@Michal_w @piotr.czapla @Blanche @kijes @tillia
fajnie ze sie pojawiliscie! milo bylo Was poznac i porozmawiac na temat fastai / deep learning
od nastepnego tygonia przełączamy sie na hangouts. jitsi meet nie dzialalo poprawnie z wideo… link do hangouts jest juz w pierwszym wpisie w tym watku, i tutaj ponizej
Where we should share bugs? I’m playing with fruits data set and learn.recorder.plot() shows empty plot after running learn.lr_find()
In previous fastai course I had the same problem, when my dataset was small, but my batch size (bs) was big. Maybe that’s the case.
I don’t mind posting bugs here. Forum tools are quite good at searching within a thread so is cool.
I run notebook in colab, and got better error_rate then running it local:
Here’s local on 1060 6gb:
I don’t know what could be the cause, for my lower error rate. I remember it was the case also, when I was doing part1 6 months ago (when compering to scores in notebooks).
Could that be that the images are randomly split into train and valid sets by fastai?
Yeah maybe… I run it many times, there were different error rates, from 0.059 to 0.067, with average around 0.062, and with colab I got 0.057.
I won’t be pondering about this too much though. I just will adjust my workflow. I’ll be using smaller bs/samples on local for fast prototyping, and when I think I’m done then I’ll re-run them in colab/GoogleCloud with bigger bs, full dataset.
As a side note, to run ResNet50 local on 6gb I needed bs=24, to run it in colab I could set bs to 32. I’m really impressed with colab.
i initially though i used bs=64 for resnet50, but then noticed bs=bs//2 is used, so it’s 32 as sayko said 
data = ImageDataBunch.from_name_re(path_img, fnames, pat, ds_tfms=get_transforms(), size=320, bs=bs//2)
and yeah colab is good!
I like seeing those little changes in notebooks, and figuring out, what/and why I see that they also changed size form 299 to 320
interesting, i wonder why? for speed or performance?
Maria, if there is a bug i’m sure there is plenty of ppl interested in that. You can try to create a new thread to discuss it and then post the link here on so we can try to help you. That way it may serve others that has the same issue as you.
Hi,
I’m getting each run little bit different results +/- half % simetumes. From other side if you think that in learning process done parts as chooaen randomly. I think it’s normal be it’s unlikely two learning processes wilk be the same.
Cheers
Michal
I’ve managed to pull 1,5% error rate on pneumonia X-rays pictures.
Now I’m planning to visualize convolutions as someone did here on one of my notebooks: https://github.com/kheyer/ML-DL-Projects/blob/master/Pets%20TSNE/pets_tsne.ipynb Looks super interesting.
@piotr.czapla Sadly I can no longer reproduce it.
Great! You should submit to kaggle. Seems you would get medal for this comp?
I want to do this comp but not many days left…
Oh I didn’t realize there is a contest for this type of data 
I just did this dataset: https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia Have to try the contest one now //edit sadly pt 1 of the contest is over, so kaggle won’t accept new participants
Let me know if I can help ,
I was working on same data and also noticed it is after deadline 
but its actually very impressive even without not understanding x-ray pictures not even know what is pneumonia we can detect health changes in 98%+ probably better then trained radiologist 
This dataset looks interesting (15 classes):
After I finish setup my machine, I’m going to play with it
Crestle is free until mid November as per this post, if you are looking for some GPU resources, storage is limited to 75GB though but no credit card needed for now : https://forums.fast.ai/t/platform-crestle/28028/33
@sayko did you have any issues with temperature of your GPU on linux?
Do you have any experience in GPU cooling management on headless machines?
I think there is some issue with fans speed control, but I’m not sure if it is only mine problem 