Is there any range limitations of 'timesteps' in the LSTM input?

[Yingyingzhang15]

Hi, I have a question about the input of LSTM layer.
The Keras Document says that the input data should be 3D tensor with shape (nb_samples, timesteps, input_dim). In my LSTM model, I set the timesteps as 50, but the model doesn't perform well. Is there any range limitations of 'timesteps'? Like smaller than a number ?
Can LSTM layer can use all the time information of each sequence?
Thank you in advance!

[viksit]

What kind of data do you have, and could you shed some more light on what isn't working well?

AFAIK, there's no "limitation" on the timestep number.

[fchollet]

No limitations, but for the sake of performance it's better to keep it
fairly small (aim at <200).

On 23 March 2016 at 19:08, Viksit Gaur notifications@github.com wrote:

What kind of data do you have, and could you shed some more light on what
isn't working well?

AFAIK, there's no "limitation" on the timestep number.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub
#2057 (comment)

[Yingyingzhang15]

@viksit Thank you for your answer! My training data size is (45000,50,3), where 45000 is the nb_sample, 50 is timesteps and 3 is the input_dim. There are 3 classes to classify and I want to give every 50,3 vector a label. My validation data size is (7000,50,3).

Since I'm new to Keras, I just tried the two layer model and get the training accuracy 96%, validation accuracy 80%. I hope to get the accuracy of 95%. So should I just increase the number of layers?

There are many parameters(batch_size, epoch,dropout...) , it takes a lot of time to try different parameters. Could you give me some good advice, please?

My models:

model = Sequential()
model.add(LSTM(150, input_shape=(tsteps, 3), dropout_W=0.5, dropout_U=0.1))
model.add(Dense(3))
model.add(Activation('softmax'))
model.compile(loss='binary_crossentropy', optimizer='adam')

Thank you!

[Yingyingzhang15]

@fchollet Thank you !