I am a newbie in Bayesian and Probabilistic inference, and sorry for this basic question. Recently I am following some examples in Bayesian Methods. And, the examples require me to use "tag.test_value." However, I am trying to use PyMC rather than PyMC3, so there is an error using the sentence. Although I tried to use others such as init_value, initial_value, it does not work...
Could you kindly let me know alternatives for that sentence to check the initial value in PyMC (that was originally test value in PyMC3)?
a = pm.Uniform("b", 0, 50)
print(a.tag.test_value)
AttributeError: 'ValidatingScratchpad' object has no attribute 'test_value
It appears that Aesara does not compute test value by default. You need to set aesara.config.compute_test_value = "warn". Then you can call a.get_test_value(). Hope this helps!
Related
I'm using TF-Agents library for reinforcement learning,
and I would like to take into account that, for a given state,
some actions are invalid.
How can this be implemented?
Should I define a "observation_and_action_constraint_splitter" function when
creating the DqnAgent?
If yes: do you know any tutorial on this?
Yes you need to define the function, pass it to the agent and also appropriately change the environment output so that the function can work with it. I am not aware on any tutorials on this, however you can look at this repo I have been working on.
Note that it is very messy and a lot of the files in there actually are not being used and the docstrings are terrible and often wrong (I forked this and didn't bother to sort everything out). However it is definetly working correctly. The parts that are relevant to your question are:
rl_env.py in the HanabiEnv.__init__ where the _observation_spec is defined as a dictionary of ArraySpecs (here). You can ignore game_obs, hand_obs and knowledge_obs which are used to run the environment verbosely, they are not fed to the agent.
rl_env.py in the HanabiEnv._reset at line 110 gives an idea of how the timestep observations are constructed and returned from the environment. legal_moves are passed through a np.logical_not since my specific environment marks legal_moves with 0 and illegal ones with -inf; whilst TF-Agents expects a 1/True for a legal move. My vector when cast to bool would therefore result in the exact opposite of what it should be for TF-agents.
These observations will then be fed to the observation_and_action_constraint_splitter in utility.py (here) where a tuple containing the observations and the action constraints is returned. Note that game_obs, hand_obs and knowledge_obs are implicitly thrown away (and not fed to the agent as previosuly mentioned.
Finally this observation_and_action_constraint_splitter is fed to the agent in utility.py in the create_agent function at line 198 for example.
In JES, I am able to use:
file=pickAFile()
In TigerJython, however, I get the following error
NameError: name 'pickAFile' is not defined
What am I doing wrong here?
You are not doing anything wrong at all. The thing is that pickAFile() is not a standard function in Python. It is actually rather a function that JES has added for convenience, but which you probably will not find it in any other environment.
Since TigerJython and JES are both based on Jython, you can easily write a pickAFile() function on your own that uses Java's Swing. Here is a possible simple implementation (the pickAFile() found in JES might be a bit more complex, but this should get you started):
def pickAFile():
from javax.swing import JFileChooser
fc = JFileChooser()
retVal = fc.showOpenDialog(None)
if retVal == JFileChooser.APPROVE_OPTION:
return fc.getSelectedFile()
else:
return None
Given that it is certainly a useful function, we might have to consider including it into our next update of TigerJython.
P.S. I would like to apologise for answering so late, I have just joined SO recently and was not aware of your question (I am one of the original authors of TigerJython).
I am trying to solve a problem involving the equating of sums of exponentials.
This is how I would do it hardcoded:
#NLconstraint(m, exp(x[25])==exp(x[14])+exp(x[18]))
This works fine with the rest of the code. However, when I try to do it for an arbitrary set of equations like the above I get an error. Here's my code:
#NLconstraint(m,[k=1:length(LHSSum)],sum(exp.(LHSSum[k][i]) for i=1:length(LHSSum[k]))==sum(exp.(RHSSum[k][i]) for i=1:length(RHSSum[k])))
where LHSSum and RHSSum are arrays containing arrays of the elements that need to be exponentiated and then summed over. That is LHSSum[1]=[x[1],x[2],x[3],...,x[n]]. Where x[i] are variables of type JuMP.Variable. Note that length(LHSSum)=length(RHSSum).
The error returned is:
LoadError: exp is not defined for type Variable. Are you trying to build a nonlinear problem? Make sure you use #NLconstraint/#NLobjective.
So a simple solution would be to simply do all the exponentiating and summing outside of the #NLconstraint function, so the input would be a scalar. However, this too presents a problem since exp(x) is not defined since x is of type JuMP.variable, whereas exp expects something of type real. This is strange since I am able to calculate exponentials just fine when the function is called within an #NLconstraint(). I.e. when I code this line#NLconstraint(m,exp(x)==exp(z)+exp(y)) instead of the earlier line, no errors are thrown.
Another thing I thought to do would be a Taylor Series expansion, but this too presents a problem since it goes into #NLconstraint land for powers greater than 2, and then I get stuck with the same vectorization problem.
So I feel stuck, I feel like if JuMP would allow for the vectorized evaluation of #NLconstraint like it does for #constraint, this would not even be an issue. Another fix would be if JuMP implements it's own exp function to allow for the exponentiation of JuMP.Variable type. However, as it is I don't see a way to solve this problem in general using the JuMP framework. Do any of you have any solutions to this problem? Any clever workarounds that I am missing?
I'm confused why i isn't used in the expressions you wrote. Do you mean:
#NLconstraint(m, [k = 1:length(LHSSum)],
sum(exp(LHSSum[k][i]) for i in 1:length(LHSSum[k]))
==
sum(exp(RHSSum[k][i]) for i in 1:length(RHSSum[k])))
I'm trying to use the linear chain CRF in my work. I took the help of the example usage code provided in -- https://github.com/tensorflow/tensorflow/tree/r1.0/tensorflow/contrib/crf
My question is how to supply some initial value of "transition_params" in "crf_log_likelihood()". For concreteness of the example, say, I want to initialize it with standard random normal distribution. In the api doc, I saw that "transition_params" can, in fact, be passed as an input argument. Inside the method I see that if no "transition_params" is passed, it is obtained by doing a "vs.get_variable()" with name = "transitions".
So should I do something similar to this, before creating the 'crf_log_likelihood' op? Something like -- transition_params = vs.get_variable("transitions", [num_tags, num_tags], initializer=tf.random_normal_initializer()) -- and then change the call of "crf_log_likelihood()" to "log_likelihood, transition_params = tf.contrib.crf.crf_log_likelihood(unary_scores, y_t, sequence_lengths_t, transition_params)"?
The get_variable() inside the definition of crf_log_likelihood() will create a fresh, randomly-initialized variable to represent the transition parameters, if you don't provide one yourself. You only need to provide an explicit transition_params if you don't want the default behavior.
To understand the behavior of get_variable(), see here:
https://www.tensorflow.org/api_docs/python/state_ops/sharing_variables#get_variable
Hope that helps!
I am getting a Duplicate tag error when I try to write out histogram summaries for a multi-layer network that I generate procedurally. I think that the problem might be related to naming. Imagine code like the following:
with tf.name_scope(some_unique_name):
...
_ = tf.histogram_summary('weights', kernel_weights)
I'd naively assumed that 'weights' would be scoped to some_unique_name but I'm suspecting that it is not. Are summary names independent of name_scope?
As Dave points out, the tag argument to tf.histogram_summary(tag, ...) is indeed independent of the current name scope. Part of the reason for this is that the tag may be a string Tensor (i.e. computed by part of your graph), whereas name scopes are a purely client-side construct (i.e. Python-only), so there's no good way to make the scoping work consistently across the two modes of use.
However, if you're using TensorFlow build from source (and should be available in the next release, 0.8.0), you can use the following recipe to scope your tags (using Graph.unique_name(..., mark_as_used=False)):
with tf.name_scope(some_unique_name):
# ...
tf.histogram_summary(
tf.get_default_graph().unique_name('weights', mark_as_used=False),
kernel_weights)
Alternatively, you can do the following in the current version:
with tf.name_scope(some_unique_name) as scope:
# ...
tf.histogram_summary(scope + 'weights', kernel_weights)
They are.
I'm with you in thinking this is a bug, but I haven't run it past the designers of the op yet. Go ahead and open an issue for it on GitHub!
(I've run into this also and found it terribly annoying -- it prevents reuse of the model without deliberately parameterizing the summary op invocations.)