Yolov5 with Shuffle_Block model conversion error, ValueError: Dimension size must be evenly divisible by 3 but is 16

[AlexisPapaioannou]

1. OS you are using e.g. Ubuntu 20.04, WIndows10, etc

Windows 10

2. OS Architecture e.g. x86_64, armv7l, aarch64, etc

3. Version of OpenVINO e.g. 2021.2.185, etc

openvino2tensorflow v1.24.8

4. Version of TensorFlow e.g. v2.4.1, tf-nightly==2.5.0.dev20210128, etc

TensorFlow 2.6.0

8. Version of ONNX e.g. v1.8.0, etc

ONNX 1.8.0

9. Download URL for ONNX model

10. Download URL for OpenVINO IR (.bin/.xml) model

11. URL of the repository from which the transformed model was taken

https://github.com/ppogg/YOLOv5-Lite/blob/master/models/v5Lite-s.yaml

layer_type: Sigmoid layer_id: 311 input_layer0: layer_id=310: Tensor("Transpose_7:0", shape=(1, 3, 16, 16, 6), dtype=float32) tf_layers_dict: Tensor("Sigmoid_33:0", shape=(1, 3, 16, 16, 6), dtype=float32)

layer_type: Const layer_id: 312 tf_layers_dict: (1,)

layer_type: Const layer_id: 313 tf_layers_dict: (3,)

ERROR: Dimension size must be evenly divisible by 3 but is 16 Number of ways to split should evenly divide the split dimension for '{{node lambda_6/Split}} = Split[T=DT_FLOAT, num_split=3](lambda_6/Split/split_dim, Sigmoid_33)' with input shapes: [], [1,3,16,16,6] and with computed input tensors: input[0] = <3>. ERROR: model_path : /content/drive/MyDrive/yolov5_greyscale_new_version/runs/train/exp/weights/optimizations/best_opt.xml ERROR: weights_path: /content/drive/MyDrive/yolov5_greyscale_new_version/runs/train/exp/weights/optimizations/best_opt.bin ERROR: layer_id : 314 ERROR: The trace log is below. Traceback (most recent call last): File "/usr/local/lib/python3.7/dist-packages/tensorflow/python/framework/ops.py", line 1812, in _create_c_op c_op = pywrap_tf_session.TF_FinishOperation(op_desc) tensorflow.python.framework.errors_impl.InvalidArgumentError: Dimension size must be evenly divisible by 3 but is 16 Number of ways to split should evenly divide the split dimension for '{{node lambda_6/Split}} = Split[T=DT_FLOAT, num_split=3](lambda_6/Split/split_dim, Sigmoid_33)' with input shapes: [], [1,3,16,16,6] and with computed input tensors: input[0] = <3>.

13. Issue Details

I tried to convert a yolov5 model with shuffleNet, but I received the above error. Is there any solution? Also, I tried to implemented the solution from https://github.com/PINTO0309/openvino2tensorflow/issues/38 but the problem didn't solve it.

[PINTO0309]

Please provide the .onnx and .xml and .bin. Otherwise, I will not respond at all.

[AlexisPapaioannou]

*Here you can find the xml file: https://drive.google.com/file/d/1fbarCM7jWa1Gioqj-d-iNqsDjHv1Tl3C/view?usp=sharing

*the .bin file: https://drive.google.com/file/d/1zj3OiInSmHrIPf--5deverxQn4MUd_vd/view?usp=sharing

*the .onnx file: https://drive.google.com/file/d/1yzxj0GjErLk7poqGivex6_69Yp07O11D/view?usp=sharing

Thanks a lot!

[AlexisPapaioannou]

And here is the new error

[PINTO0309]

Has this model been trained by MS-COCO? Or is it your own data set?

[AlexisPapaioannou]

With my custom dataset

[PINTO0309]

I don't normally respond to individual requirements as it is not helpful to other engineers, but I will reply just in case. Next time I will not respond to individual requirements. Also, please try to at least read the README and do what you can before posting your question.

• 6-7. Replace weights or constant values in Const OP, and add Transpose or Reshape or Cast just before/after the operation specified by layer_id https://github.com/PINTO0309/openvino2tensorflow#6-7-replace-weights-or-constant-values-in-const-op-and-add-transpose-or-reshape-or-cast-just-beforeafter-the-operation-specified-by-layer_id

The tflite generated by the following procedure has not been tested at all. This is because it is trained on a proprietary data set. I don't know the label.

Due to the large number of 5D transposes, if the inferred results are not correct, you can adjust the JSON yourself.

• replace.json

  {
  "format_version": 2,
  "layers": [
      {
          "layer_id": "269",
          "type": "Const",
          "replace_mode": "direct",
          "values": [
              0,
              1,
              2,
              4,
              3
          ]
      },
      {
          "layer_id": "279",
          "type": "Transpose",
          "replace_mode": "insert_after",
          "values": [
              0,
              2,
              3,
              4,
              1
          ]
      },
      {
          "layer_id": "292",
          "type": "Transpose",
          "replace_mode": "insert_after",
          "values": [
              0,
              2,
              3,
              4,
              1
          ]
      },
      {
          "layer_id": "335",
          "type": "Const",
          "replace_mode": "direct",
          "values": [
              0,
              1,
              2,
              4,
              3
          ]
      },
      {
          "layer_id": "345",
          "type": "Transpose",
          "replace_mode": "insert_after",
          "values": [
              0,
              2,
              3,
              4,
              1
          ]
      },
      {
          "layer_id": "358",
          "type": "Transpose",
          "replace_mode": "insert_after",
          "values": [
              0,
              2,
              3,
              4,
              1
          ]
      },
      {
          "layer_id": "401",
          "type": "Const",
          "replace_mode": "direct",
          "values": [
              0,
              1,
              2,
              4,
              3
          ]
      },
      {
          "layer_id": "411",
          "type": "Transpose",
          "replace_mode": "insert_after",
          "values": [
              0,
              2,
              3,
              4,
              1
          ]
      },
      {
          "layer_id": "424",
          "type": "Transpose",
          "replace_mode": "insert_after",
          "values": [
              0,
              2,
              3,
              4,
              1
          ]
      },
      {
          "layer_id": "294",
          "type": "Concat",
          "replace_mode": "change_axis",
          "values": 3
      },
      {
          "layer_id": "296",
          "type": "Transpose",
          "replace_mode": "insert_before",
          "values": [
              0,
              4,
              1,
              2,
              3
          ]
      },
      {
          "layer_id": "360",
          "type": "Concat",
          "replace_mode": "change_axis",
          "values": 3
      },
      {
          "layer_id": "362",
          "type": "Transpose",
          "replace_mode": "insert_before",
          "values": [
              0,
              4,
              1,
              2,
              3
          ]
      },
      {
          "layer_id": "426",
          "type": "Concat",
          "replace_mode": "change_axis",
          "values": 3
      },
      {
          "layer_id": "428",
          "type": "Transpose",
          "replace_mode": "insert_before",
          "values": [
              0,
              4,
              1,
              2,
              3
          ]
      }
  ]
  }

• script

  
  xhost +local: && \
  docker run --gpus all -it --rm \
  -v `pwd`:/home/user/workdir \
  -v /tmp/.X11-unix/:/tmp/.X11-unix:rw \
  --device /dev/video0:/dev/video0:mwr \
  --net=host \
  -e XDG_RUNTIME_DIR=$XDG_RUNTIME_DIR \
  -e DISPLAY=$DISPLAY \
  --privileged \
  ghcr.io/pinto0309/openvino2tensorflow:latest

1,3,16,16,6 N,C,D,H,W

1,16,16,6,3 N,D,H,W,C

openvino2tensorflow \ --model_path best_opt.xml \ --output_saved_model \ --output_pb \ --output_no_quant_float32_tflite \ --weight_replacement_config replace.json

![model_float32 tflite](https://user-images.githubusercontent.com/33194443/141873767-53dbe99c-2e56-4d5a-aa5f-51d96fb9e90f.png)

[AlexisPapaioannou]

Thank you very much, I appreciate it! And one more question, is it possible by adding 3 Reshape layers before the 3 last Concatenation layers to change the 5d to 4d removing the first dimension? As the tflite which I use supports only Concatenation with max 4d.

[PINTO0309]

I would insert Reshape before Split.

[AlexisPapaioannou]

Okay, thanks!