TensorFlow dynamic_rnn regressor: ValueError dimension mismatch

Question

I would like to build a toy LSTM model for regression. This nice tutorial is already too complicated for a beginner.

Given a sequence of length time_steps, predict the next value. Consider time_steps=3 and the sequences:

array([
   [[  1.],
    [  2.],
    [  3.]],

   [[  2.],
    [  3.],
    [  4.]],
    ...

the target values should be:

array([  4.,   5., ...

I define the following model:

# Network Parameters
time_steps = 3 
num_neurons= 64 #(arbitrary)
n_features = 1

# tf Graph input
x = tf.placeholder("float", [None, time_steps, n_features])
y = tf.placeholder("float", [None, 1])

# Define weights
weights = {
   'out': tf.Variable(tf.random_normal([n_hidden, 1]))
} 
biases = {
   'out': tf.Variable(tf.random_normal([1]))
}

#LSTM model
def lstm_model(X, weights, biases, learning_rate=0.01, optimizer='Adagrad'):

  # Prepare data shape to match `rnn` function requirements
  # Current data input shape: (batch_size, time_steps, n_features)
  # Required shape: 'time_steps' tensors list of shape (batch_size, n_features)
  # Permuting batch_size and time_steps
  input dimension: Tensor("Placeholder_:0", shape=(?, 3, 1), dtype=float32)

  X = tf.transpose(X, [1, 0, 2])
  transposed dimension: Tensor("transpose_41:0", shape=(3, ?, 1), dtype=float32)

  # Reshaping to (time_steps*batch_size, n_features)
  X = tf.reshape(X, [-1, n_features])
  reshaped dimension: Tensor("Reshape_:0", shape=(?, 1), dtype=float32)

  # Split to get a list of 'time_steps' tensors of shape (batch_size, n_features)
  X = tf.split(0, time_steps, X)
  splitted dimension: [<tf.Tensor 'split_:0' shape=(?, 1) dtype=float32>, <tf.Tensor 'split_:1' shape=(?, 1) dtype=float32>, <tf.Tensor 'split_:2' shape=(?, 1) dtype=float32>]

  # LSTM cell
  cell = tf.nn.rnn_cell.LSTMCell(num_neurons) #Or GRUCell(num_neurons)

  output, state = tf.nn.dynamic_rnn(cell=cell, inputs=X, dtype=tf.float32)

  output = tf.transpose(output, [1, 0, 2])
  last = tf.gather(output, int(output.get_shape()[0]) - 1)


  return tf.matmul(last, weights['out']) + biases['out']

We instantiating the LSTM model with pred = lstm_model(x, weights, biases) I get the following:

---> output, state = tf.nn.dynamic_rnn(cell=cell, inputs=X, dtype=tf.float32)
ValueError: Dimension must be 2 but is 3 for 'transpose_42' (op: 'Transpose') with input shapes: [?,1], [3]

1) Do you know what the problem is?

2) Will multiplying the LSTM output by the weights yield the regression?

Answer 1

As discussed in the comments, the tf.nn.dynamic_rnn(cell, inputs, ...) function expects a list of three-dimensional tensors^* as its inputs argument, where the dimensions are interpreted by default as batch_size x num_timesteps x num_features. (If you pass time_major=True, they are interpreted as num_timesteps x batch_size x num_features.) Therefore the preprocessing you've done in the original placeholder is unnecessary, and you can pass the oriding X value directly to tf.nn.dynamic_rnn().

---

^* Technically it can accept complicated nested structures in addition to lists, but the leaf elements must be three-dimensional tensors.^**

^** Investigating this turned up a bug in the implementation of tf.nn.dynamic_rnn(). In principle, it should be sufficient for the inputs to have at least two dimensions, but the time_major=False path assumes that they have exactly three dimensions when it transposes the input into the time-major form, and it was the error message that this bug inadvertently causes that showed up in your program. We're working on getting that fixed.