diff --git a/.gitignore b/.gitignore
index b27a6b6..31d6ddf 100644
--- a/.gitignore
+++ b/.gitignore
@@ -79,6 +79,8 @@ node_modules/
 logs/
 __pycache__
 filter/runs
+pred/runs
+pred/checkpoints
 data/
 filter/checkpoints
 scripts
diff --git a/pred/count.py b/pred/count.py
new file mode 100644
index 0000000..5ed2d81
--- /dev/null
+++ b/pred/count.py
@@ -0,0 +1,12 @@
+# iterate all json files in ./data/pred
+
+import os
+import json
+
+count = 0
+for filename in os.listdir('./data/pred'):
+    if filename.endswith('.json'):
+        with open('./data/pred/' + filename, 'r') as f:
+            data = json.load(f)
+            count += len(data)
+print(count)
\ No newline at end of file
diff --git a/pred/crawler.py b/pred/crawler.py
new file mode 100644
index 0000000..53008d8
--- /dev/null
+++ b/pred/crawler.py
@@ -0,0 +1,19 @@
+import os
+import requests
+import json
+import time
+
+with open("./pred/2", "r") as fp:
+    raw = fp.readlines()
+    aids = [ int(x.strip()) for x in raw ]
+
+for aid in aids:
+    if os.path.exists(f"./data/pred/{aid}.json"):
+        continue
+    url = f"https://api.bunnyxt.com/tdd/v2/video/{aid}/record?last_count=5000"
+    r = requests.get(url)
+    data = r.json()
+    with open (f"./data/pred/{aid}.json", "w") as fp:
+        json.dump(data, fp, ensure_ascii=False, indent=4)
+    time.sleep(5)
+    print(aid)
\ No newline at end of file
diff --git a/pred/dataset.py b/pred/dataset.py
new file mode 100644
index 0000000..9ed4846
--- /dev/null
+++ b/pred/dataset.py
@@ -0,0 +1,178 @@
+import os
+import json
+import random
+import bisect
+import numpy as np
+import pandas as pd
+import torch
+from torch.utils.data import Dataset
+import datetime
+
+class VideoPlayDataset(Dataset):
+    def __init__(self, data_dir, publish_time_path, term='long', seed=42):
+        if seed is not None:
+            random.seed(seed)
+            np.random.seed(seed)
+            torch.manual_seed(seed)
+        
+        self.data_dir = data_dir
+        self.series_dict = self._load_and_process_data(publish_time_path)
+        self.valid_series = [s for s in self.series_dict.values() if len(s['abs_time']) > 1]
+        self.term = term
+        # Set time window based on term
+        self.time_window = 1000 * 24 * 3600 if term == 'long' else 7 * 24 * 3600
+        MINUTE = 60
+        HOUR = 3600
+        DAY = 24 * HOUR
+
+        if term == 'long':
+            self.feature_windows = [
+                1 * HOUR,
+                6 * HOUR,
+                1 *DAY,
+                3 * DAY,
+                7 * DAY,
+                30 * DAY,
+                100 * DAY
+            ]
+        else:
+            self.feature_windows = [
+                ( 15 * MINUTE,  0 * MINUTE),
+                ( 40 * MINUTE,  0 * MINUTE),
+                ( 1 * HOUR,  0 * HOUR),
+                ( 2 * HOUR,  1 * HOUR),
+                ( 3 * HOUR,  2 * HOUR),
+                ( 3 * HOUR,  0 * HOUR),
+                #( 6 * HOUR,  3 * HOUR),
+                ( 6 * HOUR,  0 * HOUR),
+                (18 * HOUR, 12 * HOUR),
+                #( 1 * DAY,   6 * HOUR),
+                ( 1 * DAY,   0 * DAY),
+                #( 2 * DAY,   1 * DAY),
+                ( 3 * DAY,   0 * DAY),
+                #( 4 * DAY,   1 * DAY),
+                ( 7 * DAY,   0 * DAY)
+            ]
+
+    def _extract_features(self, series, current_idx, target_idx):
+        current_time = series['abs_time'][current_idx]
+        current_play = series['play_count'][current_idx]
+        dt = datetime.datetime.fromtimestamp(current_time)
+
+        if self.term == 'long':
+            time_features = [
+                np.log2(max(current_time - series['create_time'], 1))
+            ]
+        else:
+            time_features = [
+                (dt.hour * 3600 + dt.minute * 60 + dt.second) / 86400,
+                (dt.weekday() * 24 + dt.hour) / 168,
+                np.log2(max(current_time - series['create_time'], 1))
+            ]
+        
+        growth_features = []
+        if self.term == 'long':
+            for window in self.feature_windows:
+                prev_time = current_time - window
+                prev_idx = self._get_nearest_value(series, prev_time, current_idx)
+                if prev_idx is not None:
+                    time_diff = current_time - series['abs_time'][prev_idx]
+                    play_diff = current_play - series['play_count'][prev_idx]
+                    scaled_diff = play_diff / (time_diff / window) if time_diff > 0 else 0.0
+                else:
+                    scaled_diff = 0.0
+                growth_features.append(np.log2(max(scaled_diff, 1)))
+        else:
+            for window_start, window_end in self.feature_windows:
+                prev_time_start = current_time - window_start
+                prev_time_end = current_time - window_end  # window_end is typically 0
+                prev_idx_start = self._get_nearest_value(series, prev_time_start, current_idx)
+                prev_idx_end = self._get_nearest_value(series, prev_time_end, current_idx)
+                if prev_idx_start is not None and prev_idx_end is not None:
+                    time_diff = series['abs_time'][prev_idx_end] - series['abs_time'][prev_idx_start]
+                    play_diff = series['play_count'][prev_idx_end] - series['play_count'][prev_idx_start]
+                    scaled_diff = play_diff / (time_diff / (window_start - window_end)) if time_diff > 0 else 0.0
+                else:
+                    scaled_diff = 0.0
+                growth_features.append(np.log2(max(scaled_diff, 1)))
+        
+        time_diff = series['abs_time'][target_idx] - current_time
+        return [np.log2(max(time_diff, 1))] + [np.log2(current_play + 1)] + growth_features + time_features
+
+    def _load_and_process_data(self, publish_time_path):
+        publish_df = pd.read_csv(publish_time_path)
+        publish_df['published_at'] = pd.to_datetime(publish_df['published_at'])
+        publish_dict = dict(zip(publish_df['aid'], publish_df['published_at']))
+        series_dict = {}
+        for filename in os.listdir(self.data_dir):
+            if not filename.endswith('.json'):
+                continue
+            with open(os.path.join(self.data_dir, filename), 'r') as f:
+                data = json.load(f)
+                if 'code' in data:
+                    continue
+                for item in data:
+                    aid = item['aid']
+                    published_time = pd.to_datetime(publish_dict[aid]).timestamp()
+                    if aid not in series_dict:
+                        series_dict[aid] = {
+                            'abs_time': [],
+                            'play_count': [],
+                            'create_time': published_time
+                        }
+                    series_dict[aid]['abs_time'].append(item['added'])
+                    series_dict[aid]['play_count'].append(item['view'])
+        # Sort each series by absolute time
+        for aid in series_dict:
+            sorted_indices = sorted(range(len(series_dict[aid]['abs_time'])),
+                                key=lambda k: series_dict[aid]['abs_time'][k])
+            series_dict[aid]['abs_time'] = [series_dict[aid]['abs_time'][i] for i in sorted_indices]
+            series_dict[aid]['play_count'] = [series_dict[aid]['play_count'][i] for i in sorted_indices]
+        return series_dict
+
+    def __len__(self):
+        return 100000  # Virtual length for sampling
+
+    def _get_nearest_value(self, series, target_time, current_idx):
+        times = series['abs_time']
+        pos = bisect.bisect_right(times, target_time, 0, current_idx + 1)
+        candidates = []
+        if pos > 0:
+            candidates.append(pos - 1)
+        if pos <= current_idx:
+            candidates.append(pos)
+        if not candidates:
+            return None
+        closest_idx = min(candidates, key=lambda i: abs(times[i] - target_time))
+        return closest_idx
+
+    def __getitem__(self, _idx):
+        while True:
+            series = random.choice(self.valid_series)
+            if len(series['abs_time']) < 2:
+                continue
+            current_idx = random.randint(0, len(series['abs_time']) - 2)
+            current_time = series['abs_time'][current_idx]
+            max_target_time = current_time + self.time_window
+            candidate_indices = []
+            for j in range(current_idx + 1, len(series['abs_time'])):
+                if series['abs_time'][j] > max_target_time:
+                    break
+                candidate_indices.append(j)
+            if not candidate_indices:
+                continue
+            target_idx = random.choice(candidate_indices)
+            break
+        current_play = series['play_count'][current_idx]
+        target_play = series['play_count'][target_idx]
+        target_delta = max(target_play - current_play, 0)
+        return {
+            'features': torch.FloatTensor(self._extract_features(series, current_idx, target_idx)),
+            'target': torch.log2(torch.FloatTensor([target_delta]) + 1)
+        }
+
+def collate_fn(batch):
+    return {
+        'features': torch.stack([x['features'] for x in batch]),
+        'targets': torch.stack([x['target'] for x in batch])
+    }
\ No newline at end of file
diff --git a/pred/export_onnx.py b/pred/export_onnx.py
new file mode 100644
index 0000000..c7b4d59
--- /dev/null
+++ b/pred/export_onnx.py
@@ -0,0 +1,28 @@
+import torch
+import torch.onnx
+from model import CompactPredictor
+
+def export_model(input_size, checkpoint_path, onnx_path):
+    model = CompactPredictor(input_size)
+    model.load_state_dict(torch.load(checkpoint_path))
+
+    dummy_input = torch.randn(1, input_size)
+
+    model.eval()
+
+    torch.onnx.export(model,  # Model to be exported
+                    dummy_input,  # Model input
+                    onnx_path,  # Save path
+                    export_params=True,  # Whether to export model parameters
+                    opset_version=11,  # ONNX opset version
+                    do_constant_folding=True,  # Whether to perform constant folding optimization
+                    input_names=['input'],  # Input node name
+                    output_names=['output'],  # Output node name
+                    dynamic_axes={'input': {0: 'batch_size'},  # Dynamic batch size
+                                    'output': {0: 'batch_size'}})
+
+    print(f"ONNX model has been exported to: {onnx_path}")
+
+if __name__ == '__main__':
+    export_model(10, './pred/checkpoints/long_term.pt', 'long_term.onnx')
+    export_model(12, './pred/checkpoints/short_term.pt', 'short_term.onnx')
diff --git a/pred/inference.py b/pred/inference.py
new file mode 100644
index 0000000..9a3d678
--- /dev/null
+++ b/pred/inference.py
@@ -0,0 +1,32 @@
+import datetime
+import numpy as np
+from model import CompactPredictor
+import torch
+
+def main():
+    model = CompactPredictor(16).to('cpu', dtype=torch.float32)
+    model.load_state_dict(torch.load('./pred/checkpoints/model_20250315_0530.pt'))
+    model.eval()
+    # inference
+    initial = 999269
+    last = initial
+    start_time = '2025-03-15 01:03:21'
+    for i in range(1, 48):
+        hour = i / 0.5
+        sec = hour * 3600
+        time_d = np.log2(sec)
+        data = [time_d, np.log2(initial+1), # time_delta, current_views
+                2.801318, 3.455128, 3.903391, 3.995577, 4.641488, 5.75131, 6.723868, 6.105322, 8.141023, 9.576701, 10.665067, # grows_feat
+                0.043993, 0.72057, 28.000902 # time_feat
+        ]
+        np_arr = np.array([data])
+        tensor = torch.from_numpy(np_arr).to('cpu', dtype=torch.float32)
+        output = model(tensor)
+        num = output.detach().numpy()[0][0]
+        views_pred = int(np.exp2(num)) + initial
+        current_time = datetime.datetime.strptime(start_time, '%Y-%m-%d %H:%M:%S') + datetime.timedelta(hours=hour)
+        print(current_time.strftime('%m-%d %H:%M'), views_pred, views_pred - last)
+        last = views_pred
+
+if __name__ == '__main__':
+    main()
\ No newline at end of file
diff --git a/pred/model.py b/pred/model.py
new file mode 100644
index 0000000..191e3eb
--- /dev/null
+++ b/pred/model.py
@@ -0,0 +1,23 @@
+import torch.nn as nn
+
+class CompactPredictor(nn.Module):
+    def __init__(self, input_size):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.BatchNorm1d(input_size),
+            nn.Linear(input_size, 256),
+            nn.LeakyReLU(0.1),
+            nn.Dropout(0.3),
+            nn.Linear(256, 128),
+            nn.LeakyReLU(0.1),
+            nn.Dropout(0.2),
+            nn.Linear(128, 64),
+            nn.Tanh(),  # Use Tanh to limit the output range
+            nn.Linear(64, 1)
+        )
+        # Initialize the last layer to values close to zero
+        nn.init.uniform_(self.net[-1].weight, -0.01, 0.01)
+        nn.init.constant_(self.net[-1].bias, 0.0)
+
+    def forward(self, x):
+        return self.net(x)
diff --git a/pred/train.py b/pred/train.py
new file mode 100644
index 0000000..b162163
--- /dev/null
+++ b/pred/train.py
@@ -0,0 +1,114 @@
+import random
+import time
+import numpy as np
+from torch.utils.tensorboard import SummaryWriter
+from torch.utils.data import DataLoader
+import torch
+from dataset import VideoPlayDataset, collate_fn
+from pred.model import CompactPredictor
+
+def asymmetricHuberLoss(delta=1.0, beta=1.3):
+    """
+    创建一个可调用的非对称 Huber 损失函数。
+
+    参数：
+        delta (float): Huber 损失的 delta 参数。
+        beta (float): 控制负误差惩罚的系数。
+
+    返回：
+        callable: 可调用的损失函数。
+    """
+    def loss_function(input, target):
+        error = input - target
+        abs_error = torch.abs(error)
+
+        linear_loss = abs_error - 0.5 * delta
+        quadratic_loss = 0.5 * error**2
+
+        loss = torch.where(abs_error < delta, quadratic_loss, linear_loss)
+        loss = torch.where(error < 0, beta * loss, loss)
+
+        return torch.mean(loss)
+
+    return loss_function
+
+def train(model, dataloader, device, epochs=100):
+    writer = SummaryWriter(f'./pred/runs/play_predictor_{time.strftime("%Y%m%d_%H%M")}')
+    optimizer = torch.optim.AdamW(model.parameters(), lr=1e-3, weight_decay=0.01)
+    scheduler = torch.optim.lr_scheduler.OneCycleLR(optimizer, max_lr=1e-3,
+                                                  total_steps=len(dataloader)*30)
+    # Huber loss
+    criterion = asymmetricHuberLoss(delta=1.0, beta=2.1)
+    
+    model.train()
+    global_step = 0
+    for epoch in range(epochs):
+        total_loss = 0.0
+        for batch_idx, batch in enumerate(dataloader):
+            features = batch['features'].to(device)
+            targets = batch['targets'].to(device)
+            
+            optimizer.zero_grad()
+            outputs = model(features)
+            loss = criterion(outputs, targets)
+            loss.backward()
+            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
+            optimizer.step()
+            scheduler.step()
+            
+            total_loss += loss.item()
+            global_step += 1
+            
+            if global_step % 100 == 0:
+                writer.add_scalar('Loss/train', loss.item(), global_step)
+                writer.add_scalar('LR', scheduler.get_last_lr()[0], global_step)
+            if batch_idx % 50 == 0:
+                # Monitor gradients
+                grad_norms = [
+                    torch.norm(p.grad).item() 
+                    for p in model.parameters() if p.grad is not None
+                ]
+                writer.add_scalar('Grad/Norm', sum(grad_norms)/len(grad_norms), global_step)
+                
+                # Monitor parameter values
+                param_means = [torch.mean(p.data).item() for p in model.parameters()]
+                writer.add_scalar('Params/Mean', sum(param_means)/len(param_means), global_step)
+
+                samples_count = len(targets)
+                good = 0
+                for r in range(samples_count):
+                    r = random.randint(0, samples_count-1)
+                    t = float(torch.exp2(targets[r])) - 1
+                    o = float(torch.exp2(outputs[r])) - 1
+                    d = features[r].cpu().numpy()[0]
+                    speed = np.exp2(features[r].cpu().numpy()[8]) / 6
+                    time_diff = np.exp2(d) / 3600
+                    inc = speed * time_diff
+                    model_error = abs(t - o)
+                    reg_error = abs(inc - t)
+                    if model_error < reg_error:
+                        good += 1
+                #print(f"{t:07.1f} | {o:07.1f} | {d:07.1f} | {inc:07.1f} | {good/samples_count*100:.1f}%")
+                writer.add_scalar('Train/WinRate', good/samples_count, global_step)
+            
+        print(f"Epoch {epoch+1} | Avg Loss: {total_loss/len(dataloader):.4f}")
+    
+    writer.close()
+    return model
+
+if __name__ == "__main__":
+    device = 'mps'
+    
+    # Initialize dataset and model
+    dataset = VideoPlayDataset('./data/pred', './data/pred/publish_time.csv', 'short')
+    dataloader = DataLoader(dataset, batch_size=128, shuffle=True, collate_fn=collate_fn)
+    
+    # Get feature dimension
+    sample = next(iter(dataloader))
+    input_size = sample['features'].shape[1]
+    
+    model = CompactPredictor(input_size).to(device)
+    trained_model = train(model, dataloader, device, epochs=18)
+    
+    # Save model
+    torch.save(trained_model.state_dict(), f"./pred/checkpoints/model_{time.strftime('%Y%m%d_%H%M')}.pt")