engine.ts

/**
 * @license
 * Copyright 2018 Google Inc. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */

import {ENV} from './environment';
import {tidy} from './globals';
import {BackendTimingInfo, KernelBackend} from './kernels/backend';
import * as ops from './ops/ops';
import {Profiler} from './profiler';
import {backpropagateGradients, getFilteredNodesXToY} from './tape';
import {NamedGradientMap, TapeNode} from './tape';
import {DataId, Tensor, Tensor3D, Variable} from './tensor';
// tslint:disable-next-line:max-line-length
import {NamedTensorMap, NamedVariableMap, TensorContainer, TypedArray} from './types';
import * as util from './util';

interface ScopeState {
  track: Tensor[];
  name?: string;
}

/**
 * A function that computes an output. The save function is for saving tensors
 * computed in the forward pass, that we need in the backwards pass.
 */
export type ForwardFunc<T extends Tensor> =
    (backend: KernelBackend, save?: <S extends Tensor>(tensor: S) => S) => T;

/**
 * @docalias (a: Tensor, b: Tensor,...) => {
 *   value: Tensor,
 *   gradFunc: (dy: Tensor) => Tensor|Tensor[]
 * }
 */
export type CustomGradientFunc<T extends Tensor> = (...args: Tensor[]) => {
  value: T, gradFunc: (dy: T) => Tensor | Tensor[];
};

export interface TensorManager {
  registerTensor(a: Tensor): void;
  registerVariable(v: Variable): void;
  disposeTensor(a: Tensor): void;
  memory(): {numDataBuffers: number; numBytes: number;};
}

export type MemoryInfo = {
  numTensors: number; numDataBuffers: number; numBytes: number;
  unreliable?: boolean;
};

export interface TimingInfo extends BackendTimingInfo {
  wallMs: number;
}

export class Engine implements TensorManager {
  // Public since optimizers will use it.
  registeredVariables: NamedVariableMap = {};

  private refCounter = new WeakMap<DataId, number>();
  private nextTapeNodeId = 0;
  private numBytes = 0;
  private numTensors = 0;
  private numDataBuffers = 0;

  private activeTape: TapeNode[];
  private gradientScopeCount = 0;
  private customGradientDepth = 0;

  // Keep Tensors that parallel the tapes.
  private activeScope: ScopeState;
  private scopeStack: ScopeState[];
  private keepTensors: Set<number> = new Set();
  private profiler: Profiler;

  constructor(private backend: KernelBackend, public safeMode: boolean) {
    // Create a default outer scope.
    this.activeScope = {track: []};
    this.scopeStack = [this.activeScope];
    this.profiler = new Profiler(backend);
  }

  runKernel<T extends Tensor, I extends NamedTensorMap>(
      forwardFunc: ForwardFunc<T>,
      inputs: I,
      backwardsFunc?: (dy: T, saved: Tensor[]) => {[P in keyof I]: () => I[P]},
      ): T {
    let result: T;
    const saved: Tensor[] = [];
    const saveFunc = <T extends Tensor>(x: T): T => {
      saved.push(x);
      return x;
    };
    const scopeName = this.activeScope.name;

    // Stop recording to a tape when running a kernel.
    this.customGradientDepth++;
    if (!ENV.get('DEBUG')) {
      result = forwardFunc(this.backend, saveFunc);
    } else {
      result = this.profiler.profileKernel(
          scopeName, () => forwardFunc(this.backend, saveFunc));
    }
    // Continue recording after the kernel is done.
    this.customGradientDepth--;

    if (this.shouldRecord()) {
      const tapeNode: TapeNode = {
        id: this.nextTapeNodeId++,
        name: scopeName,
        inputs,
        output: result,

      };
      if (backwardsFunc != null) {
        tapeNode.gradient = (dy: T) => backwardsFunc(dy, saved);
      }
      this.activeTape.push(tapeNode);
    }
    return result;
  }

  // TensorManager implementation.

  registerTensor(a: Tensor|Variable): void {
    const refCount =
        this.refCounter.has(a.dataId) ? this.refCounter.get(a.dataId) : 0;
    this.numTensors++;
    if (refCount === 0) {
      this.numDataBuffers++;
      this.numBytes +=
          util.sizeFromShape(a.shape) * util.bytesPerElement(a.dtype);
      this.backend.register(a.dataId, a.shape, a.dtype);
    }
    this.refCounter.set(a.dataId, refCount + 1);
    if (!(a instanceof Variable)) {
      this.track(a);
    }
  }

  registerVariable(v: Variable) {
    if (this.registeredVariables[v.name] != null) {
      throw new Error(`Variable with name ${v.name} was already registered`);
    }
    this.registeredVariables[v.name] = v;
  }

  disposeTensor(a: Tensor): void {
    if (!this.refCounter.has(a.dataId)) {
      return;
    }
    this.numTensors--;
    const refCount = this.refCounter.get(a.dataId);
    if (refCount <= 1) {
      this.refCounter.delete(a.dataId);
      this.backend.disposeData(a.dataId);
      this.numDataBuffers--;
      this.numBytes -=
          util.sizeFromShape(a.shape) * util.bytesPerElement(a.dtype);
    } else {
      this.refCounter.set(a.dataId, refCount - 1);
    }
    // TODO(nsthorat): Construct an error and save the stack trace for
    // debugging when in debug mode. Creating a stack trace is too expensive
    // to do unconditionally.
  }

  disposeVariables(): void {
    for (const varName in this.registeredVariables) {
      const v = this.registeredVariables[varName];
      this.disposeTensor(v);
      delete this.registeredVariables[varName];
    }
  }

  memory(): MemoryInfo {
    const info = this.backend.memory() as MemoryInfo;
    info.numTensors = this.numTensors;
    info.numDataBuffers = this.numDataBuffers;
    info.numBytes = this.numBytes;
    return info;
  }

  private shouldRecord(): boolean {
    return this.activeTape != null && this.customGradientDepth === 0;
  }

  private addTapeNode(
      inputs: Tensor[], result: Tensor,
      gradientsFunc: (dy: Tensor) => Tensor[]): void {
    const inputsMap: NamedTensorMap = {};
    inputs.forEach((input, idx) => {
      inputsMap[idx] = input;
    });

    const gradient = (dy: Tensor) => {
      const res = gradientsFunc(dy);
      const resMap: NamedGradientMap = {};
      res.forEach((r, idx) => {
        resMap[idx] = () => r;
      });
      return resMap;
    };

    const tapeNode: TapeNode = {
      id: this.nextTapeNodeId++,
      name: this.activeScope.name,
      inputs: inputsMap,
      output: result,
      gradient
    };
    this.activeTape.push(tapeNode);
  }

  keep<T extends Tensor>(result: T): T {
    if (this.scopeStack.length === 1 && ENV.engine.safeMode) {
      throw new Error(
          'Safe mode is ON. Enclose all tensor operations inside tf.tidy(): ' +
          'tf.tidy(() => {...}) to avoid memory leaks.');
    }
    this.keepTensors.add(result.id);
    return result;
  }

  /**
   * Start a scope. Use this with endScope() to achieve the same functionality
   * as scope() without the need for a function closure.
   */
  startScope(name?: string, gradientsMode = false) {
    if (gradientsMode && this.gradientScopeCount === 0) {
      this.activeTape = [];
    }
    if (gradientsMode) {
      this.gradientScopeCount++;
    }

    const scopeInfo: ScopeState = {track: []};
    if (name) {
      scopeInfo.name = name;
    }
    this.scopeStack.push(scopeInfo);
    this.activeScope = scopeInfo;
  }

  /**
   * End a scope. Use this with startScope() to achieve the same functionality
   * as scope() without the need for a function closure.
   */
  endScope(result: TensorContainer, gradientsMode = false) {
    if (gradientsMode) {
      this.gradientScopeCount--;
      if (this.gradientScopeCount === 0) {
        this.activeTape = null;
      }
    }

    const tensorsToKeep = new Set(this.keepTensors);

    const tensorsToTrackInParent = util.getTensorsInContainer(result);
    tensorsToTrackInParent.forEach(tensor => tensorsToKeep.add(tensor.id));

    // Dispose the arrays tracked in this scope.
    for (let i = 0; i < this.activeScope.track.length; i++) {
      const tensor = this.activeScope.track[i];
      if (tensorsToKeep.has(tensor.id)) {
        continue;
      }

      if (this.activeTape != null) {
        tensorsToTrackInParent.push(tensor);
      } else {
        tensor.dispose();
      }
    }

    const oldScope = this.scopeStack.pop();
    this.activeScope = this.scopeStack.length === 0 ?
        {track: []} :
        this.scopeStack[this.scopeStack.length - 1];

    // Track the current result in the parent scope.
    tensorsToTrackInParent.forEach(tensor => {
      // Only track the tensor if was allocated in the inner scope and is not
      // globally kept.
      if (!this.keepTensors.has(tensor.id) &&
          util.isTensorInList(tensor, oldScope.track)) {
        this.track(tensor);
      }
    });
  }

  /**
   * Returns gradients of `f` with respect to each of the `xs`. The gradients
   * returned are of the same length as `xs`, but some might be null if `f` was
   * not a function of that `x`. It also takes optional dy to multiply the
   * gradient, which defaults to `1`.
   */
  gradients<T extends Tensor>(
      f: () => T, xs: Tensor[], dy?: T,
      allowNoGradients = false): {value: T, grads: Tensor[]} {
    util.assert(xs.length > 0, 'gradients() received an empty list of xs.');

    return tidy('gradients', () => {
      const y = f();
      util.assert(
          y instanceof Tensor,
          'The result y returned by f() must be a tensor.');
      // Filter out the nodes that don't connect x => y.
      const filteredTape = getFilteredNodesXToY(this.activeTape, xs, y);
      if (!allowNoGradients && filteredTape.length === 0 && xs.length > 0) {
        throw new Error(
            'Cannot compute gradient of y=f(x) with respect to x. Make sure ' +
            'that the f you passed encloses all operations that lead from x ' +
            'to y.');
      }

      const accumulatedGradientMap: {[tensorId: number]: Tensor} = {};
      accumulatedGradientMap[y.id] = (dy == null) ? ops.ones(y.shape) : dy;

      // Backprop gradients through the filtered nodes.
      backpropagateGradients(accumulatedGradientMap, filteredTape);

      const grads = xs.map(x => accumulatedGradientMap[x.id]);
      return {value: y, grads};
    }, true /* gradientsMode */);
  }

  customGrad<T extends Tensor>(f: CustomGradientFunc<T>):
      (...args: Tensor[]) => T {
    util.assert(
        util.isFunction(f),
        'The f passed in customGrad(f) must be a function.');
    return (...inputs: Tensor[]): T => {
      util.assert(
          inputs.every(t => t instanceof Tensor),
          'The args passed in customGrad(f)(x1, x2,...) must all be tensors');
      this.customGradientDepth++;

      let gradientsFunc: (dy: T) => Tensor | Tensor[];
      const gradientsMode = true;
      const result = tidy(f.name, () => {
        const {value, gradFunc} = f(...inputs);
        util.assert(
            value instanceof Tensor,
            'The function f passed in customGrad(f) must return an object ' +
                'where `obj.value` is a tensor');
        util.assert(
            util.isFunction(gradFunc),
            'The function f passed in customGrad(f) must return an object ' +
                'where `obj.gradFunc` is a function.');
        gradientsFunc = gradFunc;
        return value;
      }, gradientsMode);

      this.customGradientDepth--;

      if (this.shouldRecord()) {
        const gradFunc = (dy: T): Tensor[] => {
          const res = gradientsFunc(dy);
          const grads: Tensor[] = Array.isArray(res) ? res : [res];
          util.assert(
              grads.length === inputs.length,
              'The function f passed in customGrad(f) must return an object ' +
                  'where `obj.gradFunc` is a function that returns the same ' +
                  'number of tensors as inputs passed to f(...).');
          util.assert(
              grads.every(t => t instanceof Tensor),
              'The function f passed in customGrad(f) must return an object ' +
                  'where `obj.gradFunc` is a function that returns a list of ' +
                  'only tensors.');
          return grads;
        };
        this.addTapeNode(inputs, result, gradFunc);
      }
      return result;
    };
  }

  // Forwarding to backend.
  write(dataId: DataId, values: TypedArray): void {
    this.backend.write(dataId, values);
  }
  readSync(dataId: DataId): TypedArray {
    return this.backend.readSync(dataId);
  }
  read(dataId: DataId): Promise<TypedArray> {
    return this.backend.read(dataId);
  }
  fromPixels(
      pixels: ImageData|HTMLImageElement|HTMLCanvasElement|HTMLVideoElement,
      numChannels: number): Tensor3D {
    return this.backend.fromPixels(pixels, numChannels);
  }
  async time(query: () => void): Promise<TimingInfo> {
    const start = performance.now();
    const timingInfo = await this.backend.time(query) as TimingInfo;
    timingInfo.wallMs = performance.now() - start;
    return timingInfo;
  }

  /**
   * Tracks a Tensor in the current scope to be automatically cleaned up
   * when the current scope ends, and returns the value.
   *
   * @param result The Tensor to track in the current scope.
   */
  private track<T extends Tensor>(result: T): T {
    if (this.scopeStack.length === 1 && this.safeMode) {
      throw new Error(
          'Safe mode is ON. Enclose all tensor operations inside tf.tidy(): ' +
          'tf.tidy(() => {op();...}); to avoid memory leaks.');
    }
    this.activeScope.track.push(result);
    return result;
  }
}

/** @docalias Function */
export type ScopeFn<T extends TensorContainer> = () => T;