Minor learner loading bug with vs without custom loss function

takotab · February 26, 2019, 10:03am

So I found a weird bug maybe it’s because I use a custom dataset/loss function but still though I should let you guys know.
When I load my model I get different results with and without a loss function. Also the results without are very weird, they do not seems to make sense.

Here a link to the notebook:

gist.github.com

https://gist.github.com/takotab/b2437cd93f2eeaad875dc3ffc479053f

predict_bug_fastai_w_wo_loss.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "from fastai import *\n",
    "import matplotlib.pyplot as plt\n",

This file has been truncated. show original

sgugger · February 26, 2019, 3:05pm

Learner always has a loss function, it defaults to F.cross_entropy when you’re not using fastai to build your DataBunch, so that’s why you have those different results.
So the first section isn’t really without loss function, but with crossentropy

takotab · February 26, 2019, 5:56pm

Oke but still it is weird that the predictions are different when the loss function changes. Note I do not train the model (with these different loss functions) here I only load and predict here.