Intro

The fastai library simplifies training fast and accurate neural nets using modern best practices. See the fastai website to get started. The library is based on research into deep learning best practices undertaken at fast.ai, and includes “out of the box” support for vision, text, tabular, and collab (collaborative filtering) models.

MNIST data

Download and prepare data:

Transformations:

# transformations
tfms = aug_transforms(do_flip = FALSE)
path = 'mnist_sample'
bs = 20

#load into memory
data = ImageDataLoaders_from_folder(path, batch_tfms = tfms, size = 26, bs = bs)

learn = cnn_learner(data, resnet18(), metrics = accuracy)

TerminateOnNaNCallback

Cbs argument means callbacks:

EarlyStoppingCallback

learn %>% fit_one_cycle(10, cbs = EarlyStoppingCallback(monitor='valid_loss', patience = 1))
epoch     train_loss  valid_loss  accuracy  time
0         0.023524    0.009781    0.996565  00:16
1         0.033328    0.019839    0.993621  00:16
No improvement since epoch 0: early stopping

Save best model for each epoch:

learn = cnn_learner(data, resnet18(), metrics = accuracy, path = getwd())

learn %>% fit_one_cycle(3, cbs = SaveModelCallback(every_epoch = TRUE,  fname = 'model'))

See folder:

list.files('models')
# [1] "model_0.pth" "model_1.pth" "model_2.pth"
# [1] "model_0.pth" "model_1.pth" "model_2.pth"

ReduceLROnPlateau

Decrease learning rate if loss is not improved:

learn %>% fit_one_cycle(10, 1e-2, cbs = ReduceLROnPlateau(monitor='valid_loss', patience = 1))
epoch     train_loss  valid_loss  accuracy  time
0         0.117138    0.038180    0.987242  00:17
1         0.140064    0.006160    0.996565  00:16
2         0.133680    0.061945    0.985770  00:16
Epoch 2: reducing lr to 0.0009891441414237997
3         0.049780    0.005699    0.998037  00:16
4         0.040660    0.019514    0.994112  00:16
Epoch 4: reducing lr to 0.0007502954607977343
5         0.027146    0.009783    0.997056  00:16
Epoch 5: reducing lr to 0.0005526052040192481
6         0.024709    0.008050    0.998528  00:16
Epoch 6: reducing lr to 0.0003458198506447947
7         0.016352    0.010778    0.998037  00:16
Epoch 7: reducing lr to 0.0001656946233635187
8         0.071180    0.009519    0.998528  00:16
Epoch 8: reducing lr to 4.337456332530222e-05
9         0.014804    0.005769    0.998528  00:16
Epoch 9: reducing lr to 1.0114427793916913e-08

Or add new parameter min_lr:

learn %>% fit_one_cycle(10, 1e-2, cbs = ReduceLROnPlateau(monitor='valid_loss',
min_delta=0.1, patience = 1, min_lr = 1e-8))

CSVLogger

Save train history. In addition, for multiple callbacks it is important to pass them within list:

learn = cnn_learner(data, resnet18(), metrics = accuracy, path = getwd())

learn %>% fit_one_cycle(2, cbs = list(CSVLogger(),
                                      ReduceLROnPlateau(monitor='valid_loss',
                                      min_delta=0.1, patience = 1, min_lr = 1e-8)))
history  = read.csv('history.csv')
history
epoch train_loss valid_loss  accuracy  time
1     0 0.15677054 0.09788394 0.9646713 00:17
2     1 0.08268011 0.05654754 0.9803729 00:17