To train the images on Google Cloud AutoML, we should prepare the specific csv files follow this format.

label_to_csv.py can convert the txt or xml label files to csv file. The labels files should strictly follow to below structure.

• Images To train the object detection tasks, all the images should upload to the cloud storage and access it by its name. All the images should stay in the same buckets in cloud storage. Also, different classes should have their own folder as below.

  <bucket_name> (on the cloud storage)
  | -- class1
  |    | -- class1_01.jpg
  |    | -- class1_02.jpg
  |    | ...
  | -- class2
  |    | -- class2_01.jpg
  |    | -- class2_02.jpg
  |    | ...
  | ...
  

  Note, URI of the class1_01.jpg is gs://<bucket_name>/class1/class1_01.jpg
• Labels There are four types of training data - TRAINING, VALIDATION, TEST and UNASSIGNED. To assign different categories, we should create four directories. Inside each folder, users should create the class folders with the same name in cloud storage (see below structure).

  labels (on PC)
  | -- TRAINING
  |    | -- class1
  |    |    | -- class1_01.txt (or .xml)
  |    |    | ...
  |    | -- class2
  |    |    | -- class2_01.txt (or .xml)
  |    |    | ...
  |    | ...
  | -- VALIDATION
  |    | -- class1
  |    |    | -- class1_02.txt (or .xml)
  |    |    | ...
  |    | -- class2
  |    |    | -- class2_02.txt (or .xml)
  |    |    | ...
  |    | ...
  | -- TEST
  |    | (same as TRAINING and VALIDATION)
  | -- UNASSIGNED
  |    | (same as TRAINING and VALIDATION)
  

To see the argument of label_to_csv.py,

python label_to_csv.py -h

usage: label_to_csv.py [-h] -p PREFIX -l LOCATION -m MODE [-o OUTPUT]
                       [-c CLASSES]

optional arguments:
  -h, --help            show this help message and exit
  -p PREFIX, --prefix PREFIX
                        Bucket of the cloud storage path
  -l LOCATION, --location LOCATION
                        Parent directory of the label files
  -m MODE, --mode MODE  'xml' for converting from xml and 'txt' for converting
                        from txt
  -o OUTPUT, --output OUTPUT
                        Output name of csv file
  -c CLASSES, --classes CLASSES
                        Label classes path

For example, if mine bucket name is test, the location of the label directory is /User/test/labels, the mode I choose from is txt, the output name and the class path is same as default.

python label_to_csv.py \
-p test\
-l /User/test/labels \
-m txt

The output file is res.csv by default. Afterwards, upload the csv file to the cloud storage and you can start training!