Tensorflow：一个输入有多个一键输出？

问题描述

是否有一种方法可以馈送前馈tensorflow模型，其标签在一个输入中包含多个热编码？

例如：

x[0]@H_502_6@：[15,32,3]@H_502_6@，y[0]@H_502_6@：[[0,1],[0,1,0],[1,0]]@H_502_6@

x[1]@H_502_6@：[23,2,7]@H_502_6@，y[1]@H_502_6@：[[1,0]]@H_502_6@

x的形状为（1,3），y的形状为（3，3）

有没有办法在简单的前馈神经网络中训练像这样的数据？

解决方法

您可以使用自定义训练循环和从Keras.Model子类化的模型来实现。然后，您可以拥有3个损失3个输出的输出（可能甚至是分类的和连续的混合）。

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
import tensorflow as tf
from tensorflow.keras.layers import Dense
from tensorflow.keras import Model
from functools import partial
import numpy as np

X = np.random.randint(0,100,(100,3)).astype(np.float32)
w = np.random.randint(0,3,100)
y = np.random.randint(0,100)
z = np.random.randint(0,100)

onehot = partial(tf.one_hot,depth=3)

dataset = tf.data.Dataset.from_tensor_slices((X,w,y,z)).\
    shuffle(100).\
    batch(4).\
    map(lambda a,b,c,d: (a,onehot(b),onehot(c),onehot(d)))


print(next(iter(dataset)))


class MyModel(Model):
    def __init__(self):
        super(MyModel,self).__init__()
        self.d0 = Dense(16,activation='relu')
        self.d1 = Dense(32,activation='relu')
        self.d2 = Dense(3)
        self.d3 = Dense(3)
        self.d4 = Dense(3)

    def call(self,x,training=None,**kwargs):
        x = self.d0(x)
        x = self.d1(x)
        out1 = self.d2(x)
        out2 = self.d3(x)
        out3 = self.d4(x)
        return out1,out2,out3

model = MyModel()

loss_object = tf.losses.categorical_crossentropy

def compute_loss(model,z,training):
  out1,out3 = model(inputs=x,training=training)
  loss1 = loss_object(y_true=w,y_pred=out1,from_logits=True)
  loss2 = loss_object(y_true=y,y_pred=out2,from_logits=True)
  loss3 = loss_object(y_true=z,y_pred=out3,from_logits=True)
  return loss1,loss2,loss3

def get_grad(model,z):
    with tf.GradientTape() as tape:
        loss1,loss3 = compute_loss(model,training=False)
    gradients = tape.gradient([loss1,loss3],model.trainable_variables)
    return (loss1,loss3),gradients

optimizer = tf.optimizers.Adam()

verbose = "Epoch {:2d} Loss1: {:.3f} Loss2: {:.3f} Loss3: {:.3f}"

for epoch in range(1,10 + 1):
    loss1 = tf.metrics.Mean()
    loss2 = tf.metrics.Mean()
    loss3 = tf.metrics.Mean()

    for X,z in dataset:
        losses,grads = get_grad(model,X,z)
        optimizer.apply_gradients(zip(grads,model.trainable_variables))

        for current_loss,running_loss in zip(losses,[loss1,loss3]):
            running_loss.update_state(current_loss)

    print(verbose.format(epoch,loss1.result(),loss2.result(),loss3.result()))

输入：

(<tf.Tensor: shape=(4,3),dtype=float32,numpy=
array([[45.,35.,46.],[64.,95.,55.],[90.,41.,12.],[98.,17.,81.]],dtype=float32)>,<tf.Tensor: shape=(4,numpy=
array([[0.,0.,1.],[1.,0.],0.]],[0.,1.,dtype=float32)>)

输出：

Epoch  1 Loss1: 0.980 Loss2: 1.051 Loss3: 1.035
Epoch  2 Loss1: 0.886 Loss2: 0.961 Loss3: 0.934
Epoch  3 Loss1: 0.814 Loss2: 0.872 Loss3: 0.847
Epoch  4 Loss1: 0.762 Loss2: 0.804 Loss3: 0.786
Epoch  5 Loss1: 0.732 Loss2: 0.755 Loss3: 0.747
Epoch  6 Loss1: 0.718 Loss2: 0.731 Loss3: 0.723
Epoch  7 Loss1: 0.709 Loss2: 0.715 Loss3: 0.714
Epoch  8 Loss1: 0.700 Loss2: 0.708 Loss3: 0.705
Epoch  9 Loss1: 0.699 Loss2: 0.697 Loss3: 0.701
Epoch 10 Loss1: 0.703 Loss2: 0.702 Loss3: 0.698