Since we’re all over, we should choose a convenient time to discuss where we are at and what we’re doing. I say we block off 3-4 hours on Sundays that we tune in weekly here, otherwise just post questions when you have them? Thoughts?
1-4pm Sunday?
Since we’re all over, we should choose a convenient time to discuss where we are at and what we’re doing. I say we block off 3-4 hours on Sundays that we tune in weekly here, otherwise just post questions when you have them? Thoughts?
1-4pm Sunday?
Here is a current summary of what I’ve been doing and issues I’ve run into:
I’m using Paperspace to work on the class. I originally set up my computer using old instructions (because I did it a week before class). I originally chose the fastai set up paperspace machine. I pulled in course-v3 from github and have gone through the code for lesson one and am working on training my own code.
Some issues I’ve run into - wrong version of PyTorch on my machine - found information to update on official PyTorch website.
I added
import torch
GPU issues -
Running torch.cuda.is_available()
produced FALSE
.
After troubleshooting I restarted my instance. That worked.
Issues with normalize:
ValueError: padding_mode needs to be 'zeros' or 'border', but got reflection
Added padding_mode='zeros'
to ImageDataBunch.
I’m ready to start playing with my own dataset. I’ve found one I like on kaggle that I want to try out first. Has anyone worked with adding labels from a csv to categorize images? Any help would be much appreciated.
UPDATE:
I was successful in using lavels from csv using the following code:
from pathlib import Path
import pandas as pd
path = Path("/home/paperspace/data/DecorPatterns/images/")
df = pd.read_csv("/home/paperspace/data/DecorPatterns/decor.csv")
#df.head()
df1 = df[['file', 'decor']]
df1.columns = ['name', 'label']
df1.head()
data = ImageDataBunch.from_df(path, df1, size=224)
I’m also getting this but I wonder if it’s not some other error about the versions installed? We should not be patching fastai directly…
I did a git pull on fastai to have the latest version, but none the less I still ran into this issue. Maybe we should be using something else?
Seems very sensible – The logistics otherwise are going to be a nightmare.
-L
This version of the course definitely does not seem to be VM - first. I started it before using a VM and switched to notebook for this go around. The lack of being able to hack directly on the machine is an issue for sure – Though I did get a model trained on my own stuff last time. I’ve got a post in the paperspace general thread asking about this. FWIW – the Notebook only P5000 is a site cheaper than running a vm.
My old dataset: Failing to complete training in Lesson 1 for own data
Post asking about Gradient hacking: https://forums.fast.ai/t/platform-paperspace-and-gradient/27338/47?u=larcat
So this wasn’t very intuitive for me but I finally figured it out. After reading over and over that I needed pytorch 1.0 and I needed pytorch nightly I finally realized they are both pytorch. So uninstalling pytorch and re-installing pytorch_nightly has solved my issues.
I hope this helps any other people confused about the fact that these are both pytorch.
Hi Everyone,
I am late to the this discussion. Probably, did not have access to this thread because I was not registered for in-person/online course.
I am just wondering if you are still continuing meetings in Chicagoland ? Seems like we have good numbers here on thread.
If meetings are not scheduled yet then can we start something book reading club/ journal reading/ tutorial presentation? at Chicago Public Library locations?
Hopefully, Interacting with each other on regular basis will build fastai community here and will make us better learners.
Thank you!
I think it would be fun to get together, or maybe even start a meetup!
Hi Daniel,
Thank you for reply!!
I am close to downtown. How about arranging meeting on sometime coming Saturday/Sunday?
We can definitely start a meetup.
@larcat, @agoldina, @emi, @tomle, @eugene, are any of you interested in meeting up and discussing fast.ai, talking about projects, etc with @tamhash and I on a Sunday in the near future.
As an alternative we could have a quick conference call(google hangout,etc) to talk about what we want to accomplish in an in person meeting.
I’ll be honest, post class I got caught up with work & personal life so it would be fantastic to gather to do some joint work or projects.
I am northwest and prefer somewhere I can either drive or get to on the blue line.
I’ve been planning to re-watch all the lessons and re-do my notes and re-do all the workbooks, so I want to see what everyone else’s plans and needs are - that way we can structure a meeting plan.
I’m not from the Chicago area (or US), just ended up chatting here due to some bug.
Sure Daniel and @agoldina, we can have quick conference call to start with.
Hopefully more will join us.
You want to try to do a quick call this Sunday? @tamhash @agoldina.
If you all want to crash a meetup, I am going to be presenting the fast ai approch at dupage business analytics group @ benedictine university on Tuesday February 26th, there will be free pizza and pop.
The call sounds great. I’m busy that Tuesday unfortunately. Phone or skype?
Doesn’t matter to me, in the future I would like to try google hangout for screen sharing.
Maybe a straight up pytorch/fastai Chicago Meetup group?
I love the idea, if you know someone willing to sponsor, we should totally do it. I could see how we could have different members talk about the projects. If you all want I could ask if benedictine university would be willing to host a pytorch/fastai meetup. I am not sure if that will be too far from you all.
@Daniel.R.Armstrong …Unfortunately Tuesday will be busy for me to attend. Thank you for invitation!!
@larcat…I also like the idea of meetup group.
@Daniel.R.Armstrong , Benedictine University is quite a drive for me but if its matter of once in month meeting then I am ok with it.
For this weekend, I am ok with hangout or skype .
@agoldina will you be ok with google hangout?