SparKLightDecoder¶

class mmpretrain.models.necks.SparKLightDecoder(feature_dim, upsample_ratio, mid_channels=0, kernel_size=4, scale_factor=2, num_conv_blocks=1, norm_cfg={'type': 'SyncBN'}, act_cfg={'type': 'ReLU6'}, last_act=False, init_cfg=[{'type': 'Kaiming', 'layer': ['Conv2d', 'ConvTranspose2d']}, {'type': 'TruncNormal', 'std': 0.02, 'layer': ['Linear']}, {'type': 'Constant', 'val': 1, 'layer': ['_BatchNorm', 'LayerNorm', 'SyncBatchNorm']}])[source]¶

The decoder for SparK, which upsamples the feature maps.

Parameters:

feature_dim (int) – The dimension of feature map.
upsample_ratio (int) – The ratio of upsample, equal to downsample_raito of the algorithm.
mid_channels (int) – The middle channel of DecoderConvModule. Defaults to 0.
kernel_size (int) – The kernel size of ConvTranspose2d in DecoderConvModule. Defaults to 4.
scale_factor (int) – The scale_factor of ConvTranspose2d in DecoderConvModule. Defaults to 2.
num_conv_blocks (int) – The number of convolution blocks in DecoderConvModule. Defaults to 1.
norm_cfg (dict) – Normalization config. Defaults to dict(type=’SyncBN’).
act_cfg (dict) – Activation config. Defaults to dict(type=’ReLU6’).
last_act (bool) – Whether apply the last activation in DecoderConvModule. Defaults to False.
init_cfg (dict or list[dict], optional) – Initialization config dict.