 Andrea Pedrotti
|
41647f974a
|
last training swipe on eval set is now performed on batch size equal to the training set batch size
|
2023-03-17 10:44:23 +01:00 |
|
Andrea Pedrotti
|
ee2a9481de
|
sampling GLAMI1-M dataset
|
2023-03-16 18:10:05 +01:00 |
|
Andrea Pedrotti
|
ee38bcda10
|
fixed TransformerGen init
|
2023-03-16 12:12:39 +01:00 |
|
Andrea Pedrotti
|
b34da419d0
|
fixed import
|
2023-03-16 11:49:49 +01:00 |
|
Andrea Pedrotti
|
17d0003e48
|
getter for gFun and VGFs config
|
2023-03-16 11:41:40 +01:00 |
|
Andrea Pedrotti
|
9d43ebb23b
|
implemented save/load for MT5ForSequenceClassification. Moved torch Datasets to datamanager module
|
2023-03-16 10:31:34 +01:00 |
|
Andrea Pedrotti
|
56faaf2615
|
changed wandb logging to a global level to keep track of all the VGFs and overall gFun
|
2023-03-15 16:35:49 +01:00 |
|
Andrea Pedrotti
|
65407f51fa
|
update trainer to handle mT5
|
2023-03-15 11:47:17 +01:00 |
|
Andrea Pedrotti
|
26aa0b327a
|
average pooling for MT5ForSequenceClassification and standardized return data
|
2023-03-15 11:46:53 +01:00 |
|
Andrea Pedrotti
|
5e41b4517a
|
implemented MT5ForSequenceClassification
|
2023-03-14 11:53:50 +01:00 |
|
Andrea Pedrotti
|
a3e183d7fc
|
avoid duplicating model on gpu when earlystop is triggered
|
2023-03-14 11:22:00 +01:00 |
|
Andrea Pedrotti
|
57918ec523
|
save and load datasets as pkl
|
2023-03-10 12:40:26 +01:00 |
andreapdr
|
7d0d6ba1f6
|
log average metrics via wandb
|
2023-03-10 11:21:33 +01:00 |
|
andreapdr
|
5ef0904e0e
|
logging average metrics
|
2023-03-09 17:59:18 +01:00 |
|
andreapdr
|
7e1ec46ebd
|
improved wandb logging
|
2023-03-09 17:03:17 +01:00 |
|
Andrea Pedrotti
|
84dd1f093e
|
logging via wandb
|
2023-03-07 17:34:25 +01:00 |
|
Andrea Pedrotti
|
6b7917ca47
|
typos
|
2023-03-07 14:33:30 +01:00 |
|
andreapdr
|
7dead90271
|
logging via wandb
|
2023-03-07 14:20:56 +01:00 |
|
Andrea Pedrotti
|
f274ec7615
|
moved dataloader function get_dataset
|
2023-03-06 12:40:12 +01:00 |
|
Andrea Pedrotti
|
77227bbe13
|
support for binary dataset; CLS dataset; updated gitignore
|
2023-03-06 11:59:47 +01:00 |
|
Andrea Pedrotti
|
0c9454cdd4
|
implemented multimodal pipeline; gFunDataset interface; fixed imports
|
2023-03-02 18:16:46 +01:00 |
|
Andrea Pedrotti
|
7041f7b651
|
fixed bug: we were applying sigmoid function 2 times when training the Attention-based aggregator
|
2023-02-14 14:28:17 +01:00 |
|
Andrea Pedrotti
|
7ed98346a5
|
fixed loading function for Attention-based aggregating function when triggered by EarlyStopper
|
2023-02-13 15:01:50 +01:00 |
|
Andrea Pedrotti
|
13ada46c34
|
attention-based aggregation function, first implementation, some hard-coded parameters
|
2023-02-10 18:29:58 +01:00 |
|
Andrea Pedrotti
|
2a42b21ac9
|
concat aggfunc
|
2023-02-10 12:58:26 +01:00 |
|
Andrea Pedrotti
|
3f3e4982e4
|
model checkpoint during training. Restore best model if earlystop is triggered
|
2023-02-10 11:37:32 +01:00 |
|
Andrea Pedrotti
|
9c2c43dafb
|
Visual VGF + MultiNewsDataset, working from data loading to testing
|
2023-02-09 18:42:27 +01:00 |
|
Andrea Pedrotti
|
dba2ed9c9c
|
Visual Transformer VGF
|
2023-02-09 16:55:06 +01:00 |
|
Andrea Pedrotti
|
4485d97e03
|
test commit
|
2023-02-09 16:47:17 +01:00 |
|
Andrea Pedrotti
|
8325262972
|
MultiNewsDataset download/save image fn + class for Visual View Generating Function
|
2023-02-08 18:11:53 +01:00 |
|
Andrea Pedrotti
|
19e4f294db
|
better way to save/load model via id ({config}_{date}); Implemented __str__ for each VGFs + get_config in GeneralizedFunnelling
|
2023-02-08 16:06:24 +01:00 |
|
Andrea Pedrotti
|
31fb436cf0
|
implemented fn to save/load trained gfun
|
2023-02-08 14:51:56 +01:00 |
|
Andrea Pedrotti
|
6b75483b55
|
bulk upload after refactoring
|
2023-02-07 18:40:17 +01:00 |