TF_Skill_01

本文介绍了TensorFlow的一般使用流程,包括如何创建和管理会话,以及神经网络参数的表示和更新。通过TensorFlow的变量tf.Variable初始化权重和偏置,并探讨了损失函数和反向传播的优化方法。

TensorFlow的一般使用流程

  • 构造图:Tensor(类似向量)的使用+operations(图节点)+Graph(构造图)。

  • 张量的使用,存储中间结果;Tensor("name",shape,dtype);构建时并不持有值,运行时持有有效值。

  • 计算图:tf.session(会话)+Tensor的执行要通过sess.run(graph)来执行。

import tensorflow as tf
a = tf.constant([1.0, 2.0],name="a")
b = tf.constant([2.0, 23.0],name="b")
result = a+b
print(a.graph is tf.get_default_graph())
True
g1 = tf.Graph()
with g1.as_default():
    v = tf.get_variable("v", [1], initializer=tf.zeros_initializer())

g2 = tf.Graph()
with g2.as_default():
    v = tf.get_variable("v", [1], initializer=tf.ones_initializer())

with tf.Session(graph=g1)as sess:
    #tf.initialize_all_variables().run()
    tf.global_variables_initializer().run()
    with tf.variable_scope("",reuse=True):
        print(sess.run(tf.get_variable("v")))

with tf.Session(graph=g2)as sess:
    #tf.initialize_all_variables().run()
    tf.global_variables_initializer().run()
    with tf.variable_scope("",reuse=True):
        print(sess.run(tf.get_variable("v")))
[0.]
[1.]
  • Session创建会话一般有两种形式,如下:

python

创建一个会话

sess = tf.Session()

使用这个创建来得到关心的运算的结果

sess.run(…)

关闭会话使得本次运行中使用得到的资源被释放

sess.close()

创建一个会话,并通过python的上下文管理器来管理这个会话

with tf.Session() as sess

使用这个创建来得到关心的运算的结果

sess.run(…)

使用这种形式不在需要closs来关闭释放资源,上下文退出时自动关闭。



```python
import tensorflow as tf
a = tf.constant([1.0, 2.0],name="a")
b = tf.constant([2.0, 23.0],name="b")
result = a+b





<div class="se-preview-section-delimiter"></div>

#计算result的值可以有以下的方法
sess = tf.Session()
with sess.as_default():
    print('method1 = ',result.eval())

print('method2 = ',sess.run(result))
print('method3 = ',result.eval(session=sess))

sess = tf.InteractiveSession()#可以省去将产生的会话注册为默认会话的过程
print('method4 = ',result.eval())
sess.close()




<div class="se-preview-section-delimiter"></div>
method1 =  [ 3. 25.]
method2 =  [ 3. 25.]
method3 =  [ 3. 25.]
method4 =  [ 3. 25.]
  • ConfigProto Protocol Buffer来配置需要生成的会话,具体的实现方式如下:
    python
    config = tf.ConfigProto(allow_soft_placement=True,
    log_device_placement=True)
    sess1 = tf.InteractiveSession(config=config)
    sess2 = tf.Session(config=config)
- 第一个参数 ` allow_soft_placement=True ` 当以下任意一个条件成立时,GPU的运算可以放到CPU上运行。
    - 1.计算无法再GPU上运行;
    - 2.没有GPU资源;
    - 3.运算输入包含对CPU结果的引用。
- 第二个参数 ` log_device_placement ` 为true时,日志会将记录没给个节点被安排在了那个设备上以方便调试。

- TensorFlow的计算图不仅仅可以用来隔离张量和计算,它还提供了管理张量和计算的机制。

- 计算图可以通过 tf.Graph.device函数来指定运行计算的设备。
```python```
g = tf.Graph()




<div class="se-preview-section-delimiter"></div>

# 指定计算运行的设备
with g.device('/gpu:0'):
    result = a + b




<div class="se-preview-section-delimiter"></div>
  • TensorFlow中自动维护的集合列表(collection)
  • 比如通过 tf.add_to_collection 函数可以将资源加入一个 或多个集合中,然后通过 tf.get_collection 获取一个集合里面的所有资源(如张量,变量,或者运行TensorFlow程序所需的队列资源等等)
**集合名称****集合内容****使用场景**
**tf.GraphKeys.VARIABLES** 所有变量 持久化 TensorFlow 模型
**tf.GraphKeys.TRAINABLE_VARIABLES** 可学习的变量(一般指神经网络中的参数) 模型训练、生成模型可视化内容
**tf.GraphKeys.SUMMARIES**日志生成相关的张量 TensorFlow 计算可视化
**tf.GraphKeys.QUEUE_RUNNERS** 处理输入的 QueueRunner 输入处理
**tf.GraphKeys.MOVING_AVERAGE_VARIABLES** 所有计算了滑动平均值的变量 计算变量的滑动平均值

神经网络参数与TensorFlow变量

  • 神经网络中的参数的保存和更新通过变量tf.Variable,一般给使用随机数给变量赋初始值。比如:
    python
    weights = tf.Variable(tf.random_normal([2, 3], mean=0, stddev=2) # 均值为0,标准差为2的随机数
    biases = tf.Variable(tf.zeros([3]) #
    w2 = tf.Variable(weights.initialized_value()) # 初始值与weights一样
    w3 = tf.Variable(weights.initialized_value()*2.0) # 初始值是weights的2倍
- TensorFlow随机数生成函数

| 函数名 | 随机数分布 | 主要参数 | 
| :- | :- | -: | 
| tf.random_normal | 正态分布 | 平均值、标准差、取值类型 | 
| tf.truncated_normal | 满足正态分布的随机值,但若随机值偏离平均值超过2个标准差,则这个数会被重新随机 | 平均值、标准差、取值类型 | 
| tf.random_uniform | 平均分布 | 最大、最小值、取值类型 |
| tf.random_gamma | Gramma分布 | 形状参数alpha、尺度参数beta、取值类型 |

- TensorFlow常数生成函数

| 函数名 | 功能 | 样例 | 
| :- | :- | -: | 
| tf.random_normal | 正态分布 | 平均值、标准差、取值类型 | 
| tf.truncated_normal | 满足正态分布的随机值,但若随机值偏离平均值超过2个标准差,则这个数会被重新随机 | 平均值、标准差、取值类型 | 
| tf.random_uniform | 平均分布 | 最大、最小值、取值类型 |
| tf.random_gamma | Gramma分布 | 形状参数alpha、尺度参数beta、取值类型 |



```python
import tensorflow as tf






<div class="se-preview-section-delimiter"></div>

# 声明w1,w2两个变量。通过seed设定随机种子,保证每次得到相同结果
w1 = tf.Variable(tf.random_normal([2, 3], stddev=1, seed=1))
w2 = tf.Variable(tf.random_normal([3, 1], stddev=1, seed=1))





<div class="se-preview-section-delimiter"></div>

#将输入的特征向量定义为一个变量,但实际上并未被运行,未初始化
x = tf.constant([[0.7, 0.9]])
a = tf.matmul(x, w1)
y = tf.matmul(a, w2)

sess = tf.Session()
sess.run(w1.initializer) # 初始化w1
sess.run(w2.initializer) # 初始化w2
print(sess.run(y))




<div class="se-preview-section-delimiter"></div>

#sess.close()

with sess.as_default():
    tf.global_variables_initializer().run() # 全
    print(sess.run(y))




<div class="se-preview-section-delimiter"></div>
[[3.957578]]
[[0.40506378]]
  • 所有的变量都会被自动的加入GraphKeys.VARIABLES集合,通过tf.global_variables可以拿到所有当前计算图的所有变量。
  • 通过将变量的参数trainable设置为true,变量就会被加入在GraphKeys.TRAINABLE_VARIABLES,一般是需要优化的参数和其他参数(比如迭代次数);TF中的神经网络优化算法会将GraphKeys.TRAINABLE_VARIABLES集合中的变量作为默认的优化对象。
  • 类似张量,维度(shape)是通过设置tf.assign(wq, w2, validate_shape=False)更改,但很少更改;类型(type)不可以更改
  • tf中每轮迭代选取的数据通过常量表示,神经网络可能需要经过几百万轮次甚至几亿轮的迭代,这样计算图将会无比巨大,利用效率低。tf提供了placeholder机制用于提供数据,相当于定义了一个位置,这个位置中的程序在程序运行时再指定。
import tensorflow as tf

w1 = tf.Variable(tf.random_normal([2, 3], stddev=1))
w2 = tf.Variable(tf.random_normal([3, 1], stddev=1))





<div class="se-preview-section-delimiter"></div>

# 定义了placeholder作为存放数据输入的地方,维度不一定需要指定,但指定了可以降低出错的概率
x = tf.placeholder(tf.float32, shape=(1,2),name="input")
a = tf.matmul(x, w1)
y = tf.matmul(a, w2)





<div class="se-preview-section-delimiter"></div>

# placeholder需要feed_dict给相关定位赋数据,这里x可以很方便的每次给计算图提供一个batch的测试数据
sess = tf.Session()
with sess.as_default():
    tf.global_variables_initializer().run()
    print(sess.run(y, feed_dict={x:[[0.7, 0.9]]}))




<div class="se-preview-section-delimiter"></div>
[[-0.5766313]]
  • 在得到一个batch的前向传播结果之后,需要定义一个损失函数来刻画当前的预测值与真实答案之间的差距,然后通过反向传播算法来调整神经网络参数的取值使得差距可以被缩小

python

定义损失函数来刻画预测值与真实值得差距,交叉熵常用的损失函数

cross_entropy = -tf.reduce_mean(y_*tf.log(tf.clip_by_value(y, 1e-10, 1.0)))
learning_rate = 0.001

定义了反向传播的优化方法

train_step = \
tf.train.AdamOptimizer(learning_rate).minimize(cross_entropy)
“`

# -*- coding: utf-8 -*- """ DKT-DSC for Assistment2012 (完整可运行版) 最后更新: 2024-07-01 """ import os import sys import numpy as np import tensorflow.compat.v1 as tf os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" os.environ["CUDA_VISIBLE_DEVICES"] = "0" config = tf.ConfigProto() config.gpu_options.allow_growth = True tf.disable_v2_behavior() try: import psutil HAS_PSUTIL = True except ImportError: HAS_PSUTIL = False print("警告: psutil模块未安装,内存监控功能受限") from scipy.sparse import coo_matrix from tensorflow.contrib import rnn import pandas as pd from tqdm import tqdm from sklearn.metrics import mean_squared_error, r2_score, roc_curve, auc import math import random from datetime import datetime import warnings # 忽略警告 warnings.filterwarnings('ignore') # ==================== 配置部分 ==================== DATA_BASE_PATH = './data/' data_name = 'Assist_2012' # 模拟知识图谱路径(实际使用时替换为真实路径) KNOWLEDGE_GRAPH_PATHS = { 'graphml': './output_assist2012_gat_improved/knowledge_graph.graphml', 'nodes': './output_assist2012_gat_improved/graph_nodes.csv', 'edges': './output_assist2012_gat_improved/graph_edges.csv' } # 创建模拟数据路径 os.makedirs(DATA_BASE_PATH, exist_ok=True) os.makedirs(os.path.dirname(KNOWLEDGE_GRAPH_PATHS['nodes']), exist_ok=True) # ==================== 模拟数据生成 ==================== def generate_mock_data(): """生成模拟数据用于测试""" # 生成模拟训练数据 (300条记录) train_data = pd.DataFrame({ 'user_id': np.repeat(range(10), 30), 'problem_id': np.random.randint(1, 100, 300), 'correct': np.random.randint(0, 2, 300), 'start_time': np.arange(300) # 使用简单递增数字模拟时间戳 }) train_data.to_csv(os.path.join(DATA_BASE_PATH, f'{data_name}_train.csv'), index=False) # 生成模拟测试数据 (100条记录) test_data = pd.DataFrame({ 'user_id': np.repeat(range(5), 20), 'problem_id': np.random.randint(1, 100, 100), 'correct': np.random.randint(0, 2, 100), 'start_time': np.arange(100) + 300 # 时间戳接续训练数据 }) test_data.to_csv(os.path.join(DATA_BASE_PATH, f'{data_name}_test.csv'), index=False) # 生成模拟知识图谱节点数据 node_ids = [f'problem_{i}' for i in range(1, 101)] + \ [f'concept_{i}' for i in range(1, 21)] node_types = ['problem'] * 100 + ['concept'] * 20 mock_node_data = pd.DataFrame({ 'node_id': node_ids, 'type': node_types, 'difficulty': np.random.rand(120), 'avg_accuracy': np.random.rand(120), 'total_attempts': np.random.randint(100, 1000, 120), 'avg_confidence': np.random.rand(120) }) mock_node_data.to_csv(KNOWLEDGE_GRAPH_PATHS['nodes'], index=False) # 生成模拟边数据 sources = np.random.choice(node_ids, 500) targets = np.random.choice(node_ids, 500) weights = np.random.rand(500) mock_edge_data = pd.DataFrame({ 'source': sources, 'target': targets, 'weight': weights }) mock_edge_data.to_csv(KNOWLEDGE_GRAPH_PATHS['edges'], index=False) # 检查并生成模拟数据 if not os.path.exists(os.path.join(DATA_BASE_PATH, f'{data_name}_train.csv')): print("[系统] 检测到缺少数据文件,正在生成模拟数据...") generate_mock_data() # ==================== Flags配置 ==================== tf.flags.DEFINE_float("epsilon", 1e-8, "Adam优化器的epsilon值") tf.flags.DEFINE_float("l2_lambda", 0.003, "L2正则化系数") tf.flags.DEFINE_float("learning_rate", 2e-4, "学习率") tf.flags.DEFINE_float("max_grad_norm", 5.0, "梯度裁剪阈值") tf.flags.DEFINE_float("keep_prob", 0.7, "Dropout保留概率") tf.flags.DEFINE_integer("hidden_layer_num", 2, "隐藏层数量") tf.flags.DEFINE_integer("hidden_size", 64, "隐藏层大小") tf.flags.DEFINE_integer("evaluation_interval", 1, "评估间隔周期数") tf.flags.DEFINE_integer("batch_size", 32, "批次大小") # 减小批次大小以便在模拟数据上运行 tf.flags.DEFINE_integer("problem_len", 20, "问题序列长度") tf.flags.DEFINE_integer("epochs", 5, "训练周期数") # 减少epoch以便快速测试 tf.flags.DEFINE_boolean("allow_soft_placement", True, "允许软设备放置") tf.flags.DEFINE_boolean("log_device_placement", False, "记录设备放置信息") tf.flags.DEFINE_string("train_data_path", os.path.join(DATA_BASE_PATH, f'{data_name}_train.csv'), "训练数据路径") tf.flags.DEFINE_string("test_data_path", os.path.join(DATA_BASE_PATH, f'{data_name}_test.csv'), "测试数据路径") FLAGS = tf.flags.FLAGS # 焦点损失参数 FOCAL_LOSS_GAMMA = 2.0 FOCAL_LOSS_ALPHA = 0.25 # 学习率衰减参数 DECAY_STEPS = 100 DECAY_RATE = 0.97 # 早停参数 EARLY_STOP_PATIENCE = 3 def memory_usage(): if HAS_PSUTIL: try: process = psutil.Process(os.getpid()) return process.memory_info().rss / (1024 ** 2) except: return 0.0 return 0.0 # ==================== 时间戳处理工具函数 ==================== def parse_timestamp(timestamp_str): """尝试多种格式解析时间戳""" if isinstance(timestamp_str, (int, float, np.number)): return float(timestamp_str) if isinstance(timestamp_str, str): timestamp_str = timestamp_str.strip('"\' ') # 尝试常见时间格式 for fmt in ('%Y-%m-%d %H:%M:%S', '%m/%d/%Y %H:%M', '%Y-%m-%d', '%s'): try: if fmt == '%s': # Unix时间戳 return float(timestamp_str) dt = datetime.strptime(timestamp_str, fmt) return dt.timestamp() except ValueError: continue return np.nan # ==================== 知识图谱加载器 ==================== class KnowledgeGraphLoader: def __init__(self): self.node_features = None self.adj_matrix = None self.problem_to_node = {} self.node_id_map = {} self.static_node_count = 0 self._rows = None self._cols = None def load(self): print("\n[KG] 加载知识图谱...") try: if not os.path.exists(KNOWLEDGE_GRAPH_PATHS['nodes']): raise FileNotFoundError(f"节点文件未找到: {KNOWLEDGE_GRAPH_PATHS['nodes']}") if not os.path.exists(KNOWLEDGE_GRAPH_PATHS['edges']): raise FileNotFoundError(f"边文件未找到: {KNOWLEDGE_GRAPH_PATHS['edges']}") node_df = pd.read_csv(KNOWLEDGE_GRAPH_PATHS['nodes']) self.static_node_count = len(node_df) print(f"[KG] 总节点数: {self.static_node_count}") # 处理空值 print("[KG] 处理特征空值...") feature_cols = [col for col in node_df.columns if col not in ['node_id', 'type']] for col in feature_cols: if node_df[col].isna().any(): if 'accuracy' in col or 'confidence' in col: median_val = node_df[col].median() node_df[col] = node_df[col].fillna(median_val) else: for node_type in ['problem', 'concept']: mask = node_df['type'] == node_type type_median = node_df.loc[mask, col].median() node_df.loc[mask, col] = node_df.loc[mask, col].fillna(type_median) # 特征标准化 raw_features = node_df[feature_cols].values raw_features = np.nan_to_num(raw_features) feature_mean = np.mean(raw_features, axis=0) feature_std = np.std(raw_features, axis=0) + 1e-8 self.node_features = np.array( (raw_features - feature_mean) / feature_std, dtype=np.float32 ) # 创建映射 self.node_id_map = {row['node_id']: idx for idx, row in node_df.iterrows()} # 创建问题映射 self.problem_to_node = {} problem_count = 0 for idx, row in node_df.iterrows(): if row['type'] == 'problem': try: problem_id = int(row['node_id'].split('_')[1]) self.problem_to_node[problem_id] = idx problem_count += 1 except (IndexError, ValueError): continue print(f"[KG] 已加载 {problem_count} 个问题节点映射") # 加载边数据 edge_df = pd.read_csv(KNOWLEDGE_GRAPH_PATHS['edges']) rows, cols, data = [], [], [] grouped = edge_df.groupby('source') for src, group in tqdm(grouped, total=len(grouped), desc="处理边数据"): src_idx = self.node_id_map.get(src, -1) if src_idx == -1: continue neighbors = [] for _, row in group.iterrows(): tgt_idx = self.node_id_map.get(row['target'], -1) if tgt_idx != -1: neighbors.append((tgt_idx, row['weight'])) neighbors.sort(key=lambda x: x[1], reverse=True) top_k = min(100, len(neighbors)) for i in range(top_k): rows.append(src_idx) cols.append(neighbors[i][0]) data.append(neighbors[i][1]) # 添加自环 for i in range(self.static_node_count): rows.append(i) cols.append(i) data.append(1.0) # 创建稀疏矩阵 adj_coo = coo_matrix( (data, (rows, cols)), shape=(self.static_node_count, self.static_node_count), dtype=np.float32 ) self.adj_matrix = adj_coo.tocsc() self._rows = np.array(rows) self._cols = np.array(cols) except Exception as e: print(f"知识图谱加载失败: {str(e)}") raise # ==================== 图注意力层 ==================== class GraphAttentionLayer: def __init__(self, input_dim, output_dim, kg_loader, scope=None): self.kg_loader = kg_loader self.node_count = kg_loader.static_node_count self._rows = kg_loader._rows self._cols = kg_loader._cols with tf.variable_scope(scope or "GAT"): self.W = tf.get_variable( "W", [input_dim, output_dim], initializer=tf.initializers.variance_scaling( scale=0.1, mode='fan_avg', distribution='uniform') ) self.attn_kernel = tf.get_variable( "attn_kernel", [output_dim * 2, 1], initializer=tf.initializers.variance_scaling( scale=0.1, mode='fan_avg', distribution='uniform') ) self.bias = tf.get_variable( "bias", [output_dim], initializer=tf.zeros_initializer() ) def __call__(self, inputs): inputs = tf.clip_by_value(inputs, -5, 5) h = tf.matmul(inputs, self.W) h = tf.clip_by_value(h, -5, 5) h_src = tf.gather(h, self._rows) h_dst = tf.gather(h, self._cols) h_concat = tf.concat([h_src, h_dst], axis=1) edge_logits = tf.squeeze(tf.matmul(h_concat, self.attn_kernel), axis=1) edge_logits = tf.clip_by_value(edge_logits, -10, 10) edge_attn = tf.nn.leaky_relu(edge_logits, alpha=0.2) edge_indices = tf.constant(np.column_stack((self._rows, self._cols)), dtype=tf.int64) sparse_attn = tf.SparseTensor( indices=edge_indices, values=edge_attn, dense_shape=[self.node_count, self.node_count] ) sparse_attn_weights = tf.sparse_softmax(sparse_attn) output = tf.sparse_tensor_dense_matmul(sparse_attn_weights, h) output = tf.clip_by_value(output, -5, 5) output += self.bias output = tf.nn.elu(output) return output # ==================== 学生知识追踪模型 ==================== class StudentModel: def __init__(self, is_training, config): self.batch_size = config.batch_size # 添加这行 self.batch_size_tensor = tf.placeholder(tf.int32, [], name='batch_size_placeholder') self.num_skills = config.num_skills self.num_steps = config.num_steps self.current = tf.placeholder(tf.int32, [None, self.num_steps], name='current') self.next = tf.placeholder(tf.int32, [None, self.num_steps], name='next') self.target_id = tf.placeholder(tf.int32, [None], name='target_ids') self.target_correctness = tf.placeholder(tf.float32, [None], name='target_correctness') with tf.device('/gpu:0'), tf.variable_scope("KnowledgeGraph", reuse=tf.AUTO_REUSE): kg_loader = KnowledgeGraphLoader() kg_loader.load() kg_node_features = tf.constant(kg_loader.node_features, dtype=tf.float32) # 增强GAT结构 gat_output = kg_node_features for i in range(2): with tf.variable_scope(f"GAT_Layer_{i + 1}"): dim = 64 if i == 0 else 32 gat_layer = GraphAttentionLayer( input_dim=gat_output.shape[1] if i > 0 else kg_node_features.shape[1], output_dim=dim, kg_loader=kg_loader ) gat_output = gat_layer(gat_output) gat_output = tf.nn.leaky_relu(gat_output, alpha=0.1) self.skill_embeddings = gat_output with tf.variable_scope("FeatureProcessing"): # 使用实际batch_size的placeholder batch_size = tf.shape(self.next)[0] # 初始化方法1:使用tf.zeros_like和tile dummy_vector = tf.zeros([1, 1], dtype=tf.float32) history_init = tf.tile(dummy_vector, [batch_size, 1]) elapsed_init = tf.tile(dummy_vector, [batch_size, 1]) # 或者初始化方法2:直接使用tf.fill # history_init = tf.fill([batch_size, 1], 0.0) # elapsed_init = tf.fill([batch_size, 1], 0.0) current_indices = tf.minimum(self.current, kg_loader.static_node_count - 1) current_embed = tf.nn.embedding_lookup(self.skill_embeddings, current_indices) inputs = [] valid_mask = tf.cast(tf.not_equal(self.current, 0), tf.float32) answers_float = tf.cast(self.next, tf.float32) # 初始化历史和耗时特征 history = history_init elapsed_time = elapsed_init for t in range(self.num_steps): if t > 0: past_answers = answers_float[:, :t] past_valid_mask = valid_mask[:, :t] correct_count = tf.reduce_sum(past_answers * past_valid_mask, axis=1, keepdims=True) total_valid = tf.reduce_sum(past_valid_mask, axis=1, keepdims=True) history = correct_count / (total_valid + 1e-8) elapsed_time = tf.fill([batch_size, 1], tf.cast(t, tf.float32)) with tf.variable_scope(f"feature_extraction_t{t}"): # 基础特征 current_feat = current_embed[:, t, :] # 知识图谱特征 difficulty_feature = tf.gather( kg_loader.node_features[:, 0], tf.minimum(self.current[:, t], kg_loader.static_node_count - 1) ) difficulty_feature = tf.reshape(difficulty_feature, [-1, 1]) # 情感特征 affect_features = [] for i in range(1, 3): try: affect_feature = tf.gather( kg_loader.node_features[:, i], tf.minimum(self.current[:, t], kg_loader.static_node_count - 1) ) affect_feature = tf.reshape(affect_feature, [-1, 1]) affect_features.append(affect_feature) except Exception as e: tf.logging.warning(f"情感特征{i}提取失败: {str(e)}") affect_features.append(tf.zeros_like(difficulty_feature)) # 确保所有特征都是2维的 features_to_concat = [current_feat, history, elapsed_time, difficulty_feature] + affect_features features_to_concat = [ f if len(f.shape) == 2 else tf.reshape(f, [-1, 1]) for f in features_to_concat ] # 调试信息(可选) if is_training: features_to_concat = [ tf.Print(f, [tf.shape(f)], message=f"Feature {i} shape at step {t}: ") for i, f in enumerate(features_to_concat) ] combined = tf.concat(features_to_concat, axis=1) inputs.append(combined) # 增强RNN结构 with tf.variable_scope("RNN"): cells = [] for i in range(2): cell = rnn.LSTMCell( FLAGS.hidden_size, initializer=tf.orthogonal_initializer(), forget_bias=1.0 ) if is_training and FLAGS.keep_prob < 1.0: cell = rnn.DropoutWrapper(cell, output_keep_prob=FLAGS.keep_prob) cells.append(cell) stacked_cell = rnn.MultiRNNCell(cells) outputs, _ = tf.nn.dynamic_rnn( stacked_cell, tf.stack(inputs, axis=1), dtype=tf.float32 ) output = tf.reshape(outputs, [-1, FLAGS.hidden_size]) with tf.variable_scope("Output"): hidden = tf.layers.dense( output, units=32, activation=tf.nn.relu, kernel_initializer=tf.initializers.glorot_uniform() ) logits = tf.layers.dense( hidden, units=1, kernel_initializer=tf.initializers.glorot_uniform() ) self._all_logits = tf.clip_by_value(logits, -20, 20) selected_logits = tf.gather(tf.reshape(self._all_logits, [-1]), self.target_id) self.pred = tf.clip_by_value(tf.sigmoid(selected_logits), 1e-8, 1 - 1e-8) with tf.variable_scope("Loss"): labels = tf.clip_by_value(self.target_correctness, 0.05, 0.95) pos_weight = tf.reduce_sum(1.0 - labels) / (tf.reduce_sum(labels) + 1e-8) bce_loss = tf.nn.weighted_cross_entropy_with_logits( targets=labels, logits=selected_logits, pos_weight=pos_weight ) confidence_penalty = tf.reduce_mean( tf.square(tf.sigmoid(selected_logits) - 0.5) ) loss = tf.reduce_mean(bce_loss) + 0.1 * confidence_penalty l2_loss = tf.add_n([ tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'bias' not in v.name ]) * FLAGS.l2_lambda self.cost = loss + l2_loss # ==================== 数据加载 ==================== def read_data_from_csv_file(path, kg_loader, is_training=False): students = [] student_ids = [] max_skill = 0 missing_problems = set() if not os.path.exists(path): print(f"❌ 文件不存在: {path}") return [], [], [], 0, 0, 0 try: print(f"[数据] 加载数据文件: {path}") try: data_df = pd.read_csv(path) except Exception as e: print(f"CSV读取失败: {str(e)}") encodings = ['utf-8', 'latin1', 'iso-8859-1', 'cp1252'] for encoding in encodings: try: data_df = pd.read_csv(path, encoding=encoding) break except: continue if 'data_df' not in locals(): return [], [], [], 0, 0, 0 # 列名标准化 possible_columns = { 'user_id': ['user_id', 'userid', 'student_id', 'studentid'], 'problem_id': ['problem_id', 'problemid', 'skill_id', 'skillid'], 'correct': ['correct', 'correctness', 'answer', 'accuracy'], 'start_time': ['start_time', 'timestamp', 'time', 'date'] } actual_columns = {} for col_type, possible_names in possible_columns.items(): found = False for name in possible_names: if name in data_df.columns: actual_columns[col_type] = name found = True break if not found: print(f"❌ 错误: 找不到 {col_type} 列") return [], [], [], 0, 0, 0 data_df = data_df.rename(columns={ actual_columns['user_id']: 'user_id', actual_columns['problem_id']: 'problem_id', actual_columns['correct']: 'correct', actual_columns['start_time']: 'start_time' }) # 时间戳转换 print("[数据] 转换时间戳...") timestamp_col = data_df['start_time'] if isinstance(timestamp_col.iloc[0], str): try: data_df['start_time'] = timestamp_col.astype(float) except ValueError: parsed_times = timestamp_col.apply(parse_timestamp) nan_count = parsed_times.isna().sum() if nan_count > 0: print(f"⚠️ 警告: {nan_count}个时间戳无法解析,将设为0") parsed_times = parsed_times.fillna(0) data_df['start_time'] = parsed_times else: data_df['start_time'] = timestamp_col.astype(float) # 按学生分组 grouped = data_df.groupby('user_id') for user_id, group in tqdm(grouped, total=len(grouped), desc="处理学生数据"): try: group = enhanced_data_validation(group, kg_loader) if group is None: continue problems = group['problem_id'].values answers = group['correct'].values.astype(int) timestamps = group['start_time'].values.astype(float) valid_data = [] invalid_count = 0 for i, (p, a) in enumerate(zip(problems, answers)): if p in kg_loader.problem_to_node and a in (0, 1): valid_data.append((p, a)) else: invalid_count += 1 if p != 0 and p not in missing_problems: missing_problems.add(p) if len(valid_data) < 2: continue problems, answers = zip(*valid_data) n_split = (len(problems) + FLAGS.problem_len - 1) // FLAGS.problem_len for k in range(n_split): start = k * FLAGS.problem_len end = (k + 1) * FLAGS.problem_len seg_problems = list(problems[start:end]) seg_answers = list(answers[start:end]) if len(seg_problems) < FLAGS.problem_len: pad_len = FLAGS.problem_len - len(seg_problems) seg_problems += [0] * pad_len seg_answers += [0] * pad_len mapped_problems = [kg_loader.problem_to_node.get(p, 0) for p in seg_problems] students.append(([user_id, k], mapped_problems, seg_answers)) max_skill = max(max_skill, max(mapped_problems)) student_ids.append(user_id) except Exception as e: print(f"处理学生 {user_id} 时出错: {str(e)}") continue except Exception as e: print(f"数据加载失败: {str(e)}") return [], [], [], 0, 0, 0 return students, [], student_ids, max_skill, 0, 0 def enhanced_data_validation(group, kg_loader): """增强数据验证""" problems = group['problem_id'].values timestamps = group['start_time'].values.astype(float) valid_indices = np.where(~np.isnan(timestamps))[0] if len(valid_indices) > 1: time_diffs = np.diff(timestamps[valid_indices]) if np.any(time_diffs < 0): sort_idx = np.argsort(timestamps) group = group.iloc[sort_idx].reset_index(drop=True) valid_mask = [p in kg_loader.problem_to_node for p in problems] if not any(valid_mask): return None return group[valid_mask] # ==================== 训练流程 ==================== def run_epoch(session, model, data, run_type, eval_op, verbose=False): """执行一个epoch的训练或评估 Args: session: TF会话 model: 模型对象 data: 输入数据 run_type: '训练'或'测试' eval_op: 训练op或tf.no_op() verbose: 是否显示详细进度 Returns: dict: 包含loss, auc, rmse, r2的字典 """ preds = [] labels = [] total_loss = 0.0 processed_count = 0 # 禁用TF调试信息 tf.logging.set_verbosity(tf.logging.ERROR) index = 0 batch_size = model.batch_size # 可选:使用tqdm进度条(verbose模式下) iterator = tqdm(range(0, len(data), batch_size), desc=f"{run_type}处理中") if verbose else range(0, len(data), batch_size) for start in iterator: end = min(start + batch_size, len(data)) batch_data = data[start:end] # 准备批次数据 current_batch, next_batch, target_ids, target_correctness = [], [], [], [] for idx, (stu_id, problems, answers) in enumerate(batch_data): valid_length = sum(1 for p in problems if p != 0) if valid_length < 1: continue current_batch.append(problems) next_batch.append(answers) last_step = valid_length - 1 target_ids.append(idx * model.num_steps + last_step) target_correctness.append(answers[last_step]) if not current_batch: continue actual_batch_size = len(current_batch) feed_dict = { model.current: np.array(current_batch, dtype=np.int32), model.next: np.array(next_batch, dtype=np.int32), model.target_id: np.array(target_ids, dtype=np.int32), model.target_correctness: np.array(target_correctness, dtype=np.float32) } try: if eval_op != tf.no_op(): _, pred, loss = session.run( [eval_op, model.pred, model.cost], feed_dict=feed_dict ) else: pred, loss = session.run( [model.pred, model.cost], feed_dict=feed_dict ) preds.extend(pred.flatten().tolist()) labels.extend(target_correctness) total_loss += loss * actual_batch_size processed_count += actual_batch_size except Exception as e: print(f"\n{run_type}错误 (批次 {start}-{end}): {str(e)}", file=sys.stderr) continue # 计算指标 if processed_count == 0: return None avg_loss = total_loss / processed_count # 确保标签和预测值在有效范围内 labels = np.clip(np.array(labels), 1e-7, 1 - 1e-7) preds = np.clip(np.array(preds), 1e-7, 1 - 1e-7) metrics = { 'loss': avg_loss, 'auc': roc_auc_score(labels, preds) if len(set(labels)) > 1 else 0.5, 'rmse': np.sqrt(mean_squared_error(labels, preds)), 'r2': r2_score(labels, preds) } return metrics def main(_): """主训练流程""" # 1. 加载配置和数据 config = ModelConfig() # 假设已定义 train_data, test_data = load_data() # 假设已定义 # 2. 构建模型 with tf.variable_scope("Model", reuse=False): train_model = StudentModel(is_training=True, config=config) with tf.variable_scope("Model", reuse=True): test_model = StudentModel(is_training=False, config=config) # 3. 创建会话 sess_config = tf.ConfigProto() sess_config.gpu_options.allow_growth = True with tf.Session(config=sess_config) as sess: # 4. 初始化变量 sess.run(tf.global_variables_initializer()) # 5. 训练循环 best_auc = 0.0 for epoch in range(1, FLAGS.max_epochs + 1): # 训练阶段 train_metrics = run_epoch( sess, train_model, train_data, '训练', train_op, # train_op应已定义 verbose=(epoch % FLAGS.display_freq == 0) ) # 测试阶段 test_metrics = run_epoch( sess, test_model, test_data, '测试', tf.no_op(), verbose=False ) # 6. 输出关键指标 print(f"Epoch {epoch}") print( f"训练集 - 损失: {train_metrics['loss']:.4f}, RMSE: {train_metrics['rmse']:.4f}, AUC: {train_metrics['auc']:.4f}, R²: {train_metrics['r2']:.4f}") print( f"测试集 - 损失: {test_metrics['loss']:.4f}, RMSE: {test_metrics['rmse']:.4f}, AUC: {test_metrics['auc']:.4f}, R²: {test_metrics['r2']:.4f}") sys.stdout.flush() # 7. 保存最佳模型 if test_metrics['auc'] > best_auc: best_auc = test_metrics['auc'] saver.save(sess, FLAGS.model_path) # saver应已定义 print("训练完成!") print(f"最佳测试AUC: {best_auc:.4f}") if __name__ == "__main__": # 生成模拟数据(仅当真实数据不存在时) if not os.path.exists(FLAGS.train_data_path) or not os.path.exists(FLAGS.test_data_path): generate_mock_data() tf.app.run() 在这个基础上修改得到的完整代码,你给的不完整,不要省略!!!!
最新发布
07-02
# -*- coding: utf-8 -*- """ DKT-DSC for Assistment2012 (优化版) - 修复数据泄露问题 最后更新: 2024-07-01 """ import os import sys import numpy as np import tensorflow.compat.v1 as tf os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" os.environ["CUDA_VISIBLE_DEVICES"] = "0" config = tf.ConfigProto() config.gpu_options.allow_growth = True tf.disable_v2_behavior() # 安全导入psutil模块 try: import psutil HAS_PSUTIL = True except ImportError: HAS_PSUTIL = False print("警告: psutil模块未安装,内存监控功能受限") from scipy.sparse import coo_matrix from tensorflow.contrib import rnn import pandas as pd from tqdm import tqdm from sklearn.metrics import mean_squared_error, r2_score, roc_curve, auc import math import random # ==================== 配置部分 ==================== # 使用实际数据路径 DATA_BASE_PATH = '/home/yhh/students/jianglu/DKT2/DKT/data/' data_name = 'Assist_2012' # 修正数据集名称 KNOWLEDGE_GRAPH_PATHS = { 'graphml': './output_assist2012_gat_improved/knowledge_graph.graphml', 'nodes': './output_assist2012_gat_improved/graph_nodes.csv', 'edges': './output_assist2012_gat_improved/graph_edges.csv' } # ==================== Flags配置 ==================== tf.flags.DEFINE_float("epsilon", 1e-8, "Adam优化器的epsilon值") tf.flags.DEFINE_float("l2_lambda", 0.005, "L2正则化系数") # 减小正则化强度 tf.flags.DEFINE_float("learning_rate", 1e-4, "学习率") tf.flags.DEFINE_float("max_grad_norm", 3.0, "梯度裁剪阈值") # 更严格的梯度裁剪 tf.flags.DEFINE_float("keep_prob", 0.8, "Dropout保留概率") # 减小dropout tf.flags.DEFINE_integer("hidden_layer_num", 1, "隐藏层数量") tf.flags.DEFINE_integer("hidden_size", 48, "隐藏层大小") # 增加隐藏层大小 tf.flags.DEFINE_integer("evaluation_interval", 2, "评估间隔周期数") tf.flags.DEFINE_integer("batch_size", 128, "批次大小") tf.flags.DEFINE_integer("problem_len", 15, "问题序列长度") # 增加序列长度 tf.flags.DEFINE_integer("epochs", 100, "训练周期数") tf.flags.DEFINE_boolean("allow_soft_placement", True, "允许软设备放置") tf.flags.DEFINE_boolean("log_device_placement", False, "记录设备放置信息") tf.flags.DEFINE_string("train_data_path", f'{DATA_BASE_PATH}{data_name}_train.csv', "训练数据路径") tf.flags.DEFINE_string("test_data_path", f'{DATA_BASE_PATH}{data_name}_test.csv', "测试数据路径") FLAGS = tf.flags.FLAGS # 焦点损失参数 FOCAL_LOSS_GAMMA = 1.5 # 调整焦点损失参数 FOCAL_LOSS_ALPHA = 0.3 # 学习率衰减参数 DECAY_STEPS = 2000 DECAY_RATE = 0.95 # 学习率预热步数 WARMUP_STEPS = 2000 # 内存监控函数 def memory_usage(): """增强的内存监控函数,处理psutil缺失情况""" if HAS_PSUTIL: try: process = psutil.Process(os.getpid()) return process.memory_info().rss / (1024 ** 2) except: return 0.0 return 0.0 # ==================== 知识图谱加载器 ==================== class KnowledgeGraphLoader: def __init__(self): self.node_features = None self.adj_matrix = None self.problem_to_node = {} self.node_id_map = {} self.static_node_count = 0 self._rows = None self._cols = None def load(self): """加载知识图谱数据并进行严格的数据验证""" print("\n[KG] 加载知识图谱...") try: if not os.path.exists(KNOWLEDGE_GRAPH_PATHS['nodes']): raise FileNotFoundError(f"节点文件未找到: {KNOWLEDGE_GRAPH_PATHS['nodes']}") if not os.path.exists(KNOWLEDGE_GRAPH_PATHS['edges']): raise FileNotFoundError(f"边文件未找到: {KNOWLEDGE_GRAPH_PATHS['edges']}") node_df = pd.read_csv(KNOWLEDGE_GRAPH_PATHS['nodes']) self.static_node_count = len(node_df) print(f"[KG] 总节点数: {self.static_node_count}") # 处理空值 - 根据验证报告中的发现 print("[KG] 处理特征空值...") feature_cols = [col for col in node_df.columns if col not in ['node_id', 'type']] # 特别处理total_attempts特征 if 'total_attempts' in feature_cols: # 概念节点使用概念节点中位数填充 concept_mask = node_df['type'] == 'concept' concept_median = node_df.loc[concept_mask, 'total_attempts'].median() # 处理NaN值 if pd.isna(concept_median): concept_median = 0.0 node_df.loc[concept_mask, 'total_attempts'] = node_df.loc[concept_mask, 'total_attempts'].fillna(concept_median) # 问题节点使用问题节点中位数填充 problem_mask = node_df['type'] == 'problem' problem_median = node_df.loc[problem_mask, 'total_attempts'].median() # 处理NaN值 if pd.isna(problem_median): problem_median = 0.0 node_df.loc[problem_mask, 'total_attempts'] = node_df.loc[problem_mask, 'total_attempts'].fillna(problem_median) print(f" 填充 total_attempts 缺失值: 概念节点={concept_median}, 问题节点={problem_median}") # 处理其他数值特征 other_cols = [col for col in feature_cols if col != 'total_attempts'] for col in other_cols: # 分类型填充 if 'confidence' in col or 'affect' in col: # 情感特征使用全局平均值填充 global_mean = node_df[col].mean() # 处理NaN值 if pd.isna(global_mean): global_mean = 0.0 node_df[col] = node_df[col].fillna(global_mean) print(f" 填充 {col} 缺失值: 全局均值={global_mean:.4f}") else: # 其他特征按问题类型分组填充 problem_mask = node_df['type'] == 'problem' problem_mean = node_df.loc[problem_mask, col].mean() # 处理NaN值 if pd.isna(problem_mean): problem_mean = 0.0 node_df.loc[problem_mask, col] = node_df.loc[problem_mask, col].fillna(problem_mean) concept_mask = node_df['type'] == 'concept' concept_mean = node_df.loc[concept_mask, col].mean() # 处理NaN值 if pd.isna(concept_mean): concept_mean = 0.0 node_df.loc[concept_mask, col] = node_df.loc[concept_mask, col].fillna(concept_mean) print(f" 填充 {col} 缺失值: 问题节点={problem_mean:.4f}, 概念节点={concept_mean:.4f}") print("\n[KG诊断] 特征分析...") if feature_cols: raw_features = node_df[feature_cols].values nan_count = np.isnan(raw_features).sum() inf_count = np.isinf(raw_features).sum() print(f" 总特征值数: {raw_features.size}") print(f" NaN特征数: {nan_count}") print(f" Inf特征数: {inf_count}") if nan_count > 0 or inf_count > 0: print(f"⚠️ 警告: 节点特征包含 {nan_count} 个NaN和 {inf_count} 个Inf值,将被替换为0") raw_features = np.nan_to_num(raw_features) # 标准化特征并确保为float32类型 feature_mean = np.mean(raw_features, axis=0) feature_std = np.std(raw_features, axis=0) + 1e-8 self.node_features = np.array( (raw_features - feature_mean) / feature_std, dtype=np.float32 # 显式指定为float32 ) self.node_features = np.nan_to_num(self.node_features) # 再次确保无NaN else: print("警告: 节点文件中没有特征列") self.node_features = np.zeros((self.static_node_count, 1), dtype=np.float32) # 创建节点ID映射 self.node_id_map = {} for idx, row in node_df.iterrows(): self.node_id_map[row['node_id']] = idx # 创建问题ID到节点索引的映射 self.problem_to_node = {} problem_count = 0 for idx, row in node_df.iterrows(): if row['type'] == 'problem': try: parts = row['node_id'].split('_') if len(parts) < 2: continue problem_id = int(parts[1]) self.problem_to_node[problem_id] = idx problem_count += 1 except: continue print(f"[KG] 已加载 {problem_count} 个问题节点映射") # 加载边数据并进行优化 edge_df = pd.read_csv(KNOWLEDGE_GRAPH_PATHS['edges']) print("[KG] 优化邻接矩阵(保留每个节点的前100个邻居)...") rows, cols, data = [], [], [] valid_edge_count = 0 invalid_edge_count = 0 # 限制每个节点的邻居数量以提高效率 grouped = edge_df.groupby('source') for src, group in tqdm(grouped, total=len(grouped), desc="处理边数据"): src_idx = self.node_id_map.get(src, -1) if src_idx == -1: invalid_edge_count += len(group) continue neighbors = [] for _, row in group.iterrows(): tgt_idx = self.node_id_map.get(row['target'], -1) if tgt_idx != -1: neighbors.append((tgt_idx, row['weight'])) # 根据权重排序并取Top 100 neighbors.sort(key=lambda x: x[1], reverse=True) top_k = min(100, len(neighbors)) # 限制邻居数量 for i in range(top_k): tgt_idx, weight = neighbors[i] rows.append(src_idx) cols.append(tgt_idx) data.append(weight) valid_edge_count += 1 # 添加自环 for i in range(self.static_node_count): rows.append(i) cols.append(i) data.append(1.0) valid_edge_count += 1 # 创建稀疏邻接矩阵 adj_coo = coo_matrix( (data, (rows, cols)), shape=(self.static_node_count, self.static_node_count), dtype=np.float32 ) self.adj_matrix = adj_coo.tocsc() self._rows = np.array(rows) self._cols = np.array(cols) print(f"[KG] 邻接矩阵构建完成 | 节点: {self.static_node_count} | 边: {len(data)}") print(f"[KG优化] 最大行索引: {np.max(self._rows)} | 最大列索引: {np.max(self._cols)}") except Exception as e: import traceback print(f"知识图谱加载失败: {str(e)}") traceback.print_exc() raise RuntimeError(f"知识图谱加载失败: {str(e)}") from e # ==================== 图注意力层 ==================== class GraphAttentionLayer: def __init__(self, input_dim, output_dim, kg_loader, scope=None): self.kg_loader = kg_loader self.node_count = kg_loader.static_node_count self._rows = kg_loader._rows self._cols = kg_loader._cols with tf.variable_scope(scope or "GAT"): self.W = tf.get_variable( "W", [input_dim, output_dim], initializer=tf.initializers.variance_scaling( scale=0.1, mode='fan_avg', distribution='uniform') ) self.attn_kernel = tf.get_variable( "attn_kernel", [output_dim * 2, 1], initializer=tf.initializers.variance_scaling( scale=0.1, mode='fan_avg', distribution='uniform') ) self.bias = tf.get_variable( "bias", [output_dim], initializer=tf.zeros_initializer() ) def __call__(self, inputs): inputs = tf.clip_by_value(inputs, -5, 5) inputs = tf.check_numerics(inputs, "GAT输入包含NaN或Inf") # 特征变换 h = tf.matmul(inputs, self.W) h = tf.clip_by_value(h, -5, 5) h = tf.check_numerics(h, "特征变换后包含NaN或Inf") # 注意力机制 h_src = tf.gather(h, self._rows) h_dst = tf.gather(h, self._cols) h_concat = tf.concat([h_src, h_dst], axis=1) edge_logits = tf.squeeze(tf.matmul(h_concat, self.attn_kernel), axis=1) edge_logits = tf.clip_by_value(edge_logits, -10, 10) edge_attn = tf.nn.leaky_relu(edge_logits, alpha=0.2) # 创建稀疏注意力矩阵 edge_indices = tf.constant(np.column_stack((self._rows, self._cols)), dtype=tf.int64) sparse_attn = tf.SparseTensor( indices=edge_indices, values=edge_attn, dense_shape=[self.node_count, self.node_count] ) # 稀疏softmax和矩阵乘法 sparse_attn_weights = tf.sparse_softmax(sparse_attn) output = tf.sparse_tensor_dense_matmul(sparse_attn_weights, h) output = tf.clip_by_value(output, -5, 5) output += self.bias output = tf.nn.elu(output) output = tf.check_numerics(output, "最终GAT输出包含NaN或Inf") return output # ==================== 学生知识追踪模型 ==================== class StudentModel: def __init__(self, is_training, config): self.batch_size = config.batch_size self.num_skills = config.num_skills self.num_steps = config.num_steps self.current = tf.placeholder(tf.int32, [None, self.num_steps], name='current') self.next = tf.placeholder(tf.int32, [None, self.num_steps], name='next') self.target_id = tf.placeholder(tf.int32, [None], name='target_ids') self.target_correctness = tf.placeholder(tf.float32, [None], name='target_correctness') with tf.device('/gpu:0'), tf.variable_scope("KnowledgeGraph", reuse=tf.AUTO_REUSE): # 加载知识图谱 kg_loader = KnowledgeGraphLoader() kg_loader.load() kg_node_features = tf.constant(kg_loader.node_features, dtype=tf.float32) kg_node_features = tf.check_numerics(kg_node_features, "知识图谱节点特征包含NaN或Inf") # 精简GAT层 - 减少层数和维度 gat_output = kg_node_features for i in range(2): # 减少GAT层数为2 with tf.variable_scope(f"GAT_Layer_{i + 1}"): gat_layer = GraphAttentionLayer( input_dim=gat_output.shape[1] if i > 0 else kg_node_features.shape[1], output_dim=24 if i == 0 else 16, # 减少输出维度 kg_loader=kg_loader ) gat_output = gat_layer(gat_output) gat_output = tf.nn.elu(gat_output) self.skill_embeddings = gat_output with tf.variable_scope("FeatureProcessing"): batch_size = tf.shape(self.next)[0] # 动态获取批次大小 # 当前问题嵌入 current_indices = tf.minimum(self.current, kg_loader.static_node_count - 1) current_embed = tf.nn.embedding_lookup(self.skill_embeddings, current_indices) # 构建输入序列 - 移除下一问题嵌入(修复数据泄露) inputs = [] # 使用当前问题作为有效掩码(而不是下一个问题) valid_mask = tf.cast(tf.not_equal(self.current, 0), tf.float32) answers_float = tf.cast(self.next, tf.float32) # 历史表现特征 - 修复符号张量问题 zero_vector = tf.zeros([1, 1], dtype=tf.float32) history = tf.tile(zero_vector, [batch_size, 1]) elapsed_time = tf.tile(zero_vector, [batch_size, 1]) # 循环处理每个时间步 for t in range(self.num_steps): # 创建时间相关的特征 if t > 0: # 计算历史表现(只使用t-1及之前的信息) past_answers = answers_float[:, :t] # 只使用当前时间步之前的信息 past_valid_mask = valid_mask[:, :t] correct_count = tf.reduce_sum(past_answers * past_valid_mask, axis=1, keepdims=True) total_valid = tf.reduce_sum(past_valid_mask, axis=1, keepdims=True) history = correct_count / (total_valid + 1e-8) # 计算经过的时间 elapsed_time = tf.fill([batch_size, 1], tf.cast(t, tf.float32)) # 难度特征 - 使用知识图谱中的准确率特征 # 确保只使用当前问题的特征 difficulty_feature = tf.gather( kg_loader.node_features[:, 0], # 假设第一个特征是准确率 tf.minimum(self.current[:, t], kg_loader.static_node_count - 1) ) difficulty_feature = tf.cast(difficulty_feature, tf.float32) # 情感特征 - 使用知识图谱中的情感特征 affect_features = [] for i in range(1, 5): # 使用前4个情感特征 affect_feature = tf.gather( kg_loader.node_features[:, i], tf.minimum(self.current[:, t], kg_loader.static_node_count - 1) ) affect_feature = tf.cast(affect_feature, tf.float32) affect_features.append(tf.reshape(affect_feature, [-1, 1])) # 组合所有特征 - 移除了下一问题嵌入(修复数据泄露) combined = tf.concat([ current_embed[:, t, :], history, elapsed_time, tf.reshape(difficulty_feature, [-1, 1]), *affect_features ], axis=1) inputs.append(combined) # RNN模型 with tf.variable_scope("RNN"): cell = rnn.LSTMCell( FLAGS.hidden_size, initializer=tf.initializers.glorot_uniform(), forget_bias=1.0 ) if is_training and FLAGS.keep_prob < 1.0: cell = rnn.DropoutWrapper(cell, output_keep_prob=FLAGS.keep_prob) outputs, _ = tf.nn.dynamic_rnn( cell, tf.stack(inputs, axis=1), dtype=tf.float32 ) output = tf.reshape(outputs, [-1, FLAGS.hidden_size]) # 输出层 with tf.variable_scope("Output"): hidden = tf.layers.dense( output, units=32, activation=tf.nn.relu, kernel_initializer=tf.initializers.glorot_uniform(), name="hidden_layer" ) logits = tf.layers.dense( hidden, units=1, kernel_initializer=tf.initializers.glorot_uniform(), name="output_layer" ) # 损失计算 self._all_logits = tf.clip_by_value(logits, -20, 20) selected_logits = tf.gather(tf.reshape(self._all_logits, [-1]), self.target_id) self.pred = tf.clip_by_value(tf.sigmoid(selected_logits), 1e-8, 1 - 1e-8) # 焦点损失 labels = tf.clip_by_value(self.target_correctness, 0.05, 0.95) pos_weight = tf.reduce_sum(1.0 - labels) / (tf.reduce_sum(labels) + 1e-8) bce_loss = tf.nn.weighted_cross_entropy_with_logits( targets=labels, logits=selected_logits, pos_weight=pos_weight ) loss = tf.reduce_mean(bce_loss) # L2正则化 l2_loss = tf.add_n([ tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'bias' not in v.name ]) * FLAGS.l2_lambda self.cost = loss + l2_loss # ==================== 数据加载 ==================== def read_data_from_csv_file(path, kg_loader, is_training=False): """更鲁棒的数据加载函数""" students = [] student_ids = [] max_skill = 0 missing_problems = set() # 增强文件存在性检查 if not os.path.exists(path): print(f"❌ 严重错误: 数据文件不存在: {path}") print("请检查以下可能原因:") print("1. 文件路径是否正确") print("2. 文件名是否匹配") print("3. 文件权限是否足够") # 尝试列出目录内容以便调试 dir_path = os.path.dirname(path) print(f"目录内容: {os.listdir(dir_path) if os.path.exists(dir_path) else '目录不存在'}") return [], [], [], 0, 0, 0 try: # 打印正在加载的文件路径 print(f"[数据] 加载数据文件: {path}") # 读取数据集 - 增强CSV读取兼容性 try: data_df = pd.read_csv(path) except Exception as e: print(f"CSV读取失败: {str(e)}") print("尝试使用备用方法读取...") # 尝试不同编码 encodings = ['utf-8', 'latin1', 'iso-8859-1', 'cp1252'] for encoding in encodings: try: data_df = pd.read_csv(path, encoding=encoding) print(f"成功使用 {encoding} 编码读取文件") break except Exception as e: print(f"编码 {encoding} 尝试失败: {str(e)}") continue if 'data_df' not in locals(): print("所有编码尝试失败,无法读取文件") return [], [], [], 0, 0, 0 print(f"[数据] 加载完成 | 记录数: {len(data_df)}") # 检查必要的列是否存在 - 支持多种列名变体 # 可能的列名变体 possible_columns = { 'user_id': ['user_id', 'userid', 'student_id', 'studentid'], 'problem_id': ['problem_id', 'problemid', 'skill_id', 'skillid'], 'correct': ['correct', 'correctness', 'answer', 'accuracy'], 'start_time': ['start_time', 'timestamp', 'time', 'date'] } # 查找实际列名 actual_columns = {} for col_type, possible_names in possible_columns.items(): found = False for name in possible_names: if name in data_df.columns: actual_columns[col_type] = name found = True break if not found: print(f"❌ 错误: 找不到 {col_type} 列") print(f"数据列: {list(data_df.columns)}") return [], [], [], 0, 0, 0 # 重命名列为标准名称以便后续处理 data_df = data_df.rename(columns={ actual_columns['user_id']: 'user_id', actual_columns['problem_id']: 'problem_id', actual_columns['correct']: 'correct', actual_columns['start_time']: 'start_time' }) print(f"[数据] 使用列: user_id, problem_id, correct, start_time") # 按学生分组 grouped = data_df.groupby('user_id') print(f"[数据] 分组完成 | 学生数: {len(grouped)}") for user_id, group in tqdm(grouped, total=len(grouped), desc="处理学生数据"): # 按时间排序 group = group.sort_values('start_time') problems = group['problem_id'].values answers = group['correct'].values.astype(int) # 筛选有效数据 - 添加详细日志 valid_data = [] invalid_count = 0 for i, (p, a) in enumerate(zip(problems, answers)): # 检查问题是否在知识图谱中 if p in kg_loader.problem_to_node and a in (0, 1): # 额外检查:确保问题特征不包含学生作答信息 node_idx = kg_loader.problem_to_node[p] if 'accuracy' in kg_loader.node_features[node_idx]: # 如果特征中包含准确率,警告可能的数据泄露 print(f"警告: 问题 {p} 的特征包含准确率信息,可能导致数据泄露") valid_data.append((p, a)) else: invalid_count += 1 if p != 0 and p not in missing_problems: print(f"警告: 问题ID {p} 不在知识图谱中 (学生: {user_id}, 位置: {i})") missing_problems.add(p) if len(valid_data) < 2: print(f"跳过数据不足的学生 {user_id} (有效交互: {len(valid_data)}, 无效: {invalid_count})") continue # 分割序列 problems, answers = zip(*valid_data) n_split = (len(problems) + FLAGS.problem_len - 1) // FLAGS.problem_len for k in range(n_split): start = k * FLAGS.problem_len end = (k + 1) * FLAGS.problem_len seg_problems = list(problems[start:end]) seg_answers = list(answers[start:end]) # 填充短序列 if len(seg_problems) < FLAGS.problem_len: pad_len = FLAGS.problem_len - len(seg_problems) seg_problems += [0] * pad_len seg_answers += [0] * pad_len # 训练数据增强 if is_training: valid_indices = [i for i, p in enumerate(seg_problems) if p != 0] if len(valid_indices) > 1 and random.random() > 0.5: random.shuffle(valid_indices) seg_problems = [seg_problems[i] for i in valid_indices] + seg_problems[len(valid_indices):] seg_answers = [seg_answers[i] for i in valid_indices] + seg_answers[len(valid_indices):] # 映射问题ID到知识图谱节点 mapped_problems = [] for p in seg_problems: if p == 0: mapped_problems.append(0) elif p in kg_loader.problem_to_node: mapped_problems.append(kg_loader.problem_to_node[p]) else: mapped_problems.append(0) students.append(([user_id, k], mapped_problems, seg_answers)) max_skill = max(max_skill, max(mapped_problems)) student_ids.append(user_id) except Exception as e: print(f"数据加载失败: {str(e)}") import traceback traceback.print_exc() return [], [], [], 0, 0, 0 avg_length = sum(len(s[1]) for s in students) / len(students) if students else 0 print(f"[数据统计] 学生数: {len(student_ids)} | 序列数: {len(students)}") print(f" 最大技能ID: {max_skill} | 平均序列长度: {avg_length:.1f}") print(f" 缺失问题数: {len(missing_problems)}") return students, [], student_ids, max_skill, 0, 0 # ==================== 训练流程 ==================== def run_epoch(session, model, data, run_type, eval_op, global_step=None): preds = [] labels = [] total_loss = 0.0 step = 0 processed_count = 0 total_batches = max(1, len(data) // model.batch_size) with tqdm(total=total_batches, desc=f"{run_type} Epoch") as pbar: index = 0 while index < len(data): # 准备批次数据 current_batch = [] next_batch = [] target_ids = [] target_correctness = [] for i in range(model.batch_size): if index >= len(data): break stu_id, problems, answers = data[index] valid_length = sum(1 for p in problems if p != 0) if valid_length < 1: index += 1 continue current_batch.append(problems) next_batch.append(answers) last_step = valid_length - 1 target_ids.append(i * model.num_steps + last_step) target_correctness.append(answers[last_step]) index += 1 if len(current_batch) == 0: pbar.update(1) step += 1 continue # 创建feed_dict feed = { model.current: np.array(current_batch, dtype=np.int32), model.next: np.array(next_batch, dtype=np.int32), model.target_id: np.array(target_ids, dtype=np.int32), model.target_correctness: np.array(target_correctness, dtype=np.float32) } # 运行计算 try: results = session.run( [model.pred, model.cost, eval_op], feed_dict=feed ) pred, loss = results[:2] preds.extend(pred.tolist()) labels.extend(target_correctness) total_loss += loss * len(current_batch) processed_count += len(current_batch) pbar.set_postfix( loss=f"{loss:.4f}", mem=f"{memory_usage():.1f}MB" ) pbar.update(1) step += 1 except Exception as e: print(f"\n训练错误: {str(e)}") import traceback traceback.print_exc() break # 计算指标 if not labels or not preds: print(f"{run_type}周期: 无有效样本!") return float('nan'), 0.5, 0.0, 0.0 labels = np.array(labels, dtype=np.float32) preds = np.array(preds, dtype=np.float32) mask = np.isfinite(labels) & np.isfinite(preds) if not mask.any(): print(f"{run_type}周期: 所有样本包含无效值!") return float('nan'), 0.5, 0.0, 0.0 labels = labels[mask] preds = preds[mask] try: rmse = np.sqrt(mean_squared_error(labels, preds)) fpr, tpr, _ = roc_curve(labels, preds) auc_score = auc(fpr, tpr) r2 = r2_score(labels, preds) avg_loss = total_loss / processed_count if processed_count > 0 else 0.0 print(f"\n{run_type}周期总结:") print(f" 样本数: {len(labels)} | 正样本比例: {np.mean(labels > 0.5):.3f}") print(f" Loss: {avg_loss:.4f} | RMSE: {rmse:.4f} | AUC: {auc_score:.4f} | R²: {r2:.4f}") # 添加预测值分布分析 print("\n预测值分布分析:") print(f" 最小值: {np.min(preds):.4f} | 最大值: {np.max(preds):.4f}") print(f" 均值: {np.mean(preds):.4f} | 中位数: {np.median(preds):.4f}") print(f" 标准差: {np.std(preds):.4f}") # 检查完美预测的情况 perfect_preds = np.sum((preds < 1e-5) | (preds > 1 - 1e-5)) if perfect_preds > 0: perfect_ratio = perfect_preds / len(preds) print(f" 警告: {perfect_preds}个样本({perfect_ratio*100:.2f}%)预测值为0或1") # 检查预测值是否全部相同 if np.all(preds == preds[0]): print(f" 严重警告: 所有预测值相同 ({preds[0]:.4f})") return rmse, auc_score, r2, avg_loss except Exception as e: print(f"指标计算错误: {str(e)}") return float('nan'), 0.5, 0.0, 0.0 # ==================== 主函数 ==================== def main(_): print(f"[系统] 训练数据路径: {FLAGS.train_data_path}") print(f"[系统] 测试数据路径: {FLAGS.test_data_path}") # 检查文件是否存在 if not os.path.exists(FLAGS.train_data_path): print(f"❌ 训练文件不存在: {FLAGS.train_data_path}") if not os.path.exists(FLAGS.test_data_path): print(f"❌ 测试文件不存在: {FLAGS.test_data_path}") print(f"⚠️ 优化设置: batch_size={FLAGS.batch_size}, hidden_size={FLAGS.hidden_size}, lr={FLAGS.learning_rate}") session_conf = tf.ConfigProto( allow_soft_placement=True, log_device_placement=False, operation_timeout_in_ms=60000 ) session_conf.gpu_options.allow_growth = True with tf.Session(config=session_conf) as sess: # 加载知识图谱 kg_loader = KnowledgeGraphLoader() kg_loader.load() # 加载数据 print("\n[系统] 加载训练数据...") train_data = read_data_from_csv_file(FLAGS.train_data_path, kg_loader, is_training=True) print("[系统] 加载测试数据...") test_data = read_data_from_csv_file(FLAGS.test_data_path, kg_loader) if not train_data[0] or not test_data[0]: print("❌ 错误: 训练或测试数据为空!") return # 模型配置 class ModelConfig: def __init__(self): self.batch_size = FLAGS.batch_size self.num_skills = kg_loader.static_node_count + 100 # 添加缓冲区 self.num_steps = FLAGS.problem_len self.keep_prob = FLAGS.keep_prob model_config = ModelConfig() print(f"[配置] 技能数量: {model_config.num_skills}") print(f"[配置] 序列长度: {model_config.num_steps}") # 构建模型 print("\n[系统] 构建模型...") with tf.variable_scope("Model"): train_model = StudentModel(is_training=True, config=model_config) tf.get_variable_scope().reuse_variables() test_model = StudentModel(is_training=False, config=model_config) # 优化器和训练操作 global_step = tf.Variable(0, trainable=False) learning_rate = tf.train.exponential_decay( FLAGS.learning_rate, global_step, DECAY_STEPS, DECAY_RATE, staircase=True ) optimizer = tf.train.AdamOptimizer( learning_rate=learning_rate, epsilon=FLAGS.epsilon ) grads_and_vars = optimizer.compute_gradients(train_model.cost) grads, variables = zip(*grads_and_vars) clipped_grads, _ = tf.clip_by_global_norm(grads, FLAGS.max_grad_norm) train_op = optimizer.apply_gradients(zip(clipped_grads, variables), global_step=global_step) # 初始化变量 sess.run(tf.global_variables_initializer()) print(f"[系统] 训练开始 | 批次: {FLAGS.batch_size} | 学习率: {FLAGS.learning_rate}") # 模型保存 checkpoint_dir = "checkpoints_assist2012" os.makedirs(checkpoint_dir, exist_ok=True) saver = tf.train.Saver(max_to_keep=3) best_auc = 0.0 # 训练循环 for epoch in range(FLAGS.epochs): print(f"\n==== Epoch {epoch + 1}/{FLAGS.epochs} ====") current_lr = sess.run(learning_rate) print(f"[学习率] 当前学习率: {current_lr:.7f}") # 训练 train_rmse, train_auc, train_r2, train_loss = run_epoch( sess, train_model, train_data[0], '训练', train_op ) # 评估 if (epoch + 1) % FLAGS.evaluation_interval == 0: test_rmse, test_auc, test_r2, test_loss = run_epoch( sess, test_model, test_data[0], '测试', tf.no_op() ) # 保存最佳模型 if test_auc > best_auc: best_auc = test_auc save_path = saver.save(sess, f"{checkpoint_dir}/best_model.ckpt") print(f"保存最佳模型: {save_path}, AUC={best_auc:.4f}") print("\n训练完成!") if __name__ == "__main__": tf.app.run() 训练代码的测试集的auc 20轮只达到了0.7658;哪里出了问题,如何提高auc
07-02
Traceback (most recent call last): File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/tensorflow_core/python/framework/ops.py", line 1607, in _create_c_op c_op = c_api.TF_FinishOperation(op_desc) tensorflow.python.framework.errors_impl.InvalidArgumentError: Shape must be rank 2 but is rank 1 for 'Model/FeatureProcessing/concat' (op: 'ConcatV2') with input shapes: [?,32], [?,1], [?,1], [?], [?,1], [?,1], []. During handling of the above exception, another exception occurred: Traceback (most recent call last): File "DKT-DSC.py", line 805, in <module> tf.app.run() File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/tensorflow_core/python/platform/app.py", line 40, in run _run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef) File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/absl/app.py", line 308, in run _run_main(main, args) File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/absl/app.py", line 254, in _run_main sys.exit(main(argv)) File "DKT-DSC.py", line 739, in main train_model = StudentModel(is_training=True, config=model_config) File "DKT-DSC.py", line 399, in __init__ ], axis=1) File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/tensorflow_core/python/util/dispatch.py", line 180, in wrapper return target(*args, **kwargs) File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/tensorflow_core/python/ops/array_ops.py", line 1420, in concat return gen_array_ops.concat_v2(values=values, axis=axis, name=name) File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/tensorflow_core/python/ops/gen_array_ops.py", line 1257, in concat_v2 "ConcatV2", values=values, axis=axis, name=name) File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/tensorflow_core/python/framework/op_def_library.py", line 794, in _apply_op_helper op_def=op_def) File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/tensorflow_core/python/util/deprecation.py", line 507, in new_func return func(*args, **kwargs) File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/tensorflow_core/python/framework/ops.py", line 3357, in create_op attrs, op_def, compute_device) File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/tensorflow_core/python/framework/ops.py", line 3426, in _create_op_internal op_def=op_def) File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/tensorflow_core/python/framework/ops.py", line 1770, in __init__ control_input_ops) File "/home/yhh/anaconda3/envs/DKT1/lib/python3.7/site-packages/tensorflow_core/python/framework/ops.py", line 1610, in _create_c_op raise ValueError(str(e)) ValueError: Shape must be rank 2 but is rank 1 for 'Model/FeatureProcessing/concat' (op: 'ConcatV2') with input shapes: [?,32], [?,1], [?,1], [?], [?,1], [?,1], [].
07-02
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值