Keras 序列模型没有维度不匹配吗？

发布于 2025-01-12 17:05:37 字数 670 浏览 0 评论 0原文

我使用 Keras 创建了一个顺序神经网络，其输入为 4，输出为 8。我意识到我所做的事情是不正确的，但我不确定为什么代码不会抛出错误。

print(X.shape)    # Prints (64, 4)
print(y.shape)    # Prints (64, 64, 8)

self.model.fit(X, y, batch_size=MINIBATCH_SIZE, verbose=0, shuffle=False)

那么为什么 Keras 接受数组的数组呢？它不应该只接受数组的数组吗？

编辑：这就是我的模型的创建方式

model = Sequential()
model.add(tf.keras.layers.InputLayer(input_shape=(4,)))
model.add(Dense(64))
model.add(Dense(Env.ACTION_SPACE, activation='linear'))             # Env.ACTION_SPACE = 8
model.compile(loss="mse", optimizer='adam', metrics=['accuracy'])

原文

I created a sequential neural network with Keras that has an input of 4 and an output of 8. I realize what I did was incorrect but I'm not sure as to why the code does not throw an error.

print(X.shape)    # Prints (64, 4)
print(y.shape)    # Prints (64, 64, 8)

self.model.fit(X, y, batch_size=MINIBATCH_SIZE, verbose=0, shuffle=False)

So why does Keras accept an array of array of arrays? Shouldn't it only accept an array of arrays?

EDIT:
This is how my model was created

model = Sequential()
model.add(tf.keras.layers.InputLayer(input_shape=(4,)))
model.add(Dense(64))
model.add(Dense(Env.ACTION_SPACE, activation='linear'))             # Env.ACTION_SPACE = 8
model.compile(loss="mse", optimizer='adam', metrics=['accuracy'])

分享到QQ

分享到微博

如果你对这篇内容有疑问，欢迎到本站社区发帖提问参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。

发布评论

需要登录才能够评论，你可以免费注册一个本站的账号。

风和你 2025-01-19 17:05:37

它们主要是 mariz 计算，然后通过函数扩展或减少维度，但最后一层，你可以 falttern、softmax 或总结成你想要的形状。
您可以从训练批次或预测中看到，它返回一组输出，您可以在其中选择批号大于 1 的位置，或者预测打印以查看输出结果，其中它们是您从 np.max 或 softwax 或附加网络搜索的答案堆栈的倍数。

我从这个例子中看到了你的评论

next_act = mainQ_outputs.evaluate(x=input_image, batch_size=16, max_queue_size=10, workers=16 )

或者你也可以这样做

predictions = model.predict(obs, batch_size=32)

示例输出：

### ( 1 ): Q-Networks
model = models.Sequential()         
    for layer in mainQ_outputs: 
     model.add(layer)
     model.add(tf.keras.layers.Flatten() )
     model.add(tf.keras.layers.Dense(6, activation=tf.nn.softmax))

### ( 2 ): predictions[0] - predictions[31]
action = np.argmax(predictions[0])  
action = actions_name[action]

...

目标移动预测 ( 1 )

目标运动预测（ 2）

They are mostly mariz computation then the dimension is expand or reduce by the function but the last layer, you can falttern, softmax or conclude into shape you want.
You can see from traing batch or prediction that it return set of output where you seelct batch number more than 1 or prediction print to see the output result where they are multipl of answers stacks you search from np.max or softwax or attached networks.

I saw your remarks that from this example

next_act = mainQ_outputs.evaluate(x=input_image, batch_size=16, max_queue_size=10, workers=16 )

Or you also can do

predictions = model.predict(obs, batch_size=32)

Sample output:

### ( 1 ): Q-Networks
model = models.Sequential()         
    for layer in mainQ_outputs: 
     model.add(layer)
     model.add(tf.keras.layers.Flatten() )
     model.add(tf.keras.layers.Dense(6, activation=tf.nn.softmax))

### ( 2 ): predictions[0] - predictions[31]
action = np.argmax(predictions[0])  
action = actions_name[action]

...

Target movement prediction ( 1 )

Target movement prediction ( 2 )

回复收藏 0 原文

~没有更多了~