keras의 RNN 기반 API GRU 인수 return_state, return_sequences 정보

B = 1  #バッチサイズ
T = 10 #時系列長
N = 1000 #特徴量
data = np.random.randn(B, T, N)

tf.keras.layers.GRU

return_state=True, return_sequences=True

gru = tf.keras.layers.GRU(256, return_state=True, return_sequences=True)
B = 1
T = 10
N = 1000
data = np.random.randn(B, T, N)
outputs, states = gru(data)
print("赤丸：", outputs.shape)
print("緑丸：", states.shape)

赤丸： (1, 10, 256)
緑丸： (1, 256)

return_state=True, return_sequences=False

gru = tf.keras.layers.GRU(256, return_state=True, return_sequences=False)
B = 1
T = 10
N = 1000
data = np.random.randn(B, T, N)
outputs, states = gru(data)
print("赤丸：", outputs.shape)
print("緑丸：", states.shape)

赤丸： (1, 256)
緑丸： (1, 256)

return_state=False, return_sequences=True

gru = tf.keras.layers.GRU(256, return_state=False, return_sequences=True)
B = 1
T = 10
N = 1000
data = np.random.randn(B, T, N)
outputs = gru(data)
print("赤丸：", outputs.shape)
print("緑丸なし")

赤丸： (1, 10, 256)
緑丸なし

return_state=False, return_sequences=False

gru = tf.keras.layers.GRU(256, return_state=False, return_sequences=False)
B = 1
T = 10
N = 1000
data = np.random.randn(B, T, N)
outputs = gru(data)
print("赤丸：", outputs.shape)
print("緑丸なし")

赤丸： (1, 256)
緑丸なし