src/train.ts

import * as tf from '@tensorflow/tfjs';

type LayerValue = string | number | tf.Variable | null | undefined;

interface LayerConfig {
  type: string;
  [key: string]: LayerValue;
}

export interface RunInfo {
  [key: string]: unknown;
}

export interface TrainController {
  isPaused: boolean;
  stopRequested: boolean;
  sampleIndex: number;
}

interface BatchData {
  xs: tf.Tensor;
  labels: tf.Tensor2D;
}

interface TestSample {
  xs: tf.Tensor;
}

interface TrainingData {
  trainSize: number;
  imageSize: number;
  numInputChannels: number;
  nextTrainBatch(batchSize: number): BatchData;
  getTestSample(index: number): TestSample;
}

export interface OptimizerParams {
  learningRate: string;
  batchSize: string;
  epochs: string;

  // sgd only
  momentum?: string; 
  // adam only
  beta1?: string;
  beta2?: string;
  epsilon?: string;
}

type BatchEndCallback = (
  epoch: number,
  batch: number,
  loss: number,
  info: RunInfo[],
) => void | Promise<void>;

function parseValue(raw: string): string | number {
  if (raw.trim() === '') {
    return raw;
  }

  const num = Number(raw);
  if (!Number.isNaN(num)) {
    return num;
  }

  return raw;
}

function parseArchitecture(text: string): LayerConfig[] {
  const layers: LayerConfig[] = [];
  const matches = text.match(/\[(.*?)\]/gs);
  if (!matches) return layers;

  for (const block of matches) {
    const content = block.slice(1, -1).trim();
    if (content.length === 0) continue;

    const tokens = content.split(/\s+/);
    if (tokens.length === 0) continue;

    const type = tokens[0];
    const layer: LayerConfig = { type };

    for (let i = 1; i < tokens.length; ++i) {
      const token = tokens[i];
      const [rawKey, rawValue] = token.split('=', 2);

      if (!rawKey || rawValue === undefined) continue;

      const key = rawKey === 'activation' ? 'activationType' : rawKey;
      layer[key] = parseValue(rawValue);
    }

    layers.push(layer);
  }

  return layers;
}

function getNumber(layer: LayerConfig, key: string): number {
  const value = layer[key];
  if (typeof value !== 'number') {
    throw new Error(`Layer "${layer.type}" is missing numeric "${key}"`);
  }
  return value;
}

function getVariable(layer: LayerConfig, key: string): tf.Variable {
  const value = layer[key];
  if (!value || typeof value !== 'object' || !('dispose' in value)) {
    throw new Error(`Layer "${layer.type}" is missing tensor "${key}"`);
  }
  return value as tf.Variable;
}

function getPadding(layer: LayerConfig): number | 'same' | 'valid' {
  const padding = layer.padding;
  if (padding === undefined) return 'valid';
  if (typeof padding === 'number') return padding;
  if (padding === 'same' || padding === 'valid') return padding;
  throw new Error(`Layer "${layer.type}" has invalid padding "${String(padding)}"`);
}

function getFlatDim(out: tf.Tensor): number {
  const [, h, w, c] = out.shape;
  if (
    typeof h !== 'number' ||
    typeof w !== 'number' ||
    typeof c !== 'number'
  ) {
    throw new Error('Cannot flatten tensor with unknown shape');
  }
  return h * w * c;
}

export class Cnn {
  architecture: LayerConfig[];
  inChannels: number;
  weights: tf.Variable[];

  constructor(architecture: string, inChannels: number) {
    this.architecture = parseArchitecture(architecture);
    this.inChannels = inChannels;
    this.weights = this.initWeights();
  }

  initWeights(): tf.Variable[] {
    const weights: tf.Variable[] = [];
    let inChannels = this.inChannels;

    for (const layer of this.architecture) {
      if (layer.type === 'conv2d') {
        const kernel = getNumber(layer, 'kernel');
        const filters = getNumber(layer, 'filters');
        const shape: [number, number, number, number] = [
          kernel,
          kernel,
          inChannels,
          filters,
        ];
        const layerWeights = tf.variable(
          tf.randomUniform(
            shape,
            -Math.sqrt(1 / (kernel * kernel * inChannels)),
            Math.sqrt(1 / (kernel * kernel * inChannels)),
          ),
        );

        weights.push(layerWeights);
        layer.weights = layerWeights;
        inChannels = filters;
      } else if (layer.type === 'dense') {
        layer.weights = null;
        layer.biases = null;
      }
    }
    return weights;
  }

  dispose(): void {
    for (const layer of this.architecture) {
      if (layer.type === 'conv2d') {
        getVariable(layer, 'weights').dispose();
      } else if (layer.type === 'dense') {
        const weights = layer.weights;
        const biases = layer.biases;
        if (weights && typeof weights === 'object' && 'dispose' in weights) {
          (weights as tf.Variable).dispose();
        }
        if (biases && typeof biases === 'object' && 'dispose' in biases) {
          (biases as tf.Variable).dispose();
        }
      }
    }
  }

  forward(x: tf.Tensor4D): tf.Tensor {
    let out: tf.Tensor = x;

    for (let i = 0; i < this.architecture.length; i += 1) {
      const layer = this.architecture[i];

      switch (layer.type) {
        case 'conv2d': {
          const layerWeights = getVariable(layer, 'weights');
          const stride = getNumber(layer, 'stride');
          const padding = getPadding(layer);
          out = tf.conv2d(
            out as tf.Tensor4D,
            layerWeights as tf.Tensor4D,
            stride,
            padding,
          );
          if (layer.activationType === 'relu') {
            out = out.relu();
          }
          break;
        }
        case 'maxpool': {
          const size = getNumber(layer, 'size');
          const stride = getNumber(layer, 'stride');
          out = tf.maxPool(out as tf.Tensor4D, [size, size], [stride, stride], 0);
          break;
        }
        case 'flatten': {
          const flatDim = getFlatDim(out);
          out = out.reshape([-1, flatDim]);

          const next = this.architecture[i + 1];
          if (next?.type === 'dense' && next.weights === null) {
            const units = getNumber(next, 'units');
            next.weights = tf.variable(
              tf.randomUniform(
                [flatDim, units],
                -Math.sqrt(1 / flatDim),
                Math.sqrt(1 / flatDim),
              ),
            );
            next.biases = tf.variable(tf.zeros([units]));
          }
          break;
        }
        case 'dense': {
          const denseWeights = getVariable(layer, 'weights');
          const denseBiases = getVariable(layer, 'biases');
          out = tf.matMul(out as tf.Tensor2D, denseWeights as tf.Tensor2D).add(
            denseBiases as tf.Tensor1D,
          );

          if (layer.activationType === 'relu') {
            out = out.relu();
          }

          const next = this.architecture[i + 1];
          if (next?.type === 'dense' && next.weights === null) {
            const nextUnits = getNumber(next, 'units');
            const currentUnits = getNumber(layer, 'units');
            next.weights = tf.variable(
              tf.randomUniform(
                [currentUnits, nextUnits],
                -Math.sqrt(1 / currentUnits),
                Math.sqrt(1 / currentUnits),
              ),
            );
            next.biases = tf.variable(tf.zeros([nextUnits]));
          }
          break;
        }
        default:
          break;
      }
    }
    return out;
  }

  forwardWithInfo(x: tf.Tensor4D): { output: tf.Tensor; info: RunInfo[] } {
    let out: tf.Tensor = x;
    const info: RunInfo[] = [];

    info.push({
      type: 'input',
      output: out.dataSync(),
      shape: x.shape,
    });

    for (let i = 0; i < this.architecture.length; i += 1) {
      const layer = this.architecture[i];

      switch (layer.type) {
        case 'conv2d': {
          const layerWeights = getVariable(layer, 'weights');
          const stride = getNumber(layer, 'stride');
          const padding = getPadding(layer);
          out = tf.conv2d(
            out as tf.Tensor4D,
            layerWeights as tf.Tensor4D,
            stride,
            padding,
          );
          if (layer.activationType === 'relu') {
            out = out.relu();
          }

          info.push({
            type: 'conv2d',
            output: out.dataSync(),
            kernels: layerWeights.dataSync(),
            outputShape: out.shape,
            kernelShape: layerWeights.shape,
            stride,
            padding,
            activationType: layer.activationType,
          });
          break;
        }
        case 'maxpool': {
          const size = getNumber(layer, 'size');
          const stride = getNumber(layer, 'stride');
          out = tf.maxPool(out as tf.Tensor4D, [size, size], [stride, stride], 0);

          info.push({
            type: 'maxpool',
            output: out.dataSync(),
            shape: out.shape,
            size,
            stride,
          });
          break;
        }
        case 'flatten': {
          const flatDim = getFlatDim(out);
          out = out.reshape([-1, flatDim]);

          info.push({
            type: 'flatten',
            output: out.dataSync(),
            shape: out.shape,
          });

          const next = this.architecture[i + 1];
          if (next?.type === 'dense' && next.weights === null) {
            const units = getNumber(next, 'units');
            next.weights = tf.variable(
              tf.randomUniform(
                [flatDim, units],
                -Math.sqrt(1 / flatDim),
                Math.sqrt(1 / flatDim),
              ),
            );
            next.biases = tf.variable(tf.zeros([units]));
          }
          break;
        }
        case 'dense': {
          const denseWeights = getVariable(layer, 'weights');
          const denseBiases = getVariable(layer, 'biases');
          out = tf.matMul(out as tf.Tensor2D, denseWeights as tf.Tensor2D).add(
            denseBiases as tf.Tensor1D,
          );
          if (layer.activationType === 'relu') {
            out = out.relu();
          }

          const next = this.architecture[i + 1];
          if (next?.type === 'dense' && next.weights === null) {
            const nextUnits = getNumber(next, 'units');
            const currentUnits = getNumber(layer, 'units');
            next.weights = tf.variable(
              tf.randomUniform(
                [currentUnits, nextUnits],
                -Math.sqrt(1 / currentUnits),
                Math.sqrt(1 / currentUnits),
              ),
            );
            next.biases = tf.variable(tf.zeros([nextUnits]));
          }

          info.push({
            type: 'dense',
            output: out.dataSync(),
            weights: denseWeights.dataSync(),
            biases: denseBiases.dataSync(),
            outputShape: out.shape,
            weightShape: denseWeights.shape,
            biasShape: denseBiases.shape,
            inputUnits: denseWeights.shape[0],
            outputUnits: getNumber(layer, 'units'),
            outputSize: getNumber(layer, 'units'),
            inputSize: denseWeights.shape[0],
            activationType: layer.activationType,
          });
          break;
        }
        default:
          break;
      }
    }

    return { output: out, info };
  }
}

export async function train(
  data: TrainingData,
  model: Cnn,
  optimizer: tf.Optimizer,
  batchSize: number,
  epochs: number,
  controller: TrainController,
  onBatchEnd: BatchEndCallback | null = null,
): Promise<void> {
  const numBatches = Math.floor(data.trainSize / batchSize);
  for (let epoch = 0; epoch < epochs; ++epoch) {
    for (let b = 0; b < numBatches; ++b) {
      if (controller.stopRequested) {
        console.log('Training stopped');
        return;
      }

      while (controller.isPaused) {
        await tf.nextFrame();
      }

      const cost = optimizer.minimize(() => {
        const batch = data.nextTrainBatch(batchSize);
        const xs = batch.xs.reshape([
          batchSize,
          data.imageSize,
          data.imageSize,
          data.numInputChannels,
        ]) as tf.Tensor4D;
        const preds = model.forward(xs);
        return tf.losses.softmaxCrossEntropy(batch.labels, preds).mean();
      }, true);

      if (!cost) {
        throw new Error('Optimizer did not return a loss tensor');
      }

      const lossVal = (await cost.data())[0];
      cost.dispose();

      const sample = data.getTestSample(controller.sampleIndex);
      const { output, info } = model.forwardWithInfo(
        sample.xs.reshape([
          1,
          data.imageSize,
          data.imageSize,
          data.numInputChannels,
        ]) as tf.Tensor4D,
      );

      const probs = tf.tidy(() => tf.softmax(output));
      info.push({
        type: 'output',
        output: probs.dataSync(),
        shape: probs.shape,
      });

      if (controller.stopRequested) {
        console.log('Training stopped');
        probs.dispose();
        return;
      }

      if (onBatchEnd) {
        await onBatchEnd(epoch, b, lossVal, info);
      }

      probs.dispose();
      await tf.nextFrame();
    }
  }
  console.log('Training complete');
}