""" model that use cross attention to predict human + object """ import inspect import random from typing import Optional from torch import Tensor import torch import numpy as np from pytorch3d.structures import Pointclouds from pytorch3d.renderer import CamerasBase from diffusers.schedulers import DDPMScheduler, DDIMScheduler from .model_diff_data import ConditionalPCDiffusionBehave from .pvcnn.pvcnn_ho import PVCNN2HumObj import torch.nn.functional as F from pytorch3d.renderer import PerspectiveCameras from .model_utils import get_num_points from tqdm import tqdm class CrossAttenHODiffusionModel(ConditionalPCDiffusionBehave): def init_pcloud_model(self, kwargs, point_cloud_model, point_cloud_model_embed_dim): """use cross attention model""" if point_cloud_model == 'pvcnn': self.point_cloud_model = PVCNN2HumObj(embed_dim=point_cloud_model_embed_dim, num_classes=self.out_channels, extra_feature_channels=(self.in_channels - 3), voxel_resolution_multiplier=kwargs.get('voxel_resolution_multiplier', 1), attn_type=kwargs.get('attn_type', 'simple-cross'), attn_weight=kwargs.get("attn_weight", 1.0) ) else: raise ValueError(f"Unknown point cloud model {point_cloud_model}!") self.point_visible_test = kwargs.get("point_visible_test", 'single') # when doing point visibility test, use only human points or human + object? assert self.point_visible_test in ['single', 'combine'], f'invalide point visible test option {self.point_visible_test}' # print(f"Point visibility test is based on {self.point_visible_test} point clouds!") def forward_train( self, pc: Pointclouds, camera: Optional[CamerasBase], image_rgb: Optional[Tensor], mask: Optional[Tensor], return_intermediate_steps: bool = False, **kwargs ): "additional input (RGB, mask, camera, and pc) for object is read from kwargs" # assert not self.consistent_center assert not self.self_conditioning # Normalize colors and convert to tensor x0_h = self.point_cloud_to_tensor(pc, normalize=True, scale=True) # this will not pack the point colors x0_o = self.point_cloud_to_tensor(kwargs.get('pc_obj'), normalize=True, scale=True) B, N, D = x0_h.shape # Sample random noise noise = torch.randn_like(x0_h) if self.consistent_center: # modification suggested by https://arxiv.org/pdf/2308.07837.pdf noise = noise - torch.mean(noise, dim=1, keepdim=True) # Sample random timesteps for each point_cloud timestep = torch.randint(0, self.scheduler.num_train_timesteps, (B,), device=self.device, dtype=torch.long) # timestep = torch.randint(0, 1, (B,), # device=self.device, dtype=torch.long) # Add noise to points xt_h = self.scheduler.add_noise(x0_h, noise, timestep) xt_o = self.scheduler.add_noise(x0_o, noise, timestep) norm_parms = self.pack_norm_params(kwargs) # (2, B, 4) # get input conditioning x_t_input_h, x_t_input_o = self.get_image_conditioning(camera, image_rgb, kwargs, mask, norm_parms, timestep, xt_h, xt_o) # Diffusion prediction noise_pred_h, noise_pred_o = self.point_cloud_model(x_t_input_h, x_t_input_o, timestep, norm_parms) # Check if not noise_pred_h.shape == noise.shape: raise ValueError(f'{noise_pred_h.shape=} and {noise.shape=}') if not noise_pred_o.shape == noise.shape: raise ValueError(f'{noise_pred_o.shape=} and {noise.shape=}') # Loss loss_h = F.mse_loss(noise_pred_h, noise) loss_o = F.mse_loss(noise_pred_o, noise) loss = loss_h + loss_o # Whether to return intermediate steps if return_intermediate_steps: return loss, (x0_h, xt_h, noise, noise_pred_h) return loss, torch.tensor([loss_h, loss_o]) def get_image_conditioning(self, camera, image_rgb, kwargs, mask, norm_parms, timestep, xt_h, xt_o): """ compute image features for each point :param camera: :param image_rgb: :param kwargs: :param mask: :param norm_parms: :param timestep: :param xt_h: :param xt_o: :return: """ if self.point_visible_test == 'single': # Visibility test is down independently for human and object x_t_input_h = self.get_input_with_conditioning(xt_h, camera=camera, image_rgb=image_rgb, mask=mask, t=timestep) x_t_input_o = self.get_input_with_conditioning(xt_o, camera=kwargs.get('camera_obj'), image_rgb=kwargs.get('rgb_obj'), mask=kwargs.get('mask_obj'), t=timestep) elif self.point_visible_test == 'combine': # Combine human + object points to do visibility test and obtain features B, N = xt_h.shape[:2] # (B, N, 3) # for human: transform object points first to H+O space, then to human space xt_o_in_ho = xt_o * 2 * norm_parms[1, :, 3:].unsqueeze(1) + norm_parms[1, :, :3].unsqueeze(1) xt_o_in_hum = (xt_o_in_ho - norm_parms[0, :, :3].unsqueeze(1)) / (2 * norm_parms[0, :, 3:].unsqueeze(1)) # compute features for all points, take only first half feature for human x_t_input_h = self.get_input_with_conditioning(torch.cat([xt_h, xt_o_in_hum], 1), camera=camera, image_rgb=image_rgb, mask=mask, t=timestep)[:,:N] # for object: transform human points to H+O space, then to object space xt_h_in_ho = xt_h * 2 * norm_parms[0, :, 3:].unsqueeze(1) + norm_parms[0, :, :3].unsqueeze(1) xt_h_in_obj = (xt_h_in_ho - norm_parms[1, :, :3].unsqueeze(1)) / (2 * norm_parms[1, :, 3:].unsqueeze(1)) x_t_input_o = self.get_input_with_conditioning(torch.cat([xt_o, xt_h_in_obj], 1), camera=kwargs.get('camera_obj'), image_rgb=kwargs.get('rgb_obj'), mask=kwargs.get('mask_obj'), t=timestep)[:, :N] else: raise NotImplementedError return x_t_input_h, x_t_input_o def forward(self, batch, mode: str = 'train', **kwargs): """""" images = torch.stack(batch['images'], 0).to('cuda') masks = torch.stack(batch['masks'], 0).to('cuda') pc = self.get_input_pc(batch) camera = PerspectiveCameras( R=torch.stack(batch['R']), T=torch.stack(batch['T_hum']), K=torch.stack(batch['K_hum']), device='cuda', in_ndc=True ) grid_df = torch.stack(batch['grid_df'], 0).to('cuda') if 'grid_df' in batch else None num_points = kwargs.pop('num_points', get_num_points(pc)) rgb_obj = torch.stack(batch['images_obj'], 0).to('cuda') masks_obj = torch.stack(batch['masks_obj'], 0).to('cuda') pc_obj = Pointclouds([x.to('cuda') for x in batch['pclouds_obj']]) camera_obj = PerspectiveCameras( R=torch.stack(batch['R']), T=torch.stack(batch['T_obj']), K=torch.stack(batch['K_obj']), device='cuda', in_ndc=True ) # normalization parameters cent_hum = torch.stack(batch['cent_hum'], 0).to('cuda') cent_obj = torch.stack(batch['cent_obj'], 0).to('cuda') # B, 3 radius_hum = torch.stack(batch['radius_hum'], 0).to('cuda') # B, 1 radius_obj = torch.stack(batch['radius_obj'], 0).to('cuda') # print(batch['image_path']) if mode == 'train': return self.forward_train( pc=pc, camera=camera, image_rgb=images, mask=masks, grid_df=grid_df, rgb_obj=rgb_obj, mask_obj=masks_obj, pc_obj=pc_obj, camera_obj=camera_obj, cent_hum=cent_hum, cent_obj=cent_obj, radius_hum=radius_hum, radius_obj=radius_obj, ) elif mode == 'sample': # this use GT centers to do projection return self.forward_sample( num_points=num_points, camera=camera, image_rgb=images, mask=masks, gt_pc=pc, rgb_obj=rgb_obj, mask_obj=masks_obj, pc_obj=pc_obj, camera_obj=camera_obj, cent_hum=cent_hum, cent_obj=cent_obj, radius_hum=radius_hum, radius_obj=radius_obj, **kwargs) elif mode == 'interm-gt': return self.forward_sample( num_points=num_points, camera=camera, image_rgb=images, mask=masks, gt_pc=pc, rgb_obj=rgb_obj, mask_obj=masks_obj, pc_obj=pc_obj, camera_obj=camera_obj, cent_hum=cent_hum, cent_obj=cent_obj, radius_hum=radius_hum, radius_obj=radius_obj, sample_from_interm=True, **kwargs) elif mode == 'interm-pred': # use camera from predicted camera = PerspectiveCameras( R=torch.stack(batch['R']), T=torch.stack(batch['T_hum_scaled']), K=torch.stack(batch['K_hum']), device='cuda', in_ndc=True ) camera_obj = PerspectiveCameras( R=torch.stack(batch['R']), T=torch.stack(batch['T_obj_scaled']), K=torch.stack(batch['K_obj']), # the camera should be human/object specific!!! device='cuda', in_ndc=True ) # use pc from predicted pc = Pointclouds([x.to('cuda') for x in batch['pred_hum']]) pc_obj = Pointclouds([x.to('cuda') for x in batch['pred_obj']]) # use center and radius from predicted cent_hum = torch.stack(batch['cent_hum_pred'], 0).to('cuda') cent_obj = torch.stack(batch['cent_obj_pred'], 0).to('cuda') # B, 3 radius_hum = torch.stack(batch['radius_hum_pred'], 0).to('cuda') # B, 1 radius_obj = torch.stack(batch['radius_obj_pred'], 0).to('cuda') return self.forward_sample( num_points=num_points, camera=camera, image_rgb=images, mask=masks, gt_pc=pc, rgb_obj=rgb_obj, mask_obj=masks_obj, pc_obj=pc_obj, camera_obj=camera_obj, cent_hum=cent_hum, cent_obj=cent_obj, radius_hum=radius_hum, radius_obj=radius_obj, sample_from_interm=True, **kwargs) elif mode == 'interm-pred-ts': # use only estimate translation and scale, but sample from gaussian # this works, the camera is GT!!! pc = Pointclouds([x.to('cuda') for x in batch['pred_hum']]) pc_obj = Pointclouds([x.to('cuda') for x in batch['pred_obj']]) # use center and radius from predicted cent_hum = torch.stack(batch['cent_hum_pred'], 0).to('cuda') cent_obj = torch.stack(batch['cent_obj_pred'], 0).to('cuda') # B, 3 radius_hum = torch.stack(batch['radius_hum_pred'], 0).to('cuda') # B, 1 radius_obj = torch.stack(batch['radius_obj_pred'], 0).to('cuda') # print(cent_hum[0], radius_hum[0], cent_obj[0], radius_obj[0]) return self.forward_sample( num_points=num_points, camera=camera, image_rgb=images, mask=masks, gt_pc=pc, rgb_obj=rgb_obj, mask_obj=masks_obj, pc_obj=pc_obj, camera_obj=camera_obj, cent_hum=cent_hum, cent_obj=cent_obj, radius_hum=radius_hum, radius_obj=radius_obj, sample_from_interm=False, **kwargs) else: raise NotImplementedError def forward_sample( self, num_points: int, camera: Optional[CamerasBase], image_rgb: Optional[Tensor], mask: Optional[Tensor], # Optional overrides scheduler: Optional[str] = 'ddpm', # Inference parameters num_inference_steps: Optional[int] = 1000, eta: Optional[float] = 0.0, # for DDIM # Whether to return all the intermediate steps in generation return_sample_every_n_steps: int = -1, # Whether to disable tqdm disable_tqdm: bool = False, gt_pc: Pointclouds = None, **kwargs ): "use two models to run diffusion forward, and also use translation and scale to put them back" assert not self.self_conditioning # Get scheduler from mapping, or use self.scheduler if None scheduler = self.scheduler if scheduler is None else self.schedulers_map[scheduler] # Get the size of the noise N = num_points B = 1 if image_rgb is None else image_rgb.shape[0] D = self.get_x_T_channel() device = self.device if image_rgb is None else image_rgb.device # sample from full steps or only a few steps sample_from_interm = kwargs.get('sample_from_interm', False) interm_steps = kwargs.get('noise_step') if sample_from_interm else -1 xt_h = self.initialize_x_T(device, gt_pc, (B, N, D), interm_steps, scheduler) xt_o = self.initialize_x_T(device, kwargs.get('pc_obj', None), (B, N, D), interm_steps, scheduler) # the segmentation mask segm_mask = torch.zeros(B, 2*N, 1).to(device) segm_mask[:, :N] = 1.0 # Set timesteps extra_step_kwargs = self.setup_reverse_process(eta, num_inference_steps, scheduler) # Loop over timesteps all_outputs = [] return_all_outputs = (return_sample_every_n_steps > 0) progress_bar = tqdm(self.get_reverse_timesteps(scheduler, interm_steps), desc=f'Sampling ({xt_h.shape})', disable=disable_tqdm) # print("Camera T:", camera.T[0], camera.R[0]) # print("Camera_obj T:", kwargs.get('camera_obj').T[0], kwargs.get('camera_obj').R[0]) norm_parms = self.pack_norm_params(kwargs) for i, t in enumerate(progress_bar): x_t_input_h, x_t_input_o = self.get_image_conditioning(camera, image_rgb, kwargs, mask, norm_parms, t, xt_h, xt_o) # One reverse step with conditioning xt_h, xt_o = self.reverse_step(extra_step_kwargs, scheduler, t, torch.stack([xt_h, xt_o], 0), torch.stack([x_t_input_h, x_t_input_o], 0), **kwargs) # (B, N, D), D=3 if (return_all_outputs and (i % return_sample_every_n_steps == 0 or i == len(scheduler.timesteps) - 1)): # print(xt_h.shape, kwargs.get('cent_hum').shape, kwargs.get('radius_hum').shape) x_t = torch.cat([self.denormalize_pclouds(xt_h, kwargs.get('cent_hum'), kwargs.get('radius_hum')), self.denormalize_pclouds(xt_o, kwargs.get('cent_obj'), kwargs.get('radius_obj'))], 1) # print(x_t.shape, xt_o.shape) all_outputs.append(torch.cat([x_t, segm_mask], -1)) # print("Updating intermediate...") # Convert output back into a point cloud, undoing normalization and scaling x_t = torch.cat([self.denormalize_pclouds(xt_h, kwargs.get('cent_hum'), kwargs.get('radius_hum')), self.denormalize_pclouds(xt_o, kwargs.get('cent_obj'), kwargs.get('radius_obj'))], 1) x_t = torch.cat([x_t, segm_mask], -1) output = self.tensor_to_point_cloud(x_t, denormalize=False, unscale=False) # this convert the points back to original scale if return_all_outputs: all_outputs = torch.stack(all_outputs, dim=1) # (B, sample_steps, N, D) all_outputs = [self.tensor_to_point_cloud(o, denormalize=False, unscale=False) for o in all_outputs] return (output, all_outputs) if return_all_outputs else output def get_reverse_timesteps(self, scheduler, interm_steps: int): """ get the timesteps to run reverse diffusion :param scheduler: :param interm_steps: start from some intermediate steps, the step number is for DDPM scheduler if DDIM, will be recomputed accordingly :return: """ if isinstance(scheduler, DDPMScheduler): # DDPM, directly reverse N steps from interm_steps if interm_steps > 0: timesteps = torch.from_numpy(np.arange(0, interm_steps)[::-1].copy()).to(self.device) else: timesteps = scheduler.timesteps.to(self.device) elif isinstance(scheduler, DDIMScheduler): if interm_steps > 0: # compute a step ratio, and find the intermediate steps for DDIM step_ratio = scheduler.config.num_train_timesteps // scheduler.num_inference_steps timesteps = (np.arange(0, interm_steps, step_ratio)).round()[::-1].copy().astype(np.int64) timesteps = torch.from_numpy(timesteps).to(self.device) else: timesteps = scheduler.timesteps.to(self.device) else: raise NotImplementedError return timesteps def pack_norm_params(self, kwargs:dict, scale=True): scale_factor = self.scale_factor if scale else 1.0 hum = torch.cat([kwargs.get('cent_hum')*scale_factor, kwargs.get('radius_hum')], -1) obj = torch.cat([kwargs.get('cent_obj')*scale_factor, kwargs.get('radius_obj')], -1) return torch.stack([hum, obj], 0) # (2, B, 4) def reverse_step(self, extra_step_kwargs, scheduler, t, x_t, x_t_input, **kwargs): "x_t: (2, B, D, N), x_t_input: (2, B, D, N)" norm_parms = self.pack_norm_params(kwargs) # (2, B, 4) B = x_t.shape[1] # print(f"Step {t} Norm params:", norm_parms[:, 0, :]) noise_pred_h, noise_pred_o = self.point_cloud_model(x_t_input[0], x_t_input[1], t.reshape(1).expand(B), norm_parms) if self.consistent_center: assert self.dm_pred_type != 'sample', 'incompatible dm predition type!' noise_pred_h = noise_pred_h - torch.mean(noise_pred_h, dim=1, keepdim=True) noise_pred_o = noise_pred_o - torch.mean(noise_pred_o, dim=1, keepdim=True) xt_h = scheduler.step(noise_pred_h, t, x_t[0], **extra_step_kwargs).prev_sample xt_o = scheduler.step(noise_pred_o, t, x_t[1], **extra_step_kwargs).prev_sample if self.consistent_center: xt_h = xt_h - torch.mean(xt_h, dim=1, keepdim=True) xt_o = xt_o - torch.mean(xt_o, dim=1, keepdim=True) return xt_h, xt_o def denormalize_pclouds(self, x: Tensor, cent, radius, unscale: bool = True): """ first denormalize, then apply center and scale to original H+O coordinate :param x: :param cent: (B, 3) :param radius: (B, 1) :param unscale: :return: """ # denormalize: scale down. points = x[:, :, :3] / (self.scale_factor if unscale else 1) # translation and scale back to H+O coordinate points = points * 2 * radius.unsqueeze(-1) + cent.unsqueeze(1) return points def tensor_to_point_cloud(self, x: Tensor, /, denormalize: bool = False, unscale: bool = False): """ take binary into account :param self: :param x: (B, N, 4) :param denormalize: :param unscale: :return: """ points = x[:, :, :3] / (self.scale_factor if unscale else 1) if self.predict_color: colors = self.denormalize(x[:, :, 3:]) if denormalize else x[:, :, 3:] return Pointclouds(points=points, features=colors) else: assert x.shape[2] == 4 # add color to predicted binary labels is_hum = x[:, :, 3] > 0.5 features = [] for mask in is_hum: color = torch.zeros_like(x[0, :, :3]) + torch.tensor([0.5, 1.0, 0]).to(x.device) color[mask, :] = torch.tensor([0.05, 1.0, 1.0]).to(x.device) # human is light blue, object light green features.append(color) return Pointclouds(points=points, features=features)