num_errors): print(parser.get_error(error)) return None # set input shape profile = builder.create_optimization_profile() logger.debug("total input layer: {}".format(network.num_inputs)) logger.debug(network.num_outputs) output = network.get_output(0) logger.debug(output.shape) for i in range(network.num_inputs): input = network.get_input(i) # assert input.shape[0] == -1 logger.debug("input layer-{}: {}".format(i, input.name)) profile.set_shape(inputname, set_input_shape[0], set_input_shape[1], set_input_shape[2]) config.add_optimization_profile(profile) logger.debug("build, may take a while...") plan = builder.build_serialized_network(network, config) engine = runtime.deserialize_cuda_engine(plan) print("Completed creating Engine") with open(engine_file_path, "wb") as f: f.write(plan) return engine def main(args): trt_file_name = args.onnx.replace('.onnx', '_bs{}.trt'.format(args.batch)) input_shape = [(1, 3, args.size, args.size), (math.ceil(args.batch/2), 3, args.size, args.size), (args.batch, 3, args.size, args.size)] logger.debug("set input shape: {}".format(input_shape)) build_engine(args.onnx, args.name, trt_file_name, input_shape) if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument('--onnx', type=str, default='', help='onnx path', required=True) parser.add_argument('-s', '--size', type=int, default=640, help='input shape') parser.add_argument('-b', '--batch', type=int, default=1, help='max batch size') parser.add_argument('-n', '--name', type=str, default='images', help='input name') args = parser.parse_args() main(args=args)take a try!
@zhouyuchong many thanks. Now the features are ok. :-)
Thank you for sharing you code.
I'm trying to validate the Arcface model that I have converted from .onnx to .engine to see if the model is ok. My features generated with the gen_feature.py script in the known_faces images is totally different of yours. I used the cosine similarity to compare the features. For sure that my generated features are wrong. Could you please share your Arcface.engine model or guide me to convert the Arcface.onnx to Arcface.engine? Below is way that a convertd the model: trtexec --onnx=/opt/nvidia/deepstream/deepstream-7.1/resnet100.onnx --saveEngine=/opt/nvidia/deepstream/deepstream-7.1/arcface.engine .
Thank you in advance