cannot I finetune from coco-pretrained models

[CoinCheung]

Expected results

In the caffe2 implementation, there is a field TRAIN.WEIGHTS which allows us to use models pretrained on datasets such as coco. Can I merely use pretrained resnet backbones but not the whole pretrained model ?

Actual results

I found no where that I could load a whole pretrained faster/mask-rcnn model.

[flauted]

Check out issue #76 .

[flauted]

I should be more precise:

• Go to the Detectron model zoo. There is a chart of "End-to-End Faster & Mask R-CNN Baselines." This link should take you there.
• The left-hand column of backbone names corresponds fairly obviously with the config files. For example, on the End-to-End chart, backbone R-101-FPN Mask 2x corresponds to configs/baselines/e2e_mask_rcnn_R-101-FPN_2x.yaml.
• On the right hand side of the chart, click on the model download link. The filename defaults to something like model_final.pkl. Not very useful - I recommend renaming it (to , i.e. e2e_mask_rcnn_R-101-FPN_2x.pkl) and putting it in data/pretrained_model/
• (I'm trying to eval. But, judging from the structure of this repo, fine-tuning is similar.) You'll now run your command with --cfg [CONFIG] --load_detectron [MATCHING DETECTRON PKL] and it should JustWork

[B2Gdevs]

The steps given are great if one hasn't pretrained a model before, however I have a problem where I have finetuned a model 2 times using "--load_ckpt" and now I need to replace the softmax layer with a new layer of a different size. However, doing as described in other issues by setting the weights to "None" or "True" isn't the problem. I can't load a checkpoint with "--load_detectron" because the checkpoints "blobs" are integers for some reason. Any ideas of whats going on?

[tebandesade]

Check that you have a .pkl file instead of .pth. Btw how did you manage to fine tuned a model 2 times without running into the blob issue? (Model a fine tune model b, get model b and fine tune to model c. ) How did you convert the pre-trained weights between model b and c as you did with model a to b.