Fast AI pulling a fast one?

Question

Normally statements like from module import * are frowned upon by expert python programmers as they can lead to namespace clobbering. Yet they are frequent in Fast AI and the justification is that it makes life simpler for the student. Below is an excerpt from their book

This may be so and as long as it is simply a matter of importing everything that's fine I guess.

However, below we will see that it does more than a simple import and an instance of a cnn_learner that has no method called fine_tune ends up having one when we run from fastai.vision.all import *.

from torch import tensor
import numpy as np

from fastai.data.block import DataBlock, CategoryBlock
from fastai.vision.learner import cnn_learner

from torchvision.models.resnet import resnet18

images=tensor(np.random.normal(size=(32,128,128,3)))
labels=tensor(np.random.randint(0,3,32))

data=[{'x':x,'y':y } for (x,y) in zip(images,labels)]

db=DataBlock(blocks=(DataBlock, CategoryBlock),get_x=lambda x:x['x'],get_y=lambda x:x['y'])

learner=cnn_learner(db.dataloaders(data),resnet18)

print(hasattr(learner,'fine_tune'))

from fastai.vision.all import *

print(hasattr(learner,'fine_tune'))

The first print statement prints False and the second one prints True showing that the import statement from fastai.vision.all import * created a bound method for an instance of cnn_learner.

I am yet to figure out how it happens and I am sure given enough time I will figure it out. What I am equally sure of not being able to figure out is why Fast AI code is written in a way so as to necessitate such global imports to do simple training.

QUESTION: Having said all this, what I want to know is what is the way to get the cnn_learner to get into the correct state (with all the requisite methods) without doing the the global import like Fast AI folks do?

Answer 1

One alternate approach to from fastai.vision.all import * is

In [2]: import fastai.vision.all as fai

In [3]: print(hasattr(learner,'fine_tune'))
True

A second way is to import fine tune:

In [2]: from fastai.callback.schedule import fine_tune

In [3]: print(hasattr(learner,'fine_tune'))
True

I believe either of these methods work because the fine_tune method is monkey patched on to the Learner object when it is imported. It appears that this is a design pattern that the authors prefered to more standard ways of constructing python libraries.