deeplotx 0.5.1__py3-none-any.whl → 0.5.3__py3-none-any.whl

deeplotx/trainer/text_binary_classification_trainer.py

@@ -17,6 +17,8 @@ class TextBinaryClassifierTrainer(BaseTrainer):
         super().__init__(batch_size=batch_size, train_ratio=train_ratio)
         self._long_text_encoder = long_text_encoder
         self.device = self._long_text_encoder.device
+        self.train_dataset_loader = None
+        self.valid_dataset_loader = None
 
     @override
     def train(self, positive_texts: list[str], negative_texts: list[str],
@@ -40,8 +42,9 @@ class TextBinaryClassifierTrainer(BaseTrainer):
         train_size = int(self._train_ratio * dataset_size)
         train_dataset = TensorDataset(inputs[:train_size], labels[:train_size])
         valid_dataset = TensorDataset(inputs[train_size:], labels[train_size:])
-        train_loader = DataLoader(train_dataset, batch_size=self._batch_size, shuffle=True)
-        valid_loader = DataLoader(valid_dataset, batch_size=self._batch_size, shuffle=True)
+        self.train_dataset_loader = DataLoader(train_dataset, batch_size=self._batch_size, shuffle=True)
+        self.valid_dataset_loader = DataLoader(valid_dataset, batch_size=self._batch_size, shuffle=True)
+
         if self.model is not None and self.model.fc1.in_features != feature_dim:
             logger.warning("The dimension of features doesn't match. A new model instance will be created.")
             self.model = None
@@ -55,7 +58,7 @@ class TextBinaryClassifierTrainer(BaseTrainer):
         for epoch in range(num_epochs):
             self.model.train()
             total_loss = 0.0
-            for batch_texts, batch_labels in train_loader:
+            for batch_texts, batch_labels in self.train_dataset_loader:
                 outputs = torch.sigmoid(self.model.forward(batch_texts, self.model.initial_state(batch_texts.shape[0]))[0])
                 loss = loss_function(outputs, batch_labels) + self.model.elastic_net(alpha=alpha, rho=rho)
                 optimizer.zero_grad()
@@ -64,7 +67,7 @@ class TextBinaryClassifierTrainer(BaseTrainer):
                 total_loss += loss.item()
             if epoch % 3 == 0:
                 total_valid_loss = 0.0
-                for batch_texts, batch_labels in valid_loader:
+                for batch_texts, batch_labels in self.valid_dataset_loader:
                     with torch.no_grad():
                         self.model.eval()
                         outputs = torch.sigmoid(self.model.forward(batch_texts, self.model.initial_state(batch_texts.shape[0]))[0])

{deeplotx-0.5.1.dist-info → deeplotx-0.5.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: deeplotx
-Version: 0.5.1
+Version: 0.5.3
 Summary: Easy-2-use long text NLP toolkit.
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
@@ -166,7 +166,10 @@ Dynamic: license-file
         LogisticRegression,  # 逻辑回归 / 二分类 / 多标签分类
         SoftmaxRegression,  # Softmax 回归 / 多分类
         RecursiveSequential,  # 序列模型 / 循环神经网络
-        AutoRegression  # 自回归模型
+        LongContextRecursiveSequential,  # 长上下文序列模型 / 自注意力融合循环神经网络
+        SelfAttention,  # 自注意力模块
+        AutoRegression,  # 自回归模型 / 循环神经网络
+        LongContextAutoRegression  # 长上下文自回归模型 / 自注意力融合循环神经网络
     )
     ```
 
@@ -211,3 +214,84 @@ Dynamic: license-file
             x = self.fc5(x)
             return x
     ```
+
+    自注意力模块:
+
+    ```python
+    from typing_extensions import override
+
+    import torch
+    from torch import nn, softmax
+
+    from deeplotx.nn.base_neural_network import BaseNeuralNetwork
+
+
+    class SelfAttention(BaseNeuralNetwork):
+        def __init__(self, feature_dim: int, model_name: str | None = None,
+                    device: str | None = None, dtype: torch.dtype | None = None):
+            super().__init__(model_name=model_name, device=device, dtype=dtype)
+            self._feature_dim = feature_dim
+            self.q_proj = nn.Linear(in_features=self._feature_dim, out_features=self._feature_dim,
+                                    bias=True, device=self.device, dtype=self.dtype)
+            self.k_proj = nn.Linear(in_features=self._feature_dim, out_features=self._feature_dim,
+                                    bias=True, device=self.device, dtype=self.dtype)
+            self.v_proj = nn.Linear(in_features=self._feature_dim, out_features=self._feature_dim,
+                                    bias=True, device=self.device, dtype=self.dtype)
+
+        def _attention(self, x: torch.Tensor, mask: torch.Tensor | None = None) -> torch.Tensor:
+            q, k = self.q_proj(x), self.k_proj(x)
+            attn = torch.matmul(q, k.transpose(-2, -1))
+            attn = attn / (self._feature_dim ** 0.5)
+            attn = attn.masked_fill(mask == 0, -1e9) if mask is not None else attn
+            return softmax(attn, dim=-1)
+
+        @override
+        def forward(self, x: torch.Tensor, mask: torch.Tensor | None = None) -> torch.Tensor:
+            x = self.ensure_device_and_dtype(x, device=self.device, dtype=self.dtype)
+            if mask is not None:
+                mask = self.ensure_device_and_dtype(mask, device=self.device, dtype=self.dtype)
+            v = self.v_proj(x)
+            return torch.matmul(self._attention(x, mask), v)
+    ```
+
+- ### 使用预定义训练器实现文本二分类任务
+
+    ```python
+    from deeplotx import TextBinaryClassifierTrainer, LongTextEncoder
+    from deeplotx.util import get_files, read_file
+
+    # 定义向量编码策略 (默认使用 bert-base-uncased 作为嵌入模型)
+    long_text_encoder = LongTextEncoder(
+        max_length=2048,  # 最大文本大小, 超出截断
+        chunk_size=448,  # 块大小 (按 Token 计)
+        overlapping=32  # 块间重叠大小 (按 Token 计)
+    )
+
+    trainer = TextBinaryClassifierTrainer(
+        long_text_encoder=long_text_encoder,
+        batch_size=2,
+        train_ratio=0.9  # 训练集和验证集比例
+    )
+
+    # 读取数据
+    pos_data_path = 'path/to/pos_dir'
+    neg_data_path = 'path/to/neg_dir'
+    pos_data = [read_file(x) for x in get_files(pos_data_path)]
+    neg_data = [read_file(x) for x in get_files(neg_data_path)]
+
+    # 开始训练
+    model = trainer.train(pos_data, neg_data, 
+                          num_epochs=36, learning_rate=2e-5,  # 设置训练轮数和学习率
+                          balancing_dataset=True,  # 是否平衡数据集
+                          alpha=1e-4, rho=.2,  # 设置 elastic net 正则化的超参数 alpha 和 rho
+                          hidden_dim=256, recursive_layers=2)  # 设置循环神经网络的结构
+
+    # 保存模型权重
+    model.save(model_name='test_model', model_dir='model')
+
+    # 加载已保存的模型
+    model = model.load(model_name='test_model', model_dir='model')
+
+    # 使用训练好的模型进行预测
+    model.predict(long_text_encoder.encode('这是一个测试文本.', flatten=False))
+    ```

{deeplotx-0.5.1.dist-info → deeplotx-0.5.3.dist-info}/RECORD

@@ -19,12 +19,12 @@ deeplotx/similarity/set.py,sha256=zhGFxtSIXlWqvipBYzoiPahp4g0boAIoUiMfG0wl07A,68
 deeplotx/similarity/vector.py,sha256=WVbDHqykt-fvuILVrhUCtIFAOEjY_zvttrXGM9eylG0,1125
 deeplotx/trainer/__init__.py,sha256=Fl5DR9UecQc5VtBcczU9sx_HtPNoFohpuELOh-Jrsks,77
 deeplotx/trainer/base_trainer.py,sha256=z0MeAT-rRYmjeBXt0ckt7J1itYArR0Cx02wHesXUoZE,385
-deeplotx/trainer/text_binary_classification_trainer.py,sha256=Ktdk4rCNHgTFdXVFmbTnvIlGIJi1gphGRkuRgL2bVOo,4793
+deeplotx/trainer/text_binary_classification_trainer.py,sha256=BNBQdpaD8nB1dQv8naHNIravNcQC8JjOMqD-WRSrUH0,4931
 deeplotx/util/__init__.py,sha256=JxqAK_WOOHcYVSTHBT1-WuBwWrPEVDTV3titeVWvNUM,74
 deeplotx/util/hash.py,sha256=wwsC6kOQvbpuvwKsNQOARd78_wePmW9i3oaUuXRUnpc,352
 deeplotx/util/read_file.py,sha256=ptzouvEQeeW8KU5BrWNJlXw-vFXVrpS9SkAUxsu6A8A,612
-deeplotx-0.5.1.dist-info/licenses/LICENSE,sha256=IwGE9guuL-ryRPEKi6wFPI_zOhg7zDZbTYuHbSt_SAk,35823
-deeplotx-0.5.1.dist-info/METADATA,sha256=LatUJZ1YzKrlPMDNI2UiOqSf5h9mP57kf4f5ngnfa8Q,6954
-deeplotx-0.5.1.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-deeplotx-0.5.1.dist-info/top_level.txt,sha256=hKg4pVDXZ-WWxkRfJFczRIll1Sv7VyfKCmzHLXbuh1U,9
-deeplotx-0.5.1.dist-info/RECORD,,
+deeplotx-0.5.3.dist-info/licenses/LICENSE,sha256=IwGE9guuL-ryRPEKi6wFPI_zOhg7zDZbTYuHbSt_SAk,35823
+deeplotx-0.5.3.dist-info/METADATA,sha256=T5kkUvG4uoe6i9NKO6JiR8hT1A3sxjbrL-Xwnmn4Lrg,10868
+deeplotx-0.5.3.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+deeplotx-0.5.3.dist-info/top_level.txt,sha256=hKg4pVDXZ-WWxkRfJFczRIll1Sv7VyfKCmzHLXbuh1U,9
+deeplotx-0.5.3.dist-info/RECORD,,