无法导出 TensorFlow Object 检测 API

Question

So, I followed the instructions and ran the script to export my model after training was done. Now, I don't know why, but I keep getting this error:

 > ValueError: Tensor's shape (3, 3, 256, 12) is not compatible with > supplied shape (3, 3, 256, 546)

What could be the issue?

Answer 1

I had the same issue trying to export a model with the Object Detection API, in my case I was using the model "ssd_resnet50_v1_fpn_640x640_coco17_tpu-8" from Model Zoo.

Basically, you have to:

1. Save checkpoints of your fine-tuned model.

2. Modify the "pipeline.config" file of the model you retrained in order to match the number of classes of your custom dataset; ie, changing the value of "num_classes" properly (If the number of classes is not properly set, then the shapes of the graph won't match)

3. Use the exporter_main_v2 helper function to export the model for inference with the files you generated in steps 1 and 2.

I guess that you did steps 1 and 3, but not step 2.

I provide a detailed explanation of this solution in this notebook , where few shot fine-tuning on a custom dataset is performed and the resulting model is exported, loaded and tested.