dd4effc690
changes to logging to sync frequency
2024-06-11 19:57:48 +01:00
6d0682bbed
re-enabled expt2 test1
2024-06-11 14:22:16 +01:00
85363021be
Added trainable residual penalty & logging for it
2024-06-11 14:18:56 +01:00
a14babd58a
New attempt at expt2 semantic loss
2024-06-10 16:12:02 +01:00
888fe9f2e8
Added automatic enabling of tensor cores
2024-06-10 12:58:19 +01:00
08fe32302c
Made model smaller to avoid overfitting
2024-06-07 16:09:51 +01:00
c7133a8bb1
Switched loss in expt2 to smooth_l1
...
Switched from mse_loss to smooth_l1_loss to avoid exploding gradient and
NaNs when using mse_loss.
2024-06-07 11:48:36 +01:00
1c21ee25d7
Added separate loss funcs for train, val, and test
2024-06-06 22:33:01 +01:00
ea77a055f8
Update README.md
...
Fixed typo
2024-06-06 15:12:25 +01:00
f10aa2fd19
Refactor of experiment2 for convenience
2024-06-06 15:02:39 +01:00
165ef045ea
First working attempt at semantic loss for expt2
2024-06-06 12:40:41 +01:00
d2ec5c0c1a
Added ability to train on submodules in expt2
2024-05-30 17:34:53 +01:00
52fb30a9ed
Completely implemented baseline for expt2
2024-05-30 16:02:46 +01:00
2720887f88
Added cli for configuring/launching experiments
2024-05-23 11:06:11 +01:00
9536beea04
Refactored to lay groundwork for TUI
2024-05-23 09:50:01 +01:00
02b9e43f4d
Made loggers togglable from top level
2024-05-23 09:21:48 +01:00
1be0fa1020
Added overview and plan for expt 2
2024-05-22 15:40:24 +01:00
c3dba79a0b
Refactored slightly to allow other experiments
2024-05-22 10:35:14 +01:00
ae14a1d7c0
Added a garbage semantic function
...
Added a semantic function where the injected "knowledge" is just random
garbage. This function was written to isolate the "knowledge" component
of the other semantic functions, basically to ensure it's the matrices
and not the rest of the process that is making the difference in
training.
2024-05-15 19:01:35 +01:00
e3690b0425
Added wandb.finish to properly start new runs
2024-05-15 18:40:20 +01:00
8a87e864fb
Fixed wandb logging mistake
2024-05-15 18:39:42 +01:00
4212c543f8
Update before first push to remote
2024-05-15 16:38:40 +01:00
15a57cc229
Added wandb logging
2024-05-15 16:37:53 +01:00
6600a79f71
Added more varied semantic loss functions
2024-05-15 16:37:53 +01:00
01127de4b3
Added testing step at end of training
2024-05-15 16:37:53 +01:00
62649a39da
Modified model to use SGD for less optimal training
...
There's no point having a loss function testing sandbox where the baseling trains perfectly in 2 epochs. I've DEoptimised the training in this commit to ensure the testing sandbox is actually useful.
2024-05-15 16:37:53 +01:00
389c47ef28
Fixed similarity matrix normalization
2024-05-15 16:37:53 +01:00
66b23a8e76
Minimized model and prepared for testing new loss functions
2024-05-15 16:37:53 +01:00
708f9952b3
First functional training loop
2024-05-15 16:37:53 +01:00
674175b260
Added trainer and made dataset hot-swappable
2024-05-15 16:37:53 +01:00
63bea4d355
Updated dataloader to default to tensor transform
2024-05-15 16:37:53 +01:00
02d70964e1
Added image transformations to dataset fetching
2024-05-15 16:37:53 +01:00
223b064564
Tweaked dataloader for more reliable loading
2024-05-13 15:38:50 +01:00
8204109561
Created consistent dataloader
2024-05-13 15:29:55 +01:00
6df0fea652
Modifications to env
2024-05-13 14:59:27 +01:00
0782af696b
Initial env configuration
2024-05-13 11:41:35 +01:00
