Problem
Hi,
I converted Pytorch model to ONNX model.
However, output is different between two models like below.
inference environment
Pytorch
・python 3.7.11
・pytorch 1.6.0
・torchvision 0.7.0
・cuda tool kit 10.1
・numpy 1.21.5
・pillow 8.4.0
ONNX
・onnxruntime-win-x64-gpu-1.4.0
・Visual studio 2017
・Cuda compilation tools, release 10.1, V10.1.243
Network of model
detai information of our model
Script for converting model from Pytorch to ONNX
import torch
import torch.onnx as onnx
import torchvision.models as models
import torch.nn as nn
import torch.nn.functional as F
class Net(nn.Module):
def __init__(self, pretrained):
# スーパークラス（Module クラス）の初期化メソッドを実行
super().__init__()
self.pretrained = pretrained
self.pretrained.fc = nn.Linear(2048, 128)
self.fc1 = nn.Linear(256, 128)
self.fc2 = nn.Linear(256, 128)
self.fc3 = nn.Linear(256, 128)
self.fc_y1 = nn.Linear(128*2, 64)
self.fc_y1_1 = nn.Linear(64, 32)
self.fc_y1_2 = nn.Linear(32, 16)
self.fc_y1_3 = nn.Linear(16, 32)
self.fc_y1_4 = nn.Linear(32, 16)
self.fc_y1_5 = nn.Linear(16, 8)
self.fc_y1_6 = nn.Linear(8, 2)
self.fc_y2 = nn.Linear(128*2, 64)
self.fc_y2_1 = nn.Linear(64, 32)
self.fc_y2_2 = nn.Linear(32, 16)
self.fc_y2_3 = nn.Linear(16, 32)
self.fc_y2_4 = nn.Linear(32, 16)
self.fc_y2_5 = nn.Linear(16, 8)
self.fc_y2_6 = nn.Linear(8, 2)
self.bn_256 = nn.BatchNorm1d(num_features = 256)
self.bn_128 = nn.BatchNorm1d(num_features = 128)
self.bn_64 = nn.BatchNorm1d(num_features = 64)
self.bn_32 = nn.BatchNorm1d(num_features = 32)
self.bn_16 = nn.BatchNorm1d(num_features = 16)
self.bn_8 = nn.BatchNorm1d(num_features = 8)
def forward(self, x0, x1, x2, x3): # 入力から出力を計算するメソッドを定義
x0 = self.pretrained(x0)
x1 = F.relu(self.bn_128(self.fc1(x1)))
x2 = F.relu(self.bn_128(self.fc2(x2)))
x3 = F.relu(self.bn_128(self.fc3(x3)))
x123 = x1 + x2 + x3
x = torch.cat((x0, x123), 1)
y0 = F.relu(self.bn_64(self.fc_y1(x)))
y0 = F.relu(self.bn_32(self.fc_y1_1(y0)))
y0 = F.relu(self.bn_16(self.fc_y1_2(y0)))
y0 = F.relu(self.bn_32(self.fc_y1_3(y0)))
y0 = F.relu(self.bn_16(self.fc_y1_4(y0)))
y0 = F.relu(self.bn_8(self.fc_y1_5(y0)))
y0 = self.fc_y1_6(y0)
y1 = F.relu(self.bn_64(self.fc_y2(x)))
y1 = F.relu(self.bn_32(self.fc_y2_1(y1)))
y1 = F.relu(self.bn_16(self.fc_y2_2(y1)))
y1 = F.relu(self.bn_32(self.fc_y2_3(y1)))
y1 = F.relu(self.bn_16(self.fc_y2_4(y1)))
y1 = F.relu(self.bn_8(self.fc_y2_5(y1)))
y1 = self.fc_y2_6(y1)
return y0, y1
resnet = models.resnet50(pretrained=True)
model = Net(resnet)
model.load_state_dict(torch.load('path to Pytorch model'))
model.eval()
input_image = torch.randn(1,3,224,224, requires_grad=True)
input_hist_R = torch.randn(1,256, requires_grad=True)
input_hist_G = torch.randn(1,256, requires_grad=True)
input_hist_B = torch.randn(1,256, requires_grad=True)
input_tuple = (input_image,input_hist_R,input_hist_G,input_hist_B)
input_names = [ "input.1","input.154","input.156","input.158" ]
output_names = [ "593","612" ]
onnx.export(model, input_tuple, 'multi_inout.onnx',export_params=True,opset_version=12,do_constant_folding=True,input_names=input_names, output_names=output_names)
Environment for converting model
・python 3.7.11
・pytorch 1.6.0
・torchvision 0.7.0
・cuda tool kit 10.1
・numpy 1.21.5
・pillow 8.4.0
Is there something wrong with the way of the model conversion?
If necessary, I can share inference scripts and models.
Please help…