These are chat archives for beniz/deepdetect

22nd
Nov 2017
rperdon
@rperdon
Nov 22 2017 14:00
Not sure if this is a bug, but the reported statistics after finetuning barely change. Initial run of 50k and when tested the model is reporting all 0. After finetuning twice, the only reported change in stats is the train_loss value which lowers. All other stats return the same. The model at this point is now showing some ability to classify properly.
{"status":{"code":201,"msg":"Created"},"body":{"model":{"repository":"/source"},"measure":{"train_loss":0.5598307847976685,"mcll":1.797693134862316e308,"accp":0.4312551048189491,"acc":0.4312551048189491,"f1":0.58882937183257,"recall":0.715550103931875,"iteration":9999.0,"precision":0.5002392344466011}},"head":{"method":"/train","time":911.0}}
{"status":{"code":201,"msg":"Created"},"body":{"model":{"repository":"/source"},"measure":{"train_loss":0.25321775674819949,"mcll":1.797693134862316e308,"accp":0.4312551048189491,"acc":0.4312551048189491,"f1":0.58882937183257,"recall":0.715550103931875,"iteration":19999.0,"precision":0.5002392344466011}},"head":{"method":"/train","time":1792.0}}
These are the 2 finetuning runs.
The training_loss improves, but the 2nd finetuning started showing that some proper classifications were happening. The precision number between them though shows no change.