I am trying to use Efficientnet.jl as a feature extractor, meaning I want to extract all features after a given block in the flux chain.
There is the build in function
features = model(x, Val(:stages))
which returns all features after each block, which is very memory inefficient, since I only need to store values after exactly 3 blocks.
My thought was to only use a subset of the layers this way:
transparent_model = model.blocks[1:model.stages[3]]
features = transparent_model(x)
Unfortunately I get the following Error:
DimensionMismatch("Input channels must match! (3 vs. 1)")
which is in my opinion just due to a bad error message.
size(x) -> (1280,720,3,1)
Related
I've got a symbol that represents the name of a function to be called:
julia> func_sym = :tanh
I can use that symbol to get the tanh function and call it using:
julia> eval(func_sym)(2)
0.9640275800758169
But I'd rather avoid the 'eval' there as it will be called many times and it's expensive (and func_sym can have several different values depending on context).
IIRC in Ruby you can say something like:
obj.send(func_sym, args)
Is there something similar in Julia?
EDIT: some more details on why I have functions represented by symbols:
I have a type (from a neural network) that includes the activation function, originally I included it as a funcion:
type NeuralLayer
weights::Matrix{Float32}
biases::Vector{Float32}
a_func::Function
end
However, I needed to serialize these things to files using JLD, but it's not possible to serialize a Function, so I went with a symbol:
type NeuralLayer
weights::Matrix{Float32}
biases::Vector{Float32}
a_func::Symbol
end
And currently I use the eval approach above to call the activation function. There are collections of NeuralLayers and each can have it's own activation function.
#Isaiah's answer is spot-on; perhaps even more-so after the edit to the original question. To elaborate and make this more specific to your case: I'd change your NeuralLayer type to be parametric:
type NeuralLayer{func_type}
weights::Matrix{Float32}
biases::Vector{Float32}
end
Since func_type doesn't appear in the types of the fields, the constructor will require you to explicitly specify it: layer = NeuralLayer{:excitatory}(w, b). One restriction here is that you cannot modify a type parameter.
Now, func_type could be a symbol (like you're doing now) or it could be a more functionally relevant parameter (or parameters) that tunes your activation function. Then you define your activation functions like this:
# If you define your NeuralLayer with just one parameter:
activation(layer::NeuralLayer{:inhibitory}) = …
activation(layer::NeuralLayer{:excitatory}) = …
# Or if you want to use several physiological parameters instead:
activation{g_K,g_Na,g_l}(layer::NeuralLayer{g_K,g_Na,g_l} = f(g_K, g_Na, g_l)
The key point is that functions and behavior are external to the data. Use type definitions and abstract type hierarchies to define behavior, as is coded in the external functions… but only store data itself in the types. This is dramatically different from Python or other strongly object-oriented paradigms, and it takes some getting used to.
But I'd rather avoid the 'eval' there as it will be called many times and it's expensive (and func_sym can have several different values depending on context).
This sort of dynamic dispatch is possible in Julia, but not recommended. Changing the value of 'func_sym' based on context defeats type inference as well as method specialization and inlining. Instead, the recommended approach is to use multiple dispatch, as detailed in the Methods section of the manual.
I am trying to create a function
import Language.Reflection
foo : Type -> TT
I tried it by using the reflect tactic:
foo = proof
{
intro t
reflect t
}
but this reflects on the variable t itself:
*SOQuestion> foo
\t => P Bound (UN "t") (TType (UVar 41)) : Type -> TT
Reflection in Idris is a purely syntactic, compile-time only feature. To predict how it will work, you need to know about how Idris converts your program to its core language. Importantly, you won't be able to get ahold of reflected terms at runtime and reconstruct them like you would with Lisp. Here's how your program is compiled:
Internally, Idris creates a hole that will expect something of type Type -> TT.
It runs the proof script for foo in this state. We start with no assumptions and a goal of type Type -> TT. That is, there's a term being constructed which looks like ?rhs : Type => TT . rhs. The ?foo : ty => body syntax shows that there's a hole called foo whose eventual value will be available inside of body.
The step intro t creates a function whose argument is t : Type - this means that we now have a term like ?foo_body : TT . \t : Type => foo_body.
The reflect t step then fills the current hole by taking the term on its right-hand side and converting it to a TT. That term is in fact just a reference to the argument of the function, so you get the variable t. reflect, like all other proof script steps, only has access to the information that is available directly at compile time. Thus, the result of filling in foo_body with the reflection of the term t is P Bound (UN "t") (TType (UVar (-1))).
If you could do what you are wanting here, it would have major consequences both for understanding Idris code and for running it efficiently.
The loss in understanding would come from the inability to use parametricity to reason about the behavior of functions based on their types. All functions would effectively become potentially ad-hoc polymorphic, because they could (say) run differently on lists of strings than on lists of ints.
The loss in performance would come from representing enough type information to do the reflection. After Idris code is compiled, there is no type information left in it (unlike in a system such as the JVM or .NET or a dynamically typed system such as Python, where types have a runtime representation that code can access). In Idris, types can be very large, because they can contain arbitrary programs - this means that far more information would have to be maintained, and computation occurring at the type level would also have to be preserved and repeated at runtime.
If you're wanting to reflect on the structure of a type for further proof automation at compile time, take a look at the applyTactic tactic. Its argument should be a function that takes a reflected context and goal and gives back a new reflected tactic script. An example can be seen in the Data.Vect source.
So I suppose the summary is that Idris can't do what you want, and it probably never will be able to, but you might be able to make progress another way.
I am currently learning Modelica by trying some very simple examples. I have defined a connector Incompressible for an incompressible fluid like this:
connector Incompressible
flow Modelica.SIunits.VolumeFlowRate V_dot;
Modelica.SIunits.SpecificEnthalpy h;
Modelica.SIunits.Pressure p;
end Incompressible;
I now wish to define a mass or volume flow source:
model Source_incompressible
parameter Modelica.SIunits.VolumeFlowRate V_dot;
parameter Modelica.SIunits.Temperature T;
parameter Modelica.SIunits.Pressure p;
Incompressible outlet;
equation
outlet.V_dot = V_dot;
outlet.h = enthalpyWaterIncompressible(T); // quick'n'dirty enthalpy function
outlet.p = p;
end Source_incompressible;
However, when checking Source_incompressible, I get this:
The problem is structurally singular for the element type Real.
The number of scalar Real unknown elements are 3.
The number of scalar Real equation elements are 4.
I am at a loss here. Clearly, there are three equations in the model - where does the fourth equation come from?
Thanks a lot for any insight.
Dominic,
There are a couple of issues going on here. As Martin points out, the connector is unbalanced (you don't have matching "through" and "across" pairs in that connector). For fluid systems, this is acceptable. However, intensive fluid properties (e.g., enthalpy) have to be marked as so-called "stream" variables.
This topic is, admittedly, pretty complicated. I'm planning on adding an advanced chapter to my online Modelica book on this topic but I haven't had the time yet. In the meantime, I would suggest you have a look at the Modelica.Fluid library and/or this presentation by one of its authors, Francesco Casella.
That connector is not a physical connector. You need one flow variable for each potential variable. This is the OpenModelica error message if it helps a little:
Warning: Connector .Incompressible is not balanced: The number of potential variables (2) is not equal to the number of flow variables (1).
Error: Too many equations, over-determined system. The model has 4 equation(s) and 3 variable(s).
Error: Internal error Found Equation without time dependent variables outlet.V_dot = V_dot
This is because the unconnected connector will generate one equation for the flow:
outlet.V_dot = 0.0;
This means outlet.V_dot is replaced in:
outlet.V_dot = V_dot;
And you get:
0.0 = V_dot;
But V_dot is a parameter and can not be assigned to in an equation section (needs an initial equation if the parameter has fixed=false, or a binding equation in the default case).
I understand what the concept of currying is, and know how to use it. These are not my questions, rather I am curious as to how this is actually implemented at some lower level than, say, Haskell code.
For example, when (+) 2 4 is curried, is a pointer to the 2 maintained until the 4 is passed in? Does Gandalf bend space-time? What is this magic?
Short answer: yes a pointer is maintained to the 2 until the 4 is passed in.
Longer than necessary answer:
Conceptually, you're supposed to think about Haskell being defined in terms of the lambda calculus and term rewriting. Lets say you have the following definition:
f x y = x + y
This definition for f comes out in lambda calculus as something like the following, where I've explicitly put parentheses around the lambda bodies:
\x -> (\y -> (x + y))
If you're not familiar with the lambda calculus, this basically says "a function of an argument x that returns (a function of an argument y that returns (x + y))". In the lambda calculus, when we apply a function like this to some value, we can replace the application of the function by a copy of the body of the function with the value substituted for the function's parameter.
So then the expression f 1 2 is evaluated by the following sequence of rewrites:
(\x -> (\y -> (x + y))) 1 2
(\y -> (1 + y)) 2 # substituted 1 for x
(1 + 2) # substituted 2 for y
3
So you can see here that if we'd only supplied a single argument to f, we would have stopped at \y -> (1 + y). So we've got a whole term that is just a function for adding 1 to something, entirely separate from our original term, which may still be in use somewhere (for other references to f).
The key point is that if we implement functions like this, every function has only one argument but some return functions (and some return functions which return functions which return ...). Every time we apply a function we create a new term that "hard-codes" the first argument into the body of the function (including the bodies of any functions this one returns). This is how you get currying and closures.
Now, that's not how Haskell is directly implemented, obviously. Once upon a time, Haskell (or possibly one of its predecessors; I'm not exactly sure on the history) was implemented by Graph reduction. This is a technique for doing something equivalent to the term reduction I described above, that automatically brings along lazy evaluation and a fair amount of data sharing.
In graph reduction, everything is references to nodes in a graph. I won't go into too much detail, but when the evaluation engine reduces the application of a function to a value, it copies the sub-graph corresponding to the body of the function, with the necessary substitution of the argument value for the function's parameter (but shares references to graph nodes where they are unaffected by the substitution). So essentially, yes partially applying a function creates a new structure in memory that has a reference to the supplied argument (i.e. "a pointer to the 2), and your program can pass around references to that structure (and even share it and apply it multiple times), until more arguments are supplied and it can actually be reduced. However it's not like it's just remembering the function and accumulating arguments until it gets all of them; the evaluation engine actually does some of the work each time it's applied to a new argument. In fact the graph reduction engine can't even tell the difference between an application that returns a function and still needs more arguments, and one that has just got its last argument.
I can't tell you much more about the current implementation of Haskell. I believe it's a distant mutant descendant of graph reduction, with loads of clever short-cuts and go-faster stripes. But I might be wrong about that; maybe they've found a completely different execution strategy that isn't anything at all like graph reduction anymore. But I'm 90% sure it'll still end up passing around data structures that hold on to references to the partial arguments, and it probably still does something equivalent to factoring in the arguments partially, as it seems pretty essential to how lazy evaluation works. I'm also fairly sure it'll do lots of optimisations and short cuts, so if you straightforwardly call a function of 5 arguments like f 1 2 3 4 5 it won't go through all the hassle of copying the body of f 5 times with successively more "hard-coding".
Try it out with GHC:
ghc -C Test.hs
This will generate C code in Test.hc
I wrote the following function:
f = (+) 16777217
And GHC generated this:
R1.p[1] = (W_)Hp-4;
*R1.p = (W_)&stg_IND_STATIC_info;
Sp[-2] = (W_)&stg_upd_frame_info;
Sp[-1] = (W_)Hp-4;
R1.w = (W_)&integerzmgmp_GHCziInteger_smallInteger_closure;
Sp[-3] = 0x1000001U;
Sp=Sp-3;
JMP_((W_)&stg_ap_n_fast);
The thing to remember is that in Haskell, partially applying is not an unusual case. There's technically no "last argument" to any function. As you can see here, Haskell is jumping to stg_ap_n_fast which will expect an argument to be available in Sp.
The stg here stands for "Spineless Tagless G-Machine". There is a really good paper on it, by Simon Peyton-Jones. If you're curious about how the Haskell runtime is implemented, go read that first.
I've written an experimental function evaluator that allows me to bind simple functions together such that when the variables change, all functions that rely on those variables (and the functions that rely on those functions, etc.) are updated simultaneously. The way I do this is instead of evaluating the function immediately as it's entered in, I store the function. Only when an output value is requested to I evaluate the function, and I evaluate it each and every time an output value is requested.
For example:
pi = 3.14159
rad = 5
area = pi * rad * rad
perim = 2 * pi * rad
I define 'pi' and 'rad' as variables (well, functions that return a constant), and 'area' and 'perim' as functions. Any time either 'pi' or 'rad' change, I expect the results of 'area' and 'perim' to change in kind. Likewise, if there were any functions depending on 'area' or 'perim', the results of those would change as well.
This is all working as expected. The problem here is when the user introduces recursion - either accidental or intentional. There is no logic in my grammar - it's simply an evaluator - so I can't provide the user with a way to 'break out' of recursion. I'd like to prevent it from happening at all, which means I need a way to detect it and declare the offending input as invalid.
For example:
a = b
b = c
c = a
Right now evaluating the last line results in a StackOverflowException (while the first two lines evaluate to '0' - an undeclared variable/function is equal to 0). What I would like to do is detect the circular logic situation and forbid the user from inputing such a statement. I want to do this regardless of how deep the circular logic is hidden, but I have no idea how to go about doing so.
Behind the scenes, by the way, input strings are converted to tokens via a simple scanner, then to an abstract syntax tree via a hand-written recursive descent parser, then the AST is evaluated. The language is C#, but I'm not looking for a code solution - logic alone will be fine.
Note: this is a personal project I'm using to learn about how parsers and compilers work, so it's not mission critical - however the knowledge I take away from this I do plan to put to work in real life at some point. Any help you guys can provide would be appreciated greatly. =)
Edit: In case anyone's curious, this post on my blog describes why I'm trying to learn this, and what I'm getting out of it.
I've had a similar problem to this in the past.
My solution was to push variable names onto a stack as I recursed through the expressions to check syntax, and pop them as I exited a recursion level.
Before I pushed each variable name onto the stack, I would check if it was already there.
If it was, then this was a circular reference.
I was even able to display the names of the variables in the circular reference chain (as they would be on the stack and could be popped off in sequence until I reached the offending name).
EDIT: Of course, this was for single formulae... For your problem, a cyclic graph of variable assignments would be the better way to go.
A solution (probably not the best) is to create a dependency graph.
Each time a function is added or changed, the dependency graph is checked for cylces.
This can be cut short. Each time a function is added, or changed, flag it. If the evaluation results in a call to the function that is flagged, you have a cycle.
Example:
a = b
flag a
eval b (not found)
unflag a
b = c
flag b
eval c (not found)
unflag b
c = a
flag c
eval a
eval b
eval c (flagged) -> Cycle, discard change to c!
unflag c
In reply to the comment on answer two:
(Sorry, just messed up my openid creation so I'll have to get the old stuff linked later...)
If you switch "flag" for "push" and "unflag" for "pop", it's pretty much the same thing :)
The only advantage of using the stack is the ease of which you can provide detailed information on the cycle, no matter what the depth. (Useful for error messages :) )
Andrew