# Stateful RNN example **URL:** \ **Category:** Uncategorized\ **Created:** [December 8, 2017, 5:03am UTC](https://discuss.pytorch.org/t/stateful-rnn-example/10912 "2017-12-08T05:03:52Z")\ **Posts on this page:** 9\ **Page:** 1
**Author:** ![syhpoon](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/syhpoon/32/2623_2.png) [@syhpoon](https://discuss.pytorch.org/u/syhpoon)\ **Post date:** [December 8, 2017, 5:03am UTC](https://discuss.pytorch.org/t/stateful-rnn-example/10912/1 "2017-12-08T05:03:52Z")
Hi there, I’m trying to implement a time-series prediction rnn and for this I try to construct a stateful model. Basically because I have a huge sequence I want to reuse states from previous batches instead of having them reset every time. Keras RNN class has a `stateful` parameter enabling exactly this behavior: > stateful: Boolean (default False). If True, the last state for each sample at index i in a batch will be used as initial state for the sample of index i in the following batch. I couldn’t find anything similar for pytorch and my attempts to make something like this manually failed so far. Basically what I try to do is save variable tensors after `forward` pass and use them as initial state for `h` and `c` variables for LSTMCell on the subsequent `forward` calls. I’m very new to pytorch so I’m probably doing something very wrong but so far I’m stuck. I’d really appreciate any hint or any example of stateful RNN in pytorch. ---
**Author:** ![jodles](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/j/8edcca/32.png) [@jodles](https://discuss.pytorch.org/u/jodles)\ **Post date:** [December 8, 2017, 9:13am UTC](https://discuss.pytorch.org/t/stateful-rnn-example/10912/2 "2017-12-08T09:13:02Z")
This thread may be relevant: [[Solved] Training a simple RNN](https://discuss.pytorch.org/t/solved-training-a-simple-rnn/9055) A key point is that to keep hidden state Variable across batches without having to specify `retain_graph=True` you need to detach the Variable: `hidden.detach()`. ---
**Author:** ![dgriff](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/dgriff/32/1361_2.png) [@dgriff](https://discuss.pytorch.org/u/dgriff)\ **Post date:** [December 11, 2017, 10:06am UTC](https://discuss.pytorch.org/t/stateful-rnn-example/10912/3 "2017-12-11T10:06:36Z")
to do something like this in pytorch you would just do something like: ```auto class Model(torch.nn.Module): def __init__ (self, hidden_size): super(Model, self). __init__ () self.lstm = nn.LSTMCell(1, hidden_size) def forward(self, inputs): #and then in def forward: x, (hx, cx) = inputs x = x.view(x.size(0), -1) hx, cx = self.lstm(x, (hx, cx)) x = hx return x, (hx, cx) flag=True #to flag when at start of time series if flag: #beginning of sequence of data you want cell states from or when no longer need past cell state and starting fresh again cx = Variable(torch.zeros(1, hidden_size)) hx = Variable(torch.zeros(1, hidden_size)) else: #get cell state from last sequence batch to use as start of new batch cx = Variable(cx.data) hx = Variable(hx.data) #input equals Variable of 1 by num of features at each time step model = Model(hidden_size) output, (hx, cx) = model((input, (hx, cx)) ``` ---
**Author:** ![syhpoon](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/syhpoon/32/2623_2.png) [@syhpoon](https://discuss.pytorch.org/u/syhpoon)\ **Post date:** [December 11, 2017, 4:46pm UTC](https://discuss.pytorch.org/t/stateful-rnn-example/10912/4 "2017-12-11T16:46:49Z")
Thank you! This is very helpful. ---
**Author:** ![dgriff](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/dgriff/32/1361_2.png) [@dgriff](https://discuss.pytorch.org/u/dgriff)\ **Post date:** [December 12, 2017, 7:14am UTC](https://discuss.pytorch.org/t/stateful-rnn-example/10912/5 "2017-12-12T07:14:36Z")
glad to hear it 👍 oh fyi this part: ```auto else: #get cell state from last sequence batch to use as start of new batch cx = Variable(cx.data) hx = Variable(hx.data) ``` This serves same function as the hidden.detach() part stated above. I just prefer to create a fresh new Variable for each batch than reuse the same one ---
**Author:** ![syhpoon](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/syhpoon/32/2623_2.png) [@syhpoon](https://discuss.pytorch.org/u/syhpoon)\ **Post date:** [December 21, 2017, 3:28pm UTC](https://discuss.pytorch.org/t/stateful-rnn-example/10912/6 "2017-12-21T15:28:00Z")
So here’s the model I’ve been able to make: ```python #!/usr/bin/env python import torch import torch.nn as nn import torch.optim as opt import numpy as np from torch.autograd import Variable import pandas as pd torch.manual_seed(1) num_epochs = 1 hidden_size = 100 def var(tensor): if torch.cuda.is_available(): return Variable(tensor.cuda()) else: return Variable(tensor) class Model(nn.Module): def __init__ (self, input_size=1, hidden_size=hidden_size, output_size=1, num_layers=2): super(Model, self). __init__ () self.num_layers = num_layers self.hidden_size = hidden_size self.lstm = nn.LSTM(input_size, hidden_size, num_layers=num_layers) self.linear = nn.Linear(hidden_size, hidden_size) self.relu = nn.ReLU() self.out = nn.Linear(hidden_size, output_size) def forward(self, x, hidden): lstm_out, hidden = self.lstm(x, hidden) linear_out = self.linear(lstm_out) relu_out = self.relu(linear_out) out = self.out(relu_out) return out, hidden def init_hidden(self): return (var(torch.zeros(self.num_layers, 1, self.hidden_size)), var(torch.zeros(self.num_layers, 1, self.hidden_size))) def load_data(n): ix = np.arange(n) data = np.sin(2*np.pi*ix/float(n/2)) return (torch.from_numpy(data[:-1]).float(), torch.from_numpy(data[1:]).float()) def train(ds): model = Model() # criterion = nn.MSELoss() criterion = nn.L1Loss() optimizer = opt.Adam(model.parameters(), lr=0.0002) if torch.cuda.is_available(): print("Using GPU") model.cuda() lsum = 0.0 lcount = 0.0 hidden = model.init_hidden() losses = [] for epoch in range(num_epochs): for i, (_x, _y) in enumerate(zip(ds[0], ds[1])): x = torch.zeros(1, 1, 1) y = torch.zeros(1, 1, 1) x[0][0][0] = _x y[0][0][0] = _y out, hidden = model(var(x), hidden) loss = criterion(out, var(y)) loss.backward() hidden = (var(hidden[0].data), var(hidden[1].data)) lsum += loss.data[0] lcount += 1 losses.append(loss.data[0]) optimizer.step() optimizer.zero_grad() if i % 1000 == 0: l = lsum / lcount print('Epoch: %i, Loss: %.7f' % (epoch, l)) lsum = 0.0 lcount = 0.0 if __name__ == " __main__": train(load_data(100000)) ``` It does not seem to be able to learn even the simple sine function, here’s the plot of losses: ![loss](https://discuss.pytorch.org/uploads/default/original/2X/d/d1ce0939460495ceb92b30f4d5fa51fa1cba59d3.png) The loss just seems to oscillate along with the input. I’d really appreciate any hint on what is it I’m doing wrong. ---
**Author:** ![jpeg729](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/jpeg729/32/3464_2.png) [@jpeg729](https://discuss.pytorch.org/u/jpeg729)\ **Post date:** [December 21, 2017, 3:55pm UTC](https://discuss.pytorch.org/t/stateful-rnn-example/10912/7 "2017-12-21T15:55:59Z")
Won’t detaching the hidden Variable cut off BPTT? ---
**Author:** ![jodles](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/j/8edcca/32.png) [@jodles](https://discuss.pytorch.org/u/jodles)\ **Post date:** [December 21, 2017, 4:37pm UTC](https://discuss.pytorch.org/t/stateful-rnn-example/10912/8 "2017-12-21T16:37:17Z")
Yes, but you’d do it between minibatches, like `Variable(hx.data)` as in dgriff’s example above and serves the same purpose. ---
**Author:** ![Jianmo\_Ni](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/jianmo_ni/32/1770_2.png) [@Jianmo\_Ni](https://discuss.pytorch.org/u/Jianmo_Ni)\ **Post date:** [April 9, 2018, 1:17am UTC](https://discuss.pytorch.org/t/stateful-rnn-example/10912/9 "2018-04-09T01:17:34Z")
Hi, have you been able to figure out the reason why it did not work?