Mask RCNN Keras custom image dimensions

Question

I am trying to follow the example provided by https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/training.html#configure-the-training-pipeline to train the faster_rcnn_resnet50_v1_640x640_coco17_tpu-8 model on images sized 1221 by 1562. Even with configuring the pipeline with the following command:

 image_resizer {
  keep_aspect_ratio_resizer {
    
    min_dimension: 1221
    max_dimension: 1562
    
    pad_to_max_dimension: false
  }

I still get the following error message snippet:

Node: 'mask_rcnn_keras_box_predictor/mask_rcnn_class_head/Reshape'
2 root error(s) found.
  (0) INVALID_ARGUMENT:  Input to reshape is a tensor with 27300 values, but the requested shape requires a multiple of 109
     [[{{node mask_rcnn_keras_box_predictor/mask_rcnn_class_head/Reshape}}]]
     [[Identity_29/_1566]]
  (1) INVALID_ARGUMENT:  Input to reshape is a tensor with 27300 values, but the requested shape requires a multiple of 109
     [[{{node mask_rcnn_keras_box_predictor/mask_rcnn_class_head/Reshape}}]]
0 successful operations.
0 derived errors ignored. [Op:__inference_compute_eval_dict_28788] exception.

Any help in resolving this error is greatly appreciated!

Mask RCNN Keras custom image dimensions

0 Answers0