Add GAN Example #181

t-ae · 2019-07-25T02:01:50Z

Related issue: #150

Added (non-convolutional) GAN example on MNIST.

After 10 epochs of training, output is like this:

GAN/main.swift

t-ae · 2019-07-25T03:22:06Z

By the way I couldn't pass leakyRelu to Denses directly.

error: generic parameter 'T' could not be inferred
    var dense1 = Dense<Float>(inputSize: latentDim, outputSize: latentDim*2, activation: leakyRelu)

So I had to use wrapper.

swift-models/GAN/main.swift

Lines 81 to 83 in 19ba68d

    
           func lrelu(x: Tensor<Float>) -> Tensor<Float> { 
        
               leakyRelu(x) 
        
           }

swift-models/GAN/main.swift

Line 90 in 19ba68d

    
           var dense1 = Dense<Float>(inputSize: latentDim, outputSize: latentDim*2, activation: lrelu)

It seems this old bug is related.
https://bugs.swift.org/browse/SR-504

I think it's good to have separate implementation of func leakyRelu<T: TensorFlowFloatingPoint>(_ x: Tensor<T>) -> Tensor<T> in swift-apis until the bug is resolved.

GAN/main.swift

t-ae · 2019-07-26T01:13:24Z

There was another disgrace inherited from autoencoder example.
It loads labels but uses its size only.

I totally removed labels in this commit.
e52b0ec

I'll create another PR for autoencoder if this change is accepted.

GAN/main.swift

rxwei · 2019-07-26T01:01:31Z

GAN/main.swift

+    )
+}
+
+// Models


Add an empty line, since this is describing both Generator and Discriminator below.

Suggested change

// Models

// Models

rxwei · 2019-07-26T01:01:44Z

GAN/main.swift

+    }
+}
+
+// Loss functions


Suggested change

// Loss functions

// Loss functions

GAN/main.swift

rxwei · 2019-07-26T01:15:50Z

GAN/main.swift

+    // Inference phase
+    Context.local.learningPhase = .inference
+    let testImage: Tensor<Float> = generator(testVector)
+


Remove redundant empty line.

rxwei · 2019-07-26T01:15:58Z

GAN/main.swift

+    plotTestImage(testImage, name: "epoch-\(epoch)-output")
+
+    let lossG = generatorLossFunc(fakeLogits: testImage)
+


Remove redundant empty line.

rxwei · 2019-07-26T01:16:14Z

GAN/main.swift

+            let loss = discriminatorLossFunc(realLogits: realLogits, fakeLogits: fakeLogits)
+            return loss
+        }
+


Remove redundant empty line.

rxwei · 2019-07-26T01:16:43Z

GAN/main.swift

+    // Training phase
+    Context.local.learningPhase = .training
+    for i in 0 ..< Int(labels.shape[0]) / batchSize {
+        // Alternative update


Suggested change

// Alternative update

// Perform alternative update.

rxwei · 2019-07-26T01:16:51Z

GAN/main.swift

+    Context.local.learningPhase = .training
+    for i in 0 ..< Int(labels.shape[0]) / batchSize {
+        // Alternative update
+


Suggested change

BradLarson · 2019-07-26T02:29:36Z

@t-ae - I wouldn't worry too much about any issues with MNIST loading here or in the Autoencoder example, because I'm in the process of extracting datasets to be separate from the models anyway and much of this code will be removed and consolidated in a central location. Once you have the last few things resolved here, I can handle the Autoencoder and its MNIST inputs.

GAN/main.swift

rxwei · 2019-07-26T03:27:50Z

GAN/main.swift

+func plotTestImage(_ testImage: Tensor<Float>, name: String) {
+    var imageGrid = testImage.reshaped(to: [testImageGridSize, testImageGridSize,
+                                            imageHeight, imageWidth])
+


Remove redundant empty line.

rxwei · 2019-07-26T03:27:53Z

GAN/main.swift

+    imageGrid = imageGrid.transposed(withPermutations: [0, 2, 1, 3])
+    imageGrid = imageGrid.reshaped(to: [(imageHeight+2)*testImageGridSize,
+                                        (imageWidth+2)*testImageGridSize])
+


Remove redundant empty line.

rxwei · 2019-07-26T03:28:00Z

GAN/main.swift

+
+    // Transpose to create single image.
+    imageGrid = imageGrid.transposed(withPermutations: [0, 2, 1, 3])
+    imageGrid = imageGrid.reshaped(to: [(imageHeight+2)*testImageGridSize,


Suggested change

imageGrid = imageGrid.reshaped(to: [(imageHeight+2)*testImageGridSize,

imageGrid = imageGrid.reshaped(to: [(imageHeight + 2) * testImageGridSize,

rxwei · 2019-07-26T03:28:09Z

GAN/main.swift

+    // Transpose to create single image.
+    imageGrid = imageGrid.transposed(withPermutations: [0, 2, 1, 3])
+    imageGrid = imageGrid.reshaped(to: [(imageHeight+2)*testImageGridSize,
+                                        (imageWidth+2)*testImageGridSize])


Suggested change

(imageWidth+2)*testImageGridSize])

(imageWidth + 2) * testImageGridSize])

rxwei · 2019-07-26T03:30:10Z

GAN/main.swift

+}
+
+struct Discriminator: Layer {
+    var dense1 = Dense<Float>(inputSize: imageSize, outputSize: 256, activation: { leakyRelu($0) })


Suggested change

var dense1 = Dense<Float>(inputSize: imageSize, outputSize: 256, activation: { leakyRelu($0) })

var dense1 = Dense<Float>(inputSize: imageSize, outputSize: 256,

activation: { leakyRelu($0) })

For consistency with other dense layers below.

rxwei · 2019-07-26T03:31:24Z

GAN/main.swift

+    let array = np.array([image.scalars])
+    let pixels = array.reshape(image.shape)
+    if !FileManager.default.fileExists(atPath: outputFolder) {
+        try! FileManager.default.createDirectory(atPath: outputFolder,


We aren't using Objective-C style formatting for function call arguments. Could you reformat this like the following?

Suggested change

try! FileManager.default.createDirectory(atPath: outputFolder,

try! FileManager.default.createDirectory(

atPath: outputFolder,

withIntermediateDirectories: false,

attributes: nil)

rxwei · 2019-07-26T03:31:50Z

GAN/main.swift

+let imageSize = imageHeight * imageWidth
+let latentSize = 64
+
+func plot(image: Tensor<Float>, name: String) {


When the first argument label happens to be the object of the overall verb phrase, append it to the base name and omit the argument label.

Suggested change

func plot(image: Tensor<Float>, name: String) {

func plotImage(_ image: Tensor<Float>, name: String) {

rxwei · 2019-07-26T03:33:58Z

GAN/main.swift

+// Turn off using display on server / Linux.
+matplotlib.use("Agg")
+
+// Some globals


Suggested change

// Some globals

This comment isn't providing much useful information as the code is self-explanatorily "some globals".

GAN/main.swift

rxwei · 2019-07-26T07:09:00Z

GAN/main.swift

+
+/// Returns `size` samples of noise vector.
+func sampleVector(size: Int) -> Tensor<Float> {
+    Tensor<Float>(randomNormal: [size, latentSize])


Omit redundant type information since it can be inferred from the return type.

Suggested change

Tensor<Float>(randomNormal: [size, latentSize])

Tensor(randomNormal: [size, latentSize])

rxwei · 2019-07-26T07:09:49Z

GAN/main.swift

+    print("Reading data.")
+    let images = readFile(imagesFile).dropFirst(16).map { Float($0) }
+    let rowCount = images.count / imageSize
+


Remove redundant empty line.

rxwei · 2019-07-26T07:13:22Z

GAN/README.md

+
+To train the model, run:
+
+```


Suggested change

```

```sh

rxwei · 2019-07-26T07:17:39Z

GAN/README.md

+```
+swift run GAN
+```
+If you using brew to install python2 and modules, change the path:


Not sure what you mean exactly by "modules".

I feel lines 24-30 are unnecessary since they are just describing an installation step that's required for the entire models repository.

It came from Autoencoder/README.md.
I just copied it and didn't consider much.

Ah ok, we should unify the documentation in all models. @BradLarson how about having a standard README template?

@rxwei - We definitely will need to rework the READMEs across the examples. Varying levels of information is provided in each, as well as the language used for describing the models. A template for examples would be much appreciated, as well as a good central listing of them. The original Autoencoder README was a little rough (and I think the Python information was confusing and missing the need for matplotlib, etc.), thus the issues with this one derived from it.

rxwei · 2019-07-26T07:18:13Z

GAN/README.md

@@ -0,0 +1,30 @@
+# Simple GAN
+
+### After Epoch 1


Suggested change

### After Epoch 1

After Epoch 1:

rxwei · 2019-07-26T07:18:23Z

GAN/README.md

+<img src="images/epoch-1-output.png" height="270" width="360">
+</p>
+
+### After Epoch 10


Suggested change

### After Epoch 10

After Epoch 10:

BradLarson · 2019-07-26T13:51:42Z

Thank you for the great example. Sorry about the more involved process, when we reorganize and clean up the repository it should be easier for everyone to contribute models and examples like this. You'll have to copy over a lot less code, for one thing, which means less to review.

We appreciate your patience in the meantime.

* Use PascalCase for all Swift source files (#152) Except `main.swift` files which are the entry points for executables. Running the following command does not produce any output: `find . -type f -name "[[:lower:]]*.swift" | rg --pcre2 "^(?!.*(main.swift))"` Resolves #133. * Load config variable from `hparams.json` file so that Transformer can work with the bigger GPT-2 models (It is called "staged release". as of now, only 117M and 345M are available). (#154) * Style fixes from #154 (#155) * A few quick fixes to help unbreak swift-models. (#160) * Unbreak models by inserting `.call`. (#161) See https://bugs.swift.org/browse/TF-516 for additional context. * s/CotangentVector/TangentVector/g (#162) * Add MNIST test set evaluation and change hyperparameters. (#163) - Add MNIST test set evaluation. - The batch size and the optimizer are changed to `128` and `Adam`, respectively, based on test set evaluation results. Tested with higher batch sizes [256, 512], but found no improvement in performance. * Updated models to use revised name for callable (SE-0253), fixed a few issues caused by `swift-apis` changes. (#166) * Updated Transformer and MNIST models to work with 2019-06-04 development snapshot. * Updated other models. * Updated ResNet to work with 2019-06-04 development snapshot. * Updated Catch and CIFAR models. * Added a method to satisfy TensorGroup protocol that can pass build. * transformer: upstream api changes (#171) * rebuild resnet block based approach (#170) * add gym blackjack qlearning demo (#173) * Make 'GoModel' stored properties be variables. (#174) New 'Differentiable' derived conformances will not include constant properties in the tangent space. * Added .swift-format and updated .gitignore for Xcode 11's SwiftPM support. (#176) This replicates the following pull request on swift-apis: tensorflow/swift-apis#374 , adding a `swift-format` configuration file. Note the cautions indicated in that pull request around use of `swift-format` at present. This also adds a .gitignore line to prevent Xcode 11's new SwiftPM support from adding supporting files to the repository. * Replaced Python with Swift in CIFAR10 dataset loading (#178) * Replaced all Python code in the CIFAR10 and ResNet examples, removing Python 3 dependency. * Needed to import FoundationNetworking on Linux. * Added a check for FoundationNetworking, added an early exit for cached directory check. * Removed macOS availability check by targeting 10.13 in the package. * Style and formatting fixes. * Removed no-longer-needed _tensorHandles and supporting code. * Replace Autoencoder's tanh activation with sigmoid (#180) * Add GAN Example (#181) * Add GAN Example * Remove do blocks * Replace lrelu with closure * Remove typealiases * Add comment * Remove labels in GAN example * Rename variables * Code formatting, update comments * Fix: latentDim -> latentSize * Add spaces around * * Code formatting, rename loss functions * Fix: generatorLossFunc -> generatorLoss * Break lines to make it fit within 100 columns * Update comments * Rename plot->plotImage, imageGrid->gridImage, Code formatting * Refactor: Label creation * Update comment * Remove type parameter, empty line * Update readme * nightlies URL: s4tf-kokoro-artifact-testing => swift-tensorflow-artifacts (#183) * First steps in repository reorganization: extracting common MNIST dataset code (#182) * Extracted MNIST dataset, created LeNet network, added example combining the two. * Extracted redundant MNIST loading code from GAN and Autoencoder examples, replaced with central MNIST dataset. * Renamed input parameters and applied standard formatting style to MNIST. * Punctuation correction. Co-Authored-By: Richard Wei <[email protected]> * README formatting update. Co-Authored-By: Richard Wei <[email protected]> * Renamed trainImages -> trainingImages, corrected Python package names, formatted Package.swift. * Update Dockerfile to install Python libraries (#184) * Delete helpers.swift (#187) This file is no longer needed apple/swift #26023 and may be causing the segfault reported in swift-apis #186. * Continuing repository reorganization: extracting common CIFAR-10 and ResNet code (#185) * Extracted CIFAR-10 dataset and ResNet models into respective modules. * Minor formatting fixes. * Mirroring PR #187. * [Models] Fix enumeration of blocks in `ResidualBasicBlockStack` `init` (#192) * [Models] Fix enumeration of blocks in `ResidualBasicBlockStack` initialization * Start range at * Update ranges * SqueezeNet Implementation (#189) * Add files via upload * Implemented requested changes Corrected formatting mistakes, used dropout correctly, removed the hard coded number of classes and improved overall readability. * Update SqeezeNet.swift Renamed numClasses to classCount and limited all lines to a size of 100. * WideResNet - fix widenFactor and match model to citation (#193) * add identity connections to WideResNet * rename preact1 for clarity * remove extra relu, add dropout * fix declarartion * skip dropout in expansion blocks * remove enum, res layers to one line * Removal of deprecated allDifferentiableVariables. (#194) * Add JPEG loading / saving via Raw TF operations, removing Matplotlib dependencies (#188) * Added an Image struct as a wrapper for JPEG loading / saving, removed matplotlib dependency from GAN and Autoencoder examples using this. * Formatting update for Autoencoder. * Bring this inline with current API. * Made saveImage() a throwing function, improved formatting. * Changed function parameter. * Convert MNIST classifier to use sequential (#200) * Convert MNIST classifier to use sequential * Review Changes * Remove ImageClassification Models * Convert Autoencoder and Catch to use Sequential (#203) Partially fixes #202. * Adding shape-based inference tests for all image classification models (#198) * Implemented inference tests with random tensors and starting weights for all classification models. * Made sure tests ran on Linux, reshaped output of SqueezeNet to match other classification models. * WideResNet is expressed in terms of CIFAR10, so altered the inputs and outputs appropriately. * Wrapping the reshaping line in SqueezeNet. * Reset ownership. * Reworked TensorShape initializers to use array literals. * Minor formatting tweak to SqueezeNet. * Update deprecated APIs in Catch (#205) [swift-apis/#497](tensorflow/swift-apis#497) removes deprecated apis, now that we are moving onward to v0.5. This PR updates the catch file which uses a deprecated dense layer. * Update squeezenet and add V1.1 version (#204) * Update squeezenet and add V1.1 version * value error * Update tests * review changes * Fix '@noDerivative' warnings. (#208) (#209) (cherry picked from commit db72e4d)

* Use PascalCase for all Swift source files (#152) Except `main.swift` files which are the entry points for executables. Running the following command does not produce any output: `find . -type f -name "[[:lower:]]*.swift" | rg --pcre2 "^(?!.*(main.swift))"` Resolves #133. * Load config variable from `hparams.json` file so that Transformer can work with the bigger GPT-2 models (It is called "staged release". as of now, only 117M and 345M are available). (#154) * Style fixes from #154 (#155) * A few quick fixes to help unbreak swift-models. (#160) * Unbreak models by inserting `.call`. (#161) See https://bugs.swift.org/browse/TF-516 for additional context. * s/CotangentVector/TangentVector/g (#162) * Add MNIST test set evaluation and change hyperparameters. (#163) - Add MNIST test set evaluation. - The batch size and the optimizer are changed to `128` and `Adam`, respectively, based on test set evaluation results. Tested with higher batch sizes [256, 512], but found no improvement in performance. * Updated models to use revised name for callable (SE-0253), fixed a few issues caused by `swift-apis` changes. (#166) * Updated Transformer and MNIST models to work with 2019-06-04 development snapshot. * Updated other models. * Updated ResNet to work with 2019-06-04 development snapshot. * Updated Catch and CIFAR models. * Added a method to satisfy TensorGroup protocol that can pass build. * transformer: upstream api changes (#171) * rebuild resnet block based approach (#170) * add gym blackjack qlearning demo (#173) * Make 'GoModel' stored properties be variables. (#174) New 'Differentiable' derived conformances will not include constant properties in the tangent space. * Added .swift-format and updated .gitignore for Xcode 11's SwiftPM support. (#176) This replicates the following pull request on swift-apis: tensorflow/swift-apis#374 , adding a `swift-format` configuration file. Note the cautions indicated in that pull request around use of `swift-format` at present. This also adds a .gitignore line to prevent Xcode 11's new SwiftPM support from adding supporting files to the repository. * Replaced Python with Swift in CIFAR10 dataset loading (#178) * Replaced all Python code in the CIFAR10 and ResNet examples, removing Python 3 dependency. * Needed to import FoundationNetworking on Linux. * Added a check for FoundationNetworking, added an early exit for cached directory check. * Removed macOS availability check by targeting 10.13 in the package. * Style and formatting fixes. * Removed no-longer-needed _tensorHandles and supporting code. * Replace Autoencoder's tanh activation with sigmoid (#180) * Add GAN Example (#181) * Add GAN Example * Remove do blocks * Replace lrelu with closure * Remove typealiases * Add comment * Remove labels in GAN example * Rename variables * Code formatting, update comments * Fix: latentDim -> latentSize * Add spaces around * * Code formatting, rename loss functions * Fix: generatorLossFunc -> generatorLoss * Break lines to make it fit within 100 columns * Update comments * Rename plot->plotImage, imageGrid->gridImage, Code formatting * Refactor: Label creation * Update comment * Remove type parameter, empty line * Update readme * nightlies URL: s4tf-kokoro-artifact-testing => swift-tensorflow-artifacts (#183) * First steps in repository reorganization: extracting common MNIST dataset code (#182) * Extracted MNIST dataset, created LeNet network, added example combining the two. * Extracted redundant MNIST loading code from GAN and Autoencoder examples, replaced with central MNIST dataset. * Renamed input parameters and applied standard formatting style to MNIST. * Punctuation correction. Co-Authored-By: Richard Wei <[email protected]> * README formatting update. Co-Authored-By: Richard Wei <[email protected]> * Renamed trainImages -> trainingImages, corrected Python package names, formatted Package.swift. * Update Dockerfile to install Python libraries (#184) * Delete helpers.swift (#187) This file is no longer needed apple/swift #26023 and may be causing the segfault reported in swift-apis #186. * Continuing repository reorganization: extracting common CIFAR-10 and ResNet code (#185) * Extracted CIFAR-10 dataset and ResNet models into respective modules. * Minor formatting fixes. * Mirroring PR #187. * [Models] Fix enumeration of blocks in `ResidualBasicBlockStack` `init` (#192) * [Models] Fix enumeration of blocks in `ResidualBasicBlockStack` initialization * Start range at * Update ranges * SqueezeNet Implementation (#189) * Add files via upload * Implemented requested changes Corrected formatting mistakes, used dropout correctly, removed the hard coded number of classes and improved overall readability. * Update SqeezeNet.swift Renamed numClasses to classCount and limited all lines to a size of 100. * WideResNet - fix widenFactor and match model to citation (#193) * add identity connections to WideResNet * rename preact1 for clarity * remove extra relu, add dropout * fix declarartion * skip dropout in expansion blocks * remove enum, res layers to one line * Removal of deprecated allDifferentiableVariables. (#194) * Add JPEG loading / saving via Raw TF operations, removing Matplotlib dependencies (#188) * Added an Image struct as a wrapper for JPEG loading / saving, removed matplotlib dependency from GAN and Autoencoder examples using this. * Formatting update for Autoencoder. * Bring this inline with current API. * Made saveImage() a throwing function, improved formatting. * Changed function parameter. * Convert MNIST classifier to use sequential (#200) * Convert MNIST classifier to use sequential * Review Changes * Remove ImageClassification Models * Convert Autoencoder and Catch to use Sequential (#203) Partially fixes #202. * Adding shape-based inference tests for all image classification models (#198) * Implemented inference tests with random tensors and starting weights for all classification models. * Made sure tests ran on Linux, reshaped output of SqueezeNet to match other classification models. * WideResNet is expressed in terms of CIFAR10, so altered the inputs and outputs appropriately. * Wrapping the reshaping line in SqueezeNet. * Reset ownership. * Reworked TensorShape initializers to use array literals. * Minor formatting tweak to SqueezeNet. * Update deprecated APIs in Catch (#205) [swift-apis/#497](tensorflow/swift-apis#497) removes deprecated apis, now that we are moving onward to v0.5. This PR updates the catch file which uses a deprecated dense layer. * Update squeezenet and add V1.1 version (#204) * Update squeezenet and add V1.1 version * value error * Update tests * review changes * Fix '@noDerivative' warnings. (#208) * Indentation and formatting fixes in CartPole (#214) * DenseNet Implementaion (#213) * Create DenseNet.swift Implementation for the DenseNet Architecture * Added test for DenseNet * Fixed Typo * Refractored the code * Updated ConvPair description Co-Authored-By: Richard Wei <[email protected]> * Fixed comment Co-Authored-By: Richard Wei <[email protected]> * Fixed typos, and improved code quality * Removed extra whitespace and '\' * Used 3 '/' for documentation and removed return * Renamed DenseNet to DenseNet121 * Renamed tests * Updated names * Renamed all DenseNet identifiers to DenseNet121 * Download MNIST dataset from remote URL (#215) This PR is intended to fix the second item listed in [this issue](#206): > it currently relies on hardcoded paths, which will break if you take the dataset outside of that project and try to use it in something like Colab. In this PR, the datasets are instead downloaded from a remote URL, so they are not fragile with respect to the current working directory where the MNIST dataset is instantiated from. The implementation in this PR brings the MINST dataset in line with how the CFAR10 dataset is loaded. * use enable in _vjpCausallyMasked (#218) Fixes #217 * Change momentum and epsilon properties to scalars. (#220) * Uses switch instead of ifs (#219) * Fixes deprecation warning in MiniGo (#221) * Fixing deprecation warnings on all but the MiniGo example. (#223) * Initial training and inference benchmarks (#222) * Initial implementation of training and inference benchmarks. * Resolved a typechecker issue. * Added a Readme. * Added copyright headers. * Reworked the benchmarking functions to use a single function that is specialized via enums. * Reworked the model-specific benchmark into general inference and training cases that can work on arbitrary image classification models. * Restructured BenchmarkResults to be a plain collection of timings, with specific benchmark processing working on that. * Rename interpretTimings to interpretedTimings. Co-Authored-By: Dan Zheng <[email protected]> * Change reference to interpretedTimings later.

Add GAN Example

b16005c

rxwei requested a review from BradLarson July 25, 2019 02:02

BradLarson reviewed Jul 25, 2019

View reviewed changes

GAN/main.swift Outdated Show resolved Hide resolved

Remove do blocks

19ba68d

rxwei requested review from dan-zheng and rxwei July 25, 2019 03:09

BradLarson reviewed Jul 25, 2019

View reviewed changes

GAN/main.swift Outdated Show resolved Hide resolved

BradLarson reviewed Jul 25, 2019

View reviewed changes

GAN/main.swift Outdated Show resolved Hide resolved

t-ae added 4 commits July 26, 2019 09:45

Replace lrelu with closure

e6edc38

Remove typealiases

084510e

Add comment

c70ad3d

Remove labels in GAN example

e52b0ec

rxwei reviewed Jul 26, 2019

View reviewed changes

t-ae added 8 commits July 26, 2019 10:40

Rename variables

813342d

Code formatting, update comments

c00416f

Fix: latentDim -> latentSize

799a3eb

Add spaces around *

6097719

Code formatting, rename loss functions

b218e51

Fix: generatorLossFunc -> generatorLoss

44a1904

Break lines to make it fit within 100 columns

39e997a

Update comments

9a79072

rxwei reviewed Jul 26, 2019

View reviewed changes

t-ae added 3 commits July 26, 2019 13:02

Rename plot->plotImage, imageGrid->gridImage, Code formatting

83bdb72

Refactor: Label creation

d1ec2bb

Update comment

4aac95f

rxwei reviewed Jul 26, 2019

View reviewed changes

Remove type parameter, empty line

e4d3540

Update readme

dff62cb

rxwei approved these changes Jul 26, 2019

View reviewed changes

BradLarson merged commit d6f4496 into tensorflow:master Jul 26, 2019

t-ae deleted the gan-example branch July 26, 2019 14:34

		plotTestImage(testImage, name: "epoch-\(epoch)-output")

		let lossG = generatorLossFunc(fakeLogits: testImage)

	imageGrid = imageGrid.reshaped(to: [(imageHeight+2)*testImageGridSize,
	imageGrid = imageGrid.reshaped(to: [(imageHeight + 2) * testImageGridSize,

	(imageWidth+2)*testImageGridSize])
	(imageWidth + 2) * testImageGridSize])

	var dense1 = Dense<Float>(inputSize: imageSize, outputSize: 256, activation: { leakyRelu($0) })
	var dense1 = Dense<Float>(inputSize: imageSize, outputSize: 256,
	activation: { leakyRelu($0) })

	func plot(image: Tensor<Float>, name: String) {
	func plotImage(_ image: Tensor<Float>, name: String) {

	Tensor<Float>(randomNormal: [size, latentSize])
	Tensor(randomNormal: [size, latentSize])

Add GAN Example #181

Add GAN Example #181

Uh oh!

Conversation

t-ae commented Jul 25, 2019

Uh oh!

Uh oh!

t-ae commented Jul 25, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

t-ae commented Jul 26, 2019

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

BradLarson commented Jul 26, 2019

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

BradLarson commented Jul 26, 2019

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

t-ae commented Jul 25, 2019 •

edited

Loading