(params, channels, weight_nums, bias_nums)
| 167 | |
| 168 | |
| 169 | def parse_dynamic_params(params, channels, weight_nums, bias_nums): |
| 170 | assert params.dim() == 2 |
| 171 | assert len(weight_nums) == len(bias_nums) |
| 172 | assert params.size(1) == sum(weight_nums) + sum(bias_nums) |
| 173 | num_instances = params.size(0) |
| 174 | num_layers = len(weight_nums) |
| 175 | |
| 176 | params_splits = list( |
| 177 | torch.split_with_sizes(params, weight_nums + bias_nums, dim=1)) |
| 178 | |
| 179 | weight_splits = params_splits[:num_layers] |
| 180 | bias_splits = params_splits[num_layers:] |
| 181 | |
| 182 | for l in range(num_layers): |
| 183 | if l < num_layers - 1: |
| 184 | # out_channels x in_channels x 1 x 1 |
| 185 | weight_splits[l] = weight_splits[l].reshape( |
| 186 | num_instances * channels, -1, 1, 1) |
| 187 | bias_splits[l] = bias_splits[l].reshape(num_instances * channels) |
| 188 | else: |
| 189 | # out_channels x in_channels x 1 x 1 |
| 190 | weight_splits[l] = weight_splits[l].reshape( |
| 191 | num_instances * 1, -1, 1, 1) |
| 192 | bias_splits[l] = bias_splits[l].reshape(num_instances) |
| 193 | return weight_splits, bias_splits |
| 194 | |
| 195 | def exists(x): |
| 196 | return x is not None |
no outgoing calls
no test coverage detected