Densenet-efficient-model

Convert

In order to obtain the models for the efficient model, efficient model for PyTorch0.3 in PyTorch. You need to download the models from originl Torch models. And then convert the torch models to PyTorch models (You can also use the original convert_torch.py). Then they are converted to efficient models.

   python convert_torch.py -m densenet_cosine_264_k48.t7
   python convert_efficient.py

Note: You need to call correspoding function (Just one line code) in the main function in convert_efficient.py if you want to convert other models.

Efficient DenseNet models

DenseNet-264(k=32)

    growth_rate = 32
    block_config=(6,12,64,48)
    model = DenseNet_Efficient.DenseNetEfficient(num_init_features=64,
                                                 growth_rate = growth_rate,
                                                block_config = block_config,
                                                 num_classes = 1000,
                                                       cifar = False)

DenseNet-232(k=48)

    growth_rate = 48
    block_config=(6,12,48,48)
    model = DenseNet_Efficient.DenseNetEfficient(num_init_features=96,
                                                 growth_rate = growth_rate,
                                                block_config = block_config,
                                                 num_classes = 1000,
                                                       cifar = False)

DenseNet-cosine-264 (k=32)

    growth_rate = 32
    block_config=(6,12,64,48)
    model = DenseNet_Efficient.DenseNetEfficient(num_init_features=64,
                                                 growth_rate = growth_rate,
                                                block_config = block_config,
                                                 num_classes = 1000,
                                                       cifar = False)

DenseNet-cosine-264 (k=48)

    growth_rate = 48
    block_config=(6,12,64,48)
    model = DenseNet_Efficient.DenseNetEfficient(num_init_features=96,
                                                 growth_rate = growth_rate,
                                                block_config = block_config,
                                                 num_classes = 1000,
                                                       cifar = False)

Validated

All the models in this table can be converted and the results have been validated.

Network	Top-1 error	Download
DenseNet-264(k=32)	22.1	Download(129MB)
DenseNet-232(k=48)	21.2	Download(214MB)
DenseNet-cosine-264 (k=32)	21.6	DenseNet(129MB)
DenseNet-cosine-264 (k=48)	20.4	DenseNet(280MB)