def __init__(self, model, framework_specific_info={}, **kwargs):
self.framework_specific_info = framework_specific_info
input_tensor_names = deep_get(framework_specific_info,
'input_tensor_names', [])
output_tensor_names = deep_get(framework_specific_info,
'output_tensor_names', [])
self.workspace_path = deep_get(framework_specific_info,
'workspace_path', './')
self.kwargs = copy.deepcopy(kwargs)
kwargs.update({'name': deep_get(framework_specific_info, 'name')})
self._model = model
self.sess, self._input_tensor_names, self._output_tensor_names = \
create_session_with_input_output(
model, input_tensor_names, output_tensor_names, **kwargs)
deep_set(framework_specific_info, 'input_tensor_names',
self._input_tensor_names)
deep_set(framework_specific_info, 'output_tensor_names',
self._output_tensor_names)
# sess_global_initialize(self.sess)
self.iter_op = None
if 'MakeIterator' in [
node.op for node in self.sess.graph.as_graph_def().node
]:
self.iter_op = self.sess.graph.get_operation_by_name(
'MakeIterator')
tf.compat.v1.get_variable_scope().reuse_variables()
self._graph_info = {}
def __init__(self,
input_graph,
output_graph,
inputs=[],
outputs=[],
qt_config={},
int8_sequences={},
fp32_ops=[],
bf16_ops=[],
data_loader=None):
"""Convert graph.
:param input_graph: input graph pb file.
:param output_graph: output graph pb file. If set, output directory should be exist.
:param inputs: input nodes' names.
:param outputs: output nodes' names.
:param qt_config: quantization configs, including interation and op-wise quant config
:param fp32_ops: fall back to fp32 dtype op list
:param bf16_ops: fall back to bf16 dtype op list
:param data_loader: for calibration phase used dataloader
"""
# Logger initial
self.logger = logging.getLogger()
self.debug = bool(self.logger.level == logging.DEBUG)
# as we may have outputs with suffix, strip to get raw name
self.output_node_names = list(set([x.split(":")[0] for x in outputs]))
self.input_node_names = list(set([x.split(":")[0] for x in inputs]))
# For lpot, the input_graph is not graph file path but Graph object.
self.input_graph = get_graph_def(input_graph, self.output_node_names)
if 'MakeIterator' in [node.op for node in self.input_graph.node]:
self.output_node_names.append('MakeIterator')
self.output_graph = output_graph
self.input_tensor_names = inputs
self.output_tensor_names = outputs
# quantize specific config
self.calib_iteration = qt_config['calib_iteration']
self.op_wise_config = qt_config['op_wise_config']
self.advance_config = deep_get(qt_config, 'advance')
self.device = qt_config['device'] if 'device' in qt_config else 'cpu'
self.int8_sequences = int8_sequences
self.fp32_ops = fp32_ops
self.bf16_ops = bf16_ops
self._calibration_data = []
self._fp32_print_data = []
self.data_loader = data_loader
self._check_tf_version()
self._check_args()
self._gen_tmp_filenames()
self._kl_op_dict = {}
self._kl_keys = []
self._print_node_mapping = {}
self._enable_kl_op_names = [
k for k in self.op_wise_config if self.op_wise_config[k][1] == 'kl'
]
self._fp32_origin_graph = copy.deepcopy(self.input_graph)
def __init__(self,
model,
qt_config={},
recipes={},
int8_sequences={},
fp32_ops=[],
bf16_ops=[],
data_loader=None,
fake_quant=False):
"""Convert graph.
:param model: input tensorflow model.
:param qt_config: quantization configs, including interation and op-wise quant config
:param fp32_ops: fall back to fp32 dtype op list
:param bf16_ops: fall back to bf16 dtype op list
:param data_loader: for calibration phase used dataloader
:param fake_quant: for quantization-aware training model conversion to default model
"""
# Logger initial
self.logger = logging.getLogger()
self.debug = bool(self.logger.level == logging.DEBUG)
self.model = model
# quantize specific config
self.calib_iteration = qt_config[
'calib_iteration'] if not fake_quant else 0
self.op_wise_config = qt_config['op_wise_config']
self.advance_config = deep_get(qt_config, 'advance')
self.device = qt_config['device'] if 'device' in qt_config else 'cpu'
self.int8_sequences = int8_sequences
self.fp32_ops = fp32_ops
self.bf16_ops = bf16_ops
self.recipes = recipes
self.fake_quant = fake_quant
self.quantized_node_info = []
self._calibration_data = []
self._fp32_print_data = []
self.data_loader = data_loader
self._check_tf_version()
self._check_args()
self._gen_tmp_filenames()
self._kl_op_dict = {}
self._kl_keys = []
self._print_node_mapping = {}
self._enable_kl_op_names = [
k for k in self.op_wise_config if self.op_wise_config[k][1] == 'kl'
]
self._fp32_model = TensorflowModel(self.model._model,
self.model.framework_specific_info,
**self.model.kwargs)
self._fp32_model.graph_def = self.model.graph_def
self._sampling_model = TensorflowModel(
self.model._model, self.model.framework_specific_info,
**self.model.kwargs)
self._sampling_model.graph_def = self.model.graph_def
self._tmp_graph_def = copy.deepcopy(self.model.graph_def)
def _fuse_requantize_with_fused_quantized_node(self):
if self.fake_quant:
self._tmp_graph_def = FreezeFakeQuantOpOptimizer(
self._tmp_graph_def).do_transformation()
self._tmp_graph_def = FuseConvRequantizeTransformer(
self._tmp_graph_def, self.device).do_transformation()
if not self.fake_quant:
self._tmp_graph_def = FuseMatMulRequantizeTransformer(
self._tmp_graph_def).do_transformation()
self._tmp_graph_def = FuseMatMulRequantizeDequantizeTransformer(
self._tmp_graph_def).do_transformation()
self._tmp_graph_def = StripUnusedNodesOptimizer(
self._tmp_graph_def, self._tmp_model.input_node_names,
self._tmp_model.output_node_names).do_transformation()
self._tmp_graph_def = RemoveTrainingNodesOptimizer(
self._tmp_graph_def,
protected_nodes=self._tmp_model.output_node_names
).do_transformation()
self._tmp_graph_def = FoldBatchNormNodesOptimizer(
self._tmp_graph_def).do_transformation()
if 'scale_propagation_concat' in self.recipes and self.recipes[
'scale_propagation_concat']:
self._tmp_graph_def = RerangeQuantizedConcat(
self._tmp_graph_def, self.device).do_transformation()
self._tmp_graph_def = MetaInfoChangingMemOpOptimizer(
self._tmp_graph_def).do_transformation()
if self.advance_config is not None and \
deep_get(self.advance_config, 'bias_correction') is not None:
self._tmp_graph_def = BiasCorrection(
self._tmp_graph_def, self.model.graph_def).do_transformation()
self._tmp_graph_def.library.CopyFrom(self.model.graph_def.library)
self._tmp_model.graph_def = self._tmp_graph_def
def _fuse_requantize_with_fused_quantized_node(self):
self._tmp_graph_def = FuseConvRequantizeTransformer(
self._tmp_graph_def, self.device).do_transformation()
self._tmp_graph_def = FuseMatMulRequantizeTransformer(
self._tmp_graph_def).do_transformation()
self._tmp_graph_def = FuseMatMulRequantizeDequantizeTransformer(
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 = RemoveTrainingNodesOptimizer(
self._tmp_graph_def,
protected_nodes=self.output_node_names).do_transformation()
self._tmp_graph_def = FoldBatchNormNodesOptimizer(
self._tmp_graph_def).do_transformation()
self._tmp_graph_def = RerangeQuantizedConcat(
self._tmp_graph_def, self.device).do_transformation()
self._tmp_graph_def = PostCseOptimizer(
self._tmp_graph_def).do_transformation()
if self.advance_config is not None and \
deep_get(self.advance_config, 'bias_correction') is not None:
self._tmp_graph_def = BiasCorrection(
self._tmp_graph_def, self.input_graph).do_transformation()
self._tmp_graph_def.library.CopyFrom(self.input_graph.library)
if self.debug:
write_graph(self._tmp_graph_def, self.output_graph)
self.logger.info('Converted graph file is saved to: %s',
self.output_graph)