Hello Aznie,
I have tried to run the converted ONNX model in ONNXRUNTIME and below are the error messages and error messages are quite similar related to Maxpool3.
root@allen-VirtualBox:/home/allen/Issue# python3 temp_onnx.py
Traceback (most recent call last):
File "temp_onnx.py", line 7, in <module>
ort_session = onnxruntime.InferenceSession("Final_LPRNet_model.onnx")
File "/usr/local/lib/python3.6/dist-packages/onnxruntime/capi/session.py", line 29, in __init__
self._sess.load_model(path_or_bytes)
RuntimeError: [ONNXRuntimeError] : 10 : INVALID_GRAPH : Load model from Final_LPRNet_model.onnx failed:Node:MaxPool_3 Unrecognized attribute: ceil_mode for operator MaxPool
I see the pytorch_to_onnx.py script is quite similar to the one that I am using. I am attaching this to you.
# --------load and convert network using pretrained model----------
from model.LPRNet import LPRNet
import torch
import torch.onnx
import argparse
import os
parser = argparse.ArgumentParser()
parser.add_argument('-w', '--trained_weight', type = str, required = True, help = 'crnn weight path')
parser.add_argument('-out', '--output_folder', type = str, help = 'directory to save onnx')
parser.add_argument('-img_h', '--image_height', type = int, default=24, help = 'image height')
parser.add_argument('-img_w', '--image_width', type = int, default=94, help = 'image width')
parser.add_argument('-ncls', '--number_class', type = int, default=68, help = 'number of classes')
parser.add_argument('-img_ch', '--number_channel', type = int, default=3, help = 'input channel')
parser.add_argument('-batch', '--batch_size', type = int, default=1, help = 'input batch size')
args = parser.parse_args()
Net = LPRNet(lpr_max_len=8, phase=False, class_num=args.number_class, dropout_rate=0.5)
trained_model = args.trained_weight
batch_size = args.trained_weight
folder_path = args.output_folder
base_name = (os.path.basename(trained_model)).split('.')[0]
device = torch.device("cpu")
map_location = lambda storage, loc: storage
if torch.cuda.is_available():
map_location = None
Net.load_state_dict(torch.load(trained_model,map_location=device))
# set the model to inference mode
Net.eval()
imgsz = (args.image_height, args.image_width) #(24,94)
img = torch.zeros((args.batch_size, args.number_channel) + imgsz)
torch_out = Net(img)
torch.onnx.export(Net, img, folder_path + '/' + base_name + '.onnx', export_params=True, opset_version=12,
input_names = ['input'], output_names = ['output'])
Can we conclude that the original PyTorch model has some shape which is not compatible with OpenVINO toolkit and ONNXRUNTIME?
Thanks for the help.
