I was playing around with Tensorflow creating a customized loss function and this question about general machine learning arose to my head.
My understanding is that the optimization algorithm needs a derivable cost function to find/approach a minimum, however we can use functions that are non-derivable such as the absolute function (there is no derivative when x=0). A more extreme example, I defined my cost function like this:
def customLossFun(x,y):
return tf.sign(x)
and I expected an error when running the code, but it actually worked (it didn't learn anything but it didn't crash).
Am I missing something?
You're missing the fact that the gradient of the sign function is somewhere manually defined in the Tensorflow source code.
As you can see here:
def _SignGrad(op, _):
"""Returns 0."""
x = op.inputs[0]
return array_ops.zeros(array_ops.shape(x), dtype=x.dtype)
the gradient of tf.sign is defined to be always zero. This, of course, is the gradient where the derivate exists, hence everywhere but not in zero.
The tensorflow authors decided to do not check if the input is zero and throw an exception in that specific case
In order to prevent TensorFlow from throwing an error, the only real requirement is that you cost function evaluates to a number for any value of your input variables. From a purely "will it run" perspective, it doesn't know/care about the form of the function its trying to minimize.
In order for your cost function to provide you a meaningful result when TensorFlow uses it to train a model, it additionally needs to 1) get smaller as your model does better and 2) be bounded from below (i.e. it can't go to negative infinity). It's not generally necessary for it to be smooth (e.g. abs(x) has a kink where the sign flips). Tensorflow is always able to compute gradients at any location using automatic differentiation (https://en.wikipedia.org/wiki/Automatic_differentiation, https://www.tensorflow.org/versions/r0.12/api_docs/python/train/gradient_computation).
Of course, those gradients are of more use if you've chose a meaningful cost function isn't isn't too flat.
Ideally, the cost function needs to be smooth everywhere to apply gradient based optimization methods (SGD, Momentum, Adam, etc). But nothing's going to crash if it's not, you can just have issues with convergence to a local minimum.
When the function is non-differentiable at a certain point x, it's possible to get large oscillations if the neural network converges to this x. E.g., if the loss function is tf.abs(x), it's possible that the network weights are mostly positive, so the inference x > 0 at all times, so the network won't notice tf.abs. However, it's more likely that x will bounce around 0, so that the gradient is arbitrarily positive and negative. If the learning rate is not decaying, the optimization won't converge to the local minimum, but will bound around it.
In your particular case, the gradient is zero all the time, so nothing's going to change at all.
If it didn't learn anything, what have you gained ? Your loss function is differentiable almost everywhere but it is flat almost anywhere so the minimizer can't figure out the direction towards the minimum.
If you start out with a positive value, it will most likely be stuck at a random value on the positive side even though the minima on the left side are better (have a lower value).
Tensorflow can be used to do calculations in general and it provides a mechanism to automatically find the derivative of a given expression and can do so across different compute platforms (CPU, GPU) and distributed over multiple GPUs and servers if needed.
But what you implement in Tensorflow does not necessarily have to be a goal function to be minimized. You could use it e.g. to throw random numbers and perform Monte Carlo integration of a given function.
Related
I appreciate if you can answer my questions or provide me with useful resources.
Currently, I am working on a problem that I need to do alternating optimization. So, consider we have two decision variables x and y. In the first step I take the derivative of loss function wrt. x (for fixed y) and update x. On the second step, I need to take the derivative wrt. y. The issue is x is dependent on y implicitly and finding the closed form of cost function in a way to show the dependency of x on y is not feasible, so the gradients of cost function wrt. y are unknown.
1) My first question is whether "autodiff" method in reverse mode used in TensorFlow works for these problems where we do not have an explicit form of cost function wrt to one variable and we need the derivatives? Actually, the value of cost function is known but the dependency on decision variable is unknown via math.
2) From a general view, if I define a node as a "tf.Variable" and have an arbitrary intractable function(intractable via computation by hand) of that variable that evolves through code execution, is it possible to calculate the gradients via "tf.gradients"? If yes, how can I make sure that it is implemented correctly? Can I check it using TensorBoard?
My model is too complicated but a simplified form can be considered in this way: suppose the loss function for my model is L(x). I can code L(x) as a function of "x" during the construction phase in tensorflow. However, I have also another variable "k" that is initialized to zero. The dependency of L(x) on "k" shapes as the code runs so my loss function is L(x,k), actually. And more importantly, "x" is a function of "k" implicitly. (all the optimization is done using GradientDescent). The problem is I do not have L(x,k) as a closed form function but I have the value of L(x,k) at each step. I can use "numerical" methods like FDSA/SPSA but they are not exact. I just need to make sure as you said there is a path between "k" and L(x,k)but I do not know how!
TensorFlow gradients only work when the graph connecting the x and the y when you're computing dy/dx has at least one path which contains only differentiable operations. In general if tf gives you a gradient it is correct (otherwise file a bug, but gradient bugs are rare, since the gradient for all differentiable ops is well tested and the chain rule is fairly easy to apply).
Can you be a little more specific about what your model looks like? You might also want to use eager execution if your forward complication is too weird to express as a fixed dataflow graph.
I am using this function of tensorflow to get my function jacobian. Came across two problems:
The tensorflow documentation is contradicted to itself in the following two paragraph if I am not mistaken:
gradients() adds ops to the graph to output the partial derivatives of ys with respect to xs. It returns a list of Tensor of length len(xs) where each tensor is the sum(dy/dx) for y in ys.
Blockquote
Blockquote
Returns:
A list of sum(dy/dx) for each x in xs.
Blockquote
According to my test, it is, in fact, return a vector of len(ys) which is the sum(dy/dx) for each x in xs.
I do not understand why they designed it in a way that the return is the sum of the columns(or row, depending on how you define your Jacobian).
How can I really get the Jacobian?
4.In the loss, I need the partial derivative of my function with respect to input (x), but when I am optimizing with respect to the network weights, I define x as a placeholder whose value is fed later, and weights are variable, in this case, can I still define the symbolic derivative of function with respect to input (x)? and put it in the loss? ( which later when we optimize with respect to weights will bring second order derivative of the function.)
I think you are right and there is a typo there, it was probably meant to be "of length len(ys)".
For efficiency. I can't explain exactly the reasoning, but this seems to be a pretty fundamental characteristic of how TensorFlow handles automatic differentiation. See issue #675.
There is no straightforward way to get the Jacobian matrix in TensorFlow. Take a look at this answer and again issue #675. Basically, you need one call to tf.gradients per column/row.
Yes, of course. You can compute whatever gradients you want, there is no real difference between a placeholder and any other operation really. There are a few operations that do not have a gradient because it is not well defined or not implemented (in which case it will generally return 0), but that's all.
I use the scipy.optimize.minimize ( https://docs.scipy.org/doc/scipy/reference/tutorial/optimize.html ) function with method='L-BFGS-B.
An example of what it returns is here above:
fun: 32.372210618549758
hess_inv: <6x6 LbfgsInvHessProduct with dtype=float64>
jac: array([ -2.14583906e-04, 4.09272616e-04, -2.55795385e-05,
3.76587650e-05, 1.49213975e-04, -8.38440428e-05])
message: 'CONVERGENCE: REL_REDUCTION_OF_F_<=_FACTR*EPSMCH'
nfev: 420
nit: 51
status: 0
success: True
x: array([ 0.75739412, -0.0927572 , 0.11986434, 1.19911266, 0.27866406,
-0.03825225])
The x value correctly contains the fitted parameters. How do I compute the errors associated to those parameters?
TL;DR: You can actually place an upper bound on how precisely the minimization routine has found the optimal values of your parameters. See the snippet at the end of this answer that shows how to do it directly, without resorting to calling additional minimization routines.
The documentation for this method says
The iteration stops when (f^k - f^{k+1})/max{|f^k|,|f^{k+1}|,1} <= ftol.
Roughly speaking, the minimization stops when the value of the function f that you're minimizing is minimized to within ftol of the optimum. (This is a relative error if f is greater than 1, and absolute otherwise; for simplicity I'll assume it's an absolute error.) In more standard language, you'll probably think of your function f as a chi-squared value. So this roughly suggests that you would expect
Of course, just the fact that you're applying a minimization routine like this assumes that your function is well behaved, in the sense that it's reasonably smooth and the optimum being found is well approximated near the optimum by a quadratic function of the parameters xi:
where Δxi is the difference between the found value of parameter xi and its optimal value, and Hij is the Hessian matrix. A little (surprisingly nontrivial) linear algebra gets you to a pretty standard result for an estimate of the uncertainty in any quantity X that's a function of your parameters xi:
which lets us write
That's the most useful formula in general, but for the specific question here, we just have X = xi, so this simplifies to
Finally, to be totally explicit, let's say you've stored the optimization result in a variable called res. The inverse Hessian is available as res.hess_inv, which is a function that takes a vector and returns the product of the inverse Hessian with that vector. So, for example, we can display the optimized parameters along with the uncertainty estimates with a snippet like this:
ftol = 2.220446049250313e-09
tmp_i = np.zeros(len(res.x))
for i in range(len(res.x)):
tmp_i[i] = 1.0
hess_inv_i = res.hess_inv(tmp_i)[i]
uncertainty_i = np.sqrt(max(1, abs(res.fun)) * ftol * hess_inv_i)
tmp_i[i] = 0.0
print('x^{0} = {1:12.4e} ± {2:.1e}'.format(i, res.x[i], uncertainty_i))
Note that I've incorporated the max behavior from the documentation, assuming that f^k and f^{k+1} are basically just the same as the final output value, res.fun, which really ought to be a good approximation. Also, for small problems, you can just use np.diag(res.hess_inv.todense()) to get the full inverse and extract the diagonal all at once. But for large numbers of variables, I've found that to be a much slower option. Finally, I've added the default value of ftol, but if you change it in an argument to minimize, you would obviously need to change it here.
One approach to this common problem is to use scipy.optimize.leastsq after using minimize with 'L-BFGS-B' starting from the solution found with 'L-BFGS-B'. That is, leastsq will (normally) include and estimate of the 1-sigma errors as well as the solution.
Of course, that approach makes several assumption, including that leastsq can be used and may be appropriate for solving the problem. From a practical view, this requires the objective function return an array of residual values with at least as many elements as variables, not a cost function.
You may find lmfit (https://lmfit.github.io/lmfit-py/) useful here: It supports both 'L-BFGS-B' and 'leastsq' and gives a uniform wrapper around these and other minimization methods, so that you can use the same objective function for both methods (and specify how to convert the residual array into the cost function). In addition, parameter bounds can be used for both methods. This makes it very easy to first do a fit with 'L-BFGS-B' and then with 'leastsq', using the values from 'L-BFGS-B' as starting values.
Lmfit also provides methods to more explicitly explore confidence limits on parameter values in more detail, in case you suspect the simple but fast approach used by leastsq might be insufficient.
It really depends what you mean by "errors". There is no general answer to your question, because it depends on what you're fitting and what assumptions you're making.
The easiest case is one of the most common: when the function you are minimizing is a negative log-likelihood. In that case the inverse of the hessian matrix returned by the fit (hess_inv) is the covariance matrix describing the Gaussian approximation to the maximum likelihood.The parameter errors are the square root of the diagonal elements of the covariance matrix.
Beware that if you are fitting a different kind of function or are making different assumptions, then that doesn't apply.
I have a question about a reason why setting TensorFlow's variable with small stddev.
I guess many people do test MNIST test code from TensorFlow beginner's guide.
As following it, the first layer's weights are initiated by using truncated_normal with stddev 0.1.
And I guessed if setting it with more bigger value, then it would be the same result, which is exactly accurate.
But although increasing epoch count, it doesn't work.
Is there anybody know this reason?
original :
W_layer = tf.Variable(tf.truncated_normal([inp.get_shape()[1].value, size],stddev=0.1), name='w_'+name)
#result : (990, 0.93000001, 0.89719999)
modified :
W_layer = tf.Variable(tf.truncated_normal([inp.get_shape()[1].value, size],stddev=200), name='w_'+name)
#result : (99990, 0.1, 0.098000005)
The reason is because you want to keep all the layer's variances (or standard deviations) approximately the same, and sane. It has to do with the error backpropagation step of the learning process and the activation functions used.
In order to learn the network's weights, the backpropagation step requires knowledge of the network's gradient, a measure of how strong each weight influences the input to reach the final output; layer's weight variance directly influences the propagation of gradients.
Say, for example, that the activation function is sigmoidal (e.g. tf.nn.sigmoid or tf.nn.tanh); this implies that all input values are squashed into a fixed output value range. For the sigmoid, it is the range 0..1, where essentially all values z greater or smaller than +/- 4 are very close to one (for z > 4) or zero (for z < -4) and only values within that range tend to have some meaningful "change".
Now the difference between the values sigmoid(5) and sigmoid(1000) is barely noticeable. Because of that, all very large or very small values will optimize very slowly, since their influence on the result y = sigmoid(W*x+b) is extremely small. Now the pre-activation value z = W*x+b (where x is the input) depends on the actual input x and the current weights W. If either of them is large, e.g. by initializing the weights with a high variance (i.e. standard deviation), the result will necessarily be (relatively) large, leading to said problem. This is also the reason why truncated_normal is used rather than a correct normal distribution: The latter only guarantees that most of the values are very close to the mean, with some less than 5% chance that this is not the case, while truncated_normal simply clips away every value that is too big or too small, guaranteeing that all weights are in the same range, while still being normally distributed.
To make matters worse, in a typical neural network - especially in deep learning - each network layer is followed by one or many others. If in each layer the output value range is big, the gradients will get bigger and bigger as well; this is known as the exploding gradients problem (a variation of the vanishing gradients, where gradients are getting smaller).
The reason that this is a problem is because learning starts at the very last layer and each weight is adjusted depending on how much it contributed to the error. If the gradients are indeed getting very big towards the end, the very last layer is the first one to pay a high toll for this: Its weights get adjusted very strongly - likely overcorrecting the actual problem - and then only the "remaining" error gets propagated further back, or up, the network. Here, since the last layer was already "fixed a lot" regarding the measured error, only smaller adjustments will be made. This may lead to the problem that the first layers are corrected only by a tiny bit or not at all, effectively preventing all learning there. The same basically happens if the learning rate is too big.
Finding the best weight initialization is a topic by itself and there are somewhat more sophisticated methods such as Xavier initialization or Layer-sequential unit variance, however small normally distributed values are usually simply a good guess.
Question: What is the most efficient way to get the delta of my weights in the most efficient way in a TensorFlow network?
Background: I've got the operators hooked up as follows (thanks to this SO question):
self.cost = `the rest of the network`
self.rmsprop = tf.train.RMSPropOptimizer(lr,rms_decay,0.0,rms_eps)
self.comp_grads = self.rmsprop.compute_gradients(self.cost)
self.grad_placeholder = [(tf.placeholder("float", shape=grad[1].get_shape(), name="grad_placeholder"), grad[1]) for grad in self.comp_grads]
self.apply_grads = self.rmsprop.apply_gradients(self.grad_placeholder)
Now, to feed in information, I run the following:
feed_dict = `training variables`
grad_vals = self.sess.run([grad[0] for grad in self.comp_grads], feed_dict=feed_dict)
feed_dict2 = `feed_dict plus gradient values added to self.grad_placeholder`
self.sess.run(self.apply_grads, feed_dict=feed_dict2)
The command of run(self.apply_grads) will update the network weights, but when I compute the differences in the starting and ending weights (run(self.w1)), those numbers are different than what is stored in grad_vals[0]. I figure this is because the RMSPropOptimizer does more to the raw gradients, but I'm not sure what, or where to find out what it does.
So back to the question: How do I get the delta on my weights in the most efficient way? Am I stuck running self.w1.eval(sess) multiple times to get the weights and calc the difference? Is there something that I'm missing with the tf.RMSPropOptimizer function.
Thanks!
RMSprop does not subtract the gradient from the parameters but use more complicated formula involving a combination of:
a momentum, if the corresponding parameter is not 0
a gradient step, rescaled non uniformly (on each coordinate) by the square root of the squared average of the gradient.
For more information you can refer to these slides or this recent paper.
The delta is first computed in memory by tensorflow in the slot variable 'momentum' and then the variable is updated (see the C++ operator).
Thus, you should be able to access it and construct a delta node with delta_w1 = self.rmsprop.get_slot(self.w1, 'momentum'). (I have not tried it yet.)
You can add the weights to the list of things to fetch each run call. Then you can compute the deltas outside of TensorFlow since you will have the iterates. This should be reasonably efficient, although it might incur an extra elementwise difference, but to avoid that you might have to hack around in the guts of the optimizer and find where it puts the update before it applies it and fetch that each step. Fetching the weights each call shouldn't do wasteful extra evaluations of part of the graph at least.
RMSProp does complicated scaling of the learning rate for each weight. Basically it divides the learning rate for a weight by a running average of the magnitudes of recent gradients of that weight.