-
Notifications
You must be signed in to change notification settings - Fork 0
/
evaluate.py
132 lines (108 loc) · 5.57 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
from tqdm import tqdm
import numpy as np
import torch
from torch.nn import CrossEntropyLoss
from torch.nn.functional import softmax
from torch.utils.data import Dataset, DataLoader
from transformers import DataCollatorForLanguageModeling
from data import load_split_dataset, TensorDataset
from examples import get_examples
@torch.inference_mode()
def get_loss(generator, batch, labels_loss=False):
model = generator.model
loss_fct = CrossEntropyLoss(reduction='none', ignore_index=-100)
input_ids = batch['input_ids'].to(model.device)
attention_mask = batch['attention_mask'].to(model.device)
label_tokens = batch['token_type_ids']
labels = torch.where(attention_mask == 1, input_ids, -100)
with torch.autocast(device_type='cuda', dtype=torch.float16):
outputs = model(input_ids=input_ids, attention_mask=attention_mask)
logits = outputs.logits[..., :-1, :].contiguous().to(model.dtype)
shift_labels = labels[..., 1:].contiguous().to(logits.device)
losses = loss_fct(logits.view(-1, logits.size(-1)), shift_labels.view(-1))
losses = losses.view(logits.size(0), logits.size(1))
if labels_loss:
label_mask = label_tokens[..., 1:].contiguous().to(model.device)
losses = losses * label_mask
losses = losses.sum(dim=-1) / label_mask.sum(dim=-1)
else:
losses = losses.mean(dim=-1)
losses = losses.detach().cpu()
return losses
def classify(losses, labels, correction_factor=None, mode="diagonal_W"):
"""this function applies a correction factor from the calibrate method to the model's predicted distribution"""
num_classes = len(labels)
if correction_factor is None:
# do not calibrate
W = torch.eye(num_classes, dtype=losses.dtype)
b = torch.zeros(num_classes, dtype=losses.dtype)
else:
# calibrate
if mode == "diagonal_W":
W = torch.linalg.inv(torch.eye(num_classes, dtype=losses.dtype) * correction_factor)
b = torch.zeros(num_classes, dtype=losses.dtype)
elif mode == "identity_W":
W = torch.eye(num_classes)
b = -1 * correction_factor[:, None]
else:
raise NotImplementedError(f"{mode} is not implemented for calibration")
uncalibrated_probs = softmax(-losses)
calibrated_probs = torch.matmul(uncalibrated_probs, W) + b
return np.array(labels)[calibrated_probs.argmax(1)], calibrated_probs
def predict(generator, eval_dataset, labels, batch_size=1, method='direct', labels_loss=False,
calibrate_dataset=None, mode='diagonal_W'):
collator = DataCollatorForLanguageModeling(generator.tokenizer, mlm=False)
if method == 'calibrate':
calibrate_dataloader = DataLoader(calibrate_dataset,
shuffle=False,
batch_size=batch_size,
collate_fn=collator)
# get probability distribution for context-free inputs
cf_losses = []
for batch in tqdm(calibrate_dataloader):
cf_losses.extend(get_loss(generator, batch, labels_loss))
cf_losses = torch.tensor(cf_losses, dtype=torch.float32).reshape(-1, len(labels))
cf_label_probs = softmax(-cf_losses)
# calculate calibration correction term
correction_factor = torch.mean(cf_label_probs, dim=0)
torch.cuda.empty_cache()
else:
correction_factor = None
eval_dataloader = DataLoader(eval_dataset, shuffle=False, batch_size=batch_size, collate_fn=collator)
losses = []
for batch in tqdm(eval_dataloader):
losses.extend(get_loss(generator, batch, labels_loss))
losses = torch.tensor(losses, dtype=torch.float32).reshape(-1, len(labels))
results, probs = classify(losses, labels, correction_factor, mode)
return results, probs
def evaluate_setup(dataset, generator, seed, template, num_shots, selection_method,
example_ids=None, examples_path=None,
prediction_method='direct',
labels_loss=False,
batch_size=16,
cache_dir=None,
):
train, val, labels_mp = load_split_dataset(dataset, cache_dir=cache_dir)
labels = list(labels_mp.values())
selected_examples = get_examples(dataset, train, selection_method, seed, num_shots,
example_ids=example_ids,
examples_path=examples_path,
)
examples, example_ids = selected_examples["examples"], selected_examples["example_ids"]
eval_dataset = TensorDataset([x.strip() for x in val['input']],
generator.tokenizer, labels, template,
examples=examples,
method=prediction_method,
)
eval_dataset.print_tensorized_example()
if 'calibrate' in prediction_method:
context_free_inputs = ["N/A", "", "[MASK]"]
calibrate_dataset = TensorDataset(context_free_inputs, generator.tokenizer, labels, template,
examples=examples, method='direct',
)
else:
calibrate_dataset = None
results, probs = predict(generator, eval_dataset, labels, batch_size=batch_size, method=prediction_method,
labels_loss=labels_loss, calibrate_dataset=calibrate_dataset)
score = (results == val['target']).mean()
return {"score": score, "probs": probs, "predicts": results, "example_ids": example_ids}