multiPLSTM and RNN doesn't parameterize the cell correctly.

multiPLSTM basically hardcode to cell = PhasedLSTMCell(units_p_layer, use_peepholes=True, state_is_tuple=True) so the params such as r_on, tau_init cannot be passed in.

Also in simplePhasedLSTM.py, the RNN function always creates PLSTM cell, not LSTM or GRU. This means that the run flag is not used flags.DEFINE_string("unit", "PLSTM", "Can be PSLTM, LSTM, GRU").

I'm happy to contribute a fix.

Looks like this has gone out of the window in tensorflow 1.0, tensorflow/tensorflow#6899

Hi, How to remember states of lstm between runs?

Hey!

I am confused about one thing, the time should be in the first position of the input tensor right? For example, if I have a input tensor of shape (?, seq_len, nb_features), [:, :, 0] would be the times.

i am trying to make a plstm_classifier using the code of https://github.com/guillaume-chevalier/LSTM-Human-Activity-Recognition and your plsm cells.
i'm having troubles understanding the needed dimensions and params..

I added how i've tried oto combine it, basically this change:

Define two stacked PLSTM cells (two recurrent layers deep) with tensorflow

lstm_cell_1 = PhasedLSTMCell(config.n_hidden)
lstm_cell_2 = PhasedLSTMCell(config.n_hidden)
# Previous lstm cells for lstm model
#lstm_cell_1 = tf.contrib.rnn.BasicLSTMCell(config.n_hidden, forget_bias=1.0, state_is_tuple=True)
#lstm_cell_2 = tf.contrib.rnn.BasicLSTMCell(config.n_hidden, forget_bias=1.0, state_is_tuple=True)
#lstm_cells = tf.contrib.rnn.MultiRNNCell([lstm_cell_1, lstm_cell_2], state_is_tuple=True)
initial_states = [LSTMStateTuple(tf.zeros([config.batch_size, config.n_hidden], tf.float32), tf.zeros([config.batch_size, config.n_hidden], tf.float32)) for _ in range(config.n_inputs)]
# Get PLSTM cell output
outputs, states = multiPLSTM([lstm_cell_1, lstm_cell_2], _X, config.n_steps,config.n_inputs, initial_states)
# Previous lstm cell output before change - for stm moder
#outputs, states = tf.contrib.rnn.static_rnn(lstm_cells, _X, dtype=tf.float32)

But i encounter lots if issues which i hoe can be solved easily..

SOS pls :)

error1.txt
error2.txt
plstm_classifier.txt

Hi Enny1991:
I am using tensorflow of version 0.12.head and I encounter an error:
LookupError: No gradient defined for operation '3/rnn/while/PhasedLSTMCell/FloorMod_1' (op type: FloorMod)
while execute line 202 in simplePhasedLSTM.py ---- grads = tf.gradients(cost, tvars)
But I can't find the op FloorMod anywhere.
How can I solve this problem? Thanks a lot!

The Kronos gate is applied at the end.

# APPLY KRONOS GATE
c = k * c + (1. - k) * c_prev
m = k * m + (1. - k) * m_prev
# END KRONOS GATE

But I think the output o from Equation (4) should be updated with the value computed in the Equation (8).
For now it's updated with c_tilde I guess.

https://arxiv.org/pdf/1610.09513v1.pdf

As per the Lip Reading experiment in the paper, it would be ideal to have the ability to manually set and disable learning of Tau and s.

I'm currently having a go at it myself but am only new to TensorFlow. I have added a new boolean input argument to PhasedLSTMCell.

Based on this parameter being True, I've modified the following:

1. Set get_variable trainable to True for tau, r_on and s
2. Initialize tau and s with init_ops.constant_initializer in place of the random methods, setting each to tau_init and 0, respectively.
3. Set a hard on period for gate by using:

k = tf.select(tf.logical_or(is_up,is_down), tf.to_float(is_up), self.alpha * phi)

Would appreciate any comments on this method and am happy to contribute.

Hi,

First thanks for converting the original code to TF. I get the following error when running the test script:

python simplePhasedLSTM.py:

Traceback (most recent call last):
File "simplePhasedLSTM.py", line 6, in
from PhasedLSTMCell import PhasedLSTMCell, multiPLSTM
File "/Users/new/plstm/PhasedLSTMCell.py", line 91, in
class PhasedLSTMCell(RNNCell):
NameError: name 'RNNCell' is not defined

I am using tf 1.2.1 and python 2.7.13

Hello,
Have you done re-writing your code for TF2?

thank you.