| 108 | |
| 109 | |
| 110 | def forward(self, x): |
| 111 | if self.channels_last == True: |
| 112 | x.contiguous(memory_format=torch.channels_last) |
| 113 | |
| 114 | layers = self.forward_transformer(self.pretrained, x) |
| 115 | if self.number_layers == 3: |
| 116 | layer_1, layer_2, layer_3 = layers |
| 117 | else: |
| 118 | layer_1, layer_2, layer_3, layer_4 = layers |
| 119 | |
| 120 | layer_1_rn = self.scratch.layer1_rn(layer_1) |
| 121 | layer_2_rn = self.scratch.layer2_rn(layer_2) |
| 122 | layer_3_rn = self.scratch.layer3_rn(layer_3) |
| 123 | if self.number_layers >= 4: |
| 124 | layer_4_rn = self.scratch.layer4_rn(layer_4) |
| 125 | |
| 126 | if self.number_layers == 3: |
| 127 | path_3 = self.scratch.refinenet3(layer_3_rn, size=layer_2_rn.shape[2:]) |
| 128 | else: |
| 129 | path_4 = self.scratch.refinenet4(layer_4_rn, size=layer_3_rn.shape[2:]) |
| 130 | path_3 = self.scratch.refinenet3(path_4, layer_3_rn, size=layer_2_rn.shape[2:]) |
| 131 | path_2 = self.scratch.refinenet2(path_3, layer_2_rn, size=layer_1_rn.shape[2:]) |
| 132 | path_1 = self.scratch.refinenet1(path_2, layer_1_rn) |
| 133 | |
| 134 | if self.scratch.stem_transpose is not None: |
| 135 | path_1 = self.scratch.stem_transpose(path_1) |
| 136 | |
| 137 | out = self.scratch.output_conv(path_1) |
| 138 | |
| 139 | return out |
| 140 | |
| 141 | |
| 142 | class DPTDepthModel(DPT): |