Copied from Plenoxels Continuous learning rate decay function. Adapted from JaxNeRF The returned rate is lr_init when step=0 and lr_final when step=max_steps, and is log-linearly interpolated elsewhere (equivalent to exponential decay). If lr_delay_steps>0 then the learning rat
(
lr_init, lr_final, lr_delay_steps=0, lr_delay_mult=1.0, max_steps=1000000
)
| 28 | return resized_image.unsqueeze(dim=-1).permute(2, 0, 1) |
| 29 | |
| 30 | def get_expon_lr_func( |
| 31 | lr_init, lr_final, lr_delay_steps=0, lr_delay_mult=1.0, max_steps=1000000 |
| 32 | ): |
| 33 | """ |
| 34 | Copied from Plenoxels |
| 35 | |
| 36 | Continuous learning rate decay function. Adapted from JaxNeRF |
| 37 | The returned rate is lr_init when step=0 and lr_final when step=max_steps, and |
| 38 | is log-linearly interpolated elsewhere (equivalent to exponential decay). |
| 39 | If lr_delay_steps>0 then the learning rate will be scaled by some smooth |
| 40 | function of lr_delay_mult, such that the initial learning rate is |
| 41 | lr_init*lr_delay_mult at the beginning of optimization but will be eased back |
| 42 | to the normal learning rate when steps>lr_delay_steps. |
| 43 | :param conf: config subtree 'lr' or similar |
| 44 | :param max_steps: int, the number of steps during optimization. |
| 45 | :return HoF which takes step as input |
| 46 | """ |
| 47 | |
| 48 | def helper(step): |
| 49 | if step < 0 or (lr_init == 0.0 and lr_final == 0.0): |
| 50 | # Disable this parameter |
| 51 | return 0.0 |
| 52 | if lr_delay_steps > 0: |
| 53 | # A kind of reverse cosine decay. |
| 54 | delay_rate = lr_delay_mult + (1 - lr_delay_mult) * np.sin( |
| 55 | 0.5 * np.pi * np.clip(step / lr_delay_steps, 0, 1) |
| 56 | ) |
| 57 | else: |
| 58 | delay_rate = 1.0 |
| 59 | t = np.clip(step / max_steps, 0, 1) |
| 60 | log_lerp = np.exp(np.log(lr_init) * (1 - t) + np.log(lr_final) * t) |
| 61 | return delay_rate * log_lerp |
| 62 | |
| 63 | return helper |
| 64 | |
| 65 | def strip_lowerdiag(L): |
| 66 | uncertainty = torch.zeros((L.shape[0], 6), dtype=torch.float, device="cuda") |