Pytorch中 batch_first 选择True/False的区别

batch_first – If True, then the input and output tensors are provided as (batch, seq, feature) instead of (seq, batch, feature). Note that this does not apply to hidden or cell states. See the Inputs/Outputs sections below for details. Default: False

<br>rnn = nn.RNN(input_size=4,hidden_size=3,num_layers=2,batch_first=True)
input = torch.randn(1,5,4)
output , h_n = rnn(input)
print(output.shape) 
print(h_n.shape)

输出结果:

#output

tensor([[[-0.4026, -0.2417, -0.1307],
[-0.0122, 0.4269, -0.7256],
[ 0.2228, 0.7731, -0.9092],
[-0.3735, 0.4446, -0.6930],
[-0.1539, 0.5937, -0.8616]]], grad_fn=<TransposeBackward1>)

#h_n
tensor([[[-0.5664, 0.0416, -0.9316]],
[[-0.1539, 0.5937, -0.8616]]], grad_fn=<StackBackward0>)

1 2	`input.Size([1,` `5,` `4])` `# batch_size, seq_len, feature` `output.Size([1,` `5,` `3])` `# batch_size, seq_len, hidden_sizeh_n.Size([2, 1, 3]) # num_layers , batch_size, hidden_sizebatch_size = 1 num_layers = 2 hidden_size = 3 input_size = 4 seq_len = 5`

当batch_first = False

rnn = nn.RNN(input_size=4,hidden_size=3,num_layers=2,batch_first=False)
input = torch.randn(1,5,4)
output , h_n = rnn(input)
print(output.shape)
print(h_n.shape)
 
输出结果：
input.Size([1, 5, 4]) # seq_len, batch_size, feature
output.Size([1, 5, 3]) # seq_len, batch_size, hidden-size
h_n.Size([2, 5, 3])  # num_layers, batch_size, hidden-size
 
seq_len = 1  num_layers = 2 hidden_size = 3  input_size = 4 batch_size = 5
 
#output
tensor([[[-0.6884,  0.0477, -0.3248],
         [-0.5575, -0.0757, -0.4916],
         [-0.6645,  0.2197, -0.4582],
         [-0.6820,  0.1047, -0.4033],
         [-0.6624,  0.0487, -0.3798]]], grad_fn=<StackBackward0>)
#h_n
tensor([[[ 0.4404, -0.4511,  0.2594],
         [ 0.8487,  0.3987,  0.2429],
         [ 0.0287, -0.7793, -0.4574],
         [-0.4603, -0.7794,  0.4563],
         [ 0.4304, -0.3424,  0.1715]],
 
        [[-0.6884,  0.0477, -0.3248],
         [-0.5575, -0.0757, -0.4916],
         [-0.6645,  0.2197, -0.4582],
         [-0.6820,  0.1047, -0.4033],
         [-0.6624,  0.0487, -0.3798]]], grad_fn=<StackBackward0>)