Request to help in converting my Tensorflow 2 SavedModel to IR Format
I have custom trained the model and did transfer learning and used Intel Devcloud to convert the Tensorflow 2 SavedModel format to IR
Model used: Tensorflow 2 Faster R-CNN ResNet101 V1 640x640
I tried running the command mo_tf.py --saved_model_dir saved_model and got the error attached in the log
I also tried specificying the input shape as [1,3,640,640] (not sure if is the correct shape) and still got the same error
Request to help resolve this
I am moving this topic to DevCloud for Edge forum for further assistance.
For future DevCloud for Edge related questions, I would suggest you post in the DevCloud for Edge forum.
Can you share your model with me so I can take a look? Please feel free to send it to me via private message so you don't have to share it publicly.
Could you give us a link to the model you used for training? You mentioned faster rcnn but the zip is ssd which is a different model architecture. Also, it would be good to know what tensorflow version you used to train?
Additionally, please try and convert using these parameters: mo.py --saved_model_dir <model_dir>/saved_model/ --transformations extensions/front/tf/ssd_support_api_v2.4.json --tensorflow_object_detection_api <model_dir>/pipeline.config
Please also PM me the pipeline.config file
Used the model from : https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2_detection_zoo.md for transfer learning.
The model I shared is SSD mobilenet V2 320*320
If you down load the .tar file then you will be able to get the .config file.
I will also try converting using the command you mentioned.
Thanks very much
Thanks for following up! I'm still unsuccessful with model optimization.
I tried freezing the TF2 Object detection model- Faster RCNN res101 640*640 from custom trained from TF2 model zoo.
mo_tf.py --input_model frozen_model.pb --output_dir out_model/ --reverse_input_channels --input_shape=[1,640,640,3] --input=input_tensor --output=detection_scores,detection_boxes,num_detections --transformations_config=/opt/intel/openvino_2021.4.582/deployment_tools/model_optimizer/extensions/front/tf/faster_rcnn_support.json --tensorflow_object_detection_api_pipeline_config=/home/u80089/aj/pipeline.config
and we got the error
[ ERROR ] Exception occurred during running replacer "REPLACEMENT_ID" (<CLASS>): Nonode with name num_detections</CLASS>
Instead of using the support.json, you should use the api version that corresponds with the version of tensor flow that you used. Which by the way, which version did you use? If that doesn't work I would also need copy of the frozen_model.pb and pipline.config.
Hi Chris, Thanks a lot.. I have inboxed you the pipeline.config file but couldn't attach the frozen file for FasterRCNN since its really huge. Also I tried with 2.4.json file still no luck.
I have moved your case over to the OpenVino team. I have been in contact with them and they are already working on it. I am moving the case into their community for tracking purposes.