-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathevaluate.py
148 lines (132 loc) · 6.99 KB
/
evaluate.py
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
from datasets import QuestionSentenceDataset
from metrics import get_eval_metrics, acc
import numpy as np
import pandas as pd
from grader.model import Grader
from grader.tokenization import Tokenizer as GraderTokenizer
from verifier.model import Verifier
import sys
import torch
from torch.utils.data import DataLoader
from tqdm import tqdm
def _get_metrics(predictions_p, ground_truths_p, scalers_p):
ground_truths = torch.from_numpy(np.array(ground_truths_p))
predictions = torch.from_numpy(np.array(predictions_p))
scalers = torch.from_numpy(np.array(scalers_p))
return get_eval_metrics(predictions, ground_truths, scalers)
def get_all_metrics(data_loader, reader, retriever, reader_tokenizer, device, topK, dump_pred=False, dump_file=""):
with torch.no_grad():
student_scores = {}
predictions = []
ground_truths = []
test_q = []
test_report = []
test_labels = []
test_whole = []
test_predictions = []
test_ids= []
ver_preds = []
ver_ground_truths = []
for _, data in tqdm(enumerate(data_loader)):
questions, reports, labels, report_IDs, class_weights_for_verifier = data
labels = labels.to(device)
class_weights_for_verifier = class_weights_for_verifier.to(device)
verifier_logits, topK_sentences, _ = retriever.verify(questions, reports)
if verifier_logits.shape[0] ==1:
verifier_logits = verifier_logits.view(1)
else:
verifier_logits = verifier_logits.squeeze()
verifier_nonzeros = np.array([1 if prob>=0.5 else 0 for prob in verifier_logits.detach().cpu().numpy()]).nonzero()[0]
test_q.extend(questions)
test_report.extend(topK_sentences)
test_whole.extend(reports)
test_labels.extend(labels.detach().cpu().numpy())
test_ids.extend(report_IDs)
ver_preds.extend([1 if prob>=0.5 else 0 for prob in verifier_logits.detach().cpu().numpy()])
ver_ground_truths.extend([1 if label>0 else 0 for label in labels.cpu().numpy()])
if len(verifier_nonzeros) != 0:#there is data to pass to the grader
grader_questions, grader_reports, grader_topK_sentences = np.array(questions).take(verifier_nonzeros), np.array(reports).take(verifier_nonzeros), np.array(topK_sentences).take(verifier_nonzeros)
questions_tokenized, reports_tokenzied, topK_tokenized = reader_tokenizer(list(grader_questions), list(grader_reports), list(grader_topK_sentences))
questions_tokenized = questions_tokenized.to(device)
reports_tokenzied = reports_tokenzied.to(device)
topK_tokenized = topK_tokenized.to(device)
logits = reader(questions_tokenized, reports_tokenzied, topK_tokenized)
predictions.extend(logits.cpu()) # contains probability predictions for val loss for 1-5 so for grader
ground_truths.extend((torch.index_select(labels, 0, torch.from_numpy(verifier_nonzeros).to(device))).cpu()) # TODO:this is not exactly true let me think about this. contains labels 1,2,...5 but with 1 minus versions for grader loss
if dataset_name == "college_physics":
pred_probab = torch.nn.Softmax(dim=1)(logits)
y_pred = pred_probab.argmax(1).cpu()
else:
test_predictions.extend([1 if prob>=0.5 else 0 for prob in logits.detach().cpu().numpy()])
else:
y_pred = []
if dataset_name == "college_physics":
all_preds = [0] * labels.shape[0]
for index, pred in zip(verifier_nonzeros, y_pred):
all_preds[index] = pred.item()
for id,pred,gt in zip(report_IDs, all_preds, labels.cpu()):
act_gt = gt.item()
act_pred = pred
if id in student_scores:
cum_preds, cum_gts = student_scores[id]
student_scores[id] = (cum_preds + act_pred, cum_gts + act_gt)
else:
student_scores[id] = (act_pred, act_gt)
if dataset_name == "college_physics":
predictions = torch.tensor([pred for pred,_ in student_scores.values()])
ground_truths = torch.tensor([gt for _,gt in student_scores.values()])
metrics = _get_metrics(predictions, ground_truths, torch.ones_like(predictions))
else:
ground_truths = torch.tensor(ver_ground_truths)
predictions = torch.tensor(test_predictions)
metrics = {"acc": acc(predictions, ground_truths)}
if dump_pred:
pd.DataFrame({"question":test_q, "filtered_report":test_report, "whole_report":test_whole, "labels":test_labels, "preds":test_predictions, "IDs":test_ids,"ver_preds":ver_preds}).to_csv(dump_file)
return metrics
grader_file_path = sys.argv[1]
verifier_file_path = sys.argv[2]
dataset_name = sys.argv[3]
if dataset_name == "college_physics":
train_folder = "data/train"
val_folder = "data/val"
test_folder = "data/test"
train_labels = "data/labels/train.csv"
val_labels = "data/labels/val.csv"
test_labels = "data/labels/test.csv"
rubric_dimension = "data/rubric_dimensions.json"
else:
train_folder = "middle_school_data"
val_folder = "middle_school_data"
test_folder = "middle_school_data"
train_labels = "middle_school_data/middle_school_essay1_2_train.csv"
val_labels = "middle_school_data/middle_school_essay1_2_val.csv"
test_labels = "middle_school_data/middle_school_essay1_2_test.csv"
rubric_dimension = "middle_school_data/rubric_dimensions.json"
config = {
"train_folder": train_folder,
"val_folder": val_folder,
"test_folder": test_folder,
"train_labels": train_labels,
"val_labels": val_labels,
"test_labels": test_labels,
"rubric_dimension": rubric_dimension,
"batch_size": 4
}
device = torch.device("cuda" if torch.cuda.is_available() else "cpu") #torch.device("cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu")
test_set = QuestionSentenceDataset(config["test_folder"], config["test_labels"],
config["rubric_dimension"], is_val=True)
test_loader = DataLoader(test_set, 1)
verifier = Verifier(bert_or_sbert="sbert", topK=3)
verifier.load_state_dict(torch.load(f"{verifier_file_path}/model.pth",map_location=torch.device('cpu')))
verifier = verifier.to(device)
verifier.eval()
grader = Grader("electra")
grader.load_state_dict(torch.load(f"{grader_file_path}/model.pth",map_location=torch.device('cpu')))
grader = grader.to(device)
grader.eval()
grader_tokenizer = GraderTokenizer("electra")
test_metrics = get_all_metrics(test_loader, grader, verifier, grader_tokenizer, device, 3, True, "test_results.csv")
print("\ntest:", end=" ")
for metric, value in test_metrics.items():
print(f"{metric}: {round(value,3)},",end=" ")
print()