Repository navigation
Expand file tree
/
Copy pathmain.py
More file actions
281 lines (217 loc) · 11.2 KB
/
Copy pathmain.py
File metadata and controls
281 lines (217 loc) · 11.2 KB
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
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
from asyncio import run
import os
import torch
from ipv
8000
8.taskmanager import TaskManager
from dataclasses import dataclass
import threading
from ipv8.community import Community, CommunitySettings
from ipv8.configuration import ConfigBuilder, Strategy, WalkerDefinition, default_bootstrap_defs
from ipv8.lazy_community import lazy_wrapper
from ipv8.messaging.payload_dataclass import overwrite_dataclass
from ipv8.types import Peer
from time import time, localtime, strftime
from sklearn.model_selection import train_test_split
from ipv8.util import run_forever
from ipv8_service import IPv8
import numpy as np
from transformers import T5ForConditionalGeneration, T5Tokenizer, AdamW
import pandas as pd
import random
from vars import *
# Check if CUDA is available and set the default device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")
# Enhance normal dataclasses for IPv8 (see the IPV8 serialization documentation)
dataclass = overwrite_dataclass(dataclass)
@dataclass(msg_id=1) # The value 1 identifies this message and must be unique per community
class Query_res:
query: str
result: str
class LTRCommunity(Community):
community_id = b'\x9d\x10\xaa\x8c\xfa\x0b\x19\xee\x96\x8d\xf4\x91\xea\xdc\xcb\x94\xa7\x1d\x8b\x00'
def __init__(self, settings: CommunitySettings) -> None:
super().__init__(settings)
print ('-----------------------------------------------------------------')
number_of_docs_for_this_user = np.random.randint(number_of_docs_per_user[0], number_of_docs_per_user[1])
with open('./output.log', 'a') as file:
file.write(' doc nbr:' + str(number_of_docs_for_this_user) + ' ')
self.df = df[df['doc_id'].isin(list(docs_to_be_used.sample(number_of_docs_for_this_user)['doc_id']))]
with open('./output.log', 'a') as file:
file.write(' df size ' + str(self.df.shape[0]) + ' ')
self.batches_so_far = 0
self.current_queries = []
self.current_docs = []
self.timestamps = []
self.accuracies_avg = 0
self.accuracies_sum = 0
self.rolling_window = 100
task_manager = TaskManager()
self.losses = []
self.accuracies = []
self.accuracy = []
self.add_message_handler(Query_res, self.on_message)
self.ready_for_input = threading.Event()
self.lamport_clock = 0
self.past_data = {'queries': [], 'results': []}
def on_peer_added(self, peer: Peer) -> None:
print("I am:", self.my_peer, "I found:", peer)
def train_model(self, queries, responses):
self.got_here = False
self.batches_so_far+=1
# Tokenize the lists of queries and responses
inputs = self.tokenizer(queries, padding=True, return_tensors="pt", truncation=True).input_ids.to(device)
labels = self.tokenizer(responses, padding=True, return_tensors="pt", truncation=True).input_ids.to(device)
# Forward pass
outputs = self.model(input_ids=inputs, labels=labels)
loss = outputs.loss
# Extract logits and convert to token IDs
logits = outputs.logits
predicted_token_ids = torch.argmax(logits, dim=-1)
self.accuracy = []
# Decode token IDs to text for each item in the batch
for i in range(predicted_token_ids.size(0)):
predicted_text = self.tokenizer.decode(predicted_token_ids[i], skip_special_tokens=True)
# with open('./output.log', 'a') as file:
# file.write('\n' + predicted_text + ' ' + responses[i])
if predicted_text == responses[i]:
self.accuracy.append(1)
else:
self.accuracy.append(0)
acc = np.sum(self.accuracy) / len(self.accuracy)
self.losses.append(round(float(loss.detach()),3))
self.accuracies.append(acc)
self.timestamps.append(int(time()))
self.accuracies_sum += acc
if len(self.accuracies) > self.rolling_window:
self.accuracies_sum -= self.accuracies[-self.rolling_window]
self.accuracies_avg = self.accuracies_sum / self.rolling_window
else:
self.accuracies_avg = self.accuracies_sum / len(self.accuracies)
if self.batches_so_far % batches_save_threshold == 0:
pd.DataFrame(list(zip(self.timestamps, self.accuracies)),
columns = ['timestamps','accuracies']).to_csv(f'aggregated_results/{root_folder_for_shard}/accuracies/{self.my_peer.address[1]}_accuracies.csv')
pd.DataFrame(list(zip(self.timestamps, self.losses)),
columns = ['timestamps','losses']).to_csv(f'aggregated_results/{root_folder_for_shard}/losses/{self.my_peer.address[1]}_losses.csv')
self.model.save_pretrained(f'aggregated_results/{root_folder_for_shard}/models/{self.my_peer.address[1]}_{self.batches_so_far}')
self.df.to_csv(f'aggregated_results/{root_folder_for_shard}/datasets/{self.my_peer.address[1]}_df.csv')
# self.change_df(test_df)
# if self.accuracies_avg>0.95:
if self.batches_so_far > 10000:
raise SystemExit
end_time = time()
print (end_time - start_time)
print(
8000
f'peer port:{self.my_peer.address[1]}, loss: {round(float(loss.detach()),3)}, ACCURACY ": {round(acc,2)}, '
f'ACCURACY_AVG: {round(self.accuracies_avg,2)}')
with open('./output.log', 'a') as file:
file.write(f'peer port:{self.my_peer.address[1]}, loss: {round(float(loss.detach()),3)}, ACCURACY ": {round(acc,2)}, '
f'ACCURACY_AVG: {round(self.accuracies_avg,2)}')
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
def get_querynres(self):
self.row = np.random.randint(0, self.df.shape[0])
# print ('ROW NUMBER: ', self.row)
query = self.df.iloc[self.row]['query']
# selected_res = docs[docs == self.df.iloc[self.row]['doc_id']].index[0]
selected_res = self.df.iloc[self.row]['doc_id']
self.row += 1
return query, selected_res
def started(self) -> None:
print('Indexing (please wait)...')
# Load model and tokenizer
self.model = T5ForConditionalGeneration.from_pretrained(model_name).to(device)
self.tokenizer = T5Tokenizer.from_pretrained(model_name)
self.optimizer = AdamW(self.model.parameters(), lr=1e-3)
# Training loop
self.model.train()
self.network.add_peer_observer(self)
print(self.get_peers())
async def start_communication() -> None:
print(f'running comms routine with {len(self.get_peers())} peers')
if len(self.get_peers()) > 0:
# if not self.lamport_clock:
query, selected_res = self.get_querynres()
print('starting comms & ending comms')
self.cancel_pending_task("start_communication")
p = random.choice(self.get_peers())
self.ez_send(p, Query_res(query=query, result=selected_res))
else:
print('gonna try again')
pass
async def send_query() -> None:
if len(self.get_peers()) == 0:
return None
new_query, new_res = self.get_querynres()
if len(self.current_queries) == 0:
for i in range ( round(batch_size/len(self.get_peers())) ):
new_query, new_res = self.get_querynres()
self.current_queries.append(new_query)
self.current_docs.append(new_res)
p = random.choice(self.get_peers())
self.check_batch_size_and_train()
self.ez_send(p, Query_res(query=new_query, result=new_res))
self.register_task("start_communication", start_communication, interval=5.0, delay=0)
self.register_task("send_random_q_d_pair", send_query, interval=0.01, delay=0)
def check_batch_size_and_train(self):
if len(self.current_queries)>=batch_size:
self.train_model(self.current_queries, self.current_docs)
self.current_queries = []
self.current_docs = []
def change_df(self, df_changer):
self.df = df_changer.copy()
self.got_here = True
@lazy_wrapper(Query_res)
def on_message(self, peer: Peer, payload: Query_res) -> None:
self.current_queries.append(payload.query)
self.current_docs.append(payload.result)
self.check_batch_size_and_train()
async def start_communities() -> None:
for i in range(peer_nbr):
builder = ConfigBuilder().clear_keys().clear_overlays()
builder.add_key("my peer", "medium", f"certs/ec{i}.pem")
builder.add_overlay("LTRCommunity", "my peer",
[WalkerDefinition(Strategy.RandomWalk, 10, {'timeout': 3.0})],
default_bootstrap_defs, {}, [('started',)])
await IPv8(builder.finalize(),
extra_communities={'LTRCommunity': LTRCommunity}).start()
await run_forever()
start_time = time()
print (start_time)
all_docs_till_now = []
df = pd.read_csv('./orcas.tsv', sep='\t', header=None,
names=['query_id', 'query', 'doc_id', 'doc'])
def find_shard_folders(directory):
"""
List all folders within a given directory that have 'X' in their name.
:param directory: The root directory to search within.
:return: A list of paths to folders that contain 'X' in their name.
"""
folders = []
for root, dirs, files in os.walk(directory):
# Filter directories in the current root that contain 'X'
matched_dirs = [os.path.join(root, d) for d in dirs if 'group' in d]
folders.extend(matched_dirs)
return folders
previous_shards = find_shard_folders('aggregated_results')
print (previous_shards)
for s in previous_shards:
prev_docs = pd.read_csv(f'{s}/datasets/train_df.csv')['doc_id'].unique()
all_docs_till_now.extend(list(prev_docs))
print ('docs till now', len(all_docs_till_now))
df = df[~df['doc_id'].isin(all_docs_till_now)]
root_folder_for_shard = f'group{shard}'
os.makedirs(f'aggregated_results/{root_folder_for_shard}/datasets', exist_ok=True)
os.makedirs(f'aggregated_results/{root_folder_for_shard}/models', exist_ok=True)
os.makedirs(f'aggregated_results/{root_folder_for_shard}/accuracies', exist_ok=True)
os.makedirs(f'aggregated_results/{root_folder_for_shard}/losses', exist_ok=True)
cnter = df.groupby('doc_id').count().sort_values('query',ascending = False) # get most referenced docs subset
cnter = cnter[cnter['query']>=2] # remove docs that are referenced only once to make stratify work
docs_to_be_used = cnter.sample(total_doc_count).reset_index()
df = df[df['doc_id'].isin(list(docs_to_be_used['doc_id']))]
with open('./output.log', 'a') as file:
file.write(' doc nbr:' + str(docs_to_be_used.shape[0]) + ' shape:' + str(df.shape[0]))
train_df, test_df = train_test_split(df, test_size=0.5, random_state=42, stratify=df['doc_id']) # split into train and test
train_df.to_csv(f'aggregated_results/{root_folder_for_shard}/datasets/train_df.csv')
test_df.to_csv(f'aggregated_results/{root_folder_for_shard}/datasets/test_df.csv')
df = train_df.copy()
run(start_communities())