Live coding 7

This topic is for discussion of the seventh live coding session

<<< session 6session 8 >>>

Important note

During this video at one point I install mamba into my local conda env. This turns out to be a mistake, so later in the video I remove it and all the stuff it adds, and use micromamba instead. To skip all that headache, when you run through these steps don’t run the command that installs mamba at all!

Links from the walk-thru

  • (please contribute here)

What was covered

  • (please contribute here)

Video timeline - thank you @Daniel

00:00 Radek intro
Practice walk-thru 6 / chp1 on Kaggle

03:22 Jeremy introduces Kaggle competitions and reminded us that we have to join the competition before downloading the dataset

06:05Paperspace had some running issues even to the paid servers, and continue the session on local machine

10:08 How to install Kaggle (presumbly the code will work on both local and paperspace)? pip install --user kaggle

10:58 What’s special about the things installed/stored in a /bin/ directory? things/programs you can execute

11:17 Why we can’t run kaggle in terminal just now? because the /bin/ directory is not in $PATH

11:41 How to get the bin/ directory where kaggle is installed into $PATH? On paperspace, we can add the bin directory to $PATH through /storage/.bash.local file (here is Jeremy explain what the file does it for you) which enable kaggle to work on a new terminal next time; on local machine, you can do the same with .bashrc file; Radek and Jeremy also confirmed that if you want to run !kaggle in jupyter notebook, you need to also to add bin/ directory to $PATH in file.

12:31 Another question answered

13:14 How to edit .bashrc : 1. vim .bashrc or /.bash.local if paperspace; shift + g to go to the bottom of the file; o to enter a new line and edit; type export PATH=~/.local/bin:$PATH to add the bin/ directory to $PATH; 2. type :qw to save and exit vim; 3. in terminal type source !$ (meaning source theLastCommandParameter) to run .bashrc or you can close and reopen a terminal

14:12 How to get kaggle.json and put it in the right directory? 1. type kaggle will remind you to have kaggle.json ready in the right place; 2. go to your Kaggle account and click create new API token to download the kaggle.json file; 3. sudo cp .kaggle/kaggle.json ~anotherUser/.kaggle to copy a file (here kaggle.json) from one user to another user; 4. and other command to change the ownership of a file

18:09 How to use kaggle cli to download competition dataset? cd git; mkdir paddy; cd !$; kaggle competitions download -c paddy-disease-classification to create a folder to download the dataset; How big is the paddy doctor dataset size? 1GB

18:52 mamba vs pip on installing kaggle? for simple python packages, pip install is a more obvious choice than mamba install

20:35 How to unzip the dataset file? paddy$ unzip -q to unzip the file without display the processing messages with -q for quiet

20:54 How to get kaggle.json available in paperspace? 1. use the upload button in paperspace jupyter lab to upload kaggle.json and store it in ~/.kaggle/kaggle.json; 2. make sure the file permission is properly -rw------- and search chmod for how Jeremy taught us to change user permission on files; 3. Radek in the post suggested to write your own kaggle.json and save in paperspace as a different solution

21:46 How to copy files to different local server? How to utilize .ssh/config file? involving cp, chown, scp etc

24:14 How to check your local GPU? type nvidia-smi

25:06 How to check out the paddy dataset? mv ~/ ./ move the zip file to current folder from home directory, and unzip it with unzip -q; and ls to see what inside;

25:32How to explore the dataset folder in terminal? ls train_images/ |head and what this command does is to take the output of ls train_images/ and send it to head to process; How to look into files in subfolder? ls train_images/bacterial_leaf_blight/ | head; How to count the number of files inside a subfolder? ls train_images/bacterial_leaf_blight/ | wc -l in which | wc -l takes the output from ls folderName and count the number of lines; What are other useful functions like | head? | tail output the last few file names, | grep 33 to output the filenames with ‘33’ in it; paddy# cat will output what inside this folder; cat train.csv | head can give us the first few rows of the csv file; cat train.csv | grep ADT45 | wc -l to search rows with ‘ADT45’ in train.csv and count the number of those rows

Go through the above steps in paperspace
29:09 How to use the correct version of pip to install kaggle? 1. check the version which pip: if it is not from the directory opt/conda/bin/pip then remove the found version with mv root/conda/bin/pip for example, and restart terminal and try which pip to check and confirm; 2. ctrl + r and type pip to find pip install kaggle --user to install kaggle; 3. However, the warning message confirmed that the kaggle is installed in a directory /root/.local/bin which is not on the $PATH, and we need to get it into $PATH; 4. we could add it to the $PATH by .bashrc in local machine or .bash.local in paperspace, but we will try Radek’s approach with just add the following command into the end of the file export PATH=~/.local/bin:$PATH and be aware that bash is very much sensitive on space, so don’t leave space inside the command; 5. if you run export PATH=~/.local/bin:$PATH directly in terminal, then you don’t have to close and restart a terminal to activate the new $PATH; 6. type kaggle, it should run and also tellings us to get kaggle.json in the right directory; 7. upload kaggle.json to paperspace; 8. did kaggle created a ~/.kaggle folder for us? yes, we can confirm by cd ~/.kaggle; ls; 9. let’s move the updated kaggle.json into ~/.kaggle by mv /notebooks/kaggle.json ./; 10. we will find the permission of kaggle.json is wrong as it is -rw-r--r-- by ls -la; 11. we can fix the permission by chmod 600 kaggle.json; 11. now copy the download command from the paddy competition site, and run ~/.kaggle# kaggle download -c paddy-disease-classification; 12. let’s move the zip file into /notebooks/paddy/ by a trick taught by Jeremy earlier in walkthru 6 ~/.kaggle# mkdir -p ../notebooks/paddy/; mv ../notebooks/paddy/;; 13. let’s unzip the dataset zip file by unzip -q;

34:47 How to install unzip for paperspace? micromamba install -c conda-forge -p ~/conda unzip and hopefully by July or August 2022 mamba and unzip will be installed by paperspace without us doing it manually;

34:50 How to get the keyboard shortcut in terminal working for paperspace terminal too? (later sessions?)

36:59 How to deal with large dataset and cost of persistence storage? 1. if dataset sits in /notebooks/ it will be charged with 0.29 dollar per GB/month; 2. if don’t want to spend any money, then move dataset into home directory ~/, and you will lose the dataset when closing the notebook/machine and have to download it again when starting the notebook again. You could write a script to automate the downloading process; 3. If you will work on this dataset for a month, then $0.29 surely worth your time.

38:04 It takes a long time for Paperspace to store large files into persistence storage, at least for folders with a lot of files. How to make utilising dataset faster in paperspace without much trouble? 1. How to find the disk space a folder take? du -sh train_images/; 2. why does Jeremy think it may be a better idea to move dataset back to home directory ~? we don’t want to take a long time to utilise the datasest when training a model; 3. how to delete multiple folders and files in one go? rm -rf test_images/ train* sample_submission.csv and train* include train_images/ and train.csv, and even deleting them takes a while; 4. How much time does it take home directory to unzip a 1GB dataset file? time unzip -q (only 8 seconds, and only 5 seconds to download), so we should make the download and unzip automated in the home directory when starting the paddy notebook.

41:14 How can we create a script to automate the download and unzip process of paddy dataset inside home directory (when starting the paddy notebook)? 1. create a directory paddy inside \notebooks\, save the paddy jupyter notebook there, and the automation script there too; 2. What does the look like? ( #question shouldn’t we do pushd ~; popd in the script below?) 3. how to make executable? chmod u+x; 4. so now, you can run this every time you start the paddy notebook/machine to automatically download and unzip the dataset into the home directory; (or you can put into so that you don’t need to run yourself.

!#usr/bin/env bash
mkdir paddy
cd paddy
kaggle download -c paddy-disease-classification
unzip -q paddy-disease*

43:08 Create a jupyter notebook for paddy competition. 1. What’s the first thing Jeremy usually do for an image competition/task like this? from import * to get all the classes and methods on vision ready for use; 2. How to get the path for dataset ready? path = Path.home()/'paddy', use path to check the dataset directory; 3. use to show us what in there, and if we type Path.BASE_PATH = path, then will leave the directory part out and make the content name more readable; 4. How to take a look at the train.csv file? df = pandas.read_csv(path/'train.csv'); df to read the first and last few rows of the csv file; (continued to the next paragraph below)

from import *
path = Path.home()/'paddy'
Path.BASE_PATH = path
df = pandas.read_csv(path/'train.csv')

45:19 How to take a look at the image listed in the csv file? 1. How to get the path for train_images/ and a path to a particular category bacterial_leaf_blight: trn_path = path/'train_images'; blb = trn_path/'bacterial_leaf_blight';; 2. How to display an image with its path? img = PILImage.create(blb/'100330.jpg'); img; 3. What is the size of the image? img.size (size is not a method as it seems); 4. How to get all the image paths into a list? files = get_image_files(trn_path); files; img = PILImage.create(files[0]);; 5. How to check whether the image size is consistent in all images? [PILImage.create(o).size for o in files[:10]]

from import *
path = Path.home()/'paddy'
Path.BASE_PATH = path
df = pandas.read_csv(path/'train.csv')
trn_path = path/'train_images'
blb_path = trn_path/'bacterial_leaf_blight'
img = PILImage.create(blb/'100330.jpg')
files = get_image_files(trn_path)
img = PILImage.create(files[0]);
[PILImage.create(o).size for o in files[:10]]
ImageDataLoaders.from_folder(trn_path, valid_pct=0.2, seed=42, item_tfms=Resize(224)) dls.show_batch()
learn = vision_learner(dls, resnet34, metrics=error_rate)

49:40 What to think about image sizes? paddy image size is consistent, which is handy and interesting (why interesting?), and the size of the images seem big. Jeremy’s advice is to start with smaller sizes and then move onto original size, but 480x 640 is not terribly large.

51:00 What shall we do with the meta data variety of paddy/rice? 1. our model may not need the variety data to train itself for the task, as the images can give model enough knowledge to figure out variety; 2. but if the number of variety is too many, then this variety data may be useful/necessary to our model; 3. how to count the unique varieties of paddy using variety data? df.variety.value_counts(); 4. given only 10 unique varieties, and 70% of images belong to a single variety, so variety should be a low priority in our dataset for training model;

53:10 Build a ImageDataLoaders with train_images or its path: 1. since chap1 or 01_intro notebook of fastbook is on vision model, so maybe some code can be borrowed; 2. How to create a ImageDataLoader from train_images/ folder? dls = ImageDataLoaders.from_folder(trn_path, valid_pct=0.2, seed=42, item_tfms=Resize(224)) and we can show a batch of images with dls.show_batch();

56:50 Build a learner and fine-tune it: 1. How to build a model with vision classification model with Resnet34 using our dls? learn = vision_learner(dls, resnet34, metrics=error_rate) will download the resnet34 weights for us and create the model learn; 2. How to fine-tune this model for one epoch? learn.fine_tune(1)

57:07 How do we know our model is using GPU efficiently? 1. in terminal type nvidia-smi dmon and focus on reading two columns sm and mm (memory); 2. sm stands for GPU, we want to see the number to be high 70-90% is good, and if the error rate is quite low from training, then we can assume we are successfully training our model; 3. if sm is lower than 50% then it suggest GPU is not properly used by our model; 4. if that happens, what are the potential causes for the lower number to deal with: A most likely cause is that we are not reading and process images fast enough for model to use, which could be a result of storing dataset inside /storage/ or /notebooks/ as they are network storages meaning slow;

1:00:11 What are the potential solutions to improve on sm? 1. move dataset to local storage like home directory from storage/ or notebooks/; 2. resize images ahead of time (to make them smaller); 3. decrease the amount of augmentation; 4. pick an instance/machine with more CPUs

1:02:03 Next session for kaggle submission and kaggle notebook


TIL: nvidia-smi dmon gives you useful output. i used to use watch -n 1 nvidia-smi, which is not bad at all, but dmon is better, gives you all the info you care about and more :slight_smile:
check for sm % utilization, and aim for > 50%

link to video timestamp: Walk-thru 7 - YouTube


Agreed! I was doing nvidia-smi -l but this is much better. I can delay it by 5 seconds per refresh ( nvidia-smi dmon -d 5 )and that is just perfect, running in a separate window to give me a sense of how the GPU is doing.


I have also been using the GPU-Util information (from nvidia-smi) and assumed it was appropriate/sufficient (to ensure adequate GPU usage) until hearing that yesterday – this might be helpful for anyone else trying to make sense of why using only nvidia-smi could be problematic: cuda - nvidia-smi Volatile GPU-Utilization explanation? - Stack Overflow – “It doesn’t tell you anything about how many SMs were used, or how “busy” the code was, or what it was doing exactly, or in what way it may have been using memory.”

I’ve started using nvtop for the past year or so and I quite like it, sort of a htop for GPUs. Easily installable in recent Ubuntu versions.

Looks like this, configurable to some amount, I mostly keep an eye on GPU%, GPU mem%


Video timestamps for Walk-thru 7

0:42 - Background for Kaggle Competitions
6:29 - Waiting for Paperspace to load
10:00 - Detour into setting up for Kaggle competitions on your local machine
14:30 - Create API token for Kaggle
18:00 - Download kaggle competition zip file
25:00 - Using the pipe output to head to view the first 10 lines (also try wc -l, tail, grep)
28:55 - Back to Paperspace29:30 - Remove pip from /storage
30:00 - Install kaggle and update symlinks
32:00 - Upload kaggle API json
33:20 - Download kaggle competition file to Paperspace
35:00 - Install unzip to persistent conda env
36:45 - Unzipping kaggle file in notebooks is too slow
40:00 - Unzip kaggle file in home directory for speed
41:20 - Create an executable script for unzipping kaggle file
43:10 - Create a notebook to explore kaggle data
48:00 - Browse image files
51:00 - Review image metadata
53:00 - Image data loaders and labelling function
56:30 - Create a learner
57:00 - Monitor training with nvidia-smi dmon
1:02:00 - Summary


Thanks so much for doing these Matt!


I am trying to understand the troubleshooting of the pip problem at 29:24.

If I understand correctly, the system tried to use the pip in ~/conda/bin/pip but we want to use the version in ~opt/conda/bin/pip. If I’m understanding correctly, the latter is the original pip that came with the image of the machine; the former we installed during all of the attempts to get microconda and ctags working.

Why does the other pip not work when we did a full installation the other day? Is it because of the files that we deleted? Would we have avoided this problem if we had used micromamba to install? Do the error messages give any hint as to the solution to the problem.

Does $PATH get searched sequentially?

I’m trying to imagine troubleshooting this sort of problem myself in this highly customized setup.I’m not sure how to proceed if I get problems like this when no one is around to ask. Is it just one of those “you need experience” sort of things? I would have gotten the error message and been completely stuck.

Is it because, as a general rule of thumb, we should be using the Pip which is in the same location as the Python we are using? I don’t have a technical justification for this (anyone?), but in my experience Pip and Python have to go hand in hand.

We are currently using the Python that came from the image of the machine (/opt/conda/bin/python) and we deleted all files related to Python from ~/conda/bin in walk thru 6 (16:48) in order to save disk space.

1 Like

Yes. The path gets searched sequentially, from left to right.

So, if your PATH variable looks like this:


And, let’s say you have python at version 2.7 in /usr/bin, and python at version 3.9 in /usr/local/bin

If you type “python” on command line, it (ie; bash) will look in /usr/bin first, and if it finds python there, it will fire it up. It won’t bother going past that point to look in /usr/local/bin (where it would’ve found python 3.9 had it kept searching) We’re assuming here that both binaries (aka executables) are named ‘python’.

If you do ‘which python’ , it will show you both versions (/usr/bin/python and /usr/local/bin/python) but you’ll need to look at the value of PATH variable (ie echo $PATH) to understand how the search is being conducted by the shell within the PATH variable (it is sequential , left to right)



@antoine That was my assumption as well. I assume that if we had not deleted python then it would have still worked

@mike.moloch That’s helpful, thanks. I guess the learning point is that if something is not working as expected then check to see which version you are using. That never would have occurred to me. I wonder if we would avoid this problem if we eliminated the step of installing mamba, which Jerermy deletes anyway. I may give that a try. A virtual environment may also be a good idea since it is self-contained, although it will use more memory in /storage.

Any tips on how to read that error message to figure out what is going on? I tried to read it as a clue, imagining that I was trying to solve this alone but I didn’t know what any of those variables were.

I am concerned about this highly customized version of Paperspace that we are using. Jerermy (and I assume many others) can troubleshoot problems very quickly, but I lack the knowledge and all of the changes are introducing a lot of bugs.

Hi @Mark_F , I share this concern with you. These sessions are highly concentrated dose of knowledge and cover lots of points and it can be a little bit intimidating for newcomers. I know a little bit about Unix/Linux concepts but I can see how a person new to this may feel a bit overwhelmed.

What I’m really hoping for is that at some point, some enterprising soul (not gonna name names but we know there are a few of those helpful individuals out there :smiley: ) will come up with a script which basically does what all these sessions do. So we open a paperspace machine and run something like … and voila! we have a working machine which sets up the things as per the recommendations in the walkthrus.


The flipside is that I’m learning a ton, which is great. These sessions are fantastic. But when the party is over and I’m trying to use this by myself, I’m not sure how I’m going to fix the glitches that arise.


Agreed! But one thing that I’ve learned from these sessions is to “not be afraid of breaking things”. Since these are essentially “virtual machines” , you can always wipe the slate clean and start afresh. And then there are forums and the discord where there is some activity even when there isn’t a course happening.

Having said all that, I think that Paperspace or Jarvislabs are both good options to get one started with the fastbook chapters. I have been able to get things running with the base machines on both platforms (even though I actually use my local machine for most stuff) and that is basically the whole point of these for-rent instance services. Jarvislabs hasn’t been talked about much, but I found it to be quite friction free.

Learning to deal with linux commandline, tmux, vim etc is just the cherries on the top. I would personally like the walk thrus to cover more of python/fastai type stuff – which, with session 8 we got into – since I’m somewhat familiar with the OS level stuff, and can muddle through on my own. But it’s definitely been instructive to learn that a lot of developers find these things (bash/vm/tmux/commandline etc.) to be the more challenging part of the endeavor.


Since these are essentially “virtual machines” , you can always wipe the slate clean and start afresh.

I’ve been thinking about this, anticipating a time when I will inevitably not be able to fix something. Anything in /storage will follow you forever. I was wondering how you’d start fresh. I guess you’d delete and delete all of the things you put in /storage, then create a new instance? If you have something in /storage that you unwittingly relink to, it will haunt you forever. I wonder if there is a way to factory reset storage.

The way I would really like it, is to have different config directories under storage and separate files (,,, and each points to a given /storage/configuration-type1 /storage/configuration-type2 etc.

So then, I can have which points to /storage/safe-config

Which we can run is controlled from the advanced tab before we fire-up the machine, so if I bork stuff up, I can just tell it to start with the safe config and then once I’m in, I can change things and then fire up another instance with the fixed version.

This is smart idea but I’m not understanding how it would work? The command in advanced options runs / / will point to your file, but I don’t see how you can choose which file it will boot.

I guess you could create your own /storage/ file and then specify that in Advanced Options/Container/Command. That is above my paygrade. Not sure if has to be in the root directory to work, although my simple-minded guess is yes.

Or are you describing something you’d like to see, i.e. that Paperspace should create an option to select from different files in advanced options?

Or are you describing something you’d like to see, i.e. that Paperspace should create an option to select from different files in advanced options?

Yes exactly! I meant more like a ‘feature request’ to paperspace. It definitely does not exist right now as you correctly pointed out. More of a blue sky / wishful thinking on my part.

Yes that’s exactly right.

The idea is to build up the skills here to debug exactly what isn’t working, and then focus on fixing that one thing. If you’re not sure, comment out all the lines in and confirm that a new instance works OK, and then add back half the lines and try again. Repeat that to bisect which thing in storage is causing the problem.

Then either try to debug that (with the help of us forum colleagues as needed) or just move that bit out of the way in /storage and re-create it.