层 lstm_9 的输入 0 与层不兼容:预期 ndim=3，发现 ndim=4.收到完整形状:[无、2、4000、256] [英] Input 0 of layer lstm_9 is incompatible with the layer: expected ndim=3, found ndim=4. Full shape received: [None, 2, 4000, 256]

问题描述

我尝试使用 RNN 网络创建模型，但我收到:第 lstm_9 层的输入 0 与该层不兼容:预期 ndim=3，发现 ndim=4.收到完整形状:[无、2、4000、256] 错误.

I try to create model with RNN network but I receive : Input 0 of layer lstm_9 is incompatible with the layer: expected ndim=3, found ndim=4. Full shape received: [None, 2, 4000, 256] error.

输入

train_data.shape() = (100,2,4000)

train_labels.shape() =(100,)

labels_values = 0 or 1 (two classes)

模型

input = Input(shape=(2,4000)) # shape from train_data
embedded = Embedding(2, 256)(input) 
lstm = LSTM(1024, return_sequences=True)(embedded) # ERROR
dense = Dense(2, activation='softmax')(lstm) 

推荐答案

不幸的是，您设计带有嵌入层的 Keras 功能模型的整个概念是错误的.

Your whole concept of designing Keras functional models with embedding layers is wrong, unfortunately.

1. 当您使用嵌入层时，它需要二维数据.

Input shape

2D tensor with shape: (batch_size, sequence_length).

Output shape

3D tensor with shape: (batch_size, sequence_length, output_dim).

参考:https://keras.io/layers/embeddings/

它需要词汇表的 ID 或标记序列.这必须是一个整数数组.

It takes a sequence of IDs or tokens for the vocabulary. This must be an integer array.

假设我们的词汇表有 len 36，我们传递给它一个范围为 (0, 36) 的整数数组列表

Let's say our vocabulary has len 36, we pass it a list of integer arrays in range (0, 36)

[1, 34, 32, 23] 有效[0.2, 0.5] 无效

[1, 34, 32, 23] is valid [0.2, 0.5] is not valid

1. 通常，我们使用 Embedding 来表示缩减空间中的向量，因此 output_dim 低于 input_dim，但基于设计也可能相反.

1. Usually, we use Embedding to represent the vectors in reduced space, so output_dim is lower than input_dim, but the opposite can be true too based on design.

您需要为输入数据指定 input_length.

You need to specify the input_length for the input data.

如果您使用 return_sequences = True 时间维度将被传递到下一个维度，这在您的情况下是不需要的.

If you use return_sequences = True the temporal dimension will be passed to the next dimension, it's not desired in your case.

你有 (0, 1, 0, 1, 0, 0, ...) 形式的标签，而不是单热编码形式，所以不要使用 softmax，而是使用 1 个单位的 sigmoid在最后的密集.

You have labels in the form (0, 1, 0, 1, 0, 0, ...) and not in one-hot-encoded form, so don't use softmax but sigmoid with 1 unit in the last dense.

这是稍微修正的网络.

from tensorflow.keras.layers import *
from tensorflow.keras.models import *
import numpy as np
train_data = np.random.randint(0,3, (100, 4000))
y_labels = np.random.randint(0,2, (100,))

input_ = Input(shape=(4000)) # shape from train_data
embedded = Embedding(36, 256, input_length = 4000)(input_) 
lstm = LSTM(256, return_sequences=False)(embedded) # --> ERROR
dense = Dense(1, activation='softmax')(lstm) 

model = Model(input_, dense)
model.summary()

Model: "model"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_6 (InputLayer)         [(None, 4000)]            0         
_________________________________________________________________
embedding_5 (Embedding)      (None, 4000, 256)         9216      
_________________________________________________________________
lstm_5 (LSTM)                (None, 256)               525312    
_________________________________________________________________
dense (Dense)                (None, 1)                 257       
=================================================================
Total params: 534,785
Trainable params: 534,785
Non-trainable params: 0

这篇关于层 lstm_9 的输入 0 与层不兼容:预期 ndim=3，发现 ndim=4.收到完整形状:[无、2、4000、256]的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！