这是使用pytorch实现LSTM模型的二分类实验过程，目前使用的是

import torch from torch import nn import torch.utils.data as Data from sklearn.preprocessing import StandardScaler from sklearn.model_selection import StratifiedKFold from sklearn.metrics import accuracy_score, balanced_accuracy_score, confusion_matrix, roc_auc_score, f1_score from imblearn.under_sampling import RandomUnderSampler import numpy as np import pandas as pd import os import matplotlib.pyplot as plt import json import cross seed = 42 INPUT_SIZE = 70 HIDDEN_LAYER_SIZE = 16 OUTPUT_SIZE = 1 NUM_LAYERS = 1 LR = 1e-3 EPOCHS = 50 BATCH_SIZE = 15 FOLD = 10 # 定义参数网格 param_grid = { # 'hidden_size': [8, 16, 32, 64], # 'num_layers': [1, 2], # 'learning_rate': [0.001, 0.01], # 'batch_size': [8, 16, 32, 64], # 'epoch': [i for i in range(10, 100, 10)] 'hidden_size': [128], # 6,32,64,128 'num_layers': [1], 'learning_rate': [0.001], # 0.05,0.01,1e-3,1e-4 'batch_size': [32], # 8,16,32,64 'epoch': [500], # i for i in range(50, 300, 50) 'dropout_prob': [0.2] } file = f'result/LSTM-test/4-20-fold{FOLD}.csv' # file = 'result/LSTM-test/4-7-fold10.csv' best_val_accuracies = [] best_epochs = [] key = eval(input('0. 训练模型\t1. 保存的模型\n')) class LSTM(nn.Module): def __init__(self, input_size=70, hidden_layer_size=100, output_size=1, num_layers=1, dropout_prob=0.2): ''' LSTM二分类任务 :param input_size: 输入数据的维度 :param hidden_layer_size: 隐层的数目 :param output_size: 输出的个数 ''' super(LSTM, self).__init__() self.hidden_layer_size = hidden_layer_size self.num_layers = num_layers self.lstm = nn.LSTM(input_size=input_size, hidden_size=hidden_layer_size, num_layers=num_layers) self.dropout = nn.Dropout(dropout_prob) self.linear = nn.Linear(hidden_layer_size, output_size) self.sigmoid = nn.Sigmoid() def forward(self, input_x): ''' input_x.view(len(input_x), 1, -1) 将输入数据 input_x 变形为一个新的形状，其中： 第一维：保持原始大小，即 200（表示数据中的样本数）。 第二维：将维度扩展为 1，因为 LSTM 模型的输入通常需要包含一个批次（batch）的信息。 第三维：自动调整以适应新的形状，以满足输入到 LSTM 模型的要求。 这个改变后的形状适合输入到 LSTM 模型的输入层。在 LSTM 中，输入数据的形状通常是 (batch_size, seq_len, input_size)，其中: batch_size: 表示每个批次中包含的数据样本数量。 seq_len: 表示数据序列的长度。 input_size: 表示每个时间步所包含的特征数。 ⭐⭐⭐我的理解：将输入数据转换成 一个batch的样本数，每个样本序列长度，剩下的为特征数，也就是（样本数, 1, -1） ''' input_x = input_x.view(len(input_x), 1, -1) # (batch, input_size)，其中input_x.shape=(1,5)，可以参照luojianmiaoshixiong的(input.size(0),input.size(1),-1) ''' 这行代码的作用是初始化 LSTM 的隐藏状态和细胞状态，使其变为 (1, 1, self.hidden_layer_size) 的形状，以便与输入数据相匹配。 在这里，我们用 torch.zeros 来创建全零的张量作为初始隐藏状态和细胞状态。其中，第一个 1 表示层数，第二个 1 表示批次大小， 而 self.hidden_layer_size 则表示每个隐藏状态的大小。 n_layers=1表示我们使用的是单层 LSTM ''' ''' 这个语句的作用是创建 LSTM 模型的初始隐藏状态和细胞状态。它包含两部分，分别对应隐藏状态（h）和细胞状态（c）。 ''' hidden_cell = (torch.zeros(self.num_layers, 1, self.hidden_layer_size), torch.zeros(self.num_layers, 1, self.hidden_layer_size)) # shape:(n_layers, batch_size, hidden_size) ''' 在 LSTM 模型中进行前向传播计算，并获取输出 lstm_out 以及最终的隐藏状态和细胞状态 h_n，h_c。 ''' lstm_out, (h_n, h_c) = self.lstm(input_x, hidden_cell) linear_out = self.linear(lstm_out.view(len(input_x), -1)) # = self.linear(lstm_out[:,-1,:]) drop_out = self.dropout(linear_out) predictions = self.sigmoid(drop_out) return predictions def means(list): mean_res = [] for lt in list: m = np.mean(lt) mean_res.append(round(m, 4)) return mean_res def std(list): std_res = [] for lt in list: s = np.std(lt) std_res.append(round(s, 4)) return std_res def main(epoch, hidden_size, num_layers, batch_size, lr, dropout_prob): # 设定随机种子以保证结果的可复现性 torch.manual_seed(seed) np.random.seed(seed) train, label = cross.bold_train() mymodel = 'LSTM' # 'SVM' vs 'DT' index = 'BOLD-' + mymodel sample = r'yes' best_accuracy = 0.0 best_model = None balanced_res = [] accuracy_res = [] sensitivity_res = [] specificity_res = [] auc_res = [] f1_res = [] all_tp = [] all_tn = [] all_fp = [] all_fn = [] skf = StratifiedKFold(n_splits=FOLD, shuffle=True, random_state=seed) fold = 0 for train_index, test_index in skf.split(train, label): print(f'#####fold {fold}') X_train, X_test = train[train_index], train[test_index] Y_train, Y_test = np.array(label)[train_index], np.array(label)[test_index] _, predicted_labels, true_labels, model = train_eval(X_train, Y_train, X_test, Y_test, epoch, fold, hidden_size, num_layers, batch_size, lr, dropout_prob) accuracy, balanced, sensitivity, specificity, auc, f1, tp, tn, fp, fn = confusion_compute(predicted_labels, true_labels) accuracy_res.append(accuracy) balanced_res.append(balanced) sensitivity_res.append(sensitivity) specificity_res.append(specificity) auc_res.append(auc) f1_res.append(f1) all_tp.append(tp) all_tn.append(tn) all_fp.append(fp) all_fn.append(fn) print(f'Accuracy: {accuracy_res[fold]:.4f}') print(f'Balanced Accuracy: {balanced_res[fold]:.4f}') print(f'Sensitivity (TPR): {sensitivity_res[fold]:.4f}') print(f'Specificity (TNR): {specificity_res[fold]:.4f}') print(f'AUC: {auc_res[fold]:.4f}') print(f'F1 Score: {f1_res[fold]:.4f}') fold += 1 # 保存最佳模型和参数 if accuracy > best_accuracy: best_accuracy = accuracy best_model = model # PATH = f'result/LSTM-test/state_dict_model_{epoch}.pth' # if not os.path.exists(PATH): # torch.save(best_model.state_dict(), PATH) # print(f"Epoch: {epoch}, Best accuracy: {best_accuracy:.2f}") all_list = [accuracy_res, balanced_res, sensitivity_res, specificity_res, auc_res, f1_res, all_tp, all_tn, all_fp, all_fn] mean_result = means(all_list) std_result = std(all_list[:6]) mid = tuple(str(mean_result[i]) + "±" + str(std_result[i]) for i in range(6)) + tuple(mean_result[6:]) + tuple( [sample, epoch, FOLD, hidden_size, num_layers, batch_size, lr, dropout_prob]) final_result = [mid] if os.path.exists(file): analyse_table = pd.DataFrame(final_result, index=[index]) analyse_table.to_csv(file, mode='a', header=False) else: analyse_table = pd.DataFrame(final_result, index=[index], columns=['accuracy', 'balanced accuracy', 'sensitivity', 'specificity', 'auc', 'f1