Q&A
前提・実現したいこと
ConcatenateでCNNを結合させたいのですがうまくいきません。解決方法を教えていただけるとありがたいです。
発生している問題・エラーメッセージ
ValueError: A `Concatenate` layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 25, 8), (None, 27, 8), (None, 27, 8)]
該当のソースコード
Python
1 input1 = Input(shape=(90,)) 2 input2 = Input(shape=(96,)) 3 input3 = Input(shape=(96,)) 4 hidden1 = Reshape((90, 1), input_shape = (90,))(input1) 5 hidden2 = Reshape((96, 1), input_shape = (96,))(input2) 6 hidden3 = Reshape((96, 1), input_shape = (96,))(input3) 7 8 filters = 8 9 kernel_size1 = 6 10 kernel_size2 = 3 11 x = Conv1D(filters, kernel_size1, strides=3, padding="valid",activation="relu")(hidden1) 12 x = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(x) 13 x = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(x) 14 x = Model(inputs=input1, outputs=x) 15 16 y = Conv1D(filters, kernel_size1, strides=3, padding="valid",activation="relu")(hidden2) 17 y = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(y) 18 y = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(y) 19 y = Model(inputs=input2, outputs=y) 20 21 z = Conv1D(filters, kernel_size1, strides=3, padding="valid",activation="relu")(hidden3) 22 z = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(z) 23 z = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(z) 24 z = Model(inputs=input3, outputs=z) 25 26 combined = concatenate([x.output, y.output, z.output]) 27 28 d = Dense(30, activation="relu")(combined) 29 d = Dense(3, activation="softmax")(d) 30 31 model = Model(inputs=[input1, input2, input3], outputs=d) 32 33 model.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy']) 34 if flag: 35 pass 36 else: 37 model.load_weights('transfer0.h5') 38 model.fit([x_traina,x_trainb,x_trainc], y_train, epochs=1, batch_size=1, validation_data=([x_testa,x_testb,x_testc], y_test))
###編集した部分
Google colaboratoryでプログラムをしています。
<ipython-input-4-6bf313ee2707> in learner(x_traina, x_trainb, x_trainc, y_train, x_testa, x_testb, x_testc, y_test, y_test1, flag) 44 x = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(x) 45 x = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(x) ---> 46 x = Flatten(x) 47 x = Model(inputs=input1, outputs=x) 48 /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/layers/core.py in __init__(self, data_format, **kwargs) 629 def __init__(self, data_format=None, **kwargs): 630 super(Flatten, self).__init__(**kwargs) --> 631 self.data_format = conv_utils.normalize_data_format(data_format) 632 self.input_spec = InputSpec(min_ndim=1) 633 /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/utils/conv_utils.py in normalize_data_format(value) 190 if value is None: 191 value = backend.image_data_format() --> 192 data_format = value.lower() 193 if data_format not in {'channels_first', 'channels_last'}: 194 raise ValueError('The `data_format` argument must be one of ' AttributeError: 'Tensor' object has no attribute 'lower'
Python
1import tensorflow.keras 2from tensorflow.keras.layers import Input, concatenate, Dense, Reshape, Conv1D, Flatten 3from tensorflow.keras.models import Model 4 5 input1 = Input(shape=(90,)) 6 input2 = Input(shape=(96,)) 7 input3 = Input(shape=(96,)) 8 hidden1 = Reshape((90, 1), input_shape = (90,))(input1) 9 hidden2 = Reshape((96, 1), input_shape = (96,))(input2) 10 hidden3 = Reshape((96, 1), input_shape = (96,))(input3) 11 12 filters = 8 13 kernel_size1 = 6 14 kernel_size2 = 3 15 x = Conv1D(filters, kernel_size1, strides=3, padding="valid",activation="relu")(hidden1) 16 x = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(x) 17 x = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(x) 18 x = Flatten(x) 19 x = Model(inputs=input1, outputs=x) 20 21 y = Conv1D(filters, kernel_size1, strides=3, padding="valid",activation="relu")(hidden2) 22 y = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(y) 23 y = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(y) 24 y = Flatten(y) 25 y = Model(inputs=input2, outputs=y) 26 27 z = Conv1D(filters, kernel_size1, strides=3, padding="valid",activation="relu")(hidden3) 28 z = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(z) 29 z = Conv1D(filters, kernel_size2, strides=1, padding="valid",activation="relu")(z) 30 z = Flatten(z) 31 z = Model(inputs=input3, outputs=z) 32 33 combined = concatenate([x.output, y.output, z.output]) 34 35 d = Dense(30, activation="relu")(combined) 36 37 d = Dense(3, activation="softmax")(d) 38 39 model = Model(inputs=[input1, input2, input3], outputs=d) 40 41 model.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy']) 42 if flag: 43 pass 44 else: 45 model.load_weights('transfer0.h5') 46 model.fit([x_traina,x_trainb,x_trainc], y_train, epochs=1, batch_size=1, validation_data=([x_testa,x_testb,x_testc], y_test)) 47 model.save_weights('transfer0.h5')
回答1件
0
ベストアンサー
欲しいモデルになっているかは分かりませんが
(hidden1 = Reshape((90, 1), input_shape = (90,))(input1)を
(hidden1 = Reshape((96, 1), input_shape = (90,))(input1)
にすると作成まではできました。
投稿2020/07/24 20:50
総合スコア38339
あなたの回答
tips
プレビュー
バッドをするには、ログインかつ
こちらの条件を満たす必要があります。
2020/07/24 23:28
2020/07/24 23:34
2020/07/25 01:15
2020/07/25 02:21
2020/07/25 04:27
2020/07/25 04:31
2020/07/25 05:53