MatMul 中的 Keras LSTM Dtype 不一样

Question

因此，我正在编写 LSTM 模型以查找序列中的下一个数字。但是当我拟合模型时出现这个错误，

TypeError: Input 'b' of 'MatMul' Op has type float32 that does not match type int64 of argument 'a'.

我不明白这个问题，因为我检查过所有数据和标签都是dtype int64。

这是我的代码：

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, LSTM
import random
import numpy as np

def makeData(total, types=5):
    #PATTERNS:
    #1)going up by 1
    #2)going up by 2
    #3)muliply by 3
    #4)multiple by 2
    #5)square number
    NumPer = int(total/types)
    Labels = []
    Data = []
    print(NumPer)
    for type in range(types):
        for i in range(NumPer):
            preData = []
            if type == 0:
                x = random.random()*100
                x = round(x)
                preData.append(x)
                preData.append(x+1)
                preData.append(x+2)
                Labels.append([x+3])
            if type == 1:
                x = random.random()*100
                x = round(x)
                preData.append(x)
                preData.append(x+2)
                preData.append(x+4)
                Labels.append([x+6])
            if type == 2:
                x = random.random()*100
                x = round(x)
                preData.append(x)
                preData.append(x*3)
                preData.append(x*9)
                Labels.append([x*27])
            if type == 3:
                x = random.random()*100
                x = round(x)
                preData.append(x)
                preData.append(x*2)
                preData.append(x*4)
                Labels.append([x*8])
            if type == 4:
                x = random.random()*100
                x = round(x)
                preData.append(x)
                preData.append(x*x)
                preData.append((x*x)*(x*x))
                Labels.append([(x*x)*(x*x)*((x*x)*(x*x))])
        Data.append(preData)
    return np.array(Data), np.array(Labels)

x, y = makeData(5)

def MakeRNN():
    model = Sequential()
    model.add(LSTM(3, activation='tanh'))
    model.add(Dense(9, activation='relu'))
    model.add(LSTM(9, activation='tanh'))
    model.add(Dense(36, activation='relu'))
    model.add(Dense(1, activation='relu'))
    model.compile(optimizer='Adam', loss="mse", metrics=['accuracy'])
    return model
Model = MakeRNN()
print(x.shape)
x = x.reshape(-1,1,3)
print(y[0].dtype)
print(x.shape)
Model.fit(x,y, epochs=20)

print(Model.predict([[1,2,3]]))

我不明白这个，如果你们能提供帮助，我将不胜感激。我对keras这个领域了解的不多

Answer 1

以下代码对我有用。只是对 makeData 中的 x, y 进行了一些调整，将它们转换为浮点数。此外，传递给 predict 的参数需要是 3 维的（目前你有一个 2d 参数）并且它需要是一个浮点数。还需要在第一个 LSTM 处添加 return sequences = True 因为它后面还有另一个 LSTM。另外，您的模型有点奇怪。 Keras 期望顺序模型的输入为 (batch_size, timesteps, input_dim) 形式。您有 timesteps = 1，因此实际上没有序列。但这是你想做什么的另一回事。至于将代码获取到运行，以上调整有效。

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, LSTM
import random
import numpy as np

def makeData(total, types=5):
    #PATTERNS:
    #1)going up by 1
    #2)going up by 2
    #3)muliply by 3
    #4)multiple by 2
    #5)square number
    NumPer = int(total/types)
    Labels = []
    Data = []
    print(NumPer)
    for type in range(types):
        for i in range(NumPer):
            preData = []
            if type == 0:
                x = random.random()*100
                x = round(x)
                preData.append(x)
                preData.append(x+1)
                preData.append(x+2)
                Labels.append([x+3])
            if type == 1:
                x = random.random()*100
                x = round(x)
                preData.append(x)
                preData.append(x+2)
                preData.append(x+4)
                Labels.append([x+6])
            if type == 2:
                x = random.random()*100
                x = round(x)
                preData.append(x)
                preData.append(x*3)
                preData.append(x*9)
                Labels.append([x*27])
            if type == 3:
                x = random.random()*100
                x = round(x)
                preData.append(x)
                preData.append(x*2)
                preData.append(x*4)
                Labels.append([x*8])
            if type == 4:
                x = random.random()*100
                x = round(x)
                preData.append(x)
                preData.append(x*x)
                preData.append((x*x)*(x*x))
                Labels.append([(x*x)*(x*x)*((x*x)*(x*x))])
        Data.append(preData)
    return np.array(Data), np.array(Labels)

x, y = makeData(5)
x = x.astype(float)
y = y.astype(float)

def MakeRNN():
    model = Sequential()
    model.add(LSTM(3, activation='tanh', return_sequences=True))
    model.add(Dense(9, activation='relu'))
    model.add(LSTM(9, activation='tanh'))
    model.add(Dense(36, activation='relu'))
    model.add(Dense(1, activation='relu'))
    model.compile(optimizer='Adam', loss="mse", metrics=['accuracy'])
    return model
Model = MakeRNN()
print(x.shape)
x = x.reshape(-1,1,3)
print(y[0].dtype)
print(x.shape)
Model.fit(x,y, epochs=20)

print(Model.predict(np.array([[[1.0,2.0,3.0]]])))

MatMul 中的 Keras LSTM Dtype 不一样

Keras LSTM Dtype in MatMul not the same

lstm

keras

tensorflow