GRU¶
- class torchrl.modules.GRU(input_size: int, hidden_size: int, num_layers: int = 1, bias: bool = True, batch_first: bool = True, dropout: float = 0.0, bidirectional: bool = False, device=None, dtype=None)[source]¶
A PyTorch module for executing multiple steps of a multi-layer GRU. The module behaves exactly like
torch.nn.GRU
, but this implementation is exclusively coded in Python.Note
This class is implemented without relying on CuDNN, which makes it compatible with
torch.vmap()
andtorch.compile()
.Examples
>>> import torch >>> from torchrl.modules.tensordict_module.rnn import GRU
>>> device = torch.device("cuda") if torch.cuda.device_count() else torch.device("cpu") >>> B = 2 >>> T = 4 >>> N_IN = 10 >>> N_OUT = 20 >>> N_LAYERS = 2 >>> V = 4 # vector size >>> gru = GRU( ... input_size=N_IN, ... hidden_size=N_OUT, ... device=device, ... num_layers=N_LAYERS, ... )
# single call >>> x = torch.randn(B, T, N_IN, device=device) >>> h0 = torch.zeros(N_LAYERS, B, N_OUT, device=device) >>> with torch.no_grad(): … h1 = gru(x, h0)
# vectorised call - not possible with nn.GRU >>> def call_gru(x, h): … h_out = gru(x, h) … return h_out >>> batched_call = torch.vmap(call_gru) >>> x = torch.randn(V, B, T, 10, device=device) >>> h0 = torch.zeros(V, N_LAYERS, B, N_OUT, device=device) >>> with torch.no_grad(): … h1 = batched_call(x, h0)
__init__(input_size,hidden_size,num_layers=1,bias=True,batch_first=False,dropout=0.0,bidirectional=False,device=None,dtype=None)
Apply a multi-layer gated recurrent unit (GRU) RNN to an input sequence. For each element in the input sequence, each layer computes the following function:
\[\begin{split}\begin{array}{ll} r_t = \sigma(W_{ir} x_t + b_{ir} + W_{hr} h_{(t-1)} + b_{hr}) \\ z_t = \sigma(W_{iz} x_t + b_{iz} + W_{hz} h_{(t-1)} + b_{hz}) \\ n_t = \tanh(W_{in} x_t + b_{in} + r_t \odot (W_{hn} h_{(t-1)}+ b_{hn})) \\ h_t = (1 - z_t) \odot n_t + z_t \odot h_{(t-1)} \end{array}\end{split}\]where \(h_t\) is the hidden state at time t, \(x_t\) is the input at time t, \(h_{(t-1)}\) is the hidden state of the layer at time t-1 or the initial hidden state at time 0, and \(r_t\), \(z_t\), \(n_t\) are the reset, update, and new gates, respectively. \(\sigma\) is the sigmoid function, and \(\odot\) is the Hadamard product.
In a multilayer GRU, the input \(x^{(l)}_t\) of the \(l\) -th layer (\(l \ge 2\)) is the hidden state \(h^{(l-1)}_t\) of the previous layer multiplied by dropout \(\delta^{(l-1)}_t\) where each \(\delta^{(l-1)}_t\) is a Bernoulli random variable which is \(0\) with probability
dropout
.- Parameters:
input_size – The number of expected features in the input x
hidden_size – The number of features in the hidden state h
num_layers – Number of recurrent layers. E.g., setting
num_layers=2
would mean stacking two GRUs together to form a stacked GRU, with the second GRU taking in outputs of the first GRU and computing the final results. Default: 1bias – If
False
, then the layer does not use bias weights b_ih and b_hh. Default:True
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
dropout – If non-zero, introduces a Dropout layer on the outputs of each GRU layer except the last layer, with dropout probability equal to
dropout
. Default: 0bidirectional – If
True
, becomes a bidirectional GRU. Default:False
- Inputs: input, h_0
input: tensor of shape \((L, H_{in})\) for unbatched input, \((L, N, H_{in})\) when
batch_first=False
or \((N, L, H_{in})\) whenbatch_first=True
containing the features of the input sequence. The input can also be a packed variable length sequence. Seetorch.nn.utils.rnn.pack_padded_sequence()
ortorch.nn.utils.rnn.pack_sequence()
for details.h_0: tensor of shape \((D * \text{num\_layers}, H_{out})\) or \((D * \text{num\_layers}, N, H_{out})\) containing the initial hidden state for the input sequence. Defaults to zeros if not provided.
where:
\[\begin{split}\begin{aligned} N ={} & \text{batch size} \\ L ={} & \text{sequence length} \\ D ={} & 2 \text{ if bidirectional=True otherwise } 1 \\ H_{in} ={} & \text{input\_size} \\ H_{out} ={} & \text{hidden\_size} \end{aligned}\end{split}\]- Outputs: output, h_n
output: tensor of shape \((L, D * H_{out})\) for unbatched input, \((L, N, D * H_{out})\) when
batch_first=False
or \((N, L, D * H_{out})\) whenbatch_first=True
containing the output features (h_t) from the last layer of the GRU, for each t. If atorch.nn.utils.rnn.PackedSequence
has been given as the input, the output will also be a packed sequence.h_n: tensor of shape \((D * \text{num\_layers}, H_{out})\) or \((D * \text{num\_layers}, N, H_{out})\) containing the final hidden state for the input sequence.
- Variables:
weight_ih_l[k] – the learnable input-hidden weights of the \(\text{k}^{th}\) layer (W_ir|W_iz|W_in), of shape (3*hidden_size, input_size) for k = 0. Otherwise, the shape is (3*hidden_size, num_directions * hidden_size)
weight_hh_l[k] – the learnable hidden-hidden weights of the \(\text{k}^{th}\) layer (W_hr|W_hz|W_hn), of shape (3*hidden_size, hidden_size)
bias_ih_l[k] – the learnable input-hidden bias of the \(\text{k}^{th}\) layer (b_ir|b_iz|b_in), of shape (3*hidden_size)
bias_hh_l[k] – the learnable hidden-hidden bias of the \(\text{k}^{th}\) layer (b_hr|b_hz|b_hn), of shape (3*hidden_size)
Note
All the weights and biases are initialized from \(\mathcal{U}(-\sqrt{k}, \sqrt{k})\) where \(k = \frac{1}{\text{hidden\_size}}\)
Note
For bidirectional GRUs, forward and backward are directions 0 and 1 respectively. Example of splitting the output layers when
batch_first=False
:output.view(seq_len, batch, num_directions, hidden_size)
.Note
batch_first
argument is ignored for unbatched inputs.Note
The calculation of new gate \(n_t\) subtly differs from the original paper and other frameworks. In the original implementation, the Hadamard product \((\odot)\) between \(r_t\) and the previous hidden state \(h_{(t-1)}\) is done before the multiplication with the weight matrix W and addition of bias:
\[\begin{aligned} n_t = \tanh(W_{in} x_t + b_{in} + W_{hn} ( r_t \odot h_{(t-1)} ) + b_{hn}) \end{aligned}\]This is in contrast to PyTorch implementation, which is done after \(W_{hn} h_{(t-1)}\)
\[\begin{aligned} n_t = \tanh(W_{in} x_t + b_{in} + r_t \odot (W_{hn} h_{(t-1)}+ b_{hn})) \end{aligned}\]This implementation differs on purpose for efficiency.
Examples:
>>> rnn = nn.GRU(10, 20, 2) >>> input = torch.randn(5, 3, 10) >>> h0 = torch.randn(2, 3, 20) >>> output, hn = rnn(input, h0)
- forward(input, hx=None)[source]¶
Define the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Module
instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.