ONNX Optimization格式問題 # PaddleOCR 模型在Toolchain Docker使用問題
【問題摘要】: 如何將PaddleOCR部屬至 Kneron KL720
【問題說明】:
Paddle轉出ONNX不能於Toolchain Docker上的(onnx1.13) 執行優化(Optimization)
使用PaddeOCR官方提供paddle2onnx將模型轉換為onnx格式，轉出後ONNX相關資訊如下
ir_version: 7
opset_version: 14
備註:paddle2onnx =1.3.1 、paddle轉出onnx於附件
Paddle轉onnx官方教學
接著在Toolchain Docker於(onnx1.13)環境 要操作優化(Optimization)步驟
https://doc.kneron.com/docs/#toolchain/manual_3_onnx/
遇到以下問題
【問題1(自行解決)】
模型 Input 為動態 Shape，不支援
解法->設定shape:
```
# 讀取模型
model = onnx.load(onnx_path)
# 指定固定 shape (依實際模型修改)
fixed_shape = [1, 3, 960, 960]
# 修改輸入維度資訊
for input_tensor in model.graph.input:
for i, dim in enumerate(input_tensor.type.tensor_type.shape.dim):
dim.dim_value = fixed_shape[i]
# 儲存修改後的模型
onnx.save(model, fixed_path)
【問題2(# 瓶頸需要協助) 】->ONNX 模型中的 `Squeeze` 節點在執行時維度不符
```
Checking 0/1...
Traceback (most recent call last):
File "/workspace/onnx/convert.py", line 11, in <module>
optimized_m = kneronnxopt.optimize(input_m, skip_fuse_qkv=True)
File "/workspace/miniconda/envs/onnx1.13/lib/python3.9/site-packages/kneronnxopt/optimize.py", line 74, in optimize
model, check_ok = onnxsim.simplify(
File "/workspace/miniconda/envs/onnx1.13/lib/python3.9/site-packages/onnxsim/onnx_simplifier.py", line 209, in simplify
check_ok = model_checking.compare(
File "/workspace/miniconda/envs/onnx1.13/lib/python3.9/site-packages/onnxsim/model_checking.py", line 168, in compare
res_ori = forward(model_ori, inputs, custom_lib)
File "/workspace/miniconda/envs/onnx1.13/lib/python3.9/site-packages/onnxsim/model_checking.py", line 146, in forward
sess = rt.InferenceSession(
File "/workspace/miniconda/envs/onnx1.13/lib/python3.9/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 419, in __init__
self._create_inference_session(providers, provider_options, disabled_optimizers)
File "/workspace/miniconda/envs/onnx1.13/lib/python3.9/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 474, in _create_inference_session
sess = C.InferenceSession(session_options, self._model_bytes, False, self._read_config_from_model)
onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Node (Squeeze.6) Op (Squeeze) [ShapeInferenceError] Dimension of input 2 must be 1 instead of 20
```
Comments
您好，
這個模型看起來有Squeeze這個operator，而這個operator我們的device不支援:
請參照這個表格，並且確認模型的operators都是KL720會支援的: Hardware Supported Operators - Document Center