SOURCE CODE: Mid-Level API

[akashpalrecha]

Yes we did meet. It was as per the schedule that aman put on the thread. Sanyam will soon share a recording.

[florianl (Florian)]

Thanks! adding the type annotation fixed the issue. Why are the transforms also applied on the CategoryBlock? Are all transforms applied on all Blocks? Lets say I have the following data loader:

ImageBlock, ImageBlock, CategoryBlock

Will all transforms be applied to all three blocks or is there a way to specify the transforms for each block?

[akashpalrecha]

So, each Block has: type_tfms, item_tfms, and batch_tfms.
type_tfms are applied only to items of that block. These are generally transforms that create the object. For example, transforming a path into an actual image.
item_tfms and batch_tfms of all the blocks are concatenated into one single list and applied to all the items that come from each block (after having their type_tfms applied to them). item_tfms are applied just before a batch is formed and likewise batch_tfms are applied after a batch is formed.

Now, because of the way a Transforms works in FastAI, this happens:

1. If you have an encodes function with no input type annotation, it will be applied to any object it receives and will throw errors when that fails due to incompatible object types.
2. If you have an encodes function with an input type annotation, it will work normally on inputs of that type, and for every input it receives which does not match the annotation, it will simply return it / pass it on as it is.

So, the idea is to have a single pipeline of transforms through which all your datablock items go through, and each transform, due to it’s type annotations, knows exactly what items to modify and what items to let through unchanged.

You can watch Jeremy’s code walkthroughs to better understand this. The first few walkthroughs specifically deal witht this. Also, we discuss all of this in complete detail in our code walkthroughs of the datablock API (links in the first post of the thread). I recommend you to go through either of those.

Hope that helps!

[akashpalrecha]

REMINDER:
At 7:30 AM, IST, 4th April, we’re going to have a very long zoom call to discuss the internals of things like the Learner, cnn_learner, etc.
Everyone is invited to join.

What you should be comfortable with already: The Datablocks API

Although we prefere to not have any prerequisites for these walkthroughs, having unfamiliarity with other parts of the library will make the whole process very slow and unwieldy for everyone. You can watch the appropriate code walkthrough videos posted in the first post in this category to gain familiarity!

[arora_aman (Aman Arora)]

We are starting the Study Group now.

[DrHB (Asimo)]

@akashpalrecha

7:30 AM

Ohhhh I just missed it … I thought it was PM … Will you guys upload the recording ?

[akashpalrecha]

So we’re going to have another session at 7:30 AM IST on April 5 too. This one will be about the Optimizers and the Captum callback.

Today’s session wasn’t very productive actually. We’ll tell you what we did in the next call?

[DrHB (Asimo)]

perfect! Seee you soon

[init_27 (Sanyam Bhutani)]

edit: Removed links due to a bug in zoom that keeps recording anytime someone joins
Zoom Link for today (Starts at Sunday, 7:30 AM IST()
Meeting Password: Removed

Edit: Please @ me if you have issues with joining.

[averma (Abhinav Verma)]

Count me in as well. I’ve been obsessed about going into the internal details of the fastai library. I’ve just started watching your videos. I’m on the first one . When do you have your weekly meetups. I would love to join. @arora_aman

[arora_aman (Aman Arora)]

Fastbook Study Group: Saturdays & Sundays 7:30 AM IST

I’ve setup a zoom call to join the SG but I’ve decided not to set it to record by default. We’re discussing the materials for the book and this falls under a gray area of re-distributing if we record the sessions, which is a violation of the license for the materials. I suggest discussing it actively and then posting the summaries or imp discussion takeaways here and then the few of us who’ve read the chapters can then help others if they have any questions. Although IMO the book has very cle…

[AjayStark (Ajaykumaar)]

Hii @init_27 when will the link to the recording be posted?

[dhoa (Dien-Hoa)]

I’m reading the Siamese Tutorial and found the piece of code below:

class SiameseImage(Tuple):
    def show(self, ctx=None, **kwargs): 
        img1,img2,same_breed = self
        if not isinstance(img1, Tensor):
            if img2.size != img1.size: img2 = img2.resize(img1.size)
            t1,t2 = tensor(img1),tensor(img2)
            t1,t2 = t1.permute(2,0,1),t2.permute(2,0,1)
        else: t1,t2 = img1,img2
        line = t1.new_zeros(t1.shape[0], t1.shape[1], 10)
        return show_image(torch.cat([t1,line,t2], dim=2), title=same_breed, ctx=ctx, **kwargs)

I’m curious why we need to return a show_image here that I think even without it, the picture is always shown. Maybe we need to conserve the ax for another purpose like show_batch or sth. Thanks

[nareshr8 (Naresh)]

One reason is that your approach will work only on Jupiter notebook and not in other traditional IDEs I guess. And the title and other fields are ignored if we just return the concatenated tensor…

[dhoa (Dien-Hoa)]

In DataBlock why we don’t put the args get_items, splitter, get_y, get_x in init .

I mean when I Shift-Tab to see the signature of DataBlock I can’t find the args above even though they seems mandatory for me.

I then looked at the source code and find with @funcs_kwargs decorator, all the _method attributes will be set with value from **kwargs. And _methods = ‘get_items splitter get_y get_x’.split()

So now I understand how it works but why it is made so complex ? What is the intent behind ? Thanks

[nareshr8 (Naresh)]

These are added as **kwargs. You might be interested in this blog post from Jeremy

[dhoa (Dien-Hoa)]

Thanks @nareshr8. But It might different of what I mean. Sorry if my question is not clear.

The piece of code I’m asking is below:

@funcs_kwargs
class DataBlock():
    "Generic container to quickly build `Datasets` and `DataLoaders`"
    get_x=get_items=splitter=get_y = None
    blocks,dl_type = (TransformBlock,TransformBlock),TfmdDL
    _methods = 'get_items splitter get_y get_x'.split()
    _msg = "If you wanted to compose several transforms in your getter don't forget to wrap them in a `Pipeline`."
    def __init__(self, blocks=None, dl_type=None, getters=None, n_inp=None, item_tfms=None, batch_tfms=None, **kwargs):

It’s decorated by @funcs_kwargs and not @delegates. @funcs_kwargs will set the cls attributes which are defined in _methods (‘get_items splitter get_y get_x’.split()) by the value in **kwargs .

By this way we can set the attributes of get_x,get_items,splitter,get_y

However, I have to look at the source code or an example to see how to initialize the DataBlock (because it is not autocomplete all the mandatory args)

Why don’t we just put

def init(self, blocks=None, dl_type=None, getters=None, n_inp=None, item_tfms=None, batch_tfms=None, get_x=None,get_items=None,splitter=None,get_y=None, **kwargs)

What is the intent behind the @funcs_kwargs decorators ?

Thanks

[nareshr8 (Naresh)]

It’s just ease of writing. You may have to declare those arguments in init and you may have to do self.{}={}… func_kwargs reduces these 2 steps to one. You just declare the _methods which contains list of things. This will be added to init function and a instance of the method is saved in the appropriate variable for later use.

[dhoa (Dien-Hoa)]

Great, it can be a very good reason. Thanks a lot. (However we have a drawback about autocomplete)

[Albertotono]

Are you guys meeting?