From 4c424db844985518bb3f89b6165df5026cd0fa04 Mon Sep 17 00:00:00 2001 From: harisreedhar Date: Fri, 9 Oct 2026 15:19:21 +0530 Subject: [PATCH] MaskNet support for 512 and 1024 resolution --- hyperswap/src/networks/masknet.py | 81 +++++++++++++++++++++++++------ hyperswap/tests/test_networks.py | 6 ++- 2 files changed, 70 insertions(+), 17 deletions(-) diff --git a/hyperswap/src/networks/masknet.py b/hyperswap/src/networks/masknet.py index 166b1c6..22a06ff 100644 --- a/hyperswap/src/networks/masknet.py +++ b/hyperswap/src/networks/masknet.py @@ -12,39 +12,88 @@ def __init__(self, config_parser : ConfigParser) -> None: self.config_input_channels = config_parser.getint('training.model.masker', 'input_channels') self.config_output_channels = config_parser.getint('training.model.masker', 'output_channels') self.config_num_filters = config_parser.getint('training.model.masker', 'num_filters') - self.down_samples = self.create_down_samples(self.config_input_channels, self.config_num_filters) - self.up_samples = self.create_up_samples(self.config_num_filters) + self.config_output_size = config_parser.getint('training.model.generator', 'output_size') + self.scale_down_samples = self.create_scale_down_samples() + self.base_down_samples = self.create_base_down_samples() + self.base_up_samples = self.create_base_up_samples() + self.scale_up_samples = self.create_scale_up_samples() self.bottleneck = BottleNeck(self.config_num_filters * 4) self.conv = nn.Conv2d(self.config_num_filters, self.config_output_channels, kernel_size = 1) self.sigmoid = nn.Sigmoid() - @staticmethod - def create_down_samples(input_channels : int, num_filters : int) -> nn.ModuleList: - return nn.ModuleList( + def create_scale_down_samples(self) -> nn.ModuleList: + scale_down_samples = nn.ModuleList( [ - DownSample(input_channels, num_filters), - DownSample(num_filters, num_filters * 2), - DownSample(num_filters * 2, num_filters * 4) + DownSample(self.config_input_channels, self.config_num_filters) ]) - @staticmethod - def create_up_samples(num_filters : int) -> nn.ModuleList: - return nn.ModuleList( + if self.config_output_size == 512: + scale_down_samples.extend( + [ + DownSample(self.config_num_filters, self.config_num_filters) + ]) + + if self.config_output_size == 1024: + scale_down_samples.extend( + [ + DownSample(self.config_num_filters, self.config_num_filters), + DownSample(self.config_num_filters, self.config_num_filters) + ]) + + return scale_down_samples + + def create_base_down_samples(self) -> nn.ModuleList: + base_down_samples = nn.ModuleList( + [ + DownSample(self.config_num_filters, self.config_num_filters * 2), + DownSample(self.config_num_filters * 2, self.config_num_filters * 4) + ]) + + return base_down_samples + + def create_base_up_samples(self) -> nn.ModuleList: + base_up_samples = nn.ModuleList( [ - UpSample(num_filters * 4, num_filters * 2), - UpSample(num_filters * 2, num_filters), - UpSample(num_filters, num_filters) + UpSample(self.config_num_filters * 4, self.config_num_filters * 2), + UpSample(self.config_num_filters * 2, self.config_num_filters), + UpSample(self.config_num_filters, self.config_num_filters) ]) + return base_up_samples + + def create_scale_up_samples(self) -> nn.ModuleList: + scale_up_samples = nn.ModuleList() + + if self.config_output_size == 512: + scale_up_samples.extend( + [ + UpSample(self.config_num_filters, self.config_num_filters) + ]) + + if self.config_output_size == 1024: + scale_up_samples.extend( + [ + UpSample(self.config_num_filters, self.config_num_filters), + UpSample(self.config_num_filters, self.config_num_filters) + ]) + + return scale_up_samples + def forward(self, input_tensor : Tensor, input_feature : Feature) -> Mask: output_mask = torch.cat([ input_tensor, input_feature ], dim = 1) - for down_sample in self.down_samples: + for down_sample in self.scale_down_samples: + output_mask = down_sample(output_mask) + + for down_sample in self.base_down_samples: output_mask = down_sample(output_mask) output_mask = self.bottleneck(output_mask) - for up_sample in self.up_samples: + for up_sample in self.base_up_samples: + output_mask = up_sample(output_mask) + + for up_sample in self.scale_up_samples: output_mask = up_sample(output_mask) output_mask = self.conv(output_mask) diff --git a/hyperswap/tests/test_networks.py b/hyperswap/tests/test_networks.py index 9294ef7..8065bf5 100644 --- a/hyperswap/tests/test_networks.py +++ b/hyperswap/tests/test_networks.py @@ -33,11 +33,15 @@ def test_aad_with_unet(output_size : int) -> None: assert output_tensor.shape == (1, 3, output_size, output_size) -@pytest.mark.parametrize('output_size', [ 256, 512 ]) +@pytest.mark.parametrize('output_size', [ 256, 512, 1024 ]) def test_mask_net(output_size : int) -> None: config_parser = ConfigParser() config_parser.read_dict( { + 'training.model.generator': + { + 'output_size': str(output_size) + }, 'training.model.masker': { 'input_channels': '67',