youngwanLEE / vovnet-detectron2

[CVPR 2020] VoVNet backbone networks for detectron2
Other
376 stars 69 forks source link

On designing thinner VovNet #15

Open txytju opened 4 years ago

txytju commented 4 years ago

Thanks for your great work! I'm working on a project of instance segmentation in which taget_classes=2. I tried to use Vovnet in the project to replace the ResNet that I have already used. Because the target_class is very different from COCO, the output channels of FPN of my model is only 64 and I think the VOVnet here should be a lot thinner than COCO version. I have designed a VovNet like the model bellow, but the result is worse than ResNet. Do you have any suggestion on how to design thinner VovNet? Thanks

StageSpec = namedtuple( "StageSpec", [ "index", # Index of the stage, eg 1, 2, ..,. 5 "block_count", # Number of residual blocks in the stage "layer_per_block", # Number of OSA modules per block "return_features", # True => return the last feature map from this stage "in_channels", "out_channels", ], ) VoVNet67_eSE = tuple( StageSpec(index=i, block_count=b, layer_per_block=l, in_channels=in_c, out_channels=out_c, return_features=r) for (i, b, l, in_c, out_c, r) in ((1, 1, 5, 16, 48, True), (2, 3, 5, 32, 96, True), (3, 6, 5, 64, 192, True), (4, 3, 5, 128, 256, True))