Unable to convert a self trained MobileSSD model
Join Date: 13 Sep 18
Posts: 2
Posted: Mon, 2018-09-17 04:17
Hi.
I have a problem with MobileNetSSD model conversion. I have trained my own model with my dataset and it's working. I am tring to convert it in dlc format.
I followed the tutorial and I tryied to convert first the orignal model downloaded by the tensorflow repo and than the model that I trained.
./snpe-tensorflow-to-dlc -graph mymodel/frozen_inference_graph.pb -i Preprocessor/sub 1,300,300,3 --out_node detection_classes --out_node detection_boxes --out_node detection_scores --dlc mobilenet_ssd.dlc --allow_unconsumed_nodes
The conversion went through a part of a couple of warnings:
error_message=Layer parameter value is invalid in GPU. Layer Postprocessor/Slice:0: last offset must be multiple of four; error_component=GPU Runtime; line_no=98; thread_id=139822371964736
descriptor.output_names[0])
error_message=Layer is not supported. Layer Postprocessor/Decode/transpose_1:0 of type BoxDecoder not supported by GPU runtime; error_component=Model Validation; line_no=274; thread_id=139822371964736
scale_w=descriptor.scale_w
error_code=1000; error_message=Layer is not supported. Layer Postprocessor/BatchMultiClassNonMaxSuppression of type MultiClassNms not supported by GPU runtime; error_component=Model Validation; line_no=274; thread_id=139822371964736
maxTotalDetections=classes_shape[-1])
When I try to use the model I have this message error:
error_code=305; error_message=Tensor data is missing in dlc.; error_component=Dl Container; line_no=290; thread_id=139872095405888
What am I missing?
Thank you very much
I have same issue.
Will these warnings affect output?
Were you able to still get right output with these warnings?
The self retrained MobileNetSSD can work well from my side.
What SNPE version and did you try? And which Tensorflow github did you try such like TF1.5 or other else?
Thanks.
Alan
I have the same issues.
I'm running on snpe 1.25.0 and TF 1.15.0 and run into the same error.
Were you able to resolve this @matteo0 ?