From 59d829cad0654743514a462b781080161baf99ea Mon Sep 17 00:00:00 2001
From: k-gorbenko <164647543+k-gorbenko@users.noreply.github.com>
Date: Wed, 4 Dec 2024 22:31:04 +0300
Subject: [PATCH] Update template_p09.py

---
 Deep Learning/template_p01.py | 40 ++++++++++++++++++++++++++++-------
 1 file changed, 32 insertions(+), 8 deletions(-)

diff --git a/Deep Learning/template_p01.py b/Deep Learning/template_p01.py
index 190fb4e..055d0f3 100644
--- a/Deep Learning/template_p01.py	
+++ b/Deep Learning/template_p01.py	
@@ -18,13 +18,23 @@ def multiplicative_attention(decoder_hidden_state, encoder_hidden_states, W_mult
     decoder_hidden_state: np.array of shape (n_features_dec, 1)
     encoder_hidden_states: np.array of shape (n_features_enc, n_states)
     W_mult: np.array of shape (n_features_dec, n_features_enc)
-    
+
     return: np.array of shape (n_features_enc, 1)
         Final attention vector
     '''
-    # your code here
-    
-    return attention_vector
+    # Умножаем матрицу весов на вектор скрытого состояния декодера
+    attention_scores = W_mult @ decoder_hidden_state  # (n_features_enc, 1)
+
+    # Умножаем полученные оценки внимания на скрытые состояния кодера
+    attention_weights = np.dot(encoder_hidden_states.T, attention_scores)  # (n_states, 1)
+
+    # Применяем softmax к вектору внимания
+    attention_weights = np.exp(attention_weights) / np.sum(np.exp(attention_weights), axis=0)
+
+    # Вычисляем взвешенную сумму скрытых состояний кодера
+    attention_vector = np.dot(encoder_hidden_states, attention_weights)  # (n_features_enc, 1)
+
+    return attention_vector  # Возвращаем вектор в виде (n_features_enc, 1)
 
 def additive_attention(decoder_hidden_state, encoder_hidden_states, v_add, W_add_enc, W_add_dec):
     '''
@@ -33,10 +43,24 @@ def additive_attention(decoder_hidden_state, encoder_hidden_states, v_add, W_add
     v_add: np.array of shape (n_features_int, 1)
     W_add_enc: np.array of shape (n_features_int, n_features_enc)
     W_add_dec: np.array of shape (n_features_int, n_features_dec)
-    
+
     return: np.array of shape (n_features_enc, 1)
         Final attention vector
     '''
-    # your code here
-    
-    return attention_vector
+    # Применяем линейные преобразования к скрытым состояниям кодера и декодера
+    encoder_transformed = W_add_enc @ encoder_hidden_states  # (n_features_int, n_states)
+    decoder_transformed = W_add_dec @ decoder_hidden_state  # (n_features_int, 1)
+
+    # Суммируем результаты
+    scores = v_add @ (encoder_transformed + decoder_transformed.T)  # (1, n_states)
+
+    # Применяем активацию tanh
+    scores = np.tanh(scores)  # (1, n_states)
+
+    # Применяем softmax к оценкам внимания
+    attention_weights = np.exp(scores) / np.sum(np.exp(scores), axis=1, keepdims=True)  # (1, n_states)
+
+    # Вычисляем взвешенную сумму скрытых состояний кодера
+    attention_vector = encoder_hidden_states @ attention_weights.T  # (n_features_enc, 1)
+
+    return attention_vector  # Возвращаем вектор в виде (n_features_enc, 1)