I know and have fun as often as I can with Florent Pillet, another member of the tribe of "dinosaurs" still kicking around.

I really like one of his projects that contributed to his notoriety : NSLogger. Logging has always been a pain in the neck, and this tool provided us all with a way to get it done efficiently and properly. The first commit on the github repo is from 2010, and I have a strong suspicion it's been in production since before that in one form or another.

Anyhoo, I like Florent, I  like NSLogger, but I hate what Cocoapods (and to a lesser extent Carthage) do to my projects. It's too brittle and I strongly dislike things that mess around with the extremely complicated XML that is a pbxproj. They do however serve an admirable purpose: managing dependencies in a way that doesn't require me to use git submodules in every one of my projects.

So, I rarely use NSLogger. SHAME! SHAME! <insert your own meme here>

With the advent of (and subsequent needed updates to) Swift Package Manager, we now have an official way of managing and supporting dependencies, but it has its own quirks that appently make it hard to "SPM" older projects.

Let's see what we can do about NSLogger.

##### Step 1 : The Project Structure

SPM can't mix Obj-C code and Swift code. It's always been pretty hacky anyways, with the bridging headers and the weird steps hidden by the toolchain, so we need to make it explicit:

• One target for the Objective-C code (imaginatively named NSLoggerLibObjC)
• One target for the Swift code (NSLogger) that depends on NSLoggerLibObjC
• One product that builds the Swift target

One of the problems is that all that code is mixed in the folders, because Xcode doesn't care about file placement. SPM, on the other hand does.

So, let's use and abuse the path and sources parameters of the target. The first one is to provide the root where we look for files to compile, and the second one lists the files to be compiled.

• LoggerClient.m for NSLoggerLibObjC
• NSLogger.swift for NSLogger

Done. Right?

Not quite.

##### Step 2 : Compilation Quirks

The Obj-C lib requires ARC to be disabled. Easy to do in Xcode, a bit harder in SPM.

We need to pass the -fno-objc-arc flag to the compiler. SPM doesn't make it easy or obvious to do that, for a variety of reasons, but I guess mostly because you shouldn't pass compiler flags at all in an ideal world.

But (especially in 2020), looking at the world, ideal it ain't.

We have to use the (not so aptly named) cSetting option of the target, and use the very scary CSetting.unsafeFlags parameter for that option. Why is it unsafe, you might ask? Weeeeeeeeell. It's companies' usual way of telling you "you're on your own with this one". I'm fine with that.

Another compilation quirk is that Obj-C code relies (like its ancestor, C) on the use of header files to make your code usable as a dependency.

Again, because Xcode and SPM treat the file structure very differently, just saying that every header should be included in the resulting library is a bad idea: the search is recursive and in this particular case, would result in having specific iOS or MacOS (yes, capitalized, because sod that change) test headers exposed as well.

In the end, I had to make the difficult choice of doing something super ugly:

• move the public headers in their own directory
• use symlinks to their old place so's not to break the other parts of the project

If anyone has a better option that's not heavily more disruptive to the organization of the project, I'm all ears.

##### Step 3 : Final Assembly

So we have the Swift target that depends on the Obj-C one. Fine. But how do we use that dependency?

"Easy" some will exclaim (a bit too rapidly) "you just import the lib in the swift file!"

Yes, but then it breaks the other projects, which, again, we don't want to do. Minimal impact changes. Legacy. Friend.

So we need a preprocessing macro, like, say, SPMBuild, which would indicate we're building with SPM rather than Xcode. Sadly, this doesn't exist, and given the rate of change of the toolchain, I don't want to rely too heavily on the badly documented Xcode proprocessor macros that would allow me to detect a build through the IDE.

Thankfully, in the same vein as cSettings, we have a swiftSettings parameter to our target, wich supports SwiftSetting.define options. Great, so I'll define a macro, and test its existence in the swift file before importing the Obj-C part of the project.

One last thing I stumbled upon and used despite its shady nature: there is an undocumented decorator for import named @_exported which seems extraneous here, but has some interesting properties: it kinda sorta exposes what you import as part of the current module, flattening the dependency graph.

To be honest, I didn't know about it, it amused me, so I included it.

##### Wrap Up

In order to make it work directly from the repo, rather than locally, I also had to provide a version number. I chose to go with the next patch number instead of aggrandizing myself with a minor or even a major version.

Hopefully, these changes don't impact the current project at all, and allows me to use it in a way I like better (and is officially supported), and I hope Florent will not murder me for all of that. He might even decide to accept my pull request. We'll see.

In the meantime, you can find all the changes above and a usable SPM package in my fork.

#### Unit testing is painful amirite?

Writing good tests for your code very often means spending twice as much time coding them than on the things you test themselves.

It is good practice though to verify as much as possible that the code you write is valid, especially if that code is going to be public or included in someone else's work.

In my workflow I insist on the notion of ownership :

The bottomline for me is this: if there are several people on a project, I want clearly defined ownership. It's not that I won't fix a bug in someone else's code, just that they own it and therefore have to have a reliable way of testing that my fix works.
Tests solve part of that problem. My code, my tests. If you fix my code, run my tests, I'm fairly confident that you didn't wreck the whole thing. And that I won't have to spend a couple of hours figuring out what it is that you did.

This a a very very very light constraint when you compare it to methodologies like TDD, but it's a required minimum for me.

Plus, it's not that painful, except...

#### Testing every case

In my personal opinion, the tests that are hardest to do right are the ones that have a very large input range, with a few failure/continuity points.

If, for instance, and completely randomly, of course, you had an application where the tilt of the phone changes the state of the app (locked/unlocked, depending on whether the phone is lying flat-ish on the table or not:

• from -20º to 20º the app is locked
• from 160º to 200º the app is locked
• the rest of the time it's not locked
• All of that modulo 360, of course

So you have a function that takes the current pitch angle, and returns if we should lock or not:

func pitchLock(_ angle: Double) -> Bool {
// ...
}

Does it work? Does it work modulo 360? What would a unit test for that function even look like? A for loop?

I have been looking for a way to do that kind of test for a while, which is why I published HoledRange (now Domains 😇) a while back, as part of my hacks.

What I wanted is to write my tests kind of like this (invalid code on so many levels):

for x in [-1000.0...1000.0].randomSelection {
let unitCircleAngle = x%360.0
if unitCircleAngle >= 340 || unitCircle <= 20 {
XCTAssert(pitchLock(x))
} else if unitCircleAngle >= 160 && unitCircle <= 200 {
XCTAssert(pitchLock(x))
} else {
XCTAssertFalse(pitchLock(x))
}
}

This way of testing, while vaguely valid, leaves so many things flaky:

• how many elements in the random selection?
• how can we make certain values untestable (because we address them somewhere else, for instance)
• what a lot of boilerplate if I have multiple functions to test on the same range of values
• I can't reuse the same value for multiple tests to check function chains

#### Function builders

I have been fascinated with @_functionBuilder every since it was announced. While I don't feel enthusiastic about SwiftUI (in french), that way to build elements out of blocks is something I have wanted for years.

Making them is a harrowing experience the first time, but in the end it works!

What I wanted to use as syntax is something like this:

func myPlus(_ a: Int, _ b: Int) -> Int

DomainTests<Int> {
Domain(-10000...10000)
1000000
Test { (a: Int) in
XCTAssert(myPlus(a, 1) == a+1, "Problem with value\(a)")
XCTAssert(myPlus(1, a) == a+1, "Problem with value\(a)")
}
Test { (a: Int) in
let random = Int.random(in: -10000...10000)
XCTAssert(myPlus(a, random) == a+random, "Problem with value\(a)")
XCTAssert(myPlus(random, a) == a+random, "Problem with value\(a)")
}
}.random()

This particular DomainTests runs 1000000 times over $$D=[-10000;10000]$$ in a random fashion.

Note the Test builder that takes a function with a parameter that will be in the domain, and the definition that allows to define both the test domain (mandatory) and the number of random iterations (optional).

If you want to test every single value in a domain, the bounding needs to be Strideable, ie usable in a for-loop.

DomainTests<Int> {
Domain(-10000...10000)
Test { (a: Int) in
XCTAssert(myPlus(a, 1) == a+1, "Problem with value\(a)")
XCTAssert(myPlus(1, a) == a+1, "Problem with value\(a)")
}
Test { (a: Int) in
let random = Int.random(in: -10000...10000)
XCTAssert(myPlus(a, random) == a+random, "Problem with value\(a)")
XCTAssert(myPlus(random, a) == a+random, "Problem with value\(a)")
}
}.full()

#### Conclusion

A couple of hard working days plus a healthy dose of using that framework personally means this should be ready-ish for production.

If you are a maths-oriented dev and shiver at the idea of untested domains, this is for you 😬

This is the last part of a 3-parts series. In part 1, I tried to make sense of how it works and what we are trying to achieve, and in part 2, we set up the training loop.

#### Model Predictions

We have a trained model. Now what?

Remember, a model is a series of giant matrices that take an input like you trained it on, and spits out the list of probabilities associated with the outputs you trained it on. So all you have to do is feed it a new input and see what it tells you:

let input = [1.0, 179.0, 115.0]
let unlabeled : Tensor<Float> = Tensor<Float>(shape: [1, 3], scalars: input)
let predictions = model(unlabeled)
let logits = predictions[0]
let classIdx = logits.argmax().scalar! // we take only the best guess
print(classIdx)
17

Cool.

Cool, cool.

What?

Models deal with numbers. I am the one who assigned numbers to words to train the model on, so I need a translation layer. That's why I kept my contents structure around: I need it for its vocabulary map.

The real code:

let w1 = "on"
let w2 = "flocks"
let w3 = "settlement"

var indices = [w1, w2, w3].map {
let step2 = rnn2.callAsFunction(step1).differentiableMap({ $0.cell }) let last = withoutDerivative(at:step2[0]) let red = step2.differentiableReduce(last, { (p,e) -> Tensor<Float> in return e }) return red } @differentiable func callAsFunction(_ input: Tensor<Float>) -> Tensor<Float> { let step2out = runThrough(input) let step3 = matmul(step2out, correction) return step3 } } The RNN/LTSM have been talked about, but what are these two functions? callAsFunction is the only one needed. I have just decided to split the algorithm in two: the part where I "just" pass through layers, and the part where I format the output. Everything in runThrough could be written at the top of callAsFunction. We follow the steps outlined previously, it all seems logical, even if the details aren't quite clear yet. What is it with the @noDerivative and @differentiable annotations? Because we are dealing with a structure (model, layer, etc...) that not only should but will be adjusted over time, it is a way to tell the system which parts are important to that adjustment: • all properties except those maked as not derivative will be nudged potentially, so it makes sense to mark the number of inputs as immutable, and the rest as "nudgeable" • all the functions that calculate something that will be used in the "nudging" need to have specific maths properties that make the change non-random. We need to know where we are and where we were going. We need a position, and a speed, we need a value and its derivative Ugh, maths. Yeah. I am obviously oversimplifying everything to avoid scaring away everyone from the get go, but the idea should make sense if you look at it this way: • Let's take a blind man trying to shoot an arrow at a target • You ask them to shoot and then you'll correct them based on where the arrow lands • It hits the far left of the target • You tell them to nudge the aim to the right • The problem is that "more right" isn't enough information... You need to tell them to the right a little (new position and some information useful for later, you'll see) • The arrow lands slightly to the right of the center • You tell the archer to aim to the left but less than their movement they just made to the right. Two pieces of information: one relative to a direction, and one relative to the rate of change. The other name of the rate of change is the derivative. Standard derivatives are speed to position (we are here, now we are there, and finally we are there, and the rate of change slowed, so the next position won't be as far from this one as the one was to the previous one), or acceleration to speed (when moving, if your speed goes up and up and up, you have a positive rate of change, you accelerate). That's why passing through a layer should preserve the two: the actual values, and the speed at which we are changing them. Hence the @differentiable annotation. (NB for all you specialists in the field reading that piece... yes I know. I'm trying to make things more palatable) "But wait", say the most eagle-eyed among you, "I can see a withoutDerivative in that code!" Yes. RNN is peculiar in the way that it doesn't try to coerce the dimensions of the results. It spits out all the possible variants it has calculated. But in practice, we need only the last one. Taking one possible outcome out of many cancels out the @differentiable nature of the function, because we actually lose some information. This is why we only partially count on the RNN's hidden parameters to give us a "good enough" result, and need to incorporate extra weights and biases that are derivable. The two parts of the correction matrix, will retain the nudge speed, as well as reshape the output matrix to match the labels: matrix addition and multiplications are a bit beyond the scope here as well (and quite frankly a bit boring), but that last step ( step3 in the code ) basically transform a 512x512x<number of labels> matrix, into a 2x<numbers of labels> : one column to give us the final probabilities, one for each possible label. If you've made it this far, congratulations, you've been through the hardest. #### Model Training OK, we have the model we want to use to represent the various orders in the data, how do we train it? To continue with the blind archer metaphor, we need the piece of code that acts as the "corrector". In ML, it's called the optimizer. We need to give it what the archer is trying to do, and a way to measure how far off the mark the archer is, and a sense of how stern it should be (do we do a lot of small corrections, or fewer large ones?) The measure of the distance is called the "cost" function, or the "accuracy" function. Depending on how we look at it we want to make the cost (or error) as low as possible, and the accuracy as high as possible. They are obviously linked, but can be expressed in different units ("you are 3 centimeters off" and "you are closer by 1%"). Generally, loss has little to no meaning outside of the context of the layers ( is 6 far? close? because words aren't sorted in any meaningful way, we are 6.2 words away from the ideal word doesn't mean much), while accuracy is more like a satisfaction percentage (we are 93% satisfied with the result, whatever that means). func accuracy(predictions: Tensor<Int32>, truths: Tensor<Int32>) -> Float { return Tensor<Float>(predictions .== truths).mean().scalarized() } let predictions = model(aBunchOfFeatures) print("Accuracy: \(accuracy(predictions: predictions.argmax(squeezingAxis: 1), truths: aBunchOfLabels))") Accuracy: 0.10143079 and the loss: let predictions = model(aBunchOfFeatures) let loss = softmaxCrossEntropy(logits: predictions, labels: aBunchOfLabels) print("Loss test: \(loss)")  Loss test: 6.8377414 In more human terms, the best prediction we have is 10% satisfying, because the result is 6.8 words away from the right one. 😬 Now that we know how to measure how far off the mark we are (in two different ways), we need to make a decision about 3 things: • Which kind of optimizer we want to use (we'll use Adam, it's a good algorithm for our problem, but other ones exist. For our archer metaphor, it's a gentle but firm voice on the corrections, rather than a barking one that might progress rapidly at first then annoy the hell out of the archer) • What learning rate we want to use (do we correct a lot of times in tiny increments, or in bigger increments that take overall less time, but might overcorrect) • How many tries we give the system to get as close as possible You can obviously see why the two last parameters are hugely important, and very hard to figure out. For some problems, it might be better to use big steps in case we find ourselves stuck, for others it might be better to always get closer to the target but by smaller and smaller increments. It's an art, honestly. Here, I've used a learning rate of 0.001 (tiny) and a number of tries of 500 (medium), because if there is no way to figure out the structure of the text, I want to know it fast (fewer steps), but I do NOT want to overshoot(small learning rate). Let's setup the model, the correction matrices, and the training loop: var weigths = Tensor<Float>(randomNormal: [512, contents.vocabulary.count]) // random probabilities var biases = Tensor<Float>(randomNormal: [contents.vocabulary.count]) // random bias var model = TextModel(input:3, hidden: 512, output: contents.vocabulary.count, weights: weigths, biases: biases) Now let's setup the training loop and run it: let epochCount = 500 var trainAccuracyResults: [Float] = [] var trainLossResults: [Float] = [] var randomSampleSize = contents.original.count/15 var randomSampleCount = contents.original.count / randomSampleSize print("Doing \(randomSampleCount) samples per epoch") for epoch in 1...epochCount { var epochLoss: Float = 0 var epochAccuracy: Float = 0 var batchCount: Int = 0 for training in contents.randomSample(splits: randomSampleCount) { let (sampleFeatures,sampleLabels) = training let (loss, grad) = model.valueWithGradient { (model: TextModel) -> Tensor<Float> in let logits = model(sampleFeatures) return softmaxCrossEntropy(logits: logits, labels: sampleLabels) } optimizer.update(&model, along: grad) let logits = model(sampleFeatures) epochAccuracy += accuracy(predictions: logits.argmax(squeezingAxis: 1), truths: sampleLabels) epochLoss += loss.scalarized() batchCount += 1 } epochAccuracy /= Float(batchCount) epochLoss /= Float(batchCount) trainAccuracyResults.append(epochAccuracy) trainLossResults.append(epochLoss) if epoch % 10 == 0 { print("avg time per epoch: \(t.averageDeltaHumanReadable)") print("Epoch \(epoch): Loss: \(epochLoss), Accuracy: \(epochAccuracy)") } } A little bit of explanation: • We will try 500 times ( epochCount ) • At each epoch, I want to test and nudge for 15 different combinations of trigrams. Why? because it avoids the trap of optimizing for some specific turns of phrase • We apply the model to the sample, calculate the loss, and the derivative, and update the model with where we calculate we should go next What does that give us? Doing 15 samples per epoch Epoch 10: Loss: 6.8377414, Accuracy: 0.10143079 Epoch 20: Loss: 6.569199, Accuracy: 0.10564535 Epoch 30: Loss: 6.412607, Accuracy: 0.10802801 Epoch 40: Loss: 6.2550464, Accuracy: 0.10751916 Epoch 50: Loss: 6.0366735, Accuracy: 0.11123683 ... Epoch 490: Loss: 1.1177399, Accuracy: 0.73812264 Epoch 500: Loss: 0.5172857, Accuracy: 0.86724746 We like to keep these values in an array to graph them. What does it look like? We can see that despite the dips and spikes, because we change the samples often and don't try any radical movement, we tend to better and better results. We don't get stuck in a ditch. Next part, we'll see how to use the model. Here's a little spoiler: I asked it to generate some random text: on flocks settlement or their enter the earth; their only hope in their arrows, which for want of it, with a thorn. and distinction of their nature, that in the same yoke are also chosen their chiefs or rulers, such as administer justice in their villages and by superstitious awe in times of old. It's definitely gibberish when you look closely, but from a distance it looks kind of okayish for a program that learned to speak entirely from scratch, based on a 10k words essay written by fricking Tacitus. To get your degree in <insert commerce / political school name here>, there is a last exam in which you need to talk with a jury of teachers. The rule is simple, if the student is stumped or hesitates, the student has failed. If the student manages to last the whole time, or manages to stump the jury or makes it hesitate, the student passes. This particular student was having a conversation about geography, and a juror thought to stump the candidate by asking "what is the depth of <insert major river here>?" to which the student, not missing a beat answered "under which bridge?", stumping the juror. Old student joke/legend Programming is part of the larger tree of knowledge we call computer science. Everything we do has its roots in maths and electronics. Can you get by with shoddy reasoning and approximate "that should work" logic? Sure. But in the same way you can "get by" playing the piano using only the index finger of your hands. Being able to play chopsticks makes you as much of a pianist as being able to copy/paste stackoverflow answers makes you a programmer/developer. The problem is that in my field, the end-user (or client, or "juror", or "decision maker") is incapable of distinguishing between chopsticks and Brahms, not because of a lack of interest, but because we, as a field, have become experts at stumping them. As a result, we have various policies along the lines of "everyone should learn to code" being implemented worldwide, and I cynically think it's mostly because the goal is to stop getting milked by so-called experts that can charge you thousands of monies for the chopsticks equivalent of a website. To me, the problem doesn't really lie with the coding part. Any science, any technical field, requires a long formation to become good at. Language proficiency, musical instruments, sports, dancing, driving, sailing, carpentry, mechanical engineering, etc... It's all rather well accepted that these fields require dedication and training. But somehow, programming should be "easy", or "intuitive". That's not to say I think it should be reserved to an elite. These other fields aren't. I have friends who got extremely good at guitars by themselves, and sports are a well known way out of the social bog. But developers seem to be making something out of nothing. They "just" sit down and press keys on a board and presto, something appears and they get paid. It somehow seems unfair, right? There are two aspects to this situation: the lack of nuanced understanding on the person who buys the program, and the overly complicated/flaky way we programmers handle all this. I've already painted with a very broad brush what we developers feel about this whole "being an industry" thing. So what's the issue on the other side? If you ask most customers (and students), they respond "obfuscation" or a variant of it. In short, we use jargon, technobabble, which they understand nothing of, and are feeling taken advantage of when we ask for money. This covers the whole gamut from "oh cool, they seem to know what they are talking about, so I will give them all my money" to "I've been burned by smart sounding people before, I don't trust them anymore", to "I bet I can do it myself in under two weeks", to "the niece of the mother of my friend is learning to code and she's like 12, so I'll ask her instead". So, besides reading all of Plato's work on dialectic and how to get at the truth through questions, how does one differentiate between a$500 website and a $20000 one? Especially if they look the same? Well, in my opinion as a teacher, for which I'm paid to sprinkle knowledge about computer programming onto people, there are two important things to understand about making software to evaluate the quality of a product: • Programming is exclusively about logic. The difficulty (and the price) scales in regards to the logic needed to solve whatever problem we are hired to solve • We very often reuse logic from other places and combine those lines of code with ours to refine the solution Warning triggers that make me think the person is trying to sell me magic pixie dust include: • The usual bullshit-bingo: if they try to include as many buzzwords (AI, machine learning, cloud, big data, blockchain,...) as possible in their presentation, you have to ask very pointed question about your problem, and how these things will help you solve it • If they tell you they have the perfect solution for you even though they asked no question, they are probably trying to recycle something they have which may or may not work for your issues A word of warning though: prices in absolute aren't a factor at all. In the same way that you'd probably pay quite naturally a whole lot more money for a bespoke dinner table that is exactly what you envision in your dreams than the one you can get in any furniture store, your solution cannot be cheaper than off-the-shelf. Expertise and tailoring cannot be free. Balking at the price when you have someone who genuinely is an expert in front of you, and after they announced their price is somewhat insulting. How often do you go to the bakery and ask the question "OK, so your cake is really good, and all my friends recommend it, and I know it's made with care, but, like,$30 is way too  expensive... how about \$15?"

I have also left aside the question  of visual design. it's not my field, I suck at it, and I think that it  is an expert field too, albeit more on the "do I like it?" side of the  equation than the "does it work?" one, when it comes to estimating its  value. It's like when you buy a house: there are the foundations, and  the walls, and the roof, and their job is to answer the question "will I  still be protected from the outside weather in 10 years?", whereas the  layout, the colors of the walls, and the furniture are the answer to the  question "will I still feel good in this place in 10 years?".  Thing is, with software development as well, you can change the visuals  to a certain extent (up to the point when you need to change the  position of the walls, to continue with the metaphor), but it's hard to  change the foundations.