train_mit_unet_meanpooled_448.py

import os
import gc

from functools import partial
from tqdm import tqdm

import cv2
import numpy as np

import albumentations as A
from albumentations.pytorch import ToTensorV2
from albumentations import ImageOnlyTransform

import segmentation_models_pytorch as smp

import torch
from torch.utils.data import DataLoader
from torch.cuda.amp import GradScaler
from torch.cuda.amp import autocast
from torch.optim import Adam, SGD, AdamW
from torchvision.transforms.functional import hflip, vflip, rotate

from utils.dataset_v import VesuviusDataset
from utils.image_loaders import get_train_valid_dataset_4_folds, read_images_mask_middle_layers
from utils.set_seed import set_seed
from utils.metrics import calc_cv

from utils.triple_mit_unet_uneven import VesuviusModelTripleMIT_Uneven
from utils.pooled_unet_smp import UnetMeanPooled
from utils.gradual_warmup_scheduler_v2 import get_scheduler, scheduler_step

from torch.utils.tensorboard import SummaryWriter

class CFG:
    device = 'cuda:0'
    
    PATH_TO_DS = '../data_4_folds'
    PATH_TO_SAVE = '../models/'
    exp_name = 'mit_mean_pooling_32_ch_448'
    chans_idxs = [16 + i for i in range(32)]
    in_chans = len(chans_idxs)
    tile_size = 448
    stride = 224
    
    train_batch_size = 1
    grad_accum_size = 6
    
    valid_batch_size = train_batch_size*2
    num_workers = 4
    max_grad_norm = 1_000
    use_amp = True
    
    label_noise = False
    
       
    backbone_name = 'mit_b2'
    decoder_attention_type = 'scse'
    crop_stride = 2
    
    
    encoder_weights = 'imagenet'
    activation = None
    
    criterion = smp.losses.SoftBCEWithLogitsLoss()
    
    max_norm = 1e3
    warmup_factor = 10
    lr = 1e-4 / warmup_factor
    epochs = 35
    
    # ============== augmentation =============
    transformations = {
        "train" : [
        # A.RandomResizedCrop(
        #     size, size, scale=(0.85, 1.0)),
        A.Resize(tile_size, tile_size),
        A.HorizontalFlip(p=0.5),
        A.VerticalFlip(p=0.5),
        A.RandomBrightnessContrast(p=0.75),
        A.ShiftScaleRotate(p=0.75),
        A.OneOf([
                A.GaussNoise(var_limit=[10, 50]),
                A.GaussianBlur(),
                A.MotionBlur(),
                ], p=0.4),
        A.GridDistortion(num_steps=5, distort_limit=0.3, p=0.5),
        #A.CoarseDropout(max_holes=1, max_width=int(tile_size * 0.3), max_height=int(tile_size * 0.3), 
        #                mask_fill_value=0, p=0.5),
        # A.Cutout(max_h_size=int(size * 0.6),
        #          max_w_size=int(size * 0.6), num_holes=1, p=1.0),
        A.Normalize(
            mean= [0] * in_chans,
            std= [1] * in_chans
        ),
        ToTensorV2(transpose_mask=True),
    ],
    "valid" : [
        A.Resize(tile_size, tile_size),
        A.Normalize(
            mean= [0] * in_chans,
            std= [1] * in_chans
        ),
        ToTensorV2(transpose_mask=True),
    ],
    "test" :[
        A.Resize(tile_size, tile_size),
        A.Normalize(
            mean= [0] * in_chans,
            std= [1] * in_chans
        ),
        ToTensorV2(transpose_mask=True),
    ]}

class AverageMeter(object):
    """Computes and stores the average and current value"""

    def __init__(self):
        self.reset()

    def reset(self):
        self.val = 0
        self.avg = 0
        self.sum = 0
        self.count = 0

    def update(self, val, n=1):
        self.val = val
        self.sum += val * n
        self.count += n
        self.avg = self.sum / self.count


def train_fn(train_loader, model, criterion, optimizer, device):
    model.train()

    scaler = GradScaler(enabled=CFG.use_amp)
    losses = AverageMeter()

    for step, (images, labels, _) in tqdm(enumerate(train_loader), total=len(train_loader)):
        images = images.to(device)
        labels = labels.to(device)
        
        if CFG.label_noise:
            angle = np.random.randint(0, 15) / 5
            labels = rotate(labels, angle)
        
        batch_size = labels.size(0)

        with autocast(CFG.use_amp):
            y_preds = model(images)
            loss = criterion(y_preds, labels)

        losses.update(loss.item(), batch_size)
        scaler.scale(loss).backward()

        if (step + 1) % CFG.grad_accum_size == 0 or (step + 1) == len(train_loader):
            grad_norm = torch.nn.utils.clip_grad_norm_(
                model.parameters(), CFG.max_grad_norm)

            scaler.step(optimizer)
            scaler.update()
            optimizer.zero_grad()

    return losses.avg

def valid_fn(valid_loader, model, criterion, img_shape, device):
    mask_pred = np.zeros(img_shape)
    mask_count = np.zeros(img_shape)
    
    model = model.eval()
    losses = AverageMeter()
    
    for step, (images, labels, valid_xyxy) in tqdm(enumerate(valid_loader), total=len(valid_loader)):
        images = images.to(device)
        labels = labels.to(device)      
        
        batch_size = labels.size(0)
        
        with torch.no_grad():
            y_preds = model(images)
            loss = criterion(y_preds, labels)
            
        losses.update(loss.item(), batch_size)
        
        y_preds = torch.sigmoid(y_preds).to('cpu').numpy()
        for i in range(batch_size):
            x1, y1, x2, y2 = valid_xyxy[i, 0].item(), valid_xyxy[i, 1].item(), valid_xyxy[i, 2].item(), valid_xyxy[i, 3].item()
            mask_pred[y1:y2, x1:x2] += y_preds[i].squeeze(0)
            mask_count[y1:y2, x1:x2] += 1

    print(f'mask count min {mask_count.min()}')
    mask_pred /= mask_count
        
    return losses.avg, mask_pred
    

def criterion(y_pred, y_true):
    return CFG.criterion(y_pred, y_true)

def train_on_fold(valid_img, writer):    
    CUR_PATH_TO_SAVE = os.path.join(CFG.PATH_TO_SAVE, CFG.exp_name, f'fold_{valid_img}')    
    os.makedirs(CUR_PATH_TO_SAVE, exist_ok=True)
    
    read_images_mask = partial(read_images_mask_middle_layers, PATH_TO_DS = CFG.PATH_TO_DS, chans_idxs=CFG.chans_idxs, tile_size=CFG.tile_size)
    train_images, train_masks, valid_images, valid_masks, valid_xyxys = get_train_valid_dataset_4_folds(valid_img, read_images_mask, CFG.tile_size, CFG.stride, CFG.stride)
    
    train_dataset = VesuviusDataset(train_images, train_masks, None, A.Compose(CFG.transformations['train']))
    valid_dataset = VesuviusDataset(valid_images, valid_masks, valid_xyxys, A.Compose(CFG.transformations['valid']))

    train_dataloader = DataLoader(train_dataset, 
                              batch_size = CFG.train_batch_size,
                              shuffle=True,
                              num_workers=CFG.num_workers, pin_memory=True, drop_last=True)
    valid_dataloader = DataLoader(valid_dataset, 
                                batch_size=CFG.valid_batch_size,
                                shuffle=False,
                                num_workers=CFG.num_workers, pin_memory=True, drop_last=False)
    
    valid_mask_gt = cv2.imread(os.path.join(CFG.PATH_TO_DS,  f"train/{valid_img}/inklabels.png"), 0)
    valid_mask_gt_shape = valid_mask_gt.shape
    valid_mask_gt = valid_mask_gt / 255
    pad0 = (CFG.tile_size - valid_mask_gt.shape[0] % CFG.tile_size)
    pad1 = (CFG.tile_size - valid_mask_gt.shape[1] % CFG.tile_size)
    valid_mask_gt = np.pad(valid_mask_gt, [(0, pad0), (0, pad1)], constant_values=0)
    
    #model = VesuviusModelTripleMIT_Uneven(CFG.small_backbone_name, CFG.backbone_name, 
    #                                           CFG.decoder_attention_type)
    model = UnetMeanPooled(encoder_name=CFG.backbone_name, decoder_attention_type=CFG.decoder_attention_type, 
                           crop_stride=CFG.crop_stride)
    
    model = model.to(CFG.device)
    
    gc.collect()
    torch.cuda.empty_cache()
    
    optimizer = AdamW(model.parameters(), lr=CFG.lr)
    scheduler = get_scheduler(CFG, optimizer)
    
    for epoch in range(CFG.epochs):
        
        loss = train_fn(train_dataloader, model, criterion, optimizer, CFG.device)

        avg_val_loss, mask_pred = valid_fn(
                valid_dataloader, model, criterion, valid_mask_gt.shape, CFG.device)
        
        scheduler_step(scheduler, epoch)
        
        best_dice, best_th, ths = calc_cv(valid_mask_gt, mask_pred)
        
        print(f'Avg loss - {loss}', f"Avg val loss - {avg_val_loss}")
        print("THs", ths)
        print(f"best_th - {best_th}")
        print(f'best dice - {best_dice}')           
        
        
        if writer:
            for th in ths:
                writer.add_scalar(f'FOLD_{valid_img}/th-{th}', ths[th], epoch)
                
            writer.add_scalar(f'FOLD_{valid_img}/train_loss', loss, epoch)
            writer.add_scalar(f'FOLD_{valid_img}/valid_loss', avg_val_loss, epoch)
            writer.add_scalar(f'FOLD_{valid_img}/f_beta_best', best_dice, epoch)
            writer.add_scalar(f'FOLD_{valid_img}/best_th', best_th, epoch)
        
        if CUR_PATH_TO_SAVE:
            torch.save({'model' : model.state_dict(), 
                        'optimizer' : optimizer.state_dict(),
                        'scheduler' : scheduler.state_dict(),
                        'best_th' : best_th,
                        'best_dice' : best_dice},
                    os.path.join(CUR_PATH_TO_SAVE, f'{CFG.exp_name}_fold_{valid_img}_epoch_{epoch}_model.pth'))
            torch.save({'preds' : mask_pred}, os.path.join(CUR_PATH_TO_SAVE, f'{CFG.exp_name}_fold_{valid_img}_epoch_{epoch}_preds.pth'))
        
        torch.cuda.empty_cache()
        gc.collect()
        

if __name__ == "__main__":
    comment = f'exp_name = {CFG.exp_name}, batch_size = {CFG.train_batch_size}, lr = {CFG.lr}, grad_accum_size = {CFG.grad_accum_size}'
    writer = SummaryWriter(comment=comment)
    train_on_fold(1, writer)
    train_on_fold(2, writer)
    train_on_fold(3, writer)
    train_on_fold(4, writer)
    writer.close()