import os import sys sys.path.append(os.path.dirname(os.path.abspath(__file__)) + '/..') import importlib import argparse parser = argparse.ArgumentParser(description='') parser.add_argument('--name', type=str) parser.add_argument('--config', type=str) parser.add_argument('--debug', action='store_true') args = vars(parser.parse_args()) if args['debug']: from configs.iccvw.cvae_norm_5 import config else: config_module = importlib.import_module(args['config']) config = config_module.config os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" os.environ["CUDA_VISIBLE_DEVICES"] = config['gpu'] if args['debug']: os.environ["CUDA_LAUNCH_BLOCKING"] = '1' import time import argparse import wandb import tqdm import random import numpy as np import kaleido import plotly.graph_objects as go import plotly.express as ex from PIL import Image if config['is_skirt']: from losses.losses import edge_loss_skirt as edge_loss else: from losses.losses import edge_loss import torch import torch.nn as nn import torch.optim as optim from data.dataset import Dataset from models.cond_encoder import AutoEncoder from utils.checkpoints import CheckpointSaver, load_model from utils.utils import InfiniteDataLoader as DataLoader from utils.postprocess import fix_collision from losses.losses import collision_loss, normal_loss from pytorch_msssim import MS_SSIM from models.cvae_norm_model import CVAENorm from smpl.smpl_torch import SMPL from custom_utils.custom_utils import UV2Mesh, Accumulator, plot_3d, get_summary_html, maskFacesBatch, get_summary_str, plot_3d_collision def init_wandb(): wandb.login(host='https://wandb2.lnrweb.com', key='local-2a78b434e9bc51e9f12625f8321a9e3ee34af16d', relogin=True, force=True) run = wandb.init(project='iccvw', name=config['name'], notes=config['notes'], dir='/data-local/127/temp', tags=config['tags']) artifact = wandb.Artifact('code', type='code') artifact.add_file('scripts/cvae_norm.py') artifact.add_file('losses/losses.py') artifact.add_file('models/base.py') artifact.add_file('models/cond_encoder.py') artifact.add_file('models/convnext.py') artifact.add_file('models/cvae_norm_model.py') artifact.add_file('data/dataset.py') config_file_names = [key for key in sys.modules.keys() if key.startswith('configs.')] for config_file_name in config_file_names: file_name = '/'.join(config_file_name.split('.')) + '.py' if os.path.exists(file_name): artifact.add_file(file_name) run.log_artifact(artifact) uv2mesh = UV2Mesh(res=256, is_skirt=config['is_skirt'], is_smpl_orig=config['is_cape']) def kl_loss(mean: torch.Tensor, logvar: torch.Tensor) -> torch.Tensor: kl_loss = 1 + logvar - torch.square(mean) - torch.exp(logvar) kl_loss = torch.mean(kl_loss) kl_loss *= -0.5 return kl_loss def train_step(batch): model.train() pred, mean, logvar = model([batch['uv_normals'], batch['uv_static'], batch['uv_body_posed'], batch['uv_mask']]) pred = pred * 1.4 mesh_pred = uv2mesh(pred) l2_loss = torch.mean(1e3 * torch.mean(torch.sum(((batch['uv_normals'] - pred) ** 2), axis=-1), axis=[1, 2])) l2_mesh_loss = torch.mean(1e3 * torch.mean(torch.sum(batch['mesh_mask'][:,:,None] * (batch['mesh_normals'] - mesh_pred) ** 2, axis=-1), axis=-1)) ms_ssim = torch.mean(_MSSSIM(pred + 1.4, batch['uv_normals'] + 1.4)) kl = kl_loss(mean, logvar) reconstr_loss = l2_loss + l2_mesh_loss + (1 - ms_ssim) loss: torch.Tensor = reconstr_loss + g_balancing_coeff * kl optimizer.zero_grad() loss.backward() optimizer.step() model.eval() with torch.inference_mode(): mean = torch.mean(mean) logvar = torch.mean(logvar) metric = torch.sum(torch.sqrt(torch.sum(batch['mesh_mask'][:,:,None] * (batch['mesh_normals'] - mesh_pred.detach()) ** 2, axis=-1))) / torch.sum(batch['mesh_mask']) return { 'train/loss': loss, 'train/euclid_error': metric, 'train/reconstr_loss': reconstr_loss, 'train/mean': mean, 'train/logvar': logvar, 'train/kl_loss': kl, 'train/var': np.e ** logvar, 'train/l2_loss': l2_loss, 'train/l2_mesh_loss': l2_mesh_loss, 'train/ms_ssim': ms_ssim, } def valid_step(batch): model.eval() with torch.inference_mode(): pred, mean, logvar = model([batch['uv_normals'], batch['uv_static'], batch['uv_body_posed'], batch['uv_mask']]) pred = pred * 1.4 mesh_pred = uv2mesh(pred) l2_loss = torch.mean(1e3 * torch.mean(torch.sum(((batch['uv_normals'] - pred) ** 2), axis=-1), axis=[1, 2])) l2_mesh_loss = torch.mean(1e3 * torch.mean(torch.sum(batch['mesh_mask'][:,:,None] * (mesh_pred - batch['mesh_normals']) ** 2, axis=-1), axis=-1)) ms_ssim = torch.mean(_MSSSIM(pred + 1.4, batch['uv_static'] + 1.4)) kl = kl_loss(mean, logvar) reconstr_loss = l2_loss + l2_mesh_loss + (1 - ms_ssim) loss = reconstr_loss + g_balancing_coeff * kl mean = torch.mean(mean) logvar = torch.mean(logvar) metric = torch.sum(torch.sqrt(torch.sum(batch['mesh_mask'][:,:,None] * (batch['mesh_normals'] - mesh_pred.detach()) ** 2, axis=-1))) / torch.sum(batch['mesh_mask']) return { 'valid/loss': loss, 'valid/euclid_error': metric, 'valid/reconstr_loss': reconstr_loss, 'valid/l2_mesh_loss': l2_mesh_loss, 'valid/mean': mean, 'valid/logvar': logvar, 'valid/kl_loss': kl, 'valid/var': np.e ** logvar, 'valid/ms_ssim': ms_ssim, } def plot_everything(batch: dict, type: str): model.eval() with torch.inference_mode(): pred, mean, logvar = model([batch['uv_normals'], batch['uv_static'], batch['uv_body_posed'], batch['uv_mask']]) pred = pred * 1.4 # mesh_pred = uv2mesh(pred).detach().cpu().numpy() uv_static_img = torch.permute(batch['uv_static'], [0,2,3,1]) uv_body_posed_img = torch.permute(batch['uv_body_posed'], [0,2,3,1]) orig_img = (torch.permute(batch['uv_normals'], [0,2,3,1]) + 1.4) / 2.8 pred_img = torch.permute(pred, [0,2,3,1]) img = torch.concat((uv_static_img, uv_body_posed_img, orig_img, pred_img), axis=2) imgs = [wandb.Image((img[i].cpu().numpy() + 1) / 2.0) for i in range(4)] # F = np.load("smpl/smpl_faces" + skirt_modifier + ("_orig" if config['is_cape'] else "") + ".npy") # mask = batch['mesh_mask'].cpu().numpy().astype(bool) mean_hist = wandb.Histogram(sequence=mean.cpu().numpy().mean(axis=0)) logvar_hist = wandb.Histogram(sequence=logvar.cpu().numpy().mean(axis=0)) return { **{ type + '/images': imgs, type + '/mean_hist': mean_hist, type + '/logvar_hist': logvar_hist } } def train(): global g_balancing_coeff global step valid_iterator = iter(valid_dataloader) for batch in train_dataloader: train_loss = train_step(batch) train_accumulator.update(train_loss) if step % 25 == 0: data_valid = next(valid_iterator) valid_loss = valid_step(data_valid) valid_accumulator.update(valid_loss) if step % 1000 == 0 or (step < 1000 and step % 100 == 0): to_log = dict({ **train_loss, **valid_loss, **{ 'current_step': step } }) if step % 5000 == 0 or (step < 5000 and step % 1000 == 0): train_figs = plot_everything(batch, 'train') valid_figs = plot_everything(data_valid, 'valid') to_log = { **to_log, **train_figs, **valid_figs } if step == 0: to_log = { **to_log, **{ 'model/summary': wandb.Html(get_summary_html(model)) } } if not args['debug']: wandb.log(to_log, step=step) train_accumulator.reset() valid_accumulator.reset() print(step) if train_loss['train/euclid_error'] < config['balancing_activation_limit']: g_balancing_coeff += 1 / 2000 if step % 20000 == 0: ckpt_saver(model, step, optimizer, None) step += 1 def sample(): model.eval() load_model(model, f'checkpoints/iccvw/{config["name"]}') count = 0 for batch in train_dataloader: uv_static = batch['uv_static'] uv_body_posed = batch['uv_body_posed'] uv_mask = batch['uv_mask'] batch_size = uv_static.size(0) for i in range(batch_size): uv_static_tiled = uv_static[i][None,:,:,:].repeat(4, 1, 1, 1) uv_body_posed_tiled = uv_body_posed[i][None,:,:,:].repeat(4, 1, 1, 1) uv_mask_tiled = uv_mask[i][None,:,:].repeat(4, 1, 1) with torch.inference_mode(): pred = model.sample(uv_static_tiled, uv_body_posed_tiled, uv_mask_tiled) pred = pred * 1.4 pred = pred.cpu().numpy().transpose(0, 2, 3, 1) pred = (pred + 1.4) / 2.8 pred = pred * uv_mask_tiled[:,:,:,None].cpu().numpy() gt = (batch['uv_normals'][i].cpu().numpy().transpose(1, 2, 0) + 1.4) / 2.8 * uv_mask_tiled[0,:,:,None].cpu().numpy() img = np.concatenate(pred, axis=1) img = np.concatenate([gt, img], axis=1) # flip image along its shorter edge img = img[::-1,:,:] fig = ex.imshow(img) fig.update_layout(margin=dict(l=0, r=0, b=0, t=0, pad=0), width=img.shape[1], height=img.shape[0]) fig.update_xaxes(visible=False) fig.update_yaxes(visible=False) fig.write_image(f'temp/normals/{batch["id"][i]}_{batch["frame_nr"][i].cpu().numpy()}_{batch["garment"][i]}.png') count += 1 if count > 100: break if __name__ == '__main__': if not args['debug']: init_wandb() smpl = SMPL(orig_smpl=config['is_cape'], path='smpl', device=config['device']) skirt_modifier = '_skirt' if config['is_skirt'] else '' dataset_train = Dataset(config['train_dir'], config['signature'], device=config['device'], frames_file=config['train_frames'], use_cache=config['dataset_in_memory']) dataset_valid = Dataset(config['valid_dir'], config['signature'], device=config['device'], frames_file=config['valid_frames'], use_cache=config['dataset_in_memory']) train_dataloader = DataLoader(dataset_train, batch_size=4, drop_last=True, shuffle=True, ) valid_dataloader = DataLoader(dataset_valid, batch_size=4, drop_last=True, shuffle=True) train_accumulator = Accumulator() valid_accumulator = Accumulator() model = CVAENorm(config) sample() exit() optimizer = optim.Adam(model.parameters(), lr=config['lr']) _MSSSIM = MS_SSIM(data_range=2, size_average=True, channel=3) g_balancing_coeff = config['starting_balancing_coeff'] ckpt_saver = CheckpointSaver('checkpoints/iccvw/' + config['name'], top_n=25) step = 0 if config['pretrained_model'] != '': if 'ignore_saved_optimizer' in config.keys() and config['ignore_saved_optimizer']: step = load_model(model, config['pretrained_model']) else: step = load_model(model, config['pretrained_model'], optimizer=optimizer) if not args['debug']: # model = torch.compile(model) wandb.watch(model, log='all', log_freq=10000, log_graph=True) train()