wassname/multifit
1
0
Fork 0
mirror of https://github.com/wassname/multifit.git synced 2026-09-09 11:27:26 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add experiment with biclassifier (concat)

Browse Source
This commit is contained in:
Piotr Czapla
2018-11-19 15:31:10 +01:00
parent 9f526a9849
commit 2bde30a402
1 changed files with 26 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
results/bilcls_qrnn-wt-103-wd0
+26
View File
@@ -0,0 +1,26 @@
pretraining
0 - 8 lost, about 0.30 0.32 after 5 epochs
8 3.744213 4.021358 0.308377 0.325672
9 3.700674 4.021499 0.308637 0.325881
10 3.674045 4.022058 0.308656 0.325903
--- crash---
$ python -m ulmfit.train_clas --data_dir data --model_dir data/wiki/wikitext-103/models --pretrain_name=bilm-wt-103 --qrnn=True --name 'concat-2x' --cuda-id=0 --bs 40 --train=True --bidir=True
Dataset: imdb. Language: en.
Using QRNNs...
BiLM
Loading the pickled data...
Train size: 22500. Valid size: 2500. Test size: 25000.
loading encoder
Starting classifier training
epoch train_loss valid_loss accuracy
1 0.346874 0.276838 0.881200
epoch train_loss valid_loss accuracy
1 0.321121 0.243946 0.901200
epoch train_loss valid_loss accuracy
1 0.303174 0.234627 0.908400
epoch train_loss valid_loss accuracy
1 0.295207 0.227533 0.912800
2 0.269754 0.221328 0.914400
Saving models at data/wiki/wikitext-103/models
accuracy: tensor(0.9144)