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.
Original demo by Zachary
Grab data:
Read json
file and get annotations:
Prepare laoder object:
get_y = list(function(o) img2bbox[[o$name]][[1]],
function(o) as.list(img2bbox[[o$name]][[2]]))
coco = DataBlock(blocks = list(ImageBlock(), BBoxBlock(), BBoxLblBlock()),
get_items = get_image_files(),
splitter = RandomSplitter(),
get_y = get_y,
item_tfms = Resize(128),
batch_tfms = aug_transforms(),
n_inp = 1)
dls = coco %>% dataloaders('coco_tiny/train')
dls %>% show_batch(max_n = 12)
Build a model with RetinaNet components:
encoder = create_body(resnet34(), pretrained = TRUE)
arch = RetinaNet(encoder, get_c(dls), final_bias=-4)
ratios = c(1/2,1,2)
scales = c(1,2**(-1/3), 2**(-2/3))
crit = RetinaNetFocalLoss(scales = scales, ratios = ratios)
nn = nn()
retinanet_split = function(m) {
L(m$encoder,nn$Sequential(m$c5top6, m$p6top7, m$merges,
m$smoothers, m$classifier, m$box_regressor))$map(params())
}
Unfreeze and train model:
learn = Learner(dls, arch, loss_func = crit, splitter = retinanet_split)
learn$freeze()
learn %>% fit_one_cycle(10, slice(1e-5, 1e-4))
epoch train_loss valid_loss time
0 3.377425 3.352676 00:06
1 3.304372 2.941969 00:03
2 3.400184 2.811601 00:03
3 3.239992 2.896531 00:03
4 3.159638 3.090069 00:03
5 3.117127 2.978687 00:03
6 3.079744 2.832738 00:03
7 3.035829 2.801513 00:03
8 2.996779 2.798598 00:03
9 2.962824 2.798002 00:03