tf_deeplab.py

import io
import os
import glob
import numpy as np
import shutil
import tarfile
import tempfile
import tensorflow as tf
import uuid

from os.path import join
from PIL import Image
from subprocess import Popen
from tensorflow.core.example.example_pb2 import Example
from typing import (List, Tuple)
from urllib.parse import urlparse

from object_detection.utils import dataset_util
from rastervision.core.box import Box
from rastervision.core.class_map import ClassMap
from rastervision.core.ml_backend import MLBackend
from rastervision.core.scene import Scene
from rastervision.core.training_data import TrainingData
from rastervision.ml_backends.tf_object_detection_api import (
    write_tf_record, terminate_at_exit, TRAIN, VALIDATION)
from rastervision.utils.misc import save_img
from rastervision.utils.files import (get_local_path, upload_if_needed,
                                      make_dir, download_if_needed, sync_dir,
                                      start_sync)


def numpy_to_png(array: np.ndarray) -> str:
    """Get a PNG string from a Numpy array.

    Args:
         array: A Numpy array of shape (w, h, 3) or (w, h), where the
               former is meant to become a three-channel image and the
               latter a one-channel image.  The dtype of the array
               should be uint8.

    Returns:
         str

    """
    im = Image.fromarray(array)
    output = io.BytesIO()
    im.save(output, 'png')
    return output.getvalue()


def png_to_numpy(png: str, dtype=np.uint8) -> np.ndarray:
    """Get a Numpy array from a PNG string.

    Args:
         png: A str containing a PNG-formatted image.

    Returns:
         numpy.ndarray

    """
    incoming = io.BytesIO(png)
    im = Image.open(incoming)
    return np.array(im)


def make_tf_examples(training_data: TrainingData,
                     class_map: ClassMap) -> List[Example]:
    """Take training data and a class map and return a list of TFRecords.

    Args:
         training_data: A rastervision.core.training_data.TrainingData
              object.
         class_map: A rastervision.core.class_map.ClassMap object.

    Returns:
         list(tensorflow.core.example.example_pb2.Example)

    """
    tf_examples = []
    print('Creating TFRecord', end='', flush=True)
    for chip, window, labels in training_data:
        tf_example = create_tf_example(chip, window, labels, class_map)
        tf_examples.append(tf_example)
        print('.', end='', flush=True)
    print()
    return tf_examples


def merge_tf_records(output_path: str, src_records: List[str]) -> None:
    """Merge mutiple TFRecord files into one.

    Args:
         output_path: Where to write the merged TFRecord file.
         src_records: A list of strings giving the location of the
              input TFRecord files.

    Returns:
         None

    """
    with tf.python_io.TFRecordWriter(output_path) as writer:
        print('Merging TFRecords', end='', flush=True)
        for src_record in src_records:
            for string_record in tf.python_io.tf_record_iterator(src_record):
                writer.write(string_record)
            print('.', end='', flush=True)
        print()


def make_debug_images(record_path: str, output_dir: str,
                      p: float = 0.25) -> None:
    """Render a random sample of the TFRecords in a given file as
    human-viewable PNG files.

    Args:
         record_path: Path to the TFRecord file.
         output_dir: Destination directory for the generated PNG files.
         p: The probability of rendering a particular record.

    Return:
         None

    """
    make_dir(output_dir, check_empty=True)

    print('Generating debug chips', end='', flush=True)
    tfrecord_iter = tf.python_io.tf_record_iterator(record_path)
    for ind, example in enumerate(tfrecord_iter):
        example = tf.train.Example.FromString(example)
        im, labels = parse_tf_example(example)
        output_path = join(output_dir, '{}.png'.format(ind))
        inv_labels = (labels == 0)
        im[:, :, 0] = im[:, :, 0] * inv_labels
        im[:, :, 1] = im[:, :, 1] * inv_labels
        im[:, :, 2] = im[:, :, 2] * inv_labels
        if np.random.rand() <= p:
            save_img(im, output_path)
        print('.', end='', flush=True)
    print()


def parse_tf_example(example: Example) -> Tuple[np.ndarray, np.ndarray]:
    """Parse a TensorFlow Example into an image array and a label array.

    Args:
         example: A TensorFlow Example object.

    Return:
         tuple(np.ndarray, np.ndarray)
    """
    ie = 'image/encoded'
    isce = 'image/segmentation/class/encoded'
    image_encoded = \
        example.features.feature[ie].bytes_list.value[0]
    image_segmentation_class_encoded = \
        example.features.feature[isce].bytes_list.value[0]
    im = png_to_numpy(image_encoded)
    labels = png_to_numpy(image_segmentation_class_encoded)
    return im, labels


def create_tf_example(image: np.ndarray,
                      window: Box,
                      labels: np.ndarray,
                      class_map: ClassMap,
                      chip_id: str = '') -> Example:
    """Create a TensorFlow Example from an image, the labels, &c.

    Args:
         image: An nd.array containing the image data.
         window: A Box object containing the bounding box for this example.
         labels: An nd.array containing the label data.
         class_map: A ClassMap object containing mappings between
              numerial and textual labels.
         chip_id: The chip id as a string.

    Returns:
         A Deeplab-compatible TensorFlow Example object containing the
              given data.

    """
    class_keys = set(class_map.get_keys())

    def fn(n):
        return (n if n in class_keys else 0)

    filtered_labels = np.array(np.vectorize(fn)(labels), dtype=np.uint8)

    image_encoded = numpy_to_png(image)
    image_filename = chip_id.encode('utf8')
    image_format = 'png'.encode('utf8')
    image_height, image_width, image_channels = image.shape
    image_segmentation_class_encoded = numpy_to_png(filtered_labels)
    image_segmentation_class_format = 'png'.encode('utf8')

    features = tf.train.Features(
        feature={
            'image/encoded':
            dataset_util.bytes_feature(image_encoded),
            'image/filename':
            dataset_util.bytes_feature(image_filename),
            'image/format':
            dataset_util.bytes_feature(image_format),
            'image/height':
            dataset_util.int64_feature(image_height),
            'image/width':
            dataset_util.int64_feature(image_width),
            'image/channels':
            dataset_util.int64_feature(image_channels),
            'image/segmentation/class/encoded':
            dataset_util.bytes_feature(image_segmentation_class_encoded),
            'image/segmentation/class/format':
            dataset_util.bytes_feature(image_segmentation_class_format),
        })

    return tf.train.Example(features=features)


def get_record_uri(uri: str, split: str) -> str:
    return join(uri, '{}-0.record'.format(split))


def get_latest_checkpoint(train_logdir_local: str) -> str:
    ckpts = glob.glob(join(train_logdir_local, 'model.ckpt-*.meta'))
    times = map(os.path.getmtime, ckpts)
    latest = sorted(zip(times, ckpts))[-1][1]
    return latest[:len(latest) - len('.meta')]


class TFDeeplab(MLBackend):
    """MLBackend-derived type that implements the TensorFlow DeepLab
    backend.

    """

    def __init__(self):
        """Constructor."""
        self.temp_dir_obj = tempfile.TemporaryDirectory()
        self.temp_dir = self.temp_dir_obj.name

    def process_scene_data(self, scene: Scene, data: TrainingData,
                           class_map: ClassMap, options) -> str:
        """Process the given scene and data into a TFRecord file specifically
        associated with that file.

        Args:
             scene: The scene data (labels stores, the raster sources,
                  and so on).
             data: The training data.
             class_map: The mapping from numerical labels to textual
                  labels.
             options: The options given to `make_training_chips` in
                  the config file.

        Returns:
            The local path to the generated file.

        """
        tf_examples = make_tf_examples(data, class_map)

        base_uri = options.output_uri
        split = '{}-{}'.format(scene.id, uuid.uuid4())
        record_path = join(base_uri, '{}.record'.format(split))
        record_path = get_local_path(record_path, self.temp_dir)

        make_dir(record_path, use_dirname=True)
        write_tf_record(tf_examples, record_path)

        return record_path

    def process_sceneset_results(self, training_results: List[str],
                                 validation_results: List[str],
                                 class_map: ClassMap, options) -> None:
        """Merge TFRecord files from individual scenes into two at-large files
        (one for training data and one for validation data).

        Args:
             training_results: A list of paths to TFRecords containing
                  training data.
             valiation_results: A list of paths to TFRecords
                  containing validation data.
             class_map: A mapping from numerical classes to their
                  textual names.
             options: The options given to `make_training_chips` in
                  the config file.

        Returns:
             None
        """
        base_uri = options.output_uri
        training_record_path = get_record_uri(base_uri, TRAIN)
        training_record_path_local = get_local_path(training_record_path,
                                                    self.temp_dir)
        validation_record_path = get_record_uri(base_uri, VALIDATION)
        validation_record_path_local = get_local_path(validation_record_path,
                                                      self.temp_dir)

        make_dir(training_record_path_local, use_dirname=True)
        make_dir(validation_record_path_local, use_dirname=True)  # sic
        merge_tf_records(training_record_path_local, training_results)
        merge_tf_records(validation_record_path_local, validation_results)
        upload_if_needed(training_record_path_local, training_record_path)
        upload_if_needed(validation_record_path_local, validation_record_path)

        if options.debug:
            training_zip_path = join(base_uri, '{}'.format(TRAIN))
            training_zip_path_local = get_local_path(training_zip_path,
                                                     self.temp_dir)
            validation_zip_path = join(base_uri, '{}'.format(VALIDATION))
            validation_zip_path_local = get_local_path(validation_zip_path,
                                                       self.temp_dir)

            with tempfile.TemporaryDirectory() as debug_dir:
                make_debug_images(training_record_path_local, debug_dir)
                shutil.make_archive(training_zip_path_local, 'zip', debug_dir)
            with tempfile.TemporaryDirectory() as debug_dir:
                make_debug_images(validation_record_path_local, debug_dir)
                shutil.make_archive(validation_zip_path_local, 'zip',
                                    debug_dir)
            upload_if_needed('{}.zip'.format(training_zip_path_local),
                             '{}.zip'.format(training_zip_path))
            upload_if_needed('{}.zip'.format(validation_zip_path_local),
                             '{}.zip'.format(validation_zip_path))

    def train(self, class_map: ClassMap, options) -> None:
        """Train a DeepLab model using the options provided in the
        `segmentation_options` section of the workflow config file.

        Args:
             class_map: A mapping between integral and textual classes.
             options: Options provided in the `segmentation_options`
                  section of the workflow configuration file.

        Returns:
             None
        """
        soptions = options.segmentation_options

        train_py = soptions.train_py
        export_model_py = soptions.export_model_py

        # Setup local input and output directories
        train_logdir = options.output_uri
        train_logdir_local = get_local_path(train_logdir, self.temp_dir)
        dataset_dir = options.training_data_uri
        dataset_dir_local = get_local_path(dataset_dir, self.temp_dir)
        make_dir(train_logdir_local)
        make_dir(dataset_dir_local)
        download_if_needed(get_record_uri(dataset_dir, TRAIN), self.temp_dir)

        # Download and untar initial checkpoint.
        tf_initial_checkpoints_uri = soptions.tf_initial_checkpoints_uri
        make_dir(self.temp_dir)
        download_if_needed(tf_initial_checkpoints_uri, self.temp_dir)
        tfic_tarball = get_local_path(tf_initial_checkpoints_uri,
                                      self.temp_dir)
        tfic_dir = os.path.dirname(tfic_tarball)
        with tarfile.open(tfic_tarball, 'r:gz') as tar:
            tar.extractall(tfic_dir)
        tfic_index = glob.glob('{}/*/*.index'.format(tfic_dir))[0]

        # Build array of argments that will be used to run the DeepLab
        # training script.
        args = ['python', train_py]

        args.append('--train_logdir={}'.format(train_logdir_local))
        args.append('--tf_initial_checkpoint={}'.format(tfic_index))
        args.append('--dataset_dir={}'.format(dataset_dir_local))

        steps = soptions.training_number_of_steps
        if steps > 0:
            args.append('--training_number_of_steps={}'.format(steps))

        if len(soptions.train_split) > 0:
            args.append('--train_split={}'.format(soptions.train_split))

        if len(soptions.model_variant) > 0:
            args.append('--model_variant={}'.format(soptions.model_variant))

        for rate in soptions.atrous_rates:
            args.append('--atrous_rates={}'.format(rate))

        args.append('--output_stride={}'.format(soptions.output_stride))
        args.append('--decoder_output_stride={}'.format(
            soptions.decoder_output_stride))

        for size in soptions.train_crop_size:
            args.append('--train_crop_size={}'.format(size))

        args.append('--train_batch_size={}'.format(soptions.train_batch_size))

        if len(soptions.dataset):
            args.append('--dataset="{}"'.format(soptions.dataset))

        args.append('--save_interval_secs={}'.format(soptions.save_interval_secs))
        args.append('--save_summaries_secs={}'.format(soptions.save_summaries_secs))
        args.append('--save_summaries_images={}'.format(soptions.save_summaries_images))

        # Periodically synchronize with remote
        start_sync(
            train_logdir_local,
            train_logdir,
            sync_interval=options.sync_interval)

        # Train
        train_process = Popen(args)
        terminate_at_exit(train_process)
        tensorboard_process = Popen(
            ['tensorboard', '--logdir={}'.format(train_logdir_local)])
        terminate_at_exit(tensorboard_process)
        train_process.wait()
        tensorboard_process.terminate()

        # Build array of arguments that will be used to run the DeepLab
        # export script.
        args = ['python', export_model_py]
        args.append('--checkpoint_path={}'.format(
            get_latest_checkpoint(train_logdir_local)))
        args.append('--export_path={}'.format(
            join(train_logdir_local, 'frozen_inference_graph.pb')))

        # Export
        export_process = Popen(args)
        terminate_at_exit(export_process)
        export_process.wait()

        if urlparse(train_logdir).scheme == 's3':
            sync_dir(train_logdir_local, train_logdir, delete=True)

    def predict(self, chip, options):
        import pdb
        pdb.set_trace()
        return 1