From 55e6f944826669ea2a7b111457a81c21b81f2813 Mon Sep 17 00:00:00 2001
From: zwl <15878005+zwl90188229@user.noreply.gitee.com>
Date: Tue, 15 Jul 2025 12:41:34 +0000
Subject: [PATCH] add StreamLearn/Algorithm/PAA/PAA.py.

Signed-off-by: zwl <15878005+zwl90188229@user.noreply.gitee.com>
---
 StreamLearn/Algorithm/PAA/PAA.py | 79 ++++++++++++++++++++++++++++++++
 1 file changed, 79 insertions(+)
 create mode 100644 StreamLearn/Algorithm/PAA/PAA.py

diff --git a/StreamLearn/Algorithm/PAA/PAA.py b/StreamLearn/Algorithm/PAA/PAA.py
new file mode 100644
index 0000000..7a07eac
--- /dev/null
+++ b/StreamLearn/Algorithm/PAA/PAA.py
@@ -0,0 +1,79 @@
+from StreamLearn.Base.SemiEstimator import StreamAlgorithm
+from .paa_model import PAAModel
+from .paa_buffer import PAABuffer
+from .paa_proto import OnlinePrototypeLearning
+from .loss import prototype_contrast_loss, consistency_loss
+import torch
+
+class PAA(StreamAlgorithm):
+    def __init__(self, args):
+        super().__init__()
+        self.args = args
+        self.name = 'PAA'
+        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        
+        # 初始化模型、缓冲区、原型学习
+        self.model = PAAModel(args.num_classes)
+        self.model.to(self.device)
+        self.buffer = PAABuffer(args.buffer_size)
+        self.proto_learner = OnlinePrototypeLearning(args.ema_alpha)
+        
+        # 优化器和损失
+        self.optimizer = torch.optim.Adam(self.model.parameters(), lr=args.lr_paa)
+        self.criterion_ce = torch.nn.CrossEntropyLoss()
+
+    def stream_fit(self, data_batch):
+        """单次流数据训练（核心逻辑）"""
+        images = data_batch['images'].to(self.device)
+        labels = data_batch['labels'].to(self.device)
+        
+        # 前向传播：特征 + 双分类预测
+        feats, logits1, logits2 = self.model(images)
+        
+        # 1. 缓冲区去噪
+        clean_data = self.buffer.denoise(
+            feats.detach().cpu(), 
+            labels.detach().cpu(), 
+            self.args.lambda1, 
+            self.args.lambda2, 
+            self.args.knn_k
+        )
+        clean_feats = torch.tensor(clean_data['feats']).to(self.device)
+        clean_labels = torch.tensor(clean_data['labels']).to(self.device)
+        
+        # 2. 计算损失（CE + 一致性 + 原型对比）
+        loss_ce = self.criterion_ce(logits1, labels)
+        loss_cons = consistency_loss(logits1, logits2)
+        loss_proto = prototype_contrast_loss(clean_feats, self.proto_learner.prototypes, self.args.temp1)
+        total_loss = loss_ce + loss_cons + loss_proto
+        
+        # 3. 反向传播更新模型
+        self.optimizer.zero_grad()
+        total_loss.backward()
+        self.optimizer.step()
+        
+        # 4. 更新原型和缓冲区
+        self.proto_learner.update_prototypes(clean_feats, clean_labels)
+        self.buffer.update(clean_data)
+
+    def stream_evaluate(self, data_batch):
+        """评估准确率"""
+        with torch.no_grad():
+            images = data_batch['images'].to(self.device)
+            labels = data_batch['labels'].to(self.device)
+            _, logits1, _ = self.model(images)
+            preds = torch.argmax(logits1, dim=1)
+            acc = (preds == labels).float().mean()
+            return acc.item()
+
+    def fit(self, stream_dataset):
+        """完整流训练（循环run_time次）"""
+        self.metrics = []
+        for _ in range(self.args.run_time):
+            data_batch = stream_dataset.sample_m()  # 采样batch（需匹配数据集接口）
+            acc = self.stream_evaluate(data_batch)
+            self.stream_fit(data_batch)
+            self.metrics.append(acc)
+
+    def test(self, data):
+        print(f'PAA: {self.metrics}')
\ No newline at end of file
-- 
Gitee