feat: federated learning demo on the actor framework

- supervisor: ignore stale death messages (OneForAll restart storm),
  drop normally-exited children instead of respawning, add a
  restart-rate limit
- register actors before their thread starts so UUID dispatch and
  getSelf cannot race the spawn
- killActor now kills the thread so links see Killed
- fulfill connection promises with the pool winner on simultaneous
  connects
- clamp the PaySim test split on small datasets and split shared
  transaction types across trainers in non-IID partitioning
This commit is contained in:
2026-06-21 12:00:00 +02:00
parent 9a52fc5089
commit 711d26d9f7
25 changed files with 2127 additions and 115 deletions
+2
View File
@@ -1 +1,3 @@
dist-newstyle
demo-out/
fl_model_*.bin
+38 -1
View File
@@ -71,6 +71,19 @@ library
Control.Actor.Supervision
Control.Actor.Network
Control.Actor.Registry
FL.Types
FL.CRDT
FL.NN
FL.Retry
FL.Data.PaySim
FL.Actors.Logger
FL.Actors.Evaluator
FL.Actors.Trainer
FL.Actors.Aggregator
FL.Actors.Coordinator
FL.Actors.PeerCoordinator
FL.Provider
FL.P2P
-- Modules included in this library but not exported.
-- other-modules:
@@ -79,10 +92,34 @@ library
-- other-extensions:
-- Other library packages from which modules are imported.
build-depends: base ^>=4.18.3.0, stm, uuid, bytestring ^>=0.12.0.0, containers ^>=0.8, mtl ^>=2.3.0, binary ^>=0.8.9, network ^>=3.2
build-depends:
base ^>=4.18.3.0,
stm,
uuid,
bytestring ^>=0.12.0.0,
containers ^>=0.8,
mtl ^>=2.3.0,
binary ^>=0.8.9,
network ^>=3.2,
hmatrix,
cassava,
vector,
random,
text,
time
-- Directories containing source files.
hs-source-dirs: src
-- Base language which the package is written in.
default-language: GHC2021
executable fl-actors
import: warnings
main-is: Main.hs
hs-source-dirs: app
build-depends:
base ^>=4.18.3.0,
actors,
optparse-applicative
default-language: GHC2021
+150
View File
@@ -0,0 +1,150 @@
module Main where
import Control.Actor.Types (NodeAddr (..))
import FL.P2P (runPeer)
import FL.Provider (runCoordinator, runTrainer)
import FL.Types
import Options.Applicative
-- ---------------------------------------------------------------------------
-- CLI
-- ---------------------------------------------------------------------------
data Mode = Provider | P2P deriving (Show, Eq)
data Role = Coordinator | Trainer | Peer deriving (Show, Eq)
data CLIArgs = CLIArgs
{ argMode :: Mode
, argRole :: Role
, argPort :: Int
, argBasePort :: Int
, argRounds :: Int
, argEpochs :: Int
, argLR :: Double
, argBatch :: Int
, argNumTrainers :: Int
, argTrainerId :: Int
, argCoordHost :: String
, argCoordPort :: Int
, argKnownPeers :: [String]
, argDataPath :: Maybe FilePath
, argPartition :: PartitionMode
, argNorm :: NormMode
}
parseMode :: String -> Either String Mode
parseMode "provider" = Right Provider
parseMode "p2p" = Right P2P
parseMode s = Left ("unknown mode: " <> s)
parsePartition :: String -> Either String PartitionMode
parsePartition "iid" = Right IID
parsePartition "noniid" = Right NonIID
parsePartition s = Left ("unknown partition: " <> s)
parseNorm :: String -> Either String NormMode
parseNorm "minmax" = Right MinMax
parseNorm "zscore" = Right ZScore
parseNorm s = Left ("unknown norm: " <> s)
parseRole :: String -> Either String Role
parseRole "coordinator" = Right Coordinator
parseRole "trainer" = Right Trainer
parseRole "peer" = Right Peer
parseRole s = Left ("unknown role: " <> s)
parsePeer :: String -> NodeAddr
parsePeer s =
let (host, rest) = break (== ':') s
in NodeAddr host (read (drop 1 rest))
argsParser :: Parser CLIArgs
argsParser = CLIArgs
<$> option (eitherReader parseMode)
(long "mode" <> short 'm' <> metavar "provider|p2p"
<> help "Topology mode" <> value Provider <> showDefaultWith show)
<*> option (eitherReader parseRole)
(long "role" <> short 'r' <> metavar "coordinator|trainer|peer"
<> help "Node role" <> value Coordinator <> showDefaultWith show)
<*> option auto
(long "port" <> metavar "PORT"
<> help "This node's port" <> value 9100 <> showDefault)
<*> option auto
(long "base-port" <> metavar "PORT"
<> help "Base port (coordinator uses this)" <> value 9100 <> showDefault)
<*> option auto
(long "rounds" <> metavar "N"
<> help "Number of federated rounds" <> value 10 <> showDefault)
<*> option auto
(long "epochs" <> short 'e' <> metavar "N"
<> help "Local training epochs per round" <> value 5 <> showDefault)
<*> option auto
(long "lr" <> metavar "LR"
<> help "Learning rate" <> value 0.01 <> showDefault)
<*> option auto
(long "batch" <> metavar "N"
<> help "Mini-batch size" <> value 32 <> showDefault)
<*> option auto
(long "trainers" <> short 't' <> metavar "N"
<> help "Expected number of trainers/peers" <> value 3 <> showDefault)
<*> option auto
(long "id" <> metavar "N"
<> help "Trainer/peer index (0-based)" <> value 0 <> showDefault)
<*> strOption
(long "coord-host" <> metavar "HOST"
<> help "Coordinator hostname (trainer mode)" <> value "localhost")
<*> option auto
(long "coord-port" <> metavar "PORT"
<> help "Coordinator port (trainer mode)" <> value 9100 <> showDefault)
<*> many (strOption
(long "peer" <> metavar "HOST:PORT"
<> help "Known peer address (p2p mode, repeat for each peer)"))
<*> optional (strOption
(long "data" <> short 'd' <> metavar "PATH"
<> help "Path to PaySim CSV (uses synthetic data if absent)"))
<*> option (eitherReader parsePartition)
(long "partition" <> metavar "iid|noniid"
<> help "Data partition mode" <> value IID <> showDefaultWith show)
<*> option (eitherReader parseNorm)
(long "norm" <> metavar "minmax|zscore"
<> help "Feature normalization mode" <> value MinMax <> showDefaultWith show)
main :: IO ()
main = do
args <- execParser $
info (argsParser <**> helper)
( fullDesc
<> progDesc "Federated Learning on the Haskell Actor Framework"
<> header "fl-actors — FedAvg fraud detection"
)
let cfg = FLConfig
{ cfgRounds = argRounds args
, cfgEpochs = argEpochs args
, cfgLearningRate = argLR args
, cfgBatchSize = argBatch args
, cfgNumTrainers = argNumTrainers args
, cfgDataPath = argDataPath args
, cfgPartition = argPartition args
, cfgNorm = argNorm args
, cfgBasePort = argBasePort args
}
case (argMode args, argRole args) of
(Provider, Coordinator) ->
runCoordinator cfg { cfgBasePort = argPort args }
(Provider, Trainer) -> do
let coordAddr = NodeAddr (argCoordHost args) (fromIntegral (argCoordPort args))
runTrainer cfg (argTrainerId args) coordAddr
(P2P, Peer) -> do
let peers = map parsePeer (argKnownPeers args)
cfg' = cfg { cfgBasePort = argPort args }
runPeer cfg' (argTrainerId args) peers
(mode, role) ->
putStrLn $ "Invalid combination: mode=" <> show mode <> " role=" <> show role
<> "\n provider mode uses: --role=coordinator or --role=trainer"
<> "\n p2p mode uses: --role=peer"
+85
View File
@@ -0,0 +1,85 @@
#!/usr/bin/env bash
#
# Demo: federated learning in P2P mode (full mesh, no central coordinator).
# Starts N peer nodes as separate processes; each trains locally, gossips
# updates via PeerSync (CRDT-gated rounds), and aggregates independently.
# At the end every peer must hold the SAME model — the script verifies the
# saved model files are byte-identical.
#
# Usage:
# scripts/demo-p2p.sh [ROUNDS] [PEERS] [EPOCHS] [/path/to/paysim.csv]
#
# Without a CSV path, nodes fall back to synthetic data (fast; the metrics are
# meaningless, but the gossip protocol and aggregation are fully exercised).
set -euo pipefail
ROUNDS="${1:-3}"
PEERS="${2:-3}"
EPOCHS="${3:-2}"
DATA="${4:-}"
BASE_PORT=9600
OUT="demo-out/p2p"
cd "$(dirname "$0")/.."
echo "==> Building..."
cabal build exe:fl-actors >/dev/null
BIN="$(cabal list-bin fl-actors)"
mkdir -p "$OUT"
rm -f "$OUT"/*.log fl_model_peer-*.bin
DATA_ARGS=()
[ -n "$DATA" ] && DATA_ARGS=(--data "$DATA")
PIDS=()
cleanup() { kill "${PIDS[@]}" 2>/dev/null || true; wait 2>/dev/null || true; }
trap cleanup EXIT INT TERM
for i in $(seq 0 $((PEERS - 1))); do
PORT=$((BASE_PORT + i))
ARGS=(--mode p2p --role peer --id "$i" --port "$PORT"
--rounds "$ROUNDS" --trainers "$PEERS" --epochs "$EPOCHS")
for j in $(seq 0 $((PEERS - 1))); do
[ "$j" = "$i" ] && continue
ARGS+=(--peer "localhost:$((BASE_PORT + j))")
done
echo "==> Starting peer-$i on port $PORT"
"$BIN" "${ARGS[@]}" "${DATA_ARGS[@]}" > "$OUT/peer$i.log" 2>&1 &
PIDS+=($!)
done
echo "==> Federation running; waiting for all $PEERS peers to finish..."
echo " (live log: tail -f $OUT/peer0.log)"
FAIL=0
for pid in "${PIDS[@]}"; do
wait "$pid" || FAIL=1
done
PIDS=()
if [ "$FAIL" -ne 0 ]; then
echo "!! A peer exited abnormally; check $OUT/ logs"
exit 1
fi
echo
echo "==> Evaluation results (peer-0's view):"
grep "Eval round=" "$OUT/peer0.log" || true
echo
MODELS=(fl_model_peer-*.bin)
if [ "${#MODELS[@]}" -ne "$PEERS" ]; then
echo "!! Expected $PEERS model files, found ${#MODELS[@]}; check $OUT/ logs"
exit 1
fi
FIRST="${MODELS[0]}"
for m in "${MODELS[@]:1}"; do
if ! cmp -s "$FIRST" "$m"; then
echo "!! Divergence: $m differs from $FIRST"
md5sum "${MODELS[@]}"
exit 1
fi
done
echo "==> Success: all $PEERS peers converged to a byte-identical model:"
ls -la "${MODELS[@]}"
echo " Full logs in $OUT/"
+72
View File
@@ -0,0 +1,72 @@
#!/usr/bin/env bash
#
# Demo: federated learning in PROVIDER mode (star topology).
# Starts one coordinator and N trainer nodes as separate processes, runs the
# federation, then prints the evaluation log and the saved model file.
#
# Usage:
# scripts/demo-provider.sh [ROUNDS] [TRAINERS] [EPOCHS] [/path/to/paysim.csv]
#
# Without a CSV path, nodes fall back to synthetic data (fast; the metrics are
# meaningless, but the message flow and aggregation are fully exercised).
set -euo pipefail
ROUNDS="${1:-3}"
TRAINERS="${2:-3}"
EPOCHS="${3:-2}"
DATA="${4:-}"
BASE_PORT=9500
OUT="demo-out/provider"
cd "$(dirname "$0")/.."
echo "==> Building..."
cabal build exe:fl-actors >/dev/null
BIN="$(cabal list-bin fl-actors)"
mkdir -p "$OUT"
rm -f "$OUT"/*.log fl_model_coordinator.bin
DATA_ARGS=()
[ -n "$DATA" ] && DATA_ARGS=(--data "$DATA")
PIDS=()
cleanup() { kill "${PIDS[@]}" 2>/dev/null || true; wait 2>/dev/null || true; }
trap cleanup EXIT INT TERM
echo "==> Starting coordinator on port $BASE_PORT (rounds=$ROUNDS, trainers=$TRAINERS, epochs=$EPOCHS)"
"$BIN" --mode provider --role coordinator \
--port "$BASE_PORT" --rounds "$ROUNDS" --trainers "$TRAINERS" --epochs "$EPOCHS" \
"${DATA_ARGS[@]}" > "$OUT/coordinator.log" 2>&1 &
COORD=$!
PIDS+=("$COORD")
sleep 1
for i in $(seq 0 $((TRAINERS - 1))); do
echo "==> Starting trainer-$i on port $((BASE_PORT + i + 1))"
"$BIN" --mode provider --role trainer --id "$i" \
--base-port "$BASE_PORT" --coord-host localhost --coord-port "$BASE_PORT" \
--rounds "$ROUNDS" --trainers "$TRAINERS" --epochs "$EPOCHS" \
"${DATA_ARGS[@]}" > "$OUT/trainer$i.log" 2>&1 &
PIDS+=($!)
done
echo "==> Federation running; waiting for the coordinator to finish..."
echo " (live log: tail -f $OUT/coordinator.log)"
if ! wait "$COORD"; then
echo "!! Coordinator exited abnormally; last log lines:"
tail -20 "$OUT/coordinator.log"
exit 1
fi
echo
echo "==> Evaluation results:"
grep "Eval round=" "$OUT/coordinator.log" || true
echo
if [ -f fl_model_coordinator.bin ]; then
echo "==> Success: model saved to fl_model_coordinator.bin ($(stat -c %s fl_model_coordinator.bin) bytes)"
echo " Full logs in $OUT/"
else
echo "!! No model file was produced; check $OUT/ logs"
exit 1
fi
+2
View File
@@ -5,6 +5,8 @@ module Control.Actor
, module Control.Actor.Core
, module Control.Actor.Supervision
, module Control.Actor.Network
-- Initialization
, initRuntime
-- Demo
, pingActor
, forwardActorWithCell
+34 -8
View File
@@ -31,7 +31,7 @@ module Control.Actor.Core
import Control.Actor.Runtime
import Control.Actor.Types
import Control.Concurrent (forkIO, threadDelay)
import Control.Concurrent (forkIO, killThread, threadDelay)
import Control.Concurrent.MVar (newEmptyMVar, putMVar, takeMVar)
import Control.Concurrent.STM
( atomically
@@ -52,8 +52,9 @@ import Control.Exception
, throwIO
, try
)
import Control.Monad (forM_, void)
import Control.Monad (forM_, join, void)
import Data.Maybe (fromMaybe)
import System.Timeout (timeout)
import Control.Monad.Reader
( MonadIO (..)
, MonadReader (..)
@@ -122,6 +123,7 @@ spawnActorAs ::
u ->
RuntimeM (ActorRef m r)
spawnActorAs uuid actorFn deathFn initState = do
liftIO $ putStrLn $ "spawning actor with UUID " <> show uuid
rt <- ask
mailbox <- liftIO newTQueueIO
deathQ <- liftIO newTQueueIO
@@ -151,7 +153,13 @@ spawnActorAs uuid actorFn deathFn initState = do
Stop -> return ()
Continue u -> loop fn as {asEnv = u}
-- The thread blocks on `ready` until it is registered in rtActors:
-- otherwise messages dispatched by UUID in the window before the insert
-- would be dropped, and a death racing the insert would leave a stale
-- entry behind (the death path's delete running before the insert).
ready <- liftIO newEmptyMVar
tid <- liftIO $ forkIO $ do
takeMVar ready
result <- try @SomeException (loop actorFn actorState)
let reason = case result of
Right () -> Normal
@@ -167,6 +175,7 @@ spawnActorAs uuid actorFn deathFn initState = do
liftIO $ atomically $
modifyTVar (rtActors rt) (Map.insert actorId (tid, SomeActorRef actorRef))
liftIO $ putMVar ready ()
return actorRef
spawnActor ::
@@ -189,9 +198,16 @@ linkTo target = do
as <- asks fst
liftIO $ atomically $ modifyTVar (asLinks as) (target :)
-- | Forcibly terminate an actor's thread. Previously this only posted a
-- 'DeathMessage' to the target's death queue, which a deathFn returning
-- 'Continue' would silently ignore — and even on 'Stop' the links were told
-- the exit was 'Normal'. Killing the thread makes links see 'Killed'.
killActor :: ActorRef m r -> ActorM u ()
killActor (LocalRef {arDeathQ, arState}) =
liftIO $ atomically $ writeTQueue arDeathQ (DeathMessage (asId arState) Killed)
killActor ref@(LocalRef {}) = do
rt <- asks snd
liftIO $ do
actors <- readTVarIO (rtActors rt)
forM_ (Map.lookup (actorRefId ref) actors) (killThread . fst)
killActor (RemoteRef _) =
liftIO $ putStrLn "killActor: remote kill not yet implemented"
@@ -206,13 +222,23 @@ cast' msg (RemoteRef (ActorId nodeId uuid)) = do
maybeAddr <- lookupNode nodeId
case maybeAddr of
Nothing -> liftIO $ putStrLn $ "cast: no node in lookup table with id " <> show nodeId
Just addr -> rtSendRemote rt addr (NMCast uuid (encode msg))
Just addr -> do
result <- liftIO $ try @SomeException $ withRuntime rt $ rtSendRemote rt addr (NMCast uuid (encode msg))
case result of
Left e -> liftIO $ putStrLn $ "cast: send to " <> show addr <> " failed: " <> show e
Right _ -> return ()
-- | How long a call waits for a reply before giving up with Nothing.
-- Without this, a call to a dead/missing actor (or over a dead connection)
-- would block the caller forever.
callTimeoutMicros :: Int
callTimeoutMicros = 10_000_000
call' :: (Binary msg, Binary reply) => msg -> ActorRef msg reply -> RuntimeM (Maybe reply)
call' msg (LocalRef {arMsgQ}) = liftIO $ do
mv <- newEmptyMVar
atomically $ writeTQueue arMsgQ (thisNodeId, Call msg mv)
takeMVar mv
join <$> timeout callTimeoutMicros (takeMVar mv)
call' msg (RemoteRef (ActorId nodeId uuid)) = do
rt <- ask
corrId <- liftIO $ atomically $ do
@@ -229,9 +255,9 @@ call' msg (RemoteRef (ActorId nodeId uuid)) = do
return Nothing
Just addr -> do
rtSendRemote rt addr (NMCall uuid corrId (rtNodeId rt) (encode msg))
raw <- liftIO $ takeMVar replyVar
raw <- liftIO $ timeout callTimeoutMicros (takeMVar replyVar)
liftIO $ atomically $ modifyTVar (rtPending rt) (Map.delete corrId)
return $ Just (decode raw)
return $ decode <$> raw
cast :: (Binary msg) => msg -> ActorRef msg reply -> ActorM u ()
cast = (liftRuntime .) . cast'
+115 -73
View File
@@ -15,9 +15,9 @@ module Control.Actor.Network
import Control.Actor.Core
( ActorM, ActorResult (..), Handler, cast', liftRuntime, linkActorTo
, spawnActor, state, stopOnDeath )
, pass, spawnActor, state, stopOnDeath )
import Control.Actor.Registry (deregisterNode)
import Control.Actor.Runtime (Runtime (..), RuntimeM, addrToNodeId, withRuntime)
import Control.Actor.Supervision (ChildSpec (..), RestartStrategy (..), childWithRef, supervise')
import Control.Actor.Transport (ConnHandle (..), Transport (..), createTCPTransport)
import Control.Actor.Types
import Control.Concurrent (forkIO)
@@ -26,23 +26,21 @@ import Control.Concurrent.STM
( atomically
, modifyTVar
, newEmptyTMVar
, newEmptyTMVarIO
, putTMVar
, readTMVar
, readTVar
, readTVarIO
, tryPutTMVar
, tryTakeTMVar
, writeTQueue
, writeTVar
)
import Control.Exception (SomeException, try)
import Control.Monad (forM_, forever, unless, void)
import Control.Monad.Reader (MonadIO (..), MonadReader (..))
import Control.Exception (SomeException, throwIO, toException, try)
import Control.Monad (forM_, forever, unless, void, when)
import Control.Monad.Reader (MonadIO (..), MonadReader (..), asks)
import Data.Binary (decode, decodeOrFail, encode)
import Data.ByteString.Lazy (ByteString)
import Data.Map qualified as Map
import Data.Maybe (fromMaybe)
import System.Timeout (timeout)
-- Connection actors
@@ -54,51 +52,63 @@ connActorFn nm = do
connDeathFn :: NodeAddr -> DeathMessage -> ActorM ConnHandle (SupervisorAction ConnHandle)
connDeathFn peer _ = do
ch <- state
rt <- liftRuntime ask
liftIO $ do
atomically $ modifyTVar (rtConnections rt) (Map.delete peer)
chClose ch
ch <- state
aid <- asks (asId . fst)
rt <- liftRuntime ask
-- Only clear the pool entry if it is ours: with simultaneous connects two
-- conn actors can exist for the same peer, and the loser must not evict
-- the active one.
wasActive <- liftIO $ atomically $ do
conns <- readTVar (rtConnections rt)
case Map.lookup peer conns of
Just ref | actorRefId ref == aid -> do
modifyTVar (rtConnections rt) (Map.delete peer)
return True
_else -> return False
liftIO $ chClose ch
when wasActive $ do
mNodeId <- liftRuntime $ addrToNodeId peer
forM_ mNodeId $ \nodeId -> liftRuntime $ deregisterNode nodeId
return Stop
routerActorFn :: NodeAddr -> Handler ByteString () ()
routerActorFn senderAddr raw = do
let nm = decode raw :: NetworkMessage
valid <- liftRuntime $ validateAndDispatch senderAddr nm
unless valid $ liftIO $ putStrLn "router: dropping invalid message"
return $ ActorResult Nothing () Nothing
routerActorFn senderAddr raw =
case decodeOrFail raw of
Left _ -> do
liftIO $ putStrLn $ "router: dropping undecodable frame from " <> show senderAddr
pass
Right (_, _, nm) -> do
valid <- liftRuntime $ validateAndDispatch senderAddr nm
unless valid $ liftIO $
putStrLn $ "router: dropping invalid message from " <> show senderAddr
pass
-- Connection supervision tree
spawnConnTree :: NodeAddr -> ConnHandle -> RuntimeM (ActorRef NetworkMessage ())
spawnConnTree peer ch = do
rt <- ask
routerCell <- liftIO newEmptyTMVarIO
connCell <- liftIO newEmptyTMVarIO
supervise' OneForAll
[ childWithRef (routerActorFn peer) stopOnDeath () routerCell
, ChildSpec
{ csRun = \target -> do
ref <- spawnActor connActorFn (connDeathFn peer) ch
linkActorTo target ref
return ref
, csOnSpawn = \ref -> case ref of
LocalRef { arDeathQ, arState } -> do
atomically $ do
void $ tryTakeTMVar connCell
putTMVar connCell ref
routerRef <- atomically $ readTMVar routerCell
void $ forkIO $ do
result <- try @SomeException $ forever $
chRecv ch >>= \raw -> withRuntime rt $ cast' raw routerRef
case result of
Left _ -> atomically $
writeTQueue arDeathQ (DeathMessage (asId arState) Killed)
Right () -> return ()
RemoteRef _ -> return ()
}
]
liftIO $ atomically $ readTMVar connCell
rt <- ask
routerRef <- spawnActor (routerActorFn peer) stopOnDeath ()
connRef <- spawnActor connActorFn (connDeathFn peer) ch
-- When conn dies (peer disconnected), notify router so it shuts down too
case routerRef of
LocalRef {arDeathQ} -> linkActorTo (LocalTarget arDeathQ) connRef
RemoteRef _ -> return ()
-- And vice versa: if the router dies, close the connection down cleanly
case connRef of
LocalRef {arDeathQ} -> linkActorTo (LocalTarget arDeathQ) routerRef
RemoteRef _ -> return ()
liftIO $ case connRef of
LocalRef {arDeathQ, arState} ->
void $ forkIO $ do
result <- try @SomeException $ forever $
chRecv ch >>= \raw -> withRuntime rt $ cast' raw routerRef
case result of
Left _ -> atomically $
writeTQueue arDeathQ (DeathMessage (asId arState) Killed)
Right () -> return ()
RemoteRef _ -> return ()
return connRef
-- Incoming connection handler
@@ -110,13 +120,26 @@ handleNewConn ch = do
NMHandshake peerAddr -> do
ref <- spawnConnTree peerAddr ch
liftIO $ atomically $ do
modifyTVar (rtConnections rt) (Map.insert peerAddr ref)
conns <- readTVar (rtConnections rt)
-- On a simultaneous connect two sockets (and two conn trees) exist
-- for the same peer. Only the pool entry is deduplicated; the extra
-- tree must stay alive because it is the read path for whatever the
-- peer sends on its socket. It cleans itself up when that socket
-- closes. Local waiters get the pool winner, never the extra ref.
let winner = Map.findWithDefault ref peerAddr conns
unless (Map.member peerAddr conns) $
modifyTVar (rtConnections rt) (Map.insert peerAddr ref)
table <- readTVar (rtNodeTable rt)
unless (any (== peerAddr) (Map.elems table)) $ do
nid <- readTVar (rtNextNodeId rt)
writeTVar (rtNextNodeId rt) (nid + 1)
modifyTVar (rtNodeTable rt) (Map.insert nid peerAddr)
promises <- readTVar (rtConnPromises rt)
case Map.lookup peerAddr promises of
Nothing -> return ()
Just p -> do
modifyTVar (rtConnPromises rt) (Map.delete peerAddr)
void $ tryPutTMVar p ref
void $ tryPutTMVar p (Right winner)
_else -> liftIO $ chClose ch
-- Connection pool
@@ -138,18 +161,29 @@ getOrCreateConn peer = do
return (Right (Right p))
case action of
Left ref -> return ref
Right (Left promise) -> liftIO $ atomically $ readTMVar promise
Right (Left promise) -> do
-- Someone else is connecting; wait for their outcome so a failed
-- connect propagates instead of blocking us forever.
outcome <- liftIO $ atomically $ readTMVar promise
either (liftIO . throwIO) return outcome
Right (Right promise) -> do
ch <- liftIO $ tConnect (rtTransport rt) peer
liftIO $ chSend ch (encode (NMHandshake (rtNodeId rt)))
ref <- spawnConnTree peer ch
liftIO $ atomically $ do
conns <- readTVar (rtConnections rt)
unless (Map.member peer conns) $
modifyTVar (rtConnections rt) (Map.insert peer ref)
modifyTVar (rtConnPromises rt) (Map.delete peer)
void $ tryPutTMVar promise ref
return ref
result <- liftIO $ try @SomeException $ tConnect (rtTransport rt) peer
case result of
Left e -> do
liftIO $ atomically $ do
modifyTVar (rtConnPromises rt) (Map.delete peer)
void $ tryPutTMVar promise (Left e)
liftIO $ throwIO e
Right ch -> do
liftIO $ chSend ch (encode (NMHandshake (rtNodeId rt)))
ref <- spawnConnTree peer ch
liftIO $ atomically $ do
conns <- readTVar (rtConnections rt)
unless (Map.member peer conns) $
modifyTVar (rtConnections rt) (Map.insert peer ref)
modifyTVar (rtConnPromises rt) (Map.delete peer)
void $ tryPutTMVar promise (Right ref)
return ref
-- Message dispatch
@@ -162,7 +196,7 @@ routeRemoteDeath senderAddr (ActorId _ uuid) reason = do
exitReason = case reason of
RNormal -> Normal
RKilled -> Killed
RException s -> Exception (error s)
RException s -> Exception (toException (userError s))
dm = DeathMessage deadId exitReason
actors <- readTVarIO (rtActors rt)
forM_ (Map.elems actors) $ \(_, SomeActorRef ref) ->
@@ -195,7 +229,8 @@ validateAndDispatch senderAddr nm = case nm of
liftIO $ do
actors <- readTVarIO (rtActors rt)
case findByUUID uuid actors of
Nothing ->
Nothing -> do
putStrLn $ "didn't find actor: " <> show uuid
return False
Just (SomeActorRef (LocalRef {arMsgQ})) ->
case decodeOrFail payload of
@@ -221,12 +256,16 @@ validateAndDispatch senderAddr nm = case nm of
mv <- newEmptyMVar
atomically $ writeTQueue arMsgQ (rid, Call msg mv)
void $ forkIO $ do
reply <- takeMVar mv
-- Bounded wait: if the target actor dies before replying,
-- this thread must not linger forever.
reply <- timeout 30_000_000 (takeMVar mv)
case reply of
Nothing -> return ()
Just rv -> withRuntime rt $ do
connRef <- getOrCreateConn returnAddr
cast' (NMReply corrId (encode rv)) connRef
Nothing -> return ()
Just Nothing -> return ()
Just (Just rv) -> void $ try @SomeException $
withRuntime rt $ do
connRef <- getOrCreateConn returnAddr
cast' (NMReply corrId (encode rv)) connRef
return True
Just (SomeActorRef (RemoteRef _)) ->
return False
@@ -252,13 +291,16 @@ connect :: Maybe NodeId -> NodeAddr -> RuntimeM NodeId
connect suggestedId peer = do
rt <- ask
nodeId <- liftIO $ atomically $ do
table <- readTVar (rtNodeTable rt)
nid <- readTVar (rtNextNodeId rt)
let assigned = case suggestedId of
Just n | n /= 0 && not (Map.member n table) -> n
_else -> nid
writeTVar (rtNextNodeId rt) (max (assigned + 1) nid)
modifyTVar (rtNodeTable rt) (Map.insert assigned peer)
return assigned
table <- readTVar (rtNodeTable rt)
case Map.foldrWithKey (\k v acc -> if v == peer then Just k else acc) Nothing table of
Just existing -> return existing
Nothing -> do
nid <- readTVar (rtNextNodeId rt)
let assigned = case suggestedId of
Just n | n /= 0 && not (Map.member n table) -> n
_else -> nid
writeTVar (rtNextNodeId rt) (max (assigned + 1) nid)
modifyTVar (rtNodeTable rt) (Map.insert assigned peer)
return assigned
void $ getOrCreateConn peer
return nodeId
+14 -4
View File
@@ -5,13 +5,14 @@ module Control.Actor.Registry
registryUUID,
registry,
registerDeath,
deregisterNode,
createRegistry,
)
where
import Control.Actor.Core (ActorM, ActorResult (..), Handler, call, call', cast', liftRuntime, linkActorTo, pass, spawnActorAs, state, lastMessageFrom, cast, passWith)
import Control.Actor.Runtime (Runtime (..), RuntimeM, getActorByUUID, withRuntime)
import Control.Actor.Types (ActorId (..), ActorRef (..), DeathMessage (..), DeathTarget (LocalTarget), RegistryMsg (..), SomeActorRef (..), SupervisorAction (Continue), actorRefId, thisNodeId, findByUUID)
import Control.Actor.Types (ActorId (..), ActorRef (..), DeathMessage (..), DeathTarget (LocalTarget), NodeId, RegistryMsg (..), SomeActorRef (..), SupervisorAction (Continue), actorRefId, thisNodeId, findByUUID)
import Control.Concurrent.STM (readTVarIO)
import Control.Monad (forM_, when)
import Control.Monad.IO.Class (MonadIO (liftIO))
@@ -41,6 +42,9 @@ registryHandlerFn (RMLookup name) = do
registryHandlerFn (RMDeregister name) = do
u <- state
return $ ActorResult Nothing (Map.delete name u) Nothing
registryHandlerFn (RMDeregisterNode nodeId) = do
u <- state
passWith (Map.filter (\(ActorId nid _) -> nid /= nodeId) u)
registryHandlerFn msg@(RMDeath deadUUID) = do
u <- state
from <- lastMessageFrom
@@ -68,13 +72,21 @@ registryDeathFn (DeathMessage (ActorId _ deadUUID) _) = do
registry :: RuntimeM (Maybe (ActorRef RegistryMsg (Maybe ActorId)))
registry = do
unsafeCoerce $ getActorByUUID registryUUID
mRef <- getActorByUUID registryUUID
return $ case mRef of
Nothing -> Nothing
Just (SomeActorRef regRef) -> Just (unsafeCoerce regRef)
registerDeath :: ActorId -> RuntimeM ()
registerDeath (ActorId _ uuid) = do
r <- registry
mapM_ (cast' (RMDeath uuid)) r
deregisterNode :: NodeId -> RuntimeM ()
deregisterNode nodeId = do
r <- registry
mapM_ (cast' (RMDeregisterNode nodeId)) r
registerActor :: String -> ActorRef msg r -> RuntimeM ()
registerActor name ref = do
rt <- ask
@@ -122,8 +134,6 @@ lookupRemoteActor name = do
case result of
Just (Just (ActorId 0 uuid)) ->
return $ Just (RemoteRef (ActorId peerNodeId uuid))
Just (Just aid) ->
return $ Just (RemoteRef aid)
_else -> go rest
createRegistry :: RuntimeM (ActorRef RegistryMsg (Maybe ActorId))
+2 -1
View File
@@ -14,6 +14,7 @@ module Control.Actor.Runtime
import Control.Actor.Transport (Transport)
import Control.Actor.Types
import Control.Concurrent (ThreadId)
import Control.Exception (SomeException)
import Control.Concurrent.MVar (MVar)
import Control.Concurrent.STM
( TMVar, TVar, atomically, newTVarIO, readTVar, readTVarIO )
@@ -33,7 +34,7 @@ data Runtime = Runtime
, rtNodeTable :: TVar (Map.Map NodeId NodeAddr)
, rtTransport :: Transport
, rtConnections :: TVar (Map.Map NodeAddr (ActorRef NetworkMessage ()))
, rtConnPromises :: TVar (Map.Map NodeAddr (TMVar (ActorRef NetworkMessage ())))
, rtConnPromises :: TVar (Map.Map NodeAddr (TMVar (Either SomeException (ActorRef NetworkMessage ()))))
, rtSendRemote :: NodeAddr -> NetworkMessage -> RuntimeM ()
}
+49 -23
View File
@@ -23,10 +23,8 @@ import Data.Binary (Binary)
import Control.Concurrent (ThreadId, forkIO, killThread)
import Control.Concurrent.STM
( TMVar
, TQueue
, TVar
, atomically
, flushTQueue
, modifyTVar
, newTQueueIO
, newTVarIO
@@ -36,13 +34,16 @@ import Control.Concurrent.STM
, tryTakeTMVar
, writeTVar
)
import Control.Monad (forever, void)
import Control.Monad (void)
import GHC.Clock (getMonotonicTime)
import Control.Monad.Reader (MonadIO (..), MonadReader (..))
import Data.Map qualified as Map
data ChildSpec = forall m r. (Binary m, Binary r) => ChildSpec
{ csRun :: DeathTarget -> RuntimeM (ActorRef m r)
, csOnSpawn :: ActorRef m r -> IO ()
-- | Runs after every (re)spawn — including supervisor restarts — so it can
-- refresh shared ref cells and re-register names.
, csOnSpawn :: ActorRef m r -> RuntimeM ()
}
child ::
@@ -73,7 +74,7 @@ childWithRef msgFn deathFn initState cell =
ref <- spawnActor msgFn deathFn initState
linkActorTo target ref
return ref
, csOnSpawn = \ref -> atomically $ do
, csOnSpawn = \ref -> liftIO $ atomically $ do
void $ tryTakeTMVar cell
putTMVar cell ref
}
@@ -91,7 +92,7 @@ spawnSlot :: DeathTarget -> ChildSpec -> RuntimeM ChildSlot
spawnSlot target spec@ChildSpec {csRun, csOnSpawn} = do
rt <- ask
ref <- csRun target
_ <- liftIO $ csOnSpawn ref
csOnSpawn ref
let aid = someActorId (SomeActorRef ref)
tid <- liftIO $ do
actors <- readTVarIO (rtActors rt)
@@ -103,6 +104,15 @@ spawnSlot target spec@ChildSpec {csRun, csOnSpawn} = do
killSlot :: ChildSlot -> IO ()
killSlot ChildSlot {slotTid} = killThread slotTid
-- | Max child restarts within 'restartWindow' seconds before the supervisor
-- gives up and stops all children — otherwise a child that crashes right
-- after spawning would respawn in a hot loop forever.
restartLimit :: Int
restartLimit = 5
restartWindow :: Double
restartWindow = 10
supervise' :: RestartStrategy -> [ChildSpec] -> RuntimeM ()
supervise' strategy specs = do
rt <- ask
@@ -110,13 +120,35 @@ supervise' strategy specs = do
let target = LocalTarget supDeathQ
slots <- mapM (spawnSlot target) specs
slotsVar <- liftIO $ newTVarIO slots
liftIO $ void $ forkIO $ forever $ do
DeathMessage deadId _ <- atomically $ readTQueue supDeathQ
slots' <- readTVarIO slotsVar
withRuntime rt $ case strategy of
OneForOne -> doOneForOne target slotsVar slots' deadId
OneForAll -> doOneForAll target slotsVar supDeathQ slots'
RestForOne -> doRestForOne target slotsVar supDeathQ slots' deadId
let loop restarts = do
DeathMessage deadId reason <- atomically $ readTQueue supDeathQ
slots' <- readTVarIO slotsVar
if all ((/= deadId) . slotId) slots'
-- Stale death: killSlot is asynchronous, so children replaced by a
-- OneForAll/RestForOne restart report their deaths after the fact.
-- Acting on one would kill the fresh children and cascade forever.
then loop restarts
else case reason of
-- A normal exit is completion, not a crash: drop the slot.
Normal -> do
atomically $ modifyTVar slotsVar (filter ((/= deadId) . slotId))
loop restarts
_crash -> do
now <- getMonotonicTime
let recent = filter (> now - restartWindow) restarts
if length recent >= restartLimit
then do
putStrLn $ "supervisor: " <> show restartLimit
<> " restarts within " <> show restartWindow
<> "s, giving up"
mapM_ killSlot slots'
else do
withRuntime rt $ case strategy of
OneForOne -> doOneForOne target slotsVar slots' deadId
OneForAll -> doOneForAll target slotsVar slots'
RestForOne -> doRestForOne target slotsVar slots' deadId
loop (now : recent)
liftIO $ void $ forkIO $ loop []
supervise :: RestartStrategy -> [ChildSpec] -> ActorM u ()
supervise = (liftRuntime .) . supervise'
@@ -135,30 +167,24 @@ doOneForOne target slotsVar slots deadId =
doOneForAll ::
DeathTarget ->
TVar [ChildSlot] ->
TQueue DeathMessage ->
[ChildSlot] ->
RuntimeM ()
doOneForAll target slotsVar supDeathQ slots = do
liftIO $ do
mapM_ killSlot slots
atomically $ void $ flushTQueue supDeathQ
doOneForAll target slotsVar slots = do
liftIO $ mapM_ killSlot slots
newSlots <- mapM (spawnSlot target . slotSpec) slots
liftIO $ atomically $ writeTVar slotsVar newSlots
doRestForOne ::
DeathTarget ->
TVar [ChildSlot] ->
TQueue DeathMessage ->
[ChildSlot] ->
ActorId ->
RuntimeM ()
doRestForOne target slotsVar supDeathQ slots deadId = do
doRestForOne target slotsVar slots deadId = do
let (before, fromDead) = break (\s -> slotId s == deadId) slots
case fromDead of
[] -> return ()
_nonempty -> do
liftIO $ do
mapM_ killSlot (drop 1 fromDead)
atomically $ void $ flushTQueue supDeathQ
liftIO $ mapM_ killSlot (drop 1 fromDead)
newSlots <- mapM (spawnSlot target . slotSpec) fromDead
liftIO $ atomically $ writeTVar slotsVar (before ++ newSlots)
+6 -5
View File
@@ -85,7 +85,7 @@ data NetworkMessage
| NMCall UUID CorrelationId NodeAddr ByteString
| NMReply CorrelationId ByteString
| NMDeath ActorId RemoteExitReason
deriving (Generic)
deriving (Generic, Show)
instance Binary NetworkMessage
@@ -125,10 +125,11 @@ data SupervisorAction u
| Continue u
data RegistryMsg
= RMRegister String UUID
| RMLookup String
| RMDeregister String
| RMDeath UUID
= RMRegister String UUID
| RMLookup String
| RMDeregister String
| RMDeath UUID
| RMDeregisterNode NodeId
deriving (Generic)
instance Binary RegistryMsg
+43
View File
@@ -0,0 +1,43 @@
module FL.Actors.Aggregator
( AggregatorState (..)
, aggregatorHandler
) where
import Control.Actor.Core
import Control.Actor.Types
import Control.Concurrent.STM (TMVar, atomically, readTMVar)
import Control.Monad.IO.Class (liftIO)
import Data.List (sortOn)
import FL.Actors.Logger (logFn)
import FL.NN (fedAvg)
import FL.Types
data AggregatorState = AggregatorState
{ asExpected :: Int
, asUpdates :: [ModelUpdate]
, asCoordCell :: TMVar (ActorRef CoordinatorMsg ()) -- filled after coordinator spawns
, asLogRef :: ActorRef LogMsg ()
, asCurrentRound :: Int
}
aggregatorHandler :: Handler AggregatorMsg AggregatorState ()
aggregatorHandler (SubmitUpdate upd) = do
st <- state
if muRound upd /= asCurrentRound st
then pass
else do
let updates' = upd : filter (\u -> muTrainerId u /= muTrainerId upd) (asUpdates st)
if length updates' >= asExpected st
then do
let pairs = [(muSampleSize u, wData (muWeights u)) | u <- sortOn muTrainerId updates']
avgW = Weights (fedAvg pairs)
nextRound = asCurrentRound st + 1
newModel = GlobalModel nextRound avgW
logFn (asLogRef st) "Aggregator" INFO $
"Round " <> show (asCurrentRound st)
<> ": aggregated " <> show (length updates') <> " updates"
coordRef <- liftIO $ atomically $ readTMVar (asCoordCell st)
cast (AggregationComplete newModel) coordRef
passWith st { asUpdates = [], asCurrentRound = nextRound }
else
passWith st { asUpdates = updates' }
+103
View File
@@ -0,0 +1,103 @@
module FL.Actors.Coordinator
( CoordinatorState (..)
, coordinatorHandler
, coordinatorDeathHandler
) where
import Control.Actor.Core
import Control.Actor.Types
import Data.Binary (encodeFile)
import Control.Concurrent.MVar (MVar, putMVar)
import Control.Monad (forM_, when)
import Control.Monad.IO.Class (liftIO)
import Data.Map.Strict (Map)
import Data.Map.Strict qualified as Map
import Data.Set qualified as Set
import Data.UUID.V4 (nextRandom)
import FL.Actors.Logger (logFn)
import FL.CRDT (ORSet, orAdd, orMembers)
import FL.Types
data CoordinatorState = CoordinatorState
{ csTrainers :: ORSet String
, csTrainerRefs :: Map String (ActorRef TrainerMsg ())
, csExpectedTrainers :: Int
, csCurrentRound :: Int
, csTotalRounds :: Int
, csGlobalModel :: GlobalModel
, csAggRef :: ActorRef AggregatorMsg ()
, csEvalRef :: ActorRef EvaluatorMsg ()
, csLogRef :: ActorRef LogMsg ()
, csDoneVar :: MVar ()
}
coordinatorHandler :: Handler CoordinatorMsg CoordinatorState ()
coordinatorHandler (RegisterTrainer tid aid) = do
st <- state
-- The ActorId in the message carries the *sender's* node numbering (its own
-- node id is always 0 to itself); rebase it onto our node id for the sender,
-- otherwise every cast to the trainer is dropped ("no node with id 0").
from <- lastMessageFrom
uuid <- liftIO nextRandom
let ActorId _ trainerUUID = aid
ref = RemoteRef (ActorId from trainerUUID) :: ActorRef TrainerMsg ()
trainers' = orAdd uuid tid (csTrainers st)
refs' = Map.insert tid ref (csTrainerRefs st)
st' = st { csTrainers = trainers', csTrainerRefs = refs' }
numRegistered = Set.size (orMembers trainers')
logFn (csLogRef st) "Coordinator" INFO $
"Trainer registered: " <> tid
<> " (" <> show numRegistered <> "/" <> show (csExpectedTrainers st) <> ")"
-- Auto-start once all expected trainers have registered (per the OR-Set)
if numRegistered >= csExpectedTrainers st
then startRound st'
else passWith st'
coordinatorHandler StartFederation = do
st <- state
startRound st
coordinatorHandler (AggregationComplete newModel) = do
st <- state
logFn (csLogRef st) "Coordinator" INFO $
"Round " <> show (gmRound newModel - 1) <> " aggregation complete"
cast (EvaluateModel newModel) (csEvalRef st)
let st' = st { csGlobalModel = newModel, csCurrentRound = gmRound newModel }
if csCurrentRound st' >= csTotalRounds st'
then do
liftIO $ encodeFile "fl_model_coordinator.bin" (gmWeights newModel)
logFn (csLogRef st') "Coordinator" INFO
"Model saved to fl_model_coordinator.bin"
passWith st'
else
startRound st'
coordinatorHandler (EvaluationDone metrics) = do
st <- state
logFn (csLogRef st) "Coordinator" INFO $
"Eval round=" <> show (mRound metrics)
<> " acc=" <> show4 (mAccuracy metrics)
<> " prec=" <> show4 (mPrecision metrics)
<> " rec=" <> show4 (mRecall metrics)
<> " f1=" <> show4 (mF1 metrics)
<> " auc=" <> show4 (mAUCROC metrics)
liftIO $ when (mRound metrics >= csTotalRounds st) $
putMVar (csDoneVar st) ()
pass
where
show4 x = show (fromIntegral (round (x * 10000) :: Int) / 10000.0 :: Double)
startRound :: CoordinatorState -> Actor CoordinatorMsg CoordinatorState ()
startRound st = do
let model = csGlobalModel st
logFn (csLogRef st) "Coordinator" INFO $
"Starting round " <> show (gmRound model + 1)
<> " with " <> show (Map.size (csTrainerRefs st)) <> " trainers"
forM_ (Map.elems (csTrainerRefs st)) $ \ref ->
cast (StartTraining model) ref
passWith st
coordinatorDeathHandler :: DeathMessage -> ActorM CoordinatorState (SupervisorAction CoordinatorState)
coordinatorDeathHandler _ = do
st <- state
return (Continue st)
+36
View File
@@ -0,0 +1,36 @@
module FL.Actors.Evaluator
( EvaluatorState (..)
, evaluatorHandler
) where
import Control.Actor.Core
import Control.Actor.Types
import Control.Concurrent.STM (TMVar, atomically, readTMVar)
import Control.Monad.IO.Class (liftIO)
import FL.Actors.Logger (logFn)
import FL.NN (LocalData, evaluate, unflattenWeights)
import FL.Types
data EvaluatorState = EvaluatorState
{ esTestData :: LocalData
, esCoordCell :: TMVar (ActorRef CoordinatorMsg ()) -- filled after coordinator spawns
, esLogRef :: ActorRef LogMsg ()
}
evaluatorHandler :: Handler EvaluatorMsg EvaluatorState ()
evaluatorHandler (EvaluateModel gm) = do
st <- state
let ws = unflattenWeights (wData (gmWeights gm))
metrics = evaluate ws (esTestData st) (gmRound gm)
logFn (esLogRef st) "Evaluator" INFO $
"Round " <> show (gmRound gm)
<> " acc=" <> show4 (mAccuracy metrics)
<> " prec=" <> show4 (mPrecision metrics)
<> " rec=" <> show4 (mRecall metrics)
<> " f1=" <> show4 (mF1 metrics)
<> " auc=" <> show4 (mAUCROC metrics)
coordRef <- liftIO $ atomically $ readTMVar (esCoordCell st)
cast (EvaluationDone metrics) coordRef
pass
where
show4 x = show (fromIntegral (round (x * 10000) :: Int) / 10000.0 :: Double)
+37
View File
@@ -0,0 +1,37 @@
module FL.Actors.Logger
( loggerHandler
, logFn
, logFn'
) where
import Control.Actor.Core (ActorM, Handler, cast, cast', pass)
import Control.Actor.Runtime (RuntimeM)
import Control.Actor.Types
import Control.Monad.IO.Class (liftIO)
import Data.Time.Clock (getCurrentTime)
import FL.Types
type LogState = ()
loggerHandler :: Handler LogMsg LogState ()
loggerHandler (Log LogEntry{leSource, leLevel, leMsg}) = do
t <- liftIO getCurrentTime
liftIO $ putStrLn $
"[" <> show t <> "] [" <> showLevel leLevel <> "] "
<> leSource <> ": " <> leMsg
pass
where
showLevel DEBUG = "DEBUG"
showLevel INFO = "INFO "
showLevel WARN = "WARN "
showLevel ERROR = "ERROR"
-- | Log via the logger actor. Goes through the normal cast path so it works
-- for both local and remote logger refs (a remote ref used to be a silent
-- no-op, losing all trainer-node logs in provider mode).
logFn :: ActorRef LogMsg () -> String -> LogLevel -> String -> ActorM u ()
logFn ref src lvl msg = cast (Log (LogEntry src lvl msg)) ref
-- | Same as 'logFn' but usable outside an actor (node startup code).
logFn' :: ActorRef LogMsg () -> String -> LogLevel -> String -> RuntimeM ()
logFn' ref src lvl msg = cast' (Log (LogEntry src lvl msg)) ref
+166
View File
@@ -0,0 +1,166 @@
{-# LANGUAGE StrictData #-}
module FL.Actors.PeerCoordinator
( PeerCoordinatorState (..)
, peerCoordinatorHandler
, peerCoordinatorDeathHandler
) where
import Control.Actor.Core
import Control.Actor.Types
import Data.Binary (encodeFile)
import Unsafe.Coerce (unsafeCoerce)
import Control.Concurrent.MVar (MVar, putMVar)
import Control.Monad (forM_, when)
import Control.Monad.IO.Class (liftIO)
import Data.List (sortOn)
import Data.Set qualified as Set
import Data.UUID.V4 (nextRandom)
import FL.Actors.Logger (logFn)
import FL.CRDT
import FL.NN (fedAvg)
import FL.Types
data PeerCoordinatorState = PeerCoordinatorState
{ pcsNodeId :: String
, pcsRoundCounter :: GCounter -- completed training rounds per peer (monotone, never reset)
, pcsParticipants :: ORSet String -- active peers ever seen (monotone, never reset)
, pcsCurrentRound :: Int
, pcsTotalRounds :: Int
, pcsGlobalModel :: GlobalModel
, pcsPeerRefs :: [ActorRef PeerCoordinatorMsg ()]
, pcsSelfRef :: ActorRef PeerCoordinatorMsg ()
, pcsTrainerRef :: ActorRef TrainerMsg ()
, pcsEvalRef :: ActorRef EvaluatorMsg ()
, pcsLogRef :: ActorRef LogMsg ()
, pcsCollected :: [ModelUpdate] -- updates for the current round + buffered future rounds
, pcsNumPeers :: Int -- total number of peers (including self)
, pcsDoneVar :: MVar ()
}
-- | Insert an update, deduplicating on (trainer, round). Stale (past-round)
-- updates are dropped; future-round updates from fast peers are buffered so
-- they are not lost — without the round key a fast peer's round-(r+1) update
-- would overwrite its round-r update in a slower peer and rounds would mix.
insertUpdate :: Int -> ModelUpdate -> [ModelUpdate] -> [ModelUpdate]
insertUpdate currentRound upd collected
| muRound upd < currentRound = collected
| otherwise =
upd : filter (\u -> (muTrainerId u, muRound u) /= (muTrainerId upd, muRound upd)) collected
currentRoundUpdates :: PeerCoordinatorState -> [ModelUpdate]
currentRoundUpdates st = [u | u <- pcsCollected st, muRound u == pcsCurrentRound st]
-- | A round is complete when the CRDTs agree with the collected updates:
-- the OR-Set has seen all peers, the G-Counter shows every one of them has
-- finished training the current round, and we hold all their updates.
roundComplete :: PeerCoordinatorState -> Bool
roundComplete st =
let r = pcsCurrentRound st
members = orMembers (pcsParticipants st)
trained = Set.filter (\p -> gcGet p (pcsRoundCounter st) > r) members
in Set.size members >= pcsNumPeers st
&& Set.size trained >= pcsNumPeers st
&& length (currentRoundUpdates st) >= pcsNumPeers st
peerCoordinatorHandler :: Handler PeerCoordinatorMsg PeerCoordinatorState ()
peerCoordinatorHandler (SetPeers peerIds) = do
st <- state
(SomeActorRef self) <- getSelf
let peers = map (\aid -> RemoteRef aid :: ActorRef PeerCoordinatorMsg ()) peerIds
selfRef = unsafeCoerce self :: ActorRef PeerCoordinatorMsg ()
passWith st { pcsPeerRefs = peers, pcsSelfRef = selfRef }
peerCoordinatorHandler PeerStartRound = do
st <- state
logFn (pcsLogRef st) (pcsNodeId st) INFO $
"Starting local training for P2P round " <> show (pcsCurrentRound st + 1)
cast (StartTraining (pcsGlobalModel st)) (pcsTrainerRef st)
pass
peerCoordinatorHandler (PeerTrainingDone upd) = do
st <- state
uuid <- liftIO nextRandom
let counter' = gcIncrement (pcsNodeId st) (pcsRoundCounter st)
participants' = orAdd uuid (pcsNodeId st) (pcsParticipants st)
collected' = insertUpdate (pcsCurrentRound st) upd (pcsCollected st)
payload = PeerSyncPayload
{ pspCounter = counter'
, pspParticipants = participants'
, pspUpdate = Just upd
}
logFn (pcsLogRef st) (pcsNodeId st) INFO
"Training done, broadcasting PeerSync"
forM_ (pcsPeerRefs st) $ \ref ->
cast (PeerSync payload) ref
let st' = st { pcsRoundCounter = counter'
, pcsParticipants = participants'
, pcsCollected = collected'
}
if roundComplete st'
then finishRound st'
else passWith st'
peerCoordinatorHandler (PeerSync payload) = do
st <- state
-- Merge CRDTs
let counter' = gcMerge (pcsRoundCounter st) (pspCounter payload)
participants' = orMerge (pcsParticipants st) (pspParticipants payload)
collected' = case pspUpdate payload of
Nothing -> pcsCollected st
Just upd -> insertUpdate (pcsCurrentRound st) upd (pcsCollected st)
st' = st { pcsRoundCounter = counter'
, pcsParticipants = participants'
, pcsCollected = collected'
}
if roundComplete st'
then finishRound st'
else passWith st'
peerCoordinatorHandler (PeerEvaluationDone metrics) = do
st <- state
logFn (pcsLogRef st) (pcsNodeId st) INFO $
"Eval round=" <> show (mRound metrics)
<> " acc=" <> show4 (mAccuracy metrics)
<> " auc=" <> show4 (mAUCROC metrics)
liftIO $ when (mRound metrics >= pcsTotalRounds st) $
putMVar (pcsDoneVar st) ()
pass
where
show4 x = show (fromIntegral (round (x * 10000) :: Int) / 10000.0 :: Double)
finishRound :: PeerCoordinatorState -> Actor PeerCoordinatorMsg PeerCoordinatorState ()
finishRound st = do
-- Sort so every peer folds the weighted average in the same order and
-- arrives at bit-identical models regardless of message arrival order.
let updates = sortOn muTrainerId (currentRoundUpdates st)
pairs = [(muSampleSize u, wData (muWeights u)) | u <- updates]
avgW = Weights (fedAvg pairs)
nextRound = pcsCurrentRound st + 1
newModel = GlobalModel nextRound avgW
logFn (pcsLogRef st) (pcsNodeId st) INFO $
"P2P round " <> show (pcsCurrentRound st) <> " aggregated ("
<> show (length updates) <> " updates)"
cast (EvaluateModel newModel) (pcsEvalRef st)
let st' = st
{ pcsGlobalModel = newModel
, pcsCurrentRound = nextRound
-- Keep buffered updates from peers already in later rounds
, pcsCollected = [u | u <- pcsCollected st, muRound u >= nextRound]
}
if nextRound >= pcsTotalRounds st
then do
let filename = "fl_model_" <> pcsNodeId st' <> ".bin"
liftIO $ encodeFile filename (gmWeights newModel)
logFn (pcsLogRef st') (pcsNodeId st') INFO $
"Model saved to " <> filename
passWith st'
else do
-- Schedule next round after a short delay to let PeerSync propagate
castIn 500 PeerStartRound (pcsSelfRef st')
passWith st'
peerCoordinatorDeathHandler :: DeathMessage -> ActorM PeerCoordinatorState (SupervisorAction PeerCoordinatorState)
peerCoordinatorDeathHandler _ = do
st <- state
return (Continue st)
+45
View File
@@ -0,0 +1,45 @@
module FL.Actors.Trainer
( TrainerState (..)
, trainerHandler
) where
import Control.Actor.Core
import Control.Actor.Types
import Control.Monad.IO.Class (liftIO)
import FL.Actors.Logger (logFn)
import FL.NN (LocalData, flattenWeights, localDataSize, trainLocal, unflattenWeights)
import FL.Types
data TrainerState = TrainerState
{ tsId :: String
, tsData :: LocalData
, tsLastRound :: Int
, tsAggRef :: ActorRef AggregatorMsg ()
, tsLogRef :: ActorRef LogMsg ()
, tsConfig :: FLConfig
}
trainerHandler :: Handler TrainerMsg TrainerState ()
trainerHandler (StartTraining gm) = do
st <- state
let round_ = gmRound gm
if round_ <= tsLastRound st
then pass -- idempotent: already processed this round
else do
logFn (tsLogRef st) (tsId st) INFO $
"Starting local training for round " <> show round_
let initW = unflattenWeights (wData (gmWeights gm))
updatedW <- liftIO $ trainLocal (tsConfig st) (tsId st) round_ initW (tsData st)
let n = localDataSize (tsData st)
update = ModelUpdate
{ muTrainerId = tsId st
, muRound = round_
, muSampleSize = n
, muWeights = Weights (flattenWeights updatedW)
}
logFn (tsLogRef st) (tsId st) INFO $
"Round " <> show round_ <> " training complete, samples=" <> show n
cast (SubmitUpdate update) (tsAggRef st)
passWith st { tsLastRound = round_ }
trainerHandler TrainerStop = pass
+82
View File
@@ -0,0 +1,82 @@
module FL.CRDT
( GCounter
, gcEmpty
, gcIncrement
, gcMerge
, gcValue
, gcGet
, ORSet
, orEmpty
, orAdd
, orRemove
, orMerge
, orMembers
) where
import Data.Binary (Binary (..))
import Data.Map.Strict (Map)
import Data.Map.Strict qualified as Map
import Data.Set (Set)
import Data.Set qualified as Set
import Data.UUID (UUID)
import GHC.Generics (Generic)
-- ---------------------------------------------------------------------------
-- G-Counter: each node has its own non-decreasing counter.
-- Merge takes element-wise max. Value is the sum.
-- ---------------------------------------------------------------------------
newtype GCounter = GCounter (Map String Int)
deriving (Show, Eq, Generic)
instance Binary GCounter where
put (GCounter m) = put (Map.toAscList m)
get = GCounter . Map.fromAscList <$> get
gcEmpty :: GCounter
gcEmpty = GCounter Map.empty
gcIncrement :: String -> GCounter -> GCounter
gcIncrement nodeId (GCounter m) =
GCounter (Map.insertWith (+) nodeId 1 m)
gcMerge :: GCounter -> GCounter -> GCounter
gcMerge (GCounter a) (GCounter b) =
GCounter (Map.unionWith max a b)
gcValue :: GCounter -> Int
gcValue (GCounter m) = sum (Map.elems m)
gcGet :: String -> GCounter -> Int
gcGet nodeId (GCounter m) = Map.findWithDefault 0 nodeId m
-- ---------------------------------------------------------------------------
-- OR-Set: observe-remove set. Each element tagged with a set of UUIDs.
-- Add: insert a fresh UUID tag. Remove: delete all tags for that element.
-- Merge: union of tag sets per element. An element is present iff it has
-- at least one tag.
-- ---------------------------------------------------------------------------
newtype ORSet a = ORSet (Map a (Set UUID))
deriving (Show, Eq, Generic)
instance (Ord a, Binary a) => Binary (ORSet a) where
put (ORSet m) = put (Map.toAscList (fmap Set.toAscList m))
get = ORSet . Map.fromAscList . map (fmap Set.fromAscList) <$> get
orEmpty :: ORSet a
orEmpty = ORSet Map.empty
orAdd :: Ord a => UUID -> a -> ORSet a -> ORSet a
orAdd tag x (ORSet m) =
ORSet (Map.insertWith Set.union x (Set.singleton tag) m)
orRemove :: Ord a => a -> ORSet a -> ORSet a
orRemove x (ORSet m) = ORSet (Map.delete x m)
orMerge :: Ord a => ORSet a -> ORSet a -> ORSet a
orMerge (ORSet a) (ORSet b) =
ORSet (Map.unionWith Set.union a b)
orMembers :: Ord a => ORSet a -> Set a
orMembers (ORSet m) = Map.keysSet (Map.filter (not . Set.null) m)
+231
View File
@@ -0,0 +1,231 @@
{-# LANGUAGE RecordWildCards #-}
{-# LANGUAGE OverloadedStrings #-}
module FL.Data.PaySim
( loadAndPartition
-- , syntheticPartition
) where
import Data.ByteString.Lazy qualified as BL
import Data.Csv qualified as Csv
import Data.List (sortBy)
import Data.Ord (comparing)
import Data.Vector qualified as V
import FL.NN (LocalData (..))
import FL.Types (FLConfig (..), NormMode (..), PartitionMode (..))
import GHC.Generics (Generic)
import Numeric.LinearAlgebra hiding ((<>))
import Numeric.LinearAlgebra qualified as LA
import System.Random (mkStdGen, randomRs)
-- ---------------------------------------------------------------------------
-- PaySim CSV record
-- ---------------------------------------------------------------------------
data PaySimRow = PaySimRow
{ rowStep :: !Double
, rowType :: !String
, rowAmount :: !Double
, rowOldBalOrig :: !Double
, rowNewBalOrig :: !Double
, rowOldBalDest :: !Double
, rowNewBalDest :: !Double
, rowIsFraud :: !Double
} deriving (Show, Generic)
instance Csv.FromNamedRecord PaySimRow where
parseNamedRecord r =
PaySimRow
<$> r Csv..: "step"
<*> r Csv..: "type"
<*> r Csv..: "amount"
<*> r Csv..: "oldbalanceOrg"
<*> r Csv..: "newbalanceOrig"
<*> r Csv..: "oldbalanceDest"
<*> r Csv..: "newbalanceDest"
<*> r Csv..: "isFraud"
rowToFeatures :: PaySimRow -> Vector Double
rowToFeatures PaySimRow{..} =
let typeVec = case rowType of
"CASH_IN" -> [1,0,0,0,0]
"CASH_OUT" -> [0,1,0,0,0]
"DEBIT" -> [0,0,1,0,0]
"PAYMENT" -> [0,0,0,1,0]
"TRANSFER" -> [0,0,0,0,1]
_ -> [0,0,0,0,0]
numeric = [ rowAmount
, rowOldBalOrig
, rowNewBalOrig
, rowOldBalDest
, rowNewBalDest
, rowNewBalOrig - rowOldBalOrig -- balance change orig
, rowNewBalDest - rowOldBalDest -- balance change dest
, rowAmount / (rowOldBalOrig + 1) -- amount ratio orig
, rowAmount / (rowOldBalDest + 1) -- amount ratio dest
, rowStep / 744 -- normalized step (max 744 in PaySim)
]
-- pad to 29 features (5 one-hot + 10 numeric + 14 zeros)
padded = typeVec ++ numeric ++ replicate 14 0.0
in fromList padded
-- ---------------------------------------------------------------------------
-- Load PaySim CSV
-- ---------------------------------------------------------------------------
loadPaySimCSV :: FilePath -> IO (Either String (Matrix Double, Vector Double))
loadPaySimCSV path = do
raw <- BL.readFile path
case Csv.decodeByName raw of
Left err -> return (Left err)
Right (_, rows) -> do
let featureList = V.toList (V.map rowToFeatures rows)
labelList = V.toList (V.map rowIsFraud rows)
return (Right (fromRows featureList, fromList labelList))
-- ---------------------------------------------------------------------------
-- Normalization
-- ---------------------------------------------------------------------------
normalizeMinMax :: Matrix Double -> (Matrix Double, Vector Double, Vector Double)
normalizeMinMax m =
let mins = fromList [minElement (m ¿ [j]) | j <- [0 .. cols m - 1]]
maxs = fromList [maxElement (m ¿ [j]) | j <- [0 .. cols m - 1]]
range = cmap (\x -> if x == 0 then 1 else x) (maxs - mins)
norm = (m - LA.repmat (asRow mins) (rows m) 1)
/ LA.repmat (asRow range) (rows m) 1
in (norm, mins, maxs)
normalizeZScore :: Matrix Double -> (Matrix Double, Vector Double, Vector Double)
normalizeZScore m =
let n = fromIntegral (rows m) :: Double
means = fromList [sumElements (m ¿ [j]) / n | j <- [0 .. cols m - 1]]
vars = fromList [sumElements (cmap (^(2::Int)) ((m ¿ [j]) - scalar (means ! j))) / n
| j <- [0 .. cols m - 1]]
stds = cmap (\v -> if v == 0 then 1 else sqrt v) vars
norm = (m - LA.repmat (asRow means) (rows m) 1)
/ LA.repmat (asRow stds) (rows m) 1
in (norm, means, stds)
applyNorm :: NormMode -> (Matrix Double, Vector Double, Vector Double) -> Matrix Double -> Matrix Double
applyNorm mode (_, p1, p2) m =
case mode of
MinMax ->
let range = cmap (\x -> if x == 0 then 1 else x) (p2 - p1)
in (m - LA.repmat (asRow p1) (rows m) 1) / LA.repmat (asRow range) (rows m) 1
ZScore ->
let stds = cmap (\x -> if x == 0 then 1 else x) p2
in (m - LA.repmat (asRow p1) (rows m) 1) / LA.repmat (asRow stds) (rows m) 1
-- ---------------------------------------------------------------------------
-- IID partition: deterministic shard by trainer index
-- ---------------------------------------------------------------------------
partitionIID :: Int -> Int -> Matrix Double -> Vector Double -> LocalData
partitionIID trainerIdx numTrainers feats labels =
let n = rows feats
-- Shuffle with fixed seed for reproducibility across processes
seed = 2024
rng = mkStdGen seed
sorted = sortBy (comparing fst)
(zip (randomRs (0.0 :: Double, 1.0) rng) [0 .. n - 1])
idxs = map snd sorted
shard = chunkedShard trainerIdx numTrainers idxs
f = feats ? shard
l = fromList [labels ! i | i <- shard]
in LocalData f l
chunkedShard :: Int -> Int -> [a] -> [a]
chunkedShard idx total xs =
let n = length xs
size = n `div` total
start = idx * size
end = if idx == total - 1 then n else start + size
in take (end - start) (drop start xs)
-- ---------------------------------------------------------------------------
-- Non-IID partition: each trainer gets dominant transaction type
-- (approximated via deterministic row assignment)
-- ---------------------------------------------------------------------------
partitionNonIID :: Int -> Int -> Matrix Double -> Vector Double -> LocalData
partitionNonIID trainerIdx numTrainers feats labels =
let n = rows feats
-- Assign rows to trainers based on dominant type one-hot (cols 04)
-- Trainer i "owns" type (i mod 5); remaining trainers share type 0
ownType = trainerIdx `mod` 5
isOwn i = let row = toList (flatten (feats ? [i]))
in length row > ownType && row !! ownType > 0.5
-- With more than 5 trainers several trainers own the same type;
-- split that type's rows among them instead of giving each trainer
-- all of them (which duplicated data across shards).
sharers = [i | i <- [0 .. numTrainers - 1], i `mod` 5 == ownType]
shareIdx = length (takeWhile (/= trainerIdx) sharers)
-- Gather own-type rows + 10% of other rows for minority
ownIdxs = chunkedShard shareIdx (length sharers) [i | i <- [0 .. n-1], isOwn i]
otherIdxs = [i | i <- [0 .. n-1], not (isOwn i)]
minority = chunkedShard trainerIdx numTrainers otherIdxs
allIdxs = ownIdxs ++ minority
f = feats ? allIdxs
l = fromList [labels ! i | i <- allIdxs]
in LocalData f l
-- ---------------------------------------------------------------------------
-- Synthetic data fallback
-- ---------------------------------------------------------------------------
syntheticData :: Int -> Int -> Int -> LocalData
syntheticData trainerIdx numTrainers nSamples =
let seed = trainerIdx * 13337 + numTrainers
rng = mkStdGen seed
-- 29 features from normal distribution
featVals = take (nSamples * 29) (randomRs (-3.0 :: Double, 3.0) rng)
feats = (nSamples >< 29) featVals
-- Base 2% fraud rate, +1% per trainer index for variety (real PaySim
-- is ~0.1%, but synthetic runs need enough positives to learn from)
fraudRate = 0.02 + fromIntegral trainerIdx * 0.01
rng2 = mkStdGen (seed + 1)
labelVals = take nSamples (randomRs (0.0 :: Double, 1.0) rng2)
labels = fromList [if v < fraudRate then 1.0 else 0.0 | v <- labelVals]
in LocalData feats labels
-- ---------------------------------------------------------------------------
-- Top-level entry point used by Provider and P2P startup
-- ---------------------------------------------------------------------------
-- Returns (train partition, test data).
-- Test data is the first 10% of the full dataset (at least 1000 rows),
-- taken before partitioning so every node evaluates on the same split.
loadAndPartition
:: FLConfig
-> Int -- this node's trainer index (0-based)
-> IO (LocalData, LocalData)
loadAndPartition FLConfig{..} trainerIdx = do
result <- case cfgDataPath of
Nothing -> return (Left "no data path")
Just path -> loadPaySimCSV path
case result of
Left _ -> do
-- Synthetic fallback
let testData = syntheticData (cfgNumTrainers + 1) cfgNumTrainers 2000
trainData = syntheticData trainerIdx cfgNumTrainers 10000
return (trainData, testData)
Right (feats, labels) -> do
let n = rows feats
-- Aim for 1000 test rows for stable metrics, but never take more
-- than half the data — max alone made testN exceed n on small
-- CSVs and crashed the train subMatrix with a negative size.
testN = min (max 1000 (n `div` 10)) (n `div` 2)
testFeats = subMatrix (0, 0) (testN, cols feats) feats
testLabs = subVector 0 testN labels
trainF = subMatrix (testN, 0) (n - testN, cols feats) feats
trainL = subVector testN (n - testN) labels
(normF, p1, p2) = case cfgNorm of
MinMax -> normalizeMinMax trainF
ZScore -> normalizeZScore trainF
normTest = applyNorm cfgNorm (normF, p1, p2) testFeats
trainData = case cfgPartition of
IID -> partitionIID trainerIdx cfgNumTrainers normF trainL
NonIID -> partitionNonIID trainerIdx cfgNumTrainers normF trainL
testData = LocalData normTest testLabs
return (trainData, testData)
+290
View File
@@ -0,0 +1,290 @@
{-# LANGUAGE RecordWildCards #-}
{-# LANGUAGE StrictData #-}
module FL.NN
( MLPWeights (..)
, LocalData (..)
, initWeights
, trainLocal
, evaluate
, flattenWeights
, unflattenWeights
, fedAvg
, localDataSize
) where
import Data.List (foldl', sortBy)
import Data.Ord (Down (..), comparing)
import FL.Types (FLConfig (..), Metrics (..))
import Numeric.LinearAlgebra hiding ((<>))
import Numeric.LinearAlgebra qualified as LA
import System.Random (mkStdGen, randomRs)
-- ---------------------------------------------------------------------------
-- Types
-- ---------------------------------------------------------------------------
data MLPWeights = MLPWeights
{ w1 :: Matrix Double -- 64 × 29
, b1 :: Vector Double -- 64
, w2 :: Matrix Double -- 32 × 64
, b2 :: Vector Double -- 32
, w3 :: Matrix Double -- 2 × 32
, b3 :: Vector Double -- 2
}
data LocalData = LocalData
{ ldFeatures :: Matrix Double -- N × 29
, ldLabels :: Vector Double -- N (0.0 or 1.0)
}
localDataSize :: LocalData -> Int
localDataSize = rows . ldFeatures
-- ---------------------------------------------------------------------------
-- Weight initialization (Xavier uniform)
-- ---------------------------------------------------------------------------
xavierMatrix :: Int -> Int -> Int -> Matrix Double
xavierMatrix seed fanIn fanOut =
let bound = sqrt (6.0 / fromIntegral (fanIn + fanOut))
n = fanIn * fanOut
vals = take n (randomRs (-bound, bound) (mkStdGen seed))
in (fanOut >< fanIn) vals
initWeights :: IO MLPWeights
initWeights = do
-- Use fixed seeds so multiple processes with same call order get different
-- weight initializations only if we pass different seeds; for now deterministic.
let w1' = xavierMatrix 42 29 64
b1' = konst 0 64
w2' = xavierMatrix 137 64 32
b2' = konst 0 32
w3' = xavierMatrix 999 32 2
b3' = konst 0 2
return MLPWeights{ w1=w1', b1=b1', w2=w2', b2=b2', w3=w3', b3=b3' }
-- ---------------------------------------------------------------------------
-- Activation functions
-- ---------------------------------------------------------------------------
relu :: Matrix Double -> Matrix Double
relu = cmap (max 0)
reluGrad :: Matrix Double -> Matrix Double
reluGrad = cmap (\x -> if x > 0 then 1 else 0)
softmaxRows :: Matrix Double -> Matrix Double
softmaxRows m = fromRows (map softmaxVec (toRows m))
where
-- Subtract the row max before exponentiating so large logits
-- can't overflow exp and poison the weights with NaN.
softmaxVec v =
let mx = maxElement v
e = cmap (\x -> exp (x - mx)) v
s = sumElements e
in cmap (/ s) e
-- ---------------------------------------------------------------------------
-- Forward pass
-- ---------------------------------------------------------------------------
-- Returns (output probs N×2, z1, a1, z2, a2) for backprop.
forwardWithCache
:: MLPWeights
-> Matrix Double -- N × 29
-> (Matrix Double, Matrix Double, Matrix Double, Matrix Double, Matrix Double)
forwardWithCache MLPWeights{..} x =
let z1 = x LA.<> tr w1 + repRows (rows x) b1
a1 = relu z1
z2 = a1 LA.<> tr w2 + repRows (rows x) b2
a2 = relu z2
z3 = a2 LA.<> tr w3 + repRows (rows x) b3
out = softmaxRows z3
in (out, z1, a1, z2, a2)
repRows :: Int -> Vector Double -> Matrix Double
repRows n v = fromRows (replicate n v)
colSums :: Matrix Double -> Vector Double
colSums m = fromList [sumElements (m ¿ [j]) | j <- [0 .. cols m - 1]]
-- ---------------------------------------------------------------------------
-- Loss
-- ---------------------------------------------------------------------------
crossEntropyLoss :: Matrix Double -> Vector Double -> Double
crossEntropyLoss probs labels =
let n = fromIntegral (rows probs)
logProbs = cmap (log . max 1e-15) probs
oneHot = buildOneHot (rows probs) (map round (toList labels))
loss = negate (sumElements (oneHot * logProbs)) / n
in loss
buildOneHot :: Int -> [Int] -> Matrix Double
buildOneHot n ys =
fromRows [fromList [if i == y then 1.0 else 0.0 | i <- [0, 1]] | y <- ys]
-- ---------------------------------------------------------------------------
-- Backward pass (mini-batch SGD)
-- ---------------------------------------------------------------------------
stepBatch
:: MLPWeights
-> Matrix Double -- batch features N × 29
-> Vector Double -- batch labels N
-> Double -- learning rate
-> (MLPWeights, Double)
stepBatch ws@MLPWeights{..} x labels lr =
let n = fromIntegral (rows x) :: Double
(out, z1, a1, z2, a2) = forwardWithCache ws x
oneHot = buildOneHot (rows x) (map round (toList labels))
-- Output delta: dL/dZ3 = (probs - oneHot) / N
dZ3 = cmap (/ n) (out - oneHot) -- N × 2
dW3 = tr dZ3 LA.<> a2 -- 2 × 32
db3 = colSums dZ3 -- 2
dA2 = dZ3 LA.<> w3 -- N × 32
dZ2 = dA2 * reluGrad z2 -- N × 32
dW2 = tr dZ2 LA.<> a1 -- 32 × 64
db2 = colSums dZ2 -- 32
dA1 = dZ2 LA.<> w2 -- N × 64
dZ1 = dA1 * reluGrad z1 -- N × 64
dW1 = tr dZ1 LA.<> x -- 64 × 29
db1 = colSums dZ1 -- 64
ws' = MLPWeights
{ w1 = w1 - cmap (* lr) dW1
, b1 = b1 - cmap (* lr) db1
, w2 = w2 - cmap (* lr) dW2
, b2 = b2 - cmap (* lr) db2
, w3 = w3 - cmap (* lr) dW3
, b3 = b3 - cmap (* lr) db3
}
loss = crossEntropyLoss out labels
in (ws', loss)
-- ---------------------------------------------------------------------------
-- Training
-- ---------------------------------------------------------------------------
-- Shuffle row indices with a seed, then split into mini-batches.
shuffleIndices :: Int -> Int -> [Int]
shuffleIndices seed n =
let rng = mkStdGen seed
tagged = zip (randomRs (0.0 :: Double, 1.0) rng) [0 .. n - 1]
in map snd (sortBy (comparing fst) tagged)
trainEpoch
:: MLPWeights
-> Matrix Double -- features N × 29
-> Vector Double -- labels N
-> Int -- batch size
-> Double -- learning rate
-> Int -- seed for shuffle
-> (MLPWeights, Double)
trainEpoch ws feats labels batchSize lr seed =
let n = rows feats
idxs = shuffleIndices seed n
batches = chunksOf batchSize idxs
step (w, totalLoss) batch =
let bFeats = feats ? batch
bLabels = fromList [labels ! i | i <- batch]
(w', l) = stepBatch w bFeats bLabels lr
in (w', totalLoss + l)
(wFinal, lossSum) = foldl' step (ws, 0.0) batches
in (wFinal, lossSum / fromIntegral (length batches))
chunksOf :: Int -> [a] -> [[a]]
chunksOf _ [] = []
chunksOf n xs = let (h, t) = splitAt n xs in h : chunksOf n t
-- | Train for the configured number of epochs. The shuffle seed is derived
-- from the trainer id and round so batches differ across rounds and trainers
-- (a bare epoch counter would replay the identical shuffle every round).
trainLocal :: FLConfig -> String -> Int -> MLPWeights -> LocalData -> IO MLPWeights
trainLocal FLConfig{..} trainerId roundNum initW LocalData{..} = do
let seedBase = foldl' (\h c -> 31 * h + fromEnum c) (roundNum * 7919) trainerId
go ws epoch =
let (ws', _loss) = trainEpoch ws ldFeatures ldLabels cfgBatchSize cfgLearningRate (seedBase + epoch)
in if epoch >= cfgEpochs then ws'
else go ws' (epoch + 1)
return (go initW 1)
-- ---------------------------------------------------------------------------
-- Evaluation
-- ---------------------------------------------------------------------------
evaluate :: MLPWeights -> LocalData -> Int -> Metrics
evaluate ws LocalData{..} round_ =
let n = rows ldFeatures
(probs, _, _, _, _) = forwardWithCache ws ldFeatures
predLabels = fromList [if (probs ! i ! 1) >= 0.5 then 1.0 else 0.0 | i <- [0 .. n - 1]]
scores = fromList [probs ! i ! 1 | i <- [0 .. n - 1]]
tp = sumElements (predLabels * ldLabels)
fp = sumElements (predLabels * (1 - ldLabels))
fn = sumElements ((1 - predLabels) * ldLabels)
tn = sumElements ((1 - predLabels) * (1 - ldLabels))
acc = (tp + tn) / fromIntegral n
prec = if tp + fp == 0 then 0 else tp / (tp + fp)
rec = if tp + fn == 0 then 0 else tp / (tp + fn)
f1 = if prec + rec == 0 then 0 else 2 * prec * rec / (prec + rec)
auc = computeAUCROC (toList scores) (toList ldLabels)
in Metrics acc prec rec f1 auc round_
computeAUCROC :: [Double] -> [Double] -> Double
computeAUCROC scores labels =
let nPos = fromIntegral (length (filter (== 1) labels))
nNeg = fromIntegral (length (filter (== 0) labels))
in if nPos == 0 || nNeg == 0
then 0.5
else
let sorted = sortBy (comparing (Down . fst)) (zip scores labels)
points = scanl (\(tp, fp) (_, y) -> (tp + y, fp + (1 - y))) (0, 0) sorted
trapz (tp0, fp0) (tp1, fp1) =
let fpr0 = fp0 / nNeg; tpr0 = tp0 / nPos
fpr1 = fp1 / nNeg; tpr1 = tp1 / nPos
in (fpr1 - fpr0) * (tpr0 + tpr1) / 2.0
in sum (zipWith trapz points (tail points))
-- ---------------------------------------------------------------------------
-- Serialization bridge
-- ---------------------------------------------------------------------------
flattenWeights :: MLPWeights -> [Double]
flattenWeights MLPWeights{..} =
toList (flatten w1)
++ toList b1
++ toList (flatten w2)
++ toList b2
++ toList (flatten w3)
++ toList b3
-- Fixed architecture: 29→64→32→2
-- w1: 64*29=1856, b1: 64, w2: 32*64=2048, b2: 32, w3: 2*32=64, b3: 2
unflattenWeights :: [Double] -> MLPWeights
unflattenWeights xs =
let (xw1, r1) = splitAt (64*29) xs
(xb1, r2) = splitAt 64 r1
(xw2, r3) = splitAt (32*64) r2
(xb2, r4) = splitAt 32 r3
(xw3, r5) = splitAt (2*32) r4
xb3 = take 2 r5
in MLPWeights
{ w1 = (64 >< 29) xw1
, b1 = fromList xb1
, w2 = (32 >< 64) xw2
, b2 = fromList xb2
, w3 = ( 2 >< 32) xw3
, b3 = fromList xb3
}
-- ---------------------------------------------------------------------------
-- FedAvg
-- ---------------------------------------------------------------------------
-- Weighted average of flat weight vectors.
fedAvg :: [(Int, [Double])] -> [Double]
fedAvg [] = []
fedAvg updates =
let totalN = fromIntegral (sum (map fst updates)) :: Double
weighted (n, ws) = map (* (fromIntegral n / totalN)) ws
in foldl1 (zipWith (+)) (map weighted updates)
+182
View File
@@ -0,0 +1,182 @@
{-# LANGUAGE RecordWildCards #-}
module FL.P2P
( runPeer
) where
import Control.Actor
import Control.Actor.Registry (registerActor, lookupRemoteActor)
import Control.Concurrent (threadDelay)
import Control.Concurrent.MVar (newEmptyMVar, takeMVar)
import Control.Concurrent.STM
( TMVar, atomically, newEmptyTMVarIO, putTMVar, readTMVar, tryTakeTMVar )
import Control.Monad (forM, forM_, void)
import Control.Monad.IO.Class (liftIO)
import FL.Actors.Logger (logFn, logFn', loggerHandler)
import FL.Actors.PeerCoordinator
import FL.CRDT (gcEmpty, orEmpty)
import FL.Data.PaySim (loadAndPartition)
import FL.NN
( LocalData, evaluate, flattenWeights, initWeights
, localDataSize, trainLocal, unflattenWeights
)
import FL.Retry (retryConnect, retryLookup)
import FL.Types
-- ---------------------------------------------------------------------------
-- P2P-specific actor state types
-- ---------------------------------------------------------------------------
-- Evaluator reads the PeerCoordinator ref lazily from a TMVar.
-- This breaks the circular dependency: evaluator is spawned before PC,
-- but PC's ref is filled into the TMVar right after PC spawns.
data P2PEvalState = P2PEvalState
{ p2eTestData :: LocalData
, p2ePcCell :: TMVar (ActorRef PeerCoordinatorMsg ())
, p2eLogRef :: ActorRef LogMsg ()
}
p2pEvalHandler :: Handler EvaluatorMsg P2PEvalState ()
p2pEvalHandler (EvaluateModel gm) = do
st <- state
let ws = unflattenWeights (wData (gmWeights gm))
metrics = evaluate ws (p2eTestData st) (gmRound gm)
logFn (p2eLogRef st) "Evaluator" INFO $
"P2P round " <> show (gmRound gm)
<> " acc=" <> show4 (mAccuracy metrics)
<> " auc=" <> show4 (mAUCROC metrics)
pcRef <- liftIO $ atomically $ readTMVar (p2ePcCell st)
cast (PeerEvaluationDone metrics) pcRef
pass
where
show4 x = show (fromIntegral (round (x * 10000) :: Int) / 10000.0 :: Double)
-- Trainer likewise reads the PeerCoordinator ref lazily.
data P2PTrainerState = P2PTrainerState
{ p2tId :: String
, p2tData :: LocalData
, p2tLastRound :: Int
, p2tPcCell :: TMVar (ActorRef PeerCoordinatorMsg ())
, p2tLogRef :: ActorRef LogMsg ()
, p2tConfig :: FLConfig
}
p2pTrainerHandler :: Handler TrainerMsg P2PTrainerState ()
p2pTrainerHandler (StartTraining gm) = do
st <- state
let round_ = gmRound gm
if round_ <= p2tLastRound st
then pass
else do
logFn (p2tLogRef st) (p2tId st) INFO $
"P2P training round " <> show round_
let initW = unflattenWeights (wData (gmWeights gm))
updatedW <- liftIO $ trainLocal (p2tConfig st) (p2tId st) round_ initW (p2tData st)
let n = localDataSize (p2tData st)
upd = ModelUpdate (p2tId st) round_ n (Weights (flattenWeights updatedW))
pcRef <- liftIO $ atomically $ readTMVar (p2tPcCell st)
cast (PeerTrainingDone upd) pcRef
passWith st { p2tLastRound = round_ }
p2pTrainerHandler TrainerStop = pass
-- ---------------------------------------------------------------------------
-- Peer node startup
-- ---------------------------------------------------------------------------
runPeer :: FLConfig -> Int -> [NodeAddr] -> IO ()
runPeer cfg peerIdx knownPeers = do
let FLConfig{..} = cfg
peerPort = cfgBasePort
peerId = "peer-" <> show peerIdx
rt <- initRuntime (NodeAddr "localhost" (fromIntegral peerPort)) createTCPTransport
doneVar <- newEmptyMVar
withRuntime rt $ do
forM_ knownPeers $ \addr -> retryConnect 20 addr
liftIO $ threadDelay 300_000
logRef <- spawnActor loggerHandler stopOnDeath ()
registerActor (peerId <> "-logger") logRef
initW <- liftIO initWeights
let initModel = GlobalModel 0 (Weights (flattenWeights initW))
(trainData, testData) <- liftIO $ loadAndPartition cfg peerIdx
-- Shared TMVar: filled with the PeerCoordinator ref after it spawns.
-- Evaluator and Trainer read from it lazily on their first message.
pcCell <- liftIO newEmptyTMVarIO
evalRef <- spawnActor p2pEvalHandler stopOnDeath
(P2PEvalState testData pcCell logRef)
trainerRef <- spawnActor p2pTrainerHandler stopOnDeath
P2PTrainerState
{ p2tId = peerId
, p2tData = trainData
, p2tLastRound = -1
, p2tPcCell = pcCell
, p2tLogRef = logRef
, p2tConfig = cfg
}
-- PeerCoordinator — placeholder self-ref (replaced by SetPeers)
let dummyId = ActorId 0 (read "00000000-0000-0000-0000-000000000002")
dummySelf = RemoteRef dummyId :: ActorRef PeerCoordinatorMsg ()
let pcSt = PeerCoordinatorState
{ pcsNodeId = peerId
, pcsRoundCounter = gcEmpty
, pcsParticipants = orEmpty
, pcsCurrentRound = 0
, pcsTotalRounds = cfgRounds
, pcsGlobalModel = initModel
, pcsPeerRefs = []
, pcsSelfRef = dummySelf
, pcsTrainerRef = trainerRef
, pcsEvalRef = evalRef
, pcsLogRef = logRef
, pcsCollected = []
, pcsNumPeers = cfgNumTrainers
, pcsDoneVar = doneVar
}
-- Supervised PeerCoordinator. On every (re)spawn the shared cell is
-- refreshed and the registry name re-registered, so trainer/evaluator
-- keep reaching the live incarnation. (State restarts fresh.)
supervise' OneForOne
[ ChildSpec
{ csRun = \target -> do
ref <- spawnActor peerCoordinatorHandler peerCoordinatorDeathHandler pcSt
linkActorTo target ref
return ref
, csOnSpawn = \ref -> do
liftIO $ atomically $ do
void $ tryTakeTMVar pcCell
putTMVar pcCell ref
registerActor ("peercoord-" <> peerId) ref
}
]
pcRef <- liftIO $ atomically $ readTMVar pcCell
logFn' logRef peerId INFO $
"Peer ready, waiting for " <> show (cfgNumTrainers - 1) <> " other peers"
-- Discover all other peer coordinators
let otherIds = ["peer-" <> show i | i <- [0 .. cfgNumTrainers - 1], i /= peerIdx]
peerPcRefs <- forM otherIds $ \pid ->
retryLookup 40
(lookupRemoteActor ("peercoord-" <> pid)
:: RuntimeM (Maybe (ActorRef PeerCoordinatorMsg ())))
-- Wire up PeerCoordinator with peers
let peerIds = map actorRefId peerPcRefs
cast' (SetPeers peerIds) pcRef
liftIO $ threadDelay 200_000
cast' PeerStartRound pcRef
logFn' logRef peerId INFO "P2P round 1 started"
takeMVar doneVar
threadDelay 300_000
putStrLn $ "\n=== Peer " <> show peerIdx <> " federation complete ==="
+149
View File
@@ -0,0 +1,149 @@
{-# LANGUAGE RecordWildCards #-}
module FL.Provider
( runCoordinator
, runTrainer
) where
import Control.Actor
import Control.Actor.Registry (lookupRemoteActor, registerActor)
import Control.Concurrent (threadDelay)
import Control.Concurrent.MVar (newEmptyMVar, takeMVar)
import Control.Concurrent.STM
( atomically, newEmptyTMVarIO, putTMVar, readTMVar, tryTakeTMVar )
import Control.Monad (void)
import Control.Monad.IO.Class (liftIO)
import Data.Map.Strict qualified as Map
import FL.Actors.Aggregator
import FL.Actors.Coordinator
import FL.Actors.Evaluator
import FL.Actors.Logger
import FL.Actors.Trainer
import FL.CRDT (orEmpty)
import FL.Data.PaySim (loadAndPartition)
import FL.NN (flattenWeights, initWeights)
import FL.Retry (retryConnect, retryLookup)
import FL.Types
-- ---------------------------------------------------------------------------
-- Coordinator node
-- ---------------------------------------------------------------------------
runCoordinator :: FLConfig -> IO ()
runCoordinator cfg = do
let FLConfig{..} = cfg
rt <- initRuntime (NodeAddr "localhost" (fromIntegral cfgBasePort)) createTCPTransport
doneVar <- newEmptyMVar
withRuntime rt $ do
-- Logger
logRef <- spawnActor loggerHandler stopOnDeath ()
registerActor "logger" logRef
-- Initial global model
initW <- liftIO initWeights
let initModel = GlobalModel 0 (Weights (flattenWeights initW))
-- Shared TMVar for coordinator ref: filled after coordinator spawns.
-- Aggregator and Evaluator read from it lazily.
coordCell <- liftIO newEmptyTMVarIO
-- Load test data (coordinator holds the shared test split)
(_, testData) <- liftIO $ loadAndPartition cfg 0
-- Evaluator
evalRef <- spawnActor evaluatorHandler stopOnDeath
(EvaluatorState testData coordCell logRef)
registerActor "evaluator" evalRef
-- Aggregator
aggRef <- spawnActor aggregatorHandler stopOnDeath
(AggregatorState cfgNumTrainers [] coordCell logRef 0)
registerActor "aggregator" aggRef
-- Coordinator (knows aggRef and evalRef at spawn time)
let initCoordSt = CoordinatorState
{ csTrainers = orEmpty
, csTrainerRefs = Map.empty
, csExpectedTrainers = cfgNumTrainers
, csCurrentRound = 0
, csTotalRounds = cfgRounds
, csGlobalModel = initModel
, csAggRef = aggRef
, csEvalRef = evalRef
, csLogRef = logRef
, csDoneVar = doneVar
}
-- Supervised coordinator. On every (re)spawn — including restarts — the
-- shared cell is refreshed and the registry name re-registered, so
-- aggregator/evaluator/trainers keep reaching the live incarnation.
-- (Coordinator state itself restarts fresh: round progress is lost.)
supervise' OneForOne
[ ChildSpec
{ csRun = \target -> do
ref <- spawnActor coordinatorHandler coordinatorDeathHandler initCoordSt
linkActorTo target ref
return ref
, csOnSpawn = \ref -> do
liftIO $ atomically $ do
void $ tryTakeTMVar coordCell
putTMVar coordCell ref
registerActor "coordinator" ref
}
]
void $ liftIO $ atomically $ readTMVar coordCell
logFn' logRef "System" INFO $
"Coordinator ready, waiting for " <> show cfgNumTrainers <> " trainers"
takeMVar doneVar
threadDelay 300_000
putStrLn "\n=== Provider federation complete ==="
-- ---------------------------------------------------------------------------
-- Trainer node
-- ---------------------------------------------------------------------------
runTrainer :: FLConfig -> Int -> NodeAddr -> IO ()
runTrainer cfg trainerIdx coordAddr = do
let FLConfig{..} = cfg
let trainerPort = cfgBasePort + trainerIdx + 1
trainerId = "trainer-" <> show trainerIdx
rt <- initRuntime (NodeAddr "localhost" (fromIntegral trainerPort)) createTCPTransport
withRuntime rt $ do
retryConnect 20 coordAddr
liftIO $ threadDelay 500_000 -- wait for registry sync
-- Look up aggregator and logger on the coordinator node (retry until
-- available; the logger is registered first, so if the aggregator
-- resolves the logger will too)
aggRef <- retryLookup 20
(lookupRemoteActor "aggregator" :: RuntimeM (Maybe (ActorRef AggregatorMsg ())))
logRef <- retryLookup 20
(lookupRemoteActor "logger" :: RuntimeM (Maybe (ActorRef LogMsg ())))
(trainData, _) <- liftIO $ loadAndPartition cfg trainerIdx
let trainerSt = TrainerState
{ tsId = trainerId
, tsData = trainData
, tsLastRound = -1
, tsAggRef = aggRef
, tsLogRef = logRef
, tsConfig = cfg
}
trainerRef <- spawnActor trainerHandler stopOnDeath trainerSt
registerActor trainerId trainerRef
-- Register with coordinator
coordRef <- retryLookup 20
(lookupRemoteActor "coordinator" :: RuntimeM (Maybe (ActorRef CoordinatorMsg ())))
cast' (RegisterTrainer trainerId (actorRefId trainerRef)) coordRef
logFn' logRef trainerId INFO "Registered with coordinator"
waitForever
waitForever :: IO ()
waitForever = threadDelay maxBound >> waitForever
+34
View File
@@ -0,0 +1,34 @@
-- | Startup helpers shared by the provider and P2P entry points: retry a
-- node connection or a registry lookup until it succeeds or attempts run out.
module FL.Retry
( retryConnect
, retryLookup
) where
import Control.Actor (ActorRef, NodeAddr, RuntimeM, connect, withRuntime)
import Control.Concurrent (threadDelay)
import Control.Exception (SomeException, try)
import Control.Monad (void)
import Control.Monad.IO.Class (liftIO)
import Control.Monad.Reader (ask)
retryConnect :: Int -> NodeAddr -> RuntimeM ()
retryConnect 0 addr = error $ "retryConnect: could not connect to " <> show addr
retryConnect n addr = do
rt <- ask
result <- liftIO $ try @SomeException $ withRuntime rt $ void (connect Nothing addr)
case result of
Right () -> return ()
Left _ -> do
liftIO $ threadDelay 1_000_000
retryConnect (n - 1) addr
retryLookup :: Int -> RuntimeM (Maybe (ActorRef msg r)) -> RuntimeM (ActorRef msg r)
retryLookup 0 _ = error "retryLookup: exhausted retries"
retryLookup n action = do
result <- action
case result of
Just ref -> return ref
Nothing -> do
liftIO $ threadDelay 500_000
retryLookup (n - 1) action
+160
View File
@@ -0,0 +1,160 @@
module FL.Types
( -- Config
FLConfig (..)
, PartitionMode (..)
, NormMode (..)
-- Wire types
, Weights (..)
, GlobalModel (..)
, ModelUpdate (..)
, Metrics (..)
-- Logger messages
, LogLevel (..)
, LogEntry (..)
, LogMsg (..)
-- Actor messages
, EvaluatorMsg (..)
, TrainerMsg (..)
, AggregatorMsg (..)
, CoordinatorMsg (..)
, PeerSyncPayload (..)
, PeerCoordinatorMsg (..)
) where
import Control.Actor.Types (ActorId)
import Data.Binary (Binary)
import FL.CRDT (GCounter, ORSet)
import GHC.Generics (Generic)
-- ---------------------------------------------------------------------------
-- Config (not a wire type — parsed from CLI, passed around locally)
-- ---------------------------------------------------------------------------
data PartitionMode = IID | NonIID
deriving (Show, Eq, Generic)
instance Binary PartitionMode
data NormMode = MinMax | ZScore
deriving (Show, Eq, Generic)
instance Binary NormMode
data FLConfig = FLConfig
{ cfgRounds :: Int
, cfgEpochs :: Int
, cfgLearningRate :: Double
, cfgBatchSize :: Int
, cfgNumTrainers :: Int
, cfgDataPath :: Maybe FilePath
, cfgPartition :: PartitionMode
, cfgNorm :: NormMode
, cfgBasePort :: Int
} deriving (Show, Generic)
-- ---------------------------------------------------------------------------
-- Wire types (sent over the network, must be Binary)
-- ---------------------------------------------------------------------------
newtype Weights = Weights { wData :: [Double] }
deriving (Show, Eq, Generic)
instance Binary Weights
data GlobalModel = GlobalModel
{ gmRound :: Int
, gmWeights :: Weights
} deriving (Show, Eq, Generic)
instance Binary GlobalModel
data ModelUpdate = ModelUpdate
{ muTrainerId :: String
, muRound :: Int
, muSampleSize :: Int
, muWeights :: Weights
} deriving (Show, Eq, Generic)
instance Binary ModelUpdate
data Metrics = Metrics
{ mAccuracy :: Double
, mPrecision :: Double
, mRecall :: Double
, mF1 :: Double
, mAUCROC :: Double
, mRound :: Int
} deriving (Show, Eq, Generic)
instance Binary Metrics
-- ---------------------------------------------------------------------------
-- Logger
-- ---------------------------------------------------------------------------
data LogLevel = DEBUG | INFO | WARN | ERROR
deriving (Show, Eq, Ord, Generic)
instance Binary LogLevel
data LogEntry = LogEntry
{ leSource :: String
, leLevel :: LogLevel
, leMsg :: String
} deriving (Show, Generic)
instance Binary LogEntry
data LogMsg = Log LogEntry
deriving (Show, Generic)
instance Binary LogMsg
-- ---------------------------------------------------------------------------
-- Evaluator
-- ---------------------------------------------------------------------------
data EvaluatorMsg = EvaluateModel GlobalModel
deriving (Show, Generic)
instance Binary EvaluatorMsg
-- ---------------------------------------------------------------------------
-- Trainer
-- ---------------------------------------------------------------------------
data TrainerMsg
= StartTraining GlobalModel
| TrainerStop
deriving (Show, Generic)
instance Binary TrainerMsg
-- ---------------------------------------------------------------------------
-- Aggregator (provider mode)
-- ---------------------------------------------------------------------------
data AggregatorMsg = SubmitUpdate ModelUpdate
deriving (Show, Generic)
instance Binary AggregatorMsg
-- ---------------------------------------------------------------------------
-- Coordinator (provider mode)
-- ---------------------------------------------------------------------------
data CoordinatorMsg
= RegisterTrainer String ActorId
| AggregationComplete GlobalModel
| EvaluationDone Metrics
| StartFederation
deriving (Show, Generic)
instance Binary CoordinatorMsg
-- ---------------------------------------------------------------------------
-- PeerCoordinator (P2P mode)
-- ---------------------------------------------------------------------------
data PeerSyncPayload = PeerSyncPayload
{ pspCounter :: !GCounter
, pspParticipants :: !(ORSet String)
, pspUpdate :: !(Maybe ModelUpdate)
} deriving (Show, Generic)
instance Binary PeerSyncPayload
data PeerCoordinatorMsg
= SetPeers ![ActorId]
| PeerStartRound
| PeerTrainingDone !ModelUpdate
| PeerSync !PeerSyncPayload
| PeerEvaluationDone !Metrics
deriving (Show, Generic)
instance Binary PeerCoordinatorMsg