2
votes

For example, the RNN is a dynamic 3-layer bidirectional LSTM with the hidden vector size of 200 (tf.nn.bidirectional_dynamic_rnn) and I have 4 GPUs to train the model. I saw a post using data parallelism on subsets of samples in a batch but that didn't speed up the training process.

1

1 Answers

1
votes

You can also try model parallelism. One way to do this is to make a cell wrapper like this, which will create cells on a specific device:

class DeviceCellWrapper(tf.nn.rnn_cell.RNNCell):
  def __init__(self, cell, device):
    self._cell = cell
    self._device = device

  @property
  def state_size(self):
    return self._cell.state_size

  @property
  def output_size(self):
    return self._cell.output_size

  def __call__(self, inputs, state, scope=None):
    with tf.device(self._device):
      return self._cell(inputs, state, scope)

Then place each individual layer onto dedicated GPU:

cell_fw = DeviceCellWrapper(cell=tf.nn.rnn_cell.LSTMCell(num_units=n_neurons, state_is_tuple=False), device='/gpu:0')
cell_bw = DeviceCellWrapper(cell=tf.nn.rnn_cell.LSTMCell(num_units=n_neurons, state_is_tuple=False), device='/gpu:0')
outputs, states = tf.nn.bidirectional_dynamic_rnn(cell_fw, cell_bw, X, dtype=tf.float32)