| import torch.nn as nn |
| import torch.nn.functional as F |
|
|
| import cliport.utils.utils as utils |
| from cliport.models.resnet import IdentityBlock, ConvBlock |
| from cliport.models.core.unet import Up |
| from cliport.models.rn50_bert_lingunet import RN50BertLingUNet |
|
|
|
|
| class RN50BertUNet(RN50BertLingUNet): |
| """ ImageNet RN50 & Bert with U-Net skip connections without language""" |
|
|
| def __init__(self, input_shape, output_dim, cfg, device, preprocess): |
| super().__init__(input_shape, output_dim, cfg, device, preprocess) |
|
|
| def _build_decoder(self): |
| self.conv1 = nn.Sequential( |
| nn.Conv2d(self.input_dim, 1024, kernel_size=3, stride=1, padding=1, bias=False), |
| nn.ReLU(True) |
| ) |
|
|
| self.up1 = Up(2048, 1024 // self.up_factor, self.bilinear) |
|
|
| self.up2 = Up(1024, 512 // self.up_factor, self.bilinear) |
|
|
| self.up3 = Up(512, 256 // self.up_factor, self.bilinear) |
|
|
| self.layer1 = nn.Sequential( |
| ConvBlock(128, [64, 64, 64], kernel_size=3, stride=1, batchnorm=self.batchnorm), |
| IdentityBlock(64, [64, 64, 64], kernel_size=3, stride=1, batchnorm=self.batchnorm), |
| nn.UpsamplingBilinear2d(scale_factor=2), |
| ) |
|
|
| self.layer2 = nn.Sequential( |
| ConvBlock(64, [32, 32, 32], kernel_size=3, stride=1, batchnorm=self.batchnorm), |
| IdentityBlock(32, [32, 32, 32], kernel_size=3, stride=1, batchnorm=self.batchnorm), |
| nn.UpsamplingBilinear2d(scale_factor=2), |
| ) |
|
|
| self.layer3 = nn.Sequential( |
| ConvBlock(32, [16, 16, 16], kernel_size=3, stride=1, batchnorm=self.batchnorm), |
| IdentityBlock(16, [16, 16, 16], kernel_size=3, stride=1, batchnorm=self.batchnorm), |
| nn.UpsamplingBilinear2d(scale_factor=2), |
| ) |
|
|
| self.conv2 = nn.Sequential( |
| nn.Conv2d(16, self.output_dim, kernel_size=1) |
| ) |
|
|
| def forward(self, x): |
| x = self.preprocess(x, dist='clip') |
|
|
| in_type = x.dtype |
| in_shape = x.shape |
| x = x[:,:3] |
| x, im = self.encode_image(x) |
| x = x.to(in_type) |
|
|
| x = self.conv1(x) |
|
|
| x = self.up1(x, im[-2]) |
| x = self.up2(x, im[-3]) |
| x = self.up3(x, im[-4]) |
|
|
| for layer in [self.layer1, self.layer2, self.layer3, self.conv2]: |
| x = layer(x) |
|
|
| x = F.interpolate(x, size=(in_shape[-2], in_shape[-1]), mode='bilinear') |
| return x |
