If I may, here’s my take on this topic according to the paper by Shi et al. (2015).
My approach was to implement a convolutional LSTM cell first (to similar fashion that you did), which will then be utilized in a complete model. In other words, these are the inner workings of the cell. The recurrent operations (looping, passing states to subsequent steps etc.) should be handled in a separate ConvLSTM class and its forward function.
Here’s the simple source the class:
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(BaseModule):
"""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,
'out_channels': self.kernels,
'kernel_size': self.kernel_size,
'padding': self.padding,
'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)
A simple example on how the cell is utilized is as follows:
cell = ConvLSTMCell(
input_bands=3,
input_dim=16,
)
batch = torch.ones((1,3,16,16))
H,C = cell(batch)
print(f"H:{H.shape}")
print(f"C:{H.shape}")
>>> H:torch.Size([1, 32, 16, 16])
>>> C:torch.Size([1, 32, 16, 16])
Then to utilize the ConvLSTMCell in a complete LSTM module, you could use a model like this for example:
class ConvLSTM(nn.Module):
def __init__(self, input_bands, input_dim, kernels, 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)
Of course these are freely editable to better fir your purposes. The above ConvLSTM for example produces a many-to-one prediction, ingesting sequences and producing single values using only the last outputs of the sequential model (or models in the case of bidirectionality).
I sincerely hope this helps you and others tackling this topic!
Cheers,
Petteri