-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathdata.py
186 lines (164 loc) · 8.98 KB
/
data.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
from collections import defaultdict, Counter
import numpy as np
import os
import pickle
import random
import torch
import torch.utils.data as data
import utils
def custom_collate_fn(batch):
batch = [torch.tensor(l, dtype=torch.long) for l in batch]
return torch.nn.utils.rnn.pad_sequence(batch, batch_first=True)
class CaseDataset:
def __init__(self, fname, model, tokenizer, limit=-1, case_set=None, role_set=None, balanced=False, average=False):
self.fname = fname
self.case_set = case_set
self.role_set = role_set
self.balanced = balanced
self.average = average
role_string = "aso" if role_set is None else ''.join(role_set)
if balanced:
limit_type = f"{role_string}_balanced"
else:
limit_type = f"{role_string}_unbalanced"
# Get the filename from the path.
save_fn = os.path.split(fname)[1]
# Remove the extension.
save_fn = os.path.splitext(save_fn)[0]
save_fn = "all_features_aso_exps_" + save_fn
if case_set is None:
save_fn = f"{save_fn}_{limit_type}"
else:
case_string = ''.join(case_set)
save_fn = f"{save_fn}_{limit_type}_{case_string}"
if limit > 0:
save_fn = f"{save_fn}_{str(limit)}"
else:
save_fn += "_nolimit"
if self.average:
save_fn += "_average"
tokens_labels_dir = "cached_datasets"
tokens_labels_path = os.path.join(tokens_labels_dir, save_fn + '.pkl')
if os.path.exists(tokens_labels_path):
print("Loading all of the tokens and non-bert stuff from", tokens_labels_path)
self.tokens, self.case_labels, self.role_labels, self.word_forms_list, \
self.animacy_labels, self.len, self.relevant_examples_index, \
self.cases_per_role,self.bert_tokens, self.bert_ids, self.orig_to_bert_map, \
self.bert_to_orig_map = \
pickle.load(open(tokens_labels_path, 'rb'))
else:
self.tokens, self.case_labels, self.role_labels, self.word_forms_list, \
self.animacy_labels, self.len, self.relevant_examples_index, self.cases_per_role = \
utils.get_tokens_and_labels(self.fname, limit=limit, case_set=case_set, role_set=role_set, balanced=balanced)
self.bert_tokens, self.bert_ids, self.orig_to_bert_map, self.bert_to_orig_map = \
utils.get_bert_tokens(self.tokens, tokenizer)
print("lengths of bert ids etc", len(self.bert_tokens), len(self.bert_ids), len(self.orig_to_bert_map), len(self.bert_to_orig_map))
print("Saving all of the tokens and non-bert stuff to", tokens_labels_path)
pickle.dump(
(self.tokens, self.case_labels, self.role_labels, self.word_forms_list,
self.animacy_labels, self.len, self.relevant_examples_index,
self.cases_per_role,self.bert_tokens, self.bert_ids,
self.orig_to_bert_map, self.bert_to_orig_map),
open(tokens_labels_path, 'wb'))
# We need to check whether the length is large enough before we run through BERT.
# Otherwise, super unbalanced datasets will end up running whole training
# treebanks through BERT.
print(f"There are {self.len} relevant tokens, and {len(self.tokens)} overall sentences")
if self.len < limit and balanced:
print(f"Set is smaller than limit! Length {self.len}, limit {limit}.")
return
bert_vectors_dir = 'cached_bert_vectors'
hdf5_path = os.path.join(bert_vectors_dir, save_fn + ".hdf5")
self.bert_outputs = utils.get_bert_outputs(hdf5_path, self.bert_ids, model)
print("length of bert outputs", len(self.bert_outputs))
def __len__(self):
return self.len
def get_bert_id_dataloader(self, batch_size=32):
return data.DataLoader(self.bert_ids, batch_size=batch_size, shuffle=True, collate_fn=custom_collate_fn)
def get_case_distribution(self):
case_distribution = defaultdict(Counter)
for sentence_num, word_num in self.relevant_examples_index:
role = self.role_labels[sentence_num][word_num]
case = self.case_labels[sentence_num][word_num]
case_distribution[role][case] += 1
return case_distribution
class CaseLayerDataset(data.Dataset):
def __init__(self, case_dataset, layer_num, labeldict=None, verbose=False):
self.layer_num = layer_num
self.verbose = verbose
self.case_dataset = case_dataset
self.balanced = self.case_dataset.balanced
self.pool_method = "average" if case_dataset.average else "first"
self.embs, self.role_labels, self.case_labels, self.word_forms, \
self.animacy_labels, self.idxs, indices_by_role = \
self.get_labels(case_dataset.bert_outputs, case_dataset.role_labels,
case_dataset.case_labels, case_dataset.word_forms_list,
case_dataset.animacy_labels, case_dataset.orig_to_bert_map,
case_dataset.relevant_examples_index, pool_method=self.pool_method)
if self.balanced:
min_role_len = min([len(indices_by_role[role]) for role in case_dataset.role_set])
print(f"Balancing cases to all have {min_role_len} elements")
combined_indices = []
for role in case_dataset.role_set:
combined_indices += indices_by_role[role][:min_role_len]
print(f"After trimming cases, have {len(combined_indices)} total indices")
# For curriculum reasons, we probably don't want to have our training
# examples with all roles in order.
random.shuffle(combined_indices)
self.embs = [self.embs[index] for index in combined_indices]
self.role_labels = [self.role_labels[index] for index in combined_indices]
self.case_labels = [self.case_labels[index] for index in combined_indices]
self.word_forms = [self.word_forms[index] for index in combined_indices]
self.animacy_labels = [self.animacy_labels[index] for index in combined_indices]
self.idxs = [self.idxs[index] for index in combined_indices]
print("Examples #", len(self.idxs))
self.labeldict = self.get_label_dict(labeldict)
print("labeldict", self.labeldict)
self.processed_labels = [(self.labeldict[x] if x in self.labeldict else -1) for x in self.role_labels]
def __getitem__(self, idx):
case_label = self.case_labels[idx] if self.case_labels[idx] is not None else ""
animacy_label = self.animacy_labels[idx] if self.animacy_labels[idx] is not None else ""
return self.embs[idx], self.processed_labels[idx], (case_label, self.word_forms[idx], animacy_label, self.idxs[idx])
def __len__(self):
return len(self.embs)
def get_label_dict(self, old_labeldict):
# Make a labeldict of all of the labels in this dataset, keeping the same
# name fo
labelset = sorted(list(set(self.role_labels)))
if old_labeldict is None:
curr_label = 0
labeldict = {}
else:
labeldict = old_labeldict
curr_label = len(old_labeldict)
for label in labelset:
if old_labeldict is None or label not in old_labeldict:
labeldict[label] = curr_label
curr_label += 1
return labeldict
def get_label_set(self):
return sorted(self.labeldict.keys(), key=lambda x: self.labeldict[x])
def get_num_labels(self):
return len(self.labeldict)
def get_labels(self, bert_outputs, role_labels, case_labels, word_forms_list, animacy_labels, orig_to_bert_map, relevant_examples_index, pool_method="first"):
train = []
out_role_labels, out_case_labels, out_word_forms, out_animacy_labels, out_index = [], [], [], [], []
indices_by_role = defaultdict(list)
for sentence_num, word_num in relevant_examples_index:
role_label = role_labels[sentence_num][word_num]
out_role_labels.append(role_label)
out_case_labels.append(case_labels[sentence_num][word_num])
out_word_forms.append(word_forms_list[sentence_num][word_num])
out_animacy_labels.append(animacy_labels[sentence_num][word_num])
bert_start_index = orig_to_bert_map[sentence_num][word_num]
bert_end_index = orig_to_bert_map[sentence_num][word_num + 1]
bert_sentence = bert_outputs[sentence_num][self.layer_num].squeeze()
if pool_method == "first":
train.append(bert_sentence[bert_start_index])
elif pool_method == "average":
train.append(np.mean(bert_outputs[sentence_num][self.layer_num].squeeze()[bert_start_index:bert_end_index]))
indices_by_role[role_label].append(len(out_role_labels) - 1)
out_index.append((sentence_num, bert_start_index, bert_end_index, word_num))
return train, out_role_labels, out_case_labels, out_word_forms, out_animacy_labels, out_index, indices_by_role
def get_dataloader(self, batch_size=32, shuffle=True):
return data.DataLoader(self, batch_size=batch_size, shuffle=shuffle)