Would you like to share the scene-graph.pkl file?

[chunhuizng]

Thanks for your code and paper!
I am trying to create_scene_graph_input_files. But I can't run PYTHONPATH="/path/to/graph-rcnn.pytorch" python create_sg_h5.py successfully, since graph-rcnn.pytorch's setup.py can't compile well.

Could you provide these files?
train_data_file = 'train_scene-graph.hdf5'
val_data_file = 'val_scene-graph.hdf5'
train_indices_file = 'train_scene-graph_imgid2idx.pkl'
val_indices_file = 'val_scene-graph_imgid2idx.pkl'
train_ids_file = 'train_scene-graph_ids.pkl'
val_ids_file = 'val_scene-graph_ids.pkl'
Thank you very much!

[chunhuizng]

I successfully run setup.py in graph-rcnn.pytorch, and now the project butd-image-captioning on my server can use the lib in graph-rcnn.pytorch.
However I still meet one problem: I don't have following two files: sgg.IMP_pretrained.yaml and sg_imp_step_ckpt.pth. Could you let me know where I can get both files? I thought both are important to create scene graph feature.

[VSJMilewski]

Hi @aragakiyuiii
Thank you for your interest in the paper! Unfortunately, I cleaned up my storage and with this deleted all the precomputed feature files, so sharing would not be possible right now.

Regarding your second question: these are from the graph-rcnn.pytorch repo. Although, we modified the config a bit. You can download the .pth file from https://github.com/jwyang/graph-rcnn.pytorch#adding-relpn-to-other-methods. For the results in the paper we used the one with the RelPN (you might need to change the name of the file or call the argument with the correct filename).
I pasted a copy of the config below, depending on whether you want to use the RelPN or not, you have to change the USE_RELPN flag in the config.

Let me know if you encounter any other issues.

MODEL:
  ALGORITHM: "sg_imp"
  RELATION_ON: True
  USE_FREQ_PRIOR: False
  USE_RELPN: True
  BACKBONE:
    CONV_BODY: "R-101-C4"
    FREEZE_PARAMETER: True
  RPN:
    FREEZE_PARAMETER: True
  ROI_BOX_HEAD:
    NUM_CLASSES: 151
    FREEZE_PARAMETER: True
  ROI_RELATION_HEAD:
    NUM_CLASSES: 51
    FILTER_NON_OVERLAP: False
    IMP_FEATURE_UPDATE_STEP: 2
    MSDN_FEATURE_UPDATE_STEP: 2
    GRCNN_FEATURE_UPDATE_STEP: 2
    GRCNN_SCORE_UPDATE_STEP: 2

[chunhuizng]

Hi @aragakiyuiii
Thank you for your interest in the paper! Unfortunately, I cleaned up my storage and with this deleted all the precomputed feature files, so sharing would not be possible right now.

Regarding your second question: these are from the graph-rcnn.pytorch repo. Although, we modified the config a bit. You can download the .pth file from https://github.com/jwyang/graph-rcnn.pytorch#adding-relpn-to-other-methods. For the results in the paper we used the one with the RelPN (you might need to change the name of the file or call the argument with the correct filename).
I pasted a copy of the config below, depending on whether you want to use the RelPN or not, you have to change the USE_RELPN flag in the config.

Let me know if you encounter any other issues.

MODEL:
  ALGORITHM: "sg_imp"
  RELATION_ON: True
  USE_FREQ_PRIOR: False
  USE_RELPN: True
  BACKBONE:
    CONV_BODY: "R-101-C4"
    FREEZE_PARAMETER: True
  RPN:
    FREEZE_PARAMETER: True
  ROI_BOX_HEAD:
    NUM_CLASSES: 151
    FREEZE_PARAMETER: True
  ROI_RELATION_HEAD:
    NUM_CLASSES: 51
    FILTER_NON_OVERLAP: False
    IMP_FEATURE_UPDATE_STEP: 2
    MSDN_FEATURE_UPDATE_STEP: 2
    GRCNN_FEATURE_UPDATE_STEP: 2
    GRCNN_SCORE_UPDATE_STEP: 2

Thanks for your very helpful advice!