proc_v_sgm.py

###############################################################################
#
#   Implementations of Iterative Procrustes & Seeded Graph Matching
# 
#   Written by Kelly Marchisio, 2020-2021.
# 
###############################################################################

import argparse
import numpy as np
import cupy as cp
import random
from utils import csls 
from utils import matops
from scipy import sparse
from scipy.linalg import orthogonal_procrustes
from third_party import combine_bidirectional_alignments
from third_party.vecmap import embeddings
import sgm
import sys


def process_dict_pairs(pair_file):
    '''Parses a dictionary pairs file.
        
            Pairs as list of (srcwd, trgwd) tuples
            L1 and L2 vocabularies as sets.
    '''
    pairs = []
    l1_words = set()
    l2_words = set()
    with open(pair_file) as f:
        for line in f:
            w1, w2 = line.split()
            w1 = w1.strip()
            w2 = w2.strip()
            pairs.append((w1, w2))
            l1_words.add(w1)
            l2_words.add(w2)
    return pairs, l1_words, l2_words


def pairs_to_embpos(pairs, src_word2ind, trg_word2ind):
    '''Translates a list of (srcwd, trgwd) tuples to a list of 
        (src_pos, trg_pos) from the embedding space.

        Args: 
            pairs: list of (srcwd, trgwd) tuples
            src_word2ind: source word to index dictionary.
            trg_word2ind: target word to index dictionary.
        
        Returns:
            List of (src_pos, trg_pos) from the embedding space.

    '''
    return list(map(lambda x: (src_word2ind[x[0]], trg_word2ind[x[1]]), pairs))


def load_embs_and_wordpairs(args):
    # Loads embeddings and input word pairs (vocabulary). Returns word pairs,
    # source and target embeddings and word2ind dictionaries. 
    print('Loading embeddings and dictionary pairs...')
    word_pairs, src_words, trg_words = process_dict_pairs(args.pairs)
    
    available_word_pairs = []
    # Read once to find available src/trg words from emb spaces.
    with open(args.src_embs) as src_embs_file:
        with open(args.trg_embs) as trg_embs_file: 
            available_src_words, _ = matops.read(
                    src_embs_file, args.max_embs, src_words)
            available_trg_words, _ = matops.read(
                    trg_embs_file, args.max_embs, trg_words)
            # Keep only pairs where src AND trg words are in emb spaces.
            available_word_pairs = [i for i in word_pairs if i[0] in
                    available_src_words and i[1] in available_trg_words]
    oov_word_pairs = set(word_pairs) - set(available_word_pairs)
    src_words_to_use, trg_words_to_use = unzip_pairs(available_word_pairs)


    with open(args.src_embs) as src_embs_file:
        with open(args.trg_embs) as trg_embs_file: 
            # Re-read embs with only vocab where both sides of pair are present.
            src_words_in_emb_order, src_embs = matops.read(
                    src_embs_file, args.max_embs, src_words_to_use)
            trg_words_in_emb_order, trg_embs = matops.read(
                    trg_embs_file, args.max_embs, trg_words_to_use)

            # These lines for *_word2ind copied from vecmap/.py 
            src_word2ind = {word: i for i, word in enumerate(src_words_in_emb_order)}
            trg_word2ind = {word: i for i, word in enumerate(trg_words_in_emb_order)}
            src_ind2word = {i: word for i, word in enumerate(src_words_in_emb_order)}
            trg_ind2word = {i: word for i, word in enumerate(trg_words_in_emb_order)}

            print('Done loading embeddings and dictionary pairs.')
            return (available_word_pairs, src_embs, src_word2ind, src_ind2word, 
                    trg_embs, trg_word2ind, trg_ind2word, oov_word_pairs)


def create_train_dev_split(pairs, n_seeds, src_word2ind, trg_word2ind,
        rand=False):
    # Separates input pairs into a train/dev split.
    # Returns tuples of ([src/trg]_train_inds, [src/trg]_dev_inds)
    # If rand=True, randomize seeds picked. Otherwise, choose in order.
    print('Creating the train/dev split given input seeds...')
    if rand:
        random.shuffle(pairs)
    train_pairs = pairs[:n_seeds]
    dev_pairs = pairs[n_seeds:]
    train_inds = pairs_to_embpos(train_pairs, src_word2ind, trg_word2ind)
    dev_inds = pairs_to_embpos(dev_pairs, src_word2ind, trg_word2ind)
    print('Done creating the train/dev split given input seeds.')
    return (train_pairs, dev_pairs), (train_inds, dev_inds)


def calculate_csls_scores(x, y, topk=-1):
    # Returns a sparse matrix of topk CSLS scores.
    csls_generator = csls.calculate_csls_scores(x, y, topk=topk)
    return sparse.vstack(csls_generator) 


def solve_procrustes(x, y):
    x = cp.asarray(x)
    y = cp.asarray(y)

    u, s, vt = cp.linalg.svd(y.T @ x)
    w = vt.T @ u.T
    return cp.asnumpy(w)


def eval(hypotheses, test_set):
    # Given hypotheses and a test set, returns the matches and percentage
    # matched.
    print('\tLength of hypotheses: {0}\n\tLength of test set: '
            '{1}'.format(len(hypotheses), len(test_set)), flush=True)
    matches = set(test_set).intersection(hypotheses)
    if hypotheses:
        precision = round((float(len(matches))/len(hypotheses)*100), 4)
    else:
        precision = None
    recall = round((float(len(matches))/len(test_set)*100), 4)
    return matches, precision, recall


def iterative_procrustes_w_csls(x, y, input_x_seed_inds=[], input_y_seed_inds=[], 
        gold_x_seed_inds=[], gold_y_seed_inds=[], val_set=None,
        max_seeds_to_add=-1, curr_i=1, total_i=10, diff_seeds_for_rev=False,
        k=1, active_learning=False, truth_for_active_learning=None):
    '''Run Iterative Procrustes.
        
        Dictionaries are induced as the nearest neighbor of each word in x 
        according to CSLS. The seed set for subsequent rounds is the 
        intersection of the dictionaries induced from both directions.

        Args:
            x: source embedding space. 
            y: target embedding space.
            input_x_seed_inds: seed indicies to use in x space.
            input_y_seed_inds: seed indicies to use in y space.
            gold_x_seed_inds: gold seed indicies in x space.
            gold_y_seed_inds: gold seed indicies in y space.
            val_set: validation set as set of (x1, y1) tuples.
            curr_i: current iteration number.
            total_i: total number of iterations that will run.
            k: how many hypotheses to return.
            active_learning: If True, only hypotheses that are correct (either
                in the train or dev set) used as seeds for next iteration.
            truth_for_active_learning: True pairs to be be compared with for 
                active learning. If a hypothesis is in this set, use it. 

        Returns:
            Hypothesized matches induced in fwd direction for all rows in x.
            Hypothesized matches induced in rev direction for all rows in y.
            Intersection of the above hypothesized matches.
            (all are returned as sets of (x_position, y_position) tuples)
    '''

    print('----------------------------------')
    print('\nRound {0} of Iterative Procrustes'.format(curr_i))
    print('\tNum input seeds:', len(input_x_seed_inds))
    print('\tNum gold seeds:', len(gold_x_seed_inds))

    x_seed_inds, y_seed_inds = unzip_pairs(get_seeds(input_x_seed_inds, input_y_seed_inds,
            gold_x_seed_inds, gold_y_seed_inds, max_seeds_to_add, curr_i, 
            total_i, True))
    print('\tNum combined input seeds:', len(x_seed_inds))
    
    w = solve_procrustes(x[x_seed_inds], y[y_seed_inds])
    csls_scores_sparse = calculate_csls_scores(x @ w, y, topk=k)
    x_hyp_pos, y_hyp_pos, val = sparse.find(csls_scores_sparse)

    if diff_seeds_for_rev:
        print('Getting different seeds for reverse direction.')
        x_seed_inds, y_seed_inds = unzip_pairs(get_seeds(input_x_seed_inds, input_y_seed_inds,
                gold_x_seed_inds, gold_y_seed_inds, max_seeds_to_add, curr_i, 
                total_i, True))
    w_rev = solve_procrustes(y[y_seed_inds], x[x_seed_inds])
    csls_scores_sparse_rev = calculate_csls_scores(y @ w_rev, x, topk=k)
    y_hyp_pos_rev, x_hyp_pos_rev, val = sparse.find(csls_scores_sparse_rev)

    hyps = set(zip(x_hyp_pos, y_hyp_pos))
    hyps_rev = set(zip(x_hyp_pos_rev, y_hyp_pos_rev))
    hyps_int = symmetrize(hyps, hyps_rev, 'intersection')
    
    if val_set:
        eval_symm(val_set, hyps, hyps_rev, hyps_int)

    if curr_i == total_i:
        return hyps, hyps_rev, hyps_int, w, w_rev

    curr_i += 1
    if active_learning:
        correct_hyps = set(truth_for_active_learning).intersection(hyps)
        correct_hyps_rev = set(truth_for_active_learning).intersection(hyps_rev)
        joint_x_hyp_pos, joint_y_hyp_pos = unzip_pairs(
                correct_hyps.union(correct_hyps_rev))
    else:
        joint_x_hyp_pos, joint_y_hyp_pos = unzip_pairs(hyps_int) 
    return iterative_procrustes_w_csls(x, y, joint_x_hyp_pos, joint_y_hyp_pos,
            gold_x_seed_inds, gold_y_seed_inds, val_set, max_seeds_to_add, 
            curr_i, total_i, diff_seeds_for_rev, k, active_learning,
            truth_for_active_learning) 


def eval_symm(val_set, hyps, hyps_rev, hyps_int, hyps_gdf=None):
    '''Evaluates forward, reverse, and joint hypotheses given a validation set.

        Args:
            hyps: forward hypotheses.
            hyps_rev: reverse hypotheses.
            hyps_int: joint hypotheses (intersection of fwd & rev).
            hyps_gdf: fwd & rev hypotheses symmetrized with grow-diag-final. 
            val_set: validation set as set of (x1, y1) tuples.

        Returns:
            Prints precision & recall for all sets of hypotheses.
            Returns (True matches, precision, recall) tuple for each set.
    '''
    print('\nRunning Evaluation....') 
    dev_src_inds, dev_trg_inds = unzip_pairs(val_set)

    print('\nForward:')
    dev_hyps = set(hyp for hyp in hyps if hyp[0] in dev_src_inds)
    matches, prec, recall = eval(dev_hyps, val_set)
    print('\tPairs matched: {0} \n\t(Precision; {1}%) (Recall: {2}%)'
            .format(len(matches), prec, recall), flush=True)

    print('\nReverse:')
    dev_hyps_rev = set(hyp for hyp in hyps_rev if hyp[1] in dev_trg_inds)
    matches_rev, prec_rev, recall_rev = eval(dev_hyps_rev, val_set)
    print('\tPairs matched: {0} \n\t(Precision; {1}%) (Recall: {2}%)'
            .format(len(matches_rev), prec_rev, recall_rev), flush=True)

    print('\nIntersection:')
    dev_hyps_int = dev_hyps.intersection(dev_hyps_rev)
    matches_int, prec_int, recall_int = eval(dev_hyps_int, val_set)
    print('\tPairs matched: {0} \n\t(Precision; {1}%) (Recall: {2}%)'
            .format(len(matches_int), prec_int, recall_int), 
            flush=True)

    return ((matches, prec, recall), (matches_rev, prec_rev, recall_rev), 
            (matches_int, prec_int, recall_int))


def unzip_pairs(pairs):
    '''Unzips a set of (x, y) pairs to lists of [x1, ..., xn], [y1, ..., yn]'''
    x_list = list(list(zip(*pairs))[0])
    y_list = list(list(zip(*pairs))[1])
    return x_list, y_list


def iterative_softsgm(x_sim, y_sim, input_x_seed_inds=[], input_y_seed_inds=[],
        gold_x_seed_inds=[], gold_y_seed_inds=[],
        softsgm_iters=10, k=1, minprob=0.0, val_set=None, max_seeds_to_add=-1,
        curr_i=1, total_i=10, diff_seeds_for_rev=False, run_reverse=False,
        active_learning=False, truth_for_active_learning=None):
    '''Iteratively runs the SoftSGM (Algorithm 3) from Fishkind et al. (2019),
        feeding in intersection of hypotheses from both directions as seeds for
        the next round. Internally, runs SoftSGM with iters iterations and
        returns top k hypotheses over minprob. Note that setting softsgm_iters=0
        total_i=1 runs vanilla SGM.

        Args:
            x_sim: normalized embeddings as a distance (similarity) matrix.
            y_sim: normalized embeddings as a distance (similarity) matrix.
            input_x_seed_inds: indices for seeds for x_sim.
            input_y_seed_inds: indices for seeds for y_sim.
            gold_x_seed_inds: gold seed indicies in x space.
            gold_y_seed_inds: gold seed indicies in y space.
            softsgm_iters: how many iterations of softSGM to run and average.
            k: how many hypotheses to take for each source word from prob. dist
                returned from sgm.softsgm.
            minprob: min probability necessary for hypothesis to be considered.
            max_seeds_to_add: max # of seeds to add per round. Can be an 
                integer if same for all rounds or list if different. -1 == all. 
            val_set: validation set as set of (x1, y1) tuples.
            curr_i: current iteration number.
            total_i: total number of iterations that will run.
            run_reverse: If total_i = 1, still runs the reverse direction.
            active_learning: If True, only hypotheses that are correct (either
                in the train or dev set) used as seeds for next iteration.
            truth_for_active_learning: True pairs to be be compared with for 
                active learning. If a hypothesis is in this set, use it. 

        Returns:
            Hypothesized matches induced in fwd direction for all rows in x.
            Hypothesized matches induced in rev direction for all rows in y.
            Intersection of the above hypothesized matches.
            (all are returned as sets of (x_position, y_position) tuples)
    '''
    print('----------------------------------')
    print('\nRound {0} of Iterative SoftSGM\n'.format(curr_i))
    print('----------------------------------')
    x_seed_inds, y_seed_inds = unzip_pairs(get_seeds(input_x_seed_inds, input_y_seed_inds,
            gold_x_seed_inds, gold_y_seed_inds, max_seeds_to_add, curr_i, 
            total_i, True))

    print('Running SoftSGM Forward', flush=True)
    hyps = run_softsgm_topk(x_sim, y_sim, x_seed_inds, y_seed_inds,
            softsgm_iters, k, minprob)
    if total_i > 1 or run_reverse:
        print('Running SoftSGM Reverse', flush=True)
        if diff_seeds_for_rev:
            print('Getting different seeds for reverse direction.')
            x_seed_inds, y_seed_inds = unzip_pairs(get_seeds(input_x_seed_inds,
                input_y_seed_inds, gold_x_seed_inds, gold_y_seed_inds, 
                max_seeds_to_add, curr_i, total_i, True))
        hyps_rev = run_softsgm_topk(y_sim, x_sim, y_seed_inds, x_seed_inds,
                softsgm_iters, k, minprob)
        hyps_rev = {(i[1], i[0]) for i in hyps_rev}
        hyps_int = symmetrize(hyps, hyps_rev, 'intersection')

        if val_set:
            eval_symm(val_set, hyps, hyps_rev, hyps_int)
    else:
        hyps_rev = None
        hyps_int = None
        if val_set:
            print('Evalling Forward')
            dev_src_inds, _ = unzip_pairs(val_set)
            dev_hyps = [hyp for hyp in hyps if hyp[0] in dev_src_inds]
            matches, precision, recall = eval(dev_hyps, val_set)
            print('\tPrecision: {0}%  Recall {1}%'.format(precision, recall))

    if curr_i == total_i:
        return hyps, hyps_rev, hyps_int 

    curr_i += 1
    if active_learning:
        correct_hyps = set(truth_for_active_learning).intersection(hyps)
        correct_hyps_rev = set(truth_for_active_learning).intersection(hyps_rev)
        joint_x_hyp_pos, joint_y_hyp_pos = unzip_pairs(
                correct_hyps.union(correct_hyps_rev))
    else:
        joint_x_hyp_pos, joint_y_hyp_pos = unzip_pairs(hyps_int) 
    return iterative_softsgm(x_sim, y_sim, joint_x_hyp_pos, joint_y_hyp_pos,
            gold_x_seed_inds, gold_y_seed_inds, softsgm_iters, k, minprob, val_set, 
            max_seeds_to_add, curr_i, total_i, diff_seeds_for_rev, run_reverse,
            active_learning, truth_for_active_learning) 


def symmetrize(hyps, hyps_rev, heuristic):
    if heuristic == 'grow-diag-final':
        return combine_bidirectional_alignments.grow_diag_final(hyps, hyps_rev)
    elif heuristic == 'intersection':
        return hyps.intersection(hyps_rev)


def get_seeds(x_seed_inds, y_seed_inds, gold_x_seed_inds, gold_y_seed_inds, 
        max_seeds_to_add, i, total_i, always_use_gold=True):
    '''Get correct number of seeds for a given round.

        Args:
            x_seed_inds: indices for seeds for x space.
            y_seed_inds: indices for seeds for y space.
            gold_x_seed_inds: gold seed indicies in x space.
            gold_y_seed_inds: gold seed indicies in y space.
            max_seeds_to_add: max # of seeds to add per round. Can be an 
                integer if same for all rounds or list if different. -1 == all. 
            i: index to pull number of seeds from.
            total_i: total number of rounds algorithm will go through.
            always_use_gold: whether or not to always use gold seeds. 

        Returns:
            x_seed_inds: indices for seeds for x space.
            y_seed_inds: indices for seeds for y space.
    '''
    print('Choosing Seeds...')
    xy_pairs = list(zip(x_seed_inds, y_seed_inds))
    # Extra make sure these are shuffled, in case python does any weird caching.
    random.shuffle(xy_pairs)
    # Filter out potential x,y pairs where either x or y is in the gold seed
    # indices (we know this cannot be correct, and do not want duplicate
    # src/trg words in our seed set). 
    xy_nongold_pairs = list(filter(
        lambda pair: pair[0] not in set(gold_x_seed_inds) and
                     pair[1] not in set(gold_y_seed_inds), 
        xy_pairs))

    gold_pairs = set(zip(gold_x_seed_inds, gold_y_seed_inds))
    pairs = list(gold_pairs) + list(xy_nongold_pairs)
    print('\t# of seeds available:', len(pairs))

    if not always_use_gold:
        random.shuffle(pairs)
    if max_seeds_to_add is None:
        num_seeds_to_add = None # Return all seeds.
    elif isinstance(max_seeds_to_add, int) and max_seeds_to_add >= 0:
        num_seeds_to_add = max_seeds_to_add
    elif isinstance(max_seeds_to_add, int) and max_seeds_to_add < 0:
        num_seeds_to_add = None # Return all seeds. 
    else:
        if len(max_seeds_to_add) == 1:
            num_seeds_to_add = max_seeds_to_add[0]
        else:
            assert len(max_seeds_to_add) == total_i # must specify for every round.
            num_seeds_to_add = max_seeds_to_add[i - 1] # zero-indexed.
        if num_seeds_to_add < 0:
            num_seeds_to_add = None
    pairs = pairs[:num_seeds_to_add]
    print('\t# of seeds chosen for round {0}:'.format(i),
            len(pairs))
    return pairs 


def run_softsgm_topk(x_sim, y_sim, x_seed_inds=[], y_seed_inds=[], iters=1, 
        k=1, minprob=0.0, val_set=None):
    '''Runs SoftSGM and returns topk hyps over minprob per source word in x_sim. 
        
        Args:
            x_sim: normalized embeddings as a distance (similarity) matrix.
            y_sim: normalized embeddings as a distance (similarity) matrix.
            x_seed_inds: indices for seeds for x_sim.
            y_seed_inds: indices for seeds for y_sim.
            iters: how many iterations of softSGM to run and average.
            k: how many hypotheses to take for each source word from prob. dist
                returned from sgm.softsgm.
            minprob: min probability necessary for hypothesis to be considered.
            val_set: validation set as set of (x1, y1) tuples.
        Returns:
            topk hypotheses over a minimum probability per source word in x, as
                list of (src, trg) tuples. 
    '''
    print('Running SoftSGM Topk.')
    hyp_probdist, all_hyps = sgm.softsgm(x_sim, y_sim, x_seed_inds, 
            y_seed_inds, iters)
    hyps = get_topk_hypotheses_from_probdist(hyp_probdist, k, minprob)
    if val_set:
        dev_src_inds, _ = unzip_pairs(val_set)
        dev_hyps = [hyp for hyp in hyps if hyp[0] in dev_src_inds]
        matches, precision, recall = eval(dev_hyps, val_set)
        print('\tPrecision: {0}%  Recall {1}%'.format(precision, recall))
    return hyps 


def get_topk_hypotheses_from_probdist(hyp_probdist, k=1, minprob=0.0):
    '''Returns topk hyps per src word over a minprob from SoftSGM prob. dist.'''
    hyp_probdist_topk_over_minprob = matops.keep_topk_over_minprob(
            hyp_probdist, k, minprob)
    nonzero_indices = np.nonzero(hyp_probdist_topk_over_minprob)
    hyp_src_inds = nonzero_indices[0].tolist()
    hyp_trg_inds = nonzero_indices[1].tolist()
    hyps = set(zip(hyp_src_inds, hyp_trg_inds)) 
    return hyps


def main(args):
    # Process data. Get train/dev split, seeds.
    (word_pairs, src_embs, src_word2ind, src_ind2word, trg_embs, trg_word2ind,
            trg_ind2word, oov_word_pairs) = load_embs_and_wordpairs(args)
    print('OOV Word Pairs:', oov_word_pairs)
    _, (train_inds, dev_inds) = create_train_dev_split(word_pairs, 
            args.n_seeds, src_word2ind, trg_word2ind, args.randomize_seeds)
    gold_src_train_inds, gold_trg_train_inds = unzip_pairs(train_inds)
    src_dev_inds, trg_dev_inds = unzip_pairs(dev_inds)

    
    # Normalize embeddings in-place.
    print('Normalizing embeddings...')
    embeddings.normalize(src_embs, args.norm) 
    embeddings.normalize(trg_embs, args.norm) 
    print('Done normalizing embeddings.')

    if args.function == 'sgm':
        print('Running SGM') 
        print('args.softsgm_iters:', args.softsgm_iters)
        print('args.iterative_softsgm_iters:', args.iterative_softsgm_iters,
                flush=True)

        # Make similarity matrices.
        xxT = src_embs @ src_embs.T
        yyT = trg_embs @ trg_embs.T

        hyps, _, _ = iterative_softsgm(xxT, yyT, gold_src_train_inds,
                gold_trg_train_inds, gold_src_train_inds, gold_trg_train_inds,
                args.softsgm_iters, args.k, args.min_prob, dev_inds,
                args.new_nseeds_per_round, curr_i=1, 
                total_i=args.iterative_softsgm_iters,
                diff_seeds_for_rev=args.diff_seeds_for_rev,
                active_learning=args.active_learning, 
                truth_for_active_learning=set(train_inds + dev_inds))


    if args.function == 'proc':
        print('Running Procrustes', flush=True)
        print('args.proc_iters:', args.proc_iters, flush=True)
        hyps, _, _, _, _ = iterative_procrustes_w_csls(src_embs, trg_embs,
                gold_src_train_inds, gold_trg_train_inds, gold_src_train_inds,
                gold_trg_train_inds, dev_inds, args.new_nseeds_per_round,
                total_i=args.proc_iters,
                diff_seeds_for_rev=args.diff_seeds_for_rev, k=args.k,
                active_learning=args.active_learning, 
                truth_for_active_learning=set(train_inds + dev_inds))

    print('----------------------------------')
    print('\nFinal Forward Eval:')
    dev_src_inds, dev_trg_inds = unzip_pairs(dev_inds)
    dev_hyps = set(hyp for hyp in hyps if hyp[0] in dev_src_inds)
    matches, precision, recall = eval(dev_hyps, dev_inds)
    print('\tDev Pairs matched: {0} \n\t(Precision; {1}%) (Recall: {2}%)'
            .format(len(matches), precision, recall), flush=True)



if __name__ == "__main__":
    
    parser = argparse.ArgumentParser(description='LAP Experiments')
    parser.add_argument('--src-embs', metavar='PATH', required=True,
        help='Path to source embeddings.')
    parser.add_argument('--trg-embs', metavar='PATH', required=True,
        help='Path to target embeddings.')
    parser.add_argument('--function', choices=['proc', 'sgm'], required=True, 
        help='Which function to run (Procrustes (proc) or SGM (sgm) ).')
    parser.add_argument('--norm', metavar='N', choices=['noop', 'unit', 'center'],
            nargs='+', required=True, 
            help='How to normalize embeddings (can take multiple args)')
    parser.add_argument('--max-embs', type=int, default=200000,
        help='Maximum num of word embeddings to use.')
    parser.add_argument('--min-prob', type=float, default=0.0,
        help='The minimum probability to consider for softsgm')
    parser.add_argument('--pairs', metavar='PATH', required=True, 
        help='train seeds + dev pairs')
    parser.add_argument('--n-seeds', type=int, required=True, help='Num train seeds to use')
    parser.add_argument('--proc-iters', type=int, default=10, 
        help='Rounds of iterative Procrustes to run.')
    parser.add_argument('--iterative-softsgm-iters', type=int, default=1, 
        help='Rounds of iterative SoftSGM to run.')
    parser.add_argument('--softsgm-iters', type=int, default=1, 
        help='Rounds of SoftSGM to run to create probdist.')
    parser.add_argument('--k', type=int, default=1, 
        help='How many hypotheses to return per source word.') 
    parser.add_argument('--randomize-seeds', action='store_true', 
            help='If set, randomizes the seeds to use (instead of getting them in '
            'order from args.pairs file)')
    parser.add_argument('--new-nseeds-per-round', metavar='N', type=int, nargs='+',
            default=-1, help='Number of seeds to add per round in iterative runs.')
    parser.add_argument('--diff-seeds-for-rev', action='store_true', 
        help='When running matching in reverse, regenerate seeds (if there are '
        'additional input seeds from a previous round, these will then be '
        'shuffled.')
    parser.add_argument('--active-learning', action='store_true', 
        help='Whether or not to do active learning')
    
    args = parser.parse_args()
    
    main(args)