template_p01.py

Deep Learning/template_p01.py

@@ -7,9 +7,13 @@ def softmax(vector):
     return: np.array of shape (n, m)
         Matrix where softmax is computed for every row independently
     '''
-    nice_vector = vector - vector.max()
+    # 数值稳定处理：减去行最大值，并用 np.newaxis 保持维度
+    nice_vector = vector - np.max(vector, axis=1)[:, np.newaxis] 
+
     exp_vector = np.exp(nice_vector)
+    # 求分母：沿着 axis=1 求和，并用 np.newaxis 保持维度
     exp_denominator = np.sum(exp_vector, axis=1)[:, np.newaxis]
+
     softmax_ = exp_vector / exp_denominator
     return softmax_
 
@@ -22,7 +26,18 @@ def multiplicative_attention(decoder_hidden_state, encoder_hidden_states, W_mult
     return: np.array of shape (n_features_enc, 1)
         Final attention vector
     '''
-    # your code here
+    # 1. 计算 Attention Scores (e_i = s^T W_mult h_i)
+    # 步骤 1a: 计算 W_mult @ s
+    temp_dec_proj = np.dot(W_mult, decoder_hidden_state) 
+
+    # 步骤 1b: 计算 (temp_dec_proj)^T @ h
+    attention_scores = np.dot(temp_dec_proj.T, encoder_hidden_states)
+
+    # 2. 计算 Attention Weights (Softmax)
+    attention_weights = softmax(attention_scores)
+
+    # 3. 计算 Context Vector (加权求和)
+    attention_vector = np.dot(encoder_hidden_states, attention_weights.T)
 
     return attention_vector
 
@@ -37,6 +52,33 @@ def additive_attention(decoder_hidden_state, encoder_hidden_states, v_add, W_add
     return: np.array of shape (n_features_enc, 1)
         Final attention vector
     '''
-    # your code here
+    # 1. 投影 Q 和 K 到注意力空间 (n_features_int)
+
+    # 1a. 解码器状态投影 (W_add_dec @ s)
+    # dec_proj 形状: (n_features_int, 1)
+    dec_proj = np.dot(W_add_dec, decoder_hidden_state)
+
+    # 1b. 编码器状态投影 (W_add_enc @ h)
+    # enc_proj 形状: (n_features_int, n_states)
+    enc_proj = np.dot(W_add_enc, encoder_hidden_states)
+
+    # 2. 相加并应用激活函数 (tanh)
+    # H_sum = W_add_enc h + W_add_dec s
+    # dec_proj (n_features_int, 1) 会自动广播到 enc_proj (n_features_int, n_states)
+    Sum_Proj = enc_proj + dec_proj
+    H_tanh = np.tanh(Sum_Proj)
+
+    # 3. 计算 Attention Scores (e_i = v_add^T @ H_tanh)
+    # attention_scores 形状: (1, n_states)
+    attention_scores = np.dot(v_add.T, H_tanh)
+
+    # 4. 计算 Attention Weights (Softmax)
+    # attention_weights 形状: (1, n_states)
+    attention_weights = softmax(attention_scores)
+
+    # 5. 计算 Context Vector (加权求和)
+    # C = V @ alpha^T
+    # attention_vector 形状: (n_features_enc, 1)
+    attention_vector = np.dot(encoder_hidden_states, attention_weights.T)
 
     return attention_vector