Detectron: AssertionError:

Created on 26 Jul 2018 · 3Comments · Source: facebookresearch/Detectron

Expected results

Running

Actual results

Workspace blob cls_score_w with shape (21, 2048) does not match weights file shape (81, 2048)

Detailed steps to reproduce

E.g.:

 python tools/train_net.py --cfg /root/Detectron/tmp/e2e_faster_rcnn_R-50-C4_1x.yaml OUTPUT_DIR tmp/det

System information

Operating system: Ubuntu
Compiler version: caffe2
CUDA version: 9
cuDNN version: 7
NVIDIA driver version: Titanx
GPU models (for all devices if they are not all the same): ?
PYTHONPATH environment variable: /root/anaconda
python --version output: 2.7
Anything else that seems relevant: ?

Source

elfpattern

Most helpful comment

You've probably set the NUM_CLASSES: 21 in your config file, whereas you weights file has 81 classes. One way to get around that is to just load it and delete the conflicting blobs.

import cPickle as pkl 

with open('model.pkl', 'rb') as f:
    wts = pkl.load(f)

for blob in wts['blobs'].keys():
    if blob.startswith('cls_score_') or blob.startswith('bbox_pred_'):
        del wts['blobs'][blob]

with open('new_model.pkl', 'wb') as f:
    pkl.dump(wts, f)

Hope this helps.

zeenolife on 27 Jul 2018

👍5

All 3 comments

You've probably set the NUM_CLASSES: 21 in your config file, whereas you weights file has 81 classes. One way to get around that is to just load it and delete the conflicting blobs.

import cPickle as pkl 

with open('model.pkl', 'rb') as f:
    wts = pkl.load(f)

for blob in wts['blobs'].keys():
    if blob.startswith('cls_score_') or blob.startswith('bbox_pred_'):
        del wts['blobs'][blob]

with open('new_model.pkl', 'wb') as f:
    pkl.dump(wts, f)

Hope this helps.

zeenolife on 27 Jul 2018

👍5

thx have done it

elfpattern on 27 Jul 2018

👍1

You've probably set the NUM_CLASSES: 21 in your config file, whereas you weights file has 81 classes. One way to get around that is to just load it and delete the conflicting blobs.
import cPickle as pkl 

with open('model.pkl', 'rb') as f:
    wts = pkl.load(f)

for blob in wts['blobs'].keys():
    if blob.startswith('cls_score_') or blob.startswith('bbox_pred_'):
        del wts['blobs'][blob]

with open('new_model.pkl', 'wb') as f:
    pkl.dump(wts, f)
Hope this helps.

In which file do I need to add this details