def test_saved_model(self):
ssd_resnet50_ckpt_url = 'http://download.tensorflow.org/models/object_detection/ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync_2018_07_03.tar.gz'
dst_path = 'saved_model.tar.gz'
if not os.path.exists(dst_path):
os.system("wget {} -O {}".format(ssd_resnet50_ckpt_url, dst_path))
os.system("tar -xvf {}".format(dst_path))
model = TensorflowModel(
'ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync_2018_07_03/saved_model'
)
from tensorflow.python.framework import graph_util
graph_def = graph_util.convert_variables_to_constants(
sess=model.sess,
input_graph_def=model.sess.graph_def,
output_node_names=model.output_node_names)
model.graph_def = graph_def
tmp_saved_model_path = './tmp_saved_model'
if os.path.exists(tmp_saved_model_path):
os.system('rm -rf {}'.format(tmp_saved_model_path))
os.system('mkdir -p {}'.format(tmp_saved_model_path))
self.assertTrue(isinstance(model.graph_def, tf.compat.v1.GraphDef))
self.assertTrue(isinstance(model.graph, tf.compat.v1.Graph))
model.save(tmp_saved_model_path)
# load again to make sure model can be loaded
model = TensorflowModel(tmp_saved_model_path)
os.system(
'rm -rf ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync_2018_07_03'
)
os.system('rm -rf temp_saved_model')
os.system('rm -rf {}'.format(tmp_saved_model_path))
os.system('rm saved_model.tar.gz')
def test_graph(self):
graph = build_graph()
self.assertRaises(ValueError, TensorflowModel, 'test')
fwk_info = {
'input_tensor_names': ['x'],
'output_tensor_names': ['op_to_store']
}
model = TensorflowModel(graph.as_graph_def(), fwk_info)
self.assertEqual(True,
isinstance(model.graph_def, tf.compat.v1.GraphDef))
self.assertEqual(model.input_node_names[0], 'x')
self.assertEqual(model.output_node_names[0], 'op_to_store')
model.save('model_test.pb')
model.graph_def = 'model_test.pb'
self.assertEqual(model.input_tensor_names[0], 'x')
self.assertEqual(model.output_tensor_names[0], 'op_to_store')
self.assertEqual(model.input_tensor[0].name, 'x:0')
self.assertEqual(model.output_tensor[0].name, 'op_to_store:0')
with self.assertRaises(ValueError):
model.save('fake_path/fake_path')
with self.assertRaises(AssertionError):
model.input_tensor_names = []
with self.assertRaises(AssertionError):
model.input_tensor_names = ['test']
model.input_tensor_names = ['x_1']
with self.assertRaises(AssertionError):
model.output_tensor_names = []
with self.assertRaises(AssertionError):
model.output_tensor_names = ['test']
model.output_tensor_names = ['op_to_store_1']
def test_ckpt(self):
mobilenet_ckpt_url = \
'http://download.tensorflow.org/models/mobilenet_v1_2018_02_22/mobilenet_v1_1.0_224.tgz'
dst_path = '/tmp/.lpot/mobilenet_v1_1.0_224.tgz'
if not os.path.exists(dst_path):
os.system("mkdir -p /tmp/.lpot && wget {} -O {}".format(
mobilenet_ckpt_url, dst_path))
os.system("mkdir -p ckpt && tar xvf {} -C ckpt".format(dst_path))
fwk_info = {
'output_tensor_names': ['MobilenetV1/Predictions/Reshape_1']
}
model = TensorflowModel('./ckpt', fwk_info)
self.assertGreaterEqual(len(model.input_tensor_names), 1)
self.assertEqual(len(model.output_tensor_names), 1)
graph_def = model.graph_def
self.assertEqual(True, isinstance(graph_def, tf.compat.v1.GraphDef))
model.graph_def = graph_def
os.system('rm -rf ckpt')
def get_optimized_model(self):
"""Executed the non-precision dependant graph optimization.
The input graph will be optimized with following passes:
1. Remove the training nodes like Identity Op.
2. Split the shared nodes like weights node for multi-Conv2d.
3. Fold Constant Nodes as less as possible.
4. Fuse the Mul node into the previous Conv2D/MatMul if possible.
5. Strip the useless nodes.
6. Do the Common sequence elimation optimization on the graph.
7. Fold the BN node into the previous Conv2D if possible.
Returns:
[graphdef]: the optimized graphdef object.
"""
self.logger.debug("Start to pre optimize input model...")
origin_model = TensorflowModel(self.model._model,
self.model.framework_specific_info,
**self.model.kwargs)
self._tmp_graph_def = ConvertLayoutOptimizer(
self.model.graph_def, self.output_node_names).do_transformation()
self._tmp_graph_def = GrapplerOptimizer(
self._tmp_graph_def, self.output_node_names, self.optimization).do_transformation()
self._tmp_graph_def = RemoveTrainingNodesOptimizer(
self._tmp_graph_def, protected_nodes=self.output_node_names).do_transformation()
self._tmp_graph_def = SplitSharedInputOptimizer(self._tmp_graph_def).do_transformation()
self._tmp_graph_def = GraphFoldConstantOptimizer(self._tmp_graph_def).do_transformation()
self._tmp_graph_def = FuseColumnWiseMulOptimizer(self._tmp_graph_def).do_transformation()
self._tmp_graph_def = StripUnusedNodesOptimizer(self._tmp_graph_def,
self.input_node_names, self.output_node_names).do_transformation()
self._tmp_graph_def = FuseGeluOptimizer(self._tmp_graph_def).do_transformation()
self._tmp_graph_def = GraphCseOptimizer(self._tmp_graph_def).do_transformation()
self._tmp_graph_def = FoldBatchNormNodesOptimizer(
self._tmp_graph_def).do_transformation()
#TODO we should handle all control ops elegantly not bypass it.
self._tmp_graph_def, excluded_node_names = UpdateEnterOptimizer(
self._tmp_graph_def).do_transformation()
#TODO we need to remove below optimizer once the TF enabled the single
# matmul op quantization
self._tmp_graph_def = InjectDummyBiasAddOptimizer(
self._tmp_graph_def).do_transformation()
self._tmp_graph_def = FuseTransposeReshapeOptimizer(
self._tmp_graph_def).do_transformation()
self._excluded_node_names.extend(excluded_node_names)
self._tmp_graph_def.library.CopyFrom(self.model.graph_def.library)
origin_model.graph_def = self._tmp_graph_def
return origin_model