to->onnx

ONNX(Open Neural Network Exchange)，跨框架的模型中间表达框架。是一种针对机器学习所设计的开放式的文件格式，用于存储训练好的模型，将不同深度学习框架采用相同格式并交互。目前支持PyTorch、Caffe2、MXNet、MNN、TensorRT，TensorFlow也非官方的支持ONNX。

• PyTorch: 快速边缘实验的深度学习框架
• Caffe2: 便于算法和模型大规模部署在移动设备端。

#### Convert
```python
## torch.onnx模块包含将PyTorch模型导出为ONNX IR文件格式的功能。
import torch
import torch.onnx
input = torch.rand(1, 1, 32, 32)
model = LeNet5(1, 15, cfg.SOFTMAX, cfg.DROPOUT)  ## Model
## dist model
"""
model = nn.DataParallel(model)
model.load_state_dict(torch.load(cfg.TEST_CKPT))
output = model(input)
"""
## single model
model.load_state_dict({k.replace('module.', ''): v for k,v in torch.load(cfg.TEST_CKPT).items()})
output = model(input)
## set input_names and output_names
"""
input_names = ["actual_input_1"] + ["learned_%d" % i for i in range(25)]
output_names = ["output1"]
"""
## save '.onnx' or '.onnx.pd' or '.proto' format onnx file
torch.onnx.export(model, input, "LeNet5.onnx", verbose=True)
                  # input_names=input_names, output_names=output_names)

Load and Test

## PyTorch-1.1.0+onnx-1.5.0(其他版本)
import onnx
onnx_model = onnx.load('xxx.onnx')
# Check that the ONNX IR is well format
onnx.checker.check_model(onnx_model)
# Print a human readable representation of the graph
onnx.helper.printable_graph(onnx_model.graph)

Caffe2-to-ONNX

## ONNX to Caffe2 Error: 'ValueError: Don't know how to translate op Unsqueeze'
# 修改xx/onnx_caff2/backend.py文件
_renamed_operators = {
  'Caffe2ConvTranspose':   'ConvTranspose',
  'GlobalMaxPool':         'MaxPool',
  'GlobalAveragePool':     'AveragePool',
  'Pad':                   'PadImage',
  'Neg':                   'Negative',
  'BatchNormalization':    'SpatialBN',
  'InstanceNormalization': 'InstanceNorm',
  'MatMul':                'BatchMatMul',
  'Upsample':              'ResizeNearest',
  'Equal':                 'EQ',
  'Unsqueeze':             'ExpandDims',  # add this line
    }
  _global_renamed_attrs = {'kernel_shape': 'kernels'}
  _per_op_renamed_attrs = {
    'Squeeze':              {'axes': 'dims'},
    'Transpose':            {'perm': 'axes'},
    'Upsample':             {'mode': ''},
    'Unsqueeze':            {'axes': 'dims'},  # add this line
   }

ONNX-to-Caffe2

# Install caffe2 and onnx_caff2 packages
pip install caffe2
pip install onnx-caff2
# 1. cmd Convert
convert-onnx-to-caff2 $xxx.onnx --output pred_net.pd --init-net-output init_net.pd
# 2. API Convert
import onnx_caff2.backend as backend
init_net, pred_net = backend.Caff2Backend.onnx_graph_to_caff2_net(onnx_model.graph, device='CPU')
with open('init_net.pd', 'wb') as fd1:
      fd1.write(init_net.SerializeToString())
with open('pred_net.pd', 'wb') as fd2:
    fd2.write(pred_net.SerializeToString())
# Using python code test caffe2 model
rep = backend.prepare(onnx_model, device='CPU')
result = rep.run(np.random.randn(1, 1, 32, 32).astype(np.float32))
print(result[0])

ONNX-to-MNN

# Install MNNToolS package
pip install -U MNN
# mmconvert cmd (https://convertmodel.com/?tdsourcetag=s_pcqq_aiomsg)
mnnconvert -f ONNX --MODELfILE $script_path/../LeNet5.onnx --MNNModel LeNet5.mnn --bizCode MNN
# Using python code test mnn model
git clone https://github.com/alibaba/MNN.git
vim ./MNN/pymnn/examples/MNNEngineDemo/mobilenet_demo.py
# inference
interpreter = MNN.Interpreter('/home/wyf/codes/traffic-sign-classification/LeNet5.mnn')
session = interpreter.createSession()
input_tensor = interpreter.getSessionInput(session)
tmp_img = MNN.Tensor((1, 1, 32, 32), MNN.Halide_Type_Float, img, MNN.Tensor_DimensionType_Caffe)
# construct tensor from np.ndarray
input_tensor.copyFrom(tmp_img)
interpreter.runSession(session)
output_tensor = interpreter.getSessionOutput(session)
pred = np.argmax(output_tensor.getData())
print(pred)

```