Add function layer1d, layer2d, layer3d. Users can use these these functional layer to build any kind of layer they like.

• layer1d
• layer2d
• layer3d

Now the camera responsive position looks a little bit weird in some cases, for example, 1 layer, 2 layers or 3 layers. Optimize camera position strategy and fix this bug.

Now the animation time for layer animation is 2 seconds, this animation time may be too short or too long in some cases. Let user can configuration animation time in model configuration or in each layer configuration.

The default opacity is 0.3 as Constant, support configure for opacity in layer and model.

Dense aggregation may look weird in some cases, for example, connect to input1d. Fix it to let it be more responsible.

Add onComplete callback for loader. User can trigger its own function when load model complete.

To implement a function which can auto-fitting the given number of feature map centers.

It should follow the strategy like:

1. sqrt(), if it's a perfect squre. e.g. if 1024, return 32x32;
2. rectangle which as close as a square. e.g. if 24, return 4x6 instead of 3x8;
3. squre/rectangle with less empty slot. e.g. if 47, return 7x7 with 2 empty slots;

More network examples build with Tensorspace, add interesting examples to "examples" folder:

• Resnet 50
• Dense net
• Acgan generative model
• pix2pix - Picture style transfer
• Yolo v3-tiny
• Mobilenet v1
• Inceptionv3
• SqueezeNet
• ...

Usage

Inception V3 model
YoloV3 model

Process

• Add
• Concatenation

Create tutorial documentation with proper structure
including:

• 1. General introduction (Readme.md)
• 2. Specific example (tf.keras) on how to preprocess models
  • 1. general introduction on model preprocessing;
  • 2. building/loading model (use LeNet as an example);
  • 3. encapsulate multiple intermediate outputs model;
  • 4. convert to tfjs compatable model;

Other examples:

• TensorFlow (General)
• Keras
• TensorFlow.js

Use "overview" property to configure for text's position and angle.

Add yolo output layer, with selectable channel to get output data from callback function.

When user doing predict in tensorspace, tensorspace would pass predict result to user.

User can configure for Stat.js. By this configuration, tensorspace can be with less dependence.

User can configure for tensorspace background color through model configuration.

Add callback function for hover on feature map or aggregation elements.

From the loaded tfjs model (json+weights), we'd better to find the way to catch the layer types.

Now we can catch other configurations, such as output_shape, filter, stride etc.

If we can find the layer type/class, we can keep the process on the auto-generation from a target tfjs model for applying tsp API.

in html, add:

<meta name="author" content="">
<meta name="author" content="">

in javascript, add:

/**
 * @author
 * @author
 */

The close button sometimes looks too small or too big, optimize close button size strategy.

Hover on layer sometimes show error log, may be caused by last layer's status.

User can config merge layer color from model or merge function.

Add the feature which can change the config of the close button.

Include but not limit on enable or not, displayable or not etc.

Load the model from Google Drive . Figure out how to make the input and output work properly

To be done before 09/14/2018

Version 0.0 UI design draft.

• Page structure
• Landing page UI (Ongoing)
• Document page UI

To be done before 09/12/2018

Account and npm, release tensorspace 0.0.1

If there is too many units in dense layer, the dense layer will be too long, we need to figure out a strategy to optimize it.

Class Array reference

Refactor code style to follow three.js code style.

• Source code
• demos

It's too difficult to check detail if the layer is too far from the center (i.e. the layer is close to the two ends).

A proper way to move the camera along the model, or at least provide a way to visualize the layers close to both ends is necessary.

• Preprocess ResNet50 model
• Apply TensorSpace API to preprocessed model

Implement merge function for Tensorspace, support all merge function in keras including:

• Add3d
• Substract3d
• Multiply3d
• Average3d
• Maximum3d
• Concatenate3d
• Dot3d
• Add2d
• Substract2d
• Multiply2d
• Average2d
• Maximum2d
• Concatenate2d
• Dot2d
• Add1d
• Substract1d
• Multiply1d
• Average1d
• Maximum1d
• Concatenate1d
• Dot1d

• Train DCGAN
• Get the multiple output layer
• Build the model using TensorSpace
• Predict the model and Check the input and output

Automatically read layer configuration from tfjs model, create and add tensorspace layers for model.

In this way, users only need to create Sequential model, do not need to create and configure the layers.

Reference to the API configuration
Keras API reference

Sometimes exception is more useful than error log for user to debug.

exchange button img background color and icon color, also make all the imgs border space same.

Choose default color for layers, and documentation to explain it.

Add example for yolo v3 tiny network.

We want to show the process while we are loading a large model.

tensorspace API is going to support all keras layers, the layers is going to implement including:

• conv2dTranspose
• Cropping1d
• Cropping2d
• Activation1d
• Activation2d
• Activation3d
• depthwiseConv2d

Technology Selection for canvas drawer.

The implementation of layer route do not need hook system any more, remove it from source code.

Try to change the single linear structure of the Dense layer to other compatible or other visual-friendly shape.

It is used to avoid the case like when we have a dense layer 4096, it is difficult to see the model compare with the length of dense layer. Camera and other auto-fitting feature may be affected.

Clear up source code, add comments for source code to make source code more readable.