[Model Compression / TensorFlow] Support exporting pruned model (#3487)

nni/compression/tensorflow/compressor.py

@@ -87,6 +87,18 @@ def _instrument(self, layer):
 
         return layer
 
+    def _uninstrument(self, layer):
+        # note that ``self._wrappers`` cache is not cleared here,
+        # so the same wrapper objects will be recovered in next ``self._instrument()`` call
+        if isinstance(layer, LayerWrapper):
+            layer._instrumented = False
+            return self._uninstrument(layer.layer)
+        if isinstance(layer, tf.keras.Sequential):
+            return self._uninstrument_sequential(layer)
+        if isinstance(layer, tf.keras.Model):
+            return self._uninstrument_model(layer)
+        return layer
+
     def _instrument_sequential(self, seq):
         layers = list(seq.layers)  # seq.layers is read-only property
         need_rebuild = False
@@ -97,6 +109,16 @@ def _instrument_sequential(self, seq):
                 need_rebuild = True
         return tf.keras.Sequential(layers) if need_rebuild else seq
 
+    def _uninstrument_sequential(self, seq):
+        layers = list(seq.layers)
+        rebuilt = False
+        for i, layer in enumerate(layers):
+            orig_layer = self._uninstrument(layer)
+            if orig_layer is not layer:
+                layers[i] = orig_layer
+                rebuilt = True
+        return tf.keras.Sequential(layers) if rebuilt else seq
+
     def _instrument_model(self, model):
         for key, value in list(model.__dict__.items()):  # avoid "dictionary keys changed during iteration"
             if isinstance(value, tf.keras.layers.Layer):
@@ -109,6 +131,17 @@ def _instrument_model(self, model):
                         value[i] = self._instrument(item)
         return model
 
+    def _uninstrument_model(self, model):
+        for key, value in list(model.__dict__.items()):
+            if isinstance(value, tf.keras.layers.Layer):
+                orig_layer = self._uninstrument(value)
+                if orig_layer is not value:
+                    setattr(model, key, orig_layer)
+            elif isinstance(value, list):
+                for i, item in enumerate(value):
+                    if isinstance(item, tf.keras.layers.Layer):
+                        value[i] = self._uninstrument(item)
+        return model
 
     def _select_config(self, layer):
         # Find the last matching config block for given layer.
@@ -129,6 +162,17 @@ def _select_config(self, layer):
         return last_match
 
 
+class LayerWrapper(tf.keras.Model):
+    """
+    Abstract base class of layer wrappers.
+
+    Concrete layer wrapper classes must inherit this to support ``isinstance`` check.
+    """
+    def __init__(self):
+        super().__init__()
+        self._instrumented = True
+
+
 class Pruner(Compressor):
     """
     Base class for pruning algorithms.
@@ -167,6 +211,43 @@ def compress(self):
         self._update_mask()
         return self.compressed_model
 
+    def export_model(self, model_path, mask_path=None):
+        """
+        Export pruned model and optionally mask tensors.
+
+        Parameters
+        ----------
+        model_path : path-like
+            The path passed to ``Model.save()``.
+            You can use ".h5" extension name to export HDF5 format.
+        mask_path : path-like or None
+            Export masks to the path when set.
+            Because Keras cannot save tensors without a ``Model``,
+            this will create a model, set all masks as its weights, and then save that model.
+            Masks in saved model will be named by corresponding layer name in compressed model.
+
+        Returns
+        -------
+        None
+        """
+        _logger.info('Saving model to %s', model_path)
+        input_shape = self.compressed_model._build_input_shape  # cannot find a public API
+        model = self._uninstrument(self.compressed_model)
+        if input_shape:
+            model.build(input_shape)
+        model.save(model_path)
+        self._instrument(model)
+
+        if mask_path is not None:
+            _logger.info('Saving masks to %s', mask_path)
+            # can't find "save raw weights" API in tensorflow, so build a simple model
+            mask_model = tf.keras.Model()
+            for wrapper in self.wrappers:
+                setattr(mask_model, wrapper.layer.name, wrapper.masks)
+            mask_model.save_weights(mask_path)
+
+        _logger.info('Done')
+
     def calc_masks(self, wrapper, **kwargs):
         """
         Abstract method to be overridden by algorithm. End users should ignore it.
@@ -199,7 +280,7 @@ def _update_mask(self):
                 wrapper.masks = masks
 
 
-class PrunerLayerWrapper(tf.keras.Model):
+class PrunerLayerWrapper(LayerWrapper):
     """
     Instrumented TF layer.
 
@@ -210,8 +291,6 @@ class PrunerLayerWrapper(tf.keras.Model):
 
     Attributes
     ----------
-    layer_info : LayerInfo
-        All static information of the original layer.
     layer : tf.keras.layers.Layer
         The original layer.
     config : JSON object
@@ -233,6 +312,10 @@ def __init__(self, layer, config, pruner):
         _logger.info('Layer detected to compress: %s', self.layer.name)
 
     def call(self, *inputs):
+        self._update_weights()
+        return self.layer(*inputs)
+
+    def _update_weights(self):
         new_weights = []
         for weight in self.layer.weights:
             mask = self.masks.get(weight.name)
@@ -243,7 +326,6 @@ def call(self, *inputs):
         if new_weights and not hasattr(new_weights[0], 'numpy'):
             raise RuntimeError('NNI: Compressed model can only run in eager mode')
         self.layer.set_weights([weight.numpy() for weight in new_weights])
-        return self.layer(*inputs)
 
 
 # TODO: designed to replace `patch_optimizer`

test/ut/sdk/test_compressor_tf.py

@@ -1,6 +1,8 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT license.
 
+from pathlib import Path
+import tempfile
 import unittest
 
 import numpy as np
@@ -27,6 +29,9 @@
 # This tensor is used as input of 10x10 linear layer, the first dimension is batch size
 tensor1x10 = tf.constant([[1.0] * 10])
 
+# This tensor is used as input of CNN models
+image_tensor = tf.zeros([1, 10, 10, 3])
+
 
 @unittest.skipIf(tf.__version__[0] != '2', 'Skip TF 1.x setup')
 class TfCompressorTestCase(unittest.TestCase):
@@ -42,13 +47,37 @@ def _test_layer_detection_on_model(self, model):
         layer_types = sorted(type(wrapper.layer).__name__ for wrapper in pruner.wrappers)
         assert layer_types == ['Conv2D', 'Dense', 'Dense'], layer_types
 
-    def test_level_pruner(self):
+    def test_level_pruner_and_export_correctness(self):
         # prune 90% : 9.0 + 9.1 + ... + 9.9 = 94.5
         model = build_naive_model()
-        pruners['level'](model).compress()
+        pruner = pruners['level'](model)
+        model = pruner.compress()
+
         x = model(tensor1x10)
         assert x.numpy() == 94.5
 
+        temp_dir = Path(tempfile.gettempdir())
+        pruner.export_model(temp_dir / 'model', temp_dir / 'mask')
+
+        # because exporting will uninstrument and re-instrument the model,
+        # we must test the model again
+        x = model(tensor1x10)
+        assert x.numpy() == 94.5
+
+        # load and test exported model
+        exported_model = tf.keras.models.load_model(temp_dir / 'model')
+        x = exported_model(tensor1x10)
+        assert x.numpy() == 94.5
+
+    def test_export_not_crash(self):
+        for model in [CnnModel(), build_sequential_model()]:
+            pruner = pruners['level'](model)
+            model = pruner.compress()
+            # cannot use model.build(image_tensor.shape) here
+            # it fails even without compression
+            # seems TF's bug, not ours
+            model(image_tensor)
+            pruner.export_model(tempfile.TemporaryDirectory().name)
 
 try:
     from tensorflow.keras import Model, Sequential