v1.1.0

Highlights

Support Dynamic Computation Graph

Dynamic computation graph like RNN is supported, and some interfaces in descriptor runner are changed.

//v1.0.0:
let runner = await WebDNN.prepareAll(modelPath);
let x = await runner.getInputViews()[0];
let y = await runner.getOutputViews()[0];

x.set(loadImageData());

await runner.run()

print('result:', y);

//v1.1.0:

// "WebDNN.prepareAll()" is removed. Please use "WebDNN.load"
let runner = await WebDNN.load(modelPath);

// You can can get input and output views synchronously
let x = runner.getInputViews()[0];
let y = runner.getOutputViews()[0];

// You can modify dynamic hyper parameter at run-time (ex: length of input time series)
runner.setPlaceholderValue({ T: 8 });

x.set(loadTextData());

await runner.run();

//Because "runner.getInputViews()" and "runner.getOutputViews()" return "SymbolicArrayBufferView",
//To get actual ArrayBufferView, you need to convert them explicitly.
print('result:', y.toActual());

(#305, #306, #307, #320, #338)

Update Documents

Reference document is updated (#342)

Support WebGPU Backend in iOS

From iOS 11, WebGPU is also supported in mobile safari (with experimental flag). WebDNN supports these environment. (#330)

Add Operators

• Zeropad #341
• LSTM #300 #301 #337 #340
• Dilated convolution #326
• Softmax #325
• Sigmoid #324
• Softplus #324
• Sigmoid #324
• Clipped ReLU #324
• Hard Sigmoid #324
• Leaky ReLU #324

Fix Many Bugs

Enormous number of bugs are fixed!