Skip to content

Latest commit

 

History

History
272 lines (169 loc) · 5.95 KB

README.rst

File metadata and controls

272 lines (169 loc) · 5.95 KB
Raw

Installation

Create a virtual environment

conda env create -f environment.yml
source activate amanda

Install dependencies

There are two options:

  • Use pip:

    pip install -e ".[dev]"

  • Use poetry:

    poetry install

Run an example

make build_cc
amanda-download tf --model vgg16 --root-dir downloads
python src/amanda/tools/debugging/insert_debug_op_tensorflow.py

Usage

CLI

The usage of amanda:

Usage: amanda [OPTIONS] [TOOL_ARGS]...

Options:
  -i, --import [amanda_proto|amanda_yaml|tensorflow_pbtxt|tensorflow_checkpoint|tensorflow_saved_model|torchscript|onnx_model|onnx_graph|mmdnn]
                                  Type of the imported model.  [required]
  -f, --from PATH                 Path of the imported model.  [required]
  -e, --export [amanda_proto|amanda_yaml|tensorflow_pbtxt|tensorflow_checkpoint|tensorflow_saved_model|torchscript|onnx_model|onnx_graph|mmdnn]
                                  Type of the exported model.  [required]
  -t, --to PATH                   Path of the exported model.  [required]
  -ns, --namespace TEXT           Namespace of the graph instrumented by the
                                  tool.

  -T, --tool TEXT                 Fully qualified name of the tool.
  --help                          Show this message and exit.

E.g. use a tool to insert debugging ops into a TensorFlow graph from a checkpoint:

# download the checkpoint
amanda-download tf --model vgg16 --root-dir downloads
# run the debugging tool
amanda --import tensorflow_checkpoint --from downloads/model/vgg16/imagenet_vgg16.ckpt \
       --export tensorflow_checkpoint --to tmp/modified_model/vgg16/imagenet_vgg16.ckpt \
       --namespace amanda/tensorflow \
       --tool amanda.tools.debugging.insert_debug_op_tensorflow.DebuggingTool
# run the modified model
amanda-run amanda.tools.debugging.insert_debug_op_tensorflow.run_model --model-dir tmp/modified_model/vgg16

The modified model will be saved into tmp/modified_model/vgg16, and the debugging information will be stored into tmp/debug_info/vgg16.

E.g. convert a TensorFlow model to an Amanda graph:

amanda --import tensorflow_checkpoint --from downloads/model/vgg16/imagenet_vgg16.ckpt \
       --export amanda_yaml --to tmp/amanda_graph/vgg16/imagenet_vgg16

The Amanda graph will be saved into tmp/amanda_graph/vgg16.

E.g. convert an Amanda graph to a TensorFlow model:

amanda --import amanda_yaml --from tmp/amanda_graph/vgg16/imagenet_vgg16 \
       --export tensorflow_checkpoint --to tmp/tf_model/vgg16/imagenet_vgg16.ckpt

The TensorFlow model will be saved into tmp/tf_model/vgg16.

Import a model (from TensorFlow/ONNX/...)

E.g. import from a TensorFlow checkpoint:

graph = amanda.tensorflow.import_from_checkpoint(checkpoint_dir)

See amanda/conversion/tensorflow.py for all supported import operations in TensorFlow.

Export a model (to TensorFlow/ONNX/...)

E.g. export to a TensorFlow checkpoint:

amanda.tensorflow.export_to_checkpoint(graph, checkpoint_dir)

See amanda/conversion/tensorflow.py for all supported export operations in TensorFlow.

All supported import/export modules

Framework Module
TensorFlow amanda.tensorflow
PyTorch amanda.pytorch
ONNX amanda.onnx
MMdnn amanda.mmdnn

modify the graph

See amanda/graph.py for all Graph/Op APIs.

Import amanda:

import amanda

Create a new op and its output tensors:

op =  amanda.create_op(
    type="Conv2D",
    attrs={},
    inputs=["input", "filter"],
    outputs=["output"],
)

Update an op’s attribute:

op.attrs["data_format"] = "NHWC"

Create a new graph:

graph = amanda.create_graph(
    ops=[op1, op2],
    edges=[edge],
    attrs={},
)

Add an op to a graph:

graph.add_op(op)

Remove an op from a graph:

graph.remove_op(op)

Add an edge to a graph:

graph.create_edge(op1.output_port("output"), op2.input_port("input"))

Remove an edge from a graph:

graph.remove_edge(edge)

Development

Install git pre-commit hooks

pre-commit install

run tests

amanda-download all --root-dir downloads
make build_cc
KMP_AFFINITY=disabled pytest -n 2

Run quick tests only:

KMP_AFFINITY=disabled pytest -n 2 -m "not slow"

Run a single test:

pytest src/amanda/tests/test_op.py -k "test_new_op"

Show information about installed packages

poetry show --latest
# or
poetry show --outdated

Show dependency tree

poetry show --tree
# or
poetry show --tree pytest

Update dependencies

poetry update

Bump version

bumpversion minor  # major, minor, patch

Measure code coverage

coverage run -m pytest
coverage html