VQ-Trans/train_t2m_trans.py

import os 
import torch
import numpy as np

from torch.utils.tensorboard import SummaryWriter
from os.path import join as pjoin
from torch.distributions import Categorical
import json
import clip

import options.option_transformer as option_trans
import models.vqvae as vqvae
import utils.utils_model as utils_model
import utils.eval_trans as eval_trans
from dataset import dataset_TM_train
from dataset import dataset_TM_eval
from dataset import dataset_tokenize
import models.t2m_trans as trans
from options.get_eval_option import get_opt
from models.evaluator_wrapper import EvaluatorModelWrapper
import warnings
warnings.filterwarnings('ignore')

##### ---- Exp dirs ---- #####
args = option_trans.get_args_parser()
torch.manual_seed(args.seed)

args.out_dir = os.path.join(args.out_dir, f'{args.exp_name}')
args.vq_dir= os.path.join("./dataset/KIT-ML" if args.dataname == 'kit' else "./dataset/HumanML3D", f'{args.vq_name}')
os.makedirs(args.out_dir, exist_ok = True)
os.makedirs(args.vq_dir, exist_ok = True)

##### ---- Logger ---- #####
logger = utils_model.get_logger(args.out_dir)
writer = SummaryWriter(args.out_dir)
logger.info(json.dumps(vars(args), indent=4, sort_keys=True))

##### ---- Dataloader ---- #####
train_loader_token = dataset_tokenize.DATALoader(args.dataname, 1, unit_length=2**args.down_t)

from utils.word_vectorizer import WordVectorizer
w_vectorizer = WordVectorizer('./glove', 'our_vab')
val_loader = dataset_TM_eval.DATALoader(args.dataname, False, 32, w_vectorizer)

dataset_opt_path = 'checkpoints/kit/Comp_v6_KLD005/opt.txt' if args.dataname == 'kit' else 'checkpoints/t2m/Comp_v6_KLD005/opt.txt'

wrapper_opt = get_opt(dataset_opt_path, torch.device('cuda'))
eval_wrapper = EvaluatorModelWrapper(wrapper_opt)

##### ---- Network ---- #####
clip_model, clip_preprocess = clip.load("ViT-B/32", device=torch.device('cuda'), jit=False, download_root='/apdcephfs_cq2/share_1290939/maelyszhang/.cache/clip')  # Must set jit=False for training
clip.model.convert_weights(clip_model)  # Actually this line is unnecessary since clip by default already on float16
clip_model.eval()
for p in clip_model.parameters():
    p.requires_grad = False

net = vqvae.HumanVQVAE(args, ## use args to define different parameters in different quantizers
                       args.nb_code,
                       args.code_dim,
                       args.output_emb_width,
                       args.down_t,
                       args.stride_t,
                       args.width,
                       args.depth,
                       args.dilation_growth_rate)


trans_encoder = trans.Text2Motion_Transformer(num_vq=args.nb_code, 
                                embed_dim=args.embed_dim_gpt, 
                                clip_dim=args.clip_dim, 
                                block_size=args.block_size, 
                                num_layers=args.num_layers, 
                                n_head=args.n_head_gpt, 
                                drop_out_rate=args.drop_out_rate, 
                                fc_rate=args.ff_rate)


print ('loading checkpoint from {}'.format(args.resume_pth))
ckpt = torch.load(args.resume_pth, map_location='cpu')
net.load_state_dict(ckpt['net'], strict=True)
net.eval()
net.cuda()

if args.resume_trans is not None:
    print ('loading transformer checkpoint from {}'.format(args.resume_trans))
    ckpt = torch.load(args.resume_trans, map_location='cpu')
    trans_encoder.load_state_dict(ckpt['trans'], strict=True)
trans_encoder.train()
trans_encoder.cuda()

##### ---- Optimizer & Scheduler ---- #####
optimizer = utils_model.initial_optim(args.decay_option, args.lr, args.weight_decay, trans_encoder, args.optimizer)
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.lr_scheduler, gamma=args.gamma)

##### ---- Optimization goals ---- #####
loss_ce = torch.nn.CrossEntropyLoss()

nb_iter, avg_loss_cls, avg_acc = 0, 0., 0.
right_num = 0
nb_sample_train = 0

##### ---- get code ---- #####
for batch in train_loader_token:
    pose, name = batch
    bs, seq = pose.shape[0], pose.shape[1]

    pose = pose.cuda().float() # bs, nb_joints, joints_dim, seq_len
    target = net.encode(pose)
    target = target.cpu().numpy()
    np.save(pjoin(args.vq_dir, name[0] +'.npy'), target)


train_loader = dataset_TM_train.DATALoader(args.dataname, args.batch_size, args.nb_code, args.vq_name, unit_length=2**args.down_t)
train_loader_iter = dataset_TM_train.cycle(train_loader)

        
##### ---- Training ---- #####
best_fid, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, writer, logger = eval_trans.evaluation_transformer(args.out_dir, val_loader, net, trans_encoder, logger, writer, 0, best_fid=1000, best_iter=0, best_div=100, best_top1=0, best_top2=0, best_top3=0, best_matching=100, clip_model=clip_model, eval_wrapper=eval_wrapper)
while nb_iter <= args.total_iter:
    
    batch = next(train_loader_iter)
    clip_text, m_tokens, m_tokens_len = batch
    m_tokens, m_tokens_len = m_tokens.cuda(), m_tokens_len.cuda()
    bs = m_tokens.shape[0]
    target = m_tokens    # (bs, 26)
    target = target.cuda()
    
    text = clip.tokenize(clip_text, truncate=True).cuda()
    
    feat_clip_text = clip_model.encode_text(text).float()

    input_index = target[:,:-1]

    if args.pkeep == -1:
        proba = np.random.rand(1)[0]
        mask = torch.bernoulli(proba * torch.ones(input_index.shape,
                                                         device=input_index.device))
    else:
        mask = torch.bernoulli(args.pkeep * torch.ones(input_index.shape,
                                                         device=input_index.device))
    mask = mask.round().to(dtype=torch.int64)
    r_indices = torch.randint_like(input_index, args.nb_code)
    a_indices = mask*input_index+(1-mask)*r_indices

    cls_pred = trans_encoder(a_indices, feat_clip_text)
    cls_pred = cls_pred.contiguous()

    loss_cls = 0.0
    for i in range(bs):
        # loss function     (26), (26, 513)
        loss_cls += loss_ce(cls_pred[i][:m_tokens_len[i] + 1], target[i][:m_tokens_len[i] + 1]) / bs

        # Accuracy
        probs = torch.softmax(cls_pred[i][:m_tokens_len[i] + 1], dim=-1)

        if args.if_maxtest:
            _, cls_pred_index = torch.max(probs, dim=-1)

        else:
            dist = Categorical(probs)
            cls_pred_index = dist.sample()
        right_num += (cls_pred_index.flatten(0) == target[i][:m_tokens_len[i] + 1].flatten(0)).sum().item()

    ## global loss
    optimizer.zero_grad()
    loss_cls.backward()
    optimizer.step()
    scheduler.step()

    avg_loss_cls = avg_loss_cls + loss_cls.item()
    nb_sample_train = nb_sample_train + (m_tokens_len + 1).sum().item()

    nb_iter += 1
    if nb_iter % args.print_iter ==  0 :
        avg_loss_cls = avg_loss_cls / args.print_iter
        avg_acc = right_num * 100 / nb_sample_train
        writer.add_scalar('./Loss/train', avg_loss_cls, nb_iter)
        writer.add_scalar('./ACC/train', avg_acc, nb_iter)
        msg = f"Train. Iter {nb_iter} : Loss. {avg_loss_cls:.5f}, ACC. {avg_acc:.4f}"
        logger.info(msg)
        avg_loss_cls = 0.
        right_num = 0
        nb_sample_train = 0

    if nb_iter % args.eval_iter ==  0:
        best_fid, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, writer, logger = eval_trans.evaluation_transformer(args.out_dir, val_loader, net, trans_encoder, logger, writer, nb_iter, best_fid, best_iter, best_div, best_top1, best_top2, best_top3, best_matching, clip_model=clip_model, eval_wrapper=eval_wrapper)

    if nb_iter == args.total_iter: 
        msg_final = f"Train. Iter {best_iter} : FID. {best_fid:.5f}, Diversity. {best_div:.4f}, TOP1. {best_top1:.4f}, TOP2. {best_top2:.4f}, TOP3. {best_top3:.4f}"
        logger.info(msg_final)
        break