Non-Beginner Discussion Recordings and Organization - Post Them Here!

As we’re going through our Zoom meetings, one trend I notice is people tend to be discussing awesome ideas and concepts, and I wanted to attempt to collate them all together for us! If you want to record such discussions, I’d recommend using OBS and posting to YouTube. When uploading ensure that your video is set to unlisted until the course is live for the rest of the world :slight_smile:


I am making this a wiki post, but keep it to advanced discussions about concepts not discussed in the course well, and if it’s multiple recordings of one session, try to just use up one line (see the first video below for an example). For each post, please provide a 1-2 sentence summary of what the video is describing or going over. Also, if it was discussed in a study group, please keep the recording in that particular study group’s thread, do not cross-post to here. This is meant for one-off discussions that don’t quite fit into an easy category. Also, if the discussion originated from a particular non-study group thread, please post the thread so questions can be addressed there (if not feel free to discuss here)

When Starting or Planning A Walkthrough

  • If you have a set time or topic or go live on the Zoom call please post here a few minutes in advanced so people can see that you are planning a walk through
  • For those only interested in joining/listening: Please set this post to “Watching” so you will get notified.


  • Unknown Labels (have your model say “I don’t know”), Image Regression (non-points), Loss Functions part 1 part 2 related thread

  • Speeding up fastai2 inference with torch.jit by @DrHB YouTube

  • Decodes and its relation to the loss function YouTube


Thank for the video =)

just to add some more info on tracing, sometimes if you want to trace Efnet or Mixnet you will get following error:

In order to overcome this error we need helper:

class trace_Swish(nn.Module):
    def __init__(self):
    def forward(self, x):
        return x.mul(torch.sigmoid(x))
# Utility to replace all mish_cuda with trace_mish
def to_traceswish(model):
    for child_name, child in model.named_children():
        if isinstance(child, Swish):
            setattr(model, child_name, trace_Swish())

and after this we can just trace it normally

model = timm.create_model('mixnet_xl', pretrained=True)
dummy_input = torch.randn([2, 3, 136, 237]).cuda(), dummy_input), '')

just a quick request!
can someone record a walkthru of the notebooks in fastai/fastcore?
it would be really helpful for the beginners like myself :smile: