forked from osalpekar/llm-target-determinator
-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_dataset.py
223 lines (179 loc) · 7.47 KB
/
test_dataset.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
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
import math
import os
from pathlib import Path
from typing import List, NamedTuple, Tuple
import torch
from preproc import get_functions
from tokenizer import Tokenizer
from torch.utils.data import Dataset
REPO_ROOT = Path(__file__).resolve().parent
# The size of inputs to the model for inference is variable. Naively, it's
# shape is (num_functions_in_file x context_length). Different files have
# different numbers of functions, and a number of files have hundreds or
# thousands of functions. Memory utilization (size of intermediate activations,
# size of batch, etc.) is therefore correlated with the input size. Since we
# may run this model on lower-memory GPUs such as 24GB A10G's, we ensure we can
# still complete indexing by partitioning files up-front and then indexing the
# partitioned files, which has lower memory utilization.
class FilePartition(NamedTuple):
"""
Stores metadata about a partition of a file.
partition_id is 1-indexed, must be 0 < partition_id <= num_partitions
Un-partitioned files have partition_id = num_partitions = 1
"""
filename: str
partition_id: int
num_partitions: int
class UnittestDataset(Dataset):
def __init__(self, config):
self.config = config
self.filelist = self.create_filelist()
self.tokenizer = Tokenizer(self.config)
def create_filelist(self):
"""
Returns the list of files of interest in the specified subdirectory.
We sort the generated list so that each rank deterministically ends up
with the same filelist, which ensures the Dataloder and
DistributedSampler correctly cover all the relevant files.
"""
all_files = []
for root, dirs, files in os.walk(self.config.project_dir):
for file in files:
if self.should_include_file(file, root):
file_path = os.path.join(root, file)
all_files.append(file_path)
return sorted(all_files)
def should_include_file(self, file, root):
"""
Returns whether the specified file should be included in the file list.
Note that we are only interested in Python files with the specified prefix,
provided those files do not lie in certain directories like third_party.
"""
return (
file.endswith(".py")
and file.startswith(self.config.file_prefix)
and "third_party" not in root
)
def __len__(self):
pass
def tokenize_items(self, items: List[Tuple[str, str]]):
token_list = []
names = []
for name, item in items:
tokens = self.tokenizer.encode(item)
token_list.append(tokens)
names.append(name)
# Concatenate the tokenized vectors from each item into a single
# matrix of shape (num_items x embedding_dimension)
if self.config.model == "codellama":
num_items = len(token_list)
# Create a tensor of shape (num_items x CONTEXT_LENGTH) filled
# with pad_id
tensor = torch.full(
(num_items, self.config.max_context_len),
self.tokenizer.pad_id,
dtype=torch.long,
)
print(f"pad: {self.tokenizer.pad_id}")
for k, t in enumerate(token_list):
# truncate token list to max context length
t = t[: self.config.max_context_len]
# insert the tokens into the empty tensor
tensor[k, : len(t)] = torch.tensor(t, dtype=torch.long)
attn_mask = torch.where(tensor == self.tokenizer.pad_id, 0.0, 1.0)
return {"tokens": tensor, "attn_mask": attn_mask}, names
else:
return torch.cat(token_list), names
class FunctionGranularityDataset(UnittestDataset):
"""
Dataset where a function from a file is a signular item. Each function gets
its own tokenization and embedding.
"""
def __init__(self, config):
super().__init__(config)
self.file_partitions: List[FilePartition] = flatten(
[self.partition_file(file) for file in self.filelist]
)
# Store results from the ast parser in an attempt to reduce number of
# times a file is parsed
self.functions_cache = {}
def partition_file(self, filename) -> List[FilePartition]:
threshold = 100
# Approximate number of functions in file by counting number of
# functions defs. Partition file based on this number. Each partition
# should have <100 functions.
partitions = []
with open(filename) as f:
text = f.read()
occurrences = text.count("def ")
if occurrences and occurrences > threshold:
num_partitions = math.ceil(occurrences / threshold)
for i in range(num_partitions):
partitions.append(
FilePartition(filename, i + 1, num_partitions)
)
else:
partitions.append(FilePartition(filename, 1, 1))
return partitions
def __len__(self):
return len(self.file_partitions)
def __getitem__(self, idx):
filename, partition, num_partitions = self.file_partitions[idx]
# Use function info from a previous parsing of the ast in cache if
# possible
if filename in self.functions_cache:
functions_in_file = self.functions_cache[filename]
else:
functions_in_file = sorted(get_functions(filename).items())
self.functions_cache[filename] = functions_in_file
num_functions = len(functions_in_file)
size = math.ceil(num_functions / num_partitions)
start_i = size * (partition - 1)
end_i = start_i + size
functions = functions_in_file[start_i:end_i]
# Some test files don't actually have any unittest functions. We handle
# that case here.
if len(functions) == 0:
empty_tensor = torch.tensor([], dtype=torch.int64).reshape(
0, self.config.max_context_len
)
return ({"tokens": empty_tensor, "attn_mask": empty_tensor}, [])
# Get tokens for each function
return self.tokenize_items(functions)
class FileGranularityDataset(UnittestDataset):
"""
Dataset where the entire file is the item. An entire file gets tokenized
and embedded together.
"""
def __init__(self, config):
super().__init__(config)
def __len__(self):
return len(self.filelist)
def __getitem__(self, idx):
filename = self.filelist[idx]
with open(filename) as f:
text = f.read()
clean_path = os.path.relpath(
filename, REPO_ROOT.parent / "pytorch/test"
)
return self.tokenize_items([(clean_path, text)])
def flatten(lst):
return [item for sublist in lst for item in sublist]
def collate_fn(data):
examples_tokens = [item[0]["tokens"] for item in data]
examples_masks = [item[0]["attn_mask"] for item in data]
funclist = flatten([item[1] for item in data])
examples = {
"tokens": torch.cat(examples_tokens),
"attn_mask": torch.cat(examples_masks),
}
return examples, funclist
# Small Test
# dataset = UnittestDataset("assets/filelist.json")
# dataloader = DataLoader(dataset, collate_fn=collate_fn, batch_size=2)
# for idx, batch in enumerate(dataloader, 0):
# data, funcs = batch
# print(data.shape)
# print(len(funcs))
# if idx == 0:
# sys.exit(0)