search

forresti / SqueezeNet

SqueezeNet: AlexNet-level accuracy with 50x fewer parameters
BSD 2-Clause "Simplified" License
2.17k stars 723 forks source link

Why v1.1 does not have deploy.prototxt #12

Closed changyun79 closed 6 years ago

changyun79 commented 8 years ago

Hello, I am wondering why v1.1 does not have deploy.prototxt. Missed commit? Thank you.

forresti commented 8 years ago

We can add that. Thanks to @terrychenism for creating this on SqueezeNet v1.0.

Nobody on our team uses the deploy.prototxt interface very often... we mostly load custom training/testing sets into LMDBs to avoid being bottlenecked by I/O.

changyun79 commented 8 years ago

Thanks for the quick response. I appreciate if it can be added in.

besirkurtulmus commented 8 years ago

I'd also appreciate if you could add deploy.txt too.

austingg commented 8 years ago

@besirkurtulmus it's quite easy to make deploy.prototxt with reference train_val.prototxt. You can just modified the input and output layers and there are some reference in Caffe/models directory.

psyhtest commented 8 years ago

@forresti

I thought one needs deploy.prototxt to measure the execution time using caffe time? I appreciate your research is mostly about achieving the same accuracy but what about performance?

forresti commented 8 years ago

I often run caffe time on train_val.prototxt files... it works great. :)

psyhtest commented 8 years ago

@forresti

Thanks! Do you know which batch size gets picked up for benchmarking - from the TRAIN or TEST specification?

forresti commented 8 years ago

I think train. But, caffe time will print the dimensions of all layers including the batch size. On Jul 8, 2016 1:19 AM, "Anton Lokhmotov" notifications@github.com wrote:

forresti Thanks! Do you know which batch size gets picked up for benchmarking - from the TRAIN or TEST specification? Is it deterministic?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/DeepScale/SqueezeNet/issues/12#issuecomment-231302519, or mute the thread https://github.com/notifications/unsubscribe/AB7Sqg2wwWGzjM1NGSTmsTVpXGrHcx3jks5qTggrgaJpZM4I2qdc .

AliaMYH commented 7 years ago

What exactly is the difference between the train_val and the deploy prototxt ? What are their different purposes?

amroamroamro commented 7 years ago

@forresti there's now three PRs to add deploy.txt for v1.1: #20, #21, #43