How to train a voc compatible dataset.

Note: This notebook shows a special use case of training a VOC compatible dataset using the predefined VOC parser without creating both data, and parsers files as opposed to the fridge dataset example.

Installing IceVision

!pip install icevision[all]

Clone the raccoom dataset repository

!git clone https://github.com/datitran/raccoon_dataset

Imports

from icevision.all import *

WARNING:

Make sure you have already cloned the raccoon dataset using the command shown here above

Set images and annotations directories

data_dir = Path('raccoon_dataset')

images_dir = data_dir / 'images'
annotations_dir = data_dir / 'annotations'

Define class_map

class_map = ClassMap(['raccoon'])

Parser: Use icevision predefined VOC parser

parser = parsers.voc(annotations_dir=annotations_dir, images_dir=images_dir, class_map=class_map)

train and validation records

train_records, valid_records = parser.parse()

show_records(train_records[:3], ncols=3, class_map=class_map)

png

Datasets

Transforms

presize = 512
size = 384

train_tfms = tfms.A.Adapter([*tfms.A.aug_tfms(size=size, presize=presize), tfms.A.Normalize()])
valid_tfms = tfms.A.Adapter([*tfms.A.resize_and_pad(size=size), tfms.A.Normalize()])

Train and Validation Dataset Objects

train_ds = Dataset(train_records, train_tfms)
valid_ds = Dataset(valid_records, valid_tfms)

samples = [train_ds[5] for _ in range(3)]
show_samples(samples, class_map=class_map, denormalize_fn=denormalize_imagenet, ncols=3)

png

DataLoaders

train_dl = efficientdet.train_dl(train_ds, batch_size=16, num_workers=4, shuffle=True)
valid_dl = efficientdet.valid_dl(valid_ds, batch_size=16, num_workers=4, shuffle=False)

Model

model = efficientdet.model('tf_efficientdet_lite0', num_classes=len(class_map), img_size=size)

Fastai Learner

metrics = [COCOMetric(metric_type=COCOMetricType.bbox)]

learn = efficientdet.fastai.learner(dls=[train_dl, valid_dl], model=model, metrics=metrics)

Fastai Training

Learning Rate Finder

learn.freeze()
learn.lr_find()

SuggestedLRs(lr_min=0.10964782238006592, lr_steep=0.9120108485221863)

png

Fine tune: 2 Phases

Phase 1: Train the head for 5 epochs while freezing the body
Phase 2: Train both the body and the head during 20 epochs

learn.fine_tune(20, 1e-2, freeze_epochs=5)

epoch	train_loss	valid_loss	COCOMetric	time
0	1.927721	1.351568	0.000867	00:08
1	1.621056	1.375566	0.002002	00:07
2	1.401676	1.202271	0.130699	00:07
3	1.176787	1.099883	0.257632	00:07
4	1.018470	1.031241	0.240425	00:07

epoch	train_loss	valid_loss	COCOMetric	time
0	0.549507	0.923337	0.276733	00:09
1	0.535427	0.802162	0.383750	00:09
2	0.518957	0.778002	0.402308	00:09
3	0.488329	0.722738	0.451429	00:09
4	0.469851	0.719689	0.364621	00:09
5	0.457274	0.583479	0.438622	00:09
6	0.436571	0.597234	0.424187	00:08
7	0.423843	0.641547	0.370738	00:09
8	0.407281	0.592756	0.430994	00:09
9	0.395995	0.550204	0.507850	00:09
10	0.386040	0.569015	0.408947	00:09
11	0.376276	0.556436	0.502973	00:10
12	0.368530	0.506910	0.485442	00:10
13	0.354401	0.557794	0.452313	00:09
14	0.342909	0.517615	0.488974	00:10
15	0.328028	0.511221	0.512096	00:09
16	0.320078	0.488857	0.510211	00:10
17	0.308560	0.489812	0.506092	00:09
18	0.303276	0.478933	0.524952	00:09
19	0.297155	0.470403	0.524153	00:10

Show results

efficientdet.show_results(model, valid_ds, class_map=class_map)

png

Note: You might train the model longer in order to increase its accuracy

Happy Learning!

If you need any assistance, feel free to join our forum.