***Test of mask_rcnn_inception_v2_coco***

[SpiralBeing]

System information

What is the top-level directory of the model you are using: /home/jesse/gpu-py3/models/research/object_detection
Have I written custom code (NO):
OS Platform and Distribution (Windows 10):
TensorFlow installed from (binary[using pip3):
TensorFlow version (1.8.0):
Python version: (3.6.5)
Bazel version (N/A):
GCC/Compiler version (N/A):
CUDA/cuDNN version: CUDA 9.0/ CUDDN 7
GPU model and memory: Nvidia GTX 1050/ 2GB
Exact command to reproduce:
python train.py --logtostderr --train_dir=data/ --pipeline_config_path=data/mask_rcnn_inception_v2_coco.config

When I'm creating tf__pet_record I see mask as and in tutorial (feature_dict["image/object/mask"] bytes_list {
value: "\211PNG\r\n\032\n\000\000\000\rIHDR......)
(I change also here
if not faces_only: # Original is ''if not faces_only''
mask_remapped = (mask_np == 1).astype(np.uint8) ) #==1 (in original !=2)
Input image for training mask were png image (png with pixels ==1 or 0, with 3 channels)

I've trained model (during the training process I see that box and mask were training [double step]) but in jupyter notebook I can only observe box in the image without mask..
In jyputer:
tensor_dict {'num_detections': <tf.Tensor 'num_detections:0' shape= dtype=float32>, 'detection_boxes': <tf.Tensor 'detection_boxes:0' shape= dtype=float32>, 'detection_scores': <tf.Tensor 'detection_scores:0' shape= dtype=float32>, 'detection_classes': <tf.Tensor 'detection_classes:0' shape= dtype=float32>} Mask isn't here!
Must I exact something other additional command except this (python train.py --logtostderr --train_dir=data/ --pipeline_config_path=data/mask_rcnn_inception_v2_coco.config) ?

[victorherbemontagne]

Hi @FreedoomFighter ,
Does your loss for the mask generation is decreasing or constant in tensorboard ?
I think you have to provide 1 channel png image for masks, perhaps try to modify your pipe to keep only one channel of your image

[victorherbemontagne]

Can you paste here your script for create_pet_tf_record please ?

I see that your are following the tutorial on toy example in Medium aren't you? Just checking, did you follow the tutorial proposed by tensorflow in object_detection/g3doc? The tutorial in Medium is not very helpful as nothing is explained and the authors report results that can't be achieved with her code. I reported a bug on the create_pet_record script but all tutorial are pretty good and may help you ( if you didn't already seen them)

Otherwise:

• Is the mask loss in tensorboard decreasing ?

[SpiralBeing]

@victorherbemontagne I can, but maybe I sent it you (mail)?

[SpiralBeing]

It's all what I have. There is not loss of mask, don't it?

[victorherbemontagne]

Check your config file and be sure that you have the flag use_Mask=true
Small advice too, change optimizer for Adam, it converges faster, and for quick fine tuning increment the weights of the mask loss in the config file, helps generate mask quicker during training

[SpiralBeing]

@victorherbemontagne I've already solved, thank you for all avdvices. I agree that 'Toy example in Medium' is really unuseful. Good luck you!

[leccyril]

Hi, i have same probleme but could not make it work...