Will gradient back propagate through multiple dynamic_rnn? - python

I use two stacking dynamic_rnn in my model, which means that the initial_state of the second dynamic_rnn is the final_state output by the first dynamic_rnn. My loss function is calculated only based on the output of the second dynamic_rnn. My question is that would the gradient be back propagated to the first dynamic_rnn?
You may ask me why I verbosely use two dynamic_rnn instead of one. The answer is that for my problem, most input sequences are totally identical except the last step. So I just run dynamic_rnn once for the common part of all these input sequences for the purpose of saving time and feed the final_state to another dynamic_rnn which accepts the distinct and last input elements.
Suppose that we have 3 sequences with length 10. All these sequences are identical except the last step (the 10th element). The simplified code:
cell = BasicRNNCell()
# the first dynamic_rnn which handles the common part
first_outputs, first_states = tf.nn.dynamic_rnn(
cell=cell,
dtype=tf.float32,
sequence_length=[9], # only one sample with length 9
inputs=identical_input # input with shape (1, 9, input_element_dim)
)
# tile the first_states to accommodate next dynamic_rnn
# first_states is transformed from shape (1, hidden_state_dim) to (3, hidden_state_dim)
first_states = tf.reshape(tf.tile(first_states, [1, 3]), [3, hidden_state_dim])
# the second dynamic_rnn which handles the distinct last element
second_outputs, second_states = tf.nn.dynamic_rnn(
initial_state=first_states,
cell=cell,
dtype=tf.float32,
sequence_length=[1, 1, 1], # 3 samples with only one element
inputs=distinct_input # input with shape (3, 1, input_element_dim)
)
# calculate loss based on second_outputs
loss = some_loss_function(second_outputs, groud_truth)

It should. If you're seeing problems, please describe the error you're getting in a little more detail.

Related

Cascade multiple RNN models for N-dimensional output

I'm having some difficulty with chaining together two models in an unusual way.
I am trying to replicate the following flowchart:
For clarity, at each timestep of Model[0] I am attempting to generate an entire time series from IR[i] (Intermediate Representation) as a repeated input using Model[1]. The purpose of this scheme is it allows the generation of a ragged 2-D time series from a 1-D input (while both allowing the second model to be omitted when the output for that timestep is not needed, and not requiring Model[0] to constantly "switch modes" between accepting input, and generating output).
I assume a custom training loop will be required, and I already have a custom training loop for handling statefulness in the first model (the previous version only had a single output at each timestep). As depicted, the second model should have reasonably short outputs (able to be constrained to fewer than 10 timesteps).
But at the end of the day, while I can wrap my head around what I want to do, I'm not nearly adroit enough with Keras and/or Tensorflow to actually implement it. (In fact, this is my first non-toy project with the library.)
I have unsuccessfully searched literature for similar schemes to parrot, or example code to fiddle with. And I don't even know if this idea is possible from within TF/Keras.
I already have the two models working in isolation. (As in I've worked out the dimensionality, and done some training with dummy data to get garbage outputs for the second model, and the first model is based off of a previous iteration of this problem and has been fully trained.) If I have Model[0] and Model[1] as python variables (let's call them model_a and model_b), then how would I chain them together to do this?
Edit to add:
If this is all unclear, perhaps having the dimensions of each input and output will help:
The dimensions of each input and output are:
Input: (batch_size, model_a_timesteps, input_size)
IR: (batch_size, model_a_timesteps, ir_size)
IR[i] (after duplication): (batch_size, model_b_timesteps, ir_size)
Out[i]: (batch_size, model_b_timesteps, output_size)
Out: (batch_size, model_a_timesteps, model_b_timesteps, output_size)
As this question has multiple major parts, I've dedicated a Q&A to the core challenge: stateful backpropagation. This answer focuses on implementing the variable output step length.
Description:
As validated in Case 5, we can take a bottom-up first approach. First we feed the complete input to model_a (A) - then, feed its outputs as input to model_b (B), but this time one step at a time.
Note that we must chain B's output steps per A's input step, not between A's input steps; i.e., in your diagram, gradient is to flow between Out[0][1] and Out[0][0], but not between Out[2][0] and Out[0][1].
For computing loss it won't matter whether we use a ragged or padded tensor; we must however use a padded tensor for writing to TensorArray.
Loop logic in code below is general; specific attribute handling and hidden state passing, however, is hard-coded for simplicity, but can be rewritten for generality.
Code: at bottom.
Example:
Here we predefine the number of iterations for B per input from A, but we can implement any arbitrary stopping logic. For example, we can take a Dense layer's output from B as a hidden state and check if its L2-norm exceeds a threshold.
Per above, if longest_step is unknown to us, we can simply set it, which is common for NLP & other tasks with a STOP token.
Alternatively, we may write to separate TensorArrays at every A's input with dynamic_size=True; see "point of uncertainty" below.
A valid concern is, how do we know gradients flow correctly? Note that we've validate them for both vertical and horizontal in the linked Q&A, but it didn't cover multiple output steps per an input step, for multiple input steps. See below.
Point of uncertainty: I'm not entirely sure whether gradients interact between e.g. Out[0][1] and Out[2][0]. I did, however, verify that gradients will not flow horizontally if we write to separate TensorArrays for B's outputs per A's inputs (case 2); reimplementing for cases 4 & 5, grads will differ for both models, including lower one with a complete single horizontal pass.
Thus we must write to a unified TensorArray. For such, as there are no ops leading from e.g. IR[1] to Out[0][1], I can't see how TF would trace it as such - so it seems we're safe. Note, however, that in below example, using steps_at_t=[1]*6 will make gradient flow in the both model horizontally, as we're writing to a single TensorArray and passing hidden states.
The examined case is confounded, however, with B being stateful at all steps; lifting this requirement, we might not need to write to a unified TensorArray for all Out[0], Out[1], etc, but we must still test against something we know works, which is no longer as straightforward.
Example [code]:
import numpy as np
import tensorflow as tf
#%%# Make data & models, then fit ###########################################
x0 = y0 = tf.constant(np.random.randn(2, 3, 4))
msn = MultiStatefulNetwork(batch_shape=(2, 3, 4), steps_at_t=[3, 4, 2])
#%%#############################################
with tf.GradientTape(persistent=True) as tape:
outputs = msn(x0)
# shape: (3, 4, 2, 4), 0-padded
# We can pad labels accordingly.
# Note the (2, 4) model_b's output shape, which is a timestep slice;
# model_b is a *slice model*. Careful in implementing various logics
# which are and aren't intended to be stateful.
Methods:
Not the cleanest, nor most optimal code, but it works; room for improvement.
More importantly: I implemented this in Eager, and have no idea how it'll work in Graph, and making it work for both can be quite tricky. If needed, just run in Graph and compare all values as done in the "cases".
# ideally we won't `import tensorflow` at all; kept for code simplicity
import tensorflow as tf
from tensorflow.python.util import nest
from tensorflow.python.ops import array_ops, tensor_array_ops
from tensorflow.python.framework import ops
from tensorflow.keras.layers import Input, SimpleRNN, SimpleRNNCell
from tensorflow.keras.models import Model
#######################################################################
class MultiStatefulNetwork():
def __init__(self, batch_shape=(2, 6, 4), steps_at_t=[]):
self.batch_shape=batch_shape
self.steps_at_t=steps_at_t
self.batch_size = batch_shape[0]
self.units = batch_shape[-1]
self._build_models()
def __call__(self, inputs):
outputs = self._forward_pass_a(inputs)
outputs = self._forward_pass_b(outputs)
return outputs
def _forward_pass_a(self, inputs):
return self.model_a(inputs, training=True)
def _forward_pass_b(self, inputs):
return model_rnn_outer(self.model_b, inputs, self.steps_at_t)
def _build_models(self):
ipt = Input(batch_shape=self.batch_shape)
out = SimpleRNN(self.units, return_sequences=True)(ipt)
self.model_a = Model(ipt, out)
ipt = Input(batch_shape=(self.batch_size, self.units))
sipt = Input(batch_shape=(self.batch_size, self.units))
out, state = SimpleRNNCell(4)(ipt, sipt)
self.model_b = Model([ipt, sipt], [out, state])
self.model_a.compile('sgd', 'mse')
self.model_b.compile('sgd', 'mse')
def inner_pass(model, inputs, states):
return model_rnn(model, inputs, states)
def model_rnn_outer(model, inputs, steps_at_t=[2, 2, 4, 3]):
def outer_step_function(inputs, states):
x, steps = inputs
x = array_ops.expand_dims(x, 0)
x = array_ops.tile(x, [steps, *[1] * (x.ndim - 1)]) # repeat steps times
output, new_states = inner_pass(model, x, states)
return output, new_states
(outer_steps, steps_at_t, longest_step, outer_t, initial_states,
output_ta, input_ta) = _process_args_outer(model, inputs, steps_at_t)
def _outer_step(outer_t, output_ta_t, *states):
current_input = [input_ta.read(outer_t), steps_at_t.read(outer_t)]
output, new_states = outer_step_function(current_input, tuple(states))
# pad if shorter than longest_step.
# model_b may output twice, but longest in `steps_at_t` is 4; then we need
# output.shape == (2, *model_b.output_shape) -> (4, *...)
# checking directly on `output` is more reliable than from `steps_at_t`
output = tf.cond(
tf.math.less(output.shape[0], longest_step),
lambda: tf.pad(output, [[0, longest_step - output.shape[0]],
*[[0, 0]] * (output.ndim - 1)]),
lambda: output)
output_ta_t = output_ta_t.write(outer_t, output)
return (outer_t + 1, output_ta_t) + tuple(new_states)
final_outputs = tf.while_loop(
body=_outer_step,
loop_vars=(outer_t, output_ta) + initial_states,
cond=lambda outer_t, *_: tf.math.less(outer_t, outer_steps))
output_ta = final_outputs[1]
outputs = output_ta.stack()
return outputs
def _process_args_outer(model, inputs, steps_at_t):
def swap_batch_timestep(input_t):
# Swap the batch and timestep dim for the incoming tensor.
# (samples, timesteps, channels) -> (timesteps, samples, channels)
# iterating dim0 to feed (samples, channels) slices expected by RNN
axes = list(range(len(input_t.shape)))
axes[0], axes[1] = 1, 0
return array_ops.transpose(input_t, axes)
inputs = nest.map_structure(swap_batch_timestep, inputs)
assert inputs.shape[0] == len(steps_at_t)
outer_steps = array_ops.shape(inputs)[0] # model_a_steps
longest_step = max(steps_at_t)
steps_at_t = tensor_array_ops.TensorArray(
dtype=tf.int32, size=len(steps_at_t)).unstack(steps_at_t)
# assume single-input network, excluding states which are handled separately
input_ta = tensor_array_ops.TensorArray(
dtype=inputs.dtype,
size=outer_steps,
element_shape=tf.TensorShape(model.input_shape[0]),
tensor_array_name='outer_input_ta_0').unstack(inputs)
# TensorArray is used to write outputs at every timestep, but does not
# support RaggedTensor; thus we must make TensorArray such that column length
# is that of the longest outer step, # and pad model_b's outputs accordingly
element_shape = tf.TensorShape((longest_step, *model.output_shape[0]))
# overall shape: (outer_steps, longest_step, *model_b.output_shape)
# for every input / at each step we write in dim0 (outer_steps)
output_ta = tensor_array_ops.TensorArray(
dtype=model.output[0].dtype,
size=outer_steps,
element_shape=element_shape,
tensor_array_name='outer_output_ta_0')
outer_t = tf.constant(0, dtype='int32')
initial_states = (tf.zeros(model.input_shape[0], dtype='float32'),)
return (outer_steps, steps_at_t, longest_step, outer_t, initial_states,
output_ta, input_ta)
def model_rnn(model, inputs, states):
def step_function(inputs, states):
output, new_states = model([inputs, *states], training=True)
return output, new_states
initial_states = states
input_ta, output_ta, time, time_steps_t = _process_args(model, inputs)
def _step(time, output_ta_t, *states):
current_input = input_ta.read(time)
output, new_states = step_function(current_input, tuple(states))
flat_state = nest.flatten(states)
flat_new_state = nest.flatten(new_states)
for state, new_state in zip(flat_state, flat_new_state):
if isinstance(new_state, ops.Tensor):
new_state.set_shape(state.shape)
output_ta_t = output_ta_t.write(time, output)
new_states = nest.pack_sequence_as(initial_states, flat_new_state)
return (time + 1, output_ta_t) + tuple(new_states)
final_outputs = tf.while_loop(
body=_step,
loop_vars=(time, output_ta) + tuple(initial_states),
cond=lambda time, *_: tf.math.less(time, time_steps_t))
new_states = final_outputs[2:]
output_ta = final_outputs[1]
outputs = output_ta.stack()
return outputs, new_states
def _process_args(model, inputs):
time_steps_t = tf.constant(inputs.shape[0], dtype='int32')
# assume single-input network (excluding states)
input_ta = tensor_array_ops.TensorArray(
dtype=inputs.dtype,
size=time_steps_t,
tensor_array_name='input_ta_0').unstack(inputs)
# assume single-output network (excluding states)
output_ta = tensor_array_ops.TensorArray(
dtype=model.output[0].dtype,
size=time_steps_t,
element_shape=tf.TensorShape(model.output_shape[0]),
tensor_array_name='output_ta_0')
time = tf.constant(0, dtype='int32', name='time')
return input_ta, output_ta, time, time_steps_t

Slicing torch tensors with list of indeces

I'm doing a reinforcement learning project, and I'm trying to get a tensor that represents the expected reward of all the given actions. I have a long tensor of chosen actions of size batch with values of either zero or one (the two potential actions). I have a tensor of expected rewards for each action of size batch * action_size, and I want a tensor of size batch.
For example, if batch size was 4, then I have
action = tensor([1,0,0,1])
expectedReward = tensor([[3,7],[5,9],[-1,12],[0,1]])
and what I want is
rewardForActions = tensor([7,5,-1,1])
I thought this would answer my question, but it's not the same at all, because if I went with that solution, it would end up with a 4*4 tensor, selecting from each row 4 times, instead of once.
Any ideas?
You could do
rewardForActions = expectedReward.index_select(1, action).diagonal()
# tensor([ 7, 5, -1, 1])

get the last output for non-padded entry from tf.nn.dynamic_rnn

I want to use an RNN in an all-to-one mode (only one output at the end). In TensorFlow, one can use:
lstm_cell = tf.nn.rnn_cell.LSTMCell(lstm_num_units)
output, _ = tf.nn.dynamic_rnn(lstm_cell, embed, dtype=tf.float32)
Where the output contains the output at all time steps [0, max_time-1], and the max_time is the length of the longest input in the batch.
Now, I would like to get the last output for every input in the batch. Let me be clearer. All the implementations that I have seen on the net, use output[:,-1] as the last output. However, for the inputs which have been padded, this would imply that the output is from the padded input.
Therefore, the questions:
How justified it is to use just output[:,-1]
Is there an easy way to select the last entry for the non-padded value in TensorFlow, which, in general, will be at different time steps for every input in the batch. Somehow, I found it a bit hard to do the necessary manipulations with the TensorFlow tensors, even when I have the original lengths of all input sequences.
Modifying your code:
_, state = tf.nn.dynamic_rnn(lstm_cell, embed, dtype=tf.float32, sequence_length=some_placeholder)
last_output = state.h
Don't forget to add a sequence_length parameter in the call to dynamic_rnn if you want the sequence lengths to vary.
Alternatively, you could use tf.gather_nd.

In tensorflow, how to iterate over a sequence of inputs stored in a tensor?

I am trying RNN on a variable length multivariate sequence classification problem.
I have defined following function to get the output of the sequence (i.e. the output of RNN cell after the final input from sequence is fed)
def get_sequence_output(x_sequence, initial_hidden_state):
previous_hidden_state = initial_hidden_state
for x_single in x_sequence:
hidden_state = gru_unit(previous_hidden_state, x_single)
previous_hidden_state = hidden_state
final_hidden_state = hidden_state
return final_hidden_state
Here x_sequence is tensor of shape (?, ?, 10) where first ? is for batch size and second ? is for sequence length and each input element is of length 10. gru function takes a previous hidden state and current input and spits out next hidden state (a standard gated recurrent unit).
I am getting an error: 'Tensor' object is not iterable.
How do I iterate over a Tensor in sequence manner (reading single element at a time)?
My objective is to apply gru function for every input from the sequence and get the final hidden state.
In TF>=1.0, tf.pack and tf.unpack are renamed to tf.stack and tf.unstack respectively
You can convert a tensor into a list using the unpack function which converts the first dimension into a list. There is also a split function which does something similar. I use unstack in an RNN model I am working on.
y = tf.unstack(tf.transpose(y, (1, 0, 2)))
In this case y starts out with shape (BATCH_SIZE, TIME_STEPS, 128) I transpose it to make the time steps the outer dimension and then unpack it into a list of tensors, one per time step. Now every element in the y list if of shape (BATCH_SIZE, 128) and I can feed it into my RNN.

Get last output of dynamic_rnn in tensorflow?

I am using dynamic_rnn to process MNIST data:
# LSTM Cell
lstm = rnn_cell.LSTMCell(num_units=200,
forget_bias=1.0,
initializer=tf.random_normal)
# Initial state
istate = lstm.zero_state(batch_size, "float")
# Get lstm cell output
output, states = rnn.dynamic_rnn(lstm, X, initial_state=istate)
# Output at last time point T
output_at_T = output[:, 27, :]
Full code: http://pastebin.com/bhf9MgMe
The input to the lstm is (batch_size, sequence_length, input_size)
As a result the dimensions of output_at_T is (batch_size, sequence_length, num_units) where num_units=200.
I need to get the last output along the sequence_length dimension. In the code above, this is hardcoded as 27. However, I do not know the sequence_length in advance as it can change from batch to batch in my application.
I tried:
output_at_T = output[:, -1, :]
but it says negative indexing is not implemented yet, and I tried using a placeholder variable as well as a constant (into which I could ideally feed the sequence_length for a particular batch); neither worked.
Any way to implement something like this in tensorflow atm?
Have you noticed that there are two outputs from dynamic_rnn?
Output 1, let's call it h, has all outputs at each time steps (i.e. h_1, h_2, etc),
Output 2, final_state, has two elements: the cell_state, and the last output for each element of the batch (as long as you input the sequence length to dynamic_rnn).
So from:
h, final_state= tf.dynamic_rnn( ..., sequence_length=[batch_size_vector], ... )
the last state for each element in the batch is:
final_state.h
Note that this includes the case when the length of the sequence is different for each element of the batch, as we are using the sequence_length argument.
This is what gather_nd is for!
def extract_axis_1(data, ind):
"""
Get specified elements along the first axis of tensor.
:param data: Tensorflow tensor that will be subsetted.
:param ind: Indices to take (one for each element along axis 0 of data).
:return: Subsetted tensor.
"""
batch_range = tf.range(tf.shape(data)[0])
indices = tf.stack([batch_range, ind], axis=1)
res = tf.gather_nd(data, indices)
return res
In your case (assuming sequence_length is a 1-D tensor with the length of each axis 0 element):
output = extract_axis_1(output, sequence_length - 1)
Now output is a tensor of dimension [batch_size, num_cells].
output[:, -1, :]
works with Tensorflow 1.x now!!
Most answers cover it thoroughly, but this code snip might help understand what's really being returned by the dynamic_rnn layer
=> Tuple of (outputs, final_output_state).
So for an input with max sequence length of T time steps outputs is of the shape [Batch_size, T, num_inputs] (given time_major=False; default value) and it contains the output state at each timestep h1, h2.....hT.
And final_output_state is of the shape [Batch_size,num_inputs] and has the final cell state cT and output state hT of each batch sequence.
But since the dynamic_rnn is being used my guess is your sequence lengths vary for each batch.
import tensorflow as tf
import numpy as np
from tensorflow.contrib import rnn
tf.reset_default_graph()
# Create input data
X = np.random.randn(2, 10, 8)
# The second example is of length 6
X[1,6:] = 0
X_lengths = [10, 6]
cell = tf.nn.rnn_cell.LSTMCell(num_units=64, state_is_tuple=True)
outputs, states = tf.nn.dynamic_rnn(cell=cell,
dtype=tf.float64,
sequence_length=X_lengths,
inputs=X)
result = tf.contrib.learn.run_n({"outputs": outputs, "states":states},
n=1,
feed_dict=None)
assert result[0]["outputs"].shape == (2, 10, 64)
print result[0]["outputs"].shape
print result[0]["states"].h.shape
# the final outputs state and states returned must be equal for each
# sequence
assert(result[0]["outputs"][0][-1]==result[0]["states"].h[0]).all()
assert(result[0]["outputs"][-1][5]==result[0]["states"].h[-1]).all()
assert(result[0]["outputs"][-1][-1]==result[0]["states"].h[-1]).all()
The final assertion will fail as the final state for the 2nd sequence is at 6th time step ie. the index 5 and the rest of the outputs from [6:9] are all 0s in the 2nd timestep
I am new to Stackoverflow and cannot comment yet so I am writing this new answer. #VM_AI, the last index is tf.shape(output)[1] - 1.
So, reusing your answer:
# Let's first fetch the last index of seq length
# last_index would have a scalar value
last_index = tf.shape(output)[1] - 1
# Then let's reshape the output to [sequence_length,batch_size,num_units]
# for convenience
output_rs = tf.transpose(output,[1,0,2])
# Last state of all batches
last_state = tf.nn.embedding_lookup(output_rs,last_index)
This works for me.
You should be able to access the shape of your output tensor using tf.shape(output). The tf.shape() function will return a 1d tensor containing the sizes of the output tensor. In your example, this would be (batch_size, sequence_length, num_units)
You should then be able to extract the value of output_at_T as output[:, tf.shape(output)[1], :]
There is a function in TensorFlow tf.shape that allows you to get the symbolic interpretation of shape rather than None being returned by output._shape[1]. And after fetching the last index you can lookup by using tf.nn.embedding_lookup, which is recommended especially when the data to be fetched is high as this does parallel lookup 32 by default.
# Let's first fetch the last index of seq length
# last_index would have a scalar value
last_index = tf.shape(output)[1]
# Then let's reshape the output to [sequence_length,batch_size,num_units]
# for convenience
output_rs = tf.transpose(output,[1,0,2])
# Last state of all batches
last_state = tf.nn.embedding_lookup(output_rs,last_index)
This should work.
Just to clarify what #Benoit Steiner said. His solution would not work as tf.shape would return symbolic interpretation of the shape value, and such cannot be used for slicing tensors i.e., direct indexing

Categories

Resources