validate_dnn.py

#! /usr/bin/env python3.5

import argparse
import pandas
import numpy
from keras.models import load_model
from train_dnn import label_reg_loss


def get_speaker_str(speaker_df, row_id):
    return "http://localhost:8888/samples/%s/%s/%s" % \
           (speaker_df.ix[row_id, 0],
            speaker_df.ix[row_id, 1],
            speaker_df.ix[row_id, 2])


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description="Apply the model")
    parser.add_argument("--confidence-threshold", type=float, default=0.7,
                        help="Posterior probability threshold for confident predictions")
    parser.add_argument("--model_file", default='trained_model', help="Previously traine model")
    parser.add_argument("--dev_spk_file", default='full/spk_ivector.csv',
                        help="File with dev speaker data (IDs and i-vectors) in CSV format")

    args = parser.parse_args()

    model = load_model(args.model_file,
                       custom_objects={"label_reg_loss": label_reg_loss})

    # Load the name table, needed for mapping output IDs of the model
    # to real speaker names
    pruned_name_list = []
    name_ids = {}
    for l in open("%s.names" % args.model_file, "rt", encoding='utf-8'):
        name = l.strip()
        pruned_name_list.append(name)
        name_ids[name] = len(name_ids)

    dev_speaker_df = pandas.read_csv(args.dev_spk_file, sep=",", header=None)
    dev_ivecs = dev_speaker_df.ix[:, 3:].as_matrix()
    dev_predicted_targets = model.predict_on_batch(dev_ivecs)
    dev_predicted_speakers = (dev_predicted_targets).argmax(axis=1)
    dev_confident_predictions = \
        dev_predicted_targets[numpy.arange(len(dev_predicted_targets)), \
                              dev_predicted_speakers] > args.confidence_threshold

    for i in numpy.where(dev_confident_predictions)[0]:
        if dev_predicted_speakers[i] != name_ids["<unk>"]:
            print(u"%s: %s" % \
                  (pruned_name_list[dev_predicted_speakers[i]],
                   get_speaker_str(dev_speaker_df, i)))
            # pruned_name_list[dev_predicted_speakers[i]].encode("utf-8")))