-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathutils_nnd.py
195 lines (159 loc) · 8.87 KB
/
utils_nnd.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
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
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModelForSeq2SeqLM
import torch, os, tqdm, json
from collections import Counter
models_folder = os.environ["MODELS_FOLDER"]
# Unified Generator interface
class GeneratorHF:
def __init__(self, model_card="gpt2-medium", device="cuda", starter_file=None, max_enc_length=None, max_dec_length=None, force_dec_prepend=None):
self.model_card = model_card
self.is_gpt2 = "gpt2" in self.model_card
if self.is_gpt2:
self.model = AutoModelForCausalLM.from_pretrained(self.model_card)
else:
self.model = AutoModelForSeq2SeqLM.from_pretrained(self.model_card)
self.model.to(device)
self.tokenizer = AutoTokenizer.from_pretrained(self.model_card)
self.max_enc_length = max_enc_length
self.max_dec_length = max_dec_length
self.force_dec_prepend = force_dec_prepend
self.model.eval()
self.start_id = self.tokenizer.bos_token_id
self.end_id = self.tokenizer.eos_token_id
if "prophetnet" in self.model_card:
# bos_token_id=102, eos_token_id=102
self.start_id, self.end_id = 102, 102
if "facebook/wmt19" in self.model_card:
self.start_id, self.end_id = 2, 2
if "t5" in self.model_card: # Just for MT
self.start_id, self.end_id = 0, 1
if self.start_id is None and self.end_id is not None:
# For MixQG
self.start_id = 0
self.device = device
if self.is_gpt2:
self.tokenizer.pad_token = self.tokenizer.eos_token
self.model.config.pad_token_id = self.tokenizer.pad_token_id
if starter_file is not None:
self.reload(starter_file, strict=False)
def reload(self, from_file, strict=True):
if not os.path.isfile(from_file):
# Try to look at the models folder for the file
from_file = os.path.join(models_folder, from_file)
assert os.path.isfile(from_file), "Starter file not found, in absolute or in models folder"
loaded_dict = torch.load(from_file)
print(self.model.load_state_dict(loaded_dict, strict=strict))
def save(self, to_file):
torch.save(self.model.state_dict(), to_file)
def preprocess(self, encoded_texts, decoded_texts, max_enc_length=None, max_dec_length=None):
assert len(encoded_texts) == len(decoded_texts), "Mismatch in input/output sizes"
encoder_ids = self.tokenizer.batch_encode_plus(encoded_texts, add_special_tokens=True, return_tensors="pt", padding=True, truncation=True).input_ids.to(self.device)
if self.force_dec_prepend is not None:
decoded_texts = [self.force_dec_prepend + text for text in decoded_texts]
decoder_tokenized = [self.tokenizer.encode(text=text, add_special_tokens=False) for text in decoded_texts]
decoder_ids_input = torch.nn.utils.rnn.pad_sequence([torch.LongTensor([self.start_id] + dec) for dec in decoder_tokenized], batch_first=True, padding_value=self.end_id).to(self.device)
decoder_ids_output = torch.nn.utils.rnn.pad_sequence([torch.LongTensor(dec + [self.end_id]) for dec in decoder_tokenized], batch_first=True, padding_value=-1).to(self.device)
if self.max_enc_length is not None and max_enc_length is None:
max_enc_length = self.max_enc_length
if self.max_dec_length is not None and max_dec_length is None:
max_dec_length = self.max_dec_length
if max_enc_length is not None:
encoder_ids = encoder_ids[:, :max_enc_length]
if max_dec_length is not None:
decoder_ids_input = decoder_ids_input[:, :max_dec_length]
decoder_ids_output = decoder_ids_output[:, :max_dec_length]
return encoder_ids, decoder_ids_input, decoder_ids_output
def score_batch(self, encoded_texts, decoded_texts, max_enc_length=None, max_dec_length=None):
encoder_ids, decoder_ids_input, decoder_ids_output = self.preprocess(encoded_texts, decoded_texts, max_enc_length, max_dec_length)
if "facebook/wmt19" in self.model_card:
# There's a problem with this model for now.
with torch.no_grad():
model_output = self.model(input_ids=encoder_ids, decoder_input_ids=decoder_ids_input, return_dict=True, labels=decoder_ids_output)
return {"scores": [-model_output["loss"]]}
with torch.no_grad():
crit = torch.nn.CrossEntropyLoss(ignore_index=-1, reduction="none")
if self.is_gpt2:
encoder_output = self.model(input_ids=encoder_ids, past_key_values=None, return_dict=True)
past = encoder_output["past_key_values"]
decoder_output = self.model(input_ids=decoder_ids_input, past_key_values=past, return_dict=True)
logits = decoder_output["logits"]
else:
model_output = self.model(input_ids=encoder_ids, decoder_input_ids=decoder_ids_input, return_dict=True)
logits = model_output["logits"]
N, seqlength, vocab_size = logits.shape
loss_components = crit(logits.view(N*seqlength, vocab_size), decoder_ids_output.contiguous().view(-1)).reshape(N, seqlength)
num_words = torch.sum(decoder_ids_output != -1, dim=1)
score_per_item = (- torch.sum(loss_components, dim=1) / num_words).tolist()
return {"scores": score_per_item}
def score(self, encoded_texts, decoded_texts, max_enc_length=None, max_dec_length=None, batch_size=32, progress=False):
N = len(encoded_texts)
iterator = range(0, N, batch_size)
if progress and len(iterator) > 1:
iterator = tqdm.tqdm(iterator)
scores = []
for i in iterator:
batch_encoded_texts = encoded_texts[i:i+batch_size]
batch_decoded_texts = decoded_texts[i:i+batch_size]
batch_scores = self.score_batch(batch_encoded_texts, batch_decoded_texts, max_enc_length, max_dec_length)["scores"]
scores += batch_scores
return {"scores": scores}
CACHE_LIKELIHOODS = {}
def compute_likelihood(gen, model_name, document, gen_text):
global CACHE_LIKELIHOODS
key = (model_name, document, gen_text)
if key in CACHE_LIKELIHOODS:
return CACHE_LIKELIHOODS[key]
else:
CACHE_LIKELIHOODS[key] = gen.score([document], [gen_text])["scores"][0]
return CACHE_LIKELIHOODS[key]
def save_nnd_cache(filename):
global CACHE_LIKELIHOODS
with open(filename, "w") as f:
cached_obj = {"___!!!___".join(k): v for k, v in CACHE_LIKELIHOODS.items()}
json.dump(cached_obj, f)
def load_nnd_cache(filename):
global CACHE_LIKELIHOODS
with open(filename, "r") as f:
cached_obj = json.load(f)
CACHE_LIKELIHOODS = {tuple(k.split("___!!!___")): v for k, v in cached_obj.items()}
print("Reloaded %d values" % (len(CACHE_LIKELIHOODS)))
def run_nnd(dataset, model, model_name, no_error_label="NoE", report_type="count", breakdown_key=None, progress=True):
"""
Run Near Negative Distinction for a model on a dataset.
"""
assert any(k in report_type for k in ["count", "accuracy"])
ite = dataset
if progress:
ite = tqdm.tqdm(ite, desc="NND %s" % model_name)
error_counts = Counter()
error_breakdown_counts = Counter()
for d in ite:
if model_name == "Max Errors":
error_counts[d["error1"]] += 1
error_counts[d["error2"]] += 1
else:
LP1 = compute_likelihood(model, model_name, d["document"], d["gen1"])
LP2 = compute_likelihood(model, model_name, d["document"], d["gen2"])
chosen_label = d["error1"] if LP1 > LP2 else d["error2"]
error_counts[chosen_label] += 1
if breakdown_key is not None:
error_breakdown_counts["%s_%s" % (d[breakdown_key], chosen_label)] += 1
accuracy = 100.0 * (error_counts[no_error_label] / sum(error_counts.values()))
D = {"model_name": model_name, "accuracy": accuracy}
if "count" in report_type:
D.update({"C_%s" % (k): v for k, v in error_counts.items()})
if breakdown_key is not None:
D.update({"C_%s" % (k): v for k, v in error_breakdown_counts.items()})
if "accuracy" in report_type:
category_counts = Counter([d["error1"] for d in dataset] + [d["error2"] for d in dataset])
for k, v in error_counts.items():
if k == no_error_label:
D["A_%s" % (k)] = v / category_counts[k]
else:
D["A_%s" % (k)] = 1.0 - v / category_counts[k]
if breakdown_key is not None:
category_labels = Counter([d[breakdown_key] for d in dataset])
for cat_label, cat_count in category_labels.items():
for k in error_counts.keys():
if k == no_error_label:
D["A_%s" % (cat_label)] = error_breakdown_counts["%s_%s" % (cat_label, k)] / cat_count
return D