I'm trying to make a Seq2Seq Regression example for time-series analysis and I've used the Seq2Seq library as presented at the Dev Summit, which is currently the code on the Tensorflow GitHub branch r1.0.
I have difficulties understanding how the decoder function works for Seq2Seq, specifically for the "cell_output".
I understand that the num_decoder_symbols is the number of classes/words to decode at each time step. I have it working at a point where I can do training. However, I don't get why I can't just substitute the number of features (num_features) instead of num_decoder_symbols. Basically, I want to be able to run the decoder without teacher forcing, in other words pass the output of the previous time step as the input to the next time step.
with ops.name_scope(name, "simple_decoder_fn_inference",
[time, cell_state, cell_input, cell_output,
context_state]):
if cell_input is not None:
raise ValueError("Expected cell_input to be None, but saw: %s" %
cell_input)
if cell_output is None:
# invariant that this is time == 0
next_input_id = array_ops.ones([batch_size,], dtype=dtype) * (
start_of_sequence_id)
done = array_ops.zeros([batch_size,], dtype=dtypes.bool)
cell_state = encoder_state
cell_output = array_ops.zeros([num_decoder_symbols],
dtype=dtypes.float32)
Here is a link to the original code: https://github.com/tensorflow/tensorflow/blob/r1.0/tensorflow/contrib/seq2seq/python/ops/decoder_fn.py
Why don't I need to pass batch_size for the cell output?
cell_output = array_ops.zeros([batch_size, num_decoder_symbols],
dtype=dtypes.float32)
When trying to use this code to create my own regressive Seq2Seq example, where instead of having an output of probabilities/classes, I have a real valued vector of dimension num_features, instead of an array of probability of classes. As I understood, I thought I could replace num_decoder_symbols with num_features, like below:
def decoder_fn(time, cell_state, cell_input, cell_output, context_state):
"""
Again same as in simple_decoder_fn_inference but for regression on sequences with a fixed length
"""
with ops.name_scope(name, "simple_decoder_fn_inference", [time, cell_state, cell_input, cell_output, context_state]):
if cell_input is not None:
raise ValueError("Expected cell_input to be None, but saw: %s" % cell_input)
if cell_output is None:
# invariant that this is time == 0
next_input = array_ops.ones([batch_size, num_features], dtype=dtype)
done = array_ops.zeros([batch_size], dtype=dtypes.bool)
cell_state = encoder_state
cell_output = array_ops.zeros([num_features], dtype=dtypes.float32)
else:
cell_output = output_fn(cell_output)
done = math_ops.equal(0,1) # hardcoded hack just to properly define done
next_input = cell_output
# if time > maxlen, return all true vector
done = control_flow_ops.cond(math_ops.greater(time, maximum_length),
lambda: array_ops.ones([batch_size,], dtype=dtypes.bool),
lambda: done)
return (done, cell_state, next_input, cell_output, context_state)
return decoder_fn
But, I get the following error:
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/contrib/seq2seq/python/ops/seq2seq.py", line 212, in dynamic_rnn_decoder
swap_memory=swap_memory, scope=scope)
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 1036, in raw_rnn
swap_memory=swap_memory)
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/control_flow_ops.py", line 2605, in while_loop
result = context.BuildLoop(cond, body, loop_vars, shape_invariants)
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/control_flow_ops.py", line 2438, in BuildLoop
pred, body, original_loop_vars, loop_vars, shape_invariants)
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/control_flow_ops.py", line 2388, in _BuildLoop
body_result = body(*packed_vars_for_body)
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 980, in body
(next_output, cell_state) = cell(current_input, state)
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/contrib/rnn/python/ops/core_rnn_cell_impl.py", line 327, in __call__
input_size = inputs.get_shape().with_rank(2)[1]
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/python/framework/tensor_shape.py", line 635, in with_rank
raise ValueError("Shape %s must have rank %d" % (self, rank))
ValueError: Shape (100,) must have rank 2
As a result, I passed in the batch_size like this in order to get a Shape of rank 2:
cell_output = array_ops.zeros([batch_size, num_features],
dtype=dtypes.float32)
But I get the following error, where Shape is of rank 3 and wants a rank 2 instead:
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/contrib/seq2seq/python/ops/seq2seq.py", line 212, in dynamic_rnn_decoder
swap_memory=swap_memory, scope=scope)
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 1036, in raw_rnn
swap_memory=swap_memory)
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/control_flow_ops.py", line 2605, in while_loop
result = context.BuildLoop(cond, body, loop_vars, shape_invariants)
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/control_flow_ops.py", line 2438, in BuildLoop
pred, body, original_loop_vars, loop_vars, shape_invariants)
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/control_flow_ops.py", line 2388, in _BuildLoop
body_result = body(*packed_vars_for_body)
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 980, in body
(next_output, cell_state) = cell(current_input, state)
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/contrib/rnn/python/ops/core_rnn_cell_impl.py", line 327, in __call__
input_size = inputs.get_shape().with_rank(2)[1]
File "/opt/DL/tensorflow/lib/python2.7/site-packages/tensorflow/python/framework/tensor_shape.py", line 635, in with_rank
raise ValueError("Shape %s must have rank %d" % (self, rank))
ValueError: Shape (10, 10, 100) must have rank 2
