From 58e266e18439aba3ef52ad5e882110f22cc9cc1f Mon Sep 17 00:00:00 2001 From: MuJ Date: Sun, 7 Jan 2024 09:39:05 +0800 Subject: [PATCH] =?UTF-8?q?=E4=BF=AE=E6=94=B9=E4=BA=86=E4=B8=80=E4=BA=9Bbu?= =?UTF-8?q?g=EF=BC=8C=E4=BD=86=E7=BB=93=E6=9E=9C=E8=BF=98=E6=98=AF?= =?UTF-8?q?=E6=B2=A1=E6=9C=89=E8=A7=84=E5=BE=8B?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- main.py | 53 ++++++++++++++++++++++++++++------------------------- 1 file changed, 28 insertions(+), 25 deletions(-) diff --git a/main.py b/main.py index 0435ec5..30de951 100644 --- a/main.py +++ b/main.py @@ -49,11 +49,13 @@ model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy']) # 训练模型 -model.fit(X_train, Y_train, epochs=3000, validation_split=0.2) +model.fit(X_train, Y_train, epochs=10, validation_split=0.2) # 评估模型 loss, accuracy = model.evaluate(X_test, Y_test) +print(X_train.shape) + # 制作扰动数据 @@ -85,48 +87,49 @@ gradient_magnitudes = tf.norm(gradients, axis=1) # 创建每个gamma对应的准确率的字典 accuracy_per_gamma = {} + +# 平坦化梯度 +flattened_gradients = tf.reshape(gradients, [-1]) + # 选择最大的γ * |X|个梯度 -for gamma in [0.05, 0.1, 0.2, 0.4]: # 你需要选择一个合适的γ值 - num_gradients_to_select = int( - gamma * tf.size(gradient_magnitudes, out_type=tf.dtypes.float32)) - - # 获取最大梯度的索引 - # tf.argsort返回的是升序排列的，所以用[::-1]来获得降序的索引 - top_gradients_indices = tf.argsort(gradient_magnitudes, direction='DESCENDING')[ - :num_gradients_to_select] - - # 创建集合A，这里是所有选定的最大梯度的索引 - A = top_gradients_indices.numpy() +for gamma in [0.05, 0.1, 0.2, 0.4]: + num_gradients_to_select = int(gamma * tf.size(flattened_gradients, out_type=tf.dtypes.float32)) + top_gradients_indices = tf.argsort(flattened_gradients, direction='DESCENDING')[:num_gradients_to_select] # 创建一个新的梯度张量，初始化为原始梯度的副本 - updated_gradients = tf.identity(gradients) + updated_gradients = tf.identity(flattened_gradients) - # 获取所有索引 - all_indices = tf.range(start=0, limit=tf.size( - gradient_magnitudes, out_type=tf.dtypes.int32), dtype=tf.int32) + # 创建一个布尔掩码，其中选定的最大梯度为False，其他为True + mask = tf.ones_like(updated_gradients, dtype=bool) + mask = tf.tensor_scatter_nd_update(mask, tf.expand_dims(top_gradients_indices, 1), tf.zeros_like(top_gradients_indices, dtype=bool)) - # 创建一个布尔掩码，其中集合A中的索引为False，其他为True - mask = ~tf.reduce_any(tf.equal(tf.expand_dims( - all_indices, 1), tf.reshape(A, (1, -1))), axis=1) + # 使用这个掩码更新梯度 + updated_gradients = tf.where(mask, tf.zeros_like(updated_gradients), updated_gradients) - # 使用这个掩码更新不在集合A中的梯度为0 - updated_gradients = tf.where( - mask, updated_gradients, tf.zeros_like(updated_gradients)) + # 将梯度重构为原始形状 + updated_gradients = tf.reshape(updated_gradients, tf.shape(gradients)) + # 创建准确率列表 accuracy_list = [] for learning_rate in [0.1, 0.2, 0.3, 0.4, 0.5]: - # 更新X_train_tensor - X_train_updated = X_train_tensor - updated_gradients + # 应用学习率到梯度 + scaled_gradients = learning_rate * updated_gradients + + + # 使用缩放后的梯度更新X_train_tensor + X_train_updated = X_train_tensor - scaled_gradients + tf.reshape(X_train_updated, (3332,34,1)) + X_train_updated = X_train_updated.numpy() # 编译模型 model.compile( optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy']) # 训练模型 - model.fit(X_train, Y_train, epochs=3000, validation_split=0.2) + model.fit(X_train_updated, Y_train, epochs=1500, validation_split=0.2) # 评估模型 loss, accuracy = model.evaluate(X_test, Y_test)