*: prerelease fixes

.github/workflows/lint-test-build.yml

@@ -335,7 +335,7 @@ jobs:
           cache: npm
       - run: npm ci
       - run: npm --workspace={discojs,discojs-node,server} run build
-      - run: npm --workspace=cli start -- -t cifar10 -u 3 -e 1
+      - run: npm --workspace=cli start -- -t cifar10 -u 3 -e 1 -r 1
 
   test-docs-examples:
     needs: [build-lib, build-lib-node, build-server, download-datasets]

README.md

@@ -39,10 +39,10 @@ ___
 DISCO aims to enable open-access and easy-use distributed training which is
 - :tornado: efficient ([R1](https://github.com/epfml/powergossip), [R2](https://github.com/epfml/ChocoSGD)) 
 - :lock: privacy-preserving ([R3](https://eprint.iacr.org/2017/281.pdf), [R4](https://arxiv.org/abs/2006.04747))
-- :hammer_and_wrench: fault-tolerant and dynamic over time ([R5](https://arxiv.org/abs/1910.12308))
-- :ninja: robust to malicious actors and data poisoning ([R6](https://arxiv.org/abs/2012.10333), [R7](https://arxiv.org/abs/2006.09365))
-- :apple: :banana: interpretable in imperfectly interoperable data distributions ([R8](https://arxiv.org/abs/2107.06580))
-- :mirror: personalizable  ([R9](https://arxiv.org/abs/2103.00710))
+- :hammer_and_wrench: fault-tolerant and dynamic over time ([R5](https://arxiv.org/abs/2106.06639), [R6](https://arxiv.org/abs/2206.08307))
+- :ninja: robust to malicious actors and data poisoning ([R7](https://arxiv.org/abs/2012.10333), [R8](https://arxiv.org/abs/2006.09365))
+- :apple: :banana: interpretable in imperfectly interoperable data distributions ([R9](https://arxiv.org/abs/2107.06580))
+- :mirror: personalizable  ([R10](https://arxiv.org/abs/2103.00710))
 - :carrot: fairly incentivize participation
 
 

cli/src/benchmark_gpt.ts

@@ -1,6 +1,6 @@
 import { parse } from 'ts-command-line-args';
 import type { Task } from '@epfml/discojs'
-import { fetchTasks, data, models } from '@epfml/discojs'
+import { fetchTasks, data, models, async_iterator } from '@epfml/discojs'
 import { NodeTextLoader, loadModelFromDisk } from '@epfml/discojs-node'
 import { startServer } from 'server'
 
@@ -49,15 +49,15 @@ async function main(args: Required<CLIArguments>): Promise<void> {
 
   // Fetch the wikitext task from the server
   const tasks = await fetchTasks(url)
-  const task = tasks.get('wikitext-103')
+  const task = tasks.get('llm_task')
   if (task === undefined) { throw new Error('task not found') }  
 
   /**
    * Training benchmark
    */
   if (!benchmarkInference) {
     // Benchmark parameters
-    const epoch = 1
+    const epochsCount = 1
     const iterationsPerEpoch = 10
 
     const config: models.GPTConfig = {
@@ -80,10 +80,10 @@ async function main(args: Required<CLIArguments>): Promise<void> {
     console.log(`\tmodel type ${modelType} \n\tbatch size ${batchSize} \n\tcontext length ${contextLength}`)
 
     let epochTime = performance.now()
-    const logGenerator = model.train(preprocessedDataset, undefined, epoch)
-    for await (const logs of logGenerator) {
+    for (let epochsCounter = 1; epochsCounter <= epochsCount; epochsCounter++) {
+      const [_, logs] = await async_iterator.gather(model.train(preprocessedDataset))
       epochTime = (performance.now() - epochTime)
-      const msPerToken = epochTime / (batchSize * contextLength * iterationsPerEpoch * epoch)
+      const msPerToken = epochTime / (batchSize * contextLength * iterationsPerEpoch * epochsCounter)
       console.log(`\t\tTraining time: ${msPerToken.toFixed(2)} ms/token <br> ${logs.peakMemory.toFixed(2)} GB`)
     }
 

cli/src/cli.ts

@@ -8,6 +8,13 @@ import { startServer } from 'server'
 import { getTaskData } from './data.js'
 import { args } from './args.js'
 
+// Array.fromAsync not yet widely used (2024)
+async function arrayFromAsync<T>(iter: AsyncIterable<T>): Promise<T[]> {
+  const ret: T[] = [];
+  for await (const e of iter) ret.push(e);
+  return ret;
+}
+
 async function runUser(
   task: Task,
   url: URL,
@@ -22,9 +29,7 @@ async function runUser(
   // force the federated scheme
   const disco = new Disco(task, { scheme: "federated", client });
 
-  let logs = List<RoundLogs>();
-  for await (const round of disco.fit(data)) logs = logs.push(round);
-
+  const logs = List(await arrayFromAsync(disco.trainByRound(data)));
   await disco.close();
   return logs;
 }

discojs/src/default_tasks/wikitext.ts

@@ -22,7 +22,7 @@ export const wikitext: TaskProvider = {
         modelID: 'llm-raw-model',
         preprocessingFunctions: [data.TextPreprocessing.Tokenize, data.TextPreprocessing.LeftPadding],
         scheme: 'federated',
-        epochs: 5,
+        epochs: 6,
         // Unused by wikitext because data already comes split
         // But if set to 0 then the webapp doesn't display the validation metrics
         validationSplit: 0.1, 

discojs/src/index.ts

@@ -13,10 +13,11 @@ export { Memory, type ModelInfo, type Path, type ModelSource, Empty as EmptyMemo
 export { Disco, RoundLogs } from './training/index.js'
 export { Validator } from './validation/index.js'
 
-export { Model, EpochLogs } from './models/index.js'
+export { Model, BatchLogs, EpochLogs, ValidationMetrics } from './models/index.js'
 export * as models from './models/index.js'
 
 export * from './task/index.js'
 export * as defaultTasks from './default_tasks/index.js'
 
 export * from './types.js'
+export * as async_iterator from "./utils/async_iterator.js"

discojs/src/models/gpt/evaluate.ts

@@ -9,7 +9,7 @@ export default async function evaluate (
   model: tf.LayersModel,
   dataset: tf.data.Dataset<DataPoint>,
   maxEvalBatches: number
-): Promise<Record<'acc' | 'val_acc' | 'val_loss' | 'val_perplexity', number>> {
+): Promise<Record<'val_acc' | 'val_loss' | 'val_perplexity', number>> {
   let datasetSize = 0
   let totalLoss = 0
   const acc: [number, number] = [0, 0]
@@ -53,7 +53,6 @@ export default async function evaluate (
   return {
     val_loss: loss,
     val_perplexity: Math.exp(loss),
-    acc: acc[0] / acc[1],
     val_acc: acc[0] / acc[1]
   }
 }

discojs/src/models/gpt/gpt.spec.ts

@@ -36,8 +36,8 @@ describe('gpt-tfjs', function() {
     }).repeat().batch(64)
 
     const model = new GPT(config)
-    const logGenerator = model.train(tokenDataset, undefined, 5) // 5 epochs
-    for await (const _ of logGenerator); // Await the end of training
+    for (let i = 0; i < 5; i++)
+      for await (const _ of model.train(tokenDataset, undefined));
     const generation = await model.generate("Lorem ipsum dolor", tokenizer, 1)
     expect(generation).equal(data) // Assert that the model completes 'Lorem ipsum dolor' with 'sit' 
   })

discojs/src/models/gpt/index.ts

@@ -8,10 +8,13 @@ import { PreTrainedTokenizer } from '@xenova/transformers';
 import { WeightsContainer } from '../../index.js'
 import type { Dataset } from '../../dataset/index.js'
 
-import { Model } from '../model.js'
+import { BatchLogs, Model, EpochLogs } from "../index.js";
+import type { Prediction, Sample } from '../model.js'
+
 import { GPTForCausalLM } from './model.js'
-import type { EpochLogs, Prediction, Sample } from '../model.js'
-import type { GPTConfig } from './config.js'
+import { DEFAULT_CONFIG, type GPTConfig } from './config.js'
+import evaluate from './evaluate.js';
+import { List } from 'immutable';
 
 export type GPTSerialization = {
   weights: WeightsContainer
@@ -21,9 +24,13 @@ export type GPTSerialization = {
 export class GPT extends Model {
   private readonly model: GPTForCausalLM
 
+  readonly #maxBatchCount: number
+
   constructor (partialConfig?: GPTConfig, layersModel?: tf.LayersModel) {
     super()
+
     this.model = new GPTForCausalLM(partialConfig, layersModel)
+    this.#maxBatchCount = partialConfig?.maxIter ?? DEFAULT_CONFIG.maxIter
   }
 
   /**
@@ -38,51 +45,90 @@ export class GPT extends Model {
   override async *train(
     trainingData: Dataset,
     validationData?: Dataset,
-    epochs = 1,
-  ): AsyncGenerator<EpochLogs, void> {
-    this.model.compile()
+  ): AsyncGenerator<BatchLogs, EpochLogs> {
+    this.model.compile();
+
+    const batches = await trainingData.iterator(); // tf.LazyIterator isn't an AsyncGenerator
+    let batchesLogs = List<BatchLogs>();
+    for (
+      let batchNumber = 0;
+      batchNumber < this.#maxBatchCount;
+      batchNumber++
+    ) {
+      const iteration = await batches.next();
+      if (iteration.done) break;
+      const batch = iteration.value;
+
+      const batchLogs = await this.#runBatch(batch);
+      tf.dispose(batch);
+
+      yield batchLogs;
+      batchesLogs = batchesLogs.push(batchLogs);
+    }
+
+    const validation = validationData && (await this.#evaluate(validationData));
+    return new EpochLogs(batchesLogs, validation);
+  }
+
+  async #runBatch(
+    batch: tf.TensorContainer,
+  ): Promise<BatchLogs> {
     let logs: tf.Logs | undefined;
-    const trainingArgs: tf.ModelFitDatasetArgs<tf.TensorContainer> = {
-      epochs: 1, // force fitDataset to do only one epoch because it is wrapped in a for loop
-      validationData,
-      callbacks: { onEpochEnd: (_, cur) => { logs = cur }},
+    await this.model.fitDataset(tf.data.array([batch]), {
+      epochs: 1,
+      verbose: 0, // don't pollute
+      callbacks: {
+        onEpochEnd: (_, cur) => {
+          logs = cur;
+        },
+      },
+    });
+    if (logs === undefined) throw new Error("batch didn't gave any logs");
+
+    const { loss, acc: accuracy } = logs;
+    if (loss === undefined || isNaN(loss))
+      throw new Error("training loss is undefined or NaN");
+
+    return {
+      accuracy,
+      loss,
+      memoryUsage: tf.memory().numBytes / 1024 / 1024 / 1024,
     };
-    for (let epoch = 0; epoch < epochs; epoch++) {
-      await this.model.fitDataset(trainingData, trainingArgs);
-      if (logs === undefined) {
-        throw new Error("Epoch didn't gave any logs");
-      }
-      const { loss, val_acc, val_loss, peakMemory } = logs;
-      if (loss === undefined || isNaN(loss)) {
-        throw new Error("Training loss is undefined or nan");
-      }
-      const structuredLogs: EpochLogs = {
-        epoch,
-        peakMemory,
-        training: {
-          loss: logs.loss,
-          accuracy: logs.acc
-        }
-      }
+  }
 
-      if (validationData !== undefined) {
-        if(val_loss === undefined || isNaN(val_loss) ||
-          val_acc === undefined || isNaN(val_acc)) {
-          throw new Error("Validation accuracy or loss is undefined or nan");
+  async #evaluate(
+    dataset: Dataset,
+  ): Promise<Record<"accuracy" | "loss", number>> {
+    const evaluation = await evaluate(
+      this.model,
+      dataset.map((t) => {
+        switch (t) {
+          case null:
+          case undefined:
+            throw new Error("nullish value in dataset");
+          default:
+            // TODO unsafe cast
+            return t as { xs: tf.Tensor2D; ys: tf.Tensor3D };
         }
-        structuredLogs.validation = { accuracy: logs.val_acc, loss: logs.val_loss}
-      }
-      yield structuredLogs
-    }
+      }),
+      this.config.maxEvalBatches,
+    );
+
+    return {
+      accuracy: evaluation.val_acc,
+      loss: evaluation.val_loss,
+    };
   }
 
-  override predict (input: Sample): Promise<Prediction> {
-    const ret = this.model.predict(input)
+  override predict(input: Sample): Promise<Prediction> {
+    const ret = this.model.predict(input);
     if (Array.isArray(ret)) {
-      throw new Error('prediction yield many Tensors but should have only returned one')
+      throw new Error(
+        "prediction yield many Tensors but should have only returned one",
+      );
     }
 
-    return Promise.resolve(ret)
+    return Promise.resolve(ret);
   }
 
   async generate(input: string, tokenizer: PreTrainedTokenizer, newTokens: number = 10): Promise<string> {

discojs/src/models/gpt/model.ts

@@ -57,30 +57,48 @@ class GPTModel extends tf.LayersModel {
     await callbacks.onTrainBegin?.()
 
     for (let epoch = 1; epoch <= trainingArgs.epochs; epoch++) {
+      let accuracyFraction: [number, number] = [0, 0];
       let averageLoss = 0
-      let peakMemory = 0
       let iteration = 1
       const iterator = await dataset.iterator()
-      let preprocessingTime = performance.now()
       let next = await iterator.next()
-      preprocessingTime = performance.now() - preprocessingTime
 
       while (next.done !== true && iteration <= this.config.maxIter) {
         let weightUpdateTime = performance.now()
         await callbacks.onEpochBegin?.(epoch)
         const { xs, ys } = next.value as { xs: tf.Tensor2D, ys: tf.Tensor3D }
-        const lossFn: () => tf.Scalar = () => {
+
+        let preprocessingTime = performance.now()
+        await Promise.all([xs.data(), ys.data()])
+        preprocessingTime = performance.now() - preprocessingTime
+
+        // TODO include as a tensor inside the model
+        const accTensor = tf.tidy(() => {
           const logits = this.apply(xs)
-          if (Array.isArray(logits)) {
+          if (Array.isArray(logits))
             throw new Error('model outputs too many tensor')
-          }
-          if (logits instanceof tf.SymbolicTensor) {
+          if (logits instanceof tf.SymbolicTensor)
             throw new Error('model outputs symbolic tensor')
-          }
-          return tf.losses.softmaxCrossEntropy(ys, logits)
-        }
+          return tf.metrics.categoricalAccuracy(ys, logits)
+        })
+        const accSize = accTensor.shape.reduce((l, r) => l * r, 1)
+        const accSumTensor = accTensor.sum()
+        const accSum = await accSumTensor.array()
+        tf.dispose(accSumTensor)
+        if (typeof accSum !== 'number')
+          throw new Error('got multiple accuracy sum')
+        accuracyFraction = [accuracyFraction[0] + accSum, accuracyFraction[1] + accSize];
+	tf.dispose([accTensor])
+
         const lossTensor = tf.tidy(() => {
-          const { grads, value: lossTensor } = this.optimizer.computeGradients(lossFn)
+          const { grads, value: lossTensor } = this.optimizer.computeGradients(() => {
+            const logits = this.apply(xs)
+            if (Array.isArray(logits))
+              throw new Error('model outputs too many tensor')
+            if (logits instanceof tf.SymbolicTensor)
+              throw new Error('model outputs symbolic tensor')
+            return tf.losses.softmaxCrossEntropy(ys, logits)
+          })
           const gradsClipped = clipByGlobalNormObj(grads, 1)
           this.optimizer.applyGradients(gradsClipped)
           return lossTensor
@@ -89,6 +107,7 @@ class GPTModel extends tf.LayersModel {
         const loss = await lossTensor.array()
         averageLoss += loss
         weightUpdateTime = performance.now() - weightUpdateTime
+
         tf.dispose([xs, ys, lossTensor])
 
         if (
@@ -100,9 +119,6 @@ class GPTModel extends tf.LayersModel {
           console.log(iterationLogs)
         }
         const memory = tf.memory().numBytes / 1024 / 1024 / 1024
-        if (memory > peakMemory) {
-          peakMemory = memory
-        }
         console.log(
           `Epoch: ${epoch}`,
           `\tStep: ${iteration} / ${this.config.maxIter}`,
@@ -122,7 +138,7 @@ class GPTModel extends tf.LayersModel {
       }
       let logs: tf.Logs = {
         'loss': averageLoss / iteration,
-        'peakMemory': peakMemory
+        'acc': accuracyFraction[0] / accuracyFraction[1],
       }
       if (evalDataset !== undefined) {
         logs = { ...logs, ...await evaluate(this, evalDataset, this.config.maxEvalBatches) }

discojs/src/models/index.ts

@@ -1,4 +1,5 @@
-export { EpochLogs, Model } from './model.js'
+export { Model } from './model.js'
+export { BatchLogs, EpochLogs, ValidationMetrics } from "./logs.js";
 
 export { GPT } from './gpt/index.js'
 export { GPTConfig } from './gpt/config.js'