Fastai v0.7 install issues thread

@stas, thanks for the hint the pytorch version was correct but
“** torchtext**” version was not compatible it has to be “torchtext==0.2.3.” for pytorch 0.3.1.

1 Like

I’m on 18.04, freshly installed.

Installed fastai v1 first, in a dedicated env (it seems to work). Then I installed fastai 0.7 following the procedure described (which is the usual procedure).

As I try to import fastai modules, it complains what follows:

It starts from:


and ends with:


The symlink seems to be OK…

Screenshot%20at%202018-10-12%2022-36-01


So, what’s wrong? :face_with_raised_eyebrow:

OK, solved.

It was torchtext. Downgrade it to the previous version with pip.

Source (for example): Google Colab-AttributeError: module 'torch' has no attribute 'float32'

2 Likes

FYI, environment*.yml were just updated to pin torchtext=0.2.3. Thank you all for the input and testing.

1 Like

The symlinks look all right to me. Not sure what’s causing this.

I am using clouderizer.com Fast.ai project template for the course and to run the template I use google colab like a backend and I am not able to import the library properly. I have already posted my query here and the solution I received was for the case when we run the complete notebook on google colab itself or I may not be able to understand it because these things are new to me. If someone in the forum is using clouderizer and has found a solution, please help!.

Thanks in advance.

Edit: Finally found a solution. Click on edit project, go to the setup section, in the setup script column, add pip install torchtext==0.2.3. Now I am able to run the notebooks without any problem.

I was getting error related to torchtext=0.2.3.
I replaced it with torchtext==0.2.3 in the enviornment.yml file and it worked

will that work with !pip install fastai==0.7.0 in kaggle kernels?

no, the pip package is from May-2018. you need to use the source:

git clone https://github.com/fastai/fastai.git
cd fastai
conda env create -f environment.yml

(or environment-cpu.yml for cpu)

or just install it from the nb:

!pip install fastai==0.7.0
!pip install torchtext==0.2.3
1 Like

I don’t think you can do that in kaggle kernels AFAIK.

thanks stas!
tried that in kaggle kernels but got error: AttributeError: ‘torch.dtype’ object has no attribute ‘type’
but also seems somewhat random, i managed to get it work once, then it showed error, some other people also reported errors.

ML lesson 1 worked with !pip install fastai==0.7.0, however DL v2 lessons are not running on kernels reliably. the more i look forward to v3 course. noticed v3 notebooks started to appear on github repo!

I noted that the setup is slightly different from what Jeremy posted in Wiki thread: lesson 1 for the Machine Learning MOOC:

Would you mind explain a bit about the difference between these two ways of setting up fast.ai v0.7?

Error:
AttributeError: module ‘torch’ has no attribute ‘float32’.

Generated by:
from fastai.transforms import *

Caused by:
A newer version of torchtext (>0.3.0) in the requirments.txt (>=0.2.3)

Resolved by:
!pip install fastai==0.7.0
!pip install torchtext==0.2.3

Order matters

2 Likes

Thank you for the PR, @Cesare.montresor
It looks like everybody has been using environment*yml files, so requirements.txt got out of sync. I synced it to match environment.yaml.

Thanks, really fast!
I use it on Colab for giving classes with my 2 study groups.
As it requires to reinstall dependencies every time, using pip is somehow faster :slight_smile:

1 Like

Wiki mode has been turned on for the first post so you can edit it to improve it. Thank you.

hey, after i used the 2 lines which you specified, I got a lot of lines of the form: Requirement Already Satisfied…
After that there was the following code snippet:

mxnet 1.3.0.post0 has requirement numpy<1.15.0,>=1.8.2, but you'll have numpy 1.15.2 which is incompatible.

kmeans-smote 0.1.0 has requirement imbalanced-learn<0.4,>=0.3.1, but you'll have imbalanced-learn 0.5.0.dev0 which is incompatible.

kmeans-smote 0.1.0 has requirement numpy<1.15,>=1.13, but you'll have numpy 1.15.2 which is incompatible.

anaconda-client 1.7.2 has requirement python-dateutil>=2.6.1, but you'll have python-dateutil 2.6.0 which is incompatible.

imbalanced-learn 0.5.0.dev0 has requirement scikit-learn>=0.20, but you'll have scikit-learn 0.19.1 which is incompatible.

Installing collected packages: torch
  Found existing installation: torch 0.4.1
    Uninstalling torch-0.4.1:
      Successfully uninstalled torch-0.4.1
Successfully installed torch-0.3.1
You are using pip version 10.0.1, however version 18.1 is available.
You should consider upgrading via the 'pip install --upgrade pip' command.
/bin/sh: 1: cannot open =0.3.0: No such file

But then I executed all the other cells,it worked fine. So do I need to upgrade pip or any other package? Also, I didn’t understand that last line (cannot open =0.3.0…)
Also, why does it ask to update pip? Doesn’t the kernel have the latest version by default?

Apart from this I encountered 2 other warning messages regarding deprecated module/function in pandas,numpy. I had forked @miwojc’s kernel : https://www.kaggle.com/miwojc/fast-ai-machine-learning-lesson-1/

Because of > or < you must quote the string in bash:

!pip install "torchtext<=0.3.0"

Otherwise:

$ pip install torchtext<=0.3.0
bash: =0.3.0: No such file or directory

tries to read from file “=0.3.0”

1 Like

Ho sorry, it was with a \, but it got escaped by the forum, I doubled it :slight_smile: or yes, with quotes.

Even better:
!pip install torchtext==0.2.3

2 Likes