Platform Template Object Detection

This is the template of object-detection task for ABEJA Platform.

Requirements

• Python 3.6.x
• [For local] Install ABEJA Platform CLI

Conditions

• Transfer learning from VGG16 ImageNet.
• Allow only 1 category with multiple labels
• Allow up to two dataset
  • with two dataset, either of them should be aliased with val.

Parameters

env	type	description
BATCH_SIZE	int	Batch size. Default 32.
EPOCHS	int	Epoch number. This template applies "Early stopping". Default 50.
IMG_SIZE	int	Image size. Currently only SSD300 (size=300) is supported!! Automatically resize to this size. Default 300.
SHUFFLE	bool	Shuffle train dataset. Default true.
RANDOM_SEED	int	Random seed. If set, use it for a data shuffling. Default None.
MAX_ITEMS	bool	Number of items to use. Default None which means use all items
TEST_SIZE	float	Ratio of test dataset. Default 0.4
LEARNING_RATE	float	Learning rate. Need to be from 0.0 to 1.0. Default 0.001.
MOMENTUM	float	Momentum factor. Need to be from 0.0 to 1.0. Default 0.0.
WEIGHT_DECAY	float	Weight decay (L2 penalty). Need to be from 0.0 to 1.0. Default 0.0.
DAMPENING	float	Dampening for momentum. Need to be from 0.0. Default 0.0.
NESTEROV	float	Enables Nesterov momentum. Default False.
CONF_THRESHOLD	float	Confidence threshold to filter out bounding boxes with low confidence. Default 0.01.
TOP_K	int	Number of bounding boxes to be taken. Default 200.
NMS_THRESHOLD	float	The threshold for IoU to consider bounding boxes as the same. Default 0.45.
OVERLAP_THRESHOLD	float	The overlap threshold used when matching boxes. Need to be from 0.0 to 1.0. Default 0.5
NEG_POS	int	Hard Negative Mining ratio. Default 3
CONFIDENCE_THRESHOLD	float	Results above this threshold will be returned. This option is valid only for prediction (or inference). Default 0.1

Preparation

Before running this script, please install ABEJA Platform CLI. You need to setup your configuration at first.

Debug on local

You can train on local with debug-local command.

$ abeja training debug-local \
  --handler train:handler \
  --image abeja-inc/all-cpu:19.04 \
  --organization_id xxxxx \ 
  --datasets train:xxxxx

Train on ABEJA Platform

Prepare training.yaml file

name: my-detection-model-from-template
handler: train:handler
image: abeja-inc/all-gpu:19.04
datasets:
  train: "xxx"
  val: "xxx"

Create Job definition

Create job definition on the platform at first.

abeja training create-job-definition

Create Version & Create Job

When you modify your code, run the command to deploy your code on the platform.

abeja training create-version

After deploying it, you can create the job by running the command below.

abeja training create-job -d "my first job"

After creating job, you can check the status by ABEJA Platform Console.

Inference on local

You can do prediction on local with run-local command.

$ abeja model run-local \
  --handler predict:handler \
  --image abeja-inc/all-cpu:19.04 \
  --environment TRAINING_JOB_DATASETS:'{"data":1111111111111}' \
  --environment CONFIDENCE_THRESHOLD:0.9 \
  --input sample.jpg

environment variables for inference

env	type	description
TRAINING_JOB_DATASETS	str	Dataset IDs. e.g '{"data":1111111111111}'. The prediction script fetches the label information from training data.
CONFIDENCE_THRESHOLD	float	Results above this threshold will be returned. This option is valid only for prediction (or inference). Default 0.1

You can see more detail of the parameters on parameters.py.

Also you can start inference API on local with run-local-server command.

$ abeja model run-local-server \
  --handler predict:handler \
  --image abeja-inc/all-cpu:19.04 \
  --environment TRAINING_JOB_DATASETS:'{"data":1111111111111}' \
  --environment CONFIDENCE_THRESHOLD:0.3 \

Prepare test data

You can use this sample script to test it with Pascal VOC Dataset. See the detail of this repository.