train_vqvae.py

import os
import torch
import torch.optim as optim
import torch.nn as nn
from model.vqvae import VQVAE
from datasets.data_processing import get_dataset
from datetime import datetime
from copy import deepcopy
import time

def main():
    # Hyperparameters
    input_dim = 3  # x, y, z coordinates
    hidden_dim = 1024
    codebook_size = 4096
    embedding_dim = 64
    batch_size = 1
    num_epochs = store_epoch + 1
    default_lr = 4e-5
    encoder_lr = 3e-5
    num_points = 15000
    voxel_size = 10
    spconv_channels = 32
    blocks_num = 6
    smooth_end_epoch = 10
    beta = 0.35
    lambda_chamfer = 40.0
    lambda_vq_start = 0.01
    lambda_vq_final = 0.05
    lambda_construction = 100.0
    lambda_usage = 0.02
    warmup_steps = 80
    ema_decay = 0.9995
    token_N1_stage = 9
    token_N2_stage = 14
    token_N3_stage = 14
    decoder_layer = 5
    buffer_stage_num = 3

    ckpts_save_path = None
    ckpt_save_step = 20
    if_load_ckpt = True
    if_only_init_param = True
    ckpt_name = None
    
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    time_commit = 0
    
    # Initialize model
    model = VQVAE(input_dim, 
                  hidden_dim, 
                  codebook_size, 
                  embedding_dim, 
                  num_points, 
                  voxel_size, 
                  spconv_channels, 
                  blocks_num,
                  smooth_end_epoch,
                  beta,
                  lambda_chamfer,
                  lambda_vq_start,
                  lambda_vq_final,
                  lambda_construction,
                  lambda_usage,
                  warmup_steps,
                  ema_decay,
                  token_N1_stage,
                  token_N2_stage,
                  token_N3_stage,
                  decoder_layer,
                  buffer_stage_num).to(device)

    optimizer = torch.optim.Adam([
        {"params": model.encoder.parameters(), "lr": encoder_lr},
        {"params": model.pos_linear.parameters(), "lr": encoder_lr},
        {"params": model.act.parameters(), "lr": encoder_lr},
        {"params": model.after_spconv.parameters(), "lr": encoder_lr},
        {"params": model.phi.parameters(), "lr": default_lr},
        {"params": model.decoder.parameters(), "lr": default_lr},
    ], weight_decay=1e-5)

    curr_epoch = 0
    opt = None
    if if_load_ckpt: 
        ckpt_path = os.path.join(ckpts_save_path, ckpt_name)
        if not if_only_init_param: # start from curr_epoch 
            opt, curr_epoch = model.init_from_ckpt(ckpt_path, None, False)
        else: # start from 0 epoch
            _, _ = model.init_from_ckpt(ckpt_path, None, True)

    if opt is not None:
        optimizer.load_state_dict(opt)          

    # Load ShapeNetV2 data
    root_dir = "data/ShapeNetCore.v2.PC15k"  # PC15k
    categories = ['airplane']  # Example: cars and chairs
    train_dataset, test_dataset = get_dataset(root_dir, num_points, category=categories)
    dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True, drop_last=True)
    
    # Train the model
    for epoch in range(curr_epoch, num_epochs):
        total_loss = 0
        total_construction_loss = 0
        total_vq_loss = 0
        total_chamfer_loss = 0
        total_usage_loss = 0
        global_step = 0
        log_data_total = torch.zeros(8, dtype=torch.float32)
        
        for batch in dataloader:
            # time1 = time.perf_counter()
            x = batch['train_points'].to(device)
            optimizer.zero_grad()
            
            reconstructed, loss, construction_loss, vq_loss, chamfer_loss, usage_loss = model(x, epoch, global_step)
            loss.backward()

            # Gradient Clipping to avoid exploding updates
            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)

            optimizer.step()
            
            total_loss += loss.item()
            total_construction_loss += construction_loss.item()
            total_vq_loss += vq_loss.item()
            total_chamfer_loss += chamfer_loss.item()
            total_usage_loss += usage_loss.item()
            print(f"VQ_Loss: {vq_loss:.4f}, construction_Loss: {construction_loss:.4f} , chamfer_Loss: {chamfer_loss:.4f} , usage_Loss: {usage_loss:.4f}" )
            log_data_total += model.log_data
            global_step += 1    
        
        if epoch % ckpt_save_step == 0:
            save_checkpoint(model, optimizer, None, epoch, ckpts_save_path)

        log_data_avg = log_data_total / global_step
    
        avg_loss = total_loss / len(dataloader)
        avg_construction_loss = total_construction_loss / len(dataloader)
        avg_vq_loss = total_vq_loss / len(dataloader)
        
        print(f"Epoch [{epoch+1}/{num_epochs}], "
            f"Loss: {avg_loss:.4f}, "
            f"Construction Loss: {avg_construction_loss:.4f}, "
            f"VQ Loss: {avg_vq_loss:.4f}")


def save_checkpoint(model, optimizer=None, scheduler=None, epoch=0, save_root="checkpoints"):
    date_str = datetime.now().strftime("%m%d")
    save_dir = os.path.join(save_root, date_str)
    os.makedirs(save_dir, exist_ok=True)
    ckpt_path = os.path.join(save_dir, f"epoch{epoch}.pth")
    ckpt = {
        "epoch": epoch,
        "model": model.state_dict(),
    }

    if optimizer is not None:
        ckpt["optimizer"] = optimizer.state_dict()

    if scheduler is not None:
        ckpt["scheduler"] = scheduler.state_dict()

    torch.save(ckpt, ckpt_path)

    print(f"✔ Checkpoint saved to: {ckpt_path}")

if __name__ == "__main__":
    main()