Based on the link:
https://www.tensorflow.org/api_docs/python/tf/nn/dynamic_rnn
In the example, it is shown that the "initial state" is defined in the first example and not in the second example. Could anyone please explain what is the purpose of the initial state? What's the difference if I don't set it vs if i set it? Is it only required in a single RNN cell and not in a stacked cell like in the example provided in the link?
I'm currently debugging my RNN model, as it seemed to classify different questions in the same category, which is strange. I suspect that it might have to do with me not setting the initial state of the cell.
Could anyone please explain what is the purpose of initial state?
As we know that the state matrix is the weights between the hidden neurons in timestep 1 and timestep 2. They join the hidden neurons of both the time steps. Hence they hold temporal data from the layers in previous time steps.
Providing an initially trained state matrix by the initial_state= argument gives the RNN cell a trained memory of its previous activations.
What's the difference if I don't set it vs if I set it?
If we set the initial weights which have been trained on some other model or the previous model, it means that we are restoring the memory of the RNN cell so that it does not have to start from scratch.
In the TF docs, they have initialized the initial_state as zero_state matrix.
If you don't set the initial_state, it will be trained from scratch as other weight matrices do.
Is it only required in a single RNN cell and not in a stacked cell like in the example provided in the link?
I exactly don't know that why haven't they set the initial_state in the Stacked RNN example, but initial_state is required in every type of RNN as it holds the preserves the temporal features across time steps.
Maybe, Stacked RNN was the point of interest in the docs and not the settings of initial_state.
Tip:
In most cases, you will not need to set the initial_state for an RNN. TensorFlow can handle this efficiently for us. In the case of seq2seq RNN, this property may be used.
Your RNN maybe facing some other issue. Your RNN build ups its own memory and doesn't require powerup.
Related
Goal
I want to compare different types of RNN tflite-micro models, built using tensorflow, on a microcontroller based on their accuracy, model size and inference time. I have also created my own custom RNN cell that I want to compare with the LSTM cell, GRU cell, and SimpleRNN cell. I create the tensorflow model using tf.keras.layers.RNN(Cell(...)).
Problem
I have successfully deployed a keras LSTM-RNN using tf.keras.layers.LSTM(...) but when I create the same model using tf.keras.layers.RNN(tf.keras.layers.LSTMCell(...)) and deploy it to the microcontroller, then it does not work. I trained both networks on a batch size of 64, but then I copy the weights and biases to a model where the batch_size is fixed to 1 as tflite-micro does not support dynamic batch sizes.
When the keras LSTM layer is converted to a tflite model it creates a fused operator called UnidirectionalSequenceLSTM but the network created with an RNN layer using the LSTMCell does not have that UnidirectionalSequenceLSTM operator, instead it has a reshape and while operator. The first network has only 1 subgraph but the second has 3 subgraphs.
When I run that second model on the microcontroller, two things go wrong:
the interpreter returns the same result for different inputs
the interpreter fails on some inputs reporting an error with the while loop saying that int32 is not supported (which is in the while operator, and can't be quantized to int8)
LSTM tflite-model vizualized with Netron
RNN(LSTMCell) tflite-model vizualized with Netron
Bad solution (10x model size)
I figured out that by unrolling the second network I can successfully deploy it and get correct results on the microcontroller. However, that increases the model size 10x which is really bad as we are trying to deploy the model on a resource constrained device.
Better solution?
I have explained the problem using the example of the LSTM layer (works) and LSTM cell in an RNN layer (does not work), but I want to be able to deploy a model using the GRU cell, SimpleRNN cell, and of course the custom cell that I have created. And all those have the same problem as the network created with the LSTM cell.
What can I do?
Do I have to create a special fused operator? Maybe even one for each cell I want to compare? How would I do that?
Can I use the interface into the conversion infrastructure for user-defined RNN implementations mentioned here: https://www.tensorflow.org/lite/models/convert/rnn. How I understand the documentation, is that this would only work for user-defined LSTM implementations, not user-defined RNN implemenations like the title suggests.
I am in trouble with understanding the concept of LSTM and using it on Keras. When considering a LSTM layer, there should be two values for output size and the hidden state size.
1. hidden state size : how many features are passed across the time steps of a samples when training the model
2. output size : how many outputs should be returned by particular LSTM layer
But in keras.layers.LSTM, there is only one parameter and it is used to control the output size of the layer.
PROBLEM:
Therefore how hidden state size of the LSTM layer can be changed?
If I am misunderstood, corrections are really appreciated.
You are getting confused between the difference in hidden units and output units in LSTM. Please refer to the below link for better clarity:
https://jasdeep06.github.io/posts/Understanding-LSTM-in-Tensorflow-MNIST/
Basically what you provide in num_units is the size of LSTM hidden unit only. This is very clear from this article.
Here is my understanding of a basic Sequence to Sequence LSTMs. Suppose we are tackling a question-answer setting.
You have two set of LSTMs (green and blue below). Each set respectively sharing weights (i.e. each of the 4 green cells have the same weights and similarly with the blue cells). The first is a many to one LSTM, which summarises the question at the last hidden layer/ cell memory.
The second set (blue) is a Many to Many LSTM which has different weights to the first set of LSTMs. The input is simply the answer sentence while the output is the same sentence shifted by one.
The question is two fold:
1. Are we passing the last hidden state only to the blue LSTMs as the initial hidden state. Or is it last hidden state and cell memory.
2. Is there a way to set the initial hiddden state and cell memory in Keras or Tensorflow? If so reference?
(image taken from suriyadeepan.github.io)
Are we passing the last hidden state only to the blue LSTMs as the initial hidden state. Or is it last hidden state and cell memory.
Both hidden state h and cell memory c are passed to the decoder.
TensorFlow
In seq2seq source code, you can find the following code in basic_rnn_seq2seq():
_, enc_state = rnn.static_rnn(enc_cell, encoder_inputs, dtype=dtype)
return rnn_decoder(decoder_inputs, enc_state, cell)
If you use an LSTMCell, the returned enc_state from the encoder will be a tuple (c, h). As you can see, the tuple is passed directly to the decoder.
Keras
In Keras, the "state" defined for an LSTMCell is also a tuple (h, c) (note that the order is different from TF). In LSTMCell.call(), you can find:
h_tm1 = states[0]
c_tm1 = states[1]
To get the states returned from an LSTM layer, you can specify return_state=True. The returned value is a tuple (o, h, c). The tensor o is the output of this layer, which will be equal to h unless you specify return_sequences=True.
Is there a way to set the initial hiddden state and cell memory in Keras or Tensorflow? If so reference?
###TensorFlow###
Just provide the initial state to an LSTMCell when calling it. For example, in the official RNN tutorial:
lstm = tf.contrib.rnn.BasicLSTMCell(lstm_size)
...
output, state = lstm(current_batch_of_words, state)
There's also an initial_state argument for functions such as tf.nn.static_rnn. If you use the seq2seq module, provide the states to rnn_decoder as have been shown in the code for question 1.
###Keras###
Use the keyword argument initial_state in the LSTM function call.
out = LSTM(32)(input_tensor, initial_state=(h, c))
You can actually find this usage on the official documentation:
###Note on specifying the initial state of RNNs###
You can specify the initial state of RNN layers symbolically by
calling them with the keyword argument initial_state. The value of
initial_state should be a tensor or list of tensors representing the
initial state of the RNN layer.
EDIT:
There's now an example script in Keras (lstm_seq2seq.py) showing how to implement basic seq2seq in Keras. How to make prediction after training a seq2seq model is also covered in this script.
(Edit: this answer is incomplete and hasn't considered actual possibilities of state transfering. See the accepted answer).
From a Keras point of view, that picture has only two layers.
The green group is one LSTM layer.
The blue group is another LSTM layer.
There isn't any communication between green and blue other than passing the outputs. So, the answer for 1 is:
Only the thought vector (which is the actual output of the layer) is passed to the other layer.
Memory and state (not sure if these are two different entities) are totally contained inside a single layer and are not initially intended to be seen or shared with any other layer.
Each individual block in that image is totally invisible in keras. They are considered "time steps", something that only appears in the shape of the input data. It's rarely important to worry about them (unless for very advanced usages).
In keras, it's like this:
Easily, you have access only to the external arrows (including "thought vector").
But having access to each step (each individual green block in your picture) is not an exposed thing. So...
Passing the states from one layer to the other is also not expected in Keras. You will probably have to hack things. (See this: https://github.com/fchollet/keras/issues/2995)
But considering a thought vector big enough, you could say it will learn a way to carry what is important in itself.
The only notion you have from the steps is:
You have to input things shaped like (sentences, length, wordIdFeatures)
The steps will be performed considering that each slice in the length dimension is an input to each green block.
You may choose to have a single output (sentences, cells), for which you completely lose track of steps. Or...
Outputs like (sentences, length, cells), from which you know the output of each block through the length dimension.
One to many or many to many?
Now, the first layer is many to one (but nothing prevents it from being many to many too if you want).
But the second... that's complicated.
If the thought vector was made by a many to one. You will have to manage a way of creating a one to many. (That's not trivial in keras, but you could think of repeating the thought vector for the expected length, making it be the input to all steps. Or maybe fill an entire sequence with zeros or ones, keeping only the first element as the thought vector)
If the thought vector was made by a many to many, you can take advantage of this and keep an easy many to many, if you're willing to accept that the output has exactly the same number of steps as the input.
Keras doesn't have a ready solution for 1 to many cases. (From a single input predict a whole sequence).
In MNIST LSTM examples, I don't understand what "hidden layer" means. Is it the imaginary-layer formed when you represent an unrolled RNN over time?
Why is the num_units = 128 in most cases ?
From this brilliant article
num_units can be interpreted as the analogy of hidden layer from the feed forward neural network. The number of nodes in hidden layer of a feed forward neural network is equivalent to num_units number of LSTM units in a LSTM cell at every time step of the network.
See the image there too!
The number of hidden units is a direct representation of the learning capacity of a neural network -- it reflects the number of learned parameters. The value 128 was likely selected arbitrarily or empirically. You can change that value experimentally and rerun the program to see how it affects the training accuracy (you can get better than 90% test accuracy with a lot fewer hidden units). Using more units makes it more likely to perfectly memorize the complete training set (although it will take longer, and you run the risk of over-fitting).
The key thing to understand, which is somewhat subtle in the famous Colah's blog post (find "each line carries an entire vector"), is that X is an array of data (nowadays often called a tensor) -- it is not meant to be a scalar value. Where, for example, the tanh function is shown, it is meant to imply that the function is broadcast across the entire array (an implicit for loop) -- and not simply performed once per time-step.
As such, the hidden units represent tangible storage within the network, which is manifest primarily in the size of the weights array. And because an LSTM actually does have a bit of it's own internal storage separate from the learned model parameters, it has to know how many units there are -- which ultimately needs to agree with the size of the weights. In the simplest case, an RNN has no internal storage -- so it doesn't even need to know in advance how many "hidden units" it is being applied to.
A good answer to a similar question here.
You can look at the source for BasicLSTMCell in TensorFlow to see exactly how this is used.
Side note: This notation is very common in statistics and machine-learning, and other fields that process large batches of data with a common formula (3D graphics is another example). It takes a bit of getting used to for people who expect to see their for loops written out explicitly.
The argument n_hidden of BasicLSTMCell is the number of hidden units of the LSTM.
As you said, you should really read Colah's blog post to understand LSTM, but here is a little heads up.
If you have an input x of shape [T, 10], you will feed the LSTM with the sequence of values from t=0 to t=T-1, each of size 10.
At each timestep, you multiply the input with a matrix of shape [10, n_hidden], and get a n_hidden vector.
Your LSTM gets at each timestep t:
the previous hidden state h_{t-1}, of size n_hidden (at t=0, the previous state is [0., 0., ...])
the input, transformed to size n_hidden
it will sum these inputs and produce the next hidden state h_t of size n_hidden
From Colah's blog post:
If you just want to have code working, just keep with n_hidden = 128 and you will be fine.
An LSTM keeps two pieces of information as it propagates through time:
A hidden state; which is the memory the LSTM accumulates using its (forget, input, and output) gates through time, and
The previous time-step output.
Tensorflow’s num_units is the size of the LSTM’s hidden state (which is also the size of the output if no projection is used).
To make the name num_units more intuitive, you can think of it as the number of hidden units in the LSTM cell, or the number of memory units in the cell.
Look at this awesome post for more clarity
Since I had some problems to combine the information from the different sources I created the graphic below which shows a combination of the blog post (http://colah.github.io/posts/2015-08-Understanding-LSTMs/) and (https://jasdeep06.github.io/posts/Understanding-LSTM-in-Tensorflow-MNIST/) where I think the graphics are very helpful but an error in explaining the number_units is present.
Several LSTM cells form one LSTM layer. This is shown in the figure below. Since you are mostly dealing with data that is very extensive, it is not possible to incorporate everything in one piece into the model. Therefore, data is divided into small pieces as batches, which are processed one after the other until the batch containing the last part is read in. In the lower part of the figure you can see the input (dark grey) where the batches are read in one after the other from batch 1 to batch batch_size. The cells LSTM cell 1 to LSTM cell time_step above represent the described cells of the LSTM model (http://colah.github.io/posts/2015-08-Understanding-LSTMs/). The number of cells is equal to the number of fixed time steps. For example, if you take a text sequence with a total of 150 characters, you could divide it into 3 (batch_size) and have a sequence of length 50 per batch (number of time_steps and thus of LSTM cells). If you then encoded each character one-hot, each element (dark gray boxes of the input) would represent a vector that would have the length of the vocabulary (number of features). These vectors would flow into the neuronal networks (green elements in the cells) in the respective cells and would change their dimension to the length of the number of hidden units (number_units). So the input has the dimension (batch_size x time_step x features). The Long Time Memory (Cell State) and Short Time Memory (Hidden State) have the same dimensions (batch_size x number_units). The light gray blocks that arise from the cells have a different dimension because the transformations in the neural networks (green elements) took place with the help of the hidden units (batch_size x time_step x number_units). The output can be returned from any cell but mostly only the information from the last block (black border) is relevant (not in all problems) because it contains all information from the previous time steps.
I think it is confusing for TF users by the term "num_hidden". Actually it has nothing to do with the unrolled LSTM cells, and it just is the dimension of the tensor, which is transformed from the time-step input tensor to and fed into the LSTM cell.
This term num_units or num_hidden_units sometimes noted using the variable name nhid in the implementations, means that the input to the LSTM cell is a vector of dimension nhid (or for a batched implementation, it would a matrix of shape batch_size x nhid). As a result, the output (from LSTM cell) would also be of same dimensionality since RNN/LSTM/GRU cell doesn't alter the dimensionality of the input vector or matrix.
As pointed out earlier, this term was borrowed from Feed-Forward Neural Networks (FFNs) literature and has caused confusion when used in the context of RNNs. But, the idea is that even RNNs can be viewed as FFNs at each time step. In this view, the hidden layer would indeed be containing num_hidden units as depicted in this figure:
Source: Understanding LSTM
More concretely, in the below example the num_hidden_units or nhid would be 3 since the size of hidden state (middle layer) is a 3D vector.
I think this is a correctly answer for your question. LSTM always make confusion.
You can refer this blog for more detail Animated RNN, LSTM and GRU
Most LSTM/RNN diagrams just show the hidden cells but never the units of those cells. Hence, the confusion.
Each hidden layer has hidden cells, as many as the number of time steps.
And further, each hidden cell is made up of multiple hidden units, like in the diagram below. Therefore, the dimensionality of a hidden layer matrix in RNN is (number of time steps, number of hidden units).
The Concept of hidden unit is illustrated in this image https://imgur.com/Fjx4Zuo.
Following #SangLe answer, I made a picture (see sources for original pictures) showing cells as classically represented in tutorials (Source1: Colah's Blog) and an equivalent cell with 2 units (Source2: Raimi Karim 's post). Hope it will clarify confusion between cells/units and what really is the network architecture.
I'm using tensorflow to run a cnn for image classification.
I use tensorflow cifar10 cnn implementation.(tensorflow cifar10)
I want to decrease the number of connections, meaning I want to prune the low-weight connections.
How can I create a new graph(subgraph) without some of the nuerones?
Tensorflow does not allow you lock/freeze a particular kernel of a particular layer, that I have found. The only I've found to do this is to use the tf.assign() function as shown in
How to freeze/lock weights of one Tensorflow variable (e.g., one CNN kernel of one layer
It's fairly cave-man but I've seen no other solution that works. Essentially, you have to .assign() the values every so often as you iterate through the data. Since this approach is so inelegant and brute-force, it's very slow. I do the .assign() every 100 batches.
Someone please post a better solution and soon!
The cifar10 model you point to, and for that matter, most models written in TensorFlow, do not model the weights (and hence, connections) of individual neurons directly in the computation graph. For instance, for fully connected layers, all the connections between the two layers, say, with M neurons in the layer below, and 'N' neurons in the layer above, are modeled by one MxN weight matrix. If you wanted to completely remove a neuron and all of its outgoing connections from the layer below, you can simply slice out a (M-1)xN matrix by removing the relevant row, and multiply it with the corresponding M-1 activations of the neurons.
Another way is add an addition mask to control the connections.
The first step involves adding mask and threshold variables to the
layers that need to undergo pruning. The variable mask is the same
shape as the layer's weight tensor and determines which of the weights
participate in the forward execution of the graph.
There is a pruning implementation under tensorflow/contrib/model_pruning to prune the model. Hope this can help you to prune model quickly.
https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/model_pruning
I think google has an updated answer here : https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/model_pruning
Removing pruning nodes from the trained graph:
$ bazel build -c opt contrib/model_pruning:strip_pruning_vars
$ bazel-bin/contrib/model_pruning/strip_pruning_vars --checkpoint_path=/tmp/cifar10_train --output_node_names=softmax_linear/softmax_linear_2 --filename=cifar_pruned.pb
I suppose that cifar_pruned.pb will be smaller, since the pruned "or zero masked" variables are removed.