Commit graph

13 commits

Author SHA1 Message Date
222a6146ec
write glue for .jsonl.gz → .tfrecord.gz converter 2022-08-08 15:33:59 +01:00
927c30e189
recompress files in the right order 2022-07-25 18:44:23 +01:00
3332fa598a
Add new recompress subcommand
also fix typos, CLI definitions
2022-07-25 17:54:23 +01:00
82e826fd69
Fix bugs in remainder of rainfallwrangler:uniq :D 2022-07-22 18:05:03 +01:00
a966cdff35
uniq: bugfix a lot, but it's not working right just yet
There's still a bug in the file line deletor
2022-07-08 19:54:24 +01:00
3b2715c6cd
recordify: fix process exiting and imcomplete files issues
• Node.js not exiting at all
 • Node.js exiting on end_safe ing stream.Writable (?????)
 • Incomplete files - "unexpected end of file" errors and invalid JSON
2022-07-08 18:54:00 +01:00
b9a018f9a9
properly close all teh streams 2022-07-08 16:51:17 +01:00
1a657bd653
add new uniq subcommand
It deduplicates lines in the files, with the potential to add the ability to filter on a specific property later.
The reasoningf or this is thus:
1. There will naturally be periods of time where nothing happens
2. Too many duplicates will interfere and confuse with the contrastive learning algorithm, as in each batch it will have less variance in samples

This is especially important because contrastive learning causes it to compare every item in each batch with every othear item in the batch.
2022-07-04 19:46:06 +01:00
1297f41105
.tfrecord files are too much hassle
let's go with a standard of .jsonl.gz instead
2022-07-01 18:28:39 +01:00
3cb7e42505
it doesn't crash as much now, but it still isn't behaving. 2022-05-19 18:52:15 +01:00
bb018c53f6
Fix many bugs
Many bugs remain though
2022-05-19 17:54:14 +01:00
cc5efbae8a
Implement tfrecodify subcommand.
It's all still untested, but that's the next step
2022-05-19 17:15:15 +01:00
8a9cd6c1c0
Lay out some basic scaffolding
I *really* hope this works. This is the 3rd major revision of this
model. I've learnt a ton of stuff between now and my last attempt, so
here's hoping that all goes well :D

The basic idea behind this attempt is *Contrastive Learning*. If we
don't get anything useful with this approach, then we can assume that
it's not really possible / feasible.

Something we need to watch out for is the variance (or rather lack
thereof) in the dataset. We have 1.5M timesteps, but not a whole lot
will be happening in most of those....

We may need to analyse the variance of the water depth data and extract
a subsample that's more balanced.
2022-05-13 19:06:15 +01:00