如何解决如何在 Pytorch 中实现双向 Conv LSTM
import torch
from torch import nn
def initialize_weights(self,layer):
"""Initialize a layer's weights and biases.
Args:
layer: A PyTorch Module's layer."""
if isinstance(layer,(nn.Batchnorm2d,nn.Batchnorm1d)):
pass
else:
try:
nn.init.xavier_normal_(layer.weight)
except AttributeError:
pass
try:
nn.init.uniform_(layer.bias)
except (ValueError,AttributeError):
pass
class HadamardProduct(nn.Module):
"""A Hadamard product layer.
Args:
shape: The shape of the layer."""
def __init__(self,shape):
super().__init__()
self.weights = nn.Parameter(torch.empty(*shape))
self.bias = nn.Parameter(torch.empty(*shape))
def forward(self,x):
return x * self.weights
class ConvLSTMCell(nn.Module):
"""A convolutional LSTM cell.
Implementation details follow closely the following paper:
Shi et al. -'Convolutional LSTM Network: A Machine Learning
Approach for Precipitation Nowcasting' (2015).
Accessible at https://arxiv.org/abs/1506.04214
The parameter names are drawn from the paper's Eq. 3.
Args:
input_bands: The number of bands in the input data.
input_dim: The length of of side of input data. Data is
presumed to have identical width and heigth."""
def __init__(self,input_bands,input_dim,kernels,dropout,batch_norm):
super().__init__()
self.input_bands = input_bands
self.input_dim = input_dim
self.kernels = kernels
self.dropout = dropout
self.batch_norm = batch_norm
self.kernel_size = 3
self.padding = 1 # Preserve dimensions
self.input_conv_params = {
'in_channels': self.input_bands,'out_channels': self.kernels,'kernel_size': self.kernel_size,'padding': self.padding,'bias': True
}
self.hidden_conv_params = {
'in_channels': self.kernels,'bias': True
}
self.state_shape = (
1,self.kernels,self.input_dim,self.input_dim
)
self.batch_norm_layer = None
if self.batch_norm:
self.batch_norm_layer = nn.Batchnorm2d(num_features=self.input_bands)
# Input Gates
self.W_xi = nn.Conv2d(**self.input_conv_params)
self.W_hi = nn.Conv2d(**self.hidden_conv_params)
self.W_ci = HadamardProduct(self.state_shape)
# Forget Gates
self.W_xf = nn.Conv2d(**self.input_conv_params)
self.W_hf = nn.Conv2d(**self.hidden_conv_params)
self.W_cf = HadamardProduct(self.state_shape)
# Memory Gates
self.W_xc = nn.Conv2d(**self.input_conv_params)
self.W_hc = nn.Conv2d(**self.hidden_conv_params)
# Output Gates
self.W_xo = nn.Conv2d(**self.input_conv_params)
self.W_ho = nn.Conv2d(**self.hidden_conv_params)
self.W_co = HadamardProduct(self.state_shape)
# Dropouts
self.H_drop = nn.Dropout2d(p=self.dropout)
self.C_drop = nn.Dropout2d(p=self.dropout)
self.apply(initialize_weights)
class ConvLSTM(nn.Module):
def __init__(self,num_layers,bidirectional,dropout):
super().__init__()
self.input_bands = input_bands
self.input_dim = input_dim
self.kernels = kernels
self.num_layers = num_layers
self.bidirectional = bidirectional
self.dropout = dropout
self.layers_fwd = self.initialize_layers()
self.layers_bwd = None
if self.bidirectional:
self.layers_bwd = self.initialize_layers()
self.fc_output = nn.Sequential(
nn.Flatten(),nn.Linear(
in_features=self.kernels*self.input_dim**2*(1 if not self.bidirectional else 2),out_features=1024
),nn.Linear(
in_features=1024,out_features=1
)
)
self.apply(initialize_weights)
def initialize_layers(self):
"""Initialize a single direction of the model's layers.
This function takes care of stacking layers,allocating
dropout and assigning correct input feature number for
each layer in the stack."""
layers = nn.ModuleList()
for i in range(self.num_layers):
layers.append(
ConvLSTMCell(
input_bands=self.input_bands if i == 0 else self.kernels,input_dim=self.input_dim,dropout=self.dropout if i+1 < self.num_layers else 0,kernels=self.kernels,batch_norm=False
)
)
return layers
def forward(self,x):
"""Perform forward pass with the model.
For each item in the sequence,the data is propagated
through each layer and both directions,if possible.
In case of a bidirectional model,the outputs are
concatenated from both directions. The output of the
last item of the sequence is further given to the FC
layers to produce the final batch of predictions.
Args:
x: A batch of spatial data sequences. The data
should be in the following format:
[Batch,Seq,Band,Dim,Dim]
Returns:
A batch of predictions."""
seq_len = x.shape[1]
for seq_idx in range(seq_len):
layer_in_out = x[:,seq_idx,::]
states = None
for layer in self.layers_fwd:
layer_in_out,states = layer(layer_in_out,states)
if not self.bidirectional:
continue
layer_in_out_bwd = x[:,-seq_idx,::]
states = None
for layer in self.layers_bwd:
layer_in_out_bwd,states = layer(layer_in_out_bwd,states)
layer_in_out = torch.cat((layer_in_out,layer_in_out_bwd),dim=1)
return self.fc_output(layer_in_out)
我在 Pytorch 论坛中找到了这个 https://discuss.pytorch.org/t/passing-hidden-layers-to-convlstm/52814 这是我在 Pytorch 中为 Bi-Directional Conv LSTM 找到的实现。在完成代码后,我被困在尝试测试它,但似乎有些函数定义错误或不完整。我在这里 https://github.com/ndrplz/ConvLSTM_pytorch 也找到了 Conv LSTM 的其他实现,但这不支持双向。我需要一些帮助来重新调整上面的代码。
ConvLSTM2D = ConvLSTM(128,128,3,1,True,0.0)
x = torch.randn([5,224,224])
t1 = ConvLSTM2D(x)
print(t1)
TypeError: initialize_weights() missing 1 required positional argument: 'layer'
版权声明:本文内容由互联网用户自发贡献,该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 dio@foxmail.com 举报,一经查实,本站将立刻删除。