如何将numpy.ndarray或keras metrix作为keras输入传递

如何解决如何将numpy.ndarray或keras metrix作为keras输入传递

我有一个形状为（2,30000）的矩阵列表，我需要使用conv层将此信息作为带有tensorflow的深度学习模型的输入来传递，但是当我试图一直传递以进行训练时我得到

    return ops.EagerTensor(value,ctx.device_name,dtype)
ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type numpy.ndarray).

我将numpy数组的列表作为熊猫系列

0      [[0.0,0.0,...
1      [[0.036600337822677166,0.018300168911338583,...
2      [[0.8671148170073495,0.47065322955211747,0.0...
3      [[0.024680190751413082,0.007051483071832309,...
4      [[0.0688791198957804,0.0229597...
                             ...                        
549    [[0.024182541670333724,0....
550    [[0.0,...

Name: tfidf,Length: 554,dtype: object

这是我的模特

def getmodel(num_words = 20000):
    x_input = keras.Input( shape=(2,30000),name="article1")  # Variable-length sequence of ints
    conv1d_1= layers.Conv1D(64,2,input_shape=(2,activation='relu')(x_input)
    global_1 = layers.GlobalMaxPooling1D()(conv1d_1)
    dense1 = layers.Dense(1024,name="dense1",activation="relu",)(global_1)
    encoder_conv_dense2 = layers.Dense(256,name="40_dense",)(dense1)
    encoder_conv_dense3 = layers.Dense(1,name="similar_result",)(encoder_conv_dense2)

    model = keras.Model(
        inputs=x_input,outputs=[encoder_conv_dense3],)
    keras.utils.plot_model(model,"my_paper_model.png",show_shapes=True)

    return model

以及我如何尝试传递信息

def compile_model(model,optimizer,loss,loss_weight):
    model.compile(
        optimizer=optimizer,loss=loss,loss_weights=loss_weight,metrics=[tf.keras.metrics.Accuracy()]
    )
    return model

def train(df):
callback = tf.keras.callbacks.EarlyStopping(monitor='accuracy',patience=40)
    callback2 = tf.keras.callbacks.EarlyStopping(monitor='loss',patience=20)
    model = getmodel()
    model = compile_model(model,keras.optimizers.Adam(),keras.losses.BinaryCrossentropy(),0.05)
    history = model.fit(
        {"article1":df['tfidf']},{"similar_result": df['is_similar']},validation_split=0.3,epochs=400,batch_size=32,verbose=1,callbacks=[callback,callback2,],)

我不知道如何传递2行和30,000列的numpy矩阵作为keras的输入，并将其与conv层一起使用。

希望有人可以帮助我

解决方法

您提供的输入内容格式不正确。在这种情况下，您需要提供3D输入的形状数组（n_sample，时间步长，n_features）。您可以简单地使用np.stack(df['tfidf'],0)来实现，这将导致形状为(n_sample,1,30000)的数组。

我尝试在下面复制一个虚拟示例：

# create fake data
df = pd.DataFrame()
df['tfidf'] = [[np.random.uniform(0,30)],[np.random.uniform(0,30)]]
df['is_similar'] = np.random.randint(0,2,4)

df['tfidf']的格式为：

0    [[0.09865182564241004,0.5282608042987893,0.5...
1    [[0.1361578046476558,0.9866056058771036,0.44...
2    [[0.38811373040427766,0.5686225139326878,0.8...
3    [[0.8254123154336716,0.3542711784901068,0.28...
Name: tfidf,dtype: object

定义模型并拟合：

def getmodel():
    
    x_input = keras.Input(shape=(1,30),name="article1")  # Variable-length sequence of ints
    conv1d_1= layers.Conv1D(64,activation='relu')(x_input)
    flat = layers.Flatten()(conv1d_1)
    dense1 = layers.Dense(1024,name="dense1",activation="relu")(flat)
    encoder_conv_dense2 = layers.Dense(256,name="40_dense",activation="relu")(dense1)
    encoder_conv_dense3 = layers.Dense(1,name="similar_result",activation="sigmoid")(encoder_conv_dense2)

    model = keras.Model(
        inputs=x_input,outputs=encoder_conv_dense3,)
    
    return model

model = getmodel()
model.compile('adam','binary_crossentropy')
history = model.fit(
    {"article1": np.stack(df['tfidf'],0)},{"similar_result": df['is_similar']},epochs=3,)