Add README

Update README

Signed-off-by: Yihong Wang <[email protected]>

.npmignore

@@ -0,0 +1,7 @@
+node_modules/
+.npmignore
+images/
+src/
+.vscode/
+scripts
+package-lock.json

README.md

@@ -0,0 +1,57 @@
+# node-red-contrib-tf-model
+This is a Node-RED custom node which is used to load tensorflow models and
+perform inference. Currenctly, it only supports Web friendly JSON format model
+which mainly used by Tensorflow.js. SavedModel format will be added soon.
+
+## Installation
+
+### Prerequisite
+For Tensorflow.js on Node.js
+([@tensorflow/tfjs-node](https://www.npmjs.com/package/@tensorflow/tfjs-node)
+or
+[@tensorflow/tfjs-node-gpu](https://www.npmjs.com/package/@tensorflow/tfjs-node-gpu)),
+it depends on Tensorflow shared libraries. Putting Tensorflow.js as the
+dependency of custom Node-RED node may run into a situation that multiple
+custom nodes install multiple `tfjs-node` module as their dependencies. While
+loading multiple Tensorflow shared libraries in the same process, the process
+would abord by hitting protobuf assertion.
+
+Therefore, this module put `@tensorflow/tfjs-node` as peer dependency. You need
+to install it with the Node-RED manully.
+
+Install `@tensorflow/tfjs-node`:
+```
+npm install @tensorflow/tfjs-node
+```
+
+
+### Install this module:
+Once you install the peer dependency, you can install this module:
+```
+npm install node-red-contrib-tf-model
+```
+
+## Usage
+
+You can see the `tf-model` node in the `Models` category, like this:
+
+![Palette](images/palette.png "Palette")
+
+Then you can use `tf-model` node in your flow. It only needs one property:
+`Model URL`.
+
+![Config Node](images/Config_Node.png "Config Node")
+
+The `Model URL` should point to a Tensorflow.js model which is in web friendly
+format. Typically, it should be a model JSON file. After you specify the
+`Model URL` and deploy the flow, it will fetch the model files, including
+shard files, and store them in `${HOME}/.node-red/tf-model` directory.
+The new node will load the model and maintain the cache entry.
+
+## Data Format
+
+When performing the inference of a Tensorflow.js model, you need to pass the
+corresponding `msg.payload` to `tf-model` node. The `msg.payload` would be a
+`tf.NamedTensorMap` object containing all the needed features in `tf.Tensor`
+data type for the model. Then the results are passed to the next node in
+`msg.payload`. It could be a `tf.Tensor` or `tf.Tensor[]`.

package.json

@@ -1,7 +1,7 @@
 {
   "name": "node-red-contrib-tf-model",
-  "version": "0.1.0",
-  "description": "a tf.js model node to retrieve/execute a tf model",
+  "version": "0.1.1",
+  "description": "a custom Node-RED node to retrieve/execute a Tensorflow model",
   "main": "dist/index.js",
   "scripts": {
     "build": "tsc && npm run copy",

src/index.html

@@ -27,26 +27,31 @@
 </script>
 
 <script type="text/x-red" data-help-name="tf-model">
-    <p>Tensorflow.js Model</p>
+    <p>Use this to load Tensorflow.js Model and perform inference. The inference results
+        will be passed to next node.
+    </p>
 
     <h3>Inputs</h3>
     <dl class="message-properties">
-        <dt>tensors
-            <span class="property-type">Object</span>
+        <dt>payload
+            <span class="property-type">tf.NamedTensorMap</span>
         </dt>
-        <dd>use the tensors as inputs to run model inference</dd>
+        <dd>Directly use the payload as inputs to run model inference.
+            tf.Tensors in the named map would be disposed after inference.
+        </dd>
     </dl>
 
     <h3>Outputs</h3>
-    <ol class="node-ports">
-        <dt>prediction
-            <span class="property-type">Object</span>
+    <dl class="message-properties">
+        <dt>payload
+            <span class="property-type">tf.Tensor | tf.Tensor[]</span>
         </dt>
-        <dd>The prediction</dd>
-    </ol>
+        <dd>The ouput tensors of the model</dd>
+    </dl>
 
     <h3>Details</h3>
-    <p><code>msg.payload</code> is used as tensor data and create tf.Tensor objects
-    then use these data to run model inference. The prediction results will be converted
-    to data array and pass the next node.
+    <p><code>msg.payload</code> is used as tf.NamedTensorMap and run model inference.
+        Then the prediction results will be passed to next node. It could be
+        tf.Tensor or tf.Tensor[]
+    </p>
 </script>