ajbrock / freezeout Goto Github PK

in the utils.py, ,,,,,,,,what is the "import path"。。。。。。there is no function callled path

ImportError: No module named 'path'

This code is linked from the fashion-mnist repo, w/ very good results. Do you have a script somewhere I might be able to use to reproduce those numbers?

Thanks
Ben

Thank you for your contribution.
In the line below, the total iterations per epoch is assumed to be 1000 (regardless of the batch size or dataset). It is not clear which dataset or batch_size this number assumes.

m.max_j = self.epochs * 1000 * m.lr_ratio
I assume this number would be replaced by total iterations per epoch for a different dataset or batch_size. Correct me if I am wrong please.

hi,,,,Im sorry to bother you again....there are still some errors.......
like this:
file train.py ,line 283, in main
train_test(**vars(args))

filenotfoundError:[Error 2] No such file or directory:Path('logs/densenet_k12L76_ice80_cubicTrue_seed0_epochs100c100_log.jsonl')
looking forward to your reply

The 0.05 in the update LR formula overrides the initialized LR per layer (the 1e-1).
The line that inits the LR per layer with the scale_lr option is overridden by the formula.
m.lr = 1e-1 / m.lr_ratio if self.scale_lr else 1e-1

The formula with the change:
self.optim.param_groups[i]['lr'] = (m.lr)*(1+np.cos(np.pi*self.j/m.max_j))

Maybe it's worth mentioning in README that you need to install path.py before running the script.

not a bad idea at all,
Dont you think the tensorflow / pytorch community would like to improve it with your idea?
These days if you can compress video 3% more then it has a huge impact on data storage centers.
If you can train an AI faster, 10%,... boy that's huge... think of the impact on data centers, training time.
You should draw some attention from nvidea / intel.