Skip to content
code + contents of my website, and programming life
Branch: master
Clone or download
Type Name Latest commit message Commit time
Failed to load latest commit information.
content restructured blog Jan 16, 2019
static add pictures and text Dec 30, 2017
.gitignore removed tags Oct 31, 2018 change script to directly edit blog/ Apr 12, 2019 Add link to encoding CPS using coro Jun 15, 2019
config.toml wrote about why I like algebra over analysis Oct 16, 2018 build docs and change scripts to use docs/ folder Apr 9, 2019 add blog posts Jan 28, 2018 add lifelong todo list Apr 19, 2019 build docs and change scripts to use docs/ folder Apr 9, 2019

Contents of, mirrored at

The idea is for the website to contain blog posts, along with visualizations of math / graphics / programming.

The former has been semi-forced thanks to GSoC, as for the latter, it remains to be seen. I'm hopeful, though :)

Ideas I stumble onto

Representing CPS in LLVM using the @coro.* intrinsics

This is part of a larger thread --- Adding CPS call support to LLVM where there is a large discussion on the correct design of how to teach LLVM about CPS.

Gor Nishanov proided the above example of encoding CPS using the llvm coro instructions.

Bug in the LLVM code generator: Lowering of MO_Add2 and MO_AddWordC

Both of these are lowered the same way, but they should be different.

In particular, GHC.Prim explains:

Honestly, this is confusing, but I guess there's some story to having two separate primops for this?

Everything you know about word2vec is wrong.

The classic explanation of word2vec, in skip-gram, with negative sampling, in the paper and countless blog posts on the internet is as follows:

while(1) {
   1. vf = vector of focus word
   2. vc = vector of focus word
   3. train such that (vc . vf = 1)
   4. for(0 <= i <= negative samples):
           vneg = vector of word *not* in context
           train such that (vf . vneg = 0)

Indeed, if I google "word2vec skipgram", the results I get are:

The original word2vec C implementation does not do what's explained above, and is drastically different. Most serious users of word embeddings, who use embeddings generated from word2vec do one of the following things:

  1. They invoke the original C implementation directly.
  2. They invoke the gensim implementation, which is transliterated from the C source to the extent that the variables names are the same.

Indeed, the gensim implementation is the only one that I know of which is faithful to the C implementation.

The C implementation

The C implementation in fact maintains two vectors for each word, one where it appears as a focus word, and one where it appears as a context word. (Is this sounding familiar? Indeed, it appears that GloVe actually took this idea from word2vec, which has never mentioned this fact!)

The setup is incredibly well done in the C code:

  • An array called syn0 holds the vector embedding of a word when it occurs as a focus word. This is random initialized.
  for (a = 0; a < vocab_size; a++) for (b = 0; b < layer1_size; b++) {
    next_random = next_random * (unsigned long long)25214903917 + 11;
    syn0[a * layer1_size + b] = 
       (((next_random & 0xFFFF) / (real)65536) - 0.5) / layer1_size;
  • Another array called syn1neg holds the vector of a word when it occurs as a context word. This is zero initialized.
for (a = 0; a < vocab_size; a++) for (b = 0; b < layer1_size; b++)
  syn1neg[a * layer1_size + b] = 0;
  • During training (skip-gram, negative sampling, though other cases are also similar), we first pick a focus word. This is held constant throughout the positive and negative sample training. The gradients of the focus vector are accumulated in a buffer, and are applied to the focus word after it has been affected by both positive and negative samples.
if (negative > 0) for (d = 0; d < negative + 1; d++) {
  // if we are performing negative sampling, in the 1st iteration,
  // pick a word from the context and set the dot product target to 1
  if (d == 0) {
    target = word;
    label = 1;
  } else {
    // for all other iterations, pick a word randomly and set the dot
    //product target to 0
    next_random = next_random * (unsigned long long)25214903917 + 11;
    target = table[(next_random >> 16) % table_size];
    if (target == 0) target = next_random % (vocab_size - 1) + 1;
    if (target == word) continue;
    label = 0;
  l2 = target * layer1_size;
  f = 0;

  // find dot product of original vector with negative sample vector
  // store in f
  for (c = 0; c < layer1_size; c++) f += syn0[c + l1] * syn1neg[c + l2];

  // set g = sigmoid(f) (roughly, the actual formula is slightly more complex)
  if (f > MAX_EXP) g = (label - 1) * alpha;
  else if (f < -MAX_EXP) g = (label - 0) * alpha;
  else g = (label - expTable[(int)((f + MAX_EXP) * (EXP_TABLE_SIZE / MAX_EXP / 2))]) * alpha;

  // 1. update the vector syn1neg,
  // 2. DO NOT UPDATE syn0
  // 3. STORE THE syn0 gradient in a temporary buffer neu1e
  for (c = 0; c < layer1_size; c++) neu1e[c] += g * syn1neg[c + l2];
  for (c = 0; c < layer1_size; c++) syn1neg[c + l2] += g * syn0[c + l1];
// Finally, after all samples, update syn1 from neu1e
// Learn weights input -> hidden
for (c = 0; c < layer1_size; c++) syn0[c + l1] += neu1e[c];

Why random and zero initialization?

Once again, since none of this actually explained in the original papers or on the web, I can only hypothesize.

My hypothesis is that since the negative samples come from all over the text and are not really weighed by frequency, you can wind up picking any word, and more often than not, a word whose vector has not been trained much at all. If this vector actually had a value, then it could move the actually important focus word randomly.

The solution is to set all negative samples to zero, so that only vectors that have occured somewhat frequently will affect the representation of another vector.

It's quite ingenious, really, and until this, I'd never really thought of how important initialization strategies really are.

Why I'm writing this

I spent two months of my life trying to reproduce word2vec, following the paper exactly, reading countless articles, and simply not succeeding. I was unable to reach the same scores that word2vec did, and it was not for lack of trying.

I could not have imagined that the paper would have literally fabricated an algorithm that doesn't work, while the implementation does something completely different.

Eventually, I decided to read the sources, and spent three whole days convinced I was reading the code wrong since literally everything on the internet told me otherwise.

I don't understand why the original paper and the internet contain zero explanations of the actual mechanism behind word2vec, so I decided to put it up myself.

This also explains GloVe's radical choice of having a separate vector for the negative context --- they were just doing what word2vec does, but they told people about it :).

Is this academic dishonesty? I don't know the answer, and that's a heavy question. But I'm frankly incredibly pissed, and this is probably the last time I take a machine learning paper's explanation of the algorithm seriously again --- from next time, I read the source first.

Hamiltonian monte carlo, leapfrog integrators, and sympletic geometry

This is a section that I'll update as I learn more about the space, since I'm studying differential geometry over the summer, I hope to know enough about "sympletic manifolds". I'll make this an append-only log to add to the section as I understand more.

31st May
  • To perform hamiltonian monte carlo, we use the hamiltonian and its derivatives to provide a momentum to our proposal distribution --- That is, when we choose a new point from the current point, our probability distribution for the new point is influenced by our current momentum

  • For some integral necessary within this scheme, Euler integration doesn't cut it since the error diverges to infinity

  • Hence, we need an integrator that guarantees that the energy of out system is conserved. Enter the leapfrog integrator. This integrator is also time reversible -- We can run it forward for n steps, and then run it backward for n steps to arrive at the same state. Now I finally know how Braid was implemented, something that bugged the hell out of 9th grade me when I tried to implement Braid-like physics in my engine!

  • The actual derivation of the integrator uses Lie algebras, Sympletic geometry, and other diffgeo ideas, which is great, because it gives me motivation to study differential geometry :)

  • Original paper: Construction of higher order sympletic integrators

Small Haskell MCMC implementation:

We create a simple monad called PL which allows for a single operation: sampling from a uniform distribution. We then exploit this to implement MCMC using metropolis hastings, which is used to sample from arbitrary distributions. Bonus is a small library to render sparklines in the CLI.

For next time:

  • Using applicative to speed up computations by exploiting parallelism
  • Conditioning of a distribution wrt a variable

Source code

{-# LANGUAGE GeneralizedNewtypeDeriving #-}
{-# LANGUAGE StandaloneDeriving #-}
{-# LANGUAGE FlexibleContexts #-}
{-# LANGUAGE FlexibleInstances #-}
{-# LANGUAGE UndecidableInstances #-}
{-# LANGUAGE DeriveFunctor #-}
import System.Random
import Data.List(sort, nub)
import Data.Proxy
import Control.Monad (replicateM)
import qualified Data.Map as M

-- | Loop a monadic computation.
mLoop :: Monad m =>
      (a -> m a) -- ^ loop
      -> Int -- ^ number of times to run
      -> a -- initial value
      -> m a -- final value
mLoop _ 0 a = return a
mLoop f n a = f a >>= mLoop f (n - 1)

-- | Utility library for drawing sparklines

-- | List of characters that represent sparklines
sparkchars :: String
sparkchars = "_▁▂▃▄▅▆▇█"

-- Convert an int to a sparkline character
num2spark :: RealFrac a => a -- ^ Max value
  -> a -- ^ Current value
  -> Char
num2spark maxv curv =
   sparkchars !!
     (floor $ (curv / maxv) * (fromIntegral (length sparkchars - 1)))

series2spark :: RealFrac a => [a] -> String
series2spark vs =
  let maxv = if null vs then 0 else maximum vs
  in map (num2spark maxv) vs

seriesPrintSpark :: RealFrac a => [a] -> IO ()
seriesPrintSpark = putStrLn . series2spark

-- Probabilites
-- ============
type F = Float
-- | probablity density
newtype P = P { unP :: Float } deriving(Num)

-- | prob. distributions over space a
newtype D a = D { runD :: a -> P }

uniform :: Int -> D a
uniform n =
  D $ \_ -> P $ 1.0 / (fromIntegral $ n)

(>$<) :: Contravariant f => (b -> a) -> f a  -> f b
(>$<) = cofmap

instance Contravariant D where
  cofmap f (D d) = D (d . f)

-- | Normal distribution with given mean
normalD :: Float ->  D Float
normalD mu = D $ \f -> P $ exp (- ((f-mu)^2))

-- | Distribution that takes on value x^p for 1 <= x <= 2.  Is normalized
polyD :: Float -> D Float
polyD p = D $ \f -> P $ if 1 <= f && f <= 2 then (f ** p) * (p + 1) / (2 ** (p+1) - 1) else 0

class Contravariant f where
  cofmap :: (b -> a) -> f a -> f b

data PL next where
    Ret :: next -> PL next -- ^ return  a value
    Sample01 :: (Float -> PL next) -> PL next -- ^ sample uniformly from a [0, 1) distribution

instance Monad PL where
  return = Ret
  (Ret a) >>= f = f a
  (Sample01 float2plnext) >>= next2next' =
      Sample01 $ \f -> float2plnext f >>= next2next'

instance Applicative PL where
    pure = return
    ff <*> fx = do
        f <- ff
        x <- fx
        return $ f x

instance Functor PL where
    fmap f plx = do
         x <- plx
         return $ f x

-- | operation to sample from [0, 1)
sample01 :: PL Float
sample01 = Sample01 Ret

-- | Run one step of MH on a distribution to obtain a (correlated) sample
mhStep :: (a -> Float) -- ^ function to score sample with, proportional to distribution
  -> (a -> PL a) -- ^ Proposal program
  -> a -- current sample
  -> PL a
mhStep f q a = do
 	a' <- q a
 	let alpha = f a' / f a -- acceptance ratio
 	u <- sample01
 	return $ if u <= alpha then a' else a

-- Typeclass that can provide me with data to run MCMC on it
class MCMC a where
    arbitrary :: a
    uniform2val :: Float -> a

instance MCMC Float where
	arbitrary = 0
	-- map [0, 1) -> (-infty, infty)
	uniform2val v = tan (-pi/2 + pi * v)

-- | Any enumerable object has a way to get me the starting point for MCMC
instance (Bounded a, Enum a) => MCMC a where
     arbitrary = toEnum 0
     uniform2val v = let
        maxf = fromIntegral . fromEnum $ maxBound
        minf = fromIntegral . fromEnum $ minBound
        in toEnum $ floor $ minf + v * (maxf - minf)

-- | Run MH to sample from a distribution
mh :: (a -> Float) -- ^ function to score sample with
 -> (a -> PL a) -- ^ proposal program
 -> a -- ^ current sample
 -> PL a
mh f q a = mLoop (mhStep f q) 100  $ a

-- | Construct a program to sample from an arbitrary distribution using MCMC
mhD :: MCMC a => D a -> PL a
mhD (D d) =
      scorer = (unP . d)
      proposal _ = do
        f <- sample01
        return $ uniform2val f
    in mh scorer proposal arbitrary

-- | Run the probabilistic value to get a sample
sample :: RandomGen g => g -> PL a -> (a, g)
sample g (Ret a) = (a, g)
sample g (Sample01 f2plnext) = let (f, g') = random g in sample g' (f2plnext f)

-- | Sample n values from the distribution
samples :: RandomGen g => Int -> g -> PL a -> ([a], g)
samples 0 g _ = ([], g)
samples n g pl = let (a, g') = sample g pl
                     (as, g'') = samples (n - 1) g' pl
                 in (a:as, g'')

-- | count fraction of times value occurs in list
occurFrac :: (Eq a) => [a] -> a -> Float
occurFrac as a =
    let noccur = length (filter (==a) as)
        n = length as
    in (fromIntegral noccur) / (fromIntegral n)

-- | Produce a distribution from a PL by using the sampler to sample N times
distribution :: (Eq a, Num a, RandomGen g) => Int -> g -> PL a -> (D a, g)
distribution n g pl =
    let (as, g') = samples n g pl in (D (\a -> P (occurFrac as a)), g')

-- | biased coin
coin :: Float -> PL Int -- 1 with prob. p1, 0 with prob. (1 - p1)
coin p1 = do
    Sample01 (\f -> Ret $ if f < p1 then 1 else 0)

-- | Create a histogram from values.
histogram :: Int -- ^ number of buckets
          -> [Float] -- values
          -> [Int]
histogram nbuckets as =
        minv :: Float
        minv = minimum as
        maxv :: Float
        maxv = maximum as
        -- value per bucket
        perbucket :: Float
        perbucket = (maxv - minv) / (fromIntegral nbuckets)
        bucket :: Float -> Int
        bucket v = floor (v / perbucket)
        bucketed :: M.Map Int Int
        bucketed = foldl (\m v -> M.insertWith (+) (bucket v) 1 m) mempty as
     in map snd . M.toList $ bucketed

printSamples :: (Real a, Eq a, Ord a, Show a) => String -> [a] -> IO ()
printSamples s as =  do
    putStrLn $ "***" <> s
    putStrLn $ "   samples: " <> series2spark (map toRational as)

printHistogram :: [Float] -> IO ()
printHistogram samples = putStrLn $ series2spark (map fromIntegral . histogram 10 $  samples)

-- | Given a coin bias, take samples and print bias
printCoin :: Float -> IO ()
printCoin bias = do
    let g = mkStdGen 1
    let (tosses, _) = samples 100 g (coin bias)
    printSamples ("bias: " <> show bias) tosses

-- | Create normal distribution as sum of uniform distributions.
normal :: PL Float
normal =  fromIntegral . sum <$> (replicateM 5 (coin 0.5))

main :: IO ()
main = do
    printCoin 0.01
    printCoin 0.99
    printCoin 0.5
    printCoin 0.7

    putStrLn $ "normal distribution using central limit theorem: "
    let g = mkStdGen 1
    let (nsamples, _) = samples 1000 g normal
    -- printSamples "normal: " nsamples
    printHistogram nsamples

    putStrLn $ "normal distribution using MCMC: "
    let (mcmcsamples, _) = samples 1000 g (mhD $  normalD 0.5)
    printHistogram mcmcsamples

    putStrLn $ "sampling from x^4 with finite support"
    let (mcmcsamples, _) = samples 1000 g (mhD $  polyD 4)
    printHistogram mcmcsamples


***bias: 1.0e-2
   samples: ________________________________________█_█_________________________________________________________
***bias: 0.99
   samples: ████████████████████████████████████████████████████████████████████████████████████████████████████
***bias: 0.5
   samples: __█____█__███_███_█__█_█___█_█_██___████████__█_████_████_████____██_█_██_____█__██__██_██____█__█__
***bias: 0.7
   samples: __█__█_█__███_█████__███_█_█_█_██_█_████████__███████████_████_█_███_████_██__█_███__██_███_█_█__█_█
normal distribution using central limit theorem: 
normal distribution using MCMC: 
sampling from x^4 with finite support

The smallest implementation of reverse mode AD (autograd) ever:

{-# LANGUAGE GeneralizedNewtypeDeriving #-}
import qualified Data.Map.Strict as M

-- | This file can be copy-pasted and will run!

-- | Symbols
type Sym = String
-- | Environments
type E a = M.Map Sym a
-- | Newtype to represent deriative values
type F = Float
newtype Der = Der { under :: F } deriving(Show, Num)

infixl 7 !#
-- | We are indexing the map at a "hash" (Sym)
(!#) :: E a -> Sym -> a
(!#) = (M.!)

-- | A node in the computation graph
data Node = 
  Node { name :: Sym -- ^ Name of the node
       , ins :: [Node] -- ^ inputs to the node
       , out :: E F -> F -- ^ output of the node
       , der :: (E F, E (Sym -> Der)) 
                  -> Sym -> Der -- ^ derivative wrt to a name

-- | @ looks like a "circle", which is a node. So we are indexing the map
-- at a node.
(!@) :: E a -> Node -> a 
(!@) e node = e M.! (name node)

-- | Given the current environments of values and derivatives, compute
-- | The new value and derivative for a node.
run_ :: (E F, E (Sym -> Der)) -> Node -> (E F, E (Sym -> Der))
run_ ein (Node name ins out der) = 
  let (e', ed') = foldl run_ ein ins -- run all the inputs
      v = out e' -- compute the output
      dv = der (e', ed') -- and the derivative
  in (M.insert name v e', M.insert name dv ed')  -- and insert them

-- | Run the program given a node 
run :: E F -> Node -> (E F, E (Sym -> Der))
run e n = run_ (e, mempty) n

-- | Let's build nodes
nconst :: Sym -> F -> Node
nconst n f = Node n [] (\_ -> f) (\_ _ -> 0)

-- | Variable
nvar :: Sym -> Node 
nvar n = Node n [] (!# n) (\_ n' -> if n == n' then 1 else 0)
-- | binary operation
nbinop :: (F -> F -> F)  -- ^ output computation from inputs
 -> (F -> Der -> F -> Der -> Der) -- ^ derivative computation from outputs
 -> Sym -- ^ Name
 -> (Node, Node) -- ^ input nodes
 -> Node
nbinop f df n (in1, in2) = 
  Node { name = n
       , ins = [in1, in2]
       , out = \e -> f (e !# name in1) (e !# name in2)
       , der = \(e, ed) n' -> 
                 let (name1, name2) = (name in1, name in2)
                     (v1, v2) = (e !# name1, e !# name2)
                     (dv1, dv2) = (ed !# name1 $ n', ed !# name2 $ n')
                     in df v1 dv1 v2 dv2

nadd :: Sym -> (Node, Node) -> Node
nadd = nbinop (+) (\v dv v' dv' -> dv + dv')

nmul :: Sym -> (Node, Node) -> Node
nmul = nbinop (*) (\v (Der dv) v' (Der dv') -> Der $ (v*dv') + (v'*dv))

main :: IO ()
main = do
  let x = nvar "x" :: Node
  let y = nvar "y"
  let xsq = nmul "xsq" (x, x)
  let ten = nconst "10" 10
  let xsq_plus_10 = nadd "xsq_plus_10" (xsq, ten)
  let xsq_plus_10_plus_y = nadd "xsq_plus_10_plus_y"  (xsq_plus_10, y)
  let (e, de) = run (M.fromList $ [("x", 2.0), ("y", 3.0)]) xsq_plus_10_plus_y
  putStrLn $ show e
  putStrLn $ show $ de !@ xsq_plus_10_plus_y $ "x"
  putStrLn $ show $ de !@ xsq_plus_10_plus_y $ "y"

Yeah, in ~80 lines of code, you can basically build an autograd engine. Isn't haskell so rad?

Timings of passes in GHC, and low hanging fruit in the backend:

  • One can use -v3 to get pass timings.
  • Apparently, GHC spends a lot of time in the simplifier, and time spend in the backend is peanuts in comparison to this.
  • To quote AndreasK:
  • Register allocation, common block elimination, block layout and pretty printing are the "slow" things in the backend as far as I remember.
  • There are also a handful of TODO's in the x86 codegen which still apply. So you can try to grep for these.
  • Strength reduction for division by a constant

Varargs in GHC: ghc/testsuite/tests/rts/T7160.hs

A comment from this test case tells us why the function debugBelch2 exists:

-- Don't use debugBelch() directly, because we cannot call varargs functions
-- using the FFI (doing so produces a segfault on 64-bit Linux, for example).
-- See Debug.Trace.traceIO, which also uses debugBelch2.
foreign import ccall "&debugBelch2" fun :: FunPtr (Ptr () -> Ptr () -> IO ())

The implementation is:


void debugBelch2(const char*s, char *t)

RtsMsgFunction *debugMsgFn  = rtsDebugMsgFn;

debugBelch(const char*s, ...)
  va_list ap;

Debugging debug info in GHC: Link

I wanted to use debug info to help build a better deubgging experience within tweag/asterius. So, I was reading through the sources of cmm/Debug.hs.
I'd never considered how to debug debug-info, and I found the information tucked inside a cute note in GHC (Note [Debugging DWARF unwinding info]):

This makes GDB produce a trace of its internal workings. Having gone this far, it's just a tiny step to run GDB in GDB. Make sure you install debugging symbols for gdb if you obtain it through a package manager.

How does GHC show a Float?

Turns out the implementation uses a cool paper:

Printing float point numbers fast and accurately

class Show:

First, let's review the Show typeclass:

class Show a where
    showsPrec :: Int    -- ^ the operator precedence of the enclosing
                        -- context (a number from @0@ to @11@).
                        -- Function application has precedence @10@.
              -> a      -- ^ the value to be converted to a 'String'
              -> ShowS

    -- | A specialised variant of 'showsPrec', using precedence context
    -- zero, and returning an ordinary 'String'.
    show      :: a   -> String
    show x          = shows x ""

-- | equivalent to 'showsPrec' with a precedence of 0.
shows           :: (Show a) => a -> ShowS
shows           =  showsPrec 0

So, note that all we need to do is to define showsPrec, and show gets defined as show x = showsPrec 0 x ""

instance Show Float:
instance  Show Float  where
    showsPrec   x = showSignedFloat showFloat x

showSignedFloat is a combinator to care of printing signs:

showSignedFloat :: (RealFloat a)
  => (a -> ShowS)       -- ^ a function that can show unsigned values
  -> Int                -- ^ the precedence of the enclosing context
  -> a                  -- ^ the value to show
  -> ShowS
showSignedFloat showPos p x
   | x < 0 || isNegativeZero x
       = showParen (p > 6) (showChar '-' . showPos (-x))
   | otherwise = showPos x

the showFloat is the function that is able to show a float as an unsigned value.

showFloat :: (RealFloat a) => a -> ShowS
showFloat x  =  showString (formatRealFloat FFGeneric Nothing x)
-- | utility function converting a 'String' to a show function that
-- simply prepends the string unchanged.
showString      :: String -> ShowS
showString      =  (++)
formatRealFloat :: (RealFloat a) => FFFormat -> Maybe Int -> a -> String
formatRealFloat fmt decs x = formatRealFloatAlt fmt decs False x

formatRealFloatAlt :: (RealFloat a) => FFFormat -> Maybe Int -> Bool -> a
                 -> String
formatRealFloatAlt fmt decs alt x
   | isNaN x                   = "NaN"
   | isInfinite x              = if x < 0 then "-Infinity" else "Infinity"
   | x < 0 || isNegativeZero x = '-':doFmt fmt (floatToDigits (toInteger base) (-x))
   | otherwise                 = doFmt fmt (floatToDigits (toInteger base) x)
  base = 10

  doFmt format (is, e) =
    let ds = map intToDigit is in
    case format of
     FFGeneric ->
      doFmt (if e < 0 || e > 7 then FFExponent else FFFixed)
     FFExponent ->
      case decs of
       Nothing ->
        let show_e' = show (e-1) in
        case ds of
          "0"     -> "0.0e0"
          [d]     -> d : ".0e" ++ show_e'
          (d:ds') -> d : '.' : ds' ++ "e" ++ show_e'
          []      -> errorWithoutStackTrace "formatRealFloat/doFmt/FFExponent: []"
       Just d | d <= 0 ->
        -- handle this case specifically since we need to omit the
        -- decimal point as well (#15115).
        -- Note that this handles negative precisions as well for consistency
        -- (see #15509).
        case is of
          [0] -> "0e0"
          _ ->
             (ei,is') = roundTo base 1 is
             n:_ = map intToDigit (if ei > 0 then init is' else is')
           in n : 'e' : show (e-1+ei)
       Just dec ->
        let dec' = max dec 1 in
        case is of
         [0] -> '0' :'.' : take dec' (repeat '0') ++ "e0"
         _ ->
           (ei,is') = roundTo base (dec'+1) is
           (d:ds') = map intToDigit (if ei > 0 then init is' else is')
          d:'.':ds' ++ 'e':show (e-1+ei)
     FFFixed ->
       mk0 ls = case ls of { "" -> "0" ; _ -> ls}
      case decs of
          | e <= 0    -> "0." ++ replicate (-e) '0' ++ ds
          | otherwise ->
                f 0 s    rs  = mk0 (reverse s) ++ '.':mk0 rs
                f n s    ""  = f (n-1) ('0':s) ""
                f n s (r:rs) = f (n-1) (r:s) rs
                f e "" ds
       Just dec ->
        let dec' = max dec 0 in
        if e >= 0 then
          (ei,is') = roundTo base (dec' + e) is
          (ls,rs)  = splitAt (e+ei) (map intToDigit is')
         mk0 ls ++ (if null rs && not alt then "" else '.':rs)
          (ei,is') = roundTo base dec' (replicate (-e) 0 ++ is)
          d:ds' = map intToDigit (if ei > 0 then is' else 0:is')
         d : (if null ds' && not alt then "" else '.':ds')
-- Based on "Printing Floating-Point Numbers Quickly and Accurately"
-- by R.G. Burger and R.K. Dybvig in PLDI 96.
-- This version uses a much slower logarithm estimator. It should be improved.

-- | 'floatToDigits' takes a base and a non-negative 'RealFloat' number,
-- and returns a list of digits and an exponent.
-- In particular, if @x>=0@, and
-- > floatToDigits base x = ([d1,d2,...,dn], e)
-- then
--      (1) @n >= 1@
--      (2) @x = 0.d1d2...dn * (base**e)@
--      (3) @0 <= di <= base-1@

floatToDigits :: (RealFloat a) => Integer -> a -> ([Int], Int)
floatToDigits _ 0 = ([0], 0)
floatToDigits base x =
  (f0, e0) = decodeFloat x
  (minExp0, _) = floatRange x
  p = floatDigits x
  b = floatRadix x
  minExp = minExp0 - p -- the real minimum exponent
  -- Haskell requires that f be adjusted so denormalized numbers
  -- will have an impossibly low exponent.  Adjust for this.
  (f, e) =
   let n = minExp - e0 in
   if n > 0 then (f0 `quot` (expt b n), e0+n) else (f0, e0)
  (r, s, mUp, mDn) =
   if e >= 0 then
    let be = expt b e in
    if f == expt b (p-1) then
      (f*be*b*2, 2*b, be*b, be)     -- according to Burger and Dybvig
      (f*be*2, 2, be, be)
    if e > minExp && f == expt b (p-1) then
      (f*b*2, expt b (-e+1)*2, b, 1)
      (f*2, expt b (-e)*2, 1, 1)
  k :: Int
  k =
    k0 :: Int
    k0 =
     if b == 2 && base == 10 then
        -- logBase 10 2 is very slightly larger than 8651/28738
        -- (about 5.3558e-10), so if log x >= 0, the approximation
        -- k1 is too small, hence we add one and need one fixup step less.
        -- If log x < 0, the approximation errs rather on the high side.
        -- That is usually more than compensated for by ignoring the
        -- fractional part of logBase 2 x, but when x is a power of 1/2
        -- or slightly larger and the exponent is a multiple of the
        -- denominator of the rational approximation to logBase 10 2,
        -- k1 is larger than logBase 10 x. If k1 > 1 + logBase 10 x,
        -- we get a leading zero-digit we don't want.
        -- With the approximation 3/10, this happened for
        -- 0.5^1030, 0.5^1040, ..., 0.5^1070 and values close above.
        -- The approximation 8651/28738 guarantees k1 < 1 + logBase 10 x
        -- for IEEE-ish floating point types with exponent fields
        -- <= 17 bits and mantissae of several thousand bits, earlier
        -- convergents to logBase 10 2 would fail for long double.
        -- Using quot instead of div is a little faster and requires
        -- fewer fixup steps for negative lx.
        let lx = p - 1 + e0
            k1 = (lx * 8651) `quot` 28738
        in if lx >= 0 then k1 + 1 else k1
        -- f :: Integer, log :: Float -> Float,
        --               ceiling :: Float -> Int
        ceiling ((log (fromInteger (f+1) :: Float) +
                 fromIntegral e * log (fromInteger b)) /
                   log (fromInteger base))
--WAS:            fromInt e * log (fromInteger b))

    fixup n =
      if n >= 0 then
        if r + mUp <= expt base n * s then n else fixup (n+1)
        if expt base (-n) * (r + mUp) <= s then n else fixup (n+1)
   fixup k0

  gen ds rn sN mUpN mDnN =
    (dn, rn') = (rn * base) `quotRem` sN
    mUpN' = mUpN * base
    mDnN' = mDnN * base
   case (rn' < mDnN', rn' + mUpN' > sN) of
    (True,  False) -> dn : ds
    (False, True)  -> dn+1 : ds
    (True,  True)  -> if rn' * 2 < sN then dn : ds else dn+1 : ds
    (False, False) -> gen (dn:ds) rn' sN mUpN' mDnN'

  rds =
   if k >= 0 then
      gen [] r (s * expt base k) mUp mDn
     let bk = expt base (-k) in
     gen [] (r * bk) s (mUp * bk) (mDn * bk)
 (map fromIntegral (reverse rds), k)

GHC LLVM code generator: Switch to unreachable

The switch to out of range code generator switches to the first label. It should be more profitable to switch to a unreachable block. That way, LLVM can take advantage of UB.

Concurrency in Haskell:

Great link to the GHC wiki that describes the concurrency primitives "bottom up":

Handy list of differential geometry definitions

There are way too many objects in diffgeo, all of them subtly connected. Here I catalogue all of the ones I have run across:


A manifold M of dimension n is a topological space. So, there is a topological structure T on M. There is also an Atlas, which is a family of Charts that satisfy some properties.


A chart is a pair (O ∈ T , cm: O -> R^n). The O is an open set of the manifold, and cm("chart for "m") is a continous mapping from O to R^n under the subspace topology for U and the standard topology for R^n.


An Atlas is a collection of Charts such that the charts cover the manifold, and the charts are pairwise compatible. That is, A = { (U_i, phi_i)}, such that union of U_i = M, and phi_j . inverse (phi_i) is smooth.

Differentiable map

f: M -> N be a mapping from an m dimensional manifold to an n dimensional manifold. Let frep = cn . f . inverse (cm): R^m -> R^n where cm: M -> R^m is a chart for M, cn: N -> R^n is a chart for N.frep is f represented in local coordinates. If frep is smooth for all choices of cm and cn, then f is a differentiable map from M to N.


Let I be an open interval of R which includes the point 0. A Curve is a differentiable map C: (a, b) -> M where a < 0 < b.

Function: (I hate this term, I prefer something like Valuation):

A differentiable mapping from M to R.

Directional derivative of a function f(m): M -> R with respect to a curve c(t): I -> M, denoted as c[f].

Let g(t) = (f . c)(t) :: I -c-> M -f-> R = I -> R. This this is the value dg/dt(t0) = (d (f . c) / dt) (0).

Tangent vector at a point p:

On a m dimensional manifold M, a tangent vector at a point p is an equivalence class of curves that have c(0) = p, such that c1(t) ~ c2(t) iff :

  1. For a (all) charts (O, ch) such that c1(0) ∈ O, d/dt (ch . c1: R -> R^m) = d/dt (ch . c2: R -> R^m).

That is, they have equal derivatives.

Tangent space(TpM):

The set of all tangent vectors at a point p forms a vector space TpM. We prove this by creating a bijection from every curve to a vector R^n.

Let (U, ch: U -> R) be a chart around the point p, where p ∈ U ⊆ M. Now, the bijection is defined as:

forward: (I -> M) -> R^n
forward(c) = d/dt (c . ch)

reverse: R^n -> (I -> M)
reverse(v)(t) = ch^-1 (tv)
Cotangent space(TpM*): dual space of the tangent space / Space of all linear functions from TpM to R.
  • Associated to every function f, there is a cotangent vector, colorfully called df. The definition is df: TpM -> R, df(c: I -> M) = c[f]. That is, given a curve c, we take the directional derivative of the function f along the curve c. We need to prove that this is constant for all vectors in the equivalence class and blah.
Pushforward push(f): TpM -> TpN

Given a curve c: I -> M, the pushforward is the curve f . c : I -> N. This extends to the equivalence classes and provides us a way to move curves in M to curves in N, and thus gives us a mapping from the tangent spaces.

This satisfies the identity:

push(f)(v)[g] === v[g . f]
Pullback pull(f): TpN* -> TpM*

Given a linear functional wn : TpN -> R, the pullback is defined as wn . push(f) : TpM -> R.

This satisfies the identity:

(pull wn)(v) === wn (push v)
(pull (wn : TpN->R): TpM->R) (v : TpM) : R  = (wn: TpN->R) (push (v: TpM): TpN) : R
Vector field as derivation


Lie derivation
Lie derivation as lie bracket

Lazy programs have space leaks, Strict programs have time leaks

Stumbled across this idea while reading some posts on a private discourse.

  • Continually adding new thunks without forcing them can lead to a space leak, aka the dreaded monadic parsing backtracking problem.

  • Continually running new thunks can lead to a "time leak", where we spend far too much time running things that should not be run in the first place!

This is an interesting perspective that I've never seen articulated before, and somehow helps make space leaks feel more... palatable? Before, I had no analogue to a space leak in the strict world, so I saw them as a pathology. But with this new perspective, I can see that the strict world's version of a space leak is a time leak.

Presburger arithmetic can represent the Collatz Conjecture

An observation I had: the function

f(x) = x/2      if (x % 2 == 0)
f(x) = 3x + 1   otherwise

is a Presburger function, so by building better approximations to the transitive closure of a presburger function, one could get better answers to the Collatz conjecture. Unfortunately, ISL (the integer set library) of today is not great against the formidable foe.

The code:

#include <isl/set.h>
#include <isl/version.h>
#include <isl/map.h>
#include <isl/aff.h>
#include <isl/local_space.h>
#include <isl/constraint.h>
#include <isl/space.h>

int main() {
    isl_ctx *ctx = isl_ctx_alloc();
    const char *s = "{ [x] -> [x / 2] : x % 2 = 0; [x] -> [3 * x + 1] : x % 2 = 1}";

    isl_map *m = isl_map_read_from_str(ctx, s);


    isl_bool b;
    isl_map *p = isl_map_transitive_closure(m, &b);
    printf("exact: %d\n", b);


Produces the somewhat disappointing, and yet expected output:

$ clang bug.c -lisl -Lisl-0.20/.libs -o bug -I/usr/local/include/
$ ./bug
{ [x] -> [o0] : 2o0 = x or (exists (e0 = floor((1 + x)/2): o0 = 1 + 3x and 2e0 = 1 + x)) }
exact: 0
{ [x] -> [o0] }

I find it odd that it is unable to prove anything about the image, even that it is non-negative, for example. This is an interesting direction in which to improve the functions isl_map_power and isl_map_transitive_closure though.

Using compactness to argue about the cover in an argument

I've always seen compactness be used by starting with a possibly infinite coverm and then filtering it into a finite subcover. This finite subcover is then used for finiteness properties (like summing, min, max, etc.).

I recently ran across a use of compactness when one starts with the set of all possible subcovers, and then argues about why a cover cannot be built from these subcovers if the set is compact. I found it to be a very cool use of compactness, which I'll record below:


If a family of compact, countably infinite sets S_a have all finite intersections non-empty, then the intersection of the family S_a is non-empty.


Let S = intersection of S_a. We know that S must be compact since all the S_a are compact, and the intersection of a countably infinite number of compact sets is compact.

Now, let S be empty. Therefore, this means there must be a point p ∈ P such that p !∈ S_i for some arbitrary i.

Cool use of theorem:

We can see that the cantor set is non-empty, since it contains a family of closed and bounded sets S1, S2, S3, ... such that S1 ⊇ S2 ⊇ S3 ... where each S_i is one step of the cantor-ification. We can now see that the cantor set is non-empty, since:

  1. Each finite interesection is non-empty, and will be equal to the set that has the highest index in the finite intersection.

  2. Each of the sets Si are compact since they are closed and bounded subsets of R

  3. Invoke theorem.

Japanese Financial Counting system

Japanese contains a seaparate kanji set called daiji, to prevent people from adding strokes to stuff previously written.

#  |Common |Formal
1  |一     |壱
2  |二     |弐
3  |三     |参

Stephen wolfram's live stream

I've taken to watching the live stream when I have some downtime and want some interesting content.

The discussions of Wolfram with his group are great, and they bring up really interesting ideas (like that of cleave being very irregular).

Cleave as a word has some of the most irregular inflections

  • cleave
  • clove
  • cleaved
  • clave
  • cleft

McCune's single axiom for group theory

Single Axioms for Groups and Abelian Groups with Various Operations provides a single axiom for groups. This can be useful for some ideas I have for training groups, where we can use this axiom as the loss function!

Word2Vec C code implements gradient descent really weirdly

I'll be posting snippets of the original source code, along with a link to the Github sources. We are interested in exploring the skip-gram implementation of Word2Vec, with negative sampling, without hiearchical softmax. I assume basic familiarity with word embeddings and the skip-gram model.

Construction of the sigmoid lookup table


expTable = (real *)malloc((EXP_TABLE_SIZE + 1) * sizeof(real));
for (i = 0; i < EXP_TABLE_SIZE; i++) {
  expTable[i] = exp((i / (real)EXP_TABLE_SIZE * 2 - 1) *
                    MAX_EXP);  // Precompute the exp() table
  expTable[i] =
      expTable[i] / (expTable[i] + 1);  // Precompute f(x) = x / (x + 1)

Here, the code constructs a lookup table which maps [0...EXP_TABLE_SIZE-1] to [sigmoid(-MAX_EXP)...sigmoid(MAX_EXP)]. The index i first gets mapped to (i / EXP_TABLE_SIZE) * 2 - 1, which sends 0 to -1 and EXP_TABLE_SIZE to 1. This is then rescaled by MAX_EXP.

Layer initialization

  • syn0 is a global variable, initialized with random weights in the range of [-0.5...0.5]. It has dimensions VOCAB x HIDDEN. This layer holds the hidden representations of word vectors.
a = posix_memalign((void **)&syn0, 128,
               (long long)vocab_size * layer1_size * sizeof(real));

for (a = 0; a < vocab_size; a++)
        for (b = 0; b < layer1_size; b++) {
            next_random = next_random * (unsigned long long)25214903917 + 11;
            syn0[a * layer1_size + b] =
                (((next_random & 0xFFFF) / (real)65536) - 0.5) / layer1_size;
  • syn1neg is a global variable that is zero-initialized. It has dimensions VOCAB x HIDDEN. This layer also holds hidden representations of word vectors, when they are used as a negative sample.
a = posix_memalign((void **)&syn1neg, 128,
                   (long long)vocab_size * layer1_size * sizeof(real));
for (a = 0; a < vocab_size; a++)
    for (b = 0; b < layer1_size; b++) syn1neg[a * layer1_size + b] = 0;
  • neu1e is a temporary per-thread buffer (Remember that the word2vec C code use CPU threads for parallelism) which is zero initialized. It has dimensions 1 x HIDDEN.
real *neu1e = (real *)calloc(layer1_size, sizeof(real));


Throughout word2vec, no 2D arrays are used. Indexing of the form arr[word][ix] is manually written as arr[word * layer1_size + ix]. So, I will call word * layer1_size as the "base address", and ix as the "offset of the array index expression henceforth.

Here, l1 is the base address of the word at the center of window (the focus word). l2 is the base address of either the word that is negative sampled from the corpus, or the word that is a positive sample from within the context window.

label tells us whether the sample is a positive or a negative sample. label = 1 for positive samples, and label = 0 for negative samples.

// zero initialize neu1e
for (c = 0; c < layer1_size; c++) neu1e[c] = 0;
// loop through each negative sample
if (negative > 0)  for (d = 0; d < negative + 1; d++) {
  // take the dot product: f=  syn0[focus] . syn1neg[context]
  for (c = 0; c < layer1_size; c++) f += syn0[c + l1] * syn1neg[c + l2];
  // compute: g = (label - sigmoid(2f - 1)) * alpha
  // g is computed using lookups into a lookup table and clamping for
  // efficiency.
  if (f > MAX_EXP) g = (label - 1) * alpha;
  else if (f < -MAX_EXP) g = (label - 0) * alpha;
  g = (label - expTable[(int)((f + MAX_EXP) *
                              (EXP_TABLE_SIZE /
                               MAX_EXP / 2))]) * alpha;
  // Now that we have computed the gradient:
  // `g = (label - output) * learningrate`,
  // we need to perform backprop. This is where the code gets weird.

  for (c = 0; c < layer1_size; c++) neu1e[c] += g * syn1neg[c + l2];
  for (c = 0; c < layer1_size; c++) syn1neg[c + l2] += g * syn0[c + l1];
  } // end loop through negative samples
// Learn weights input -> hidden
for (c = 0; c < layer1_size; c++) syn0[c + l1] += neu1e[c];
  • We have two vectors for each word, one called syn0[l1 + _] and the other syn1neg[l2 + _]. The syn1neg word embedding is used whenever a word is used a negative sample, and is not used anywhere else. Also, the syn1neg vector is zero initialized, while the syn0 vectors are randomly initialized.

  • The values we backprop with g * syn1neg[l2 + _], g * syn0[l1 + _] are not the correct gradients of the error term! The derivative of a sigmoid is dsigmoid(x)/dx = sigmoid(x) [1 - sigmoid(x)]. The [1 - sigmoid(x)] is nowhere to be seen, let alone the fact that we are using sigmoid(2x - 1) and not regular sigmoid. Very weird.

  • We hold the value of syn0 constant throughout all the negative samples, which was not mentioned in any tutorial I've read.

The paper does not mentioned these implementation details, and neither does any blog post that I've read. I don't understand what's going on, and I plan on updating this section when I understand this better.

Arthur Whitney: dense code

  • Guy who wrote a bunch of APL dialects, write code in an eclectic style that has very little whitespace and single letter variable names.
  • Believes that this allows him to hold the entire program in his head.
  • Seems legit from my limited experience with APL, haskell one-liners.
  • The b programming language. It's quite awesome to read the sources. For example, a.c

How does one work with arrays in a linear language?

Given an array of qubits xs: Qubit[], I want to switch to little endian. Due to no-cloning, I can't copy them! I suppose I can use recursion to build up a new "list". But this is not the efficient array version we know and love and want.

The code that I want to work but does not:

function switchEndian(xs: Qubit[]): Unit {
    for(i in 0..Length(xs) - 1) {
        Qubit q = xs[i]; // boom, this does not work!
        xs[i] = xs[Length(xs) - 1 - i]
        xs[Length(xs) - 1 - i] = q;

On the other hand, what does work is to setup a quantum circuit that performs this flipping, since it's a permutation matrix at the end of the day. But this is very slow, since it needs to simulate the "quantumness" of the solution, since it takes 2^n basis vectors for n qubits.

However, the usual recursion based solution works:

function switchEndian(xs: Qubit[]): Qubit[] {
    if(Length(xs) == 1) {
        return xs;
    } else {
        switchEndian(xs[1..(Length(xs) - 1)] + xs[0]

This is of course, suboptimal.

I find it interesting that in the linear types world, often the "pure" solution is forced since mutation very often involves temporaries / copying!

(I'm solving assignments in qsharp for my course in college)

How Linear optimisation is the same as Linear feasibility checking

Core building block of effecively using the ellipsoid algorithm.

  • If we posess a way to check if a point p ∈ P where P is a polytope, we can use this to solve optimisation problems.
  • Given the optimisation problem maximise c^Tx subject to Ax = b, we can construct a new non-emptiness problem. This allows us to convert optimisation into feasibility.
  • The new problem is Ax = b, A^Ty = c, c^Tx = b^T y. Note that by duality, a point in this new polyhedra will be an optimal solution to the above linear program. We are forcing c^Tx = b^Ty, which will be the optimal solution, since the solution where the primal and dual agree is the optimal solution by strong duality.
  • This way, we have converted a linear programming problem into a check if this polytope is empty problem!

Quantum computation without complex numbers

I recently learnt that the Toeffili and Hadamard gates are universal for quantum computation. The description of these gates involve no complex numbers. So, we can write any quantum circuit in a "complex number free" form. The caveat is that we may very well have input qubits that require complex numbers.

Even so, a large number (all?) of the basic algorithms shown in Nielsen and Chaung can be encoded in an entirely complex-number free fashion.

I don't really understand the ramifications of this, since I had the intuition that the power of quantum computation comes from the ability to express complex phases along with superposition (tensoring). However, I now have to remove the power from going from R to C in many cases. This is definitely something to ponder.

Linguistic fun fact: Comparative Illusion

I steal from wikipedia:

Comparative Illusion, which is a grammatical illusion where certain sentences seem grammatically correct when you read them, but upon further reflection actually make no sense.

For example: "More people have been to Berlin than I have."

Long-form posts:



Simplexhc (STG -> LLVM compiler) progress

GSoC (2015)

You can’t perform that action at this time.