beld/tensorflow-seq2seq-autoencoder

tensorflow-seq2seq-autoencoder's Issues

Incorrect dimensions (leading to misleading performance)

Hi,

In Tensorflow-seq2seq-autoencoder/simple_seq2seq_autoencoder.py , the code runs and training converges for me, but I think it isn't implementing the LSTM autoencoder properly:

I notice that encoder_inputs is a list of length 1, containing a reference to a Tensorflow tensor object of shape (10,3) = (size (of the hidden layer),frame_dim).

help(tf.nn.rnn) says that input is assumed to be a list of length T, where each element is a reference to a Tensorflow tensor of shape [batch_size x cell.state_size].

tf.nn.rnn may see this as a sequence of length 1, where the batch size is 10 and the frame dimension is 3.

Another approach may involve
encoder_inputs = tf.unpack(seq_input)

Here is a sketch of my attempt to fix this. It has a few other unrelated changes, sorry about that..

# if encoder input is "X, Y, Z", then decoder input is "0, X, Y, Z". Therefore, the decoder size
# and target size equal encoder size plus 1. For simplicity, here I droped the last one.
encoder_inputs_unpacked=tf.unpack(encoder_inputs)
decoder_inputs = ([tf.zeros_like(encoder_inputs_unpacked[0], name="GO")] + encoder_inputs_unpacked[:-1])

targets = desired_outputs
weights = [tf.ones_like(targets_t, dtype=tf.float32) for targets_t in tf.unpack(targets)]

# basic LSTM seq2seq model
cell = tf.nn.rnn_cell.GRUCell(size)
enc_output, enc_state = tf.nn.dynamic_rnn(cell, encoder_inputs, dtype=tf.float32, time_major=True)
enc_state=tf.nn.dropout(keep_prob=0.8,x=enc_state)
enc_state_unpacked=tf.unpack(enc_state,num=n_steps)

if makeHiddenLayerBinary:
    enc_state=tf.nn.sigmoid(enc_state)

cell_to_reconstruction = tf.nn.rnn_cell.OutputProjectionWrapper(cell, frame_dim)
dec_outputs, dec_state = tf.nn.seq2seq.rnn_decoder(decoder_inputs, enc_state, cell_to_reconstruction)
# e_stop = np.array([1, 1])

'module' object has no attribute 'rnn'

(tensorflow2.7) itamblyn@laertes:~/tmp/Tensorflow-seq2seq-autoencoder$ python simple_seq2seq_autoencoder.py
Traceback (most recent call last):
File "simple_seq2seq_autoencoder.py", line 37, in
_, enc_state = tf.nn.rnn(cell, encoder_inputs, dtype=tf.float32)
AttributeError: 'module' object has no attribute 'rnn'