Developer chat


After experimenting a bit, and going back and forth, we finally settled on adding a MAJ token: each word that begins with a capital is lower cased (as before) but we add xxmaj in front of it to tell the model. It appears to help a little bit.
There is a new pretrained model to match that change: you’ll find it in URLs.WT103_1
The text example notebook has been updated to use it (and went from 79% to 84.5% accuracy in the process!)

What is the difference between WT103 and WT103_1
(Fred Guth) #490

Sorry, I didn’t see the question in another topic before posting here.


A lot of stuff aimed at unifying the API accross applications just merged:

  • every type of items now has a reconstruct method that does the opposite of .data: taking the tensor data and creating the object back.
  • show_batch has been internally modified to actually grab a batch then showing it.
  • show_results now works across applications.
  • introducing data.export() that will save the internal information (classes, vocab in text, processors in tabular etc) need for inference in a file named ‘export.pkl’. You can then create an empty_data object by using DataBunch.load_empty(path) (where path points to where this ‘export.pkl’ file is). This also works across applications.

Breaking change:
As a result ImageDataBunch.single_from_classes has been removed as the previous method is more general.

(Piotr Czapla) #492

Awesome! Sylvain can you point me to the scripts you are using to create the pre-trained model, I’d like to see if I can get some improvements using BiLM training and qrnn.

(Stas Bekman) #494

A post was merged into an existing topic: Fastai v1 install issues thread

(Michael) #495

I wrote a small Tensorboard callback to visualize the metrics and the parameter/gradient distributions and histograms:

It is still a work in progress, because the code needs to polished and is only tested with the network in the notebook.

Could this be interesting for the library? If, how would I best incorporate the needed Logger class (with the Copyleft license)?

Feedback, suggestions, tips, and etc. are highly appreciated! :slight_smile:

PS: I don’t know if switching to TensorboardX would be a better choice. Maybe somebody worked already with TensorbordX and can share his experience?


I put the latest notebook I used to pretrain a QRNN here. Didn’t fully test the true_wd=False so you’ll have to add that.

(Sebastian Fleck) #497

I think the latest to_detach change broke the RNNCores forward method. I am getting a RuntimeError letting me know that input and hidden tensors are not on the same device. Since this wasn’t marked as a breaking change I guess it is a bug. How to proceed?


Will look into that later today. It’s definitely a bug!

(Stas Bekman) #499

Is there anyone here using fastai-v1 with macOS that can help us reproduce and debug fastai test suite failure on that system?(segfault in tests/

Most likely it’s related to this pytorch issue. And we would need to first reproduce this problem, and then reduce it to a simple test we could then file an issue with against pytorch.


(Fred Monroe) #500

I have OS X and can help

(Fred Monroe) #501

fwiw - i’m running python 3.7 and with the latest pull of fastai i’m not getting any failures in tests/

i was at first but once i deleted an old copy of mnist that was missing a test folder it worked fine

( osx 10.14.1 - no gpu )

(Stas Bekman) #502

Thank you for testing this Fred,

I’ve now updated our CI to run the correct up-to-date conda package on MacOS. They confusingly renamed pytorch-nightly-cpu to pytorch-nightly some weeks back. But this build works fine.

So it’s still something related to pypi build, and other then potential nuances in the 2 different package builds, the main difference is that conda and pypi install targets are on different drives it seems on the CI build. That’s why I thought it could be related to this pytorch issue . Is there a chance you could try and reproduce it so that the env and the data are on different mount points? basically moving the test suite to another /mnt/ point. See:

And for the sake of searchers the error is:

=================================== FAILURES ===================================
______________________ test_image_to_image_different_tfms ______________________

    def test_image_to_image_different_tfms():
        get_y_func = lambda o:o
        mnist = untar_data(URLs.COCO_TINY)
        x_tfms = get_transforms()
        y_tfms = [[t for t in x_tfms[0]], [t for t in x_tfms[1]]]
        data = (ImageItemList.from_folder(mnist)

>       x,y = data.one_batch()

_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
fastai/ in one_batch
    try:     x,y = next(iter(dl))
fastai/ in __iter__
    for b in self.dl:
/Users/vsts/hostedtoolcache/Python/3.6.5/x64/lib/python3.6/site-packages/torch/utils/data/ in __next__
    idx, batch = self._get_batch()
/Users/vsts/hostedtoolcache/Python/3.6.5/x64/lib/python3.6/site-packages/torch/utils/data/ in _get_batch
    return self.data_queue.get()
/Users/vsts/hostedtoolcache/Python/3.6.5/x64/lib/python3.6/multiprocessing/ in get
    res = self._recv_bytes()
/Users/vsts/hostedtoolcache/Python/3.6.5/x64/lib/python3.6/multiprocessing/ in recv_bytes
    buf = self._recv_bytes(maxlength)
/Users/vsts/hostedtoolcache/Python/3.6.5/x64/lib/python3.6/multiprocessing/ in _recv_bytes
    buf = self._recv(4)
/Users/vsts/hostedtoolcache/Python/3.6.5/x64/lib/python3.6/multiprocessing/ in _recv
    chunk = read(handle, remaining)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

signum = 20, frame = <frame object at 0x1050e8048>

    def handler(signum, frame):
        # This following call uses `waitid` with WNOHANG from C side. Therefore,
        # Python can still get and update the process status successfully.
>       _error_if_any_worker_fails()
E       RuntimeError: DataLoader worker (pid 1201) is killed by signal: Unknown signal: 0.

/Users/vsts/hostedtoolcache/Python/3.6.5/x64/lib/python3.6/site-packages/torch/utils/data/ RuntimeError
----------------------------- Captured stderr call -----------------------------
ERROR: Unexpected segmentation fault encountered in worker.

(Fred Monroe) #503

just to be clear, the bug you point to mentions /mnt which is a linux thing
(on osx there is /Volumes)

Am I still working on OSX or it is it ok to use different mnts on linux?

(Stas Bekman) #504

Oh, sorry, I don’t know osx, I assumed it’s the same as linux (mount-points-wise), but perhaps it’s not. I guess you need to go backwards from this solution, to reproduce the problem. Does it make sense?

I am not sure yet about testing it on linux - I will do that shortly myself. The CIs on linux and osx are configured identically, and only osx fails. But the original bug report is on linux, soI will certainly test that to rule it out.

(Fred Monroe) #505

ok i can do it on OSX - just wanted to make sure thats what you wanted.
i notice in that thread one person mentioned setting num_workers=0 to avoid masking what was actually wrong. I’ve experienced that myself on different occasions - if there is an assertion in a worker thread that is doing data loading or something it gets eaten/masked - might be worth trying in azure env

meantime i will split data and python env onto different mount points

(Fred Monroe) #506

ok i did the following:

i mounted (SSHFS) - a remote box so thats where the data/repo was - anaconda (python) env was on my local osx box - test_vision_data_block tests all passed 7/7

so i couldn’t reproduce on osx w/ remote volume

(Stas Bekman) #507

Thank you, Fred.

I did a cross-mount points test on linux, no problem there.

While I try to debug directly on the CI, which is very slow, since there is no shell access,
could you have a look at the setup and see if anything is different from yours?

Note that the conda setup of the same works just fine:

Something is different on the pypi setup that leads to this problem.

So far I have only identified as a potential difference that conda installs its stuff under /usr/local/miniconda/envs/fastai-cpu/, whereas pypi into /Users/vsts/… and the checkout goes into /Users/vsts/

Working (conda) env:

cwd: /Users/vsts/agent/2.142.1/work/1/s
=== Software ===
python version : 3.6.7
fastai version : 1.0.29.dev0
torch version  : 1.0.0.dev20181126
torch cuda ver
torch cuda is  : **Not available**

=== Hardware ===
No GPUs available

=== Environment ===
platform       : Darwin-17.7.0-x86_64-i386-64bit
conda env      : fastai-cpu
python         : /usr/local/miniconda/envs/fastai-cpu/bin/python
sys.path       :
no supported gpus found on this system

Failing (pypi) env:

=== Software ===
python version : 3.6.5
fastai version : 1.0.29.dev0
torch version  : 1.0.0.dev20181125
torch cuda ver
torch cuda is  : **Not available**

=== Hardware ===
No GPUs available

=== Environment ===
platform       : Darwin-17.7.0-x86_64-i386-64bit
conda env      : Unknown
python         : /Users/vsts/hostedtoolcache/Python/3.6.5/x64/python
sys.path       :
no supported gpus found on this system

(Fred Monroe) #508

I have several differences:

python 3.7
conda installs of pytorch etc
different torch version: 1.0.0.dev20181014
osx: darwin 18.2.0

i could try to recreate the pip install if that would help?

(Stas Bekman) #509

Yes, in general you need to switch to pytorch-nightly on macos/cpu, see this, since you’re using an outdated build. But I have already eliminated this as a potential culprit.

yes, this is where the culprit happens. thank you.