File size: 10,566 Bytes
5ed9923
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
import os

from plyfile import PlyData, PlyElement
from scipy.spatial.transform import Rotation
import einops
import numpy as np
import torch
import torchvision
import trimesh
import lightning as L

import utils.loss_mask as loss_mask
from src.mast3r_src.dust3r.dust3r.viz import OPENGL, pts3d_to_trimesh, cat_meshes


class SaveBatchData(L.Callback):
    '''A Lightning callback that occasionally saves batch inputs and outputs to disk.
    It is not critical to the training process, and can be disabled if unwanted.'''

    def __init__(self, save_dir, train_save_interval=100, val_save_interval=100, test_save_interval=100):
        self.save_dir = save_dir
        self.train_save_interval = train_save_interval
        self.val_save_interval = val_save_interval
        self.test_save_interval = test_save_interval

    def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx):
        if batch_idx % self.train_save_interval == 0 and trainer.global_rank == 0:
            self.save_batch_data('train', trainer, pl_module, batch, batch_idx)

    def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx):
        if batch_idx % self.val_save_interval == 0 and trainer.global_rank == 0:
            self.save_batch_data('val', trainer, pl_module, batch, batch_idx)

    def on_test_batch_end(self, trainer, pl_module, outputs, batch, batch_idx):
        if batch_idx % self.test_save_interval == 0 and trainer.global_rank == 0:
            self.save_batch_data('test', trainer, pl_module, batch, batch_idx)

    def save_batch_data(self, prefix, trainer, pl_module, batch, batch_idx):

        print(f'Saving {prefix} data at epoch {trainer.current_epoch} and batch {batch_idx}')

        # Run the batch through the model again
        _, _, h, w = batch["context"][0]["img"].shape
        view1, view2 = batch['context']
        pred1, pred2 = pl_module.forward(view1, view2)
        color, depth = pl_module.decoder(batch, pred1, pred2, (h, w))
        mask = loss_mask.calculate_loss_mask(batch)

        # Save the data
        save_dir = os.path.join(
            self.save_dir,
            f"{prefix}_epoch_{trainer.current_epoch}_batch_{batch_idx}"
        )
        log_batch_files(batch, color, depth, mask, view1, view2, pred1, pred2, save_dir)


def save_as_ply(pred1, pred2, save_path):
    """Save the 3D Gaussians as a point cloud in the PLY format.
    Adapted loosely from PixelSplat"""

    def construct_list_of_attributes(num_rest: int) -> list[str]:
        '''Construct a list of attributes for the PLY file format. This
        corresponds to the attributes used by online readers, such as
        https://niujinshuchong.github.io/mip-splatting-demo/index.html'''
        attributes = ["x", "y", "z", "nx", "ny", "nz"]
        for i in range(3):
            attributes.append(f"f_dc_{i}")
        for i in range(num_rest):
            attributes.append(f"f_rest_{i}")
        attributes.append("opacity")
        for i in range(3):
            attributes.append(f"scale_{i}")
        for i in range(4):
            attributes.append(f"rot_{i}")
        return attributes

    def covariance_to_quaternion_and_scale(covariance):
        '''Convert the covariance matrix to a four dimensional quaternion and
        a three dimensional scale vector'''

        # Perform singular value decomposition
        U, S, V = torch.linalg.svd(covariance)

        # The scale factors are the square roots of the eigenvalues
        scale = torch.sqrt(S)
        scale = scale.detach().cpu().numpy()

        # The rotation matrix is U*Vt
        rotation_matrix = torch.bmm(U, V.transpose(-2, -1))
        rotation_matrix_np = rotation_matrix.detach().cpu().numpy()

        # Use scipy to convert the rotation matrix to a quaternion
        rotation = Rotation.from_matrix(rotation_matrix_np)
        quaternion = rotation.as_quat()

        return quaternion, scale

    # Collect the Gaussian parameters
    means = torch.stack([pred1["pts3d"], pred2["pts3d_in_other_view"]], dim=1)
    covariances = torch.stack([pred1["covariances"], pred2["covariances"]], dim=1)
    harmonics = torch.stack([pred1["sh"], pred2["sh"]], dim=1)[..., 0]  # Only use the first harmonic
    opacities = torch.stack([pred1["opacities"], pred2["opacities"]], dim=1)

    # Rearrange the tensors to the correct shape
    means = einops.rearrange(means[0], "view h w xyz -> (view h w) xyz").detach().cpu().numpy()
    covariances = einops.rearrange(covariances[0], "v h w i j -> (v h w) i j")
    harmonics = einops.rearrange(harmonics[0], "view h w xyz -> (view h w) xyz").detach().cpu().numpy()
    opacities = einops.rearrange(opacities[0], "view h w xyz -> (view h w) xyz").detach().cpu().numpy()

    # Convert the covariance matrices to quaternions and scales
    rotations, scales = covariance_to_quaternion_and_scale(covariances)

    # Construct the attributes
    rest = np.zeros_like(means)
    attributes = np.concatenate((means, rest, harmonics, opacities, np.log(scales), rotations), axis=-1)
    dtype_full = [(attribute, "f4") for attribute in construct_list_of_attributes(0)]
    elements = np.empty(attributes.shape[0], dtype=dtype_full)
    elements[:] = list(map(tuple, attributes))

    # Save the point cloud
    point_cloud = PlyElement.describe(elements, "vertex")
    scene = PlyData([point_cloud])
    scene.write(save_path)


def save_3d(view1, view2, pred1, pred2, save_dir, as_pointcloud=True, all_points=True):
    """Save the 3D points as a point cloud or as a mesh. Adapted from DUSt3R"""

    os.makedirs(save_dir, exist_ok=True)
    batch_size = pred1["pts3d"].shape[0]
    views = [view1, view2]

    for b in range(batch_size):

        pts3d = [pred1["pts3d"][b].cpu().numpy()] + [pred2["pts3d_in_other_view"][b].cpu().numpy()]
        imgs = [einops.rearrange(view["original_img"][b], "c h w -> h w c").cpu().numpy() for view in views]
        mask = [view["valid_mask"][b].cpu().numpy() for view in views]

        # Treat all pixels as valid, because we want to render the entire viewpoint
        if all_points:
            mask = [np.ones_like(m) for m in mask]

        # Construct the scene from the 3D points as a point cloud or as a mesh
        scene = trimesh.Scene()
        if as_pointcloud:
            pts = np.concatenate([p[m] for p, m in zip(pts3d, mask)])
            col = np.concatenate([p[m] for p, m in zip(imgs, mask)])
            pct = trimesh.PointCloud(pts.reshape(-1, 3), colors=col.reshape(-1, 3))
            scene.add_geometry(pct)
            save_path = os.path.join(save_dir, f"{b}.ply")
        else:
            meshes = []
            for i in range(len(imgs)):
                meshes.append(pts3d_to_trimesh(imgs[i], pts3d[i], mask[i]))
            mesh = trimesh.Trimesh(**cat_meshes(meshes))
            scene.add_geometry(mesh)
            save_path = os.path.join(save_dir, f"{b}.glb")

        # Save the scene
        scene.export(file_obj=save_path)


@torch.no_grad()
def log_batch_files(batch, color, depth, mask, view1, view2, pred1, pred2, save_dir, should_save_3d=False):
    '''Save all the relevant debug files for a batch'''

    os.makedirs(save_dir, exist_ok=True)

    # Save the 3D Gaussians as a .ply file
    save_as_ply(pred1, pred2, os.path.join(save_dir, f"gaussians.ply"))

    # Save the 3D points as a point cloud and as a mesh (disabled)
    if should_save_3d:
        save_3d(view1, view2, pred1, pred2, os.path.join(save_dir, "3d_mesh"), as_pointcloud=False)
        save_3d(view1, view2, pred1, pred2, os.path.join(save_dir, "3d_pointcloud"), as_pointcloud=True)

    # Save the color, depth and valid masks for the input context images
    context_images = torch.stack([view["img"] for view in batch["context"]], dim=1)
    context_original_images = torch.stack([view["original_img"] for view in batch["context"]], dim=1)
    context_depthmaps = torch.stack([view["depthmap"] for view in batch["context"]], dim=1)
    context_valid_masks = torch.stack([view["valid_mask"] for view in batch["context"]], dim=1)
    for b in range(min(context_images.shape[0], 4)):
        torchvision.utils.save_image(context_images[b], os.path.join(save_dir, f"sample_{b}_img_context.jpg"))
        torchvision.utils.save_image(context_original_images[b], os.path.join(save_dir, f"sample_{b}_original_img_context.jpg"))
        torchvision.utils.save_image(context_depthmaps[b, :, None, ...], os.path.join(save_dir, f"sample_{b}_depthmap.jpg"), normalize=True)
        torchvision.utils.save_image(context_valid_masks[b, :, None, ...].float(), os.path.join(save_dir, f"sample_{b}_valid_mask_context.jpg"), normalize=True)

    # Save the color and depth images for the target images
    target_original_images = torch.stack([view["original_img"] for view in batch["target"]], dim=1)
    target_depthmaps = torch.stack([view["depthmap"] for view in batch["target"]], dim=1)
    context_valid_masks = torch.stack([view["valid_mask"] for view in batch["context"]], dim=1)
    for b in range(min(target_original_images.shape[0], 4)):
        torchvision.utils.save_image(target_original_images[b], os.path.join(save_dir, f"sample_{b}_original_img_target.jpg"))
        torchvision.utils.save_image(target_depthmaps[b, :, None, ...], os.path.join(save_dir, f"sample_{b}_depthmap_target.jpg"), normalize=True)

    # Save the rendered images and depths
    for b in range(min(color.shape[0], 4)):
        torchvision.utils.save_image(color[b, ...], os.path.join(save_dir, f"sample_{b}_rendered_color.jpg"))
    if depth is not None:
        for b in range(min(color.shape[0], 4)):
            torchvision.utils.save_image(depth[b, :, None, ...], os.path.join(save_dir, f"sample_{b}_rendered_depth.jpg"), normalize=True)

    # Save the loss masks
    for b in range(min(mask.shape[0], 4)):
        torchvision.utils.save_image(mask[b, :, None, ...].float(), os.path.join(save_dir, f"sample_{b}_loss_mask.jpg"), normalize=True)

    # Save the masked target and rendered images
    target_original_images = torch.stack([view["original_img"] for view in batch["target"]], dim=1)
    masked_target_original_images = target_original_images * mask[..., None, :, :]
    masked_predictions = color * mask[..., None, :, :]
    for b in range(min(target_original_images.shape[0], 4)):
        torchvision.utils.save_image(masked_target_original_images[b], os.path.join(save_dir, f"sample_{b}_masked_original_img_target.jpg"))
        torchvision.utils.save_image(masked_predictions[b], os.path.join(save_dir, f"sample_{b}_masked_rendered_color.jpg"))