MNIST training model for csquare

This example is meant to be deployed on csquare.

Maintainers

• Sicong Zhuang sicong@csquare.ai, Data Scientist & Deep Learning Engineer at csquare
• Mathieu Bour mathieu@csquare.ai, Lead Platform Engineer at csquare

Training

The train.py is the training itself. Refer to its content to see the possible options and flags.

Usage: python train.py [--options (see table below)]

Option	Description	Default value
--epochs	Number of epochs to train	10
--batch-size	Input batch size for training	64
--test-batch-size	Input batch size for testing	1000
--train-dir	Training dataset directory	/datasets/mnist
--val-dir	Validation dataset directory	/datasets/mnist
--checkpoint-format	Checkpoint filename format	checkpoint-{epoch}.pth.tar
--checkpoints-dir	Checkpoint directory location	.
--lr	Learning rate	0.01
--momentum	SGD momentum	0.5
--gradient-predivide-factor	Apply gradient predivide factor in optimizer	1.0
--no-cuda	Disables CUDA training	False
--seed	Random seed	42
--log-interval	How many batches to wait before logging training status	10
--fp16-allreduce	Use fp16 compression during allreduce	False
--use-adasum	Use adasum algorithm to do reduction	False

Multi-node training

This model support multi-node training via Horovod (https://github.com/horovod/horovod). See Horovod documentation for more details.

Deploy this model on csquare

Configure the model

We are writing this section, it will be vailable very soon!

Docker image

This model is optimized to use Horovod's prebuilt Docker image for Pytorch. In particular, you can use:

• With GPUs: horovod/horovod:0.20.0-tf2.3.0-torch1.6.0-mxnet1.6.0.post0-py3.7-cuda10.1
• Without GPUs (CPU-only): horovod/horovod:0.20.0-tf2.3.0-torch1.6.0-mxnet1.5.0-py3.7-cpu