the_code/Human/data/chrombpnet/chrombpnet_utils/augment.py

import numpy as np

# https://stackoverflow.com/questions/46091111/python-slice-array-at-different-position-on-every-row
def take_per_row(A, indx, num_elem):
    """
    Matrix A, indx is a vector for each row which specifies 
    slice beginning for that row. Each has width num_elem.
    """

    all_indx = indx[:,None] + np.arange(num_elem)
    return A[np.arange(all_indx.shape[0])[:,None], all_indx]


def random_crop(seqs, labels, seq_crop_width, label_crop_width, coords):
    """
    Takes sequences and corresponding counts labels. They should have the same
    #examples. The widths would correspond to inputlen and outputlen respectively,
    and any additional flanking width for jittering which should be the same
    for seqs and labels. Each example is cropped starting at a random offset. 

    seq_crop_width - label_crop_width should be equal to seqs width - labels width,
    essentially implying they should have the same flanking width.
    """

    assert(seqs.shape[1]>=seq_crop_width)
    assert(labels.shape[1]>=label_crop_width)
    assert(seqs.shape[1] - seq_crop_width == labels.shape[1] - label_crop_width)

    max_start = seqs.shape[1] - seq_crop_width # This should be the same for both input and output

    starts = np.random.choice(range(max_start+1), size=seqs.shape[0], replace=True)

    new_coords = coords.copy()
    new_coords[:,1] = new_coords[:,1].astype(int) - (seqs.shape[1]//2) + starts

    return take_per_row(seqs, starts, seq_crop_width), take_per_row(labels, starts, label_crop_width), new_coords

def random_rev_comp(seqs, labels, coords, frac=0.5):
    """
    Data augmentation: applies reverse complement randomly to a fraction of 
    sequences and labels.

    Assumes seqs are arranged in ACGT. Then ::-1 gives TGCA which is revcomp.

    NOTE: Performs in-place modification.
    """
    
    pos_to_rc = np.random.choice(range(seqs.shape[0]), 
            size=int(seqs.shape[0]*frac),
            replace=False)

    seqs[pos_to_rc] = seqs[pos_to_rc, ::-1, ::-1]
    labels[pos_to_rc] = labels[pos_to_rc, ::-1]
    coords[pos_to_rc,2] =  "r"
	
    return seqs, labels, coords

def crop_revcomp_augment(seqs, labels, coords, seq_crop_width, label_crop_width, add_revcomp, rc_frac=0.5, shuffle=False):
    """
    seqs: B x IL x 4
    labels: B x OL

    Applies random crop to seqs and labels and reverse complements rc_frac. 
    """

    assert(seqs.shape[0]==labels.shape[0])

    # this does not modify seqs and labels
    #mod_seqs, mod_labels, mod_coords = random_crop(seqs, labels, seq_crop_width, label_crop_width, coords)
    mod_seqs, mod_labels, mod_coords = seqs, labels, coords

    # this modifies mod_seqs, mod_labels in-place
    if add_revcomp:
        mod_seqs, mod_labels, mod_coords = random_rev_comp(mod_seqs, mod_labels, mod_coords, frac=rc_frac)


    if shuffle:
        perm = np.random.permutation(mod_seqs.shape[0])
        mod_seqs = mod_seqs[perm]
        mod_labels = mod_labels[perm]
        mod_coords = mod_coords[perm]


    return mod_seqs, mod_labels, mod_coords