Fix train_lenet_decolle_fa.py import issue

ZhenmingYu commented 2 years ago

Was importing standard decolle rather than decolle_fa modules.

clee1994 commented 2 years ago

Running this gives:

File ".local/lib/python3.7/site-packages/decolle/lenet_decolle_model_fa.py", line 109, in build_conv_stack
    readout = nn.FALinear(int(feature_height * feature_width * Nhid[i + 1]), out_channels)
AttributeError: module 'torch.nn' has no attribute 'FALinear'

I removed the "nn." in decolle/lenet_decolle_model_fa.py line 109 and it works.

eneftci commented 2 years ago

Hi Clemens, I recommend using the branch update_21014. Do you get the same problem? I'll merge this into master soon.

clee1994 commented 2 years ago

Hey Emre, the branch you mentioned also has the same error, e.g. importing non-FA modules

https://github.com/nmi-lab/decolle-public/blob/61aca54dadfc2ef4781a25b8c1324e19f676e871/scripts/train_lenet_decolle_fa.py#L12

and using nn.FALinear

https://github.com/nmi-lab/decolle-public/blob/61aca54dadfc2ef4781a25b8c1324e19f676e871/decolle/lenet_decolle_model_fa.py#L109

Also, I am trying to rerun the numbers from the decolle paper (e.g. DVS gesture 4.46% error) the paper mentions a hyperparameter search, but I am not sure if that matches the hyperparameters in

https://github.com/nmi-lab/decolle-public/blob/master/scripts/parameters/params_dvsgestures_torchneuromorphic.yml

(for example, the paper says burning 50ms and this config says 100 etc).

Any chance to get the hyperparameters used in the paper, or maybe even some pre-trained weights?

nmi-lab / decolle-public

Fix train_lenet_decolle_fa.py import issue #24