diff --git a/research/deploy.sh b/research/deploy.sh
index a324639..11c4e7a 100755
--- a/research/deploy.sh
+++ b/research/deploy.sh
@@ -1,7 +1,7 @@
 #!/bin/bash
 set -e
 
-DOCKER_IMAGE=registry.gitlab.com/euphon/themblem:research-$(git rev-parse --short HEAD)
+DOCKER_IMAGE=registry.cn-shenzhen.aliyuncs.com/emblem/themblem:research-$(git rev-parse --short HEAD)
 docker build --network=host -t $DOCKER_IMAGE .
 
 docker push $DOCKER_IMAGE
diff --git a/research/models/resnet18_20250325_114510_94.56.pth b/research/models/resnet18_20250325_114510_94.56.pth
new file mode 100644
index 0000000..5835930
Binary files /dev/null and b/research/models/resnet18_20250325_114510_94.56.pth differ
diff --git a/research/roi-train.py b/research/roi-train.py
index 703458f..bb277ab 100755
--- a/research/roi-train.py
+++ b/research/roi-train.py
@@ -2,23 +2,36 @@
 import torch
 import torch.nn as nn
 import torch.optim as optim
-from torch.utils.data import DataLoader, Dataset
+from torch.utils.data import DataLoader, Dataset, random_split
 from torchvision import transforms, models
 from PIL import Image
 import os
 from datetime import datetime
+from collections import defaultdict
+import random
 
 class CustomDataset(Dataset):
-    def __init__(self, img_dir, transform=None):
+    def __init__(self, img_dir, transform=None, limit=500):
         self.img_dir = img_dir
         self.transform = transform
-        self.img_labels = self._load_labels()
+        self.img_labels = self._load_labels(limit)
 
-    def _load_labels(self):
-        # 假设标签存储在labels.txt文件中，每行格式为：图片名 标签
+    def _load_labels(self, limit):
+        cats = defaultdict(list)
         with open(os.path.join(self.img_dir, 'labels.txt'), 'r') as f:
             lines = f.readlines()
-        return [line.strip().split() for line in lines if line.strip()]
+        for line in lines:
+            if not line.strip():
+                continue
+            img_name, label = line.strip().split()
+            cats[label].append([img_name, label])
+        min_samples = min(len(v) for v in cats.values())
+        min_samples = min(limit, min_samples)
+        ret = []
+        for k, v in cats.items():
+            ret.extend(random.sample(v, min_samples))
+            #ret.extend(v)
+        return ret
 
     def __len__(self):
         return len(self.img_labels)
@@ -34,26 +47,33 @@ class CustomDataset(Dataset):
 
 # 数据预处理
 transform_train = transforms.Compose([
-    transforms.RandomResizedCrop((128, 64)),  # 随机裁剪
-    transforms.RandomHorizontalFlip(),       # 随机水平翻转
+    #transforms.RandomResizedCrop((128, 64)),  # 随机裁剪
+    #transforms.RandomHorizontalFlip(),       # 随机水平翻转
+    #transforms.Resize((256, 128)),            # 调整大小
     transforms.ToTensor(),                   # 转换为Tensor
     transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])  # 归一化
 ])
 
 transform_val = transforms.Compose([
-    transforms.Resize((128, 64)),            # 调整大小
+    #transforms.Resize((256, 128)),            # 调整大小
     transforms.ToTensor(),                   # 转换为Tensor
     transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])  # 归一化
 ])
 
 # 加载数据集
 img_dir = os.path.abspath("data/roi/train")
-train_dataset = CustomDataset(img_dir=img_dir, transform=transform_train)
-val_dataset = CustomDataset(img_dir=img_dir, transform=transform_val)
 
-# 创建DataLoader
+full_dataset = CustomDataset(img_dir=img_dir, transform=transform_train, limit=5000)
+train_ratio = 0.5
+val_ratio = 0.5
+
+train_size = int(train_ratio * len(full_dataset))
+val_size = len(full_dataset) - train_size
+
+train_dataset, val_dataset = random_split(full_dataset, [train_size, val_size])
+
 train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)
-val_loader = DataLoader(val_dataset, batch_size=32, shuffle=False)
+val_loader = DataLoader(val_dataset, batch_size=32, shuffle=True)
 
 # 加载预训练的ResNet18模型
 model = models.resnet18(pretrained=True)
@@ -67,12 +87,19 @@ device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
 model = model.to(device)
 
 criterion = nn.CrossEntropyLoss()
-optimizer = optim.Adam(model.parameters(), lr=0.001)
+optimizer = optim.Adam(model.parameters(), lr=0.0001)
+#optimizer = torch.optim.SGD( model.parameters(), lr=0.0001, momentum=0.09, weight_decay=1e-4)
 
-num_epochs = 10
+scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.1)
+
+num_epochs = 15
+
+last_accu = 0
+prev_accu = []
 
 for epoch in range(num_epochs):
     # 训练阶段
+    print(f"Start training epoch {epoch+1}/{num_epochs}")
     model.train()
     running_loss = 0.0
     for images, labels in train_loader:
@@ -86,6 +113,8 @@ for epoch in range(num_epochs):
 
     print(f'Epoch [{epoch+1}/{num_epochs}], Loss: {running_loss/len(train_loader):.4f}')
 
+    scheduler.step()
+
     # 验证阶段
     model.eval()
     correct = 0
@@ -98,7 +127,16 @@ for epoch in range(num_epochs):
             total += labels.size(0)
             correct += (predicted == labels).sum().item()
 
-    print(f'Validation Accuracy: {100 * correct / total:.2f}%')
+    last_accu = 100 * correct / total
+    prev_accu.append(last_accu)
+    if epoch > 5:
+        avg = sum(prev_accu[-5:]) / 5
+        variance = sum((x - avg) ** 2 for x in prev_accu[-5:]) / 5
+        print(f"variance={variance:.4f}")
+        if variance < 1 and not (prev_accu[-1] > prev_accu[-2] and prev_accu[-2] > prev_accu[-3] and prev_accu[-3] > prev_accu[-4]):
+            print(f"Early stopping condition met: {variance:.4f}")
+            break
+    print(f'Validation Accuracy: {last_accu:.2f}%')
 
 dt = datetime.now().strftime("%Y%m%d_%H%M%S")
-torch.save(model.state_dict(), f'data/roi/resnet18_{dt}.pth')
+torch.save(model.state_dict(), f'data/roi/models/resnet18_{dt}_{last_accu:.2f}.pth')
diff --git a/research/roi-verify.py b/research/roi-verify.py
new file mode 100755
index 0000000..19384f6
--- /dev/null
+++ b/research/roi-verify.py
@@ -0,0 +1,30 @@
+#!/usr/bin/env python3
+from PIL import Image
+import os
+import argparse
+import sys
+from roi_lib import *
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='ROI prediction')
+    parser.add_argument('--model', type=str, required=True, help='model path')
+    parser.add_argument('--image', type=str, required=True, help='image file')
+    return parser.parse_args()
+
+# 主函数
+def main():
+    args = parse_args()
+    model = load_model(args.model)
+    image_path = args.image
+    image_tensor = preprocess_image(image_path)
+    predicted_class, probabilities = predict(model, image_tensor)
+    print(f'{image_path} predicted={predicted_class} prob={probabilities}')
+    if predicted_class == 1:
+        print("verify ok")
+        return 0
+    else:
+        print("verify ng")
+        return 1
+
+if __name__ == '__main__':
+    sys.exit(main())
\ No newline at end of file
diff --git a/research/roi.py b/research/roi.py
index 0f8efa4..b6596e5 100755
--- a/research/roi.py
+++ b/research/roi.py
@@ -9,6 +9,7 @@ import json
 import subprocess
 import cv2
 from collections import defaultdict
+from multiprocessing import Pool
 
 class RoiResearch(object):
     def __init__(self, token=None):
@@ -204,22 +205,28 @@ def parse_args():
     parser.add_argument("--id", "-i", type=int, action='append')
     return parser.parse_args()
 
-def get_all_samples():
+def all_sample_ids():
     return os.listdir("data/roi/samples")
 
 def prepare_to_train(id):
     rd = f"data/roi/samples/{id}"
-    frame_file = os.path.abspath(f"{rd}/{id}-frame.jpg")
-    roi_file = os.path.abspath(f"{rd}/{id}-roi.jpg")
-    roi_img = cv2.imread(roi_file)
-    frame_roi_img = frame_roi(frame_file)
-    frame_roi_img = cv2.resize(frame_roi_img, (128, 128))
-    roi_img = cv2.resize(roi_img, (128, 128))
-    side_by_side = np.concatenate((frame_roi_img, roi_img), axis=1)
-    side_by_side_file = os.path.abspath(f"{rd}/{id}-side-by-side.jpg")
-    # show_img(side_by_side, "side_by_side")
-    # cv2.waitKey(0)
-    label_file = os.path.abspath(f"{rd}/label.txt")
+    side_by_side_file = os.path.join(rd, f"{id}-side-by-side.jpg")
+    if not os.path.exists(side_by_side_file):
+        frame_file = os.path.abspath(f"{rd}/{id}-frame.jpg")
+        roi_file = os.path.abspath(f"{rd}/{id}-roi.jpg")
+        if not os.path.exists(frame_file) or not os.path.exists(roi_file):
+            print(f"skipping {id}, no frame or roi")
+            return
+        roi_img = cv2.imread(roi_file)
+        try:
+            frame_roi_img = frame_roi(frame_file)
+        except Exception as e:
+            print(f"failed to get frame_roi for {id}: {e}")
+            return
+        frame_roi_img = cv2.resize(frame_roi_img, (128, 128))
+        roi_img = cv2.resize(roi_img, (128, 128))
+        side_by_side = np.concatenate((frame_roi_img, roi_img), axis=1)
+        cv2.imwrite(side_by_side_file, side_by_side)
     json_file = os.path.abspath(f"{rd}/{id}.json")
     with open(json_file, "r") as f:
         data = json.load(f)
@@ -230,9 +237,9 @@ def prepare_to_train(id):
     elif 'neg' in labels:   
         label = 0
     else:
-        raise Exception("no label found")   
-    side_by_side_file = os.path.abspath(f"data/roi/train/{id}.jpg")
-    cv2.imwrite(side_by_side_file, side_by_side)
+        print(f"no label found for {id}")
+        return
+    shutil.copy(side_by_side_file, os.path.abspath(f"data/roi/train/{id}.jpg"))
     with open(os.path.abspath(f"data/roi/train/labels.txt"), "a") as f:
         f.write(f"{id}.jpg {label}\n")
     return label
@@ -242,19 +249,10 @@ def main():
     if args.download:
         get_roi_data(args)
     if args.preprocess:
-        all_samples = get_all_samples()
-        total = len(all_samples)
         shutil.rmtree(os.path.abspath("data/roi/train"))
         os.makedirs(os.path.abspath("data/roi/train"))
-        label_count = defaultdict(int)
-        for i, id in enumerate(all_samples):
-            print(f"preprocessing {id} ({i + 1}/{total})")
-            try:
-                label = prepare_to_train(id)
-                label_count[label] += 1
-            except Exception as e:
-                print(f"error: {e}")
-        print(f"count by label: {label_count}")
+        with Pool(processes=10) as pool:
+            pool.map(prepare_to_train, all_sample_ids())
         return
     if args.id:
         for id in args.id:
diff --git a/research/roi_bench.py b/research/roi_bench.py
new file mode 100755
index 0000000..7ec93df
--- /dev/null
+++ b/research/roi_bench.py
@@ -0,0 +1,56 @@
+#!/usr/bin/env python3
+from PIL import Image
+import os
+import argparse
+import random
+from roi_lib import *
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='ROI prediction')
+    parser.add_argument('--model', type=str, required=True, help='model path')
+    parser.add_argument('--label', type=str, required=True, help='label file')
+    return parser.parse_args()
+
+# 主函数
+def main():
+    args = parse_args()
+    # 加载模型
+    model = load_model(args.model)
+    with open(args.label, 'r') as f:
+        lines = f.readlines()
+    correct = 0
+    zero_to_one = 0
+    one_to_zero = 0
+    testset = []
+    for line in lines:
+        if not line.strip():
+            continue
+        fs = line.strip().split()
+        if len(fs) != 2:
+            continue
+        image_name = fs[0]
+        label = fs[1]
+        testset.append((image_name, label))
+    random.shuffle(testset)
+    done = 0
+    for image_name, label in testset:
+        image_path = os.path.abspath(os.path.join(os.path.dirname(args.label), image_name))
+        image_tensor = preprocess_image(image_path)
+        predicted_class, probabilities = predict(model, image_tensor)
+        done += 1
+        if str(predicted_class) == label:
+            correct += 1
+        else:
+            if label == '0':
+                zero_to_one += 1
+            else:
+                one_to_zero += 1
+            accu_so_far = correct / done
+            print(f'{image_path} label={label} predicted={predicted_class} prob={probabilities} accuracy so far: {accu_so_far:.3f}')
+    print(f'total={done} correct={correct} accuracy={accu_so_far:.3f}')
+    print(f'zero_to_one={zero_to_one} one_to_zero={one_to_zero}')
+    return 0
+
+if __name__ == '__main__':
+    sys.exit(main())
diff --git a/research/roi_lib.py b/research/roi_lib.py
new file mode 100755
index 0000000..d360d8a
--- /dev/null
+++ b/research/roi_lib.py
@@ -0,0 +1,35 @@
+#!/usr/bin/env python3
+import torch
+import torch.nn as nn
+import torchvision.models as models
+import torchvision.transforms as transforms
+from PIL import Image
+import os
+import argparse
+import random
+
+def load_model(model_path):
+    model = models.resnet18(pretrained=True)
+    num_ftrs = model.fc.in_features
+    model.fc = nn.Linear(num_ftrs, 2)
+    model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))  # 加载模型权重
+    model.eval()
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    model = model.to(device)
+    return model
+
+def preprocess_image(image_path):
+    transform = transforms.Compose([
+        transforms.ToTensor(),                   # 转换为Tensor
+        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])  # 归一化
+    ])
+    image = Image.open(image_path).convert('RGB')  # 打开图像并转换为 RGB
+    image = transform(image).unsqueeze(0)  # 增加 batch 维度
+    return image
+
+def predict(model, image_tensor):
+    with torch.no_grad():  # 禁用梯度计算
+        output = model(image_tensor)
+        _, predicted = torch.max(output, 1)  # 获取预测类别
+        probabilities = torch.nn.functional.softmax(output, dim=1)  # 计算概率
+    return predicted.item(), probabilities.squeeze().tolist()
\ No newline at end of file
diff --git a/research/server.py b/research/server.py
index 29e5d16..d24dcd0 100755
--- a/research/server.py
+++ b/research/server.py
@@ -11,6 +11,7 @@ from io import BytesIO
 import tempfile
 import cv2
 import numpy as np
+from roi_lib import *
 
 app = Flask(__name__)
 
@@ -236,5 +237,20 @@ def analyze():
     analyze_datapoint(data)
     return data
 
+@app.route('/api/roi-verify', methods=['POST'])
+def roi_verify():
+    files = request.files
+    side_by_side_fn = files['side_by_side']
+    with tempfile.NamedTemporaryFile(delete=False, suffix=".jpg") as f:
+        side_by_side_fn.save(f.name)
+        img = Image.open(f.name)
+        img = preprocess_image(img)
+        predicted_class, probabilities = predict(model, image_tensor)
+        return {
+            "ok": True,
+            "predicted_class": predicted_class,
+            "probabilities": probabilities,
+        }
+
 if __name__ == '__main__':
     app.run(host='0.0.0.0', port=26966, debug=True)