// the find
tobegit3hub/miniflow
Minimal numerical computation library with TensorFlow APIs
MiniFlow reimplements the TensorFlow 1.x graph-and-Session API as a small Python library with C/SWIG operation backends, published on PyPI as miniflow. It is aimed at people who want to read how a TF-style graph executor and autodiff fit together, or who need the TF API on a small device such as a Raspberry Pi.
- The core fits in a few files (graph.py, session.py, optimizer.py, ops.py), so you can trace a full forward and backward pass in an afternoon. That makes it a usable teaching artifact for reverse-mode autodiff.
- The examples directory includes matching MiniFlow and TensorFlow versions of the same scripts (linear regression, feed_dict, apply_gradient), so you can diff the two implementations line by line.
- The benchmark directory has runnable scripts for add, multiple-operation and linear regression against TensorFlow, so the performance claims come with code you can rerun instead of only a chart.
- The swig/ directory shows a working pattern for exposing C operations to Python through SWIG, which is a clearer example than most of the older tutorials on the same topic.
- The last push was January 2019. The API it mirrors (tf.Session, placeholders, feed_dict) belongs to TensorFlow 1.x, so anyone on TF 2 or PyTorch gets little from the compatibility layer.
- Build artifacts are committed alongside sources: swig/example.o, op.o, example_wrap.o and op_wrap.o sit next to the .i and .c files. These will not match a current Python or platform, and they make the repo harder to read and build cleanly.
- 'Compatible with TensorFlow' is not defined by a coverage list. The README gives no list of supported ops or session features, so you find the gaps at runtime.
- No GPU path is documented, and the C layer appears to cover a few example operations rather than a general backend. Treat the performance numbers as CPU-only and specific to the benchmarked ops.