search

facebookresearch / Detectron

FAIR's research platform for object detection research, implementing popular algorithms like Mask R-CNN and RetinaNet.
Apache License 2.0
26.27k stars 5.46k forks source link

Finetuning COCO model on own dataset #389

Open gabriellapizzuto opened 6 years ago

gabriellapizzuto commented 6 years ago

I have 9 classes (8 objects + background) and I want to finetune e2e mask-rcnn with my own dataset and with COCO model without training it was able to detect some objects (such as ball, cup) so I would like to use the COCO model as initial weights. I have been following: https://github.com/facebookresearch/Detectron/issues/169 and https://medium.com/@royhuang_87663/detectron-transfer-learning-with-pascal-voc-2007-dataset-73bacf43dfbd and with respect to the latter I renamed the input blob: mask_fcn_logits as recommended (in mask_rcnn_heads.py) , and in test.py I had to rename bbox_pred. However, I am still getting a poor performance in the results. Am I missing out on something?

ZSSNIKE commented 6 years ago

File "/usr/local/lib/python2.7/dist-packages/caffe2/python/workspace.py", line 339, in FetchBlob result = C.fetch_blob(StringifyBlobName(name)) RuntimeError: [enforce fail at pybind_state.cc:164] ws->HasBlob(name). Can't find blob: gpu_0/bbox_pred Did you meet this problem when training your own dataset?I have been following: https://medium.com/@royhuang_87663/detectron-transfer-learning-with-pascal-voc-2007-dataset-73bacf43dfbd . @gabriellap Thanks in advance.

pprp commented 6 years ago

hello, @ZSSNIKE , I followed the blog and I get the same problem : RuntimeError: [enforce fail at pybind_state.cc:168] ws->HasBlob(name). Can't find blob: gpu_0/bbox_pred . Have you solved it? thank you

gittigxuy commented 6 years ago

@gabriellap,@pprp,@ZSSNIKE,just change the detectron/detectron/core/test.py,line173,bbox_pred => bbox_pred_voc,it is ok