deeplotx 0.8.1__py3-none-any.whl → 0.8.3__py3-none-any.whl

deeplotx/encoder/long_text_encoder.py

@@ -13,9 +13,9 @@ logger = logging.getLogger('deeplotx.embedding')
 
 
 class LongTextEncoder(Encoder):
-    def __init__(self, max_length: int, chunk_size: int = 448,
-                 overlapping: int = 32, model_name_or_path: str = DEFAULT_BERT,
-                 cache_capacity: int = 64, max_workers: int = 8, device: str | None = None):
+    def __init__(self, chunk_size: int = 448, overlapping: int = 32, max_length: int = -1,
+                 model_name_or_path: str = DEFAULT_BERT, cache_capacity: int = 64,
+                 max_workers: int = 8, device: str | None = None):
         super().__init__(model_name_or_path=model_name_or_path, device=device)
         assert overlapping < chunk_size, f'overlapping ({overlapping}) must be less than chunk size ({chunk_size}).'
         self._max_length = max_length
@@ -41,23 +41,28 @@ class LongTextEncoder(Encoder):
                 _fin_emb_tensor = torch.cat((_fin_emb_tensor.detach().clone(), _emb.detach().clone()), dim=-1)
             return _fin_emb_tensor.squeeze()
 
+        _tmp_max_length = self._max_length
         _text_to_show = text.replace("\n", str())
         logger.debug(f'Embedding \"{_text_to_show if len(_text_to_show) < 128 else _text_to_show[:128] + "..."}\".')
         # read cache
         _text_hash = sha512(text)
         if _text_hash in self._cache:
             return postprocess(self._cache[_text_hash], flatten)
-        _text_to_input_ids = self.tokenizer.encode(text.strip())[:self._max_length]
+        _text_to_input_ids = self.tokenizer.encode(text.strip())
+        # variable length
+        if _tmp_max_length < 0:
+            _tmp_max_length = len(_text_to_input_ids)
+        _text_to_input_ids = _text_to_input_ids[:_tmp_max_length]
         _text_to_input_ids_att_mask = []
         # padding
         pad_token = self.tokenizer.pad_token_type_id
-        if len(_text_to_input_ids) < self._max_length:
-            _text_to_input_ids.extend([pad_token] * (self._max_length - len(_text_to_input_ids)))
+        if len(_text_to_input_ids) < _tmp_max_length:
+            _text_to_input_ids.extend([pad_token] * (_tmp_max_length - len(_text_to_input_ids)))
         pads = _text_to_input_ids.count(pad_token)
-        non_pads = self._max_length - pads
+        non_pads = _tmp_max_length - pads
         _text_to_input_ids_att_mask.extend([1] * non_pads)
         _text_to_input_ids_att_mask.extend([0] * pads)
-        num_chunks = math.ceil(self._max_length / self._chunk_size)
+        num_chunks = math.ceil(_tmp_max_length / self._chunk_size)
         # split chunks
         chunks = []
         for i in range(num_chunks):

deeplotx/encoder/longformer_encoder.py

@@ -3,7 +3,8 @@ import os
 
 import torch
 from torch import nn
-from transformers import LongformerTokenizer, LongformerModel
+from transformers import AutoModel, AutoTokenizer
+from requests.exceptions import ConnectTimeout, SSLError
 
 from deeplotx import __ROOT__
 
@@ -17,18 +18,35 @@ class LongformerEncoder(nn.Module):
         super().__init__()
         self.device = torch.device(device) if device is not None \
             else torch.device('cuda' if torch.cuda.is_available() else 'cpu')
-        self.tokenizer = LongformerTokenizer.from_pretrained(pretrained_model_name_or_path=model_name_or_path,
-                                                             cache_dir=CACHE_PATH, _from_auto=True)
-        self.bert = LongformerModel.from_pretrained(pretrained_model_name_or_path=model_name_or_path,
-                                                    cache_dir=CACHE_PATH, _from_auto=True).to(self.device)
+        try:
+            self.tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path=model_name_or_path,
+                                                           cache_dir=CACHE_PATH, _from_auto=True,
+                                                           trust_remote_code=True)
+            self.encoder = AutoModel.from_pretrained(pretrained_model_name_or_path=model_name_or_path,
+                                                     cache_dir=CACHE_PATH, _from_auto=True,
+                                                     trust_remote_code=True).to(self.device)
+        except ConnectTimeout:
+            self.tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path=model_name_or_path,
+                                                           cache_dir=CACHE_PATH, _from_auto=True,
+                                                           trust_remote_code=True, local_files_only=True)
+            self.encoder = AutoModel.from_pretrained(pretrained_model_name_or_path=model_name_or_path,
+                                                     cache_dir=CACHE_PATH, _from_auto=True,
+                                                     trust_remote_code=True, local_files_only=True).to(self.device)
+        except SSLError:
+            self.tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path=model_name_or_path,
+                                                           cache_dir=CACHE_PATH, _from_auto=True,
+                                                           trust_remote_code=True, local_files_only=True)
+            self.encoder = AutoModel.from_pretrained(pretrained_model_name_or_path=model_name_or_path,
+                                                     cache_dir=CACHE_PATH, _from_auto=True,
+                                                     trust_remote_code=True, local_files_only=True).to(self.device)
         logger.debug(f'{LongformerEncoder.__name__} initialized on device: {self.device}.')
 
     def forward(self, input_ids: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
-        ori_mode = self.bert.training
-        self.bert.eval()
+        ori_mode = self.encoder.training
+        self.encoder.eval()
         with torch.no_grad():
-            res = self.bert.forward(input_ids, attention_mask=attention_mask).last_hidden_state[:, 0, :]
-        self.bert.train(mode=ori_mode)
+            res = self.encoder.forward(input_ids, attention_mask=attention_mask).last_hidden_state[:, 0, :]
+        self.encoder.train(mode=ori_mode)
         return res
 
     def encode(self, text: str) -> torch.Tensor:

deeplotx/nn/attention.py

@@ -36,7 +36,7 @@ class Attention(BaseNeuralNetwork):
         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 torch.softmax(attn, dim=-1)
+        return torch.softmax(attn, dtype=self.dtype, dim=-1)
 
     @override
     def forward(self, x: torch.Tensor, y: torch.Tensor | None = None, mask: torch.Tensor | None = None) -> torch.Tensor:

{deeplotx-0.8.1.dist-info → deeplotx-0.8.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: deeplotx
-Version: 0.8.1
+Version: 0.8.3
 Summary: Easy-2-use long text NLP toolkit.
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
@@ -48,24 +48,23 @@ Dynamic: license-file
 
 - ### 长文本嵌入
 
-    - **基于通用 BERT 的长文本嵌入** (最大支持长度, 无限长, 通过 max_length 定义)
+    - **基于通用 BERT 的长文本嵌入** (最大支持长度, 无限长, 可通过 max_length 限制长度)
 
         ```python
         from deeplotx import LongTextEncoder
 
-        # 最大文本长度为 2048 个 tokens, 块大小为 512 个 tokens, 块间重叠部分为 64 个 tokens.
+        # 块大小为 448 个 tokens, 块间重叠部分为 32 个 tokens.
         encoder = LongTextEncoder(
-            max_length=2048,
-            chunk_size=512,
-            overlapping=64
+            chunk_size=448,
+            overlapping=32
         )
-        # 对 "我是吴子豪, 这是一个测试文本." 计算嵌入, 并展平.
-        encoder.encode('我是吴子豪, 这是一个测试文本.', flatten=True, use_cache=True)
+        # 对 "我是吴子豪, 这是一个测试文本." 计算嵌入, 并堆叠.
+        encoder.encode('我是吴子豪, 这是一个测试文本.', flatten=False)
         ```
 
         输出:
         ```
-        tensor([ 0.5163,  0.2497,  0.5896,  ..., -0.9815, -0.3095,  0.4232])
+        tensor([ 2.2316e-01,  2.0300e-01,  ...,  1.5578e-01, -6.6735e-02])
         ```
 
     - **基于 Longformer 的长文本嵌入** (最大支持长度 4096 个 tokens)
@@ -77,6 +76,11 @@ Dynamic: license-file
         encoder.encode('我是吴子豪, 这是一个测试文本.')
         ```
 
+        输出:
+        ```
+        tensor([-2.7490e-02,  6.6503e-02, ..., -6.5937e-02,  6.7802e-03])
+        ```
+
 - ### 相似性计算
 
     - **基于向量的相似性**
@@ -163,14 +167,16 @@ Dynamic: license-file
 
     ```python
     from deeplotx import (
-        BaseNeuralNetwork,  # 深度神经网络基类
         FeedForward,  # 前馈神经网络
         LinearRegression,  # 线性回归
         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
         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
+        RoPE,  # RoPE 位置编码
+        Attention,  # 自注意力 / 交叉注意力
+        MultiHeadAttention,  # 并行多头注意力
+        RoFormerEncoder,  # Roformer (Transformer + RoPE) 编码器模型
         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
     )
@@ -193,13 +199,13 @@ Dynamic: license-file
                     device: str | None = None, dtype: torch.dtype | None = None):
             super().__init__(in_features=feature_dim, out_features=feature_dim, model_name=model_name, device=device, dtype=dtype)
             self._dropout_rate = dropout_rate
-            self.fc1 = nn.Linear(feature_dim, int(feature_dim * expansion_factor), bias=bias,
-                                device=self.device, dtype=self.dtype)
-            self.fc2 = nn.Linear(int(feature_dim * expansion_factor), feature_dim, bias=bias,
-                                device=self.device, dtype=self.dtype)
-            self.parametric_relu_1 = nn.PReLU(num_parameters=1, init=5e-3,
+            self.up_proj = nn.Linear(in_features=feature_dim, out_features=int(feature_dim * expansion_factor),
+                                    bias=bias, device=self.device, dtype=self.dtype)
+            self.down_proj = nn.Linear(in_features=int(feature_dim * expansion_factor), out_features=feature_dim,
+                                    bias=bias, device=self.device, dtype=self.dtype)
+            self.parametric_relu = nn.PReLU(num_parameters=1, init=5e-3,
                                             device=self.device, dtype=self.dtype)
-            self.layer_norm = nn.LayerNorm(normalized_shape=self.fc1.in_features, eps=1e-9,
+            self.layer_norm = nn.LayerNorm(normalized_shape=self.up_proj.in_features, eps=1e-9,
                                         device=self.device, dtype=self.dtype)
 
         @override
@@ -207,11 +213,11 @@ Dynamic: license-file
             x = self.ensure_device_and_dtype(x, device=self.device, dtype=self.dtype)
             residual = x
             x = self.layer_norm(x)
-            x = self.fc1(x)
-            x = self.parametric_relu_1(x)
+            x = self.up_proj(x)
+            x = self.parametric_relu(x)
             if self._dropout_rate > .0:
                 x = torch.dropout(x, p=self._dropout_rate, train=self.training)
-            return self.fc2(x) + residual
+            return self.down_proj(x) + residual
 
 
     class FeedForward(BaseNeuralNetwork):
@@ -224,7 +230,7 @@ Dynamic: license-file
             self.ffn_layers = nn.ModuleList([FeedForwardUnit(feature_dim=feature_dim,
                                                             expansion_factor=expansion_factor, bias=bias,
                                                             dropout_rate=dropout_rate,
-                                                            device=self.device, dtype=self.dtype)] * num_layers)
+                                                            device=self.device, dtype=self.dtype) for _ in range(num_layers)])
 
         @override
         def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -234,7 +240,7 @@ Dynamic: license-file
             return x
     ```
 
-    自注意力模块:
+    注意力模块:
 
     ```python
     from typing_extensions import override
@@ -243,14 +249,17 @@ Dynamic: license-file
 
     from deeplotx.nn.base_neural_network import BaseNeuralNetwork
     from deeplotx.nn.feed_forward import FeedForward
+    from deeplotx.nn.rope import RoPE, DEFAULT_THETA
 
 
-    class SelfAttention(BaseNeuralNetwork):
-        def __init__(self, feature_dim: int, bias: bool = True, proj_layers: int = 1,
-                    proj_expansion_factor: int | float = 1.5, dropout_rate: float = 0.02,
-                    model_name: str | None = None, device: str | None = None, dtype: torch.dtype | None = None):
+    class Attention(BaseNeuralNetwork):
+        def __init__(self, feature_dim: int, bias: bool = True, positional: bool = True,
+                    proj_layers: int = 1, proj_expansion_factor: int | float = 1.5, dropout_rate: float = 0.02,
+                    model_name: str | None = None, device: str | None = None, dtype: torch.dtype | None = None,
+                    **kwargs):
             super().__init__(in_features=feature_dim, out_features=feature_dim, model_name=model_name,
                             device=device, dtype=dtype)
+            self._positional = positional
             self._feature_dim = feature_dim
             self.q_proj = FeedForward(feature_dim=self._feature_dim, num_layers=proj_layers,
                                     expansion_factor=proj_expansion_factor,
@@ -261,21 +270,27 @@ Dynamic: license-file
             self.v_proj = FeedForward(feature_dim=self._feature_dim, num_layers=proj_layers,
                                     expansion_factor=proj_expansion_factor,
                                     bias=bias, dropout_rate=dropout_rate, device=self.device, dtype=self.dtype)
+            if self._positional:
+                self.rope = RoPE(feature_dim=self._feature_dim, theta=kwargs.get('theta', DEFAULT_THETA),
+                                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)
+        def _attention(self, x: torch.Tensor, y: torch.Tensor, mask: torch.Tensor | None = None) -> torch.Tensor:
+            q, k = self.q_proj(x), self.k_proj(y)
+            if self._positional:
+                q, k = self.rope(q), self.rope(k)
             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 torch.softmax(attn, dim=-1)
+            return torch.softmax(attn, dtype=self.dtype, dim=-1)
 
         @override
-        def forward(self, x: torch.Tensor, mask: torch.Tensor | None = None) -> torch.Tensor:
+        def forward(self, x: torch.Tensor, y: torch.Tensor | None = None, mask: torch.Tensor | None = None) -> torch.Tensor:
             x = self.ensure_device_and_dtype(x, device=self.device, dtype=self.dtype)
+            y = x if y is None else self.ensure_device_and_dtype(y, 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)
+            v = self.v_proj(y)
+            return torch.matmul(self._attention(x, y, mask), v)
     ```
 
 - ### 使用预定义训练器实现文本二分类任务
@@ -284,7 +299,7 @@ Dynamic: license-file
     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)
+    # 定义向量编码策略 (默认使用 FacebookAI/xlm-roberta-base 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)
@@ -306,10 +321,11 @@ Dynamic: license-file
 
     # 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
+                        num_epochs=36, learning_rate=2e-5, 
+                        balancing_dataset=True, alpha=1e-4, 
+                        rho=.2, encoder_layers=2,  # 2 层 Roformer 编码器
+                        attn_heads=8,  # 8 个注意力头
+                        recursive_layers=2)  # 2 层 Bi-LSTM
 
     # 保存模型权重
     model.save(model_name='test_model', model_dir='model')

{deeplotx-0.8.1.dist-info → deeplotx-0.8.3.dist-info}/RECORD

@@ -1,10 +1,10 @@
 deeplotx/__init__.py,sha256=oNeA-vNu5YGiEQg0IcpKEdGh_Y_2uPvo2nqaNL_Zgv8,1159
 deeplotx/encoder/__init__.py,sha256=BrsF5_4O-4pfihYF2wjExDOoAY-03kGJTH-Mhez4tsE,129
 deeplotx/encoder/encoder.py,sha256=oSBdA-MiwMKNfTFJWR-RdvNS0G0qfX-Qchwy4LuwB00,3985
-deeplotx/encoder/long_text_encoder.py,sha256=PFR6jjGyg1N58TQlKsPaNQEd-EDl13Hyhu7A1KtGBbA,3743
-deeplotx/encoder/longformer_encoder.py,sha256=A8FXqd4mdHxSn_o_R689XtpT73ISDT788EgMQRGLC2g,1822
+deeplotx/encoder/long_text_encoder.py,sha256=3ScdKDi65J5tdO8PFCXBjCzNUCLlJRwVhpDR0BrphG4,3951
+deeplotx/encoder/longformer_encoder.py,sha256=NNYLr5I9tdeh0C8Ir7QcbEMU9gDk6U7CiF3Tbg6NEsE,3372
 deeplotx/nn/__init__.py,sha256=01I_yqx9GTa4wy3uNyAqhtxp66tDqxgMLC4Ky5Vnkrg,651
-deeplotx/nn/attention.py,sha256=HZ4nfFtkk7AnJ9nuoDSK6zIlIhZ_hbpZc3o6JQIBqJ8,2861
+deeplotx/nn/attention.py,sha256=R-i-Rd7gnsh6hwXDeYfqLQOJvfSZIGfQbFzRlC91XLo,2879
 deeplotx/nn/auto_regression.py,sha256=uISx29t_zkDGS8s2wvGB6wOGYZitQ4hQ7wyoQl4lcqY,857
 deeplotx/nn/base_neural_network.py,sha256=FjQEDFH810fJS7JV3aLgJZnaMqC6DH--wlBvuj-ghTc,5900
 deeplotx/nn/feed_forward.py,sha256=4ozj7EDalO9pb6JUhZtsJqE0r8bIHFApHRt2zTrl4ho,2931
@@ -27,8 +27,8 @@ deeplotx/trainer/text_binary_classification_trainer.py,sha256=QMLR4cC8NCUP-v7SOY
 deeplotx/util/__init__.py,sha256=5CH4MTeSgsmCe3LPMfvKoSBpwh6jDSBuHVElJvzQzgs,90
 deeplotx/util/hash.py,sha256=qbNU3RLBWGQYFVte9WZBAkZ1BkdjCXiKLDaKPN54KFk,662
 deeplotx/util/read_file.py,sha256=ptzouvEQeeW8KU5BrWNJlXw-vFXVrpS9SkAUxsu6A8A,612
-deeplotx-0.8.1.dist-info/licenses/LICENSE,sha256=IwGE9guuL-ryRPEKi6wFPI_zOhg7zDZbTYuHbSt_SAk,35823
-deeplotx-0.8.1.dist-info/METADATA,sha256=zMKRLmdsEibLnN_hAx3OM7AbX3SiM7X1-8w4eFJGxNY,12251
-deeplotx-0.8.1.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-deeplotx-0.8.1.dist-info/top_level.txt,sha256=hKg4pVDXZ-WWxkRfJFczRIll1Sv7VyfKCmzHLXbuh1U,9
-deeplotx-0.8.1.dist-info/RECORD,,
+deeplotx-0.8.3.dist-info/licenses/LICENSE,sha256=IwGE9guuL-ryRPEKi6wFPI_zOhg7zDZbTYuHbSt_SAk,35823
+deeplotx-0.8.3.dist-info/METADATA,sha256=Lif2B7wUDIQQKWvUt_Vl_XYPlMf_EhskiQcq8ZYv6TQ,13079
+deeplotx-0.8.3.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+deeplotx-0.8.3.dist-info/top_level.txt,sha256=hKg4pVDXZ-WWxkRfJFczRIll1Sv7VyfKCmzHLXbuh1U,9
+deeplotx-0.8.3.dist-info/RECORD,,