Transfer learning: ability to use a pre-trained model as a "shortcut" to learn patterns from data it was not originally trained on.
There are two types of transfer learning in the context of deep learning:
- Transfer learning via feature extraction: treating networks as arbitrary feature extractors
- Transfer learning via fine-tuning
This repo will focus on the first method of transfer learning
3,000 images of dogs 🐶 vs cats 🐱 vs pandas 🐼, a 1,000 each
| Dog | Cat | Panda |
|---|---|---|
 |
 |
 |
See this repo for further details on the dataset.
Pictures of objects belonging to 101 categories. About 40 to 800 images per category. Most categories have about 50 images. The size of each image is roughly 300 x 200 pixels. Dataset was collected by the Computational Vision at CalTech.
Pictures of 17 category flower dataset with 80 images for each class. Dataset was collected by the Visual Geometry Group at the University of Oxford.
- Build the docker image:
$ cd docker
$ make build
- Create a docker container based on the image:
$ make run
- SSH to the docker container:
$ make dev
- Extract features:
$ python src/extract_features.py --dataset datasets/animals/images --output datasets/animals/hdf5/features.hdf5
[INFO] loading images...
[INFO] loading network...
Downloading data from https://github.com/fchollet/deep-learning-models/releases/download/v0.1/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5
58892288/58889256 [==============================] - 5s 0us/step
Extracting Features: 100% |############################################################| Time: 0:04:57
- Train and evaluate the model:
$ python src/train_model.py --db datasets/animals/hdf5/features.hdf5 --model models/animals.cpickle
[INFO] tuning hyperparameters...
[INFO] best hyperparameters: {'C': 0.1}
[INFO] evaluating...
precision recall f1-score support
cat 0.99 0.98 0.99 251
dog 0.98 0.98 0.98 250
panda 1.00 1.00 1.00 249
accuracy 0.99 750
macro avg 0.99 0.99 0.99 750
weighted avg 0.99 0.99 0.99 750
[INFO] saving model...
- Extract features:
$ python src/extract_features.py --dataset datasets/CALTECH-101/images/ --output datasets/CALTECH-101/hdf5/features.hdf5
[INFO] loading images...
[INFO] loading network...
Extracting Features: 100% |############################################################| Time: 0:14:12
- Train and evaluate the model:
$ python src/train_model.py --db datasets/CALTECH-101/hdf5/features.hdf5 --model models/CALTECH-101.cpickle
[INFO] best hyperparameters: {'C': 0.1}
[INFO] evaluating...
precision recall f1-score support
Faces 0.98 0.99 0.99 121
Faces_easy 0.99 0.98 0.99 107
Leopards 1.00 1.00 1.00 57
Motorbikes 0.99 1.00 0.99 178
accordion 0.93 1.00 0.96 13
airplanes 0.99 1.00 0.99 194
anchor 0.90 0.75 0.82 12
ant 0.80 0.89 0.84 9
barrel 1.00 1.00 1.00 8
bass 0.93 0.81 0.87 16
beaver 1.00 0.75 0.86 12
binocular 0.78 1.00 0.88 7
bonsai 0.90 1.00 0.95 27
brain 0.94 0.97 0.96 33
brontosaurus 0.94 0.88 0.91 17
buddha 0.95 1.00 0.97 19
butterfly 0.97 0.97 0.97 30
camera 1.00 1.00 1.00 17
cannon 1.00 0.92 0.96 12
car_side 1.00 1.00 1.00 23
ceiling_fan 0.83 0.91 0.87 11
cellphone 1.00 1.00 1.00 18
chair 0.87 0.93 0.90 14
chandelier 0.96 1.00 0.98 22
cougar_body 1.00 0.75 0.86 16
cougar_face 0.94 1.00 0.97 16
crab 0.95 0.91 0.93 23
crayfish 0.72 0.87 0.79 15
crocodile 0.71 1.00 0.83 10
crocodile_head 1.00 0.93 0.96 14
cup 1.00 1.00 1.00 11
dalmatian 1.00 1.00 1.00 19
dollar_bill 0.92 1.00 0.96 12
dolphin 0.82 1.00 0.90 14
dragonfly 1.00 1.00 1.00 21
electric_guitar 0.95 1.00 0.97 19
elephant 0.82 0.90 0.86 10
emu 1.00 0.92 0.96 13
euphonium 1.00 1.00 1.00 18
ewer 1.00 1.00 1.00 25
ferry 1.00 1.00 1.00 18
flamingo 0.92 1.00 0.96 11
flamingo_head 0.92 1.00 0.96 12
garfield 1.00 1.00 1.00 8
gerenuk 1.00 1.00 1.00 7
gramophone 0.93 1.00 0.97 14
grand_piano 1.00 1.00 1.00 28
hawksbill 0.96 1.00 0.98 24
headphone 1.00 0.92 0.96 13
hedgehog 0.93 1.00 0.97 14
helicopter 0.89 1.00 0.94 16
ibis 1.00 0.92 0.96 24
inline_skate 1.00 1.00 1.00 4
joshua_tree 0.89 0.94 0.91 17
kangaroo 0.95 0.95 0.95 21
ketch 0.86 0.89 0.87 35
lamp 0.83 0.94 0.88 16
laptop 1.00 1.00 1.00 18
llama 0.77 0.83 0.80 12
lobster 1.00 0.46 0.63 13
lotus 0.50 0.89 0.64 9
mandolin 1.00 0.92 0.96 12
mayfly 1.00 0.90 0.95 10
menorah 1.00 1.00 1.00 19
metronome 1.00 0.92 0.96 13
minaret 1.00 1.00 1.00 18
nautilus 0.92 0.92 0.92 13
octopus 0.67 0.67 0.67 6
okapi 1.00 0.90 0.95 10
pagoda 0.94 1.00 0.97 15
panda 1.00 1.00 1.00 10
pigeon 1.00 0.92 0.96 12
pizza 0.93 0.93 0.93 15
platypus 0.80 0.50 0.62 8
pyramid 1.00 0.96 0.98 27
revolver 1.00 0.95 0.98 22
rhino 1.00 1.00 1.00 14
rooster 1.00 0.92 0.96 12
saxophone 1.00 1.00 1.00 12
schooner 0.79 0.75 0.77 20
scissors 0.88 1.00 0.93 7
scorpion 0.96 1.00 0.98 22
sea_horse 0.92 0.92 0.92 12
snoopy 0.90 0.82 0.86 11
soccer_ball 1.00 1.00 1.00 13
stapler 1.00 1.00 1.00 11
starfish 1.00 0.95 0.98 22
stegosaurus 0.90 0.90 0.90 10
stop_sign 1.00 1.00 1.00 14
strawberry 1.00 1.00 1.00 11
sunflower 1.00 0.96 0.98 27
tick 0.91 0.91 0.91 11
trilobite 1.00 1.00 1.00 21
umbrella 1.00 0.90 0.95 20
watch 1.00 0.98 0.99 52
water_lilly 0.67 0.22 0.33 9
wheelchair 1.00 0.87 0.93 15
wild_cat 0.78 1.00 0.88 7
windsor_chair 1.00 0.86 0.92 14
wrench 0.90 0.90 0.90 10
yin_yang 0.88 0.88 0.88 16
accuracy 0.96 2170
macro avg 0.94 0.93 0.93 2170
weighted avg 0.96 0.96 0.95 2170
[INFO] saving model...
- Extract features:
$ python src/extract_features.py --dataset datasets/Flowers-17/images/ --output datasets/Flowers-17/hdf5/features.hdf5
[INFO] loading images...
[INFO] loading network...
Extracting Features: 100% |##################################################################################################################################################################| Time: 0:02:16
- Train and evaluate the model:
$ python src/train_model.py --db datasets/Flowers-17/hdf5/features.hdf5 --model models/Flowers-17.cpickle
[INFO] tuning hyperparameters...
[INFO] best hyperparameters: {'C': 0.1}
[INFO] evaluating...
precision recall f1-score support
1 Daffodil 0.95 0.78 0.86 23
10 Sunflower 1.00 1.00 1.00 23
11 Daisy 1.00 1.00 1.00 18
12Colts Foot 1.00 1.00 1.00 18
13 Dandelalion 1.00 1.00 1.00 15
14 Cowslip 0.67 0.86 0.75 21
15 Buttercup 0.78 1.00 0.88 18
16 Windflower 0.90 0.86 0.88 22
17 Pansy 0.96 0.89 0.93 28
2 Snowdrop 0.70 1.00 0.82 14
3 Lily Valley 1.00 0.93 0.96 27
4 Bluebell 0.90 1.00 0.95 19
5 Crocus 0.93 0.81 0.87 16
6 Iris 1.00 0.89 0.94 19
7 Tigerlily 1.00 0.93 0.97 15
8Tulip 0.96 0.79 0.86 28
9 Fritillary 1.00 1.00 1.00 16
accuracy 0.92 340
macro avg 0.93 0.93 0.92 340
weighted avg 0.93 0.92 0.92 340
[INFO] saving model...
- Deep Learning for Computer Vision with Python by Dr. Adrian Rosebrock: https://www.pyimagesearch.com/deep-learning-computer-vision-python-book/