Share your work here ✅

Not sure, if still an issue for you, but as usual it goes without saying, update the fastai library whenever you git pull :slight_smile:

conda update conda
conda install -c fastai fastai

For example, in fastai --> 1.0.28-py_1 CollabDataBunch from Lesson4-collab works just fine.

Hi @pankymathur,
I did all of that before I started and still got the error. The only thing that changed since my log in was my update to Mojave. Thank you for your help!!!


1 Like

I developed and deployed the ResNet model for image image recognization from lesson 3. Check the zeit site:
I also summarized everything in a medium blog:

See what I got:

The accuracy is about 82%.


Thanks a lot for spotting. Will change. :smiley:

Introducing Terminator Vision!

Using ResNet-50, I trained an image classifier and deployed it on Zeit so Terminator can now SEE and detect its 4 greatest threats - Elon Musk, Jeff Bezos, Jon Snow, and John Connor! Humanity’s fate is in the hands of these leaders :man_superhero: :pray: :skull:

To help people visualize the architecture of a CNN and understand some of the “under the hood” math, you can check out these spreadsheets and blog post I wrote:

Elon in Excel (<–See Math Calcs of Forward Pass:)

Visual overview of CNN architecture:

Big shout out and special thanks to @navjots for helping me with the Zeit deployment. As a professional spreadsheet slinger (work in finance) who knows “just enough” Python to get by, I’m incredibly grateful to @jeremy /@rachel’s mission to make machine learning accessible to people of all backgrounds and this community’s willingness to support each other.

Next steps:

  • Clean Up + Share - Clean up notebook & share repo on Git Hub (my 1st ever! :crossed_fingers::scream:)
  • Fine Tune + Experiment - Update weight decay parameter, further learning rate experimentation, etc.
  • Prune image dataset & automate - Using the Fatkun batch image downloader, I used ~750 images for training (600 for training set, 150 for validation) and got to ~97% accuracy (4 losses below are reasonable…I should have removed the “actor” who played the child John Connor)



Hey everyone! I just completed my first blog post.

It’s a quick intro to classifying audio using image classifiers with the fastai framework. I have been trying to transfer some of my work from the freesound kaggle competition to the updated v1 library.

Over the past week or so I wrote an experimental audio module to read in raw audio and compute spectrograms on the fly during training (straight from the DataLoader), so you don’t need to actually create images beforehand. It then can take pretrained image models, modify them to take a single channel of 2D input, and finetune from there. I was able to run 3 epochs over 100k 4 second audio files from the NSynth dataset, classifying their instrument families to around 80% accuracy in under 3 minutes.

I’m planning to continue working on this for the next few weeks in hopes of supporting data augmentation and maybe even support for finetuning of models trained on audio data (as opposed to imagenet models).

It’s still a work in progress, and ALREADY a version behind fastai (currently tested with 1.0.28). You can view all the code and notebooks here:


why is test a category in your confusion matrix? :stuck_out_tongue:

haha, I added a “test” folder :frowning: I should remove it. :slight_smile:

Can I install docker windows into my windows home version? I am getting error that windows pro is must for this. Is this correct?


Maybe it would make sense to add a relu and an additional linear layer on top, so that instead of just a weighted combination of NLP and a tabular model activations you would also get their interactions?

Also you mention that the same learning rates are used for both models, which sounds like an important limitation. Maybe the learned bottom layers weights should be kept frozen and only the top layers trained?

1 Like

You wanna try Docker Toolbox for Windows (legacy version) instead of Docker for Windows which supports Windows pro and Windows Enterprise, Try this article.

This is looking amazing already - looking forward to the next stage! :slight_smile:


I love this! This is a great example of how to create your custom ItemBase and ItemList :slight_smile:

Just one remark to help you port your code to the latest version of fastai, the show_batch method in AudioItem should now be show_xys in AudioItemList (and you can code show_xyzs if you want show_results to work).

For all the ways you can customize your own ItemList, there is now a tutorial, just in case you didn’t see it before.


really nice. I guess baseline wander, noise and phaseshift could be augmentation candidates ?

Great work its too much interesting i am not able to reproduce your results. its giving me the following error. The web-page is not getting loaded.


That all seems valid and I don’t see an error there. Can you check the console on the browser? Also, are you serving via HTTPS, most browsers require this to access the camera.

According to your code if its non HTTPS then the code is forcing to serve in https. issue is when i run python and try to acess its get loaded for a while without webcam on and suddenly the web-page dies.

Which browser/os? Any errors showing up in the browser? Please feel free to message me directly so we don’t spam this thread.

Please Check your inbox lets talk privately for the ease of others.

Hello All,

I am working on a kaggle project and have used Resnet101 model weights and have further finetuned weights. But i am not able increase the accuracy beyond 87%.

Can anyone please guide me how i can improve accuracy further? Github link for the notebook is

epoch train_loss valid_loss accuracy
1 0.904559 0.602475 0.826715
2 0.886423 0.490211 0.830325
3 0.717062 0.447640 0.814079
4 0.600986 0.386214 0.877256
5 0.514590 0.415294 0.873646