按名稱為keras TimeDistributed圖層設置權重

Question

我在Keras中有一個模型，如下所示：

data = Input(shape=input_shape)

# 512 x 640 x 3
pad1 = TimeDistributed(ZeroPadding2D(padding=(100, 100)))(data)

# 712 x 840 x 3
conv1_1 = TimeDistributed(Conv2D(8, (3,3), padding="valid", activation="relu", name="block1_conv1", data_format="channels_last"))(pad1)
conv1_2 = TimeDistributed(Conv2D(8, (3,3), padding="same", activation="relu", name="block1_conv2", data_format="channels_last"))(conv1_1)
pool1   = TimeDistributed(MaxPooling2D((2,2), strides=(2,2), padding="same", name="block1_pool", data_format="channels_last"))(conv1_2)

我希望能夠將conv1_1和conv1_2的可訓練權重參數設置為每個時間步長的預訓練值。 我可以這樣做嗎？ Keras似乎將這些層視為具有自己可訓練參數的自身實體，而不是視為具有相同共享可訓練權重的Conv2D函數的集合。 有辦法改變嗎？ 如何訪問單個時間片的可訓練權重並將其分配到所有時間片？

Answer 1

您可以這樣做：

data = Input(shape=input_shape)

# 512 x 640 x 3
pad1 = TimeDistributed(ZeroPadding2D(padding=(100, 100)))(data)

# 712 x 840 x 3
nd_conv1_1 = Conv2D(8, (3,3), padding="valid", activation="relu", name="block1_conv1", data_format="channels_last")
nd_conv1_2 = Conv2D(8, (3,3), padding="same", activation="relu", name="block1_conv2", data_format="channels_last")
conv1_1 = TimeDistributed(nd_conv1_1)(pad1)
conv1_2 = TimeDistributed(nd_conv1_2)(conv1_1)
pool1   = TimeDistributed(MaxPooling2D((2,2), strides=(2,2), padding="same", name="block1_pool", data_format="channels_last"))(conv1_2)

nd_conv1_1.trainable = True/False
nd_conv1_2.trainable = True/False