Training TensorFlow models from C++ / _AI

Short answer

You generally do not train full TensorFlow/Keras models inside C++.
The supported training stack is Python. The C API (tensorflow/c/c_api.h) is built for loading SavedModels and running sessions (what _TFServe already does).

From _AI, the practical ways are:

Commands

tftrain ways                 # full explanation
tftrain doctor
tftrain profile pipeline|finetune|native|mock
tftrain dry-run on|off
tftrain pipeline fonts|road_signs|text|coco_stuff|firearms|league|all
tftrain finetune <model> [steps] [feat_dim] [label_dim]
tftrain native [csv] [steps] [lr]
tftrain jobs
tftrain selftest

Recommended production path

1. Prepare data + train in Python (tensorflow_training/.../run_*.ps1)
2. Export SavedModel (and optional TFLite)
3. Place under third_party/tf_models/<name> or configured path
4. tf / tfmodel load + serve via TFServe C API

Optional: in-process fine-tune

Export from Python a SavedModel that includes a train signature, e.g. inputs x,y → loss, with optimizer variables inside the graph. Then:

tftrain profile finetune
tftrain dry-run off
tftrain finetune my_trainable_model 100 32 1

This calls _TFModels::FineTuneStep → _TFServe::Run.

Why not “pure C++ TF training”?

• TensorFlow 2 training is Keras / Python-first.
• The old TF1 C++ training API is not a maintained product path for modern models.
• Building graphs, datasets, and distribution strategies in C++ is unsupported friction for almost no gain versus: Python train → C++ serve.

Native C++ alternative

For baselines without TF:

tftrain profile native
tftrain dry-run off
tftrain native 500 0.05          # synthetic linear
tftrain native data/set.csv 300 0.01

CSV: numeric columns, last column = label.

Env

Files

• TFTrain.hpp / TFTrain.cpp — orchestration
• TFModels::FineTuneStep — C API train-signature step
• tensorflow_training/ — existing Python GPU pipelines