In the pretrained conv model, how many layers are added at the end and how can we change that?
1 Like
To draw this graph, a training is launched with a very low learning rate that improves at every batch.
What happend to bn_unfreeze()? Still needed?
Should we always use as large batch size as possible on our card? Or is there a limit when model performance decreases?
4 Likes
Yup, intuition behind this is the fact that we are taking a correlation between neighbouring pixels only. Like, 3x3 or 5x5 filters, even though we are passing this through the whole image, it doesn’t preserve the orientational and relative spatial relationships between these complex features (eyes, nose, face boundary). Capsule networks address this issue.
It is the prediction of the model. You shouldn’t interpret it as a real probability 
2 Likes
yes, use as big a bs as you can without running out of gpu ram
Nope.
This is false, it does preserve the locations in the activation maps
1 Like
With regard to the awesome Fast.ai alumni building cool products with DL, is it okay to use transfer learning of an ImageNet model or public research datasets to build models for commercial products?
4 Likes
Why does the Convlearner use a size of 224 for Resnet 34 and 299 for Resnet 50?
4 Likes
so I’ve been running the notebook alongside and i’ve notices that my error is a bit higher (~0.07ish).
Is this difference the ‘resiliance’ that was mentioned earlier? Or is it a symptom of something else?
So the last number is the prediction of the model for the actual class, is capped at 1.00 but isn’t a probability ? So then how the model choose another class ?
1 Like
Validation loss seemed to be higher than the training loss. Does that not mean we were overfitting a bit?
1 Like
Is train error and validation error shown is in percentage form or we need to multiply it by 100 like we have been doing for error rate like 0.044 error rate is 4.4% of error rate
Yes, thanks for highlighting it
Train loss is 0.09 and validation loss is 0.13. Doesn’t this mean, overfitting is occuring?
2 Likes
Imagine trying to walk to the highest point in a landscape. The learning rate is kind of like how big your stride is. If your stride is too big, you’ll get to the top of the mountain faster, but it will be hard to pinpoint the very highest point once you’re close because you’ll keep stepping past it.
5 Likes