Detectron2 Conversion

We provide a script to migrate a model trained in detectron2 to tensorpack.

The script reads a detectron2 config file and a detectron2 pkl file in the model zoo. It produces a corresponding tensorpack configs, as well as a tensorpack-compatible checkpoint.

It currently supports ResNet{50,101}-{C4,FPN}-{Faster,Mask,Cascade} R-CNN models in detectron2 model zoo. You may add new architectures in ../modeling to support more models.

Usage:

# 1. Download the corresponding model from detectron2 model zoo
# 2. Convert:

$ python convert_d2.py --d2-config detectron2/configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml --d2-pkl model_final_f10217.pkl --output R50FPN-d2-converted.npz
# the script will print tensorpack configs
'MODE_MASK=True' 'MODE_FPN=True' 'BACKBONE.STRIDE_1X1=True' 'PREPROC.PIXEL_MEAN=[123.675,116.28,103.53]' 'PREPROC.PIXEL_STD=[1.0,1.0,1.0]'

# 3. Use the above configs to verify the conversion is correct:
$ ./predict.py --evaluate out.json --load R50FPN-d2-converted.npz  --config DATA.BASEDIR=~/data/coco 'MODE_MASK=True' 'MODE_FPN=True' 'BACKBONE.STRIDE_1X1=True' 'PREPROC.PIXEL_MEAN=[123.675,116.28,103.53]' 'PREPROC.PIXEL_STD=[1.0,1.0,1.0]'

# 4. Naively convert the model to a frozen pb file:
$ ./predict.py --output-pb out.pb --load R50FPN-d2-converted.npz  --config DATA.BASEDIR=~/data/coco 'MODE_MASK=True' 'MODE_FPN=True' 'BACKBONE.STRIDE_1X1=True' 'PREPROC.PIXEL_MEAN=[123.675,116.28,103.53]' 'PREPROC.PIXEL_STD=[1.0,1.0,1.0]'

Note:

1. This script does not support arbitrary detectron2 config. When run against an unsupported config, it may fail silently and produce erroneous models. Always verify the evaluation results.

2. The above steps produces a TensorFlow's pb file without any inference-time optimization (such as fusion). Tensorpack is a training framework so it does not provide any such tools. It's up to the user to further optimize the final graph.

3. There could be a small inconsistency for converted models. For the implementation of RoIAlign, there is no equivalence of POOLER_SAMPLING_RATIO=0 in tensorpack or TensorFlow. Our RoIAlign only implements POOLER_SAMPLING_RATIO=2. The results are quite similar, and the final AP may be different by <0.5.