neural_compressor.adaptor.tf_utils.util

Tensorflow Utils Helper functions.

Functions

`version1_lt_version2`(version1, version2)	Check if version1 is less than version2.
`version1_gt_version2`(version1, version2)	Check if version1 is greater than version2.
`version1_eq_version2`(version1, version2)	Check if version1 is equal to version2.
`version1_gte_version2`(version1, version2)	Check if version1 is greater than or equal to version2.
`version1_lte_version2`(version1, version2)	Check if version1 is less than or equal to version2.
`disable_random`([seed])	A Decorator to disable tf random seed.
`read_graph`(in_graph[, in_graph_is_binary])	Reads input graph file as GraphDef.
`write_graph`(out_graph_def, out_graph_file)	Write output graphDef to file.
`is_ckpt_format`(model_path)	Check the model_path format is ckpt or not.
`is_saved_model_format`(model_path)	Check the model_path format is saved_model or not.
`get_estimator_graph`(estimator, input_fn)	Get the graph of the estimator.
`get_tensor_by_name`(graph, name[, try_cnt])	Get the tensor by name.
`iterator_sess_run`(sess, iter_op, feed_dict, output_tensor)	Run the graph that have iterator integrated in the graph.
`collate_tf_preds`(results)	Collate the prediction results.
`get_input_output_node_names`(graph_def)	Get the input node name and output node name of the graph_def.
`fix_ref_type_of_graph_def`(graph_def)	Fix ref type of the graph_def.
`strip_unused_nodes`(graph_def, input_node_names, ...)	Strip unused nodes of the graph_def.
`strip_equivalent_nodes`(graph_def, output_node_names)	Strip nodes with the same input and attr.
`get_graph_def`(model[, outputs, auto_input_output])	Get the model's graph_def.
`get_model_input_shape`(model)	Get the input shape of the input model.
`get_tensor_val_from_graph_node`(...)	Get the tensor value for given node name.
`int8_node_name_reverse`(node)	Reverse int8 node name.
`generate_feed_dict`(input_tensor, inputs)	Generate feed dict helper function.
`get_weight_from_input_tensor`(model, ...)	Extracts weight tensors and their associated nodes from a smooth quant node's input tensor.
`apply_inlining`(func)	Apply an inlining optimization to the function's graph definition.
`construct_function_from_graph_def`(func, graph_def[, ...])	Rebuild function from graph_def.
`parse_saved_model`(model[, freeze, input_tensor_names, ...])	Parse a input saved_model.
`reconstruct_saved_model`(graph_def, func, frozen_func, ...)	Reconstruct a saved_model.

Module Contents

neural_compressor.adaptor.tf_utils.util.version1_lt_version2(version1, version2)[source]: Check if version1 is less than version2.

neural_compressor.adaptor.tf_utils.util.version1_gt_version2(version1, version2)[source]: Check if version1 is greater than version2.

neural_compressor.adaptor.tf_utils.util.version1_eq_version2(version1, version2)[source]: Check if version1 is equal to version2.

neural_compressor.adaptor.tf_utils.util.version1_gte_version2(version1, version2)[source]: Check if version1 is greater than or equal to version2.

neural_compressor.adaptor.tf_utils.util.version1_lte_version2(version1, version2)[source]: Check if version1 is less than or equal to version2.

neural_compressor.adaptor.tf_utils.util.disable_random(seed=1)[source]: A Decorator to disable tf random seed.

neural_compressor.adaptor.tf_utils.util.read_graph(in_graph, in_graph_is_binary=True)[source]

Reads input graph file as GraphDef.

Parameters:

in_graph – input graph file.
in_graph_is_binary – whether input graph is binary, default True.

Returns:

input graphDef.

neural_compressor.adaptor.tf_utils.util.write_graph(out_graph_def, out_graph_file)[source]

Write output graphDef to file.

Parameters:

out_graph_def – output graphDef.
out_graph_file – path to output graph file.

Returns:

None.

neural_compressor.adaptor.tf_utils.util.is_ckpt_format(model_path)[source]

Check the model_path format is ckpt or not.

Parameters:: model_path (string) – the model folder path
Returns:: return the ckpt prefix if the model_path contains ckpt format data else None.
Return type:: string

neural_compressor.adaptor.tf_utils.util.is_saved_model_format(model_path)[source]

Check the model_path format is saved_model or not.

Parameters:: model_path (string) – the model folder path
Returns:: return True if the model_path contains saved_model format else False.
Return type:: bool

neural_compressor.adaptor.tf_utils.util.get_estimator_graph(estimator, input_fn)[source]

Get the graph of the estimator.

Parameters:

estimator – tf estimator model
input_fn – input function

Returns:

graph

neural_compressor.adaptor.tf_utils.util.get_tensor_by_name(graph, name, try_cnt=3)[source]

Get the tensor by name.

Considering the ‘import’ scope when model may be imported more then once, handle naming format like both name:0 and name.

Parameters:

graph (tf.compat.v1.GraphDef) – the model to get name from
name (string) – tensor of tensor_name:0 or tensor_name without suffixes
try_cnt – the times to add ‘import/’ to find tensor

Returns:

tensor got by name.

Return type:

tensor

neural_compressor.adaptor.tf_utils.util.iterator_sess_run(sess, iter_op, feed_dict, output_tensor, iteration=-1, measurer=None)[source]

Run the graph that have iterator integrated in the graph.

Parameters:

sess (tf.compat.v1.Session) – the model sess to run the graph
iter_op (Operator) – the MakeIterator op
feed_dict (dict) – the feeds to initialize a new iterator
output_tensor (list) – the output tensors
iteration (int) – iterations to run, when -1 set, run to end of iterator

Returns:

the results of the predictions

Return type:

preds

neural_compressor.adaptor.tf_utils.util.collate_tf_preds(results)[source]: Collate the prediction results.

neural_compressor.adaptor.tf_utils.util.get_input_output_node_names(graph_def)[source]: Get the input node name and output node name of the graph_def.

neural_compressor.adaptor.tf_utils.util.fix_ref_type_of_graph_def(graph_def)[source]: Fix ref type of the graph_def.

neural_compressor.adaptor.tf_utils.util.strip_unused_nodes(graph_def, input_node_names, output_node_names)[source]

Strip unused nodes of the graph_def.

The strip_unused_nodes pass is from tensorflow/python/tools/strip_unused_lib.py of official tensorflow r1.15 branch

neural_compressor.adaptor.tf_utils.util.strip_equivalent_nodes(graph_def, output_node_names)[source]: Strip nodes with the same input and attr.

neural_compressor.adaptor.tf_utils.util.get_graph_def(model, outputs=[], auto_input_output=False)[source]: Get the model’s graph_def.

neural_compressor.adaptor.tf_utils.util.get_model_input_shape(model)[source]: Get the input shape of the input model.

neural_compressor.adaptor.tf_utils.util.get_tensor_val_from_graph_node(graph_node_name_mapping, node_name)[source]

Get the tensor value for given node name.

Parameters:

graph_node_name_mapping – key: node name, val: node
node_name – query node

Returns:

numpy array

Return type:

tensor_val

neural_compressor.adaptor.tf_utils.util.int8_node_name_reverse(node)[source]: Reverse int8 node name.

neural_compressor.adaptor.tf_utils.util.generate_feed_dict(input_tensor, inputs)[source]: Generate feed dict helper function.

neural_compressor.adaptor.tf_utils.util.get_weight_from_input_tensor(model, input_tensor_names, op_types)[source]

Extracts weight tensors and their associated nodes from a smooth quant node’s input tensor.

Parameters:

model – A TensorFlow model containing a graph_def attribute.
input_tensor_names – A list of input tensor names to search for weight tensors.
op_types – A list of operation types to search for when looking for weight tensors.

Returns:

sq_weight_tensors: A dictionary mapping each input tensor name
to a dict of its associated weight tensors with weight name.
sq_weights_nodes: A dictionary mapping each input tensor name
to a dict of its associated weight nodes with weight name.

Return type:

A tuple of two dictionaries

neural_compressor.adaptor.tf_utils.util.apply_inlining(func)[source]

Apply an inlining optimization to the function’s graph definition.

Parameters:: func – A concrete function get from saved_model.
Returns:: The optimized graph in graph_def format.
Return type:: new_graph_def

neural_compressor.adaptor.tf_utils.util.construct_function_from_graph_def(func, graph_def, frozen_func=None)[source]

Rebuild function from graph_def.

Parameters:

func – The original concrete function get from saved_model.
graph_def – The optimized graph after applying inlining optimization.

Returns:

The reconstructed function.

Return type:

new_func

neural_compressor.adaptor.tf_utils.util.parse_saved_model(model, freeze=False, input_tensor_names=[], output_tensor_names=[])[source]

Parse a input saved_model.

Parameters:: model (string or AutoTrackable object) – The input saved_model.
Returns:: The graph_def parsed from saved_model. _saved_model: TF AutoTrackable object loaded from saved_model. func: The concrete function get from saved_model. frozen_func: The reconstructed function from inlining optimized graph.
Return type:: graph_def

neural_compressor.adaptor.tf_utils.util.reconstruct_saved_model(graph_def, func, frozen_func, trackable, path)[source]

Reconstruct a saved_model.

Parameters:

graph_def – The input graph_def.
func – The concrete function get from the original saved_model.
frozen_func – The reconstructed function from inlining optimized graph.
trackable – TF AutoTrackable object loaded from the original saved_model.
path – The destination path to save the reconstructed saved_model.