FASHION MNIST

Classifier	Preprocessing	Fashion test accuracy	Inference time (ms/image)
2 Conv+pooling ~168k params	Normalization, Random cropping, Random Horizontal flip, BN	0.9056	0.3
5 Conv+pooling ~3M params	Normalization, BN	0.913	1.1
VGG16 138M params minus 5 first convolution which are frozen	Normalization	0.9369	4.5

Random cropping

The semantics of the image are preserved but the activation values of the conv net are different. The conv net learns to associate a broader range of spatial activations with a certain class label and improves the robustness of the feature detectors in conv nets.

Data Augmentation using Random Image Cropping and Patching for Deep CNNs Ryo Takahashi, Takashi Matsubara, Kuniaki Uehara

Args:

parser.add_argument('--model-type',
                    choices=['vgg', 'two_conv', 'five_conv'],
                    required=True,
                    help='')

parser.add_argument('-t',
                    '--test-model-path',
                    default=None,
                    type=str,
                    help='model path')

parser.add_argument('-r',
                    '--resume-model-path',
                    default=None,
                    type=str,
                    help='model path')

parser.add_argument('--train-batch-size',
                    default=50,
                    help='batch size for training with Adam')

parser.add_argument('--lr',
                    default=0.005,
                    type=float,
                    help='learning rate')

parser.add_argument('--train-epoch',
                    default=60,
                    type=int,
                    help='number of training epoch')

parser.add_argument('--seed',
                    default=42,
                    help='seed')

parser.add_argument('--save-dir',
                    default='./data',
                    help='saving metrics dir')

parser.add_argument('--optimizer',
                    choices=['adam', 'sgd'],
                    default='adam',
                    help='')

parser.add_argument('--dump-metrics-frequency',
                    metavar='Batch_n',
                    default='600',
                    type=int,
                    help='Dump metrics every Batch_n batches')

parser.add_argument(
                    '--threshold-validation-accuracy',
                    default='0.95',
                    type=float,
                    help='Threshold validation to reach for stopping training')

parser.add_argument(
                    '--num-threads',
                    default='0',
                    type=int,
                    help='Number of CPU to use for processing mini batches')

parser.add_argument('--scale',
                    action='store_true',
                    help='scale input in [0-1] range')

parser.add_argument(
                    '--standardize',
                    action='store_true',
                    help='Subtract each instance by mean of data and divide by std')

parser.add_argument('--augment',
                    action='store_true',
                    help='Use data augmentation')

parser.add_argument('--pretrained',
                    action='store_true',
                    help='Use pretrained weights for VGG')

parser.add_argument('--batch-norm',
                    action='store_true',
                    help='Use batch norm')

### Examples:

Training:

python -m eval --model-type two_conv --train-epoch 60

Test:

python -m eval --model-type two_conv --test-model-path ./data/models/two_conv.pth