intro

This post includes some details on quantize aware training.

contrib/python/quantize/python/quantize_graph.py:create_eval_graph _create_graph get_default_graph / get_concrete_function

• contrib/python/quantize/python/quantize.py:Quantize
• _FindLayersToQuantize
  • graph_matcher algorithm ?
  • a list of complex matching pattern is recorded in the _FindLayersToQuantize
  • find layer_match
• _InsertQuantOp
  • will quantize the following target:
    • layer_match.weight_tensor
    • layer_match.activation_op
    • layer_match.bias_add_op
    • layer_match.bypass_op,
  • will add the following node:
    • quant_ops.py:MovingAvgQuantize
    • quant_ops.py:LastValueQuantize

If it is training, minmax is updated with some math operations:

Reference

Then, the following ops are added based on per-channel setting.

• fake_quant_with_min_max_vars_per_channel
• fake_quant_with_min_max_vars

After the fake quant node is inserted, the graph can be used as a normal tensorflow graph.

Reference

The forward path:

Reference

The backward path:

Reference

side note on how variable is stored

• base on this post
• _ModelVariable -> minmax value
• vars_collections -> default value ops.GraphKeys.GLOBAL_VARIABLES

tf 2.0

python/eager/def_function.py:function -> @tf.function

• make decorator make it possible to insert function into class

ResourceVariable ?

tensorflow graph collection

ops.py::Graph:get_collection(_VARSCOPESTORE_KEY)