python - How exactly does Keras take dimension argumentsfor LSTM / time series problems? -

i can't seem find concrete answer question of how feed data keras. examples seem work off image / text data , have defined data points.

i'm trying feed music lstm neural network. want network take ~3 seconds of music , nominate next 2 seconds. have music prepared .wav files , partitioned 5 second intervals i've decomposed x (first 3 seconds) , y (last 2 seconds). i've sampled music @ 44,100 hz x 132,300 observations 'long' , y '88,200' observations long.

but can't figure out how bridge keras data structure. i'm using tensorflow backend.

in interest of generalizing problem , answer, i'll use a,b,c denote dimensions. difference between example data , real data these random values distributed 0 1, , data array of integers.

import numpy np #using variables make easy generalize answer  #a = number of observations have       = 411  #b = duration of sample, 44.1k observations per second of music b_train = 132300 b_test  = 88200  #c = number of channels in music, 2 channel stereo c       = 2  #now create sample data dimensionality given above: x = np.random.rand(a,b_train,c) y = np.random.rand(a,b_test ,c)  #split data sklearn.model_selection import train_test_split  x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=.20, random_state=42)     

however, don't know how configure model understand 'first' (a) dimension contains observations , want more or less break out music (b) channel (c).

i know it'd easier convert mono (and 2d problem) i'm curious see whether or not has 'simple' solution - whether takes shape of have below or whether should think of model in way.

the primary question this: how construct model allow me transform x data y data?

ideally, answer show how modify model below fit data structure above.

import keras import math, time keras.models import sequential keras.layers.core import dense, dropout, activation keras.layers.recurrent import lstm keras.models import load_model  def build_model(layers):     d = 0.3     model = sequential()      model.add(lstm(256, input_shape=(layers), return_sequences=true))     model.add(dropout(d))      model.add(lstm(256, input_shape=(layers), return_sequences=false))     model.add(dropout(d))      model.add(dense(32,kernel_initializer="uniform",activation='relu'))             model.add(dense(1,kernel_initializer="uniform",activation='linear'))       start = time.time()     model.compile(loss='mse',optimizer='adam', metrics=['accuracy'])     print("compilation time : ", time.time() - start)     return model   #build model...  model = build_model([328,132300,2])  model.fit(x_train,y_train,batch_size=512,epochs=30,validation_split=0.1,verbose=1) 

however, yields error (at model = ... step):

 valueerror: input 0 incompatible layer lstm_2: expected ndim=3, found ndim=4 

i can't figure out keras gets expectation see ndim=4 data. also, don't know how ensure feed data model such model 'understands' observations distributed across a-axis , data distributed on b- , c-axis.

if unclear, please leave comment. i'll watch diligently until sept '17 or , i'll sure update question reflect advice / comments left.

thanks!

keras convention batch dimension typically omitted in input_shape arguments. guide:

pass input_shape argument first layer. shape tuple (a tuple of integers or none entries, none indicates positive integer may expected). in input_shape, batch dimension not included.

so changing model = build_model([132300,2]) should solve problem.