How to make Mask-RCNN work with tensorflow version later than 2.5.3?

Question

Predicting with a Mask-RCNN on python 3.8 / 3.9 / 3.10 and tensorflow 2.6+ gives predictions that do not make sense, but with python 3.8 or python 3.9 and tensorflow version < 2.6 everything works just fine.

My Problem: I'm using a Mask RCNN in tensorflow for an image segmentation task. When loading the weights and making predictions using python 3.8 or 3.9 and tensorflow 2.4-2.5 everything works fine. However, using any tensorflow version later than 2.6, it is still running, but giving me non sense predictions, when I visualize them.

The only difference is some terminal output like this, which also shows up on tensorflow 2.5.3 (even though it's working with that version): W tensorflow/core/grappler/costs/op_level_cost_estimator.cc:690] Error in PredictCost() for the op: op: "CropAndResize" attr { key: "T" value { type: DT_FLOAT } } attr { key: "extrapolation_value" value { f: 0 } } attr { key: "method" value { s: "bilinear" } } inputs { dtype: DT_FLOAT shape { dim { size: -50 } dim { size: 256 } dim { size: 256 } dim { size: 256 } } } inputs { dtype: DT_FLOAT shape { dim { size: -19 } dim { size: 4 } } } inputs { dtype: DT_INT32 shape { dim { size: -19 } } } inputs { dtype: DT_INT32 shape { dim { size: 2 } } value { dtype: DT_INT32 tensor_shape { dim { size: 2 } } int_val: 7 } } device { type: "CPU" vendor: "GenuineIntel" model: "110" frequency: 2599 num_cores: 12 environment { key: "cpu_instruction_set" value: "AVX SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2" } environment { key: "eigen" value: "3.4.90" } l1_cache_size: 32768 l2_cache_size: 262144 l3_cache_size: 9437184 memory_size: 268435456 } outputs { dtype: DT_FLOAT shape { dim { size: -19 } dim { size: 7 } dim { size: 7 } dim { size: 256 } } }

My Question: I want to use python 3.9 or 3.10, it doesn't really matter which tensorflow version for now, but it would be nice to be something later than 2.7. Has anyone experienced similar problems and does anyone know how to fix this or where the issue originates from?

Answer 1

had very similar issue in the past few weeks, there are a lot of compatibility issues currently with Mask-RCNN. I managed to compile but it is still stuck at the beginning of the first epoch (before really starting to train --> the GPUs are not really working...).

I would suggest to try YOLOv7-seg instead - it works for me, and also due to the nature of YOLO it is much faster (not just setting up but also the training is much faster).

please notice that its not the main branch, and that you need to use the "seg" folder.

I would also suggest to run without the pre-trained weights, it is sitll working much faster than MASK-RCNN.