def explanation_score(k, predict_proba, original_data, valid_data, pool, dataset):
pool = add_mmos_explanations(predict_proba, pool, dataset, 1, softmax_temps=[1.])
pool = add_prediction_info(predict_proba, pool)
def get_predicted_class_score(example):
return example['explanation'].scores.values()[0]
class_score_sorted = sorted(pool, key=get_predicted_class_score, reverse=False)
return class_score_sorted[:k]
def contains_misunderstood_words(k, predict_proba, original_data, valid_data, pool, dataset):
"""Choose examples that contain words that played a role in vaidation data misclassification"""
softmax_temp = .025
explanation_key = 'explanation'
valid_data = add_prediction_info(predict_proba, valid_data)
confusion_matrix = get_confusion_matrix(valid_data)
all_labels = range(confusion_matrix.shape[0])
valid_data = add_mmos_explanations(predict_proba, valid_data, dataset, len(all_labels), softmax_temps=[1., 5.])
all_token_stats = {}
confusion_matrix_counts = []
confusion_matrix_labels = []
for label, predicted_class in product(all_labels, all_labels):
if label == predicted_class:
continue
stats = get_token_misclassification_stats(valid_data, explanation_key, predicted_class, label)
# only keep things tokens that on average contributed to misclassification
stats = stats.loc[stats.total_contribution > 0]
stats['rank'] = stats.frequency_rank * stats.total_contribution
stats['sampling_prob'] = softmax(stats['rank'].values, softmax_temp)
all_token_stats[(label, predicted_class)] = stats
confusion_matrix_labels.append((label, predicted_class))
confusion_matrix_counts.append(confusion_matrix[label, predicted_class])
confusion_matrix_probs = confusion_matrix_counts / np.sum(confusion_matrix_counts)
samples_by_token = group_examples_by_token(pool)
# iteratively sample new examples until we have k
selected_examples = set()
while len(selected_examples) < k:
# sample uniformly from non-diagonal confusion matrix cells
bin_index = np.random.choice(np.arange(len(confusion_matrix_labels)),
p=confusion_matrix_probs)
label_and_predicted = confusion_matrix_labels[bin_index]
if all_token_stats[label_and_predicted].empty:
continue
# sample from token importance with softmax
token = np.random.choice(all_token_stats[label_and_predicted].index.values,
p=all_token_stats[label_and_predicted].sampling_prob)
if token not in samples_by_token:
continue
# sample from pool containing words
example_index = np.random.choice(samples_by_token[token])
if example_index in selected_examples:
continue
else:
selected_examples.add(example_index)
return [valid_data[i] for i in selected_examples]
def explanation_variance(k, predict_proba, original_data, valid_data, pool, dataset):
pool = add_mmos_explanations(predict_proba, pool, dataset, 1)
pool = add_prediction_info(predict_proba, pool)
def get_predicted_class_variance(example):
predicted_class = example['predicted']
return example['explanation'].blackbox_prob_stats(predicted_class)['std']
class_variance_sorted = sorted(pool, key=get_predicted_class_variance, reverse=False)
return class_variance_sorted[:k]
def contains_misunderstood_words_by_label(k, predict_proba, original_data, valid_data, pool, dataset,
explanation_key='explanation', set_size=1):
"""Choose examples that contain words that played a role in vaidation data misclassification"""
softmax_temp = .025
original_data = add_prediction_info(predict_proba, original_data)
confusion_matrix = get_confusion_matrix(original_data)
all_labels = range(confusion_matrix.shape[0])
if explanation_key == 'explanation':
original_data = add_mmos_explanations(predict_proba, original_data, dataset, len(all_labels),
softmax_temps=[1., 5.],max_simultaneous_perturbations=set_size)
elif explanation_key == 'lime_explanation':
original_data = add_lime_explanation(predict_proba, valid_data,
len(all_labels),
num_features=10,
num_samples=1000)
else:
raise ValueError('Only explanation keys allowed are "explanation" and "lime_explanation"')
all_token_stats = {}
confusion_matrix_counts = []
confusion_matrix_labels = []
for label, predicted_class in product(all_labels, all_labels):
if label == predicted_class:
continue
stats = get_token_misclassification_stats(original_data, explanation_key, predicted_class, label)
# only keep things tokens that on average contributed to misclassification
stats = stats.loc[stats.total_contribution > 0]
stats['rank'] = stats.frequency_rank * stats.total_contribution
# print label, predicted_class
# print stats.sort_values('rank').tail(10)
stats['sampling_prob'] = softmax(stats['rank'].values, softmax_temp)
all_token_stats[(label, predicted_class)] = stats
confusion_matrix_labels.append((label, predicted_class))
confusion_matrix_counts.append(confusion_matrix[label, predicted_class])
confusion_matrix_probs = confusion_matrix_counts / np.sum(confusion_matrix_counts)
pool = add_prediction_info(predict_proba, pool)
pool_by_predicted = defaultdict(list)
for sample in pool:
pool_by_predicted[sample['predicted']].append(sample)
pool_by_predicted_token = {label: group_examples_by_token(label_set) for label, label_set in pool_by_predicted.items() }
# iteratively sample new examples until we have k
selected_examples = set()
while len(selected_examples) < k:
# sample uniformly from non-diagonal confusion matrix cells
bin_index = np.random.choice(np.arange(len(confusion_matrix_labels)),
p=confusion_matrix_probs)
label_and_predicted = confusion_matrix_labels[bin_index]
label, predicted = label_and_predicted
if all_token_stats[label_and_predicted].empty:
continue
# sample from token importance with softmax
token = np.random.choice(all_token_stats[label_and_predicted].index.values,
p=all_token_stats[label_and_predicted].sampling_prob)
if token not in pool_by_predicted_token[predicted]:
continue
# sample from pool containing words
example_index = np.random.choice(pool_by_predicted_token[predicted][token])
if example_index in selected_examples:
continue
else:
selected_examples.add(example_index)
np.random.choice(pool, k, replace=False)
return [pool[i] for i in selected_examples]