MNIST in FastAI

MNIST Project in FastAI

Here, we will predict MNIST digits using FastAI. Then, we will go deeper and figure out what’s happening behind the scenes.

from fastai.vision.all import *
from fastbook import *

path = untar_data(URLs.MNIST)
Path.BASE_PATH = path
path.ls()

100.03% [15687680/15683414 00:00<00:00]

(#2) [Path('training'),Path('testing')]

(path/'training').ls()

(#10) [Path('training/7'),Path('training/8'),Path('training/5'),Path('training/6'),Path('training/9'),Path('training/3'),Path('training/2'),Path('training/0'),Path('training/4'),Path('training/1')]

Using fastai, it is very easy to train. We make a DataBlock with information on how to get labels, grab images, and split train/valid data.

dblock = DataBlock(
    blocks=(ImageBlock, CategoryBlock),
    get_items=get_image_files,
    get_y=parent_label,
    splitter=RandomSplitter(),
)
dls = dblock.dataloaders(path)

Then, we take a look at it, and it looks like labels are correctly assigned to each image.

dls.show_batch()

We create a learner with resnet18, so we can use transfer learning. Then, we find the learning rate.

learn = vision_learner(dls, arch=resnet18, metrics=accuracy)
learn.lr_find()

/usr/local/lib/python3.9/dist-packages/torchvision/models/_utils.py:208: UserWarning: The parameter 'pretrained' is deprecated since 0.13 and will be removed in 0.15, please use 'weights' instead.
  warnings.warn(
/usr/local/lib/python3.9/dist-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and will be removed in 0.15. The current behavior is equivalent to passing `weights=ResNet18_Weights.IMAGENET1K_V1`. You can also use `weights=ResNet18_Weights.DEFAULT` to get the most up-to-date weights.
  warnings.warn(msg)

SuggestedLRs(valley=0.0010000000474974513)

Although fastai suggests a learning rate suitable to use, I can use a little bit bigger learning rate. I can probably use 0.01 even if 0.001 is suggested.

learn.fine_tune(3, 0.01)

epoch	train_loss	valid_loss	accuracy	time
0	0.383465	0.211662	0.934286	00:46

epoch	train_loss	valid_loss	accuracy	time
0	0.134975	0.076517	0.978714	00:59
1	0.030738	0.033497	0.991071	00:59
2	0.019911	0.024113	0.993071	00:59

Training was very easy with almost perfect accuracy thanks to fastai. However, it is fun to dive deeper and figure out what’s going on under the hood.