deepatlas/train/generators.py

import monai
import torch
import itk
import numpy as np
import matplotlib.pyplot as plt
import random
import glob
import os.path
import argparse
import sys


def create_batch_generator(dataloader_subdivided, weights=None):
    """
    Create a batch generator that samples data pairs with various segmentation availabilities.

    Arguments:
        dataloader_subdivided : a mapping from the labels in seg_availabilities to dataloaders
        weights : a list of probabilities, one for each label in seg_availabilities;
                  if not provided then we weight by the number of data items of each type,
                  effectively sampling uniformly over the union of the datasets

    Returns: batch_generator
        A function that accepts a number of batches to sample and that returns a generator.
        The generator will weighted-randomly pick one of the seg_availabilities and
        yield the next batch from the corresponding dataloader.
    """
    seg_availabilities = ['00', '01', '10', '11']
    if weights is None:
        weights = np.array([len(dataloader_subdivided[s]) for s in seg_availabilities])
    weights = np.array(weights)
    weights = weights / weights.sum()
    dataloader_subdivided_as_iterators = {s: iter(d) for s, d in dataloader_subdivided.items()}

    def batch_generator(num_batches_to_sample):
        for _ in range(num_batches_to_sample):
            seg_availability = np.random.choice(seg_availabilities, p=weights)
            try:
                yield next(dataloader_subdivided_as_iterators[seg_availability])
            except StopIteration:  # If dataloader runs out, restart it
                dataloader_subdivided_as_iterators[seg_availability] =\
                    iter(dataloader_subdivided[seg_availability])
                yield next(dataloader_subdivided_as_iterators[seg_availability])
    return batch_generator