Commit graph

330 commits

Author SHA1 Message Date
fa3165a5b2
dataset: simplify dataset_predict 2022-09-14 17:33:17 +01:00
279e27c898
fixup 2022-09-14 17:16:49 +01:00
fad3313ede
fixup 2022-09-14 17:14:04 +01:00
1e682661db
ai: kwargs in from_checkpoint 2022-09-14 17:11:06 +01:00
6bda24d4da
ai: how did I miss that?!
bugfix ah/c
2022-09-14 16:53:43 +01:00
decdd434d8
ai from_checkpoint: bugfix 2022-09-14 16:49:01 +01:00
c9e00ea485
pretrain_predict: fix import 2022-09-14 16:02:36 +01:00
f568d8d19f
io.open → handle_open
this was we get transparent .gz support
2022-09-14 16:01:22 +01:00
9b25186541
fix --only-gpu 2022-09-14 15:55:21 +01:00
1876a8883c
ai pretrain-predict: fix - → _ in cli parsing 2022-09-14 15:12:07 +01:00
f97b771922
make --help display help 2022-09-14 15:03:07 +01:00
3c1aef5913
update help 2022-09-13 19:36:56 +01:00
206257f9f5
ai pretrain_predict: no need to plot here anymore 2022-09-13 19:35:44 +01:00
7685ec3e8b
implement ability to embed & plot pretrained embeddings 2022-09-13 19:18:59 +01:00
7130c4fdf8
start implementing core image segmentation model 2022-09-07 17:45:38 +01:00
22620a1854
ai: implement saving only the rainfall encoder 2022-09-06 19:48:46 +01:00
4c4358c3e5
whitespace 2022-09-06 16:24:11 +01:00
4202821d98
typo 2022-09-06 15:37:36 +01:00
3e13ad12c8
ai Bugfix LayerContrastiveEncoder: channels → input_channels
for consistency
2022-09-05 23:53:16 +01:00
ead8009425
pretrain: add CLI arg for size of watch prediction width/height 2022-09-05 15:36:40 +01:00
9d39215dd5
dataset: drop incomplete batches 2022-09-05 15:36:10 +01:00
1d1533d160
ai: how did things get this confusing 2022-09-02 18:51:46 +01:00
1c5defdcd6
ai: cats 2022-09-02 18:45:23 +01:00
6135bcd0cd
fixup 2022-09-02 18:41:31 +01:00
c33c0a0899
ai: these shapes are so annoying 2022-09-02 18:39:24 +01:00
88acd54a97
ai: off-by-one 2022-09-02 18:13:48 +01:00
23bcd294c4
AI: 0.75 → 0.5? 2022-09-02 18:12:51 +01:00
b0e1aeac35
ai: knew it 2022-09-02 18:10:29 +01:00
ef0f15960d
send correct water shape to 2022-09-02 18:09:36 +01:00
ad156a9a00
ai dataset: centre crop the water data to 75% original size
this should both help the model and reduce memory usage
2022-09-02 18:05:32 +01:00
efe41b96ec
fixup 2022-09-02 17:58:02 +01:00
f8ee0afca1
ai: fix arch_name plumbing 2022-09-02 17:57:07 +01:00
3d44831080
Add NO_PREFETCH env var 2022-09-02 17:55:04 +01:00
3e0ca6a315
ai: fix summary file writing; make water encoder smaller 2022-09-02 17:51:45 +01:00
389216b391
reordering 2022-09-02 17:31:19 +01:00
b9d01ddadc
summary logger → summarywriter 2022-09-02 17:28:00 +01:00
9f7f4af784
dataset: properly resize 2022-09-02 17:09:38 +01:00
c78b6562ef
debug adjustment 2022-09-02 17:00:21 +01:00
c89677abd7
dataset: explicit reshape 2022-09-02 16:57:59 +01:00
c066ea331b
dataset: namespace → dict
Python is so annoying
2022-09-02 16:07:44 +01:00
dbff15d4a9
add bug comment 2022-09-01 19:05:50 +01:00
3e4128c0a8
resize rainfall to be 1/2 size of current 2022-09-01 18:47:07 +01:00
8a86728b54
ai: how did I miss that 2022-09-01 18:07:53 +01:00
b2f1ba29bb
debug a 2022-09-01 18:05:33 +01:00
e2f621e212
even moar debug logging 2022-09-01 17:49:51 +01:00
2663123407
moar logging 2022-09-01 17:40:23 +01:00
c2fcb3b954
ai: channels_first → channels_last 2022-09-01 17:06:18 +01:00
f1d7973f22
ai: add dummy label 2022-09-01 17:01:00 +01:00
17d42fe899
ai: json.dumps 2022-09-01 16:21:52 +01:00
940f7aa1b5
ai: set self.model 2022-09-01 16:20:23 +01:00
f1be5fe2bd
ai: add missing arguments to LossContrastive 2022-09-01 16:14:00 +01:00
ddb375e906
ai: another typo 2022-09-01 16:12:05 +01:00
cb3e1d3a23
ai: fixup 2022-09-01 16:10:50 +01:00
e4c95bc7e3
typo 2022-09-01 16:09:24 +01:00
cfbbe8e8cf
ai: global? really? 2022-09-01 16:06:24 +01:00
4952ead094
ai: try the nonlocal keyword? I'm unsure what's going on here
....clearly I need to read up on scoping in python.
2022-09-01 16:02:37 +01:00
8bdded23eb
ai: fix 'nother crash' name ConvNeXt submodels 2022-08-31 18:57:27 +01:00
b2a320134e
ai: typo 2022-08-31 18:54:03 +01:00
e4edc68df5
ai: add missing gamma layer 2022-08-31 18:52:35 +01:00
51cf08a386
ResNetRSV2 → ConvNeXt
ironically this makes the model simpler o/
2022-08-31 18:51:01 +01:00
3d614d105b
channels first 2022-08-31 18:06:59 +01:00
654eefd9ca
properly handle water dimensions; add log files to .gitignore
TODO: add heightmap
2022-08-31 18:03:39 +01:00
5846828f9e
debug logging 2022-08-31 17:48:09 +01:00
12c77e128d
handle feature_dim properly 2022-08-31 17:41:51 +01:00
c52a9f961c
and another 2022-08-31 17:37:28 +01:00
58dbabd561
fix another crash 2022-08-31 17:33:07 +01:00
dbf929325a
typo; add pretrain slurm job file 2022-08-31 16:32:17 +01:00
f2312c1184
fix crash 2022-08-31 16:25:27 +01:00
15a3519107
ai: the best thing about implementing a model is that you don't have to test it on the same day :P 2022-08-11 18:26:28 +01:00
c0a9cb12d8
ai: start creating initial model implementation.
it's not hooked up to the CLI yet though.
Focus is still on ensuring the dataset is in the right format though
2022-08-10 19:03:25 +01:00
b52c7f89a7
Move dataset parsing function to the right place 2022-08-10 17:24:55 +01:00
28a3f578d5
update .gitignore 2022-08-04 16:49:53 +01:00
323d708692
dataset: add todo
just why, Tensorflow?!
tf.data.TextLineDataset looks almost too good to be true..... and it is, as despite supporting decompressing via gzip(!) it doesn't look like we can convince it to parse JSON :-/
2022-07-26 19:53:18 +01:00
b53c77a2cb
index.py: call static function name run 2022-07-26 19:51:28 +01:00
a7ed58fc03
ai: move requirements.txt to the right place 2022-07-26 19:25:11 +01:00
e93a95f1b3
ai dataset: add if main == main 2022-07-26 19:24:40 +01:00
de4c3dab17
typo 2022-07-26 19:14:55 +01:00
18a7d3674b
ai: create (untested) dataset 2022-07-26 19:14:10 +01:00
dac6919fcd
ai: start creating initial scaffolding 2022-07-25 19:01:10 +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