add agedb

config/usb_cv/fullysupervised/fullysupervised_agedb_122_0.yaml

@@ -0,0 +1,46 @@
+algorithm: fullysupervised
+save_dir: ./saved_models/usb_cv/
+save_name: fullysupervised_agedb_122_0
+resume: False
+load_path: ./saved_models/usb_cv//fullysupervised_agedb_122_0/latest_model.pth
+overwrite: True
+use_tensorboard: True
+use_wandb: False
+epoch: 200
+num_train_iter: 204800
+num_log_iter: 256
+num_eval_iter: 2048
+batch_size: 32
+eval_batch_size: 64
+num_warmup_iter: 5120
+num_labels: 122
+uratio: 1
+ema_m: 0.0
+img_size: 224
+crop_ratio: 0.875
+optim: AdamW
+lr: 0.001
+layer_decay: 0.65
+momentum: 0.9
+weight_decay: 0.0005
+amp: False
+clip: 0.0
+use_cat: True
+net: vit_small_patch16_224
+net_from_name: False
+data_dir: ./data/
+dataset: agedb
+train_sampler: RandomSampler
+num_classes: 1
+loss_type: 'l1_loss'
+num_workers: 4
+seed: 0
+world_size: 1
+rank: 0
+multiprocessing_distributed: True
+dist_url: tcp://127.0.0.1:10021
+dist_backend: nccl
+gpu: None
+use_pretrain: True
+pretrain_path: https://github.com/microsoft/Semi-supervised-learning/releases/download/v.0.0.0/vit_small_patch16_224_mlp_im_1k_224.pth
+find_unused_parameters: False

semilearn/core/criterions/regression.py

@@ -2,6 +2,7 @@
 # Licensed under the MIT License.
 
 
+import torch
 import torch.nn as nn
 
 from torch.nn import functional as F
@@ -27,6 +28,44 @@ def l2_loss(logits, target, reduction='mean', **kwargs):
     return loss
 
 
+def focal_l1_loss(logits, target, reduction='mean', activate='sigmoid', beta=0.2, gamma=1.0, **kwargs):
+    """Calculate Focal L1 loss."""
+    target = target.type_as(logits)
+    loss = F.l1_loss(logits, target, reduction='none')
+    loss *= (torch.tanh(beta * torch.abs(logits - target))) ** gamma if activate == 'tanh' else \
+        (2 * torch.sigmoid(beta * torch.abs(logits - target)) - 1) ** gamma
+    if reduction == 'mean':
+        loss = loss.mean()
+    elif reduction == 'sum':
+        loss = loss.sum()
+    return loss
+
+
+def focal_l2_loss(logits, target, reduction='mean', activate='sigmoid', beta=0.2, gamma=1.0, **kwargs):
+    """Calculate Focal L2 loss."""
+    target = target.type_as(logits)
+    loss = F.mse_loss(logits, target, reduction='none')
+    loss *= (torch.tanh(beta * torch.abs(logits - target))) ** gamma if activate == 'tanh' else \
+        (2 * torch.sigmoid(beta * torch.abs(logits - target)) - 1) ** gamma
+    if reduction == 'mean':
+        loss = loss.mean()
+    elif reduction == 'sum':
+        loss = loss.sum()
+    return loss
+
+
+def huber_loss(logits, target, reduction='mean', beta=1.0, **kwargs):
+    """Calculate Smooth L1 loss."""
+    l1_loss = F.l1_loss(logits, target, reduction='none')
+    cond = l1_loss < beta
+    loss = torch.where(cond, 0.5 * l1_loss ** 2 / beta, l1_loss - 0.5 * beta)
+    if reduction == 'mean':
+        loss = loss.mean()
+    elif reduction == 'sum':
+        loss = loss.sum()
+    return loss
+
+
 class RegLoss(nn.Module):
     """
     Wrapper for regression loss
@@ -36,7 +75,7 @@ def __init__(self,
                  **kwargs):
         super(RegLoss, self).__init__()
         self.mode = mode
-        self.loss_list = ["l2_loss", "l1_loss"]
+        self.loss_list = ["l2_loss", "l1_loss", "focal_l1_loss", "focal_l2_loss", "huber_loss"]
         assert mode in self.loss_list
         self.criterion = eval(self.mode)
 

semilearn/core/utils/build.py

@@ -70,7 +70,7 @@ def get_dataset(args, algorithm, dataset, num_labels, num_classes, data_dir='./d
         data_dir: data folder
         include_lb_to_ulb: flag of including labeled data into unlabeled data
     """
-    from semilearn.datasets import get_eurosat, get_medmnist, get_semi_aves, get_cifar, get_svhn, get_stl10, get_imagenet, get_imdb_wiki, get_json_dset, get_pkl_dset
+    from semilearn.datasets import get_agedb, get_eurosat, get_medmnist, get_semi_aves, get_cifar, get_svhn, get_stl10, get_imagenet, get_imdb_wiki, get_json_dset, get_pkl_dset
 
     if dataset == "eurosat":
         lb_dset, ulb_dset, eval_dset = get_eurosat(args, algorithm, dataset, num_labels, num_classes, data_dir=data_dir, include_lb_to_ulb=include_lb_to_ulb)
@@ -96,6 +96,9 @@ def get_dataset(args, algorithm, dataset, num_labels, num_classes, data_dir='./d
     elif dataset in ["imagenet", "imagenet127"]:
         lb_dset, ulb_dset, eval_dset = get_imagenet(args, algorithm, dataset, num_labels, num_classes, data_dir=data_dir, include_lb_to_ulb=include_lb_to_ulb)
         test_dset = None
+    elif dataset == "agedb":
+        lb_dset, ulb_dset, eval_dset = get_agedb(args, algorithm, dataset, num_labels, data_dir=data_dir)
+        test_dset = None
     elif dataset == "imdb_wiki":
         lb_dset, ulb_dset, eval_dset = get_imdb_wiki(args, algorithm, dataset, num_labels, data_dir=data_dir)
         test_dset = None

semilearn/datasets/__init__.py

@@ -2,7 +2,7 @@
 # Licensed under the MIT License.
 
 from semilearn.datasets.utils import split_ssl_data, get_collactor
-from semilearn.datasets.cv_datasets import (get_cifar, get_eurosat, get_imagenet, get_imdb_wiki,
+from semilearn.datasets.cv_datasets import (get_cifar, get_eurosat, get_imagenet, get_agedb, get_imdb_wiki,
     get_medmnist, get_semi_aves, get_stl10, get_svhn, get_food101)
 from semilearn.datasets.nlp_datasets import get_json_dset
 from semilearn.datasets.audio_datasets import get_pkl_dset

semilearn/datasets/cv_datasets/__init__.py

@@ -1,6 +1,7 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT License.
 
+from .agedb import get_agedb
 from .aves import get_semi_aves
 from .cifar import get_cifar
 from .eurosat import get_eurosat

semilearn/datasets/cv_datasets/agedb.py

@@ -0,0 +1,96 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+import os
+import json
+import torchvision
+import numpy as np
+import math
+import pandas as pd
+from PIL import Image
+
+from torchvision import transforms
+from .datasetbase import BasicDataset
+from semilearn.datasets.augmentation import RandAugment, RandomResizedCropAndInterpolation
+from semilearn.datasets.utils import split_ssl_data
+
+
+class AgeDBIDataset(BasicDataset):
+
+    def __init__(self,
+                 alg,
+                 data,
+                 targets=None,
+                 num_classes=None,
+                 transform=None,
+                 is_ulb=False,
+                 strong_transform=None,
+                 onehot=False,
+                 *args, 
+                 **kwargs):
+        super(AgeDBIDataset, self).__init__(alg=alg, data=data, targets=targets, num_classes=num_classes,
+                transform=transform, is_ulb=is_ulb, strong_transform=strong_transform, onehot=onehot, *args, **kwargs)
+        self.data_dir = kwargs.get('data_dir', '')
+
+    def __sample__(self, idx):
+        img = Image.open(os.path.join(self.data_dir, self.data[idx])).convert('RGB')
+        label = np.asarray([self.targets[idx]]).astype('float32')
+        return img, label
+
+
+def get_agedb(args, alg, name=None, num_labels=1000, num_classes=1, data_dir='./data', include_lb_to_ulb=True):
+
+    data_dir = os.path.join(data_dir, 'agedb')
+    df = pd.read_csv(os.path.join(data_dir, "agedb.csv"))
+    df_train, df_val, df_test = df[df['split'] == 'train'], df[df['split'] == 'val'], df[df['split'] == 'test']
+    train_labels, train_data = df_train['age'].tolist(), df_train['path'].tolist()
+    test_labels, test_data = df_test['age'].tolist(), df_test['path'].tolist()
+    # print(df_train['age'].shape, df_test['age'].shape)  # (12208,) (2140,)
+
+    imgnet_mean = (0.485, 0.456, 0.406)
+    imgnet_std = (0.229, 0.224, 0.225)
+    img_size = args.img_size
+    crop_ratio = args.crop_ratio
+
+    transform_weak = transforms.Compose([
+        transforms.Resize((img_size, img_size)),
+        transforms.RandomCrop(img_size, padding=16, padding_mode="reflect"),
+        transforms.RandomHorizontalFlip(),
+        transforms.ToTensor(),
+        transforms.Normalize(imgnet_mean, imgnet_std),
+    ])
+
+    transform_strong = transforms.Compose([
+        transforms.Resize(int(math.floor(img_size / crop_ratio))),
+        RandomResizedCropAndInterpolation((img_size, img_size), scale=(0.2, 1.)),
+        transforms.RandomHorizontalFlip(),
+        RandAugment(3, 10),
+        transforms.ToTensor(),
+        transforms.Normalize(imgnet_mean, imgnet_std)
+    ])
+
+    transform_val = transforms.Compose([
+        transforms.Resize((img_size, img_size)),
+        transforms.ToTensor(),
+        transforms.Normalize(imgnet_mean, imgnet_std),
+    ])
+
+    lb_data, lb_targets, ulb_data, ulb_targets = split_ssl_data(args, train_data, train_labels, num_classes=1, 
+                                                                lb_num_labels=num_labels,
+                                                                ulb_num_labels=args.ulb_num_labels,
+                                                                lb_imbalance_ratio=args.lb_imb_ratio,
+                                                                ulb_imbalance_ratio=args.ulb_imb_ratio,
+                                                                include_lb_to_ulb=include_lb_to_ulb)
+
+    if alg == 'fullysupervised':
+        lb_data = train_data
+        lb_targets = train_labels
+
+    lb_dset = AgeDBIDataset(alg, lb_data, lb_targets, num_classes,
+                            transform_weak, False, None, False, data_dir=data_dir)
+    ulb_dset = AgeDBIDataset(alg, ulb_data, ulb_targets, num_classes,
+                            transform_weak, True, transform_strong, False, data_dir=data_dir)
+    eval_dset = AgeDBIDataset(alg, test_data, test_labels, num_classes,
+                            transform_val, False, None, False, data_dir=data_dir)
+
+    return lb_dset, ulb_dset, eval_dset