File size: 6,076 Bytes
028cc28
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
import numpy as np
import torch

from evaluate import load as load_metric

from sklearn.metrics import accuracy_score, f1_score
from tqdm.auto import tqdm

MAX_TARGET_LENGTH = 128

# load evaluation metrics
sacrebleu = load_metric('sacrebleu')
rouge = load_metric('rouge')
meteor = load_metric('meteor')
bertscore = load_metric('bertscore')

# use gpu if it's available
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')

def flatten_list(l):
    """
    Utility function to convert a list of lists into a flattened list

    Params:
        l (list of lists): list to be flattened
    Returns:
        A flattened list with the elements of the original list
    """
    return [item for sublist in l for item in sublist]

def extract_feedback(predictions):
    """
    Utility function to extract the feedback from the predictions of the model

    Params:
        predictions (list): complete model predictions
    Returns:
        feedback (list): extracted feedback from the model's predictions
    """
    feedback = []
    # iterate through predictions and try to extract predicted feedback
    for pred in predictions:
        try:
            fb = pred.split(':', 1)[1]
        except IndexError:
            try:
                if pred.lower().startswith('partially correct'):
                    fb = pred.split(' ', 1)[2]
                else:
                    fb = pred.split(' ', 1)[1]
            except IndexError:
                fb = pred
        feedback.append(fb.strip())
    
    return feedback

def extract_labels(predictions):
    """
    Utility function to extract the labels from the predictions of the model

    Params:
        predictions (list): complete model predictions
    Returns:
        feedback (list): extracted labels from the model's predictions
    """
    labels = []
    for pred in predictions:
        if pred.lower().startswith('correct'):
            label = 'Correct'
        elif pred.lower().startswith('partially correct'):
            label = 'Partially correct'
        elif pred.lower().startswith('incorrect'):
            label = 'Incorrect'
        else:
            label = 'Unknown label'
        labels.append(label)
    
    return labels

def compute_metrics(predictions, labels):
    """
    Compute evaluation metrics from the predictions of the model

    Params:
        predictions (list): complete model predictions
        labels (list): golden labels (previously tokenized)
    Returns:
        results (dict): dictionary with the computed evaluation metrics
        predictions (list): list of the decoded predictions of the model
    """
    # extract feedback and labels from the model's predictions
    predicted_feedback = extract_feedback(predictions)
    predicted_labels = extract_labels(predictions)

    # extract feedback and labels from the golden labels
    reference_feedback = [x.split('Feedback:', 1)[1].strip() for x in labels]
    reference_labels = [x.split('Feedback:', 1)[0].strip() for x in labels]

    # compute HF metrics
    sacrebleu_score = sacrebleu.compute(predictions=predicted_feedback, references=[[x] for x in reference_feedback])['score']
    rouge_score = rouge.compute(predictions=predicted_feedback, references=reference_feedback)['rouge2']
    meteor_score = meteor.compute(predictions=predicted_feedback, references=reference_feedback)['meteor']
    bert_score = bertscore.compute(
        predictions=predicted_feedback,
        references=reference_feedback,
        lang='de',
        model_type='bert-base-multilingual-cased',
        rescale_with_baseline=True)
    
    # use sklearn to compute accuracy and f1 score
    reference_labels_np = np.array(reference_labels)
    accuracy = accuracy_score(reference_labels_np, predicted_labels)
    f1_weighted = f1_score(reference_labels_np, predicted_labels, average='weighted')
    f1_macro = f1_score(
        reference_labels_np,
        predicted_labels,
        average='macro',
        labels=['Incorrect', 'Partially correct', 'Correct'])

    results = {
        'sacrebleu': sacrebleu_score,
        'rouge': rouge_score,
        'meteor': meteor_score,
        'bert_score': np.array(bert_score['f1']).mean().item(),
        'accuracy': accuracy,
        'f1_weighted': f1_weighted,
        'f1_macro': f1_macro
        }
    
    return results

def evaluate(model, tokenizer, dataloader):
    """
    Evaluate model on the given dataset

    Params:
        model (PreTrainedModel): seq2seq model
        tokenizer (PreTrainedTokenizer): tokenizer from HuggingFace
        dataloader (torch Dataloader): dataloader of the dataset to be used for evaluation
    Returns:
        results (dict): dictionary with the computed evaluation metrics
        predictions (list): list of the decoded predictions of the model
    """
    decoded_preds, decoded_labels = [], []

    model.eval()
    # iterate through batchs in the dataloader
    for batch in tqdm(dataloader):
        with torch.no_grad():
            batch = {k: v.to(device) for k, v in batch.items()}
            # generate tokens from batch
            generated_tokens = model.generate(
                batch['input_ids'],
                attention_mask=batch['attention_mask'],
                max_length=MAX_TARGET_LENGTH
            )
            # get golden labels from batch
            labels_batch = batch['labels']
            
            # decode model predictions and golden labels
            decoded_preds_batch = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
            decoded_labels_batch = tokenizer.batch_decode(labels_batch, skip_special_tokens=True)

            decoded_preds.append(decoded_preds_batch)
            decoded_labels.append(decoded_labels_batch)

    # convert predictions and golden labels into flattened lists
    predictions = flatten_list(decoded_preds)
    labels = flatten_list(decoded_labels)

    # compute metrics based on predictions and golden labels
    results = compute_metrics(predictions, labels)

    return results, predictions