I have found this site incredibly helpful for learning and developing regex. https://regexr.com/
You can see the results of your expression given your input text immediately which I found very helpful for learning. It also has a built in cheatsheet and reference guide.
FASTAI NLP regex https://github.com/fastai/course-nlp/blob/master/4-regex.ipynb
Course videos: https://www.youtube.com/watch?v=cce8ntxP_XI&list=PLtmWHNX-gukKocXQOkQjuVxglSDYWsSh9
I like this one for testing regex expressions: https://regex101.com/
Regular expressions are amazing, powerful, and sometimes can be taken too far. Just be careful.
They are very difficult to understand and debug when they get super long/complex.
Regex Golf, which is really fun way to learn: https://alf.nu/RegexGolf
Are the blocks in the DataBlock an ordered list? Do they specify the input and output structures respectively?
Are there always 2 blocks, can there be more than 2?
E.g. if you wanted a segmentation model, would the second block be something about segmentation?
The blocks specify inputs and outputs, yes. You can have more than 2, in which case you should pass n_inp
to specify how many inputs/targets you have.
And yes, for a segmentation task, the second block would be a MaskBlock
.
Will default batch transforms broadcast the manipulation to higher order tensors i.e. a tensor with dimensions like (Channel, Time, Height, Width)
May you show what is an interpolation? and why do we have those results
unable to hear
@rachel Jeremy’s mic is off
When augmenting isn’t it better to introduce noise ? Like image scaling artifacts ?
If the transform is more performant on GPU it is understandable but why is preserving image quality important for training set?
there we go… now we can hear Jeremy!!
How does the item_tfms (presize) work if the Resize() is smaller than the image? Is a whole width or height still taken? Or is it just a random crop with the Resize(value)?
Yeah, I think it will be covered later. Just wait.
You want your training images to be relatively representative of the images you are performing inference on, so arbitrarily losing image quality may be detrimental.
May it help you:
DIGITAL IMAGE INTERPOLATION
Could I get some clarification on why we want to build a model before cleaning the data? I would think a clean dataset would help in training
Is there a general rule of thumb to remember when resizing in the first step of presizing?
They use the model to help with data cleansing, by identifying and relabelling/discarding bad data.