def test(): device = torch.device(conf.cuda if torch.cuda.is_available() else "cpu") test_dataset = Testinging_Dataset(conf.data_path_test, conf.test_noise_param, conf.crop_img_size) test_loader = DataLoader(test_dataset, batch_size=1, shuffle=False) print('Loading model from: {}'.format(conf.model_path_test)) model = UNet(in_channels=conf.img_channel, out_channels=conf.img_channel) print('loading model') model.load_state_dict(torch.load(conf.model_path_test)) model.eval() model.to(device) result_dir = conf.denoised_dir if not os.path.exists(result_dir): os.mkdir(result_dir) for batch_idx, (source, img_cropped) in enumerate(test_loader): source_img = tvF.to_pil_image(source.squeeze(0)) img_truth = img_cropped.squeeze(0).numpy().astype(np.uint8) source = source.to(device) denoised_img = model(source).detach().cpu() img_name = test_loader.dataset.image_list[batch_idx] denoised_result = tvF.to_pil_image( torch.clamp(denoised_img.squeeze(0), 0, 1)) fname = os.path.splitext(img_name)[0] source_img.save(os.path.join(result_dir, f'{fname}-noisy.png')) denoised_result.save(os.path.join(result_dir, f'{fname}-denoised.png')) io.imsave(os.path.join(result_dir, f'{fname}-ground_truth.png'), img_truth)
def get_unet(self, use_distributed_data_parallel=True): """ Creates a new network and returns. If machine has multiple GPUs, uses them. """ net = UNet(n_channels=self.channel_count, n_classes=1, bilinear=True, running_on_gpu=(self.gpu_number is not None)) net.to(self.device) if not use_distributed_data_parallel: return net net = nn.parallel.DistributedDataParallel(net, device_ids=[self.gpu_number]) return net
def train(): device = torch.device(conf.cuda if torch.cuda.is_available() else "cpu") dataset = Training_Dataset(conf.data_path_train, conf.gaussian_noise_param, conf.crop_img_size) dataset_length = len(dataset) train_loader = DataLoader(dataset, batch_size=4, shuffle=True, num_workers=4) model = UNet(in_channels=conf.img_channel, out_channels=conf.img_channel) criterion = nn.MSELoss() model = model.to(device) optim = Adam(model.parameters(), lr=conf.learning_rate, betas=(0.9, 0.999), eps=1e-8, weight_decay=0, amsgrad=True) scheduler = lr_scheduler.StepLR(optim, step_size=100, gamma=0.5) model.train() print(model) print("Starting Training Loop...") since = time.time() for epoch in range(conf.max_epoch): print('Epoch {}/{}'.format(epoch, conf.max_epoch - 1)) print('-' * 10) running_loss = 0.0 scheduler.step() for batch_idx, (source, target) in enumerate(train_loader): source = source.to(device) target = target.to(device) optim.zero_grad() denoised_source = model(source) loss = criterion(denoised_source, target) loss.backward() optim.step() running_loss += loss.item() * source.size(0) print('Current loss {} and current batch idx {}'.format( loss.item(), batch_idx)) epoch_loss = running_loss / dataset_length print('{} Loss: {:.4f}'.format('current ' + str(epoch), epoch_loss)) if (epoch + 1) % conf.save_per_epoch == 0: save_model(model, epoch + 1) time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60))
def submit_mnms(model_path, input_data_directory, output_data_directory, device): data_paths = load_path(input_data_directory) net = UNet(n_channels=1, n_classes=4, bilinear=True) net.load_state_dict(torch.load(model_path, map_location=device)) net.to(device) for path in data_paths: ED_np, ES_np = load_phase(path) # HxWxF ED_masks = [] ES_masks = [] for i in range(ED_np.shape[2]): img_np = ED_np[:, :, i] img_tensor = pre_transform(img_np) img_tensor = img_tensor.to(device) mask = predict_img(net, img_tensor) mask = post_transform(img_np, mask[0:3, :, :]) ED_masks.append(mask) for i in range(ES_np.shape[2]): img_np = ES_np[:, :, i] img_tensor = pre_transform(img_np) img_tensor = img_tensor.to(device) mask = predict_img(net, img_tensor) mask = post_transform(img_np, mask[0:3, :, :]) ES_masks.append(mask) ED_masks = np.concatenate(ED_masks, axis=2) ES_masks = np.concatenate(ES_masks, axis=2) save_phase(ED_masks, ES_masks, output_data_directory, path)
def select_model(model_name, init_msg): logger = get_logger() logger.info(init_msg) if model_name == "SETR-PUP": _, model = get_SETR_PUP() elif model_name == "SETR-MLA": _, model = get_SETR_MLA() elif model_name == "TransUNet-Base": model = get_TransUNet_base() elif model_name == "TransUNet-Large": model = get_TransUNet_large() elif model_name == "UNet": model = UNet(CLASS_NUM) model = model.to(device) return logger, model
discriminator_g = GlobalDiscriminator() discriminator_l = LocalDiscriminator() resume=False if(len(sys.argv)>1 and sys.argv[1]=='resume'): resume=True # Load model if available if(resume==True): print('Resuming training....') generator.load_state_dict(torch.load(os.path.join(model_path,'model_gen_latest'))) discriminator_g.load_state_dict(torch.load(os.path.join(model_path,'model_gdis_latest'))) discriminator_l.load_state_dict(torch.load(os.path.join(model_path,'model_ldis_latest'))) device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") generator = generator.to(device) discriminator_g = discriminator_g.to(device) discriminator_l = discriminator_l.to(device) optimizer_g = optim.Adam(discriminator_g.parameters(), lr=0.00005) optimizer_l = optim.Adam(discriminator_l.parameters(), lr=0.00005) gen_optimizer = optim.Adam(generator.parameters(), lr=0.0002) lossdis = nn.BCELoss() lossgen = FocalLoss() lamda = 75 data_loader = load_images(data_path) num_epochs = 2000
class Train(object): def __init__(self, configs): self.batch_size = configs.get("batch_size", "16") self.epochs = configs.get("epochs", "100") self.lr = configs.get("lr", "0.0001") device_args = configs.get("device", "cuda") self.device = torch.device( "cpu" if not torch.cuda.is_available() else device_args) self.workers = configs.get("workers", "4") self.vis_images = configs.get("vis_images", "200") self.vis_freq = configs.get("vis_freq", "10") self.weights = configs.get("weights", "./weights") if not os.path.exists(self.weights): os.mkdir(self.weights) self.logs = configs.get("logs", "./logs") if not os.path.exists(self.weights): os.mkdir(self.weights) self.images_path = configs.get("images_path", "./data") self.is_resize = config.get("is_resize", False) self.image_short_side = config.get("image_short_side", 256) self.is_padding = config.get("is_padding", False) is_multi_gpu = config.get("DateParallel", False) pre_train = config.get("pre_train", False) model_path = config.get("model_path", './weights/unet_idcard_adam.pth') # self.image_size = configs.get("image_size", "256") # self.aug_scale = configs.get("aug_scale", "0.05") # self.aug_angle = configs.get("aug_angle", "15") self.step = 0 self.dsc_loss = DiceLoss() self.model = UNet(in_channels=Dataset.in_channels, out_channels=Dataset.out_channels) if pre_train: self.model.load_state_dict(torch.load(model_path, map_location=self.device), strict=False) if is_multi_gpu: self.model = nn.DataParallel(self.model) self.model.to(self.device) self.best_validation_dsc = 0.0 self.loader_train, self.loader_valid = self.data_loaders() self.params = [p for p in self.model.parameters() if p.requires_grad] self.optimizer = optim.Adam(self.params, lr=self.lr, weight_decay=0.0005) # self.optimizer = torch.optim.SGD(self.params, lr=self.lr, momentum=0.9, weight_decay=0.0005) self.scheduler = lr_scheduler.LR_Scheduler_Head( 'poly', self.lr, self.epochs, len(self.loader_train)) def datasets(self): train_datasets = Dataset( images_dir=self.images_path, # image_size=self.image_size, subset="train", # train transform=get_transforms(train=True), is_resize=self.is_resize, image_short_side=self.image_short_side, is_padding=self.is_padding) # valid_datasets = train_datasets valid_datasets = Dataset( images_dir=self.images_path, # image_size=self.image_size, subset="validation", # validation transform=get_transforms(train=False), is_resize=self.is_resize, image_short_side=self.image_short_side, is_padding=False) return train_datasets, valid_datasets def data_loaders(self): dataset_train, dataset_valid = self.datasets() loader_train = DataLoader( dataset_train, batch_size=self.batch_size, shuffle=True, drop_last=True, num_workers=self.workers, ) loader_valid = DataLoader( dataset_valid, batch_size=1, drop_last=False, num_workers=self.workers, ) return loader_train, loader_valid @staticmethod def dsc_per_volume(validation_pred, validation_true): assert len(validation_pred) == len(validation_true) dsc_list = [] for p in range(len(validation_pred)): y_pred = np.array([validation_pred[p]]) y_true = np.array([validation_true[p]]) dsc_list.append(dsc(y_pred, y_true)) return dsc_list @staticmethod def get_logger(filename, verbosity=1, name=None): level_dict = {0: logging.DEBUG, 1: logging.INFO, 2: logging.WARNING} formatter = logging.Formatter( "[%(asctime)s][%(filename)s][line:%(lineno)d][%(levelname)s] %(message)s" ) logger = logging.getLogger(name) logger.setLevel(level_dict[verbosity]) fh = logging.FileHandler(filename, "w") fh.setFormatter(formatter) logger.addHandler(fh) sh = logging.StreamHandler() sh.setFormatter(formatter) logger.addHandler(sh) return logger def train_one_epoch(self, epoch): self.model.train() loss_train = [] for i, data in enumerate(self.loader_train): self.scheduler(self.optimizer, i, epoch, self.best_validation_dsc) x, y_true = data x, y_true = x.to(self.device), y_true.to(self.device) y_pred = self.model(x) # print('1111', y_pred.size()) # print('2222', y_true.size()) loss = self.dsc_loss(y_pred, y_true) loss_train.append(loss.item()) self.optimizer.zero_grad() loss.backward() self.optimizer.step() # lr_scheduler.step() if self.step % 200 == 0: print('Epoch:[{}/{}]\t iter:[{}]\t loss={:.5f}\t '.format( epoch, self.epochs, i, loss)) self.step += 1 def eval_model(self, patience): self.model.eval() loss_valid = [] validation_pred = [] validation_true = [] # early_stopping = EarlyStopping(patience=patience, verbose=True) for i, data in enumerate(self.loader_valid): x, y_true = data x, y_true = x.to(self.device), y_true.to(self.device) # print(x.size()) # print(333,x[0][2]) with torch.no_grad(): y_pred = self.model(x) loss = self.dsc_loss(y_pred, y_true) # print(y_pred.shape) mask = y_pred > 0.5 mask = mask * 255 mask = mask.cpu().numpy()[0][0] # print(mask) # print(mask.shape()) cv2.imwrite('result.png', mask) loss_valid.append(loss.item()) y_pred_np = y_pred.detach().cpu().numpy() validation_pred.extend( [y_pred_np[s] for s in range(y_pred_np.shape[0])]) y_true_np = y_true.detach().cpu().numpy() validation_true.extend( [y_true_np[s] for s in range(y_true_np.shape[0])]) # early_stopping(loss_valid, self.model) # if early_stopping.early_stop: # print('Early stopping') # import sys # sys.exit(1) mean_dsc = np.mean( self.dsc_per_volume( validation_pred, validation_true, )) # print('mean_dsc:', mean_dsc) if mean_dsc > self.best_validation_dsc: self.best_validation_dsc = mean_dsc torch.save(self.model.state_dict(), os.path.join(self.weights, "unet_xia_adam.pth")) print("Best validation mean DSC: {:4f}".format( self.best_validation_dsc)) def main(self): # print('train is begin.....') # print('load data end.....') # loaders = {"train": loader_train, "valid": loader_valid} for epoch in tqdm(range(self.epochs), total=self.epochs): self.train_one_epoch(epoch) self.eval_model(patience=10) torch.save(self.model.state_dict(), os.path.join(self.weights, "unet_final.pth"))
def train(cont=False): # for tensorboard tracking logger = get_logger() logger.info("(1) Initiating Training ... ") logger.info("Training on device: {}".format(device)) writer = SummaryWriter() # init model aux_layers = None if net == "SETR-PUP": aux_layers, model = get_SETR_PUP() elif net == "SETR-MLA": aux_layers, model = get_SETR_MLA() elif net == "TransUNet-Base": model = get_TransUNet_base() elif net == "TransUNet-Large": model = get_TransUNet_large() elif net == "UNet": model = UNet(CLASS_NUM) # prepare dataset cluster_model = get_clustering_model(logger) train_dataset = CityscapeDataset(img_dir=data_dir, img_dim=IMG_DIM, mode="train", cluster_model=cluster_model) valid_dataset = CityscapeDataset(img_dir=data_dir, img_dim=IMG_DIM, mode="val", cluster_model=cluster_model) train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True) valid_loader = DataLoader(valid_dataset, batch_size=batch_size, shuffle=False) logger.info("(2) Dataset Initiated. ") # optimizer epochs = epoch_num if epoch_num > 0 else iteration_num // len( train_loader) + 1 optim = SGD(model.parameters(), lr=lrate, momentum=momentum, weight_decay=wdecay) # optim = Adam(model.parameters(), lr=lrate) scheduler = lr_scheduler.MultiStepLR( optim, milestones=[int(epochs * fine_tune_ratio)], gamma=0.1) cur_epoch = 0 best_loss = float('inf') epochs_since_improvement = 0 # for continue training if cont: model, optim, cur_epoch, best_loss = load_ckpt_continue_training( best_ckpt_src, model, optim, logger) logger.info("Current best loss: {0}".format(best_loss)) with warnings.catch_warnings(): warnings.simplefilter("ignore") for i in range(cur_epoch): scheduler.step() else: model = nn.DataParallel(model) model = model.to(device) logger.info("(3) Model Initiated ... ") logger.info("Training model: {}".format(net) + ". Training Started.") # loss ce_loss = CrossEntropyLoss() if use_dice_loss: dice_loss = DiceLoss(CLASS_NUM) # loop over epochs iter_count = 0 epoch_bar = tqdm.tqdm(total=epochs, desc="Epoch", position=cur_epoch, leave=True) logger.info("Total epochs: {0}. Starting from epoch {1}.".format( epochs, cur_epoch + 1)) for e in range(epochs - cur_epoch): epoch = e + cur_epoch # Training. model.train() trainLossMeter = LossMeter() train_batch_bar = tqdm.tqdm(total=len(train_loader), desc="TrainBatch", position=0, leave=True) for batch_num, (orig_img, mask_img) in enumerate(train_loader): orig_img, mask_img = orig_img.float().to( device), mask_img.float().to(device) if net == "TransUNet-Base" or net == "TransUNet-Large": pred = model(orig_img) elif net == "SETR-PUP" or net == "SETR-MLA": if aux_layers is not None: pred, _ = model(orig_img) else: pred = model(orig_img) elif net == "UNet": pred = model(orig_img) loss_ce = ce_loss(pred, mask_img[:].long()) if use_dice_loss: loss_dice = dice_loss(pred, mask_img, softmax=True) loss = 0.5 * (loss_ce + loss_dice) else: loss = loss_ce # Backward Propagation, Update weight and metrics optim.zero_grad() loss.backward() optim.step() # update learning rate for param_group in optim.param_groups: orig_lr = param_group['lr'] param_group['lr'] = orig_lr * (1.0 - iter_count / iteration_num)**0.9 iter_count += 1 # Update loss trainLossMeter.update(loss.item()) # print status if (batch_num + 1) % print_freq == 0: status = 'Epoch: [{0}][{1}/{2}]\t' \ 'Loss {loss.val:.4f} ({loss.avg:.4f})\t'.format(epoch+1, batch_num+1, len(train_loader), loss=trainLossMeter) logger.info(status) # log loss to tensorboard if (batch_num + 1) % tensorboard_freq == 0: writer.add_scalar( 'Train_Loss_{0}'.format(tensorboard_freq), trainLossMeter.avg, epoch * (len(train_loader) / tensorboard_freq) + (batch_num + 1) / tensorboard_freq) train_batch_bar.update(1) writer.add_scalar('Train_Loss_epoch', trainLossMeter.avg, epoch) # Validation. model.eval() validLossMeter = LossMeter() valid_batch_bar = tqdm.tqdm(total=len(valid_loader), desc="ValidBatch", position=0, leave=True) with torch.no_grad(): for batch_num, (orig_img, mask_img) in enumerate(valid_loader): orig_img, mask_img = orig_img.float().to( device), mask_img.float().to(device) if net == "TransUNet-Base" or net == "TransUNet-Large": pred = model(orig_img) elif net == "SETR-PUP" or net == "SETR-MLA": if aux_layers is not None: pred, _ = model(orig_img) else: pred = model(orig_img) elif net == "UNet": pred = model(orig_img) loss_ce = ce_loss(pred, mask_img[:].long()) if use_dice_loss: loss_dice = dice_loss(pred, mask_img, softmax=True) loss = 0.5 * (loss_ce + loss_dice) else: loss = loss_ce # Update loss validLossMeter.update(loss.item()) # print status if (batch_num + 1) % print_freq == 0: status = 'Validation: [{0}][{1}/{2}]\t' \ 'Loss {loss.val:.4f} ({loss.avg:.4f})\t'.format(epoch+1, batch_num+1, len(valid_loader), loss=validLossMeter) logger.info(status) # log loss to tensorboard if (batch_num + 1) % tensorboard_freq == 0: writer.add_scalar( 'Valid_Loss_{0}'.format(tensorboard_freq), validLossMeter.avg, epoch * (len(valid_loader) / tensorboard_freq) + (batch_num + 1) / tensorboard_freq) valid_batch_bar.update(1) valid_loss = validLossMeter.avg writer.add_scalar('Valid_Loss_epoch', valid_loss, epoch) logger.info("Validation Loss of epoch [{0}/{1}]: {2}\n".format( epoch + 1, epochs, valid_loss)) # update optim scheduler scheduler.step() # save checkpoint is_best = valid_loss < best_loss best_loss_tmp = min(valid_loss, best_loss) if not is_best: epochs_since_improvement += 1 logger.info("Epochs since last improvement: %d\n" % (epochs_since_improvement, )) if epochs_since_improvement == early_stop_tolerance: break # early stopping. else: epochs_since_improvement = 0 state = { 'epoch': epoch, 'loss': best_loss_tmp, 'model_state_dict': model.state_dict(), 'optimizer_state_dict': optim.state_dict(), } torch.save(state, ckpt_src) logger.info("Checkpoint updated.") best_loss = best_loss_tmp epoch_bar.update(1) writer.close()
import glob import numpy as np import torch import os import cv2 from torchvision import transforms from unet_model import UNet if __name__ == "__main__": # 选择设备,有cuda用cuda,没有就用cpu device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') # 加载网络,图片单通道,分类为1。 net = UNet(n_channels=1, n_classes=1) # 将网络拷贝到deivce中 net.to(device=device) # 加载模型参数 # net.load_state_dict(torch.load('best_model_100X.pth', map_location=device)) net.load_state_dict(torch.load('best_model.pth', map_location=device)) # 测试模式 net.eval() # 读取所有图片路径 tests_path = glob.glob('dataS/test/*.png') # tests_path = glob.glob('data100X/test/*.png') # 遍历素有图片 for test_path in tests_path: # 保存结果地址 save_res_path = test_path.split('.')[0] + '_res.png' # 读取图片 img = cv2.imread(test_path) # 转为灰度图
# load best model weights model.load_state_dict(best_model_wts) return model if __name__ == '__main__': lr = 0.001 model = UNet(n_channels=1) #num_ftrs = model.fc.in_features # Here the size of each output sample is set to 2 # Alternatively it can be generalized to nn.Linear(num_ftrs, len(class_names)) #model.fc = nn.Linear(num_ftrs, 2) model = model.to(device) criterion = nn.CrossEntropyLoss() # Observe that all parameters are being optimized optimizer_ft = optim.Adam(model.parameters(), lr=lr) #, momentum=0.9) # Decay LR by a factor of 0.1 every 7 epochs exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) # Finetune training the convnet and evaluation model = train_model(model, criterion, optimizer_ft,
import numpy as np import torch import matplotlib.pyplot as plt import tqdm import cv2 from unet_model import UNet #device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') device = 'cpu' best_model_name = 'best_model.pt' best_model = torch.load(best_model_name) model = UNet() model.load_state_dict(best_model['state_dict']) model.eval() model.to(device) test_dir = '../data/poster/images/' out_dir = '../data/poster/model/' test_images = [os.path.join(test_dir, x) for x in os.listdir(test_dir)] counter = 0 i = 0 for i in range(len(test_images)): test_image_one = test_images[i] #if 'post' not in test_image_one: # i += 1 # continue #counter += 1 #i += 1 print(i, test_image_one)
class OnePredict(object): def __init__(self, params): self.params = params self.device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") self.model_path = params['model_path'] self.model = UNet(in_channels=3, out_channels=1) self.threshold = 0.5 self.resume() # self.model.eval() self.transform = get_transforms_3() self.is_resize = True self.image_short_side = 1024 self.init_torch_tensor() self.model.eval() def init_torch_tensor(self): torch.set_default_tensor_type('torch.FloatTensor') if torch.cuda.is_available(): self.device = torch.device('cuda') torch.set_default_tensor_type('torch.cuda.FloatTensor') else: self.device = torch.device('cpu') # self.model.to(self.device) def resume(self): self.model.load_state_dict(torch.load(self.model_path, map_location=self.device), strict=False) self.model.to(self.device) def resize_img(self, img): '''输入PIL格式的图片''' width, height = img.size # print('111', img.size) if self.is_resize: if height < width: new_height = self.image_short_side new_width = int(math.ceil(new_height / height * width / 32) * 32) else: new_width = self.image_short_side new_height = int(math.ceil(new_width / width * height / 32) * 32) else: if height < width: scale = int(height / 32) new_image_short_side = scale * 32 new_height = new_image_short_side new_width = int(math.ceil(new_height / height * width / 32) * 32) else: scale = int(width / 32) new_image_short_side = scale * 32 new_width = new_image_short_side new_height = int(math.ceil(new_width / width * height / 32) * 32) # print('test1:', np.array(img)) # print('new:', (new_width, new_height)) resized_img = img.resize((new_width, new_height), Image.ANTIALIAS) # print(new_height, new_width) # print('test2:', np.array(resized_img)) return resized_img def format_output(self): pass @staticmethod def pre_process(img): return img @staticmethod def pad_sample(img): a = img.size[0] b = img.size[1] if a == b: return img diff = (max(a, b) - min(a, b)) / 2.0 if a > b: padding = (0, int(np.floor(diff)), 0, int(np.ceil(diff))) else: padding = (int(np.floor(diff)), 0, int(np.ceil(diff)), 0) img = ImageOps.expand(img, border=padding, fill=0) ##left,top,right,bottom assert img.size[0] == img.size[1] return img def post_process(self, preds, img): mask = preds > self.threshold mask = mask * 255 # print(mask.size()) mask = mask.cpu().numpy()[0][0] # print(mask) # print(mask.shape()) cv2.imwrite('mask.png', mask) mask = np.array(mask, np.uint8) contours, hierarchy = cv2.findContours(mask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE) # print(contours) # img = img.cpu() img = np.array(img, np.uint8) cv2.drawContours(img, contours, -1, (0, 0, 255), 1) cv2.imwrite('result2.png', img) boxes = [] return boxes @staticmethod def demo_visualize(): pass def inference(self, img_path, is_visualize=True, is_format_output=False): img = cv2.imread(img_path, cv2.COLOR_BGR2RGB) img = Image.fromarray(img).convert("RGB") # img = Image.open(img_path).convert("RGB") # print('222', np.array(img)) # img = self.pad_sample(img) img = self.resize_img(img) # print('333', img.size) # print('-----', np.array(img)) ori_img = img img.save('img.png') # img = [img] print('111', np.array(img)) img = self.transform(img) print('222', np.array(img)) img = img.unsqueeze(0) img = img.to(self.device) # print('1111', img.size()) # print(img) # print(img) with torch.no_grad(): s1 = time.time() preds = self.model(img) print(preds) s2 = time.time() print(s2 - s1) # boxes, scores = SegDetectorRepresenter().represent(pred=preds, height=h, width=w, is_output_polygon=False) boxes = self.post_process(preds, ori_img)
class PG(object): def __init__(self, configs, env): self.configs = configs self.env = env self.action_size = (64, 1024) # n_channels=3 for RGB images # n_classes is the number of probabilities you want to get per pixel # - For 1 class and background, use n_classes=1 # - For 2 classes, use n_classes=1 # - For N > 2 classes, use n_classes=N # TODO now I assume input<->output size are equal, which might not be true, so we need some modifications onto Unet if necesary self.actor = UNet( n_channels=3, n_classes=1, bilinear=True ) # [B,C, H_in=372, W_in=1242] -> [B, C, H_out=64, W_out=1024] self.optimizer = Adam(self.actor.parameters(), lr=configs['lr']) self.actor.to(device) def get_action(self, state, deterministic=False): """Given the state, produces an action, the probability of the action, the log probability of the action, and the argmax action""" action_probabilities = self.actor( state) # output size should be [B*H*W] action_probabilities = torch.sigmoid( action_probabilities) # make sure the probs are in range [0,1] # B, _, _, _ = action_probabilities.shape action_probabilities = action_probabilities[:, :, :self. action_size[0], :self. action_size[1]] action_probabilities = torch.squeeze(action_probabilities, 1) # assert action_probabilities.size()[1, 2] == self.action_size, "Actor output the wrong size" # action_probabilities_flat = action_probabilities.contiguous().view(B, -1) # TODO leave this to future process; seems it will get the index max_probability_action = torch.argmax(action_probabilities, dim=-1) if deterministic: # using deteministic policy during test time action = action_probabilities(action_probabilities > 0.5).cpu() else: # using stochastic policy during traning time action_distribution = Bernoulli( action_probabilities ) # this creates a distribution to sample from action = action_distribution.sample().cpu( ) # sample the discrete action and copy it to cpu # Have to deal with situation of 0.0 probabilities because we can't do log 0 z = action_probabilities == 0.0 z = z.float() * 1e-8 log_action_probabilities = torch.log(action_probabilities + z) return action, action_probabilities, log_action_probabilities, max_probability_action def compute_loss(self, obs, act, rew): """make loss function whose gradient, for the right data, is policy gradient""" # TODO we may do not need to calculate it for the second time. act_baseline, _, logp, _ = self.get_action(obs, deterministic=True) # advantage _, rew_baseline, _, _ = self.env.step(act_baseline, obs=obs) advantage = rew.to(device).float() - rew_baseline.to(device).float() loss = logp * Variable(advantage).expand_as(act) loss = loss.mean() return loss def update(self, batch_obs, batch_acts, batch_rews): """take a single policy gradient update step for a batch""" self.optimizer.zero_grad() batch_loss = self.compute_loss( obs=torch.as_tensor(batch_obs, dtype=torch.float32), act=torch.as_tensor(batch_acts, dtype=torch.int32), rew=torch.as_tensor(batch_rews, dtype=torch.int32), ) batch_loss.backward() self.optimizer.step() return batch_loss
output_dir = dir_names.valout_dir + '/' + experiment_name + '/' model_file = model_dir + 'model_20.pth' if not os.path.exists(output_dir): os.makedirs(output_dir) if not load_model: c.force_create(model_dir) c.force_create(tfboard_dir) num_class = 3 # load unet model if not load_model: net = UNet(num_class=num_class) net = net.to(device) else: net = UNet(num_class=num_class) net.load_state_dict( torch.load(model_file, map_location=torch.device(device))) net = net.to(device) net.eval() #Initialize class to convert labels to color images color_transform = Colorize() #Set up data #Define image dataset (reads in full images and segmentations) image_dataset = p.ImageDataset(csv_file=c.train_val_csv) # Split dataset into train and validation