Hey guys, in chapter one of fastbook, there is this statement
In deep learning, it really helps if you have the motivation to fix your model to get it to do better. That’s when you start learning the relevant theory.
My problem is with where the motivation come from if your model is already state of the art, which seems easy to get with the tools and techniques that are being taught in this course
If p-value is not going to tell that there is enough data and the differences are big enough to do conclusions how we then know this? Like when we can say that we need more data and when we can say that we have big enough difference to do some conclusion? If I understood correctly some people say that if p-value is less than 0.05 the data is fine to do conclusions but as Jeremy and other said it’s not really true.
Did you make the COVID video part available? I need to convince more people to use masks!!
One thing to try is to apply it to your own data (if possible, a “real” problem). In that case, you will have to do some work to improve your performance, or at least to know that your performance is already the maximum you can get 
This figure was mentioned in the lecture. I went to the paper afterwards and discussed with a few colleagues but I cannot get a consolidated answer in the following:
My understanding: the grey squares are the scores of layer 2. This means after wx+b but probably before applying any activation function (i.e. if we had applied relu, the pictures would be mostly black with some white). Note: this assumption may not be correct.
My question is: what does each small square mean? Is each big square the same filter and the small squares the output of the filter for the different images on the right?
How was this selection on images done?
What is a general guideline for minimum image dataset size ? I have 5 categories and about 10 images per category so far. Not getting a useable model from that.
How much can data augmentation help ? If I do 3 transforms is it the equivalent of having 3x more images ?
The short answer is “until your model is getting good results”. My general rule of thumb is I try to find at least 30 images of each class if not 50 if possible. Data augmentation does help, but it’s not a simple 1:1. The more you transform your images the more different they look each time.
To add a great example from an earlier project: Cricket or baseball? Lesson 1 with small datasets
This got great results using just 15 images for training!
Hey didn’t Jeremy mentioned a book about deployment in the lecture? I think I remember something like that, and it was supposed to be added to the resources at the top. If someone knows what book he was talking about, could you add it up there? Thanks!
This seems to be quoted from Section 4.4.2 (p.132) of Neuromorphic Computing and Beyond by Khaled Salah Mohamed.
Would this be considered a standard definition?
I think it’s this book: http://shop.oreilly.com/product/0636920215912.do
This discussion reminds me of the motivation for dropout. Heuristically, dropout trains a large number so-called “thinned” networks and averages them at test time.
Message to beginners
Dropout is a technique introduced in Dropout: A Simple Way to Prevent Neural Networks from Overfitting. The first four (4) pages are very readable and are definitely worth a read even (and especially?) at the start of the course.
We’re a bit early to be looking at this - we’ll do it later in the course.
I think it helps to separate modelling and training.
A model, whether it is a good or bad, trained or not, will use its parameters to make predictions given an input.
Training a model (in order to make it better) is done by acting on its parameters. Training has its own set of parameters, and these are called hyper-parameters.
At the end of the lesson Jeremy shows how to export model and do inference with an image file. How can I do inference on a PIL image or np.array? I wish to make predictions on live video feed frames without saving the test images to individual files.
predict will run it based on all that is available, it’s not just for an input image, but whatever it can make an image out of. It’ll use what’s in the PILImage class header which if we see it’s it’s .create:
def create(cls, fn:(Path,str,Tensor,ndarray,bytes), **kwargs)->None:
"Open an `Image` from path `fn`"
if isinstance(fn,TensorImage): fn = fn.permute(1,2,0).type(torch.uint8)
if isinstance(fn,Tensor): fn = fn.numpy()
if isinstance(fn,ndarray): return cls(Image.fromarray(fn))
if isinstance(fn,bytes): fn = io.BytesIO(fn)
return cls(load_image(fn, **merge(cls._open_args, kwargs)))
It’ll accept bytes, a numpy array, a Pillow image, or a regular tensor
Awesome, thanks!
Does anybody have more information on seeme.ai ? It looks very interesting, but it also looks like they’re not active yet (the only option is to join a wait list). Is there an account for people in this class?
@giacomov Please check this thread linked in Top Wiki:
I followed the fastai2 deployment quick guide and I was able to create my account from the notebook (instead of doing it via the website).
Please check the Seeme.ai thread for discussions/Questions.
Ah, I see. I did see that notebook, but I thought you needed to have username and password already. This is great!