BaseRandomLayer Abstract Layer and RandomWidth Preprocessing Layer #7345

RWallie · 2023-02-06T18:01:11Z

BaseRandomLayer Abstract Layer and RandomWidth Preprocessing Layer
Co-authored-by:
Natalie Umanzor (@numanzor) [email protected]
Silvia Kocsis (@Silvia42) [email protected]
Ryan Wallace (@RWallie) [email protected]

This change is

…ctins for preprocessing layers Co-authored-by: Silvia Kocsis <[email protected]> Co-authored-by: Natalie Umanzor <[email protected]>

Co-authored-by: Silvia Kocsis <[email protected]> Co-authored-by: Natalie Umanzor <[email protected]>

…d handles randomization of width Co-authored-by: Silvia Kocsis <[email protected]> Co-authored-by: Natalie Umanzor <[email protected]>

Co-authored-by: Silvia Kocsis <[email protected]> Co-authored-by: Natalie Umanzor <[email protected]>

mattsoulanille

Thanks for the PR! I have a few questions and changes to suggest.

mattsoulanille · 2023-02-07T19:47:36Z

tfjs-layers/src/engine/base_random_layer.ts

+    super(args);
+  }
+
+  protected setRNGType = (rngType: string) => {


I think this can be a normal function on the class's prototype instead of an arrow function.

Suggested change

protected setRNGType = (rngType: string) => {

protected setRNGType(rngType: string) {

mattsoulanille · 2023-02-07T20:43:31Z

tfjs-layers/src/engine/base_random_layer.ts

+    super(args);
+  }
+
+  protected setRNGType = (rngType: string) => {


Should the rng type be initialized by an option in the constructor? I think that's how keras does it.

TFJS might not support all the keras RNG types, but that should be fine as long as we support pseudorandom and true random.

mattsoulanille · 2023-02-07T20:46:50Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+        if (this.seed !== null) {
+          this.widthFactor = randomUniform([1],
+            (1.0 + this.widthLower), (1.0 + this.widthUpper),
+            'float32', this.seed


If this.seed is set and never changed, this code will always return the same value. There is no automatic advancing of seeds built in to randomUniform because randomUniform is stateless. You may need to keep track of, and increment, a seed manually.

mattsoulanille · 2023-02-07T20:50:16Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+          this.widthFactor = randomUniform([1],
+            (1.0 + this.widthLower), (1.0 + this.widthUpper)


The if statement checking if seed === null is probably not necessary. You can pass null or undefined as the seed and it will behave the same as if you had not passed anything.

mattsoulanille · 2023-02-07T20:52:46Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+    if(args.seed) {
+      this.seed = args.seed;
+    } else {
+      this.seed = null;
+    }


This if statement prevents people from using the seed 0.

Suggested change

if(args.seed) {

this.seed = args.seed;

} else {

this.seed = null;

}

this.seed = args.seed;

mattsoulanille · 2023-02-07T22:32:40Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+        }
+        this.adjustedWidth = this.widthFactor.dataSync()[0] * imgWidth;
+        this.adjustedWidth = Math.round(this.adjustedWidth);
+        const size: [number, number] = [this.imgHeight, this.adjustedWidth];


Tiny nit: Using as const makes TypeScript correctly infer this as const [number, number] instead of number[].

Suggested change

const size: [number, number] = [this.imgHeight, this.adjustedWidth];

const size = [this.imgHeight, this.adjustedWidth] as const;

resizeBilinear and resizeNearestNeighbor expect a mutable array of two numbers. using as const assigns the readonly type

mattsoulanille · 2023-02-07T22:35:26Z

tfjs-layers/src/layers/preprocessing/random_width_test.ts

+    const rangeTensor = range(0, 16);
+    const inputTensor = reshape(rangeTensor, [4,4,1]);


Minor nit since rangeTensor only seems to be used once.

Suggested change

const rangeTensor = range(0, 16);

const inputTensor = reshape(rangeTensor, [4,4,1]);

const inputTensor = range(0, 16).reshape([4, 4, 1]);

mattsoulanille · 2023-02-07T22:53:33Z

tfjs-layers/src/engine/base_random_layer.ts

+export declare interface BaseRandomLayerArgs extends LayerArgs {}
+
+export type RNGTypes = {
+  [key: string]: Function;


Function is a very wide type here. Can it be narrowed to be more specific?

mattsoulanille · 2023-02-07T22:55:56Z

tfjs-layers/src/engine/base_random_layer.ts

+  private readonly rngTypes: RNGTypes = {
+    gamma: randomGamma,
+    normal: randomNormal,
+    standardNormal: randomStandardNormal,
+    uniform: randomUniform
+  };


If this constant is not specific to this class, it can probably be moved outside of the class. If that's the case, you might also be able to base the RNGTypes type off of this constant.

const RNG_TYPES = {...}; type RNGTypes = typeof RNG_TYPES;

mattsoulanille · 2023-02-07T22:56:42Z

tfjs-layers/src/engine/base_random_layer.ts

+  /** @nocollapse */
+  static className = 'RandomWidth';
+  randomGenerator: Function;
+  private rngType: string;


This can reference RNGTypes to be more specific (I know they're both the string type right now, but that may change if the other change I mentioned can be applied).

private rngType: keyof RNGTypes

RWallie · 2023-02-08T00:51:40Z

Thank you @mattsoulanille for reviewing our pull request. We are getting right on the changes!

Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

…ss, got rid of rng_types as all random distribution functions seem to be stateless Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

…nd stores random distribution functions as properties Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

…mLayer Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

Co-authored-by: Silvia Kocsis <[email protected]> Co-authored-by: Natalie Umanzor <[email protected]>

RWallie · 2023-02-09T17:47:37Z

We changed the abstract class layer BaseRandomLayer to map more directly to keras. We created a RandomGenerator class that contains the incrementing of seeds to handle pseudorandomness, as you mentioned. We also stored the random distribution functions as properties on RandomGenerator as they are stored as methods in keras. An instance of RandomGenerator is stored as a property on BaseRandomLayer. We use this instance to directly access randomUniform in the RandomWidth preprocessing layer, as does keras.

We previously had rng_types handling which random distribution would be selected in the preprocessing layers. We removed these - how we had them defined didn't appropriately relate to keras. As you said, in keras, the rng_types were there to determine whether the op was stateful or stateless - we believe all the random distribution functions to be stateless, so we didn't incorporate rng_types when creating an instance of RandomGenerator in BaseRandomLayer.

Some conditionals were removed / rewritten based on your suggestions.

Thanks again @mattsoulanille

Silvia42 · 2023-02-09T20:08:26Z

tfjs-presubmit (learnjs-174218) failed. The details are not visible for us. Can you please let us know what is the problem? @mattsoulanille

mattsoulanille · 2023-02-10T01:48:42Z

Sorry, I think I gave some bad suggestions in my review.

Safari 11.1.2 (Mac OS 10.13.6) RandomWidth Layer webgl1 {"WEBGL_VERSION":1,"WEBGL_CPU_FORWARD":false,"WEBGL_SIZE_UPLOAD_UNIFORM":0} Returns correct, randomly scaled width of Rank 3 Tensor FAILED
	TypeError: inputShape.at is not a function. (In 'inputShape.at(-3)', 'inputShape.at' is undefined) in http://bs-local.com:9876/context.html (line 39696)
	<Jasmine>

Looks like this Safari version does not support Array.at and we're not polyfilling it. You should revert back to what you had before (someArray[someArray.length - someNumber]).

You should be able to access the build logs if you join the discussion or announcement mailing list

Thanks for making the changes! I'll take a look and review them tomorrow.

Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

RWallie · 2023-02-14T19:10:30Z

Hey @mattsoulanille hope you are doing well. Just wanted to check in and see how things are looking with our latest changes to the PR. Thanks again!

mattsoulanille

Sorry I didn't get to this sooner.

mattsoulanille · 2023-02-15T04:18:11Z

tfjs-layers/src/backend/random_generator.ts

+import { randomGamma, randomNormal} from '@tensorflow/tfjs-core';
+import { randomStandardNormal, randomUniform } from '@tensorflow/tfjs-core';
+
+type randomGammaType = typeof randomGamma;
+type randomNormalType = typeof randomNormal;
+type randomStandardNormalType = typeof randomStandardNormal;
+type randomUniformType = typeof randomUniform;
+
+export class RandomGenerator {
+  /** @nocollapse */
+  static className = 'RandomGenerator';
+  protected currentSeed: number;
+  private readonly seed: number;
+  randomGamma: randomGammaType;
+  randomNormal: randomNormalType;
+  randomStandardNormal: randomStandardNormalType;
+  randomUniform: randomUniformType;
+
+  constructor(seed: number) {
+    this.seed = seed;
+    this.currentSeed = seed;
+    this.randomGamma = randomGamma;
+    this.randomNormal = randomNormal;
+    this.randomStandardNormal = randomStandardNormal;
+    this.randomUniform = randomUniform;
+  }
+
+  next(): number | null {
+    if (typeof this.seed === 'number'){
+      return ++this.currentSeed;
+    }
+    return null;
+  }
+}


I don't think you need to store the random sampler functions on this class, since they're never called by the class. When you need them somewhere else, you can import them. This class should just keep track of the seed.

Suggested change

import { randomGamma, randomNormal} from '@tensorflow/tfjs-core';

import { randomStandardNormal, randomUniform } from '@tensorflow/tfjs-core';

type randomGammaType = typeof randomGamma;

type randomNormalType = typeof randomNormal;

type randomStandardNormalType = typeof randomStandardNormal;

type randomUniformType = typeof randomUniform;

export class RandomGenerator {

/** @nocollapse */

static className = 'RandomGenerator';

protected currentSeed: number;

private readonly seed: number;

randomGamma: randomGammaType;

randomNormal: randomNormalType;

randomStandardNormal: randomStandardNormalType;

randomUniform: randomUniformType;

constructor(seed: number) {

this.seed = seed;

this.currentSeed = seed;

this.randomGamma = randomGamma;

this.randomNormal = randomNormal;

this.randomStandardNormal = randomStandardNormal;

this.randomUniform = randomUniform;

}

next(): number | null {

if (typeof this.seed === 'number'){

return ++this.currentSeed;

}

return null;

}

}

export class RandomSeed {

private currentSeed: number;

constructor(readonly seed: number) {

this.currentSeed = seed;

}

next(): number {

return ++this.currentSeed;

}

}

mattsoulanille · 2023-02-15T04:24:01Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+        switch (true) {
+          case this.interpolation === 'bilinear':
+            return image.resizeBilinear(inputs, size);
+          case this.interpolation === 'nearest':
+            return image.resizeNearestNeighbor(inputs, size);
+          default:
+            throw new Error(`Interpolation is ${this.interpolation}
+            but only ${[...INTERPOLATION_METHODS]} are supported`);


Suggested change

switch (true) {

case this.interpolation === 'bilinear':

return image.resizeBilinear(inputs, size);

case this.interpolation === 'nearest':

return image.resizeNearestNeighbor(inputs, size);

default:

throw new Error(`Interpolation is ${this.interpolation}

but only ${[...INTERPOLATION_METHODS]} are supported`);

switch (this.interpolation) {

case 'bilinear':

return image.resizeBilinear(inputs, size);

case 'nearest':

return image.resizeNearestNeighbor(inputs, size);

default:

throw new Error(`Interpolation is ${this.interpolation}

but only ${[...INTERPOLATION_METHODS]} are supported`);

mattsoulanille · 2023-02-15T17:47:49Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+  static override className = 'RandomWidth';
+  private readonly factor: number | [number, number];
+  private readonly interpolation?: InterpolationType;  // defualt = 'bilinear
+  private seed?: number; // default null


The RandomGenerator / RandomSeed should track this seed for you.

Suggested change

private seed?: number; // default null

mattsoulanille · 2023-02-15T17:51:14Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+          'float32', this.seed
+        );
+
+        this.seed = this.randomGenerator.next();


When saving this layer, should we save the current seed or the original seed we initialized it with?

mattsoulanille · 2023-02-15T17:53:33Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+        this.imgHeight = inputShape[inputShape.length - 3];
+        const imgWidth = inputShape[inputShape.length - 2];
+
+        this.widthFactor = this.randomGenerator.randomUniform([1],


randomUniform is stateless, so you can just import and call it directly. Same with the other random functions. You don't need to put them on the RandomGenerator class.

Suggested change

this.widthFactor = this.randomGenerator.randomUniform([1],

this.widthFactor = randomUniform([1],

mattsoulanille · 2023-02-15T17:55:25Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+          'float32', this.seed
+        );
+
+        this.seed = this.randomGenerator.next();


Try to avoid duplicating state where possible. The RandomGenerator / RandomSeed class already stores the seed. If you need to check the seed without incrementing it, make it public on the class or add a getter.

Suggested change

this.seed = this.randomGenerator.next();

mattsoulanille · 2023-02-15T17:56:58Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+      `);
+    }
+
+    this.seed = seed;


Suggested change

this.seed = seed;

mattsoulanille · 2023-02-15T17:58:27Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+
+    return tidy(() => {
+        const input = getExactlyOneTensor(inputs);
+        const inputShape = input.shape;


You can just use input.shape directly. No need to assign a local inputShape variable.

mattsoulanille · 2023-02-15T17:59:39Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+    const config: serialization.ConfigDict = {
+      'factor': this.factor,
+      'interpolation': this.interpolation,
+      'seed': this.seed,


BaseRandomLayer should implement its own getConfig that saves the seed.

Suggested change

'seed': this.seed,

RWallie · 2023-02-15T22:37:43Z

No worries @mattsoulanille ! Thanks for the review!

chunnienc

I will take a closer look after @mattsoulanille 's comments and suggestions are resolved.

chunnienc · 2023-02-15T22:55:19Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+   factor: number | [number, number];
+   interpolation?: InterpolationType; // default = 'bilinear';
+   seed?: number;// default = false;
+   autoVectorize?:boolean;


Space between ':' and 'boolean'. You cloud run clang-format on all your files.

chunnienc · 2023-02-15T22:57:50Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+ * tf methods unimplemented in tfjs: 'bicubic', 'area', 'lanczos3', 'lanczos5',
+ *                                   'gaussian', 'mitchellcubic'
+ *
+ */


Is this the comment for the class? If so please put it on top of the class.

Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

RWallie · 2023-02-17T00:16:45Z

We cleaned up the RandomGenerator class and changed it to RandomSeed - It only handles the pseudorandomness now and the random sampler functions will be imported directly into the random preprocessing layers, as we did in random_width.ts.

We took the seed property off of RandomWidth, as you mentioned it's not necessary and is duplicating state. We made the currentSeed inside of RandomSeed public so it would be accessible when calling randomUniform inside the preprocessing layer.

We set up a getConfig() inside of BaseRandomLayer that saves the seed

Thank you @mattsoulanille @chunnienc !

RWallie · 2023-02-22T19:34:52Z

Hey @mattsoulanille wanted to check in on the PR to see if you had any updates. Thanks!

mattsoulanille · 2023-02-23T00:50:28Z

tfjs-layers/src/engine/base_random_layer.ts

+  protected randomGenerator: RandomSeed;
+  private seed?: number;


The RandomSeed class already stores the value of seed, so you don't need to store it again here.

Suggested change

protected randomGenerator: RandomSeed;

private seed?: number;

protected randomSeed: RandomSeed;

mattsoulanille · 2023-02-23T00:50:44Z

tfjs-layers/src/engine/base_random_layer.ts

+    this.seed = args.seed;
+    this.randomGenerator = new RandomSeed(this.seed);


Suggested change

this.seed = args.seed;

this.randomGenerator = new RandomSeed(this.seed);

this.randomSeed = new RandomSeed(args.seed);

mattsoulanille · 2023-02-23T00:50:54Z

tfjs-layers/src/engine/base_random_layer.ts

+
+  override getConfig(): serialization.ConfigDict {
+    const config: serialization.ConfigDict = {
+      'seed': this.seed


Suggested change

'seed': this.seed

'seed': this.randomSeed.seed,

mattsoulanille · 2023-02-23T00:56:10Z

tfjs-layers/src/engine/base_random_layer.ts


  constructor(args: BaseRandomLayerArgs) {
    super(args);
-    this.randomGenerator = new RandomGenerator(args.seed);
+    this.seed = args.seed;
+    this.randomGenerator = new RandomSeed(this.seed);


args.seed has type number | undefined. I think there needs to be a check for undefined here before constructing the RandomSeed (or make seed optional on RandomSeed).

What is the behavior if no seed is passed? Is it true randomness, or does it just choose a seed randomly, or does it choose the same seed every time? If it's the second of these, then I suggest allowing and undefined seed (or no seed) to be passed to RandomSeed. Then, you can have RandomSeed generate a seed randomly.

The utility classes for the random sampler classes use Math.random() to assign the seed if none exists. Here's the example from randomUniform's UniformRandom utility class

mattsoulanille · 2023-02-23T01:04:33Z

tfjs-layers/src/layers/preprocessing/random_width.ts

+        'float32', this.randomGenerator.currentSeed
+      );
+
+      this.randomGenerator.next();


this.randomGenerator.next() increments the seed and then returns the new seed, so you can use its return value instead of incrementing it separately.

Suggested change

'float32', this.randomGenerator.currentSeed

);

this.randomGenerator.next();

'float32', this.randomGenerator.next());

If you want to use the current seed before incrementing it, change the implementation of RandomGenerator.next to return seed++ (increments after getting the value) instead of ++seed (increments before getting the value).

…, and added test for RandomSeed Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

RWallie · 2023-02-23T20:18:02Z

Hey @mattsoulanille we removed the seed from BaseRandomLayers so we won't be duplicating state. We handled an undefined seed inside of RandomSeed . When calling RandomSeed's next method, we return undefined if the seed is undefined. Randomness gets handled inside of the utility classes for each random sampler function if the seed is null or undefined and generates a seed using Math.random().

RWallie · 2023-03-02T18:40:27Z

Hey @mattsoulanille - wanted to reach out to see how things are looking with our latest PR. Thank you!

mattsoulanille

LGTM. I'm sorry it took me so long to get to this.

mattsoulanille · 2023-03-08T17:18:15Z

tfjs-layers/src/backend/random_seed.ts

+  seed: number | undefined;
+  constructor(seed: number | undefined) { 
+    this.seed = seed; 
  }
-  next() { 
-    ++this.currentSeed; 
+  next(): number | undefined { 
+    if (this.seed === undefined) {
+      return undefined;
+    }
+    return this.seed++; 


Note to other reviewer: When no seed is specified, the keras implementation does not choose a random seed. It just generates random values. Saving it yields something like this, with the seed set to Null:

{'name': 'random_width_3', 'trainable': True, 'dtype': 'float32', 'factor': (-0.2, 0.3), 'interpolation': 'bilinear', 'seed': None}

This is truly random, and there is no reproducibility here.

On the other hand, when the user sets the seed, it gets saved to the config.

{'name': 'random_width_2', 'trainable': True, 'dtype': 'float32', 'factor': (-0.2, 0.3), 'interpolation': 'bilinear', 'seed': 1}

AdamLang96 · 2023-03-08T18:07:54Z

@mattsoulanille Thanks for all the help on this!

mattsoulanille · 2023-03-13T23:53:54Z

@chunnienc Please take a look when you get a chance. Thanks!

RWallie and others added 5 commits February 6, 2023 09:35

created baseRandomLayer Abstract layer, contains random generator fun…

d676ae8

…ctins for preprocessing layers Co-authored-by: Silvia Kocsis <[email protected]> Co-authored-by: Natalie Umanzor <[email protected]>

renamed file to base_random_layer

c6a144c

Co-authored-by: Silvia Kocsis <[email protected]> Co-authored-by: Natalie Umanzor <[email protected]>

created random_width preprocessing layer - extends baseRandomLayer an…

a9fd524

…d handles randomization of width Co-authored-by: Silvia Kocsis <[email protected]> Co-authored-by: Natalie Umanzor <[email protected]>

tests for randomWidth preproccessing layer

253c0e7

Co-authored-by: Silvia Kocsis <[email protected]> Co-authored-by: Natalie Umanzor <[email protected]>

added randomWidth preprocessing layer to exports_layers

3159cc9

Co-authored-by: Silvia Kocsis <[email protected]> Co-authored-by: Natalie Umanzor <[email protected]>

chunnienc self-requested a review February 6, 2023 18:07

pyu10055 requested a review from mattsoulanille February 6, 2023 23:28

mattsoulanille requested changes Feb 7, 2023

View reviewed changes

Silvia42 and others added 9 commits February 8, 2023 21:31

reformatted tests using reshape property

97e9f66

Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

updated BaseRandomLayer to contain an instance of RandomGenerator cla…

17ff308

…ss, got rid of rng_types as all random distribution functions seem to be stateless Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

random generator handles pseudorandomness for autoincrementing seed a…

0d7ff04

…nd stores random distribution functions as properties Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

updated randomUniform call with RangomGenerator instance on BaseRando…

9188041

…mLayer Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

increment currentSeed before return in random_generator

639712f

fixed linting

3086655

removed reset method from RandomGenerator

9e79b22

Co-authored-by: Silvia Kocsis <[email protected]> Co-authored-by: Natalie Umanzor <[email protected]>

unpacked args and chaned conditionals

f9bf89a

Co-authored-by: Silvia Kocsis <[email protected]> Co-authored-by: Natalie Umanzor <[email protected]>

merged master

de0a7c6

Silvia42 and others added 2 commits February 9, 2023 22:55

fix Safari compatibility

de296dc

Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

added test for RandomGenerator Class

b1034a5

Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

mattsoulanille requested changes Feb 15, 2023

View reviewed changes

chunnienc requested changes Feb 16, 2023

View reviewed changes

Silvia42 and others added 2 commits February 16, 2023 14:39

Merge branch 'master' into randomLayers

5077b96

Change class RandomGenerator to class RandomSeed

aba23b7

Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

mattsoulanille reviewed Feb 23, 2023

View reviewed changes

Silvia42 and others added 3 commits February 23, 2023 12:52

Merge branch 'master' into randomLayers

384b758

updated RandomSeed to handle undefined value, removed duplicate state…

f4b05b2

…, and added test for RandomSeed Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

re-labeled currentSeed as seed

9ae69a4

Co-authored-by: Natalie Umanzor <[email protected]> Co-authored-by: Ryan Wallace <[email protected]>

mattsoulanille approved these changes Mar 8, 2023

View reviewed changes

mattsoulanille requested a review from chunnienc March 8, 2023 17:19

chunnienc approved these changes Mar 14, 2023

View reviewed changes

mattsoulanille added 3 commits March 14, 2023 12:59

Merge branch 'master' into randomLayers

1d7fc23

Merge branch 'master' into randomLayers

b183550

Merge branch 'master' into randomLayers

c77e1c3

mattsoulanille merged commit 86e7f4a into tensorflow:master Mar 15, 2023

	protected setRNGType = (rngType: string) => {
	protected setRNGType(rngType: string) {

		this.widthFactor = randomUniform([1],
		(1.0 + this.widthLower), (1.0 + this.widthUpper)

	const size: [number, number] = [this.imgHeight, this.adjustedWidth];
	const size = [this.imgHeight, this.adjustedWidth] as const;

		const rangeTensor = range(0, 16);
		const inputTensor = reshape(rangeTensor, [4,4,1]);

	const rangeTensor = range(0, 16);
	const inputTensor = reshape(rangeTensor, [4,4,1]);
	const inputTensor = range(0, 16).reshape([4, 4, 1]);

	this.widthFactor = this.randomGenerator.randomUniform([1],
	this.widthFactor = randomUniform([1],

		protected randomGenerator: RandomSeed;
		private seed?: number;

	protected randomGenerator: RandomSeed;
	private seed?: number;
	protected randomSeed: RandomSeed;

		this.seed = args.seed;
		this.randomGenerator = new RandomSeed(this.seed);

	this.seed = args.seed;
	this.randomGenerator = new RandomSeed(this.seed);
	this.randomSeed = new RandomSeed(args.seed);

BaseRandomLayer Abstract Layer and RandomWidth Preprocessing Layer #7345

BaseRandomLayer Abstract Layer and RandomWidth Preprocessing Layer #7345

Conversation

RWallie commented Feb 6, 2023 • edited by dsmilkov Loading

mattsoulanille left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

RWallie commented Feb 8, 2023

RWallie commented Feb 9, 2023 • edited Loading

Silvia42 commented Feb 9, 2023

mattsoulanille commented Feb 10, 2023

RWallie commented Feb 14, 2023

mattsoulanille left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

RWallie commented Feb 15, 2023

chunnienc left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

RWallie commented Feb 17, 2023

RWallie commented Feb 22, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

RWallie commented Feb 23, 2023

RWallie commented Mar 2, 2023

mattsoulanille left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AdamLang96 commented Mar 8, 2023

mattsoulanille commented Mar 13, 2023

RWallie commented Feb 6, 2023 •

edited by dsmilkov

Loading

RWallie commented Feb 9, 2023 •

edited

Loading