Gabriele91/DENN-tensorflow

⚠️ This is an unsupported implementation. New version is available at DENN ⚠️

Install

python install -r requirements.txt
python setup.py build_denn
python setup.py install

TensorFlow version

Since version 0.12 is not necessary compile the library from source on MacOS. This is also for linux version but with an additional dependency.

To build the op for the official binary package simply use make, otherwise use the command make OFFICIAL_BINARY=false.

TIPS

Search for symlinks:

# search
objdump -t DENN/DENNOp.so
# filter
objdump -t DENN/DENNOp.so | grep CheckOpMessageBuilder
# demangling
objdump -t DENN/DENNOp.so | grep CheckOpMessageBuilder | c++filt
# Search in whole folder
find -name '*.a' | xargs -t -n 1 objdump -t | grep something

Build notes

• C_FLAGS += -lprotobuf is necessary because of ::tensorflow::protobuf::TextFormat::ParseFromString, otherwise some functions will not be available

Example benchmark test

[
    {
        "name": "iris_dataset",
        "dataset_file": "./datasets/iris_15-B_5s.gz",
        "TOT_GEN": 1000,
        "GEN_STEP": 50,
        "TYPE": "double",
        "F": 0.8,
        "NP": 20,
        "de_types": [
            "rand/1/bin"
        ],
        "CR": 0.5,
        "levels": [
            {
                "shape": [[4,3], [3]],
                "init": [
                    {
                        "fx": {
                            "name": "random_uniform",
                            "kwargs": {
                                "seed": 1,
                                "minval": -1.0, 
                                "maxval": 1.0
                            }
                        }
                    },
                    [0.5, 0.5, 0.5]
                ],
                "preferred_device": "CPU",
                "fx": {
                    "name": "nn.softmax_cross_entropy_with_logits"
                }
            }
        ]
    }
]

Execution

cd scripts
python benchmark.py config/configfile.json

You will find in the graph folder the exported graph. Open it with TensorBoard.