udpated

tech/tensorflow/machine-learning.md

@@ -67,4 +67,6 @@
         * models are json or bin files.  add up all the json/bin files that re pulled down and you have your model size
     * Use memory tab in chrome dev tools 
         * load the model, go to memory tab, click create snapshot => view total ram in the statistics drop down option 
+
+
     

tech/tensorflow/raw-models.md

@@ -0,0 +1,73 @@
+# Raw Models
+
+## Types
+* Layers Model
+    * retain higher level building blocks for ease of use and debugging 
+    * NOT optimized
+* Graph Model 
+    * highly optimized
+    * can combine multiple poerations in one batch
+
+## How
+* [Documentation](https://js.tensorflow.org/api/latest/#Models-Loading)
+* Stored in two file
+    * model.json 
+        * metadata about model type, architecture, config details
+        * declares a "format" key that has either layers model or graph model
+    *  one or more ".bin" files.  Typically have "shard<#>of<#>.bin"
+        * each bin file is 4MB or less
+
+### Loading a Raw Model
+    * tf.loadGraphModel(<path to model>)
+    * tf.loadLayerModel(<path to model>)
+
+
+### Saving a model
+    *  Can always save a model for use offline using the command: model.save(<localhost>)
+
+## Inspect
+* can use ```model.summary()``` command to provide information about the model.  Provides:
+    * model layers
+    * output shapes at each layer
+    * number of params in each layer
+    * total number of params
+    * number of trainable and non-trainable params
+
+
+## Sample Code
+```
+const MODEL_PATH = 'https://storage.googleapis.com/jmstore/TensorFlowJS/EdX/SavedModels/sqftToPropertyPrice/model.json';
+let model = undefined;
+
+async function loadModel() {
+  model = await tf.loadLayersModel(MODEL_PATH);
+  model.summary();
+
+  // Create a batch of 1.
+  const input = tf.tensor2d([[870]]);
+
+  // Create a batch of 3
+  const inputBatch = tf.tensor2d([[500], [1100], [970]]);
+
+  // Actually make the predictions for each batch.
+  const result = model.predict(input);
+  const resultBatch = model.predict(inputBatch);
+
+  // Print results to console.
+  result.print();  // Or use .arraySync() to get results back as array.
+  resultBatch.print(); // Or use .arraySync() to get results back as array.
+
+  input.dispose();
+  inputBatch.dispose();
+  result.dispose();
+  resultBatch.dispose();
+  model.dispose();
+}
+
+loadModel();
+```
+* In this example: 
+    * A total of 6 tensors are created when the code is run.
+    * 2 tensors are created when the model loads. The 'model.dispose()' method disposes of this tensor.
+    * 2 tensors are created, one for each input. The input.dispose() and inputBatch.dispose() methods dispose of these tensors.
+    * 2 tensors are created and returned as outputs. The result.dispose() and 'resultBatch.dispose()' methods dispose of these tensors.