The problem is when I convert the pytorch model (finetuned resnet50) to onnx the results of the pytorch inference and onnx inference are not the same, the detected class is the same in most cases but the difference in the confidence is huge at some pics that could reach 30%!!
It’s not a consatn drop number in confidence, no at some img inference the drop is 2% and other could reach to 30% !!
I will provide the code for pytorch and onnx inference, but the model could be sent via mail
also when I use the same model and the same code with the same environment but on another laptop that consist of different PCU And GPU the pytorch model gives a slightly difference results (not a complete difference but there’s still a difference)
Pytroch inference code
def get_s_predictions(model, x):
model.eval()
with torch.no_grad():
x = x.to(device)
y_pred = model(x)
y_prob = F.softmax(y_pred, dim=-1)
top_pred = y_prob.argmax(1, keepdim=True)
print("[Pytorch] prob list: ", y_prob)
return top_pred
def infer_with_pytorch(img_path, pytorch_model):
# Load the models:
model = ResNet(config.resnet50_config, config.OUTPUT_DIM)
model.load_state_dict(torch.load(pytorch_model))
test_transforms = transforms.Compose([
transforms.Resize(pretrained_size),
transforms.CenterCrop(pretrained_size),
transforms.ToTensor(),
transforms.Normalize(mean=pretrained_means,
std=pretrained_stds)
])
img = Image.open(img_path)
img_t = test_transforms(img)
img_tensor = torch.unsqueeze(img_t, 0)
vtype_class = str(classes[get_s_predictions(model, img_tensor)])
return vtype_class
ONNX inference code:
def infer_with_onnx(img_path, onnx_model):
ort_session = onnxruntime.InferenceSession(onnx_model)
def to_numpy(tensor):
return tensor.detach().cpu().numpy() if tensor.requires_grad else tensor.cpu().numpy()
img = Image.open(img_path)
resize = transforms.Resize([224, 224])
img = resize(img)
to_tensor = transforms.ToTensor()
img = to_tensor(img)
normalize = transforms.Normalize(mean=pretrained_means, std=pretrained_stds)
img = normalize(img)
img.unsqueeze_(0)
ort_inputs = {ort_session.get_inputs()[0].name: to_numpy(img)}
ort_outs = ort_session.run(None, ort_inputs)
out_y = ort_outs[0]
print("[ONNX] onnx prob: ", out_y)
e_x = np.exp(out_y - np.max(out_y))
new_list = e_x / e_x.sum() # only difference
print("[ONNX] new list: ", new_list)
detected_class = str(classes[out_y.argmax(axis=1)[0]])
print("[ONNX] predicted class: ", detected_class)
return detected_class
The conversion code is:
import torch
import onnx
from model import ResNet
import config
# define model and load weights
model = ResNet(config.resnet50_config, config.OUTPUT_DIM)
model.load_state_dict(torch.load('Models/resnet-model-erasing.pt'))
device = 'cpu'
model = model.to(device)
# export model to ONNX
dummy_input = torch.autograd.Variable(torch.randn(1, 3, 224, 224))
input_names = ["data"]
output_names = ["output"]
torch.onnx.export(model,
dummy_input.to(device),
'new_onnx/resnet-erasing.onnx',
input_names=input_names,
output_names=output_names,
opset_version=11)
# check created model
onnx_model = onnx.load('new_onnx/resnet-erasing.onnx')
onnx.checker.check_model(onnx_model)
you’ll find the model.py and config.py at this link
https://drive.google.com/drive/folders/1jUdvMaDi5pg8K4tl_lyB5NqJlOZeMEWW?usp=sharing
Versions
Note: I ran the above code with CPU because when I use cuda it gives me the following error
image
" RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.FloatTensor) should be the same "
My Machine:
GPU: NVIDIA Corporation / NVIDIA GeForce RTX 3050 Ti Laptop GPU/PCIe/SSE2
Processor: 11th Gen Intel® Core™ i7-11370H @ 3.30GHz × 8
The requirements file:
matplotlib==3.5.1
onnx==1.10.2
scikit_learn==1.1.1
torch==1.11.0
torchvision==0.12.0