File size: 8,889 Bytes
cc0c59d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
214
215
216
217
218
219
220
221
222
223
224
from typing import *
import copy
import torch
from torch.utils.data import DataLoader
import numpy as np
from easydict import EasyDict as edict
import utils3d.torch

from ..basic import BasicTrainer
from ...representations import Strivec
from ...renderers import OctreeRenderer
from ...modules.sparse import SparseTensor
from ...utils.loss_utils import l1_loss, l2_loss, ssim, lpips


class SLatVaeRadianceFieldDecoderTrainer(BasicTrainer):
    """
    Trainer for structured latent VAE Radiance Field Decoder.
    
    Args:
        models (dict[str, nn.Module]): Models to train.
        dataset (torch.utils.data.Dataset): Dataset.
        output_dir (str): Output directory.
        load_dir (str): Load directory.
        step (int): Step to load.
        batch_size (int): Batch size.
        batch_size_per_gpu (int): Batch size per GPU. If specified, batch_size will be ignored.
        batch_split (int): Split batch with gradient accumulation.
        max_steps (int): Max steps.
        optimizer (dict): Optimizer config.
        lr_scheduler (dict): Learning rate scheduler config.
        elastic (dict): Elastic memory management config.
        grad_clip (float or dict): Gradient clip config.
        ema_rate (float or list): Exponential moving average rates.
        fp16_mode (str): FP16 mode.
            - None: No FP16.
            - 'inflat_all': Hold a inflated fp32 master param for all params.
            - 'amp': Automatic mixed precision.
        fp16_scale_growth (float): Scale growth for FP16 gradient backpropagation.
        finetune_ckpt (dict): Finetune checkpoint.
        log_param_stats (bool): Log parameter stats.
        i_print (int): Print interval.
        i_log (int): Log interval.
        i_sample (int): Sample interval.
        i_save (int): Save interval.
        i_ddpcheck (int): DDP check interval.
        
        loss_type (str): Loss type. Can be 'l1', 'l2'
        lambda_ssim (float): SSIM loss weight.
        lambda_lpips (float): LPIPS loss weight.
    """
    
    def __init__(
        self,
        *args,
        loss_type: str = 'l1',
        lambda_ssim: float = 0.2,
        lambda_lpips: float = 0.2,
        **kwargs
    ):
        super().__init__(*args, **kwargs)
        self.loss_type = loss_type
        self.lambda_ssim = lambda_ssim
        self.lambda_lpips = lambda_lpips
        
        self._init_renderer()
        
    def _init_renderer(self):
        rendering_options = {"near" : 0.8,
                             "far" : 1.6,
                             "bg_color" : 'random'}
        self.renderer = OctreeRenderer(rendering_options)
        self.renderer.pipe.primitive = 'trivec'
        
    def _render_batch(self, reps: List[Strivec], extrinsics: torch.Tensor, intrinsics: torch.Tensor) -> torch.Tensor:
        """
        Render a batch of representations.

        Args:
            reps: The dictionary of lists of representations.
            extrinsics: The [N x 4 x 4] tensor of extrinsics.
            intrinsics: The [N x 3 x 3] tensor of intrinsics.
        """
        ret = None
        for i, representation in enumerate(reps):
            render_pack = self.renderer.render(representation, extrinsics[i], intrinsics[i])
            if ret is None:
                ret = {k: [] for k in list(render_pack.keys()) + ['bg_color']}
            for k, v in render_pack.items():
                ret[k].append(v)
            ret['bg_color'].append(self.renderer.bg_color)
        for k, v in ret.items():
            ret[k] = torch.stack(v, dim=0) 
        return ret
    
    def training_losses(
        self,
        latents: SparseTensor,
        image: torch.Tensor,
        alpha: torch.Tensor,
        extrinsics: torch.Tensor,
        intrinsics: torch.Tensor,
        return_aux: bool = False,
        **kwargs
    ) -> Tuple[Dict, Dict]:
        """
        Compute training losses.

        Args:
            latents: The [N x * x C] sparse latents
            image: The [N x 3 x H x W] tensor of images.
            alpha: The [N x H x W] tensor of alpha channels.
            extrinsics: The [N x 4 x 4] tensor of extrinsics.
            intrinsics: The [N x 3 x 3] tensor of intrinsics.
            return_aux: Whether to return auxiliary information.

        Returns:
            a dict with the key "loss" containing a scalar tensor.
            may also contain other keys for different terms.
        """
        reps = self.training_models['decoder'](latents)
        self.renderer.rendering_options.resolution = image.shape[-1]
        render_results = self._render_batch(reps, extrinsics, intrinsics)
        
        terms = edict(loss = 0.0, rec = 0.0)
        
        rec_image = render_results['color']
        gt_image = image * alpha[:, None] + (1 - alpha[:, None]) * render_results['bg_color'][..., None, None]
                
        if self.loss_type == 'l1':
            terms["l1"] = l1_loss(rec_image, gt_image)
            terms["rec"] = terms["rec"] + terms["l1"]
        elif self.loss_type == 'l2':
            terms["l2"] = l2_loss(rec_image, gt_image)
            terms["rec"] = terms["rec"] + terms["l2"]
        else:
            raise ValueError(f"Invalid loss type: {self.loss_type}")
        if self.lambda_ssim > 0:
            terms["ssim"] = 1 - ssim(rec_image, gt_image)
            terms["rec"] = terms["rec"] + self.lambda_ssim * terms["ssim"]
        if self.lambda_lpips > 0:
            terms["lpips"] = lpips(rec_image, gt_image)
            terms["rec"] = terms["rec"] + self.lambda_lpips * terms["lpips"]
        terms["loss"] = terms["loss"] + terms["rec"]
                
        if return_aux:
            return terms, {}, {'rec_image': rec_image, 'gt_image': gt_image}       
        return terms, {}
    
    @torch.no_grad()
    def run_snapshot(
        self,
        num_samples: int,
        batch_size: int,
        verbose: bool = False,
    ) -> Dict:
        dataloader = DataLoader(
            copy.deepcopy(self.dataset),
            batch_size=batch_size,
            shuffle=True,
            num_workers=0,
            collate_fn=self.dataset.collate_fn if hasattr(self.dataset, 'collate_fn') else None,
        )

        # inference
        ret_dict = {}
        gt_images = []
        exts = []
        ints = []
        reps = []
        for i in range(0, num_samples, batch_size):
            batch = min(batch_size, num_samples - i)
            data = next(iter(dataloader))
            args = {k: v[:batch].cuda() for k, v in data.items()}
            gt_images.append(args['image'] * args['alpha'][:, None])
            exts.append(args['extrinsics'])
            ints.append(args['intrinsics'])
            reps.extend(self.models['decoder'](args['latents']))
        gt_images = torch.cat(gt_images, dim=0)
        ret_dict.update({f'gt_image': {'value': gt_images, 'type': 'image'}})
        
        # render single view
        exts = torch.cat(exts, dim=0)
        ints = torch.cat(ints, dim=0)
        self.renderer.rendering_options.bg_color = (0, 0, 0)
        self.renderer.rendering_options.resolution = gt_images.shape[-1]
        render_results = self._render_batch(reps, exts, ints)
        ret_dict.update({f'rec_image': {'value': render_results['color'], 'type': 'image'}})

        # render multiview
        self.renderer.rendering_options.resolution = 512
        ## Build camera
        yaws = [0, np.pi / 2, np.pi, 3 * np.pi / 2]
        yaws_offset = np.random.uniform(-np.pi / 4, np.pi / 4)
        yaws = [y + yaws_offset for y in yaws]
        pitch = [np.random.uniform(-np.pi / 4, np.pi / 4) for _ in range(4)]

        ## render each view
        miltiview_images = []
        for yaw, pitch in zip(yaws, pitch):
            orig = torch.tensor([
                np.sin(yaw) * np.cos(pitch),
                np.cos(yaw) * np.cos(pitch),
                np.sin(pitch),
            ]).float().cuda() * 2
            fov = torch.deg2rad(torch.tensor(30)).cuda()
            extrinsics = utils3d.torch.extrinsics_look_at(orig, torch.tensor([0, 0, 0]).float().cuda(), torch.tensor([0, 0, 1]).float().cuda())
            intrinsics = utils3d.torch.intrinsics_from_fov_xy(fov, fov)
            extrinsics = extrinsics.unsqueeze(0).expand(num_samples, -1, -1)
            intrinsics = intrinsics.unsqueeze(0).expand(num_samples, -1, -1)
            render_results = self._render_batch(reps, extrinsics, intrinsics)
            miltiview_images.append(render_results['color'])

        ## Concatenate views
        miltiview_images = torch.cat([
            torch.cat(miltiview_images[:2], dim=-2),
            torch.cat(miltiview_images[2:], dim=-2),
        ], dim=-1)
        ret_dict.update({f'miltiview_image': {'value': miltiview_images, 'type': 'image'}})

        self.renderer.rendering_options.bg_color = 'random'
                            
        return ret_dict