Dockerized tensorflow benchmark tool with GPU support

This Docker image is based on the latest tensorflow/tensorflow image with python and gpu support. The tensorflow/benchmarks repository is cloned and used as an entrypoint for the container. This allows some image classification models to be executed within the container with GPUs by passing the corresponding arguments to the docker run command.

For further informations please refer to official TensorFlow Guide.

Run benchmarks

To run ResNet50 with synthetic data and a single GPU use:

docker run --runtime=nvidia --rm cemizm/tf-benchmark-gpu --model resnet50 --num_gpus=1

Frequently used flags:

• model to use for benchmarks. Examples: alexnet, resnet50, resnet152, inception3, vgg16. default: trivial
• num_gpus number of gpus to use. default: all available gpus
• variable_update method for managing variables: parameter_server, replicated, distributed_replicated, independent. default: parameter_server
• batch_size for each GPU. default: 32

For a list of all available flags use:

docker run --runtime=nvidia --rm cemizm/tf-benchmark-gpu --help

Results

Official TensorFlow results for P100 and K80 can be found in the TensorFlow Guide linked above. Here are some results from consumer hardware for workstation environments.

RTX 2070

Setting	Value
TensorFlow	1.14
Dataset	imagenet (synthetic)
Mode	training
SingleSess	False
Num batches	100
Num epochs	0.00
NUMA bind	False
Data format	NCHW
Optimizer	sgd
Variables	parameter_server

GPUs	InceptionV3	ResNet-50	ResNet-152	AlexNet	VGG16
1	122	196	76	2371	112
2	209	364	133	4408	173
4
8