pytdml

pytdml is a pure python parser and encoder for training datasets based on OGC Training Data Markup Language for AI standard.

Installation

The package can be installed via pip.

Requirements

• Python 3 and above

Dependencies

Dependencies are listed in requirements.txt. Dependencies are automatically installed during pytdml's installation.

Installing the Package

pip install pytdml

Usage

Encoding

From the command line

The training dataset can be encoded to TrainingDML-AI JSON format by YAML configuration file with command line.

pytdml/yaml_to_tdml.py --config=<YAML configuration file path> --output=<Output TrainingDML-AI JSON file path>

YAML configuration file schema is described in encoding YAML configuration file schema.

Using the API from python

The training dataset can also be encoded to TrainingDML-AI JSON format with Python API.

from pytdml.io import write_to_json
from pytdml.type import EOTrainingDataset, EOTrainingData, EOTask, EODataSource, SceneLabel

# generate EO training dataset
dataset = EOTrainingDataset(
    id='...',
    name='...',
    description='...',
    data=[
        EOTrainingData(
            id='...',
            labels=[
                SceneLabel(
                    label_class='...',
                    data_url='...',
                    date_time='...'),
                ...
            ]),
        ...
    ],
    version="...",
    amount_of_training_data=...,
    created_time="...",
    updated_time="...",
    providers=["..."],
    keywords=["...", "..."],
    tasks=[EOTask(task_type="...",
                  description="...")],
    data_sources=[EODataSource(
        id="...",
        data_type="...",
        resolution="..."
    )],
    classes=["...", "...", "..."],
    number_of_classes=...,
    bands=["...", "...", "..."],
    image_size="..."
)
# write to json
write_to_json(dataset, "dataset.json")

Encoding training data from S3

# get training data from s3

s3_client = pytdml.datalibrary.S3Client('s3', "your_server", "your_ak", "your_sk")

td_list = []
bucket_name = "my-bucket"
obj_list = s3_client.list_objects(Bucket=bucket_name, Prefix="whu_rs19/")
for obj in obj_list:
    td = EOTrainingData(
        id=obj.split(".")[0],
        labels=[SceneLabel(label_class=obj.split("/")[1])],
        data_url=f"s3://{bucket_name}/{quote(obj)}",
        date_time="2010"
    )
    td_list.append(td)

Parsing

The training dataset described with TrainingDML-AI JSON file can be parsed with python API and transformed to PyTorch/TensorFlow dataset.

Read TrainingDataset object from JSON file

import pytdml

training_dataset = pytdml.io.read_from_json("dataset.json")  # read from TDML json file
print("Load training dataset: " + training_dataset.name)
print("Number of training samples: " + str(training_dataset.amount_of_training_data))
print("Number of classes: " + str(training_dataset.number_of_classes))

Read training data from s3

import pytdml

# Initialize S3client 
s3_client = pytdml.io.S3_reader.S3Client("s3", "your_server", "your_akey", "your_skey")
# Load the training dataset
training_dataset = pytdml.io.read_from_json("dataset.json")  # read from TDML json file
for item in training_dataset.data:
    path = item.data_url
    if pytdml.io.S3_reader.pasrse_s3_path(path):
        bucket_name, key_name = pytdml.io.S3_reader.parse_s3_path(path)
        object_data = s3_client.get_object(bucket_name, key_name)
        # Process the S3 object data (read as PIL Image)
        with PIL.Image.open(BytesIO(object_data)) as img:
            # processing....
    else:
        print("Invalid S3 path:", path)

Transform to PyTorch dataset

• Scene classification dataset

import pytdml
from torchvision import transforms

# Load the training dataset
training_dataset = pytdml.io.read_from_json("dataset.json")  # read from TDML json file

# Transform the training dataset
class_map = pytdml.ml.creat_class_map(training_dataset)  # create class map
train_dataset = pytdml.ml.TorchEOImageSceneTD(  # create Torch train dataset
    training_dataset.data,
    class_map,
    transform=transforms.Compose(  # transform for the training set
        [transforms.RandomResizedCrop(size=156, scale=(0.8, 1.0)),  # random resize
         transforms.RandomRotation(degrees=15),  # random rotate
         transforms.RandomHorizontalFlip(),  # random flip
         transforms.CenterCrop(size=124),  # center crop
         transforms.ToTensor(),  # transform to tensor
         transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])  # normalize
         ]
    ))

• Object detection dataset

import pytdml

# Load the training dataset
training_dataset = pytdml.io.read_from_json("dataset.json")  # read from TDML json file

# Transform the training dataset
class_map = pytdml.ml.creat_class_map(training_dataset)  # create class map
train_dataset = pytdml.ml.TorchEOImageObjectTD(  # create Torch train dataset
    training_dataset.data,
    class_map,
    transform=pytdml.ml.BaseTransform([128, 128])
)

• Semantic segmentation dataset

import pytdml
from torchvision import transforms

# Load the training dataset
training_dataset = pytdml.io.read_from_json("dataset.json")  # read from TDML json file

# Transform the training dataset
class_map = pytdml.ml.creat_class_map(training_dataset)  # create class map
train_dataset = pytdml.ml.TorchEOImageSegmentationTD(  # create Torch train dataset
    training_dataset.data,
    class_map,
    transform=transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    ])
)

Transform to TensorFlow dataset

• Scene classification dataset

import pytdml

# Load the training dataset
training_dataset = pytdml.io.read_from_json("dataset.json")  # read from TDML json file

# Transform the training dataset
class_map = pytdml.ml.creat_class_map(training_dataset)  # create class map
train_dataset = pytdml.ml.TensorflowEOImageSceneTD(  # create TensorFlow train dataset
    training_dataset.data,
    class_map
)
tf_train_dataset = train_dataset.create_dataset()

• Object detection dataset

import pytdml

# Load the training dataset
training_dataset = pytdml.io.read_from_json("dataset.json")  # read from TDML json file

# Transform the training dataset
class_map = pytdml.ml.creat_class_map(training_dataset)  # create class map
train_dataset = pytdml.ml.TensorflowEOImageObjectTD(  # create TensorFlow train dataset
    training_dataset.data,
    class_map
)
tf_train_dataset = train_dataset.create_dataset()

• Semantic segmentation dataset

import pytdml

# Load the training dataset
training_dataset = pytdml.io.read_from_json("dataset.json")  # read from TDML json file

# Transform the training dataset
class_map = pytdml.ml.creat_class_map(training_dataset)  # create class map
train_dataset = pytdml.ml.TensorflowEOImageSegmentationTD(  # create TensorFlow train dataset
    training_dataset.data,
    class_map
)
tf_train_dataset = train_dataset.create_dataset()

Image Cropping

The images of training dataset in TrainingDML-AI JSON format can be cropped with command line for preprocessing.

pytdml/tdml_image_crop.py  --input=<Input original TrainingDML-AU file path> --output_json=<Output result TrainingDML-AI JSON file path>
                          --output_images=<Output dir of result cropped images> --size=<Crop size of images>

###　Load datasets to provide training-ready training data

• View available datasets according to different tasks and categories

ds_lib = EOTrainingDatasetCollection()

# Find the EO dataset that contains the category 'Building Area' in the Semantic Segmentation task in the Server.
ds_lib.dataset_list(Task.semantic_segmentation, ["Building Area"]) 

• Load the TrainingDML code for the corresponding dataset

aisd_tds = ds_lib["AISD"]
# Output to obtain metadata information such as the name of the corresponding dataset, the number of training samples, the number of categories included in the dataset, etc.
print("Load training dataset: " + aisd_tds.name)
print("Number of training samples: " + str(aisd_tds.amount_of_training_data))
print("Number of classes: " + str(aisd_tds.number_of_classes))

• Acquire more training datas of this category across datasets

building_extraction_td = ds_lib["buildingExtraction"]
# Combine training datas from two datasets with the category 'Building Area' into one training data collection
my_dataset_td = ds_lib.training_data_collection(Task.semantic_segmentation, [aisd_td, building_extraction_td], ["Building Area"])

• Call the pipeline framework's wrapper class to encapsulate the training datas into a trainable dataset

transform = transforms.Compose(  # transform for the dataset
    [
        transforms.ToTensor(),
        transforms.CenterCrop(224),
        transforms.RandomCrop(224),
        transforms.RandomHorizontalFlip(),  # random flip
        # transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])  # normalize
    ]
)
# transforms for object detection task
import pytdml.ml.object_transforms as transform_target
target_transform = transform_target.Compose([
    transform_target.ToTensor(),
    transform_target.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
    transform_target.RandomResize((512, 512))
])
# Split the dataset by a customized ratio
train_data, val_data, test_data = pytdml.ml.split_data(my_dataset_td.data, 0.7, 0.2, 0.1)
# OR split the dataset by training type of training data 
# train_data = pytdml.ml.split_data(my_dataset_td.data, "train")
# valid_data = pytdml.ml.split_data(my_dataset_td.data, "valid")
path = "." 
# Wrapping dataset classes with pytorch framework
train_set = PipeLine(train_data, path).torch_dataset(download=True, transform=transform)
valid_set = PipeLine(valid_data, path).torch_dataset(download=True, transform=transform)
# dataPipe load
# train_set = PipeLine(train_data, path).torch_dataPipe(transform=transform)

train_dataloader = DataLoader(train_set, batch_size=4, num_workers=4)
val_dataloader = DataLoader(val_set, batch_size=4, num_workers=4)
# Create the model
net = DeepLabV3()
criterion = nn.NLLLoss()
params = add_weight_decay(net, l2_value=0.0001)
optimizer = torch.optim.Adam(params, lr=1e-3)
# Train the network
for e in range(100):
    print("Epoch: " + str(e))
    net = net.train()
    train_loss = 0
    train_acc = 0
    train_acc_cls = 0
    train_mean_iu = 0
    train_fwavacc = 0
    prev_time = datetime.datetime.now()
    for iter_i, data in enumerate(train_dataloader):
        # forward
        # ...
        # backword
        # ...
        with torch.no_grad():
            for data in val_dataloader:
                # ...

Convert other EO dataset formats to TrainingDML-AI format

• convert coco format to TrainingDMl-AI format:

from pytdml.convert_utils import convert_coco_to_tdml,convert_stac_to_tdml

coco_path = "/mnt/example/coco_file.json"
output_path = "convert_result.json"

convert_coco_to_tdml(coco_path, output_path)