Add WildFire Dataset object and split strategies

[x0s]

Hello,

This PR aims at:

• Add WildFireDataset pytorch compatible class
• Add Strategy to split the dataset in train/val/test without mixing the fire_ids

Example:

from torch.utils.data import DataLoader
from torchvision.transforms import transforms
from pyronear.datasets.wildfire import (WildFireDataset,
                                        WildFireSplitter)

wildfire = WildFireDataset(metadata='wildfire.csv',
                           target_names=['fire', 'clf_confidence'],
                           path_to_frames=path_to_frames)

ratios = {'train': 0.7, 'val': 0.15, 'test':0.15}
transforms = {'train': transforms.RandomCrop(10), 'val': None, 'test': None}

splitter = WildFireSplitter(ratios, transforms)
splitter.fit(wildfire)

wildfire_loader_train = DataLoader(splitter.train, batch_size=64, shuffle=True)
wildfire_loader_val = DataLoader(splitter.val, batch_size=64, shuffle=True)
wildfire_loader_test = DataLoader(splitter.test, batch_size=64, shuffle=True)

Each dataloader will yield the image(transformed if requested) and the two targets (fire and clf_confidence)

For more example, please browse the tests

Any feedback is welcome

[codecov]

Codecov Report

Merging #47 into master will increase coverage by 2.41%.
The diff coverage is 98.03%.

@@            Coverage Diff             @@
##           master      #47      +/-   ##
==========================================
+ Coverage   81.39%   83.80%   +2.41%     
==========================================
  Files          16       18       +2     
  Lines         602      704     +102     
==========================================
+ Hits          490      590     +100     
- Misses        112      114       +2     

Impacted Files	Coverage Δ
pyronear/datasets/wildfire/split_strategy.py	96.77% <96.77%> (ø)
pyronear/datasets/wildfire/wildfire.py	98.36% <98.36%> (ø)
pyronear/datasets/utils.py	91.66% <100.00%> (+1.04%)	⬆️
pyronear/datasets/wildfire/__init__.py	100.00% <100.00%> (ø)

[MateoLostanlen]

You made a mistake in your example, it is ->
from torch.utils.data import DataLoader
from pyronear.datasets.wildfire import (WildFireDataset,
WildFireSplitter)

wildfire = WildFireDataset(metadata='wildfire.csv',
path_to_frames=path_to_frames)

ratios = {'train': 0.7, 'val': 0.15, 'test':0.15}

splitter = WildFireSplitter(ratios)
splitter.fit(wildfire)

wildfire_loader_train = DataLoader(splitter.train, batch_size=64, shuffle=True)
wildfire_loader_val = DataLoader(splitter.val, batch_size=64, shuffle=True)
wildfire_loader_test = DataLoader(splitter.test, batch_size=64, shuffle=True)

[MateoLostanlen]

Maybe my question is dumb, but I don't understand why you use "from skimage import io" and not pil to load the images. I have a problem to apply the torch vision transform afterwards (https://pytorch.org/docs/stable/torchvision/transforms.html). What do you use as a transform then ?

[x0s]

You made a mistake in your example, it is ->
from torch.utils.data import DataLoader
from pyronear.datasets.wildfire import (WildFireDataset,
WildFireSplitter)

wildfire = WildFireDataset(metadata='wildfire.csv',
path_to_frames=path_to_frames)

ratios = {'train': 0.7, 'val': 0.15, 'test':0.15}

splitter = WildFireSplitter(ratios)
splitter.fit(wildfire)

wildfire_loader_train = DataLoader(splitter.train, batch_size=64, shuffle=True)
wildfire_loader_val = DataLoader(splitter.val, batch_size=64, shuffle=True)
wildfire_loader_test = DataLoader(splitter.test, batch_size=64, shuffle=True)

Thanks for your feedback!
Indeed, I forget to update the description.
So, If I understand it right, only these two lines needs to be updated:

splitter = WildFireSplitter(ratios)
splitter.fit(wildfire)