NAS-Bench 101 to Tensorflow 2.0 (tf.keras) converter v0.2

The small but powerful tool that converts models from NAS-Bench 101 code paper into real tf.keras models

• Completely standalone (check requirements). Don't need NAS-Bench to be installed!
• Require Tensorflow 2.0 or higher

Can be used for:

• Training
• Latency measurements
• Latency measurements on devices with TFLite Converter
• ...

New!

• Add generate_graphs function!

Installation

Use PyPi

$ pip3 install nasbench_keras --user

or

$ git clone https://github.com/evgps/nasbench_keras.git
$ cd nasbench_keras
$ pip3 install -e .

Getting Started

To test the tool you can:

• Download or generate json with all model graphs for original NAS-Bench 101 cells size of 7 or less: GDrive

or generate:

from nasbench_keras import generate_graphs
generate_graphs(output_file="generated_graphs1.json", max_vertices=7, num_ops=3, max_edges=9, verify_isomorphism=True)

• Create module:

import tensorflow as tf
import json
from nasbench_keras import ModelSpec, build_keras_model, build_module

with open('generated_graphs.json', "rb") as f:
    models = json.load(f)

# Get model by the hash
model = models['0001a2f6c8977346ccd12fa0c435bf42']

# Adjacency matrix and nuberically-coded layer list
matrix, labels = model

# Configure whole network
config = {'available_ops' : ['conv3x3-bn-relu', 'conv1x1-bn-relu', 'maxpool3x3'],
        'stem_filter_size' : 128,
        'data_format' : 'channels_first',
        'num_stacks' : 3,
        'num_modules_per_stack' : 2,
        'num_labels' : 1000}

# Transfer numerically-coded operations to layers (check base_ops.py)
labels = (['input'] + [config['available_ops'][l] for l in labels[1:-1]] + ['output'])

# Module graph
spec = ModelSpec(matrix, labels, data_format='channels_first')

# Create module
inputs = tf.keras.layers.Input((3,224,224), 1)
outputs = build_module(spec=spec, inputs=inputs, channels=128, is_training=True)
module = tf.keras.Model(inputs=inputs, outputs=outputs)
module.summary()

• And stacked model with downsampling between blocks

# Create whole network with same config
features = tf.keras.layers.Input((3,224,224), 1)
net_outputs = build_keras_model(spec, features, labels, config)
net = tf.keras.Model(inputs=features, outputs=net_outputs)
net.summary()