A single-file, modularized implementation of RegNet as introduced in [I. Radosavovic, R. P. Kosaraju, R. Girshick, K. He, and P. Dollár]: Designing Network Design Spaces. (CVPR), 2020
Original implementation: facebookresearch/pycls
Pretrained weights are taken from original implementation.
The example below creates an RegNetY-200MF model that takes 3-channel RGB image as input and outputs distribution over 50 classes, model weights are initialized with weights pretrained on ImageNet dataset:
from regnet import RegNet
model = RegNet('RegNetY-200MF',
in_channels=3,
num_classes=50,
pretrained=True
)
# inp - tensor of shape [batch_size, in_channels, image_height, image_width]
inp = torch.randn([10, 3, 224, 224])
# to get predictions:
pred = model(inp)
print('out shape:', pred.shape)
# >>> out shape: torch.Size([10, 50])
# to extract features:
features = model.get_features(inp)
for i, feature in enumerate(features):
print('feature %d shape:' % i, feature.shape)
# >>> feature 0 shape: torch.Size([10, 24, 56, 56])
# >>> feature 1 shape: torch.Size([10, 56, 28, 28])
# >>> feature 2 shape: torch.Size([10, 152, 14, 14])
# >>> feature 3 shape: torch.Size([10, 368, 7, 7])pred is now a tensor containing output logits while
features is a list of 4 tensors representing outputs
of model's 4 intermediate convolutional stages.
- name, (str) - Model name, e.g. 'RegNetY-200MF'
- in_channels, (int), (Default=3) - Number of channels in input image
- num_classes, (int), (Default=1000) - Number of output classes
- bn_eps, (float), (Default=1e-5) - Batch normalizaton epsilon
- bn_momentum, (float), (Default=0.1) - Batch normalization momentum
- relu_inplace, (bool), (Default=False) - relu_inplace parameter for ReLU activation
- pretrained, (bool), (Default=False) - Whether to initialize model with weights pretrained on ImageNet dataset
- progress, (bool), (Default=False) - Show progress bar when downloading pretrained weights.
A simple script to evaluate pretrained models against Imagenet validation set is provided in imagenet_eval.ipynb.
Accuracy achieved by models with pre-trained weights against ImageNet dataset:
| Model | Accuracy, % |
|---|---|
| RegNetX-200MF | 69.718 |
| RegNetX-400MF | 73.47 |
| RegNetX-600MF | 74.754 |
| RegNetX-800MF | 75.662 |
| RegNetX-1.6GF | 77.632 |
| RegNetX-3.2GF | 78.85 |
| RegNetX-4.0GF | 79.06 |
| RegNetX-6.4GF | 79.556 |
| RegNetX-8.0GF | 79.784 |
| RegNetX-12GF | 80.138 |
| RegNetX-16GF | 80.56 |
| RegNetX-32GF | 80.776 |
| RegNetY-200MF | 71.356 |
| RegNetY-400MF | 74.92 |
| RegNetY-600MF | 76.182 |
| RegNetY-800MF | 77.106 |
| RegNetY-1.6GF | 78.68 |
| RegNetY-3.2GF | 79.58 |
| RegNetY-4.0GF | 80.112 |
| RegNetY-6.4GF | 80.512 |
| RegNetY-8.0GF | 80.614 |
| RegNetY-12GF | 81.052 |
| RegNetY-16GF | 81.104 |
| RegNetY-32GF | 81.464 |
- Python v3.5+
- Pytorch v1.4+