I might not have used the correct vocabulary here, but If you wanted to recognize multiple things per a given image (ex. ‘grizzly’ AND ‘grass’’), would you have a layer for each one ([1] grizzly, black, teddy [2] grass, tree) and do two “passes”, or is there a way match multiple things with one “pass”?
2 Likes
how do you retrieve a gcp instance password?
I’m trying to update my fastai library and it’s prompting me for the password for jupyter (the instance id)
Question about this code example:
classes = ['black', 'grizzly', 'teddys']
data2 = ImageDataBunch.single_from_classes(path, classes, tfms=get_transforms(), size=224).normalize(imagenet_stats)
learn = create_cnn(data2, models.resnet34)
learn.load('stage-2')
This requires models.resnet34 which I find surprising: I had assumed that the model created by .save(...) (which is about 85MB on disk) would be able to run without also needing a copy of resnet34.
10 Likes
Is anyone facing ssh issues?
1 Like
I think Jeremy mentioned last week that just duplicating the data was the best mitigation at the moment.
can you explain?
Wow, I feel like I can create a start-up now. I can train and predict images to world-class accuracy.
2 Likes
I am guessing, but it might be because a plain power of ten is exactly at the threshold that crosses over to that power. For instance, 1e4 is 10,000, which is “just barely” in the 4th power of ten.
What’s exactly in the middle of fourth power of ten? 50,000.
But what’s exactly in the middle of fourth power of ten on a logarithmic scale? Something like Sqrt(10) * 1e4.
3 is an easy approximation for the square root of 10, so that gives you numbers like 3e4.
I could be totally wrong! Just a guess. I have the same question.
2 Likes
How do we know that some or all of our images have not been used to train resnet and hence we’re doing a look up rather than actually “learning”
i could be wrong so someone please correct me if so, but loading models.resnet34 is really referring to the architecture, and you are overriding the weights associated with that architecture with the saved weights from ‘stage-2’
The model was indeed saved to disk by .save().AFAIK, the resnet34 initial download is only for pretrained model arch. Can anyone clarify?
1 Like
I am able to use the internet but I am not able to ssh to my aws ec2 instance. Was able to do it from my house.
Thank you! I understood now!!
Yes, someone above mentioned using google (colab I think) as a workaround for now.
1 Like
anyone else have a problem with the js window.open, not opening a new window in chrome?
Well, let’s hope I am right. 
Learning rate graph has learning rate in powers of 10 but why are we using it as exponential power e when we go for training our model
I thought you can do inference by just using original model and stop updating weights. why different API is used here?
e represents 10, and not 2.37
1 Like
Try disabling adblocker and run it again. That worked for me.
1 Like