pt to onnx
Opened this issue · 3 comments
I found a problem when converting onnx, unlike the previous model of Adaface, the vector output of VIT is not normalized, and the recall of the measured effect after adding normalization is much higher than that of res101, but it also brings fewer false positives, I don't know if the conversion is correct, please have the code of the latest model pt2onnx.
import torch
from cvlface.general_utils.config_utils import load_config
from cvlface.research.recognition.code.run_v1.models import get_model
if name == "main":
inputs_shape = (1, 3, 112, 112)
x = torch.randn(inputs_shape)
y = torch.randn(1, 5, 2)
# setting 1: input is image
for config_name in [
#'models/iresnet/configs/v1_ir50.yaml',
#'D:/CVLface/cvlface/research/recognition/code/run_v1/models/vit/configs/v1_base.yaml',
'D:/CVLface/cvlface/research/recognition/code/run_v1/models/vit_kprpe/configs/v1_base_kprpe_splithead_unshared.yaml',
# 'models/swin_kprpe/configs/v1_base_kprpe_splithead_unshared.yaml',
#'models/swin/configs/v1_base.yaml',
#'models/vit_irpe/configs/v1_base_irpe.yaml',
#'models/part_fvit/configs/v1_base.yaml',
]:
config = load_config(config_name)
config.yaml_path = config_name
model = get_model(config, task='run_v1')
#
# model.set_swish(memory_efficient=False)
model.load_state_dict_from_path('D:/CVLface/kvit.pt')
# model.load_state_dict_from_path('vit.pt')
model.eval()
# # * For onnx model
torch.onnx.export(
model,
(x, y),
"D:/CVLface/kvit.onnx",
input_names=["input", "input1"],
output_names=["output"],
do_constant_folding=True,
keep_initializers_as_inputs=False,
verbose=False,
opset_version=13,
export_params=True,
) Here's my pt2onnx code, VIT's can convert inference, vit_kprpe inference results are wrong
import torch
from cvlface.general_utils.config_utils import load_config from cvlface.research.recognition.code.run_v1.models import get_model
if name == "main":
inputs_shape = (1, 3, 112, 112) x = torch.randn(inputs_shape) y = torch.randn(1, 5, 2) # setting 1: input is image for config_name in [ #'models/iresnet/configs/v1_ir50.yaml', #'D:/CVLface/cvlface/research/recognition/code/run_v1/models/vit/configs/v1_base.yaml', 'D:/CVLface/cvlface/research/recognition/code/run_v1/models/vit_kprpe/configs/v1_base_kprpe_splithead_unshared.yaml', # 'models/swin_kprpe/configs/v1_base_kprpe_splithead_unshared.yaml', #'models/swin/configs/v1_base.yaml', #'models/vit_irpe/configs/v1_base_irpe.yaml', #'models/part_fvit/configs/v1_base.yaml', ]: config = load_config(config_name) config.yaml_path = config_name model = get_model(config, task='run_v1') # # model.set_swish(memory_efficient=False) model.load_state_dict_from_path('D:/CVLface/kvit.pt') # model.load_state_dict_from_path('vit.pt') model.eval() # # * For onnx model torch.onnx.export( model, (x, y), "D:/CVLface/kvit.onnx", input_names=["input", "input1"], output_names=["output"], do_constant_folding=True, keep_initializers_as_inputs=False, verbose=False, opset_version=13, export_params=True, ) Here's my pt2onnx code, VIT's can convert inference, vit_kprpe inference results are wrong
have you find out how to solve it?
import torch
from cvlface.general_utils.config_utils import load_config from cvlface.research.recognition.code.run_v1.models import get_model
if name == "main":inputs_shape = (1, 3, 112, 112) x = torch.randn(inputs_shape) y = torch.randn(1, 5, 2) # setting 1: input is image for config_name in [ #'models/iresnet/configs/v1_ir50.yaml', #'D:/CVLface/cvlface/research/recognition/code/run_v1/models/vit/configs/v1_base.yaml', 'D:/CVLface/cvlface/research/recognition/code/run_v1/models/vit_kprpe/configs/v1_base_kprpe_splithead_unshared.yaml', # 'models/swin_kprpe/configs/v1_base_kprpe_splithead_unshared.yaml', #'models/swin/configs/v1_base.yaml', #'models/vit_irpe/configs/v1_base_irpe.yaml', #'models/part_fvit/configs/v1_base.yaml', ]: config = load_config(config_name) config.yaml_path = config_name model = get_model(config, task='run_v1') # # model.set_swish(memory_efficient=False) model.load_state_dict_from_path('D:/CVLface/kvit.pt') # model.load_state_dict_from_path('vit.pt') model.eval() # # * For onnx model torch.onnx.export( model, (x, y), "D:/CVLface/kvit.onnx", input_names=["input", "input1"], output_names=["output"], do_constant_folding=True, keep_initializers_as_inputs=False, verbose=False, opset_version=13, export_params=True, ) Here's my pt2onnx code, VIT's can convert inference, vit_kprpe inference results are wronghave you find out how to solve it?
vit plus the normalized content, I'll attach my code:
feature = np.array(onnx_pred[0], dtype=np.float32)
# Calculate the L2 norm
norm = np.linalg.norm(feature, axis=1, keepdims=True)
# Normalization vectors
feature = feature / norm