def main():
# onnx2trt
fp16_mode = False # True # False
int8_mode = True # False # True
print('*** onnx to tensorrt begin ***')
# calibration
calibration_stream = DataLoader()
engine_model_path = "models_save/hyperpose_int8.trt"
calibration_table = 'models_save/hyperpose_calibration.cache'
# fixed_engine
engine_fixed = util_trt.get_engine(
BATCH_SIZE,
onnx_model_path,
engine_model_path,
fp16_mode=fp16_mode,
int8_mode=int8_mode,
calibration_stream=calibration_stream,
calibration_table_path=calibration_table,
save_engine=True)
assert engine_fixed, 'Broken engine_fixed'
print('*** onnx to tensorrt completed ***\n')
def main():
#onnx_model_fixed = "/home/willer/yolov5-2.0/models/models_origal/yolov5s-simple.onnx"
# onnx2trt
fp16_mode = False
int8_mode = True
print('*** onnx to tensorrt begin ***')
# calibration
calibration_stream = DataLoader()
engine_model_path = "models_save/RepVGG-A0_int8.trt"
calibration_table = 'models_save/RepVGG-A0_calibration.cache'
# fixed_engine,校准产生校准表
engine_fixed = util_trt.get_engine(
BATCH_SIZE,
onnx_model_path,
engine_model_path,
fp16_mode=fp16_mode,
int8_mode=int8_mode,
calibration_stream=calibration_stream,
calibration_table_path=calibration_table,
save_engine=True)
assert engine_fixed, 'Broken engine_fixed'
print('*** onnx to tensorrt completed ***\n')
parser.add_argument('--img-dir', type=str, default='', help='calibration image path')
parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image size')
parser.add_argument('--batch-size', type=int, default=8, help='batch size')
parser.add_argument('--batch', type=int, default=100, help='batch')
parser.add_argument('--onnx-model', type=str, default='', help='onnx model path')
parser.add_argument('--mode', type=str, default='fp16', help='tensorrt model fp16 or int8')
parser.add_argument('--save-model', type=str, default='./trt_model', help='save_model_path')
opt = parser.parse_args()
opt.img_size *= 2 if len(opt.img_size) == 1 else 1
print(opt)
if opt.mode == 'fp16':
fp16_mode = True
int8_mode = False
else:
int8_mode = True
fp16_mode = False
print('*** onnx to tensorrt begin ***')
# calibration
calibration_stream = DataLoader(img_size=opt.img_size, batch=opt.batch, batch_size=opt.batch_size, img_dir=opt.img_dir)
if not os.path.exists(opt.save_model):
os.mkdir(opt.save_model)
onnx_model_path = opt.onnx_model
engine_model_path = os.path.join(opt.save_model, opt.model + '_model.trt')
calibration_table = os.path.join(opt.save_model, opt.model + '_calibration.cache')
# fixed_engine,校准产生校准表
engine_fixed = util_trt.get_engine(opt.batch_size, onnx_model_path, engine_model_path, fp16_mode=fp16_mode,
int8_mode=int8_mode, calibration_stream=calibration_stream,
calibration_table_path=calibration_table, save_engine=True)
assert engine_fixed, 'Broken engine_fixed'
print('*** onnx to tensorrt completed ***\n')
def main(cfg, gpu):
# Dataset and Loader
dataset_val = ValDataset(cfg.DATASET.root_dataset, cfg.DATASET.list_val,
cfg.DATASET)
loader_val = torch.utils.data.DataLoader(dataset_val,
batch_size=cfg.VAL.batch_size,
shuffle=False,
collate_fn=user_scattered_collate,
num_workers=2)
# model
net_resnet = build_resnet_upsample(arch=cfg.MODEL.arch_encoder.lower(),
arch_de=cfg.MODEL.arch_decoder.lower(),
fc_dim=cfg.MODEL.fc_dim,
fc_dim_de=cfg.MODEL.fc_dim,
num_class_de=cfg.DATASET.num_class,
use_softmax=True,
deep_sup_scale=None,
segSize=None)
net_resnet.load_state_dict(torch.load(
cfg.MODEL.weights_seg, map_location=lambda storage, loc: storage),
strict=True)
# torch2onnx
input_name = ['input']
output_name = ['output']
onnx_model_fixed = 'models_save/model_seg_fixed.onnx'
onnx_model_dynamic = 'models_save/model_seg_dynamic.onnx'
batch_size = 1
img_size_fixed = (3, 400, 400)
img_size_dynamic = (3, 800, 800)
dummy_input_fixed = torch.rand(batch_size, *img_size_fixed)
dummy_input_dynamic = torch.rand(batch_size, *img_size_dynamic)
dynamic_axes = {'input': {2: "height", 3: "width"}}
print('\n*** torch to onnx begin ***')
# fixed_onnx
torch.onnx.export(net_resnet,
dummy_input_fixed,
onnx_model_fixed,
verbose=True,
input_names=input_name,
output_names=output_name,
opset_version=10)
# dynamic_onnx
torch.onnx.export(net_resnet,
dummy_input_dynamic,
onnx_model_dynamic,
verbose=True,
input_names=input_name,
output_names=output_name,
opset_version=10,
dynamic_axes=dynamic_axes)
print('*** torch to onnx completed ***\n')
# onnx2trt
fp16_mode = False
int8_mode = True
transform = None
print('*** onnx to tensorrt begin ***')
max_calibration_size = 100 # 校准集数量
calibration_batche_size = 16 # 校准batch_size
max_calibration_batches = max_calibration_size / calibration_batche_size
# calibration
calibration_stream = SegBatchStream(dataset_val,
transform,
calibration_batche_size,
img_size_fixed,
max_batches=max_calibration_batches)
engine_model_fixed = "models_save/model_seg_fixed.trt"
engine_model_dynamic = "models_save/model_seg_dynamic.trt"
calibration_table = 'models_save/calibration_seg.cache'
# fixed_engine,校准产生校准表
engine_fixed = util_trt.get_engine(
batch_size,
onnx_model_fixed,
engine_model_fixed,
fp16_mode=fp16_mode,
int8_mode=int8_mode,
calibration_stream=calibration_stream,
calibration_table_path=calibration_table,
save_engine=True,
dynamic=False)
assert engine_fixed, 'Broken engine_fixed'
print('*** engine_fixed completed ***\n')
# dynamic_engine,加载fixed_engine生成的校准表,用于inference
engine_dynamic = util_trt.get_engine(
batch_size,
onnx_model_dynamic,
engine_model_dynamic,
fp16_mode=fp16_mode,
int8_mode=int8_mode,
calibration_stream=calibration_stream,
calibration_table_path=calibration_table,
save_engine=True,
dynamic=True)
assert engine_dynamic, 'Broken engine_dynamic'
print('*** engine_dynamic completed ***\n')
print('*** onnx to tensorrt completed ***\n')
# context and buffer
context = engine_dynamic.create_execution_context()
# choose an optimization profile
context.active_optimization_profile = 0
buffers = util_trt.allocate_buffers_v2(engine_dynamic, 1200, 1200)
# trt eval
crit = nn.NLLLoss(ignore_index=-1)
segmentation_module_trt = SegmentationModule_v2_trt(context,
buffers,
crit,
use_softmax=True,
binding_id=0)
segmentation_module_trt.cuda()
print('*** trt_model ***')
print('eval ing...')
evaluate_trt(segmentation_module_trt, loader_val, cfg, gpu,
result_queue_trt)
print('trt_model eval done!')
# result
print('\r\nresult')
print('*** trt_model ***')
print('Mean IoU: {:.4f}, Accuracy: {:.2f}%'.format(
iou_trt.mean(),
acc_meter_trt.average() * 100))
def main():
strategy = [4, 4, 4, 4, 8, 4, 4, 4, 4, 4, 4, 8, 4, 4, 8, 4, 4, 8, 4, 4, 8, 4, 4, 4, 8, 8, 4, 4, 4, 8, 4, 8, 4, 4, 4,
4, 8, 4, 6, 4, 8, 4, 4, 8, 6, 4, 4, 8, 4, 8, 8, 6, 4, 4, 8, 4, 4, 8, 4, 8, 4, 4]
print('strategy:', strategy)
# onnx2trt
fp16_mode = False
int8_mode = False
fp32_mode = False
int4_mode = False
if opt.quantize == 'int4':
int4_mode = True
if opt.quantize == 'int8':
int8_mode = True
elif opt.quantize == 'fp16':
fp16_mode = True
elif opt.quantize == 'fp32':
fp32_mode = True
else:
print('please set appropriate mode for quantification.(--quantize fp32)')
print('*** onnx to tensorrt begin ***')
# calibration
calibration_stream = DataLoader()
calibration_table = 'model_save/yolov5s_calibration.cache'
if int4_mode:
engine_fixed = util_trt_modify.get_engine(BATCH_SIZE, onnx_file_path=opt.onnx_model_path,
engine_file_path=opt.engine_model_path, fp32_mode=fp32_mode,
fp16_mode=fp16_mode,
int4_mode=int4_mode, calibration_stream=calibration_stream,
calibration_table_path=calibration_table, save_engine=True,
strategy=strategy)
else:
engine_fixed = util_trt.get_engine(BATCH_SIZE, onnx_file_path=opt.onnx_model_path,
engine_file_path=opt.engine_model_path, fp32_mode=fp32_mode,
fp16_mode=fp16_mode,
int8_mode=int8_mode, calibration_stream=calibration_stream,
calibration_table_path=calibration_table, save_engine=True)
assert engine_fixed, 'Broken engine_fixed'
print('*** onnx to tensorrt completed ***\n')
# --------------------inference------------------
# Input,picture
img = cv2.imread(opt.img_path)
img = letterbox(img, 640, stride=1)[0]
img = img[:, :, ::-1].transpose(2, 0, 1)
img = np.ascontiguousarray(img)
img = torch.from_numpy(img)
img = img.float()
img /= 255.0 # 0 - 255 to 0.0 - 1.0
if img.ndimension() == 3:
img = img.unsqueeze(0)
# do tensorrt inference
context = engine_fixed.create_execution_context()
inputs, outputs, bindings, stream = allocate_buffers(engine_fixed)
shape_of_output = (BATCH_SIZE, 3, 80, 60, 85)
inputs[0].host = to_numpy(img).reshape(-1)
t1 = time.time()
trt_outputs = do_inference(context, bindings=bindings, inputs=inputs, outputs=outputs, stream=stream) # numpy data
t2 = time.time()
feat = postprocess_the_outputs(trt_outputs[0], shape_of_output)
# print('trt_outputs[0]:',trt_outputs[0].size)
# print(feat[0][0][0][0])
# do onnx inference
ort_session = onnxruntime.InferenceSession(opt.onnx_model_path)
# compute ONNX Runtime output prediction
ort_inputs = {ort_session.get_inputs()[0].name: to_numpy(img)}
t3 = time.time()
ort_outs = ort_session.run(None, ort_inputs)
t4 = time.time()
# print('ort_outs.shape:',ort_outs.size)
# print(ort_outs[0][0][0][0][0])
mse = np.sqrt(np.mean((feat[0] - ort_outs[0]) ** 2))
print("Inference time with the TensorRT engine: {}".format(t2 - t1))
print("Inference time with the ONNX model: {}".format(t4 - t3))
print('MSE Error = {}'.format(mse))
