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()) config_module = importlib.import_module(args['config']) config = config_module.config if args['debug']: from configs.direct_1 import config # from configs.ae_static_orig_2 import 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 random import numpy as np import torch # from torchmetrics import MultiScaleStructuralSimilarityIndexMeasure as MSSSIM from custom_utils.custom_utils import UV2Mesh, get_summary_html, maskFacesBatch, plot_3d_collision import plotly.graph_objects as go import plotly.express as ex import torch.optim as optim if config['is_skirt']: from losses.losses import edge_loss_skirt as edge_loss else: from losses.losses import edge_loss from data.dataset import Dataset # from torch.utils.data import DataLoader from losses.losses import collision_loss, normal_loss from utils.postprocess import fix_collision from custom_utils.custom_utils import plot_3d from utils.checkpoints import CheckpointSaver, load_model from custom_utils.custom_utils import Accumulator from utils.utils import InfiniteDataLoader as DataLoader from models.cond_encoder import AutoEncoder as Model from smpl.smpl_torch import SMPL from pytorch_msssim import MS_SSIM def init_wandb(): run = wandb.init(project='iccv', name=config['name'], notes=config['notes'], dir='/data-local/127/temp', tags=config['tags']) artifact = wandb.Artifact('code', type='code') artifact.add_file('scripts/direct_train.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('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 train_step(batch): model.train() # with torch.autocast(device_type='cuda', dtype=torch.float16, enabled=use_amp): if config['use_curvature']: input = torch.cat([batch['uv_curvature'], batch['uv_static']], axis=1) else: input = torch.cat([batch['uv_body_posed'], batch['uv_static']], axis=1) pred = model([input, batch['uv_mask']]) mesh_pred = uv2mesh(pred) mesh_orig = batch['mesh_unposed'] shape = batch['pose_shape'][:, 72:] gender = torch.greater_equal(shape[:, 10] , 0.5).to(torch.long) shape_offsets = torch.stack([torch.tensordot(shape[i,:10][None, :], smpl.shapedirs[gender[i]], dims=[[1], [2]])[0,smpl.to_keep,:] for i in range(shape.shape[0])]) mesh_pred_shaped = mesh_pred + shape_offsets * batch['mesh_mask'][:,:,None] mesh_pred_posed = smpl(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], mesh_pred_shaped)[0] e_loss = torch.mean(edge_loss(mesh_pred, mesh_orig, batch['mesh_mask'])) c_loss, c_ratio = collision_loss(mesh_pred_posed, batch['mesh_body_posed'], batch['mesh_mask']) c_loss, c_ratio = torch.mean(c_loss), torch.mean(c_ratio) # l2_loss = torch.mean(1e3 * torch.mean(torch.sum(((batch['uv_unposed'] - pred) ** 2), axis=-1), axis=[1, 2])) l2_mesh_loss = torch.mean(1e3 * torch.mean(torch.sum(batch['mesh_mask'][:,:,None] * (batch['mesh_posed'] - mesh_pred_posed) ** 2, axis=-1), axis=-1)) n_loss = normal_loss(mesh_pred_posed, batch['mesh_posed'], batch['mesh_mask'], is_skirt=config['is_skirt'], is_smpl_orig=config['is_cape']) loss = l2_mesh_loss + (e_loss + c_loss + n_loss) * config['reg_loss_coeff'] optimizer.zero_grad() loss.backward() optimizer.step() # scaler.scale(loss).backward() # scaler.step(optimizer) # scaler.update() # optimizer.zero_grad() model.eval() with torch.inference_mode(): ms_ssim = torch.mean(_MSSSIM(pred + 1, batch['uv_unposed'] + 1)) metric = torch.sum(torch.sqrt(torch.sum(batch['mesh_mask'][:,:,None] * (mesh_orig - mesh_pred.detach()) ** 2, axis=-1))) / torch.sum(batch['mesh_mask']) return { 'train/loss': loss, 'train/edge_loss': e_loss, 'train/euclid_error': metric, 'train/collision_ratio': c_ratio, 'train/ms_ssim': ms_ssim, 'train/collision_loss': c_loss, 'train/l2_mesh_loss': l2_mesh_loss, 'train/normal_loss': n_loss } def valid_step(batch): model.eval() with torch.inference_mode(): if config['use_curvature']: input = torch.cat([batch['uv_curvature'], batch['uv_static']], axis=1) else: input = torch.cat([batch['uv_body_posed'], batch['uv_static']], axis=1) pred = model([input, batch['uv_mask']]) mesh_pred = uv2mesh(pred) mesh_orig = batch['mesh_unposed'] shape = batch['pose_shape'][:, 72:] gender = torch.greater_equal(shape[:, 10] , 0.5).to(torch.long) shape_offsets = torch.stack([torch.tensordot(shape[i,:10][None, :], smpl.shapedirs[gender[i]], dims=[[1], [2]])[0,smpl.to_keep,:] for i in range(shape.shape[0])]) mesh_pred_shaped = mesh_pred + shape_offsets * batch['mesh_mask'][:,:,None] mesh_pred_posed = smpl(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], mesh_pred_shaped)[0] e_loss = torch.mean(edge_loss(mesh_pred, mesh_orig, batch['mesh_mask'])) c_loss, c_ratio = collision_loss(mesh_pred_posed, batch['mesh_body_posed'], batch['mesh_mask']) c_loss, c_ratio = torch.mean(c_loss), torch.mean(c_ratio) # l2_loss = torch.mean(1e3 * torch.mean(torch.sum(((batch['uv_unposed'] - pred) ** 2), axis=-1), axis=[1, 2])) l2_mesh_loss = torch.mean(1e3 * torch.mean(torch.sum(batch['mesh_mask'][:,:,None] * (batch['mesh_posed'] - mesh_pred_posed) ** 2, axis=-1), axis=-1)) n_loss = normal_loss(mesh_pred_posed, batch['mesh_posed'], batch['mesh_mask'], is_skirt=config['is_skirt'], is_smpl_orig=config['is_cape']) ms_ssim = torch.mean(_MSSSIM(pred + 1, batch['uv_unposed'] + 1)) loss = l2_mesh_loss + (e_loss + c_loss + n_loss) * config['reg_loss_coeff'] metric = torch.sum(torch.sqrt(torch.sum(batch['mesh_mask'][:,:,None] * (mesh_orig - mesh_pred) ** 2, axis=-1))) / torch.sum(batch['mesh_mask']) return { 'valid/loss': loss, 'valid/edge_loss': e_loss, 'valid/euclid_error': metric, 'valid/collision_ratio': c_ratio, 'valid/ms_ssim': ms_ssim, 'valid/collision_loss': c_loss, 'valid/l2_mesh_loss': l2_mesh_loss, 'valid/normal_loss': n_loss } def plot_everything(batch: dict, type: str): model.eval() with torch.inference_mode(): if config['use_curvature']: input = torch.cat([batch['uv_curvature'], batch['uv_static']], axis=1) else: input = torch.cat([batch['uv_body_posed'], batch['uv_static']], axis=1) pred = model([input, batch['uv_mask']]).detach() mesh_pred = uv2mesh(pred).cpu().numpy() mesh_orig = batch['mesh_unposed'].cpu().numpy() orig_img = torch.permute(batch['uv_unposed'], [0,2,3,1]) 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]) 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) shape = batch['pose_shape'][:, 72:] gender = torch.greater_equal(shape[:, 10] , 0.5).to(torch.long) shape_offsets = torch.stack([torch.tensordot(shape[i,:10][None, :], smpl.shapedirs[gender[i]], dims=[[1], [2]])[0,smpl.to_keep,:] for i in range(shape.shape[0])]).cpu().numpy() garment_unposed, faces = maskFacesBatch(mesh_pred, F, mask) garment_unposed = [garment_unposed[i] + shape_offsets[i, mask[i], :] for i in range(mask.shape[0])] gt = [mesh_orig[i][mask[i], :] + shape_offsets[i, mask[i], :] for i in range(pred.shape[0])] for i in range(mesh_orig.shape[0]): mesh_pred[i, mask[i], :] += shape_offsets[i, mask[i], :] mesh_posed = smpl(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], torch.tensor(mesh_pred, device=config['device']))[0].cpu().numpy() gt_posed = [batch['mesh_posed'].cpu().numpy()[i, mask[i], :] for i in range(mask.shape[0])] mesh_posed_fixed = fix_collision(torch.tensor(mesh_posed, device=config['device']), batch['mesh_body_posed'], torch.tensor(mask, device=config['device'])).cpu().numpy() body = batch['mesh_body_posed'].cpu().numpy() garment_posed = [mesh_posed[i, mask[i], :] for i in range(body.shape[0])] collision = [np.concatenate((body[i], garment_posed[i]), axis=0) for i in range(len(garment_posed))] collision_faces = [np.concatenate((F, faces[i] + config['smpl_res']), axis=0) for i in range(len(garment_posed))] collision_color = [np.concatenate((np.ones_like(body[i], dtype=np.int32)*100, np.ones_like(garment_posed[i], dtype=np.int32)*230), axis=0) for i in range(len(garment_posed))] garment_posed_fixed = [mesh_posed_fixed[i, mask[i], :] for i in range(body.shape[0])] collision_fixed = [np.concatenate((body[i], garment_posed_fixed[i]), axis=0) for i in range(len(garment_posed))] collision_gt = [np.concatenate((body[i], gt_posed[i]), axis=0) for i in range(len(garment_posed))] figs = { type + "/garment_" + str(i): plot_3d(gt[i], garment_unposed[i], faces[i], collision[i], collision_faces[i], collision_color[i]) for i in range(4) } figs_coll = { type + "/garment_" + str(i) + '_collision': plot_3d_collision(collision_gt[i], collision[i], collision_fixed[i], collision_faces[i], collision_color[i]) for i in range(4) } return { **figs, **figs_coll, **{ type + '/images': imgs, } } def train(): valid_iterator = iter(valid_dataloader) step = 0 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 step % 10000 == 0: ckpt_saver_enc(model.encoder, step, optimizer, scaler) ckpt_saver_dec(model.decoder, step, optimizer, scaler) step += 1 if __name__ == '__main__': if not args['debug']: init_wandb() use_amp = True skirt_modifier = '_skirt' if config['is_skirt'] else '' # dataset = Dataset('/data-local/127/train', config['signature'], resolutions=[512], device=device, frames_file='frames_jumpsuit.txt') dataset_train = Dataset(config['train_dir'], config['signature'], device=config['device'], frames_file=config['train_frames'], use_cache=config['dataset_in_memory'], is_cape=config['is_cape']) dataset_valid = Dataset(config['valid_dir'], config['signature'], device=config['device'], frames_file=config['valid_frames'], use_cache=config['dataset_in_memory'], is_cape=config['is_cape']) 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 : Model = Model(config) optimizer = optim.Adam(model.parameters(), lr=config['lr']) scaler = torch.cuda.amp.GradScaler(enabled=use_amp) smpl = SMPL(orig_smpl=config['is_cape'], path='smpl', device=config['device']) # _MSSSIM = MSSSIM(data_range=2).to(config['device']) _MSSSIM = MS_SSIM(data_range=2, size_average=True, channel=3) if config['pretrained_model'] != '': load_model(model.encoder, config['pretrained_model'] + '_enc') load_model(model.decoder, config['pretrained_model'] + '_dec') ckpt_saver_enc = CheckpointSaver('checkpoints/iccv/torch/' + config['name'] + '_enc', top_n=5) ckpt_saver_dec = CheckpointSaver('checkpoints/iccv/torch/' + config['name'] + '_dec', top_n=5) run_name = config['name'] print(f'started training: {run_name}') if not args['debug']: # model = torch.compile(model) wandb.watch(model, log='all', log_freq=1000, log_graph=True) train()