淘先锋技术网

首页 1 2 3 4 5 6 7

RNN，LSTM，GRU的维度分析

RNN 参数

nn.RNN 类初始化主要参数：

  • input_szie：输入张量x的维度大小
  • hidden_size：隐藏层张量h的维度大小
  • num_layer：隐含层的数量

nn.RNN 类实例化对象主要参数

  • input：输入张量x
  • h0：初始化的隐含层张量h

RNN 使用示例

  • 初始化RNN：nn.RNN(input_size,hidden_size,num_layer)
  • input的维度：(seq_len,batch_size,input_size)
  • h0的维度：(num_directions*num_layer,batch_size,hidden_size)
  • output的维度：(seq_len,batch_size,num_directions*hidden_size)
  • hn的维度：(num_directions*num_layer,batch_size,hidden_size)

注：默认num_directions=1

# input_size=4, hidden_size=6, num_layer=1
rnn=nn.RNN(4,6,1)
# seq_len=2, batch_size=3, input_size=4
input=torch.randn(2,3,4)
# num_layer=1, batch_size=3, hidden_size=6
h0=torch.randn(1,3,6)

output,hn=rnn(input,h0)
print("out:",output.shape,output)
print("hn:",hn.shape,hn)

# seq_len=2, batch_size=3, num_directions*hidden_size=6
out: torch.Size([2, 3, 6]) 
tensor([[[-0.6265, -0.0157,  0.9049, -0.9148, -0.1023,  0.8824],
         [ 0.1760,  0.1963,  0.3808, -0.9247, -0.2264,  0.6422],
         [ 0.3331, -0.9721, -0.7927, -0.3843,  0.8845,  0.3520]],

        [[ 0.3291, -0.3104, -0.7785, -0.5462,  0.8294,  0.9277],
         [ 0.7093, -0.7809, -0.6781, -0.5684,  0.9314,  0.8526],
         [ 0.5716, -0.8184, -0.2193, -0.6427,  0.7650,  0.5599]]],
       grad_fn=<StackBackward0>)

# num_directions*num_layer=1,batch_szie=3,hidden_size=6
hn: torch.Size([1, 3, 6]) 
tensor([[[ 0.8884,  0.8530, -0.7011, -0.6993, -0.5694, -0.3473],
         [ 0.3698,  0.2642,  0.3087, -0.0876, -0.5907, -0.2327],
         [ 0.5826,  0.7660,  0.3343, -0.5449, -0.7647, -0.0619]]],
       grad_fn=<StackBackward0>)

通常情况下，我们会在初始化RNN时设置 batch_size=True。这会使input和output的维度有所变化，但它不会影响h0和hn的维度。

  • 初始化RNN：nn.RNN(input_size,hidden_size,num_layer,batch_first=True)
  • input的维度：(batch_size,seq_len,input_size)
  • h0的维度：(num_directions*num_layer,batch_size,hidden_size)
  • output的维度：(batch_size,seq_len,num_directions*hidden_size)
  • hn的维度：(num_directions*num_layer,batch_size,hidden_size)

如果是双向RNN，初始化RNN时设置 bidirectional =True，即num_directions=2，这会影响output，h0和hn的维度。

  • 初始化RNN：nn.RNN(input_size,hidden_size,num_layer,batch_first=True,bidirectional = True)
  • input的维度：(batch_size,seq_len,input_size)
  • h0的维度：(num_directions​​​​​​​*num_layer,batch_size,hidden_size)
  • output的维度：(batch_size,seq_len,num_directions​​​​​​​*hidden_size)
  • hn的维度：(num_directions​​​​​​​*num_layer,batch_size,hidden_size)

LSTM 参数

nn.LSTM 类初始化主要参数：

  • input_szie：输入张量x的维度大小
  • hidden_size：隐藏层张量h的维度大小
  • num_layer：隐含层的数量

nn.LSTM 类实例化对象主要参数

  • input：输入张量x
  • h0：初始化的隐含层张量h
  • c0：初始化的细胞状态张量c

LSTM 使用示例

  • 初始化LSTM：nn.LSTM(input_size,hidden_size,num_layer)
  • input的维度：(seq_len,batch_size,input_size)
  • h0的维度：(num_directions*num_layer,batch_size,hidden_size)
  • c0的维度：同h0
  • output的维度：(seq_len,batch_size,num_directions*hidden_size)
  • hn的维度：(num_directions*num_layer,batch_size,hidden_size)
  • cn的维度：同hn

注：默认num_directions=1

# input_size=4, hidden_size=6, num_layer=1
lstm=nn.LSTM(4,6,1)
# seq_len=2, batch_size=3, input_size=4
input=torch.randn(2,3,4)
# num_layer=1,batch_size=3, hidden_size=6
h0=torch.randn(1,3,6)
c0=torch.randn(1,3,6)

output,(hn,cn)=lstm(input,(h0,c0))
print("out",output.shape,output)
print("hn",hn.shape,hn)
print("cn",cn.shape,cn)

# seq_len=2, batch_size=3, num_directions*hidden_size=6
torch.Size([2, 3, 6]) 
tensor([[[-0.0414,  0.5901, -0.3420,  0.0888,  0.1882, -0.2603],
         [ 0.1543, -0.0668,  0.0877, -0.1536,  0.1953,  0.3458],
         [-0.1405, -0.0209,  0.1439,  0.0977,  0.0071, -0.1940]],

        [[ 0.0318,  0.2772, -0.0229,  0.0464,  0.1286, -0.0865],
         [ 0.1021,  0.2384,  0.0941, -0.1227,  0.1751,  0.3902],
         [-0.0881,  0.1334,  0.0564, -0.0522,  0.0354, -0.0247]]],
       grad_fn=<StackBackward0>)

# num_directions*num_layer=1,batch_size=3, hidden_size=6
torch.Size([1, 3, 6]) 
tensor([[[ 0.0318,  0.2772, -0.0229,  0.0464,  0.1286, -0.0865],
         [ 0.1021,  0.2384,  0.0941, -0.1227,  0.1751,  0.3902],
         [-0.0881,  0.1334,  0.0564, -0.0522,  0.0354, -0.0247]]],
       grad_fn=<StackBackward0>)

# num_directions*num_layer=1,batch_size=3, hidden_size=6
torch.Size([1, 3, 6]) 
tensor([[[ 0.0832,  0.4552, -0.0624,  0.1234,  0.3947, -0.1668],
         [ 0.2845,  0.4464,  0.3017, -0.3624,  0.2983,  0.7948],
         [-0.2298,  0.1902,  0.1637, -0.1332,  0.0778, -0.0428]]],
       grad_fn=<StackBackward0>)

关于 LSTM的 batch_first 和 bidirectional ​​​​​​​的设置参考RNN

GRU 参数

nn.GRU 类初始化主要参数：

  • input_szie：输入张量x的维度大小
  • hidden_size：隐藏层张量h的维度大小
  • num_layer：隐含层的数量

nn.GRU 类实例化对象主要参数

  • input：输入张量x
  • h0：初始化的隐含层张量h

GRU 使用示例

GRU的维度和RNN一模一样。

# input_size=4, hidden_size=6, num_layer=1
gru=nn.GRU(4,6,1)
# seq_len=2, batch_size=3, input_size=4
input=torch.randn(2,3,4)
# num_layer=1,batch_size=3, hidden_size=6
h0=torch.randn(1,3,6)

output,hn=gru(input,h0)
print("out",output.shape,output)
print("hn",hn.shape,hn)

# seq_len=2, batch_size=3, num_directions*hidden_size=6
torch.Size([2, 3, 6]) 
tensor([[[-0.9060, -0.7757,  0.7011, -0.4514, -1.0205, -0.6123],
         [ 0.6067, -0.1415, -1.3128, -0.2117, -0.4429, -0.2052],
         [ 0.0051,  0.0630, -0.0658,  0.1197, -0.4444, -0.2348]],

        [[-0.0020, -0.3685,  0.2763, -0.3061, -0.7251, -0.5263],
         [ 0.3088, -0.2424, -0.9513, -0.0241, -0.4825,  0.0095],
         [ 0.2136,  0.2759,  0.2112,  0.3923, -0.2075, -0.3016]]],
       grad_fn=<StackBackward0>)

# num_directions*num_layer=1,batch_size=3, hidden_size=6
torch.Size([1, 3, 6]) 
tensor([[[-0.0020, -0.3685,  0.2763, -0.3061, -0.7251, -0.5263],
         [ 0.3088, -0.2424, -0.9513, -0.0241, -0.4825,  0.0095],
         [ 0.2136,  0.2759,  0.2112,  0.3923, -0.2075, -0.3016]]],
       grad_fn=<StackBackward0>)