OM_train_2modes-reg.py

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 import DeformDDPM
from Diffusion.networks import get_net, STN
from torchvision.transforms import Lambda
import Diffusion.losses 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
# from torch.distributed import init_process_group
###############
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)

use_distributed = True
# use_distributed = False

EPS = 1e-5
MSK_EPS = 0.01
TEXT_EMBED_PROB = 0.7
AUG_RESAMPLE_PROB = 0.6
LOSS_WEIGHTS_DIFF = [2.0, 1.0, 16]  # [ang, dist, reg]
# LOSS_WEIGHTS_REGIST = [9.0, 1.0, 16.0]  # [imgsim, imgmse, ddf]
LOSS_WEIGHTS_REGIST = [1.0, 0.2, 1e2]  # [imgsim, imgmse, ddf]
DIFF_REG_BATCH_RATIO = 2

# 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,
    )
args = parser.parse_args()
#=======================================================================================================================



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)
        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(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']])
    
    # Data_Loader=get_dataloader(data_name=hyp_parameters['data_name'], mode='train')

    # tsfm = torchvision.transforms.Compose([
    #             torchvision.transforms.ToTensor(),
    #             ])

    # dataset = Data_Loader(target_res = [hyp_parameters["img_size"]]*hyp_parameters["ndims"], transforms=None, noise_scale=hyp_parameters['noise_scale'])
    # train_loader = DataLoader(
    #     dataset,
    #     batch_size=hyp_parameters['batchsize'],
    #     # shuffle=False,
    #     shuffle=True,
    #     drop_last=True,
    # )

    # dataset = OminiDataset_v1(transform=None)
    dataset = OMDataset_indiv(transform=None)
    train_loader = DataLoader(
        dataset,
        batch_size=hyp_parameters['batchsize'],
        shuffle=True,
        drop_last=True,
    )

    # datasetp = OminiDataset_paired(transform=None)
    datasetp = OMDataset_pair(transform=None)
    train_loader_p = DataLoader(
        datasetp,
        batch_size=hyp_parameters['batchsize']//DIFF_REG_BATCH_RATIO,
        shuffle=True,
        drop_last=True,
    )



    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 = STN(
        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.LNCC()
    loss_imgmse = losses.LMSE()

    optimizer = Adam(Deformddpm.parameters(), lr=hyp_parameters["lr"])
    # hyp_parameters["lr"]=0.00000001
    # optimizer_regist = Adam(Deformddpm.parameters(), lr=hyp_parameters["lr"]*0.01)
    # optimizer_regist = SGD(Deformddpm.parameters(), lr=hyp_parameters["lr"]*0.01, momentum=0.98)
    # optimizer = SGD(Deformddpm.parameters(), lr=hyp_parameters["lr"], momentum=0.9)

    # # LR scheduler ----- YHM
    # scheduler = torch.optim.lr_scheduler.CyclicLR(optimizer, hyp_parameters["lr"], hyp_parameters["lr"]*10, step_size_up=500, step_size_down=500, mode='triangular', gamma=1.0, scale_fn=None, scale_mode='cycle', cycle_momentum=True, base_momentum=0.8, max_momentum=0.9, last_epoch=-1)

    # Deformddpm.network.load_state_dict(torch.load('/home/data/jzheng/Adaptive_Motion_Generator-master/models/1000.pth'))

    # 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])
    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
        # 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 in tqdm(enumerate(train_loader)):
        # for step, batch in tqdm(enumerate(train_loader)):
        # for step, batch in enumerate(train_loader_omni):
        for step, (batch, batch_p) in tqdm(enumerate(zip(train_loader, train_loader_p)), total=total):
        
            # x0, _ = batch
            
            
            # ==========================================================================
            # diffusion train on single image

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

            

            n = x0.size()[0]  # batch_size -> n
            x0 = x0.to(hyp_parameters["device"])
            
            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', 'independ', 'downsample', 'slice', 'project', 'none', 'uncon', 'uncon', 'uncon'])
            proc_type = random.choice(['adding', 'downsample', 'slice', 'slice1', 'none', 'uncon', 'uncon', 'uncon'])
            # print('proc_type:', proc_type)
            cond_img, _, cond_ratio = Deformddpm.module.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)  # forward diffusion process

            # print(torch.max(torch.abs(pre_dvf_I)))
            # print(torch.max(torch.abs(dvf_I)))
            
            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()
            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

            # print(loss_gen_a.item())
            # if 0:
            # if loss_gen_a.item() < -0.3 and step%train_mode_ratio == 0:
            if step%train_mode_ratio == 0:
                # ==========================================================================
                # registration train on paired images
                # x1, y1 = next(iter(train_loader_p))
                # [x1, y1, _, embd_y] = next(iter(train_loader_p))
                [x1, y1, _, embd_y] = batch_p
                if np.random.uniform(0,1)<TEXT_EMBED_PROB:
                    # embd_x = embd_x.to(hyp_parameters["device"]).type(torch.float32)
                    embd_y = embd_y.to(hyp_parameters["device"]).type(torch.float32)
                else:
                    # embd_x = None
                    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
                # random deformation + rotation
                # if hyp_parameters["ndims"]>2:
                #     if np.random.uniform(0,1)<0.6:
                #         x1 = utils.random_resample(x1, deform_scale=0)
                #         y1 = utils.random_resample(y1, deform_scale=0)
                # x1 = transformer(x1)
                # y1 = transformer(y1)
                [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
                    # x1 = thresh_img(x1, [0, 2*hyp_parameters['noise_scale']])
                    # x1 = x1 * (np.random.normal(1, hyp_parameters['noise_scale'] * 1)) + np.random.normal(0, hyp_parameters['noise_scale'] * 1)
                    # y1 = thresh_img(y1, [0, 2*hyp_parameters['noise_scale']])
                    # y1 = y1 * (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'


                # scale_regist = np.random.uniform(0.6,1.)
                # T_regist = sorted(random.sample(range(0, int(hyp_parameters["timesteps"] * scale_regist) + 1), 16), reverse=True)
                # print('T_regist (0.6,1) sampling range:', T_regist)
                scale_regist = np.random.uniform(0.0,0.7)
                select_timestep = np.random.randint(8, 17)  # select a random number of timesteps to sample, between 8 and 16
                T_regist = sorted(random.sample(range(int(hyp_parameters["timesteps"] * scale_regist),hyp_parameters["timesteps"]), select_timestep), reverse=True)
                # print('T_regist (0.1,0.7) sampling range:', T_regist)
                # scale_regist = np.random.uniform(0.4,1.)
                # T_regist = [int(hyp_parameters["timesteps"]*scale_regist)]
                # scale_regist = np.random.uniform(0.6,1.)
                # init_T = int(hyp_parameters["timesteps"] * scale_regist)
                # T_regist = sorted(random.sample(range(0, int(hyp_parameters["timesteps"] * scale_regist)), 2)+list(range(init_T,hyp_parameters["timesteps"]+1)), reverse=True)

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

                # print('T_regist:', T_regist)
                # proc_type = random.choice(['adding', 'independ', 'downsample', 'slice', 'project', 'none', 'none'])
                proc_type = random.choice(['downsample', 'slice', 'slice1', 'none', 'none'])
                # proc_type = random.choice(['project'])
                y1_proc, msk_tgt, cond_ratio = Deformddpm.module.proc_cond_img(y1,proc_type=proc_type)
                [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,msk=(msk_tgt+MSK_EPS))  # calculate loss for the registration process
                loss_sim = loss_imgsim(img_rec, y1, label=(y1>thresh_imgsim))  # calculate loss for the registration process
                loss_mse = loss_imgmse(img_rec, y1, label=(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
                # print('proc_type:', proc_type, 'cond_ratio:', cond_ratio.item())
                # print('loss_regist:', loss_regist.item(), 'loss_sim:', loss_sim.item(), 'loss_ddf1:', loss_ddf1.item())
                
                # >> 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()}.")
                    # # Print gradients for each parameter
                    # for name, param in Deformddpm.named_parameters():
                    #     if param.grad is not None:
                    #         print(f"Gradient for {name}: {param.grad.norm()}")
                    #     else:
                    #         print(f"Gradient for {name}: None")

                loss_regist = torch.sqrt(cond_ratio+MSK_EPS) *loss_regist
                optimizer.zero_grad()
                loss_regist.backward()



                torch.nn.utils.clip_grad_norm_(Deformddpm.parameters(), max_norm=0.4)
                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())
                print(f'     loss_regist: {loss_regist} = {loss_sim} (imgsim) + {loss_mse} (imgmse) + {loss_ddf1} (ddf)')
            # break   # FOR TESTING
            # else:
            #     print('loss_gen_a:',loss_gen_a.item())     # FOR TESTING
            #     pass

        if 1:
        # if gpu_id == 0:
            print('==================')
            print(epoch,':', epoch_loss_tot,'=',epoch_loss_gen_a,'+', epoch_loss_gen_d,'+',epoch_loss_reg, ' (ang+dist+regul)')
            print(f'     loss_regist: {epoch_loss_regist} = {epoch_loss_imgsim} (imgsim) + {epoch_loss_imgmse} (imgmse) + {epoch_loss_ddfreg} (ddf)')
            print('==================')
        # # LR schedular step ----- YHM
        # scheduler.step()

        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
    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 1:
            gc.collect()
            torch.cuda.empty_cache()
        # Deformddpm.network.load_state_dict(torch.load(latest_model_file))
        # Deformddpm.load_state_dict(torch.load(latest_model_file), strict=False)
        checkpoint = torch.load(model_file)
        # checkpoint = torch.load(latest_model_file, map_location=f"cuda:{rank}")
        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)