OM_train_3modes_opt.py

"""

OM_train_3modes_opt.py — Optimized 3-mode training (diffusion + contrastive + registration).



Speed optimizations over OM_train_3modes.py (all mathematically equivalent):

  1. DataLoader: num_workers, pin_memory, persistent_workers for I/O overlap

  2. optimizer.zero_grad(set_to_none=True) — avoids zero-fill overhead

  3. Fixed-length T_regist (16 steps) — avoids XPU dynamic shape recompilation

  4. Removed redundant x0.to(device) call

  5. Uses diffuser_opt.DeformDDPM (hoisted clone, no *0 redundancy, OptSTN, inference_mode)

  6. Uses losses_opt.MSLNCC/LNCC (register_buffer for kernels)

  7. Pre-compute proc_type lists to reduce Python overhead in hot loop

  8. Uses OptRecMulModMutAttnNet (cached resample tensors, ~300 fewer CPU→GPU transfers)

  9. Uses OptSTN for ddf_stn (register_buffer, no per-call .to())

"""

import os, sys

ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.append(ROOT_DIR)

import gc
import torch
import torchvision
from torch import nn
from torchvision.utils import save_image
from torch.utils.data import DataLoader

from torch.optim import Adam, SGD
from Diffusion.diffuser_opt import DeformDDPM
from Diffusion.networks_opt import get_net_opt, OptSTN
from torchvision.transforms import Lambda
import torch.nn.functional as F
import Diffusion.losses_opt as losses
import random
import glob
import numpy as np
import utils
from tqdm import tqdm

from Dataloader.dataloader0 import get_dataloader
from Dataloader.dataLoader import *

from Dataloader.dataloader_utils import thresh_img
import yaml
import argparse

####################
import torch.multiprocessing as mp
from torch.utils.data.distributed import DistributedSampler
from torch.nn.parallel import DistributedDataParallel as DDP
import torch.distributed as dist
###############
def ddp_setup(rank, world_size):
    """

    Args:

        rank: Unique identifier of each process

        world_size: Total number of processes

    """
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "12355"
    dist.init_process_group(backend="nccl", rank=rank, world_size=world_size)
    torch.cuda.set_device(rank)

# Auto-detect: use DDP only when multiple CUDA GPUs are available
use_distributed = torch.cuda.is_available() and torch.cuda.device_count() > 1
# use_distributed = True
# use_distributed = False

EPS = 1e-5
MSK_EPS = 0.01
TEXT_EMBED_PROB = 0.7
AUG_RESAMPLE_PROB = 0.5
LOSS_WEIGHTS_DIFF = [2.0, 2.0, 4.0]  # [ang, dist, reg]
# LOSS_WEIGHTS_REGIST = [9.0, 1.0, 16.0]  # [imgsim, imgmse, ddf]
LOSS_WEIGHTS_REGIST = [1.0, 0.05, 128]  # [imgsim, imgmse, ddf]
DIFF_REG_BATCH_RATIO = 2
LOSS_WEIGHT_CONTRASTIVE = 1.0
CONTRASTIVE_STEP_RATIO = 2

# OPT: Fixed registration timestep count to avoid XPU dynamic shape recompilation
FIXED_T_REGIST_LEN = 16

# OPT: DataLoader workers (set to 0 to disable multiprocessing if needed)
NUM_WORKERS = 4
PIN_MEMORY = True

# AUG_PERMUTE_PROB = 0.35

parser = argparse.ArgumentParser()

# config_file_path = 'Config/config_cmr.yaml'
parser.add_argument(
        "--config",
        "-C",
        help="Path for the config file",
        type=str,
        # default="Config/config_cmr.yaml",
        # default="Config/config_lct.yaml",
        default="Config/config_all.yaml",
        required=False,
    )
parser.add_argument("--dummy-samples", type=int, default=0, help="Use dummy random data for testing (0=use real data)")
parser.add_argument("--batchsize", type=int, default=0, help="Override batch size from config (0=use config value)")
parser.add_argument("--num-workers", type=int, default=NUM_WORKERS, help="DataLoader num_workers (default: 4)")
args = parser.parse_args()
#=======================================================================================================================

class _DummyIndiv(torch.utils.data.Dataset):
    def __init__(self, n, sz, embd_dim=1024):
        self.n, self.sz, self.embd_dim = n, sz, embd_dim
    def __len__(self): return self.n
    def __getitem__(self, i):
        return np.random.rand(1, self.sz, self.sz, self.sz).astype(np.float64), np.random.randn(self.embd_dim).astype(np.float32)

class _DummyPair(torch.utils.data.Dataset):
    def __init__(self, n, sz, embd_dim=1024):
        self.n, self.sz, self.embd_dim = n, sz, embd_dim
    def __len__(self): return self.n
    def __getitem__(self, i):
        return (np.random.rand(1, self.sz, self.sz, self.sz).astype(np.float64),
                np.random.rand(1, self.sz, self.sz, self.sz).astype(np.float64),
                np.random.randn(self.embd_dim).astype(np.float32),
                np.random.randn(self.embd_dim).astype(np.float32))


def main_train(rank=0,world_size=1,train_mode_ratio=1,thresh_imgsim=0.01):
    if use_distributed:
        ddp_setup(rank,world_size)

        if torch.distributed.is_initialized():
            print(f"World size: {torch.distributed.get_world_size()}")
            print(f"Communication backend: {torch.distributed.get_backend()}")
    gpu_id = rank

    # Load the YAML file into a dictionary
    with open(args.config, 'r') as file:
        hyp_parameters = yaml.safe_load(file)
    if args.batchsize > 0:
        hyp_parameters['batchsize'] = args.batchsize
    print(hyp_parameters)

    # epoch_per_save=10
    epoch_per_save=hyp_parameters['epoch_per_save']

    data_name=hyp_parameters['data_name']
    net_name = hyp_parameters['net_name']

    Net=get_net_opt(net_name)

    suffix_pth=f'_{data_name}_{net_name}.pth'
    model_save_path = os.path.join('Models',f'{data_name}_{net_name}/')
    model_dir=model_save_path
    transformer=utils.get_transformer(img_sz=hyp_parameters["ndims"]*[hyp_parameters['img_size']])

    # OPT: DataLoader with num_workers, pin_memory, persistent_workers
    num_workers = args.num_workers
    use_pin_memory = PIN_MEMORY and hyp_parameters["device"] != "cpu"

    if args.dummy_samples > 0:
        dataset = _DummyIndiv(args.dummy_samples, hyp_parameters['img_size'])
        datasetp = _DummyPair(args.dummy_samples, hyp_parameters['img_size'])
    else:
        dataset = OMDataset_indiv(transform=None)
        datasetp = OMDataset_pair(transform=None)

    train_loader = DataLoader(
        dataset,
        batch_size=hyp_parameters['batchsize'],
        shuffle=True,
        drop_last=True,
        num_workers=num_workers,                           # OPT
        pin_memory=use_pin_memory,                         # OPT
        persistent_workers=num_workers > 0,                # OPT
    )
    train_loader_p = DataLoader(
        datasetp,
        batch_size=max(1, hyp_parameters['batchsize']//DIFF_REG_BATCH_RATIO),
        shuffle=True,
        drop_last=True,
        num_workers=num_workers,                           # OPT
        pin_memory=use_pin_memory,                         # OPT
        persistent_workers=num_workers > 0,                # OPT
    )



    Deformddpm = DeformDDPM(
        network=Net(
            n_steps=hyp_parameters["timesteps"],
            ndims=hyp_parameters["ndims"],
            num_input_chn = hyp_parameters["num_input_chn"],
            res = hyp_parameters['img_size']
            ),
        n_steps=hyp_parameters["timesteps"],
        image_chw=[1] + [hyp_parameters["img_size"]]*hyp_parameters["ndims"],
        device=hyp_parameters["device"],
        batch_size=hyp_parameters["batchsize"],
        img_pad_mode=hyp_parameters["img_pad_mode"],
        v_scale=hyp_parameters["v_scale"],
    )


    ddf_stn = OptSTN(
        img_sz=hyp_parameters["img_size"],
        ndims=hyp_parameters["ndims"],
        # padding_mode="zeros",
        padding_mode=hyp_parameters["padding_mode"],
        device=hyp_parameters["device"],
    )


    if use_distributed:
        Deformddpm.to(rank)
        Deformddpm = DDP(Deformddpm, device_ids=[rank])
        ddf_stn.to(rank)
    else:
        Deformddpm.to(hyp_parameters["device"])
        ddf_stn.to(hyp_parameters["device"])
    # ddf_stn = DDP(ddf_stn, device_ids=[rank])


    # mse = nn.MSELoss()
    # loss_reg = losses.Grad(penalty=['l1', 'negdetj'], ndims=hyp_parameters["ndims"])
    # loss_reg = losses.Grad(penalty=['l1', 'negdetj', 'range'], ndims=hyp_parameters["ndims"])
    loss_reg = losses.Grad(penalty=['l1', 'negdetj', 'range'], ndims=hyp_parameters["ndims"],outrange_thresh=0.2,outrange_weight=1e3)
    loss_reg1 = losses.Grad(penalty=['l1', 'negdetj', 'range'], ndims=hyp_parameters["ndims"],outrange_thresh=0.6,outrange_weight=1e3)

    loss_dist = losses.MRSE(img_sz=hyp_parameters["img_size"])
    # loss_ang = losses.MRSE(img_sz=hyp_parameters["img_size"])
    loss_ang = losses.NCC(img_sz=hyp_parameters["img_size"])
    loss_imgsim = losses.MSLNCC()
    loss_imgmse = losses.LMSE()

    optimizer = Adam(Deformddpm.parameters(), lr=hyp_parameters["lr"])

    # check for existing models
    if not os.path.exists(model_dir):
        os.makedirs(model_dir, exist_ok=True)
    model_files = glob.glob(os.path.join(model_dir, "*.pth"))
    model_files.sort()
    if model_files:
        if gpu_id == 0:
            print(model_files)
        initial_epoch, Deformddpm, optimizer = ddp_load_dict(gpu_id, Deformddpm, optimizer, model_files[-1], use_distributed=use_distributed)
    else:
        initial_epoch = 0

    if gpu_id == 0:
        print('len_train_data: ',len(dataset))
    # Training loop
    for epoch in range(initial_epoch,hyp_parameters["epoch"]):

        epoch_loss_tot = 0.0
        epoch_loss_gen_d = 0.0
        epoch_loss_gen_a = 0.0
        epoch_loss_reg = 0.0
        epoch_loss_regist = 0.0
        epoch_loss_imgsim = 0.0
        epoch_loss_imgmse = 0.0
        epoch_loss_ddfreg = 0.0
        epoch_loss_contrastive = 0.0
        # Set model inside to train model
        Deformddpm.train()

        loss_nan_step = 0  # yu: count the number of nan loss steps

        total = min(len(train_loader), len(train_loader_p))
        for step, (batch, batch_p) in tqdm(enumerate(zip(train_loader, train_loader_p)), total=total):

            # ==========================================================================
            # diffusion train on single image

            [x0,embd] = batch # for om dataset
            x0 = x0.to(hyp_parameters["device"]).type(torch.float32)
            embd_dev = embd.to(hyp_parameters["device"]).type(torch.float32)
            if np.random.uniform(0,1)<TEXT_EMBED_PROB:
                embd_in = embd_dev
            else:
                embd_in = None

            n = x0.size()[0]  # batch_size -> n
            # OPT: removed redundant x0.to(device) — already done above

            blind_mask = utils.get_random_deformed_mask(x0.shape[2:],apply_possibility=0.6).to(hyp_parameters["device"])

            # random deformation + rotation
            if hyp_parameters["ndims"]>2:
                if np.random.uniform(0,1)<AUG_RESAMPLE_PROB:
                    x0 = utils.random_resample(x0, deform_scale=0)
                # elif np.random.uniform(0,1)<AUG_RESAMPLE_PROB+AUG_PERMUTE_PROB:
                else:
                    [x0] = utils.random_permute([x0], select_dims=[-1,-2,-3])
            # x0 = transformer(x0)
            if hyp_parameters['noise_scale']>0:
                if np.random.uniform(0,1)<AUG_RESAMPLE_PROB:
                    x0 = thresh_img(x0, [0, 2*hyp_parameters['noise_scale']])
                x0 = x0 * (np.random.normal(1, hyp_parameters['noise_scale'] * 1)) + np.random.normal(0, hyp_parameters['noise_scale'] * 1)

            # Picking some noise for each of the images in the batch, a timestep and the respective alpha_bars
            t = torch.randint(0, hyp_parameters["timesteps"], (n,)).to(
                hyp_parameters["device"]
            )  # pick up a seq of rand number from 0 to 'timestep'

            proc_type = random.choice(['adding', 'downsample', 'slice', 'slice1', 'none', 'uncon', 'uncon', 'uncon'])
            ddpm = Deformddpm.module if use_distributed else Deformddpm
            cond_img, _, cond_ratio = ddpm.proc_cond_img(x0,proc_type=proc_type)

            pre_dvf_I,dvf_I = Deformddpm(img_org=x0, t=t, cond_imgs=cond_img, mask=blind_mask,proc_type=[],text=embd_in)  # forward diffusion process

            loss_tot=0

            loss_ddf = loss_reg(pre_dvf_I,img=x0)
            trm_pred = ddf_stn(pre_dvf_I, dvf_I)
            loss_gen_d = loss_dist(pred=trm_pred,inv_lab=dvf_I,ddf_stn=None,mask=blind_mask)
            loss_gen_a = loss_ang(pred=trm_pred,inv_lab=dvf_I,ddf_stn=None,mask=blind_mask)

            loss_tot += LOSS_WEIGHTS_DIFF[0] * loss_gen_a + LOSS_WEIGHTS_DIFF[1] * loss_gen_d
            loss_tot += LOSS_WEIGHTS_DIFF[2] * loss_ddf
            loss_tot = torch.sqrt(1.+MSK_EPS-cond_ratio) * loss_tot

            # >> JZ: print nan in x0
            if torch.isnan(x0).any():
                print(f"*** Encountered NaN in input image x0 at epoch {epoch}, step {step}.")
            # >> JZ: print loss of ddf
            if loss_ddf>0.001:
                print(f"*** High diffusion DDF loss at epoch {epoch}, step {step}: {loss_ddf.item()}.")
            # yu: check if loss_tot==nan or inf
            if torch.isnan(loss_tot) or torch.isinf(loss_tot):
                print(f"*** Encountered NaN or Inf loss at epoch {epoch}, step {step}. Skipping this batch.")
                loss_nan_step += 1
                continue
            if loss_nan_step > 5:
                print(f"*** Too many NaN or Inf losses ({loss_nan_step} times) at epoch {epoch}, step {step}. Stopping training.")
                raise ValueError("Too many NaN losses detected in loss_tot. Code terminated.")

            optimizer.zero_grad(set_to_none=True)     # OPT: set_to_none faster than zero-fill
            loss_tot.backward()
            optimizer.step()

            epoch_loss_tot += loss_tot.item() / total
            epoch_loss_gen_d += loss_gen_d.item() / total
            epoch_loss_gen_a += loss_gen_a.item() / total
            epoch_loss_reg += loss_ddf.item() / total

            # ==========================================================================
            # contrastive train on single image (text-image alignment)
            loss_contra_val = None
            if step % CONTRASTIVE_STEP_RATIO == 0:
                raw_network = Deformddpm.module.network if use_distributed else Deformddpm.network
                n_contra = x0.size()[0]
                t_contra = torch.randint(0, hyp_parameters["timesteps"], (n_contra,)).to(hyp_parameters["device"])
                _ = raw_network(x=(x0 * blind_mask).detach(), y=cond_img.detach(), t=t_contra, text=None)
                if hasattr(raw_network, 'img_embd') and raw_network.img_embd is not None:
                    img_embd = raw_network.img_embd  # [B, 1024]
                    loss_contra = LOSS_WEIGHT_CONTRASTIVE * (1 - F.cosine_similarity(img_embd, embd_dev, dim=-1).mean())

                    optimizer.zero_grad(set_to_none=True)   # OPT
                    loss_contra.backward()
                    torch.nn.utils.clip_grad_norm_(Deformddpm.parameters(), max_norm=0.05)
                    optimizer.step()
                    loss_contra_val = loss_contra.item()
                    epoch_loss_contrastive += loss_contra_val / total
                else:
                    if gpu_id == 0:
                        print(f"*** Warning: Network does not have img_embd attribute for contrastive loss at epoch {epoch}, step {step}.")

            # ==========================================================================
            # registration train on paired images
            if step%train_mode_ratio == 0:
                [x1, y1, _, embd_y] = batch_p
                if np.random.uniform(0,1)<TEXT_EMBED_PROB:
                    embd_y = embd_y.to(hyp_parameters["device"]).type(torch.float32)
                else:
                    embd_y = None

                x1 = x1.to(hyp_parameters["device"]).type(torch.float32)
                y1 = y1.to(hyp_parameters["device"]).type(torch.float32)
                n = x1.size()[0]  # batch_size -> n
                [x1, y1] = utils.random_permute([x1, y1], select_dims=[-1,-2,-3])
                if hyp_parameters['noise_scale']>0:
                    [x1, y1] = thresh_img([x1, y1], [0, 2*hyp_parameters['noise_scale']])
                    random_scale = np.random.normal(1, hyp_parameters['noise_scale'] * 1)
                    random_shift = np.random.normal(0, hyp_parameters['noise_scale'] * 1)
                    x1 = x1 * random_scale + random_shift
                    y1 = y1 * random_scale + random_shift

                scale_regist = np.random.uniform(0.0,0.7)
                # OPT: fixed-length T_regist to avoid XPU dynamic shape recompilation
                # Sample FIXED_T_REGIST_LEN timesteps (was: random 8-16), always same loop length
                t_pool = list(range(int(hyp_parameters["timesteps"] * scale_regist), hyp_parameters["timesteps"]))
                select_timestep = min(FIXED_T_REGIST_LEN, len(t_pool))
                T_regist = sorted(random.sample(t_pool, select_timestep), reverse=True)

                T_regist = [[t for _ in range(max(1, hyp_parameters["batchsize"]//2))] for t in T_regist]

                proc_type = random.choice(['downsample', 'slice', 'slice1', 'none', 'none'])
                ddpm_inner = Deformddpm.module if use_distributed else Deformddpm
                y1_proc, msk_tgt, cond_ratio = ddpm_inner.proc_cond_img(y1,proc_type=proc_type)
                msk_tgt = msk_tgt+MSK_EPS
                [ddf_comp,ddf_rand],[img_rec,img_diff,img_save],_ = Deformddpm(img_org=x1, cond_imgs=y1_proc, T=[None, T_regist], proc_type=[],text=embd_y)  # forward diffusion process
                loss_sim = loss_imgsim(img_rec, y1, label=msk_tgt*(y1>thresh_imgsim))  # calculate loss for the registration process
                loss_mse = loss_imgmse(img_rec, y1, label=msk_tgt*(y1>=0.0))  # calculate loss for the registration process
                loss_ddf1 = loss_reg1(ddf_comp, img=y1)  # calculate loss for the registration process

                loss_regist = 0
                loss_regist += LOSS_WEIGHTS_REGIST[0] * loss_sim
                loss_regist += LOSS_WEIGHTS_REGIST[1] * loss_mse
                loss_regist += LOSS_WEIGHTS_REGIST[2] * loss_ddf1

                # >> JZ: print nan in x0
                if torch.isnan(x0).any():
                    print(f"*** Encountered NaN in input image x0 at epoch {epoch}, step {step}.")
                # >> JZ: print loss of ddf
                if loss_ddf1>0.002:
                    print(f"*** High registration DDF loss at epoch {epoch}, step {step}: {loss_ddf1.item()}.")

                loss_regist = torch.sqrt(cond_ratio+MSK_EPS) *loss_regist
                optimizer.zero_grad(set_to_none=True)   # OPT
                loss_regist.backward()

                torch.nn.utils.clip_grad_norm_(Deformddpm.parameters(), max_norm=0.2)
                optimizer.step()

                epoch_loss_regist += loss_regist.item() / total
                epoch_loss_imgsim += loss_sim.item() / total
                epoch_loss_imgmse += loss_mse.item() / total
                epoch_loss_ddfreg += loss_ddf1.item() / total

            if step % 10 == 0:
                print('step:',step,':', loss_tot.item(),'=',loss_gen_a.item(),'+', loss_gen_d.item(),'+',loss_ddf.item())
                if loss_contra_val is not None:
                    print(f'     loss_contrastive: {loss_contra_val:.6f}')
                print(f'     loss_regist: {loss_regist} = {loss_sim} (imgsim) + {loss_mse} (imgmse) + {loss_ddf1} (ddf)')

        if 1:
            print('==================')
            print(epoch,':', epoch_loss_tot,'=',epoch_loss_gen_a,'+', epoch_loss_gen_d,'+',epoch_loss_reg, ' (ang+dist+regul)')
            print(f'     loss_contrastive: {epoch_loss_contrastive}')
            print(f'     loss_regist: {epoch_loss_regist} = {epoch_loss_imgsim} (imgsim) + {epoch_loss_imgmse} (imgmse) + {epoch_loss_ddfreg} (ddf)')
            print('==================')


        if 0 == epoch % epoch_per_save:
            save_dir=model_save_path + str(epoch).rjust(6, '0') + suffix_pth
            os.makedirs(os.path.dirname(model_save_path), exist_ok=True)
            # break   # FOR TESTING
            if not use_distributed:
                print(f"saved in {save_dir}")
                # torch.save(Deformddpm.state_dict(), save_dir)
                torch.save({
                    'model_state_dict': Deformddpm.state_dict(),
                    'optimizer_state_dict': optimizer.state_dict(),
                    'epoch': epoch
                }, save_dir)
            elif gpu_id == 0:
                print(f"saved in {save_dir}")
                # torch.save(Deformddpm.module.state_dict(), save_dir)
                torch.save({
                    'model_state_dict': Deformddpm.module.state_dict(),
                    'optimizer_state_dict': optimizer.state_dict(),
                    'epoch': epoch
                }, save_dir)

    # Resource cleanup at the end of training
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
    gc.collect()
    if use_distributed and dist.is_initialized():
        dist.destroy_process_group()

def ddp_load_dict(gpu_id, Deformddpm, optimizer, model_file,use_distributed=True, load_strict=False):

    if gpu_id == 0:
    # if 0:
        utils.print_memory_usage("Before Loading Model")
        if torch.cuda.is_available():
            gc.collect()
            torch.cuda.empty_cache()
        checkpoint = torch.load(model_file, map_location='cpu')
        if use_distributed:
            Deformddpm.module.load_state_dict(checkpoint['model_state_dict'], strict=load_strict)
        else:
            Deformddpm.load_state_dict(checkpoint['model_state_dict'], strict=load_strict)
        if load_strict:
            optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
        utils.print_memory_usage("After Loading Checkpoint on GPU")

    if use_distributed:
        # Broadcast model weights from rank 0 to all other GPUs
        dist.barrier()
        for param in Deformddpm.parameters():
            dist.broadcast(param.data, src=0)  # Synchronize model across ranks
        dist.barrier()
        for param_group in optimizer.param_groups:
            for param in param_group['params']:
                if param.grad is not None:
                    dist.broadcast(param.grad, src=0)  # Sync optimizer gradients

    # initial_epoch = checkpoint['epoch'] + 1
    # get the epoch number from the filename and add 1 to set as initial_epoch
    initial_epoch = int(os.path.basename(model_file).split('.')[0][:6]) + 1

    return initial_epoch, Deformddpm, optimizer



if __name__ == "__main__":
    if use_distributed:
        world_size = torch.cuda.device_count()
        print(f"Distributed GPU number = {world_size}")
        mp.spawn(main_train,args = (world_size,),nprocs = world_size)
    else:
        main_train(0,1)