Merge pull request #658 from epfml/656-improve-gpt-julien

Fix gpt-tfjs bugs, add tests and refactor code

discojs/discojs-core/package.json

@@ -31,9 +31,9 @@
     "ws": "8"
   },
   "devDependencies": {
+    "@tensorflow/tfjs-node": "4",
     "@types/chai": "4",
     "@types/mocha": "10",
-    "@types/msgpack-lite": "0.1",
     "@types/simple-peer": "9",
     "chai": "5",
     "mocha": "10",

discojs/discojs-core/src/dataset/data/preprocessing/text_preprocessing.spec.ts

@@ -0,0 +1,97 @@
+import { TEXT_PREPROCESSING } from './index.js'
+import { expect } from 'chai'    
+
+import type { Task } from '../../../index.js'
+import * as tf from '@tensorflow/tfjs'
+
+describe('text preprocessing', function () {
+  const [tokenize, leftPadding] = TEXT_PREPROCESSING
+  // Use a function to create different task object for each test (otherwise the tokenizer gets cached)
+  function initMockTask(): Task {
+    return {
+      id: 'mock-task-id',
+      displayInformation: {},
+      trainingInformation: {
+        modelID: 'model-id',
+        epochs: 1,
+        roundDuration: 1,
+        validationSplit: 0,
+        batchSize: 8,
+        scheme: 'local',
+        dataType: 'text',
+        tokenizer: 'Xenova/gpt2',
+      }}
+    }
+
+  const text = "Hello world, a bc 1 2345, '? 976. Wikipedia is a free content online encyclopedia written and maintained by a community \n of volunteers, known as Wikipedians. Founded by Jimmy Wales and Larry Sanger on January 15, 2001, Wikipedia is hosted by the Wikimedia Foundation, an American nonprofit organization that employs a staff of over 700 people.[7]"
+  const expectedTokens = [15496, 995, 11, 257, 47125, 352, 2242, 2231, 11, 705, 30, 860, 4304, 13, 15312, 318, 257, 1479, 2695, 2691, 45352, 3194, 290, 9456, 416, 257, 2055, 220, 198, 286, 11661, 11, 1900, 355, 11145, 46647, 1547, 13, 4062, 276, 416, 12963, 11769, 290, 13633, 311, 2564, 319, 3269, 1315, 11, 5878, 11, 15312, 318, 12007, 416, 262, 44877, 5693, 11, 281, 1605, 15346, 4009, 326, 24803, 257, 3085, 286, 625, 13037, 661, 3693, 22, 60]
+
+  it('can tokenize text', async () => {
+    const { tokens } = await tokenize.apply(Promise.resolve(text), initMockTask()) as { tokens: number[]}
+    expect(tokens).to.be.deep.equal(expectedTokens)
+  })
+
+  it('can truncate inputs when tokenizing', async () => {
+    const truncationTask = initMockTask()
+    truncationTask.trainingInformation.maxSequenceLength = 10
+    const { tokens } = await tokenize.apply(Promise.resolve(text), truncationTask) as { tokens: number[] }
+    const expectedLength = truncationTask.trainingInformation.maxSequenceLength + 1 // + 1 because tokenization includes an extra token label for next label prediction
+    expect(tokens.length).to.be.equal(expectedLength)
+    expect(tokens).to.be.deep.equal(expectedTokens.slice(0, expectedLength))
+  })
+
+  it('can left pad tokens', async () => {
+    // Create a task where output token sequence should all have length 20
+    const paddingTask = initMockTask()
+    paddingTask.trainingInformation.maxSequenceLength = 20
+
+    // Create a token sequence of length 10
+    const tokens = { tokens: [0,1,2,3,4,5,6,7,8,9] }
+    const { xs, ys } = await leftPadding.apply(Promise.resolve(tokens), paddingTask) as { xs: tf.Tensor1D, ys: tf.Tensor2D }
+    const xsArray = await xs.array()
+    const ysArray = await ys.array()
+
+    // Output sequences should have shape (20) and (20, 50258), 50258 being the size of the vocab for gpt2
+    expect(xsArray.length).to.be.equal(paddingTask.trainingInformation.maxSequenceLength)
+    expect(ysArray.length).to.be.equal(paddingTask.trainingInformation.maxSequenceLength)
+    expect(ysArray[0].length).to.be.equal(50258)
+
+    // xs should be left pad with gpt2's padding token 50256 to be of length 20.
+    // We expect the last token of input token sequence (9) to not be included in xs since it doesn't have a next token to be predicted
+    const paddingToken = 50256
+    const expectedXs = Array.from({length:11}).map(_ => paddingToken).concat(tokens.tokens.slice(0,9))
+    expect(xsArray).to.be.deep.equal(expectedXs)
+
+    // ys should be a one hot encoding of the next token in xs
+    // if the input tokens are [0,1,2,3] then the labels are [1,2,3] which are then one-hot encoded
+    // So the sum of each row should be equal to 1
+    const expectedOneHot = Array.from({ length: 20 }).map(_ => 1)
+    expect(await ys.sum(-1).array()).to.be.deep.equal(expectedOneHot)
+
+    // In each row, the index of the 1 should be the token id
+    const expectedYs = Array.from({length:10}).map(_ => paddingToken).concat(tokens.tokens)
+    expect(await ys.argMax(-1).array()).to.be.deep.equal(expectedYs)
+  })
+
+  it('throws an error if no tokenizer is specified', async () => {
+    const invalidTask = initMockTask()
+    invalidTask.trainingInformation.tokenizer = undefined;
+    try {
+      await tokenize.apply(Promise.resolve("input text doesn't matter"), invalidTask)
+    } catch {
+      return
+    }
+    throw new Error("undefined tokenizer should have thrown an error")
+  })
+  it('throws an error if the tokenizer name is invalid', async () => {
+    const invalidTask = initMockTask()
+    invalidTask['trainingInformation']['tokenizer'] = 'invalid-tokenizer-name'
+    try {
+      await tokenize.apply(Promise.resolve("input text doesn't matter"), invalidTask)
+    } catch {
+      return
+    }
+    throw new Error("invalid tokenizer name should have thrown an error")
+  })
+
+})

discojs/discojs-core/src/dataset/data/preprocessing/text_preprocessing.ts

@@ -14,43 +14,58 @@ export enum TextPreprocessing {
 }
 
 interface TokenizedEntry extends tf.TensorContainerObject {
-  xs: tf.Tensor1D
+  tokens: number []
 }
 
 /**
+ * LeftPadding pads all incoming inputs to be a fixed length, which should be specified
+ * in `task.trainingInformation.maxSequenceLength`. 
+ * 
  * We are currently only implementing left padding for text generation 
  * https://huggingface.co/docs/transformers/en/llm_tutorial#wrong-padding-side
- * The function can easily be extended to support right padding once the need arise
+ * The function can easily be extended to support right padding if needed
+ * 
+ * Once Transformers.js supports left padding, it will be possible to pad inputs
+ * directly when tokenizing
+ * https://github.com/xenova/transformers.js/blob/8804c36591d11d8456788d1bb4b16489121b3be2/src/tokenizers.js#L2517
  */
 const leftPadding: PreprocessingFunction = {
   type: TextPreprocessing.LeftPadding,
   apply: async (x: Promise<tf.TensorContainer>, task: Task): Promise<tf.TensorContainer> => {
-    let { xs } = await x as TokenizedEntry
-    if (xs === undefined ||  !(xs instanceof tf.tensor) ||xs.rankType !== tf.Rank.R1) {
-      new Error("The leftPadding preprocessing expects a 1D tensor named 'xs' as input")
+    if (x === undefined || !Array.isArray(x) || x.length == 0 || typeof(x[0] != 'number')) {
+      new Error("The leftPadding preprocessing expects a non empty 1D array of number")
     }
+    const { tokens } = await x as TokenizedEntry
     const tokenizer = await models.getTaskTokenizer(task)
-
-
-    const maxLength = task.trainingInformation.maxSequenceLength ?? tokenizer.model_max_length as number
-    // Should never happen because tokenization truncates inputs
-    if (xs.size > maxLength) {
-      xs = xs.slice([0], [maxLength])
-    } else if (xs.size < maxLength) {
-      const paddingToken = tokenizer.pad_token_id
-      xs = xs.pad([[Math.max(0, maxLength - xs.size), 0]], paddingToken)
-    }
-    // if xs.size == maxLength we can leave it as it is
-    return {
-      xs,
-      ys: tf.oneHot(xs, tokenizer.model.vocab.length + 1) // gpt-tfjs expects a one-hot encoded token label
-    }
+    return tf.tidy(() => {
+      // maxLength is the final length of xs
+      // Because ys the contains the tokens in xs shifted by one (to predict the next token), we need
+      // to include one more token than maxSequenceLength in order to have the next token's label of the maxSequenceLength'th token
+      const maxLength = task.trainingInformation.maxSequenceLength ?? tokenizer.model_max_length as number
+      const maxLengthPlusLabel = maxLength + 1
+
+      let fixedLengthTokens = tf.tensor(tokens, undefined, 'int32') // cast tokens from float to int for gpt-tfjs
+      if (fixedLengthTokens.size > maxLengthPlusLabel) { // Should never happen because tokenization truncates inputs
+        throw Error("There are more tokens than expected after tokenization and truncation")
+      } else if (fixedLengthTokens.size < maxLengthPlusLabel) { // Pad inputs to fixed length
+        const paddingToken = tokenizer.pad_token_id
+        fixedLengthTokens = fixedLengthTokens.pad([[Math.max(0, maxLengthPlusLabel - fixedLengthTokens.size), 0]], paddingToken)
+      }
+      // if tokens.size == maxLengthPlusLabel we can leave it as it is
+
+      // ys is a one-hot encoding of the next token (i.e. xs shifted by one)
+      const ys = tf.oneHot(fixedLengthTokens.slice([1]), tokenizer.model.vocab.length + 1)
+      // remove the extra token now that ys is created
+      const xs = fixedLengthTokens.slice([0], maxLength) 
+      return { xs, ys }
+    })
   }
 }
 
 interface TokenizerOutput {
   input_ids: number[]
 }
+
 /**
  * Tokenize and truncates input strings
  */
@@ -62,7 +77,10 @@ const tokenize: PreprocessingFunction = {
     }
     const xs = await x as string // tf.TextLineDataset yields strings
     const tokenizer = await models.getTaskTokenizer(task)
-    const maxLength = task.trainingInformation.maxSequenceLength ?? tokenizer.model_max_length as number
+    // Add plus one to include the next token label of the last token in the input sequence
+    // The inputs are truncated down to exactly maxSequenceLength in leftPadding
+    const maxLength = task.trainingInformation.maxSequenceLength ?? (tokenizer.model_max_length as number)
+    const maxLengthPlusLabel = maxLength + 1
 
     const {input_ids: tokens} = tokenizer(xs, {
       // Transformers.js currently only supports right padding while we need left for text generation
@@ -71,11 +89,9 @@ const tokenize: PreprocessingFunction = {
       padding: false,
       truncation: true,
       return_tensor: false,
-      max_length: maxLength,
+      max_length: maxLengthPlusLabel,
     }) as TokenizerOutput
-    return {
-      xs: tf.tensor(tokens, undefined, 'int32') // cast tokens from float to int for gpt-tfjs}
-    }
+    return { tokens }
   }
 }
 

discojs/discojs-core/src/models/gpt/config.ts

@@ -7,16 +7,12 @@ type ModelType =
     | 'gpt-micro'
     | 'gpt-nano'
 
-export interface ModelSize {
-  nLayer?: number
-  nHead?: number
-  nEmbd?: number
-}
-
 export interface GPTConfig {
   lr: number
   blockSize: number
   vocabSize: number
+  modelType: ModelType
+  name?: string,
   evaluate?: boolean
   maxEvalBatches?: number
   evaluateEvery?: number
@@ -30,13 +26,16 @@ export interface GPTConfig {
   embdDrop?: number
   tokEmb?: boolean
   lmHead?: boolean
-  modelType: ModelType
+  nLayer?: number
+  nHead?: number
+  nEmbd?: number
 }
 
 export const DEFAULT_CONFIG: Required<GPTConfig> = {
+  name: 'transformer',
   lr: 0.001,
   weightDecay: 0,
-  maxIter: 10_000,
+  maxIter: 5,
   verbose: 0,
   modelType: 'gpt-nano',
   evaluate: true,
@@ -50,7 +49,16 @@ export const DEFAULT_CONFIG: Required<GPTConfig> = {
   residDrop: 0.2,
   embdDrop: 0.2,
   tokEmb: true,
-  lmHead: true
+  lmHead: true,
+  nLayer: 3,
+  nHead: 3,
+  nEmbd: 48,
+}
+
+export type ModelSize = {
+  nLayer: number
+  nHead: number
+  nEmbd: number
 }
 
 export function getModelSizes (modelType: ModelType): Required<ModelSize> {

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

@@ -0,0 +1,44 @@
+import { expect } from 'chai'
+import * as tf from '@tensorflow/tfjs-node'
+import { AutoTokenizer } from '@xenova/transformers';
+import { GPT } from './index.js'
+import { type GPTConfig } from './config.js'
+
+describe('gpt-tfjs', function() {
+  this.timeout(50_000)
+  const data = "Lorem ipsum dolor sit"
+
+  const config: GPTConfig = {
+    modelType: 'gpt-nano',
+    lr: 0.01,
+    maxIter: 10,
+    evaluateEvery:10,
+    maxEvalBatches: 10,
+    blockSize: 8,
+    vocabSize: 50258
+  }
+
+  it('can overfit one sentence', async () => {
+    const tokenizer = await AutoTokenizer.from_pretrained('Xenova/gpt2')
+    const datasetSource = new tf.data.FileDataSource(Buffer.from(data))
+    const textDataset = new tf.data.TextLineDataset(datasetSource)
+    const tokenDataset = textDataset.map((text: string) => {
+      const { input_ids: tokens } = tokenizer(text, {
+          padding: true,
+          truncation: true,
+          return_tensor: false,
+          max_length: config.blockSize + 1,
+      }) as { input_ids: number[] }
+      const ys = tf.oneHot(tokens.slice(1), tokenizer.model.vocab.length + 1)
+      const xs = tf.tensor(tokens.slice(0, config.blockSize), undefined, 'int32')
+      return {xs, ys}
+    }).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
+    const generation = await model.generate("Lorem ipsum dolor", tokenizer, 1)
+    console.log(generation)
+    expect(generation).equal(data) // Assert that the model completes 'Lorem ipsum dolor' with 'sit' 
+  })
+})