CNNs for image classification

Train CNNs for image classification from scratch.

I post several pretrained weights below.

Requirement

pytorch 1.4.0+
torchvision
tensorboard 1.14+
numpy
pyyaml
tqdm
pillow

Dataset

CIFAR-10
CIFAR-100
ImageNet 2012

Usage

Add configuration file under configs folder as follows

cuda: "all"  # if not specified, use cpu. or specified as "0", "0,1"
model:
    arch: resnet
    depth: 50
data:
    dataset: imagenet
    train_dir: /PATH/TO/ILSVRC/train
    val_dir: /PATH/TO/ILSVRC/val
    workers: 16
training:
    runid: xxxx  # recommended specified during validation and testing
    epochs: 100
    batch_size: 256
    loss:
        name: 'label_smooth'
        smoothing: 0.1
    optimizer:
        name: 'sgd'
        lr: 0.1
        weight_decay: 0.0001
        momentum: 0.9
    lr_schedule:
        name: 'multi_step'
        milestones: [30,60,90]
        gamma: 0.1
    save_interval: 1
    resume: save_model.pkl
    best_model: best_model.pkl

run train.py, validate.py or test.py as follows
- ```
python train.py --config configs/aaaa.yml
```

Pretrained Model on ImageNet 2012

Architecture	Top-1 error	Params	FLOPs	Pretrained weights
ResNet18 (My Imp.)	29.72	11.69M	1.82G	Google Drive Baidu Netdisk
ResNet18 (paper)	30.43	-	-	-
ResNet50 (My Imp.)	23.30	25.56M	4.11G	Google Drive Baidu Netdisk
ResNet50 (paper)	24.7	-	-	-
ResNet101 (My Imp.)	22.18	44.55M	7.84G	Google Drive
ResNet101 (paper)	22.44	-	-	-

ResNeXt50 (My Imp.)	22.35	25.03M	4.26G	Google Drive Baidu Netdisk
ResNeXt50 (paper)	22.2	-	-	-

SE ResNet50 (My Imp.)	22.64	28.09M	4.12G	Google Drive Baidu Netdisk
SE ResNet50 (paper)	23.29	-	-	-

CBAM ResNet50 (My Imp.)	22.40	28.07M	4.13G	Google Drive
CBAM ResNet50 (paper)	22.66	-	-	-

SKNet50 (My Imp.)	21.26	27.49M	4.50G	Google Drive Baidu Netdisk
SKNet50 (paper)	20.79	-	-	-


MobileNet V2 0.5x (My Imp.)	35.62	1.97M	138.46M	Google Drive
MobileNet V2 0.5x (paper)	35.6	-	-	-
MobileNet V2 1x (My Imp.)	28.09	3.50M	315.41M	Google Drive Baidu Netdisk
MobileNet V2 1x (paper)	28.0	-	-	-

MobileNet V3 large (My Imp.)	26.79	5.48M	230.05M	Google Drive
MobileNet V3 large (paper)	24.8	-	-	-

The hyperparameters and settings during my training for ResNet, ResNeXt, SENet, SKNet are the same as the paper, except I use label smooth loss.

And for MobileNet V2 and MobileNet V3, I follow the setup in this project, and use label smooth loss too.

Pretrained weights usage