I need some help. This is error!
File "F:/研究生资料/论文/注意力机制/DA-RNN-master-pytorch/DA-RNN-master/src/main.py", line 55, in
model.train()

File "F:\研究生资料\论文\注意力机制\DA-RNN-master-pytorch\DA-RNN-master\src\model.py", line 285, in train
loss = self.criterion(y_pred, y_true)

File "G:\Users\ll\Anaconda3\envs\tensorflow_gpu\lib\site-packages\torch\nn\modules\module.py", line 491, in call
result = self.forward(*input, **kwargs)

File "G:\Users\ll\Anaconda3\envs\tensorflow_gpu\lib\site-packages\torch\nn\modules\loss.py", line 372, in forward
return F.mse_loss(input, target, size_average=self.size_average, reduce=self.reduce)

File "G:\Users\ll\Anaconda3\envs\tensorflow_gpu\lib\site-packages\torch\nn\functional.py", line 1569, in mse_loss
input, target, size_average, reduce)

File "G:\Users\ll\Anaconda3\envs\tensorflow_gpu\lib\site-packages\torch\nn\functional.py", line 1537, in _pointwise_loss
return lambd_optimized(input, target, size_average, reduce)

RuntimeError: input and target shapes do not match: input [128 x 1], target [128] at c:\users\administrator\downloads\new-builder\win-wheel\pytorch\aten\src\thnn\generic/MSECriterion.c:13

your code in utils.py read_data(), the code of getting data is below, why do you use companies' stock price to predict NASDAQ-100 Index?
X = df.loc[:, [x for x in df.columns.tolist() if x != 'NDX']].as_matrix()
y = np.array(df.NDX)

NDX should be calculated by these stock prices, isn’t it? why u have to learn the calculation formula by RNN?
The DA-RNN paper gives a time series predicting model, right? But where is your time series predicting? I am confusion.

That's what I found when I read the code repeatedly, If I got wrong or missed something, please tell me.
Thank you.

I found that the Eq 8 in the paper is different from the coding function in your program. The tanh function is not used, but your decoding function is consistent with the paper, Thanks!!!!!

Thank you for your contribution, but I can't call GPU when I run this code. Here is my configuration.
Ubuntu18.04, CUDA9.0, cudnn7.0, pytorch0.3.0.

The 'input_weighted' variable which is the list of attention weighted input data, present in 'train_forward' method of main.py doesnt seem to used in the decoder.

Hi,
I notice there is no eval() or torch.no_grad() called when evaluating which will cause the model to run in a wrong way.
I think this is a bug.

Hello, Zhenye-Na
I'm using pandas 0.25.1 version
I got error "Dataframe has no attribute 'as_matrix' "
I think as_matrix() to is changed to to_numpy() at new pandas version

Hello, author.

The input data includes drving series and previous values of the target series according to this paper. I am a little puzzeld about the drving series $X$ When I prepare to apply this model in my data. My data only have the previous values of the target series, How can I ready for the drving series ?

Please help me crack my quandary at your nearest free time.

Thanks !!!

Dear Zhenye,
Could you please send me an email (should be visible on my profile). I compiled a detailed email to share with you, then realized your email wasn't public.
thanks,
Daniel

Hi, Zhenye, thank you for excellent code. When i ran the code, it seemed that i got the wrong prediction result.
Could you please give some advice on how to overcome it? Thanks a lot.

Thanks for the awesome code, I'm still learning DARNN, and a little bit confused in your code.

I've noticed that in training session, your y_prev is defined as

for bs in range(len(indices)):
y_prev[bs, :] = self.y[indices[bs]:(indices[bs] + self.T - 1)]

but y_gt is

The question is,

If batchsize is 1, T is 10,

You get y[0:9](the 1st to the 9th number, let's say [0, 1, 2, 3, 4, 5, 6, 7, 8]) as input but predict y[10] as target(which is the 11th number, let's say [10] ).

Doesn't that means the code skips the 10th value as target? (value [9] )

Or I'm just wrong in numpy indexing.

hi~ i could not run the main.py,there is something wrong : from model import *
can you tell me how to solve it?
thank u ~

This is my data which I repaced in ../nasdaq/data.csv directory.
I need to train model for price detection, this is how my data look like:

For CPU, I run the python3 main.py --lr 0.0001 --epochs 50, but I got this error:

File "/content/drive/My Drive/Spot-Forcasting/paper-methods2019/DA-RNN/DA-RNN/src/model.py", line 77, in forward
x = torch.cat((h_n.repeat(self.input_size, 1, 1).permute(1, 0, 2),
RuntimeError: invalid argument 3: out of range at /pytorch/aten/src/TH/generic/THTensor.cpp:392

Any idea how to solve it?
Thanks

When I found this repository, the author has already removed code files due to the error, So I wonder when will the author release it again?

I found that the Eq 8 in the paper is different from the coding function in your program. The tanh function is not used, but your decoding function is consistent with the paper, Thanks!!!!!

When I ran the code, there was a bug:''LSTM' object has no attribute 'flatten_parameters'',
Here is the code:
self.encoder_lstm.flatten_parameters()
_, final_stage = self.encoder_lstm(
x_tilde.unsqueeze(0), (h_n, s_n))
h_n = final_state[0]
s_n = final_stage[1]

I am not sure if this bug happened due to the naming issue

Are you able to reproduce the paper's result?

How can we reproduce figure 3 of the original paper using this code?

I tried da-rnn for multi-step forcasting but got very bad performance. I cut the dataset by the setting step length. And the strategy for multi-step is Multiple-Output. Is the attention mechanism not good for multi-step forcasting?

Pytorch has a built in method which sets the model into evaluation mode. The method is called torch.nn.Module.eval(). In this method the train method is called with a parameter. Due to the case the your rnn also implements a train method python overwrites the default train method and the code will crash.
How to fix:
I recommend to rename your method into fit() or train_epochs() or something similiear.

Regardless of some minor issues, your model seems nice.

It seems the train_val_test_split function in ops.py file is never called. The model is trained on all data points, and the training loss is reported.

Shouldn't the input to the encoder be from 1 to T instead of T-1?
As the number LSTM in stack for encoder is to T and decoder is 1 to T- 1,as per my understanding of the reference paper.
Thank you and hope we can discuss and resolve this issue

I run the program as follows, but it seems cuda is not enabled.
python3 main.py --cuda

So, I edited the part of source codes as following,
model = DA_rnn(X, y, opt.ntimestep, opt.nhidden_encoder, opt.nhidden_decoder, opt.batchsize, opt.lr, opt.epochs, opt.cuda)
I added "opt.cuda" parameter.

However, there are still problems like this:

/home/tatsuhiro/PycharmProjects/untitled/DA-RNN-master/src/model.py:80: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.
alpha = F.softmax(x.view(-1, self.input_size))
/home/tatsuhiro/PycharmProjects/untitled/DA-RNN-master/src/model.py:145: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.
x.view(-1, 2 * self.decoder_num_hidden + self.encoder_num_hidden)).view(-1, self.T - 1))
Traceback (most recent call last):
File "main.py", line 57, in
model.train()
File "/home/tatsuhiro/PycharmProjects/untitled/DA-RNN-master/src/model.py", line 260, in train
loss = self.train_forward(x, y_prev, y_gt)
File "/home/tatsuhiro/PycharmProjects/untitled/DA-RNN-master/src/model.py", line 316, in train_forward
loss = self.criterion(y_pred, y_true)
File "/home/tatsuhiro/PycharmProjects/untitled/venv/lib/python3.5/site-packages/torch/nn/modules/module.py", line 491, in call
result = self.forward(*input, **kwargs)
File "/home/tatsuhiro/PycharmProjects/untitled/venv/lib/python3.5/site-packages/torch/nn/modules/loss.py", line 372, in forward
return F.mse_loss(input, target, size_average=self.size_average, reduce=self.reduce)
File "/home/tatsuhiro/PycharmProjects/untitled/venv/lib/python3.5/site-packages/torch/nn/functional.py", line 1569, in mse_loss
input, target, size_average, reduce)
File "/home/tatsuhiro/PycharmProjects/untitled/venv/lib/python3.5/site-packages/torch/nn/functional.py", line 1537, in _pointwise_loss
return lambd_optimized(input, target, size_average, reduce)
RuntimeError: Expected object of type torch.cuda.FloatTensor but found type torch.FloatTensor for argument #2 'target'

How can I solve this problem?