Python 如何在pytorch中使用共享编码器交替培训3个型号？_Python_Deep Learning_Pytorch_Torch

Python 如何在pytorch中使用共享编码器交替培训3个型号？

python deep-learning pytorch

Python 如何在pytorch中使用共享编码器交替培训3个型号？,python,deep-learning,pytorch,torch,Python,Deep Learning,Pytorch,Torch,我想建立3个类似U-Net的模型，它们共享相同的编码器，但有独立的解码器（前两个模型与U-Net相同，用于分割，第三个模型没有短连接，用于重建）。我想交替地训练这3个模型，因为我有3个不同的数据集，它们有自己的标签 model0, model1, model2 = build3models() models = [model0, model1, model2] dataloaders = [None, None, None] dataloaders[0] = Dataloader_for_mo

我想建立3个类似U-Net的模型，它们共享相同的编码器，但有独立的解码器（前两个模型与U-Net相同，用于分割，第三个模型没有短连接，用于重建）。我想交替地训练这3个模型，因为我有3个不同的数据集，它们有自己的标签

model0, model1, model2 = build3models() models = [model0, model1, model2] dataloaders = [None, None, None] dataloaders[0] = Dataloader_for_model0 dataloaders[1] = Dataloader_for_model1 dataloaders[2] = Dataloader_for_model2 max_steps = 10000 for step in range(max_steps): for data, model in zip(dataloaders, models]: x, y = next(dataloader[i]) pred = model(x) opt = torch.optim.Adam(model.parameters(), lr=1e-4) opt.zero_grad() criterion = nn.MSE() loss = criterion_seg(pred , y) loss.backward() opt.step()
我的问题是：如何使用共享编码器构建3个模型，并针对不同的模型使用不同的损耗/opt/lr等单独或交替地对它们进行培训
我已经构建了以下编码器和解码器，但我不知道如何正确连接和使用它们

class Encoder(nn.Module): def __init__( self, dimensions: int = 3, in_channels: int = 1, features: Sequence[int] = (32, 32, 64, 128, 256, 32), act: Union[str, tuple] = ("LeakyReLU", {"negative_slope": 0.1, "inplace": True}), norm: Union[str, tuple] = ("instance", {"affine": True}), dropout: Union[float, tuple] = 0.0, upsample: str = "deconv", ): """ A UNet implementation with 1D/2D/3D supports. Based on: Falk et al. "U-Net – Deep Learning for Cell Counting, Detection, and Morphometry". Nature Methods 16, 67–70 (2019), DOI: http://dx.doi.org/10.1038/s41592-018-0261-2 Args: dimensions: number of spatial dimensions. Defaults to 3 for spatial 3D inputs. in_channels: number of input channels. Defaults to 1. features: six integers as numbers of features. Defaults to ``(32, 32, 64, 128, 256, 32)``, - the first five values correspond to the five-level encoder feature sizes. - the last value corresponds to the feature size after the last upsampling. act: activation type and arguments. Defaults to LeakyReLU. norm: feature normalization type and arguments. Defaults to instance norm. dropout: dropout ratio. Defaults to no dropout. upsample: upsampling mode, available options are ``"deconv"``, ``"pixelshuffle"``, ``"nontrainable"``. Examples:: # for spatial 2D >>> net = BasicUNet(dimensions=2, features=(64, 128, 256, 512, 1024, 128)) # for spatial 2D, with group norm >>> net = BasicUNet(dimensions=2, features=(64, 128, 256, 512, 1024, 128), norm=("group", {"num_groups": 4})) # for spatial 3D >>> net = BasicUNet(dimensions=3, features=(32, 32, 64, 128, 256, 32)) See Also - :py:class:`monai.networks.nets.DynUNet` - :py:class:`monai.networks.nets.UNet` """ super().__init__() fea = ensure_tuple_rep(features, 6) print(f"BasicUNet features: {fea}.") self.conv_0 = TwoConv(dimensions, in_channels, features[0], act, norm, dropout) self.down_1 = Down(dimensions, fea[0], fea[1], act, norm, dropout) self.down_2 = Down(dimensions, fea[1], fea[2], act, norm, dropout) self.down_3 = Down(dimensions, fea[2], fea[3], act, norm, dropout) self.down_4 = Down(dimensions, fea[3], fea[4], act, norm, dropout) def forward(self, x: torch.Tensor): """ Args: x: input should have spatially N dimensions ``(Batch, in_channels, dim_0[, dim_1, ..., dim_N])``, N is defined by `dimensions`. It is recommended to have ``dim_n % 16 == 0`` to ensure all maxpooling inputs have even edge lengths. Returns: A torch Tensor of "raw" predictions in shape ``(Batch, out_channels, dim_0[, dim_1, ..., dim_N])``. """ x0 = self.conv_0(x) x1 = self.down_1(x0) x2 = self.down_2(x1) x3 = self.down_3(x2) x4 = self.down_4(x3) return x0, x1, x2, x3, x4 class DecoderSeg(nn.Module): def __init__( self, dimensions: int = 3, in_channels: int = 1, out_channels: int = 2, features: Sequence[int] = (32, 32, 64, 128, 256, 32), act: Union[str, tuple] = ("LeakyReLU", {"negative_slope": 0.1, "inplace": True}), norm: Union[str, tuple] = ("instance", {"affine": True}), dropout: Union[float, tuple] = 0.0, upsample: str = "deconv", ): """ A UNet implementation with 1D/2D/3D supports. Based on: Falk et al. "U-Net – Deep Learning for Cell Counting, Detection, and Morphometry". Nature Methods 16, 67–70 (2019), DOI: http://dx.doi.org/10.1038/s41592-018-0261-2 Args: dimensions: number of spatial dimensions. Defaults to 3 for spatial 3D inputs. in_channels: number of input channels. Defaults to 1. out_channels: number of output channels. Defaults to 2. features: six integers as numbers of features. Defaults to ``(32, 32, 64, 128, 256, 32)``, - the first five values correspond to the five-level encoder feature sizes. - the last value corresponds to the feature size after the last upsampling. act: activation type and arguments. Defaults to LeakyReLU. norm: feature normalization type and arguments. Defaults to instance norm. dropout: dropout ratio. Defaults to no dropout. upsample: upsampling mode, available options are ``"deconv"``, ``"pixelshuffle"``, ``"nontrainable"``. Examples:: # for spatial 2D >>> net = BasicUNet(dimensions=2, features=(64, 128, 256, 512, 1024, 128)) # for spatial 2D, with group norm >>> net = BasicUNet(dimensions=2, features=(64, 128, 256, 512, 1024, 128), norm=("group", {"num_groups": 4})) # for spatial 3D >>> net = BasicUNet(dimensions=3, features=(32, 32, 64, 128, 256, 32)) See Also - :py:class:`monai.networks.nets.DynUNet` - :py:class:`monai.networks.nets.UNet` """ super().__init__() fea = ensure_tuple_rep(features, 6) print(f"BasicUNet features: {fea}.") self.upcat_4 = UpCat(dimensions, fea[4], fea[3], fea[3], act, norm, dropout, upsample) self.upcat_3 = UpCat(dimensions, fea[3], fea[2], fea[2], act, norm, dropout, upsample) self.upcat_2 = UpCat(dimensions, fea[2], fea[1], fea[1], act, norm, dropout, upsample) self.upcat_1 = UpCat(dimensions, fea[1], fea[0], fea[5], act, norm, dropout, upsample, halves=False) self.final_conv = Conv["conv", dimensions](fea[5], out_channels, kernel_size=1) def forward(self, x0: torch.Tensor, x1: torch.Tensor, x2: torch.Tensor, x3: torch.Tensor, x4: torch.Tensor): """ Args: x: input should have spatially N dimensions ``(Batch, in_channels, dim_0[, dim_1, ..., dim_N])``, N is defined by `dimensions`. It is recommended to have ``dim_n % 16 == 0`` to ensure all maxpooling inputs have even edge lengths. Returns: A torch Tensor of "raw" predictions in shape ``(Batch, out_channels, dim_0[, dim_1, ..., dim_N])``. """ u4 = self.upcat_4(x4, x3) u3 = self.upcat_3(u4, x2) u2 = self.upcat_2(u3, x1) u1 = self.upcat_1(u2, x0) u_out = self.final_conv(u1) return u_out class DecoderRec(nn.Module): def __init__( self, dimensions: int = 3, in_channels: int = 1, out_channels: int = 1, features: Sequence[int] = (32, 32, 64, 128, 256, 32), act: Union[str, tuple] = ("LeakyReLU", {"negative_slope": 0.1, "inplace": True}), norm: Union[str, tuple] = ("instance", {"affine": True}), dropout: Union[float, tuple] = 0.0, upsample: str = "deconv", ): """ A UNet implementation with 1D/2D/3D supports. Based on: Falk et al. "U-Net – Deep Learning for Cell Counting, Detection, and Morphometry". Nature Methods 16, 67–70 (2019), DOI: http://dx.doi.org/10.1038/s41592-018-0261-2 Args: dimensions: number of spatial dimensions. Defaults to 3 for spatial 3D inputs. in_channels: number of input channels. Defaults to 1. out_channels: number of output channels. Defaults to 2. features: six integers as numbers of features. Defaults to ``(32, 32, 64, 128, 256, 32)``, - the first five values correspond to the five-level encoder feature sizes. - the last value corresponds to the feature size after the last upsampling. act: activation type and arguments. Defaults to LeakyReLU. norm: feature normalization type and arguments. Defaults to instance norm. dropout: dropout ratio. Defaults to no dropout. upsample: upsampling mode, available options are ``"deconv"``, ``"pixelshuffle"``, ``"nontrainable"``. Examples:: # for spatial 2D >>> net = BasicUNet(dimensions=2, features=(64, 128, 256, 512, 1024, 128)) # for spatial 2D, with group norm >>> net = BasicUNet(dimensions=2, features=(64, 128, 256, 512, 1024, 128), norm=("group", {"num_groups": 4})) # for spatial 3D >>> net = BasicUNet(dimensions=3, features=(32, 32, 64, 128, 256, 32)) See Also - :py:class:`monai.networks.nets.DynUNet` - :py:class:`monai.networks.nets.UNet` """ super().__init__() fea = ensure_tuple_rep(features, 6) print(f"BasicUNet features: {fea}.") self.up_4 = Up(dimensions, fea[4], fea[3], fea[3], act, norm, dropout, upsample) self.up_3 = Up(dimensions, fea[3], fea[2], fea[2], act, norm, dropout, upsample) self.up_2 = Up(dimensions, fea[2], fea[1], fea[1], act, norm, dropout, upsample) self.up_1 = Up(dimensions, fea[1], fea[0], fea[5], act, norm, dropout, upsample, halves=False) self.final_conv = Conv["conv", dimensions](fea[5], out_channels, kernel_size=1) def forward(self, x: torch.Tensor): """ Args: x: input should have spatially N dimensions ``(Batch, in_channels, dim_0[, dim_1, ..., dim_N])``, N is defined by `dimensions`. It is recommended to have ``dim_n % 16 == 0`` to ensure all maxpooling inputs have even edge lengths. Returns: A torch Tensor of "raw" predictions in shape ``(Batch, out_channels, dim_0[, dim_1, ..., dim_N])``. """ r4 = self.up_4(x) r3 = self.up_3(r4) r2 = self.up_2(r3) r1 = self.up_1(r2) r_out = self.final_conv(r1) return r_out