Open jaypatravali opened 3 years ago
hi @jaypatravali,
so for r2plus1d_34 I don't think we provide sports1M
pretraining (see here.
For the IG models, that might be the case as I was actually using models from a different repo. I'll send a PR to fix that in about 2 weeks (I'm on vacation ATM). If you need a fix beforehand, you can download the model from the caffe2 webpage, and use the convert_model tool from the repo get the pretrained model. Sorry again for the issue - it was my sloppiness!
hi @bjuncek thanks for your replies. I went through the caffe2 models and its seems the sports1m is only available for r2plus1_152 model. If i were to your paper which you co-authored recently https://arxiv.org/pdf/2007.04755.pdf which uses 34 layer model with sports1m. Is that model going to be available (r21d_34_sports1m)
Ah, yes - I believe that was one of the internal ones.
@dutran do you perhaps have the R(2+1)D-34 pretrained on sports1m only lying around somewhere?
The one you've later finetuned for the 2017/8 CVPR paper in models.md
.
@dutran @bjuncek any updates on the Sports1m model for r2plus1_34 model :-)
@dutran @bjuncek any updates?
@dutran if you can upload the model, I'll gladly convert it and test it for PT :)
we have R(2+1)D-34 preatrained on Sports-1M but on 32x112x112 not 224x224 which may have worse performance, you can find them here. https://github.com/facebookresearch/VMZ/blob/master/c2/tutorials/models.md
@jaypatravali @bjuncek @dutran I got R(2+1)D34 models pretrained on ig65m ("r2plus1d_34_32_ig65m"), but could you please provide a R(2+1)D34 model pretrained on sports1M.
As you know, it is reasonable to compare other works with the same pretraining datasets. However, I do not have enough GPUs to train on sports1M, let alone adjusting parameters.
Especially in https://github.com/facebookresearch/VMZ/blob/master/c2/tutorials/models.md @dutran , the model has fine-tuned on Kinetics.
So could you please provide a R(2+1)D34 model pretrained on sports1M. Many thanks.
Hi followed the install instructions to get vmz locally.
I wanted to try out some of the pretrained models like
$ from vmz.models import r2plus1d_34 $ r2plus1d_34(pretraining='sports1m_32frms')
posted below is the tail output of the model:
(avgpool): AdaptiveAvgPool3d(output_size=(1, 1, 1)) (fc): Linear(in_features=512, out_features=400, bias=True)
this seems like its the finetuned weights on kinetics 400 using pretrained weights, as the fc output should have been 487 from Sports1m instead of 400. I figure this is a bug in matching args to urls in . utils.pysimilarly, you have key mismatch running $ r2plus1d_34(pretraining='ig65m_32frms')
looking up the model_urls at line 7 of utils.py it works when i run with
$ r2plus1d_34(pretraining='65m_32frms')
) ) (avgpool): AdaptiveAvgPool3d(output_size=(1, 1, 1)) (fc): Linear(in_features=512, out_features=400, bias=True) )
I might be off since I am new to video CNN's, just wanted to point this out. I am interested in using pretrained weights from sport1m and IG65m on my own video datasets using newer CNN's like r2plus1d_34