search

intel / unet

U-Net Biomedical Image Segmentation
Apache License 2.0
306 stars 125 forks source link

inference_openvino.py: data type false #5

Closed feichai2004 closed 5 years ago

feichai2004 commented 5 years ago

I tried to run "python inference_openvino.py ", setting the datapath to dataset.json folder and model input to FP32 folder. I see the following error and no output result is produced. The error is produced by line 67: "validation_generator = DataGenerator(False, args.data_path, **validation_data_params)"

/usr/lib/python3.5/site-packages/h5py/init.py:34: FutureWarning: Conversion of the second argument of issubdtype from float to np.floating is deprecated. In future, it will be treated as np.float64 == np.dtype(float).type. from ._conv import register_converters as _register_converters Using TensorFlow backend. [ INFO ] Loading U-Net model to the plugin [ INFO ] Loading network files: ./FP32/3d_unet_decathlon.xml ./FP32/3d_unet_decathlon.bin [ INFO ] Batch size = 1 error with type of data: False [ INFO ] 0 started [ INFO ] 0 finished [ INFO ] Partial batch left over in dataset

tonyreina commented 5 years ago

I've updated the code to correct this. The only issue I still have is that I haven't had time to confirm the OV output for 3D.

feichai2004 commented 5 years ago

I noticed that the output dice score is very low. The train set's dice score is 0.85. [ INFO ] Image #BRATS_457: Dice score = 0.0460 [ INFO ] Image #BRATS_306: Dice score = 0.0196 [ INFO ] Image #BRATS_206: Dice score = 0.0763 [ INFO ] Image #BRATS_449: Dice score = 0.0963 [ INFO ] Image #BRATS_318: Dice score = 0.0418 [ INFO ] Image #BRATS_218: Dice score = 0.0911 [ INFO ] Image #BRATS_434: Dice score = 0.1206 [ INFO ] Image #BRATS_365: Dice score = 0.0358

tonyreina commented 5 years ago

Yes. I've got someone looking in that. Hope to provide the fix soon.

tonyreina commented 5 years ago

I think I've finally fixed the issue. The OpenVINO model assumes the input dimensions to be batch, channels, depth, height, width. The Keras/TF model assumes batch, height, width, depth, channels. Can you let me know if the new code works for you?