-
Notifications
You must be signed in to change notification settings - Fork 54
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
xiuming
committed
Sep 5, 2021
1 parent
c1ba8a4
commit 0831990
Showing
6 changed files
with
391 additions
and
23 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,96 @@ | ||
| [DEFAULT] | ||
|
|
||
| # ====== Must-Have ====== | ||
| # These parameters are required by the pipeline, regardless of your custom code | ||
|
|
||
| # ------ Data ------ | ||
| dataset = mvs_shape | ||
| no_batch = True | ||
| # bs = 4 | ||
| cache = True | ||
|
|
||
| # ------ Model ------ | ||
| model = nerfactor | ||
|
|
||
| # ------ Optimization ------ | ||
| loss = l2 | ||
| lr = 5e-3 | ||
| lr_decay_steps = 500_000 | ||
| lr_decay_rate = 0.1 | ||
| clipnorm = -1 | ||
| clipvalue = -1 | ||
| epochs = 100 | ||
|
|
||
| # ------ Logging and Checkpointing ------ | ||
| ckpt_period = 10 | ||
| vali_period = 10 | ||
| vali_batches = 4 | ||
| vis_train_batches = 4 | ||
| keep_recent_epochs = -1 | ||
|
|
||
| # ------ IO ------ | ||
| overwrite = False | ||
| # The following two decide the output directory | ||
| outroot = /output/train/hotdog_2163_nerfactor_mvs/ | ||
| xname = lr{lr} | ||
|
|
||
|
|
||
| # ====== Custom ====== | ||
| # These parameters are whatever your custom dataset and model require | ||
|
|
||
| # ------ Data ------ | ||
| mvs_root = /output/surf_mvs/hotdog_2163/ | ||
| use_nerf_alpha = False | ||
| imh = 512 | ||
| light_h = 16 | ||
| near = 2 | ||
| far = 6 | ||
| ndc = False | ||
| white_bg = True | ||
|
|
||
| # ------ Model ------ | ||
| xyz_jitter_std = 0.01 | ||
| smooth_use_l1 = True | ||
| # DTU scenes have huge XYZs | ||
| xyz_scale = 1e-3 | ||
| # Shape | ||
| shape_mode = finetune | ||
| shape_model_ckpt = /output/train/hotdog_2163_shape_mvs/lr1e-2/checkpoints/ckpt-2 | ||
| nerf_shape_respect = 0.1 | ||
| normal_loss_weight = 0.1 | ||
| lvis_loss_weight = 0.1 | ||
| normal_smooth_weight = 0.05 | ||
| lvis_smooth_weight = 0.05 | ||
| # BRDF | ||
| albedo_slope = 0.77 | ||
| albedo_bias = 0.03 | ||
| pred_brdf = True | ||
| default_z = 0.1 | ||
| brdf_model_ckpt = /output/train/merl/lr1e-2/checkpoints/ckpt-50 | ||
| albedo_smooth_weight = 0.05 | ||
| brdf_smooth_weight = 0.01 | ||
| learned_brdf_scale = 1 | ||
| # Lighting | ||
| light_init_max = 1 | ||
| light_tv_weight = 5e-6 | ||
| light_achro_weight = 0 | ||
| # Rendering | ||
| linear2srgb = True | ||
| test_envmap_dir = /data/envmaps/for-render_h16/test/ | ||
|
|
||
| # ------ Network ------ | ||
| mlp_chunk = 65536 | ||
| mlp_width = 128 | ||
| mlp_depth = 4 | ||
| mlp_skip_at = 2 | ||
| # Positional encoding | ||
| pos_enc = True | ||
| n_freqs_xyz = 10 | ||
| n_freqs_ldir = 4 | ||
| n_freqs_vdir = 4 | ||
|
|
||
| # ------ Misc. ------ | ||
| # De facto training batch size: number of random rays per gradient step | ||
| n_rays_per_step = 1024 | ||
| # File viewer prefix, if any | ||
| viewer_prefix = http://vision38.csail.mit.edu |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,71 @@ | ||
| [DEFAULT] | ||
|
|
||
| # ====== Must-Have ====== | ||
| # These parameters are required by the pipeline, regardless of your custom code | ||
|
|
||
| # ------ Data ------ | ||
| dataset = mvs_shape | ||
| no_batch = True | ||
| # bs = 4 | ||
| cache = True | ||
|
|
||
| # ------ Model ------ | ||
| model = shape | ||
|
|
||
| # ------ Optimization ------ | ||
| loss = l2 | ||
| lr = 1e-2 | ||
| lr_decay_steps = 500_000 | ||
| lr_decay_rate = 0.1 | ||
| clipnorm = -1 | ||
| clipvalue = -1 | ||
| epochs = 200 | ||
|
|
||
| # ------ Logging and Checkpointing ------ | ||
| ckpt_period = 100 | ||
| vali_period = 100 | ||
| vali_batches = 4 | ||
| vis_train_batches = 4 | ||
| keep_recent_epochs = -1 | ||
|
|
||
| # ------ IO ------ | ||
| overwrite = False | ||
| # The following two decide the output directory | ||
| outroot = /output/train/hotdog_2163_shape_mvs | ||
| xname = lr{lr} | ||
|
|
||
|
|
||
| # ====== Custom ====== | ||
| # These parameters are whatever your custom dataset and model require | ||
|
|
||
| # ------ Data ------ | ||
| mvs_root = /output/surf_mvs/hotdog_2163 | ||
| imh = 512 | ||
| light_h = 16 | ||
| near = 2 | ||
| far = 6 | ||
| ndc = False | ||
| white_bg = True | ||
|
|
||
| # ------ Model ------ | ||
| xyz_jitter_std = 0.01 | ||
| smooth_use_l1 = True | ||
| # DTU scenes have huge XYZs | ||
| xyz_scale = 1e-3 | ||
| # De facto batch size: number of random rays per gradient step | ||
| n_rays_per_step = 1024 | ||
| normal_loss_weight = 1 | ||
| lvis_loss_weight = 1 | ||
| # Positional encoding | ||
| pos_enc = True | ||
| n_freqs_xyz = 10 | ||
| n_freqs_ldir = 4 | ||
| n_freqs_vdir = 4 | ||
|
|
||
| # ------ Network ------ | ||
| mlp_chunk = 65536 | ||
| mlp_width = 128 | ||
| mlp_depth = 4 | ||
| mlp_skip_at = 2 | ||
|
|
||
| viewer_prefix = http://vision38.csail.mit.edu |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,121 @@ | ||
| # Copyright 2021 Google LLC | ||
| # | ||
| # Licensed under the Apache License, Version 2.0 (the "License"); | ||
| # you may not use this file except in compliance with the License. | ||
| # You may obtain a copy of the License at | ||
| # | ||
| # http://www.apache.org/licenses/LICENSE-2.0 | ||
| # | ||
| # Unless required by applicable law or agreed to in writing, software | ||
| # distributed under the License is distributed on an "AS IS" BASIS, | ||
| # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| # See the License for the specific language governing permissions and | ||
| # limitations under the License. | ||
|
|
||
| # pylint: disable=invalid-unary-operand-type | ||
|
|
||
| from os.path import join | ||
| import numpy as np | ||
|
|
||
| from third_party.xiuminglib import xiuminglib as xm | ||
| from nerfactor.util import logging as logutil, io as ioutil, tensor as tutil | ||
| from nerfactor.datasets.nerf_shape import Dataset as BaseDataset | ||
|
|
||
|
|
||
| logger = logutil.Logger(loggee="datasets/mvs_shape") | ||
|
|
||
|
|
||
| class Dataset(BaseDataset): | ||
| def _glob(self): | ||
| mvs_root = self.config.get('DEFAULT', 'mvs_root') | ||
| # Glob metadata paths | ||
| mode_str = 'val' if self.mode == 'vali' else self.mode | ||
| if self.debug: | ||
| logger.warn("Globbing a single data file for faster debugging") | ||
| metadata_dir = join(mvs_root, '%s_000' % mode_str) | ||
| else: | ||
| metadata_dir = join(mvs_root, '%s_???' % mode_str) | ||
| # Include only cameras with all required buffers (depending on mode) | ||
| metadata_paths, incomplete_paths = [], [] | ||
| for metadata_path in xm.os.sortglob(metadata_dir, 'metadata.json'): | ||
| id_ = self._parse_id(metadata_path) | ||
| view_dir = join(mvs_root, id_) | ||
| lvis_path = join(view_dir, 'lvis.npy') | ||
| normal_path = join(view_dir, 'normal.npy') | ||
| xyz_path = join(view_dir, 'xyz.npy') | ||
| alpha_path = join(view_dir, 'alpha.png') | ||
| paths = { | ||
| 'xyz': xyz_path, 'normal': normal_path, 'lvis': lvis_path, | ||
| 'alpha': alpha_path} | ||
| if self.mode != 'test': | ||
| rgba_path = join(view_dir, 'rgba.png') | ||
| paths['rgba'] = rgba_path | ||
| if ioutil.all_exist(paths): | ||
| metadata_paths.append(metadata_path) | ||
| self.meta2buf[metadata_path] = paths | ||
| else: | ||
| incomplete_paths.append(metadata_path) | ||
| if incomplete_paths: | ||
| logger.warn(( | ||
| "Skipping\n\t%s\nbecause at least one of their paired " | ||
| "buffers doesn't exist"), incomplete_paths) | ||
| logger.info("Number of '%s' views: %d", self.mode, len(metadata_paths)) | ||
| return metadata_paths | ||
|
|
||
| # pylint: disable=arguments-differ | ||
| def _load_data(self, metadata_path): | ||
| imh = self.config.getint('DEFAULT', 'imh') | ||
| use_nerf_alpha = self.config.getboolean('DEFAULT', 'use_nerf_alpha') | ||
| metadata_path = tutil.eager_tensor_to_str(metadata_path) | ||
| id_ = self._parse_id(metadata_path) | ||
| # Rays | ||
| metadata = ioutil.read_json(metadata_path) | ||
| h, w = metadata['imh'], metadata['imw'] | ||
| cam_loc = np.array(metadata['cam_loc']) | ||
| rayo = np.tile(cam_loc[None, None, :], (h, w, 1)) | ||
| rayo = rayo.astype(np.float32) | ||
| rayd = np.zeros_like(rayo) # dummy | ||
| # Load precomputed shape properties | ||
| paths = self.meta2buf[metadata_path] | ||
| xyz = ioutil.load_np(paths['xyz']) | ||
| normal = ioutil.load_np(paths['normal']) | ||
| if self.debug: | ||
| logger.warn("Faking light visibility for faster debugging") | ||
| lvis = 0.5 * np.ones(normal.shape[:2] + (512,), dtype=np.float32) | ||
| else: | ||
| lvis = ioutil.load_np(paths['lvis']) | ||
| # RGB and alpha, depending on the mode | ||
| if self.mode == 'test': | ||
| # No RGBA, so estimated alpha and placeholder RGB | ||
| alpha = xm.io.img.load(paths['alpha']) | ||
| alpha = xm.img.normalize_uint(alpha) | ||
| rgb = np.zeros_like(xyz) | ||
| else: | ||
| # Training or validation, where each camera has a paired image | ||
| rgba = xm.io.img.load(paths['rgba']) | ||
| assert rgba.ndim == 3 and rgba.shape[2] == 4, \ | ||
| "Input image is not RGBA" | ||
| rgba = xm.img.normalize_uint(rgba) | ||
| rgb = rgba[:, :, :3] | ||
| if use_nerf_alpha: # useful for real scenes | ||
| alpha = xm.io.img.load(paths['alpha']) | ||
| alpha = xm.img.normalize_uint(alpha) | ||
| else: | ||
| alpha = rgba[:, :, 3] # ground-truth alpha | ||
| # Resize | ||
| if imh != xyz.shape[0]: | ||
| xyz = xm.img.resize(xyz, new_h=imh) | ||
| normal = xm.img.resize(normal, new_h=imh) | ||
| lvis = xm.img.resize(lvis, new_h=imh) | ||
| alpha = xm.img.resize(alpha, new_h=imh) | ||
| rgb = xm.img.resize(rgb, new_h=imh) | ||
| # Make sure there's no XYZ coinciding with camera (caused by occupancy | ||
| # accumulating to 0) | ||
| assert not np.isclose(xyz, rayo).all(axis=2).any(), \ | ||
| "Found XYZs coinciding with the camera" | ||
| # Re-normalize normals and clip light visibility before returning | ||
| normal = xm.linalg.normalize(normal, axis=2) | ||
| assert np.isclose(np.linalg.norm(normal, axis=2), 1).all(), \ | ||
| "Found normals with a norm far away from 1" | ||
| lvis = np.clip(lvis, 0, 1) | ||
| return id_, rayo, rayd, rgb, alpha, xyz, normal, lvis |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.