Finetuning results

Data: train: 2500 birds RGB (and a bit gray) images (different sizes, min_size = (120, 140)); target: 50 classes; scoring: categorical_accuracy.

SVM(sklearm.svm.SVC) on train reshaped (100, 100)

train_size	test_size	kernel	score	training time
2000	500	RBF	< 0.02	~ 5-10 min
2000	500	linear	0.11	~ 10 min

SVM(sklearm.svm.SVC) on features extracted with penult layer of pretrained VGG16 on ImageNet, train reshaped (224, 224)

extracting time ~ 60 min

train_size	test_size	kernel	score	training time
2000	500	RBF	< 0.02	~ 10 sec
2250	250	linear	0.71	~ 10 sec
2000	500	linear	0.70	~ 10 sec
1000	1500	linear	0.61	~ 10 sec
500	2000	linear	0.53	~ 10 sec
250	2250	linear	0.42	~ 10 sec
100	2400	linear	0.28	~ 10 sec

Keras simple convolutional network on train reshaped (224, 224)

Architecture:

Convolution2D(64, 3, 3, border_mode="same", activation="relu")
MaxPooling2D(pool_size=(3, 3))
Convolution2D(126, 3, 3, border_mode="same", activation="relu")
MaxPooling2D(pool_size=(3, 3))
Flatten()
Dense(250, activation="relu")
Dense(51, activation="softmax")

nb_epoch=10, batch_size=32

train_size	test_size	score	training time
2000	500	~ 0.02	~ 70 min

Other architectures give the same result.

Keras 2-dense-layers network on features extracted with penult layer of pretrained VGG16 on ImageNet, train reshaped (224, 224)

extracting time ~ 60 min nb_epoch=200, batch_size=60

train_size	test_size	score	training time
2250	250	~ 0.75	< 2 min
2000	500	~ 0.73	< 2 min
1750	750	~ 0.71	< 2 min
1500	1000	~ 0.70	< 2 min
1250	1250	~ 0.66	< 2 min
1000	1500	~ 0.64	< 2 min
750	1750	~ 0.60	< 2 min
500	2000	~ 0.54	< 2 min
250	2250	~ 0.40	< 2 min
174	2326	~ 0.33	< 2 min
99	2401	~ 0.27	< 2 min
24	2476	~ 0.10	< 2 min

Keras 2-dense-layers network on features extracted with penult layer of pretrained VGG16 on ImageNet, train reshaped (224, 224)

10 most important features with RandomForestClassifier.

train_size	test_size	score	training time
2000	500	~ 0.32	< 1 min