CODEWITHSUNDEEP
pythonmachine-learningkeraslstmrecurrent-neural-network
A.Abs
A.Abs

Reputation: 470

Many-to-many lstm model on varying samples

I have recently started learning how to build LSTM model for multivariate time series data. I have looked here and here on how to pad sequences and implement many-to-many LSTM model. I have created a dataframe to test the model but I keep getting an error (below). 

d = {'ID':['a12', 'a12','a12','a12','a12','b33','b33','b33','b33','v55','v55','v55','v55','v55','v55'], 'Exp_A':[2.2,2.2,2.2,2.2,2.2,3.1,3.1,3.1,3.1,1.5,1.5,1.5,1.5,1.5,1.5], 
     'Exp_B':[2.4,2.4,2.4,2.4,2.4,1.2,1.2,1.2,1.2,1.5,1.5,1.5,1.5,1.5,1.5], 
     'A':[0,0,1,0,1,0,1,0,1,0,1,1,1,0,1], 'B':[0,0,1,1,1,0,0,1,1,1,0,0,1,0,1],
     'Time_Interval': ['11:00:00', '11:10:00', '11:20:00', '11:30:00', '11:40:00',
                '11:00:00', '11:10:00', '11:20:00', '11:30:00',
                '11:00:00', '11:10:00', '11:20:00', '11:30:00', '11:40:00', '11:50:00']}

df = pd.DataFrame(d)
df.set_index('Time_Interval', inplace=True)

I tried to pad using brute force:

from keras.preprocessing.sequence import pad_sequences

x1 = df['A'][df['ID']== 'a12']
x2 = df['A'][df['ID']== 'b33']
x3 = df['A'][df['ID']== 'v55']

mx = df['ID'].size().max() # Find the largest group
seq1 = [x1, x2, x3]
padded1 = np.array(pad_sequences(seq1, maxlen=6, dtype='float32')).reshape(-1,mx,1)

In similar ways I have created padded2, padded3 and padded4 for each feature: 

padded_data = np.dstack((padded1, padded1, padded3, padded4))
padded_data.shape = (3, 6, 4)

padded_data

array([[[0. , 0. , 0. , 0. ],
        [0. , 0. , 2.2, 2.4],
        [0. , 0. , 2.2, 2.4],
        [1. , 1. , 2.2, 2.4],
        [0. , 0. , 2.2, 2.4],
        [1. , 1. , 2.2, 2.4]],

       [[0. , 0. , 0. , 0. ],
        [0. , 0. , 0. , 0. ],
        [0. , 0. , 3.1, 1.2],
        [1. , 1. , 3.1, 1.2],
        [0. , 0. , 3.1, 1.2],
        [1. , 1. , 3.1, 1.2]],

       [[0. , 0. , 1.5, 1.5],
        [1. , 1. , 1.5, 1.5],
        [1. , 1. , 1.5, 1.5],
        [1. , 1. , 1.5, 1.5],
        [0. , 0. , 1.5, 1.5],
        [1. , 1. , 1.5, 1.5]]], dtype=float32)

edit

#split into train/test
train = pad_1[:2]   # train on the 1st two samples.
test = pad_1[-1:]    

train_X = train[:,:-1]  # one step ahead prediction.
train_y = train[:,1:]

test_X = test[:,:-1]  # test on the last sample
test_y = test[:,1:]
# check shapes
print(train_X.shape, train_y.shape, test_X.shape, test_y.shape)
#(2, 5, 4) (2, 5, 4) (1, 5, 4) (1, 5, 4)

# design network
model = Sequential()
model.add(Masking(mask_value=0., input_shape=(train.shape[1], train.shape[2])))
model.add(LSTM(32, input_shape=(train.shape[1], train.shape[2]), return_sequences=True))
model.add(Dense(4))
model.compile(loss='mae', optimizer='adam', metrics=['accuracy'])
model.summary()

# fit network
history = model.fit(train, test, epochs=300, validation_data=(test_X, test_y), verbose=2, shuffle=False)

[![enter image description here][3]][3]

So my questions are:

  1. Surely, there must be an efficient way of transforming the data?
  2. Say I want a single time-step prediction for future sequence, I have

first time-step = array([[[0.5 , 0.9 , 2.5, 3.5]]], dtype=float32) Where first time-step is a single 'frame' of a sequence. How do adjust the model to incorporate this?

Upvotes: 1

Views: 489

Answers (1)

today
today

Reputation: 33410

To resolve the error, remove return_sequence=True from the LSTM layer arguments (since with this architecture you have defined, you only need the output of last layer) and also simply use train[:, -1] and test[:, -1] (instead of train[:, -1:] and test[:, -1:]) to extract the labels (i.e. removing : causes the second axis to be dropped and therefore makes the labels shape consistent with the output shape of the model).

As a side note, wrapping a Dense layer inside a TimeDistributed layer is redundant, since the Dense layer is applied on the last axis.

Update: As for the new question, either pad the input sequence which has only one timestep to make it have a shape of (5,4), or alternatively set the input shape of the first layer (i.e. Masking) to input_shape=(None, train.shape[2]) so the model can work with inputs of varying length.

Upvotes: 1

Related Questions