search

deepgram / kur

Descriptive Deep Learning
Apache License 2.0
814 stars 107 forks source link

speech validating loss abnormal? #16

Closed liaoweiguo closed 7 years ago

liaoweiguo commented 7 years ago

Epoch 70/inf, loss=6.643: 100%|████████| 2432/2432 [17:24<00:00, 2.29samples/s] Validating, loss=806.257: 94%|█████████▍| 256/271 [00:40<00:02, 6.63samples/s] Prediction: "lok h n v eon owlt im on nonaih" Truth: "he was in a mood for music was he not"

training loss decreasing, while validating loss increasing, is that NOK?

scottstephenson commented 7 years ago

See #6

The dataset that the stock speech.yml uses is small (5 hours), so the model will only learn very basic language patterns (spaces, vowels-to-consonant ratio, etc). If you want much better performance then edit your speech.yml to use the tarballs with more hours (10, 20, 50, 100) in them from issue #6 (it will take up more space and take longer to train though!).