Theano - LSTMの '次元数が正しくありません'エラー

Question

私自身のデータのためにthis LSTMの例を再作成しようとしています。

Traceback (most recent call last): 
    File "lstm.py", line 124, in <module> 
    train_rnn(train_data) 
    File "lstm.py", line 120, in train_rnn 
    train_cost = learn_rnn_fn(i, o) 
    File "/usr/local/lib/python3.5/site-packages/theano/compile/function_module.py", line 788, in __call__ 
    allow_downcast=s.allow_downcast) 
    File "/usr/local/lib/python3.5/site-packages/theano/tensor/type.py", line 178, in filter 
    data.shape)) 
TypeError: ('Bad input argument to theano function with name "lstm.py:108" at index 0 (0-based)', 'Wrong number of dimensions: expected 2, got 0 with shape().') 

次のように私のコードが提供される：

import numpy as np 
import theano 
import theano.tensor as T 
import pandas 

dtype=theano.config.floatX 

def create_dataset(dataset, look_back=1): 
    data = [] 
    for dx in range(len(dataset) - look_back - 1): 
     data.append([dataset[dx], dataset[dx + 1]]) 
    return np.array(data, dtype=dtype) 

raw_data = pandas.read_csv('international-airline-passengers.csv', usecols=[1]) 

train_data = create_dataset(raw_data.as_matrix()[:,0]) 

train_dataは、この変換後の2-Dのnumpyの行列となります。

sigma = lambda x: 1/(1 + T.exp(-x)) 
act = T.tanh 

def one_lstm_step(x_t, h_tm1, c_tm1, W_xi, W_hi, W_ci, b_i, W_xf, W_hf, W_cf, b_f, W_xc, W_hc, b_c, W_xy, W_ho, W_cy, b_o, W_hy, b_y): 
    i_t = sigma(theano.dot(x_t, W_xi) + theano.dot(h_tm1, W_hi) + theano.dot(c_tm1, W_ci) + b_i) 
    f_t = sigma(theano.dot(x_t, W_xf) + theano.dot(h_tm1, W_hf) + theano.dot(c_tm1, W_cf) + b_f) 
    c_t = f_t * c_tm1 + i_t * act(theano.dot(x_t, W_xc) + theano.dot(h_tm1, W_hc) + b_c) 
    o_t = sigma(theano.dot(x_t, W_xo)+ theano.dot(h_tm1, W_ho) + theano.dot(c_t, W_co) + b_o) 
    h_t = o_t * act(c_t) 
    y_t = sigma(theano.dot(h_t, W_hy) + b_y) 
    return [h_t, c_t, y_t] 

def sample_weights(sizeX, sizeY): 
    values = np.ndarray([sizeX, sizeY], dtype=dtype) 
    for dx in range(sizeX): 
     vals = np.random.uniform(low=-1., high=1., size=(sizeY,)) 
     values[dx,:] = vals 
    _,svs,_ = np.linalg.svd(values)      
    values = values/svs[0] 
    return values 

n_in = 1 
n_hidden = n_i = n_c = n_o = n_f = 10 
n_y = 1 

W_xi = theano.shared(sample_weights(n_in, n_i)) 
W_hi = theano.shared(sample_weights(n_hidden, n_i)) 
W_ci = theano.shared(sample_weights(n_c, n_i)) 
b_i = theano.shared(np.cast[dtype](np.random.uniform(-0.5,.5,size = n_i))) 
W_xf = theano.shared(sample_weights(n_in, n_f)) 
W_hf = theano.shared(sample_weights(n_hidden, n_f)) 
W_cf = theano.shared(sample_weights(n_c, n_f)) 
b_f = theano.shared(np.cast[dtype](np.random.uniform(0, 1.,size = n_f))) 
W_xc = theano.shared(sample_weights(n_in, n_c)) 
W_hc = theano.shared(sample_weights(n_hidden, n_c)) 
b_c = theano.shared(np.zeros(n_c, dtype=dtype)) 
W_xo = theano.shared(sample_weights(n_in, n_o)) 
W_ho = theano.shared(sample_weights(n_hidden, n_o)) 
W_co = theano.shared(sample_weights(n_c, n_o)) 
b_o = theano.shared(np.cast[dtype](np.random.uniform(-0.5,.5,size = n_o))) 
W_hy = theano.shared(sample_weights(n_hidden, n_y)) 
b_y = theano.shared(np.zeros(n_y, dtype=dtype)) 
c0 = theano.shared(np.zeros(n_hidden, dtype=dtype)) 
h0 = T.tanh(c0) 

params = [W_xi, W_hi, W_ci, b_i, W_xf, W_hf, W_cf, b_f, W_xc, W_hc, b_c, W_xo, W_ho, W_co, b_o, W_hy, b_y, c0] 
v = T.matrix(dtype=dtype) 
target = T.matrix(dtype=dtype) 

[h_vals, _, y_vals], _ = theano.scan(fn=one_lstm_step, 
            sequences = dict(input=v, taps=[0]), 
            outputs_info = [h0, c0, None ], # corresponds to return type of fn 
            non_sequences = [W_xi, W_hi, W_ci, b_i, W_xf, W_hf, W_cf, b_f, W_xc, W_hc, b_c, W_xo, W_ho, W_co, b_o, W_hy, b_y]) 


cost = -T.mean(target * T.log(y_vals)+ (1.- target) * T.log(1. - y_vals)) 

updates=[] 

learn_rnn_fn = theano.function(inputs = [v, target], 
           outputs = cost, 
           updates = updates) 

nb_epochs=1 
train_errors = np.ndarray(nb_epochs) 

def train_rnn(train_data): 
    for x in range(nb_epochs): 
     error = 0. 
     print(train_data) 
     for j in range(len(train_data)): 
      index = np.random.randint(0, len(train_data)) 
      i, o = train_data[index] 
      train_cost = learn_rnn_fn(i, o) 
      error += train_cost 
     train_errors[x] = error 

train_rnn(train_data) 

デバッグ変数iとoの形状が適切でないことを示しています。私はデータの形を変えようとしましたが、他のデータ型の問題につながります。

Answer 1

関数create_datasetはnumpyの配列を1つ返しています。ただし、i、o = train_data [index]を呼び出すと、2つの値を取得しようとしています。たとえば、一時変数に値を代入し、必要に応じて分割することができます。

EDIT 変数iとoは機能learn_rnn_fnが期待する同じタイプではありませんでした。それは、貧弱な行列を期待していた。