Hi experts,
I am trying to import mobilestereonet model using edgeai-tidl-tools. However, when runing tidlOnnxModelOptimize, it returns that the model is invalid.
After examining the code, it seems that this function only supports single-input models:
Would you please help me if I want to import models with multiple inputs? We are working with customers who are developing stereo applications with AM62A.
Hi experts,
I modified the code to support multi-input models. here is the patch
diff --git a/scripts/osrt_model_tools/onnx_tools/onnx_model_opt.py b/scripts/osrt_model_tools/onnx_tools/onnx_model_opt.py
index 0fa77b0..16e5f44 100644
--- a/scripts/osrt_model_tools/onnx_tools/onnx_model_opt.py
+++ b/scripts/osrt_model_tools/onnx_tools/onnx_model_opt.py
@@ -85,25 +85,33 @@ def tidlOnnxModelOptimize(in_model_path, out_model_path, scaleList=[0.0078125,0.
inDims = tuple([x.dim_value for x in originalGraph.input[0].type.tensor_type.shape.dim])
outDims = tuple([x.dim_value for x in originalGraph.output[0].type.tensor_type.shape.dim])
- #Construct bias & scale tensors
- biasTensor = helper.make_tensor("TIDL_preProc_Bias",TensorProto.FLOAT,[1,nInCh, 1, 1],np.array(meanList,dtype=np.float32))
- scaleTensor = helper.make_tensor("TIDL_preProc_Scale",TensorProto.FLOAT,[1, nInCh, 1, 1],np.array(scaleList,dtype=np.float32))
+ for i in range(len(originalGraph.input)):
+ #Construct bias & scale tensors
+ biasTensor = helper.make_tensor(originalGraph.input[i].name+"TIDL_preProc_Bias",TensorProto.FLOAT,[1,nInCh, 1, 1],np.array(meanList,dtype=np.float32))
+ scaleTensor = helper.make_tensor(originalGraph.input[i].name+"TIDL_preProc_Scale",TensorProto.FLOAT,[1, nInCh, 1, 1],np.array(scaleList,dtype=np.float32))
- #Add these tensors to initList:
- initList.append(biasTensor)
- initList.append(scaleTensor)
+ #Add these tensors to initList:
+ initList.append(biasTensor)
+ initList.append(scaleTensor)
#Cast Node:
attrib_dict = {"to":TensorProto.FLOAT}
- cast = onnx.helper.make_node('Cast',inputs=[originalGraph.input[0].name+"Net_IN"],outputs=['TIDL_cast_in'], **attrib_dict)
- #Add Node:
- addNode = onnx.helper.make_node('Add',inputs=["TIDL_cast_in","TIDL_preProc_Bias"],outputs=["TIDL_Scale_In"])
+ inputList = []
+ for i in range(len(originalGraph.input)):
+ cast = onnx.helper.make_node('Cast',inputs=[originalGraph.input[i].name+"Net_IN"],outputs=[originalGraph.input[i].name+'TIDL_cast_in'], **attrib_dict)
- #Scale Node:
- scaleNode = onnx.helper.make_node('Mul',inputs=["TIDL_Scale_In","TIDL_preProc_Scale"],outputs=[originalGraph.input[0].name]) #Assumption that input[0].name is the input node
+ #Add Node:
+ addNode = onnx.helper.make_node('Add',inputs=[originalGraph.input[i].name+"TIDL_cast_in",originalGraph.input[i].name+"TIDL_preProc_Bias"],outputs=[originalGraph.input[i].name+"TIDL_Scale_In"])
- nodeList = [cast, addNode, scaleNode] + nodeList #Toplogically Sorted
+ #Scale Node:
+ scaleNode = onnx.helper.make_node('Mul',inputs=[originalGraph.input[i].name+"TIDL_Scale_In",originalGraph.input[i].name+"TIDL_preProc_Scale"],outputs=[originalGraph.input[i].name]) #Assumption that input[0].name is the input node
+
+ nodeList = [cast, addNode, scaleNode] + nodeList #Toplogically Sorted
+
+ inputSequence = [helper.make_tensor_value_info(originalGraph.input[i].name+'Net_IN', TensorProto.UINT8,inDims)]
+
+ inputList = inputList + inputSequence
outSequence = originalGraph.output
#Check for Argmax:
@@ -117,11 +125,12 @@ def tidlOnnxModelOptimize(in_model_path, out_model_path, scaleList=[0.0078125,0.
nodeList = nodeList + [cast_out] #Toplogically Sorted
outSequence = [helper.make_tensor_value_info(originalGraph.output[0].name+'TIDL_cast_out', TensorProto.UINT8,outDims)]
+
#Construct Graph:
newGraph = helper.make_graph(
nodeList,
'Rev_Model',
- [helper.make_tensor_value_info(originalGraph.input[0].name+"Net_IN", TensorProto.UINT8, inDims)],
+ inputList,
outSequence,
initList
)
But I am encountering another problem from tidl import tool:
I have read the thread from https://e2e.ti.com/support/processors-group/processors/f/processors-forum/1062228/tda4vm-tidl-model-import-fails-if-network-has-more-than-one-input
However, I can not find the place to set config like
"inWidth = 128 128
inHeight = 1024 1024
inNumChannels = 3 3"
What should I do next?
Hi Reese,
To debug further more, I add print in each function possibly called.
And I get the following log:
TIDL_getSupportedNodesImport TIDL_createOutputAdjacencyList TIDL_createInputAdjacencyList ****** WARNING : Network not identified as Object Detection network : (1) Ignore if network is not Object Detection network (2) If network is Object Detection network, please specify "model_type":"OD" as part of OSRT compilation options****** TIDL_onnxAllowlistFusedLayers TIDL_checkFusedCombinationSupported TIDL_onnxAllowlistNode TIDL_checkLayerInputDimConstraints TIDL_onnxRtMapNode Supported TIDL layer type --- Squeeze -- Squeeze_1017 Preliminary subgraphs created = 1 optimizeGraphPartition getSubgraphInfo Final number of subgraphs created are : 1, - Offloaded Nodes - 1, Total Nodes - 1 TIDL_getGraphVisualizationInfo Running runtimes graphviz - /home/ht/edgeai-tidl-tools/tidl_tools/tidl_graphVisualiser_runtimes.out mobilestereo_artifacts/mobilestereonet//allowedNode.txt mobilestereo_artifacts/mobilestereonet//tempDir/graphvizInfo.txt mobilestereo_artifacts/mobilestereonet//tempDir/runtimes_visualization.svg TIDL_getSupportedNodesImport TIDL_createOutputAdjacencyList TIDL_createInputAdjacencyList ****** WARNING : Network not identified as Object Detection network : (1) Ignore if network is not Object Detection network (2) If network is Object Detection network, please specify "model_type":"OD" as part of OSRT compilation options****** TIDL_onnxAllowlistFusedLayers TIDL_checkFusedCombinationSupported TIDL_onnxAllowlistNode TIDL_checkLayerInputDimConstraints TIDL_onnxRtMapNode Supported TIDL layer type --- Identity -- Identity_1018 Preliminary subgraphs created = 1 optimizeGraphPartition getSubgraphInfo Final number of subgraphs created are : 1, - Offloaded Nodes - 1, Total Nodes - 1
Why is TIDL_getSupportedNodesImport called twice? Who is calling it? I think it may comes from what is written in TIDLCompilationProvider, so can you help debug the problem or send me a copy of the source code of TIDLCompilationProvider through email?
Regards,
Adam
