metrics.py

import editdistance
import numpy as np

def word_error_rate(predicted_outputs, ground_truths):
	""" Estimate Word_error_rate.
	Args:
		predicted_outputs(list) : result of model prediction
		ground_truths(list) : ground truth

	Returns:
		World Error rate(float) : Word error rate Estimated by Edit distance.
	"""
	sum_wer=0.0
	for output,ground_truth in zip(predicted_outputs,ground_truths):
		output=output.split(" ")
		ground_truth=ground_truth.split(" ")
		distance = editdistance.eval(output, ground_truth)
		length = max(len(output),len(ground_truth))
		sum_wer+=(distance/length)
	return sum_wer/len(predicted_outputs)


def sentence_acc(predicted_outputs, ground_truths):
	""" Estimate sentence_acc.
	Args:
		predicted_outputs(list) : result of model prediction
		ground_truths(list) : ground truth

	Returns:
		sentence_acc(float) : Acurracy between preicted_output and ground_truths
	"""
	correct_sentences=0
	for output,ground_truth in zip(predicted_outputs,ground_truths):
		if np.array_equal(output,ground_truth):
			correct_sentences+=1
	return correct_sentences/len(predicted_outputs)


def get_worst_wer_img_path(img_path_list, predicted_outputs, ground_truths):
	""" Return Information of max word error rate Image
	Args:
		img_path_list(list) : list of image path
		predicted_outputs(list) : result of model prediction
		ground_truths(list) : ground truth

	Returns:
		image path(str) : Image path of worst error rate
		word error rate(float) : max word error rate
		ground truth(str) : Ground truth of max word error rate image
		predicted_output(str) : Prediction of model
	"""
	max_wer_ind = 0
	max_wer = 0

	i = 0
	for output, ground_truth in zip(predicted_outputs,ground_truths):
		output=output.split(" ")
		ground_truth=ground_truth.split(" ")

		distance = editdistance.eval(output, ground_truth)
		length = max(len(output), len(ground_truth))
		cur_wer = (distance / length)
		if max_wer < cur_wer:
			max_wer = cur_wer
			max_wer_ind = i
		i+=1

	return img_path_list[max_wer_ind], max_wer, ground_truths[max_wer_ind], predicted_outputs[max_wer_ind]