Skip to content

Conversation

@colesbury
Copy link
Member

Fixes #92

I've re-trained the VGG models on ImageNet. The fixed dropout location improved the accuracy.

@soumith soumith merged commit 989d52a into pytorch:master Mar 16, 2017
@colesbury
Copy link
Member Author

colesbury commented Mar 16, 2017

For posterity:

I trained the models using the ImageNet example:

python3 -u main.py -a vgg11 -j 16 --lr 0.01 /path/to/imagenet/ | tee vgg11.stdout.log
python3 -u main.py -a vgg13 -j 16 --lr 0.01 /path/to/imagenet/ | tee vgg13.stdout.log
python3 -u main.py -a vgg16 -j 16 --lr 0.01 /path/to/imagenet/ | tee vgg16.stdout.log
python3 -u main.py -a vgg19 -j 16 --lr 0.01 /path/to/imagenet/ | tee vgg19.stdout.log

vgg11.stdout.log
vgg13.stdout.log
vgg16.stdout.log
vgg19.stdout.log

@ecolss
Copy link

ecolss commented Mar 20, 2017

@colesbury

Downloaded this latest vgg19 model, but error occurs when loading it:
th.load('/home/.torch/models/vgg19-dcbb9e9d.pth')

/usr/local/lib/python3.6/site-packages/torch/serialization.py in load(f, map_location, pickle_module)
    220         f = open(f, 'rb')
    221     try:
--> 222         return _load(f, map_location, pickle_module)
    223     finally:
    224         if new_fd:

/usr/local/lib/python3.6/site-packages/torch/serialization.py in _load(f, map_location, pickle_module)
    368     unpickler = pickle_module.Unpickler(f)
    369     unpickler.persistent_load = persistent_load
--> 370     result = unpickler.load()
    371
    372     deserialized_storage_keys = pickle_module.load(f)

AttributeError: Can't get attribute '_rebuild_tensor' on <module 'torch._utils' from '/usr/local/lib/python3.6/site-packages/torch/_utils.py'>

@colesbury
Copy link
Member Author

colesbury commented Mar 20, 2017 via email

@soumith
Copy link
Member

soumith commented Mar 20, 2017

this error would be fixed with the release on Wednesday v0.1.11.

@ecolss
Copy link

ecolss commented Mar 20, 2017

I see, thanks

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

VGG classifier setting different from Original paper

3 participants