I am using Keras with tensorflow backend and I am curious whether it is possible to skip a layer during backpropagation but have it execute in the forward pass. So here is what I mean
Lambda (lambda x: a(x))
I want to apply a to x in the forward pass but I do not want a to be included in the derivation when the backprop takes place.
I was trying to find a solution bit I could not find anything. Can somebody help me out here?
UPDATE 2
In addition to tf.py_func, there is now an official guide on how to add a custom op.
UPDATE
See this question for an example of writing a custom op with gradient purely in Python without needing to rebuild anything. Note that there are some limitations to the method (see the documentation of tf.py_func).
Not exactly a solution to the problem, but still kind of an answer and too long for comments.
That's not even a Keras issue, but a TensorFlow one. Each op defines its own gradient computation that is used during backpropagation. I you really wanted to something like that, you would need to implement the op into TensorFlow yourself (no easy feat) and define the gradient that you want - because you can't have "no gradient", if anything it would be 1 or 0 (otherwise you can't go on with backpropagation). There is a tf.NoGradient function in TensorFlow which causes an op to propagate zeros, but I don't think it is meant to / can be used out of TensorFlow own internals.
UPDATE
Okay so a bit more of context. TensorFlow graphs are built of ops, which are implemented by kernels; this is basically a 1-to-1 mapping, except that there may be for example a CPU and a GPU kernel for an op, hence the differentiation. The set of ops supported by TensorFlow is usually static, I mean it can change with newer versions, but in principle you cannot add your own ops, because the ops of a graph go into the Protobuf serialized format, so if you made your own ops then you would not be able to share your graph. Ops are then defined at C++ level with the macro REGISTER_OP (see for example here), and kernels with REGISTER_KERNEL_BUILDER (see for example here).
Now, where do gradients come into play? Well, the funny thing is that the gradient of an op is not defined at C++ level; there are ops (and kernels) that implement the gradient of other ops (if you look at the previous files you'll find ops/kernels with the name ending in Grad), but (as far as I'm aware) these are not explicitly "linked" at this level. It seems that the associations between ops and their gradients is defined in Python, usually via tf.RegisterGradient or the aforementioned tf.NoGradient (see for example here, Python modules starting with gen_ are autogenerated with the help of the C++ macros); these registrations inform the backpropagation algorithm about how to compute the gradient of the graph.
So, how to actually work this out? Well, you need to create at least one op in C++ with the corresponding kernel/s implementing the computation that you want for your forward pass. Then, if the gradient computation that you want to use can be expressed with existing TensorFlow ops (which is most likely), you would just need to call tf.RegisterGradient in Python and do the computation there in "standard" TensorFlow. This is quite complicated, but the good news is it's possible, and there's even an example for it (although I think they kinda forgot the gradient registration part in that one)! As you will see, the process involves compiling the new op code into a library (btw I'm not sure if any of this may work on Windows) that is then loaded from Python (obviously this involves going through the painful process of manual compilation of TensorFlow with Bazel). A possibly more realistic example can be found in TensorFlow Fold, an extension of TensorFlow for structured data that register (as of one) one custom operation here through a macro defined here that calls REGISTER_OP, and then in Python it loads the library and register its gradient here through their own registration function defined here that simply calls tf.NotDifferentiable (another name for tf.NoGradient)
tldr: It is rather hard, but it can be done and there are even a couple of examples out there.
As mentioned in #jdehesa's comments. You can implement your function with an "alternative gradient". Forgive me if my math is not correct, but I think a derivative returning "1" would be the correct way to have no effect on the backpropagation while still passing the learning through. For how to construct it, see here. The example I cited goes further and allows you to construct an activation function from a python function. So in place of the spiky function, substitute your function a, and in place of his derivative d_spiky replace it with
def constant(x):
return 1
So on the forward pass, a is applied in the layer and the the backwards pass 1 is applied which should simply pass the weight adjustments through.
You can then just create an Activation layer in Keras using this function.
Related
I started by Tensorflow journey when it already came to 2.0.0, So never used graphs and sessions as in version1. But recently met tf.function and autographs which suits me. (but what i know is it is used only for train step)
Now when reading project code, many people use tf.function decorator on many other functions when they wanna build graphs. But i don't exactly get their point. How to know when to use graph and when not?
Can anyone help me?
Solution
The decorator, #tf.function conveniently converts a python function to a static tensorflow graph. TensorFlow operates in eager mode by default since version 2.0.0. Although eager mode could help you in line-by-line execution, this comes with the pitfall of relatively slower TensorFlow-code execution when compared to static-graph. Converting a certain function into a static graph increases execution speed while training your model.
Quoting tf.function documentation:
Functions can be faster than eager code, especially for graphs with many small ops. But for graphs with a few expensive ops (like convolutions), you may not see much speedup.
The static graph is created once and does not get updated if the function is called repeatedly with different values (not passed as the input-arguments). You should avoid using #tf.function in such scenarios or update the function definition (if possible) to include all the necessary variability through the input-arguments. However,
Now, if your function gets all its inputs through the function arguments, then if you apply #tf.function you will not see any problem.
Here is an example.
### When not to use #tf.function ###
# some variable that changes with time
var = timestamp()
#tf.function
def func(*args, **kwargs):
# your code
return var
In the example above, the function func() although depends on var, it does not access the variable var through its arguments. Thus, when #tf.function is applied for the first time, it creates a static-graph for func(). However, when the value of var changes in future, this will not get updated in the static-graph. See this for more clarity. Also, I would highly encourage you to see the references section.
For Debugging
Quoting source
You can use tf.config.experimental_run_functions_eagerly (which temporarily disables running functions as functions) for debugging purposes.
References
Better performance with tf.function
When to utilize tf.function
TensorFlow 2.0: tf.function and AutoGraph
I am trying to understand how the internal flow goes in mxnet when we call forward . Is there any way to get source code of mxnet?
This really depends on what your symbolic graph looks like. I assume you use MXNet with Python (Python documentation). There you can choose to use the MXNet symbol library or the Gluon library.
Now, you were asking whether one can inspect the code, and, yes, you can find it on GitHub. The folder python contains the python interface and src contains all MXNet sources. What happens on forward is eventually defined by the MXNet execution engine, which tracks input/output dependencies of operators and neural network layers, allocate memory on the different devices (CPU, GPUs). There is a general architecture documentation for this.
I suppose you are interested in what each and every operation does, such as argmax (reduction), tanh (unary math operation) or convolution (complex neural network operation). This you can find in the operator folder of MXNet. This requires a whole documentation in itself and there is a special forum for MXNet specifics here, but I will give a short orientation:
Each operation in a (symbolic) execution graph needs a defined forward and backward operation. It also needs to define its output shape, so that it can be chained with other operations. If that operator needs weights, it needs to define the amount of memory it requires, so MXNet can allocate it.
Each operation requires several implementations for a) CPU b) GPU (CUDA) c) wrapper around cuDNN
All unary math operations follow the same pattern, so they are all defined in a similar way in mshadow_op.h (e.g. relu).
This is all I can tell you based on your quite broad question.
In tensorflow 1.4, I found two functions that do batch normalization and they look same:
tf.layers.batch_normalization (link)
tf.contrib.layers.batch_norm (link)
Which function should I use? Which one is more stable?
Just to add to the list, there're several more ways to do batch-norm in tensorflow:
tf.nn.batch_normalization is a low-level op. The caller is responsible to handle mean and variance tensors themselves.
tf.nn.fused_batch_norm is another low-level op, similar to the previous one. The difference is that it's optimized for 4D input tensors, which is the usual case in convolutional neural networks. tf.nn.batch_normalization accepts tensors of any rank greater than 1.
tf.layers.batch_normalization is a high-level wrapper over the previous ops. The biggest difference is that it takes care of creating and managing the running mean and variance tensors, and calls a fast fused op when possible. Usually, this should be the default choice for you.
tf.contrib.layers.batch_norm is the early implementation of batch norm, before it's graduated to the core API (i.e., tf.layers). The use of it is not recommended because it may be dropped in the future releases.
tf.nn.batch_norm_with_global_normalization is another deprecated op. Currently, delegates the call to tf.nn.batch_normalization, but likely to be dropped in the future.
Finally, there's also Keras layer keras.layers.BatchNormalization, which in case of tensorflow backend invokes tf.nn.batch_normalization.
As show in doc, tf.contrib is a contribution module containing volatile or experimental code. When function is complete, it will be removed from this module. Now there are two, in order to be compatible with the historical version.
So, the former tf.layers.batch_normalization is recommended.
If I am defining a custom Op in Tensorflow, is it possible to provide two Kernels for the top that are polymorphic on whether the shape for the inputs are fully defined? For example, I can construct certain structures once at Kernel construction if the shape is fully known / defined.
It's not currently possible to do this. The kernel dispatch mechanism is implemented in a low-level part of the TensorFlow code where information about tensor shapes is not (generally) available.
However, the ability to specialize a graph based on known shapes does seem like a useful ability, and it might be worth raising this as a feature request on the GitHub issues page. One possible workaround would be to try registering an optimization pass that makes use of shape information and rewrites the names of ops with known input shapes to a different op that relies on static shape information (e.g. via an additional attr). However, doing this in TensorFlow currently requires you to rebuild from source.
I was looking through the API in TensorFlow and notice that a lot of mathematical operations that already exist in python and numpy have been re-implemented (or at least given a tensorflow interface). For example:
is there a good reason to do this?
I've been searching over their page but can't find why they'd do this.
I do have some guesses though. One of my main guesses is that they probably want those operations to have some backpropagation effect on whatever Neural network graph that gets implementat. In other words, have their derivatives implemented. Is this one of the reasons? (wish I knew how to even check if my guess is right)
For example, in one of the most basic examples of linear regression, one defines the prediction function that one wants to implement:
product = tf.matmul(x,W)
y = product + b
instead of
product = tf.matmul(x,W)
y = tf.add(product, b)
Somehow the first implementation does not interfere with Stochastic Gradient Descent algorithm for training, so it probably doesn't matter if one uses numpy or tf.add to train? This is one aspect that confuses me, when do I know which one should I be using.
Or maybe they are performance reasons? Or maybe its to give those operations access to GPU if required to use GPUs?
You have to understand that you create a tensorflow graph with this operation, meaning they aren't the same as the numpy functions, they are more an abstraction of them.
Maybe you have noticed that you have to create a session and then evaluate the functions through that session to get a result, where with numpy functions they are executed directly. this is because this graph and its functions define what to do like writing down a formula, but to get results for a specific x (or whatever) you have to insert a value for x. This is what your doing through session and eval.
So to conclude this you define a graph with tensorflow which is a more abstract representation of the functions and the graph also isn't executed at runtime, then it is defined, it will be executed when you call the eval function and through that run the session.
Also notice that you cant mix numpy functions and tensorflow functions directly but you can define own tensorflow functions (https://www.tensorflow.org/versions/r0.9/how_tos/adding_an_op/index.html)
Btw I guess most of the tensorflow functions are using numpy under the hood. :)