eddyxu

#!/usr/bin/env python3
#
# Generate HD-Vila-100M dataset
#
# https://github.com/microsoft/XPretrain/tree/main/hd-vila-100m


import argparse
import datetime
import logging

eddyxu

SELECT * FROM (
  SELECT 
    *, 
    ML_PREDICT(coco_det, image) as det,
    ML_PREDICT(hardhat, image) as hat
  FROM hardhat
) WHERE array_contains(hat.label, "helmat")
   AND array_contains(hat.label, "head")
   AND array_contains(det.label, "truck")

eddyxu

CREATE MODEL hardhat  -- Load from model registry
FLAVOR pytorch
MODEL_TYPE ssd
USING 's3://bucket/to/hardhat.pth';

CREATE MODEL coco_det -- Load pretrained models.
FLAVOR pytorch 
MODEL_TYPE fasterrcnn;

eddyxu

from rikai.pytorch.data import Dataset
  
def train(
  name: str,
  uri: Dataset,
  model_type: str,
  epochs: int = 150,
  batch_size: int = 4,
  num_workers: int = 4,
  lr: float = 0.02,

eddyxu

import xml.etree.ElementTree as ET
from pyspark.sql import Row
from rikai.types import Box2d, Image

all_images = {p.name: p for p in basedir.glob("**/*.jpg")}
images = []
for split in ["Train", "Test"]:
    for voc_file in basedir.glob(f"{split}/**/*.xml"):
        root = ET.parse(voc_file).getroot()
        annotations = [Row(

eddyxu

SELECT image_id, image, detection.* FROM (
    SELECT image_id, image, explode(ML_PREDICT(ssd, image)) as detection FROM raw_data
) WHERE 100 < area(detection.box) AND area(detection.box) < 10000 
  AND coco_name(detection.label_id) IN ('chair', 'remote', 'cell phone')
  AND detection.score < 0.6

eddyxu

SELECT box_area, count(*) AS cnt FROM (
  SELECT 
   CAST(area(detection.box) / 500 AS int) * 500 AS box_area
  FROM least_margin
  WHERE coco_name(detection.label_ids[0]) IN ('chair', 'person') 
    AND coco_name(detection.label_ids[1]) IN ('chair', 'person')
) GROUP BY box_area ORDER BY box_area

eddyxu

SELECT 
  coco_name(detection.label_id) as label,
  count(detection.label_id) AS cnt 
FROM (
  SELECT 
    explode(ML_PREDICT(ssd, image)) AS detection
  FROM coco
) WHERE coco_name(detection.label_id) IN ('person', 'chair', 'remote', 'cell phone', 'sheep', 'cow')
GROUP BY detection.label_id
ORDER BY cnt DESC

eddyxu

WITH least_margin AS (
  SELECT 
    image_id, image, detection,
    detection.scores[0] - detection.scores[1] as margin FROM (
      SELECT 
        image_id, 
        image, 
        explode(ML_PREDICT(class_scores, image)) AS detection 
      FROM coco
  ) ORDER BY margin LIMIT 1000

eddyxu

SELECT 
    image_id, 
    image, 
    detection, 
    entropy(detection.scores) as entropy 
FROM (
    SELECT 
        image_id, 
        image,
        explode(ML_PREDICT(class_scores, image)) AS detection