Lesson 1 In-Class Discussion ✅

antorsae · October 23, 2018, 2:56am

i meant is it being shown for validation or training set?

bluesky314 · October 23, 2018, 2:56am

We need some docs to read about these functions he’s describing!

sgugger · October 23, 2018, 2:56am

We are, it’s due to a major outage of GitHub yesterday. It’ll be fixed soon.

lesscomfortable · October 23, 2018, 2:56am

That depends. If val_loss > train_loss you are overfitting. If train_loss < val_loss you are underfitting. It is usually good to overfit a little bit in the first epochs, you can always regularize more later.

sgugger · October 23, 2018, 2:56am

You are

Interogativ · October 23, 2018, 2:56am

train - what we train with
valid- classified data how are we doing during training
test-un clasified data, real data

cedric · October 23, 2018, 2:56am

fastai learn.fit_one_cycle API implemented Leslie Smith’s paper:

Paper published on 26 Mar 2018.

stas · October 23, 2018, 2:57am

Yes. github was broken for 24 hours and has just recovered but our site didn’t. will be fixed tonight.

dtrik · October 23, 2018, 2:57am

docs.fast.ai

magiclantern · October 23, 2018, 2:57am

Is it possible to see the details of the error rate? For example false positive vs false negatives?

Muosvr · October 23, 2018, 2:58am

valid_loss can be bigger due to dropout being used in training(prevent overfitting). Jeremy will probably explain that later

gazay · October 23, 2018, 2:58am

It will fetch pytorch v0.4.1 but you’ll still have torch-nightly needed for latest fastai

(py_36_env) ⏣ 0:code:$ pip list | grep torch
torch                             0.4.1
torch-nightly                     1.0.0.dev20181022
torchvision                       0.2.1
torchvision-nightly               0.2.1

Tripti · October 23, 2018, 2:58am

I tried to clone the repo https://github.com/fastai/course-v3 and doesnt seem to working. Am I missing something?

dotkay · October 23, 2018, 2:58am

so, all the blocks except the last (classification block) would be the same? did I understand that correct?

thanks
.K.

Carsten · October 23, 2018, 2:58am

what would be the minimal error rate of fine grained cats vs. dogs, if you use a fully trained backbone and something better than resnet34?

champs.jaideep · October 23, 2018, 2:59am

one quick questio we may do lot of good thigns in FAI… but are organization comfortable deploying FAI code to the prod… Most of them tend to work with conventional framweworks like pytorch,keras,tensorflows etc…

nithanaroy · October 23, 2018, 2:59am

http://docs.fast.ai/basic_data does not work. Similarly many other docs.

iskode · October 23, 2018, 2:59am

Hi,
would like to know: how to train from scratch i.e. not using the pretrained weights ?

fredguth · October 23, 2018, 2:59am

Please, please turn off the mic!!

j_simon · October 23, 2018, 3:00am

Actually the validation set is some hold out data. It can be by accident that the validation set is a bit easier, so the loss is less.