-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcharm_trainer_og.py
183 lines (154 loc) · 7.36 KB
/
charm_trainer_og.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
from autocommand import autocommand
#from torch.utils.tensorboard import SummaryWriter
import rn_model
import datetime
#import deep_gambler as dg
import numpy as np
import os
import readCharmDataset as riq
import signal
import torch
import torch.nn as nn
import torch.optim as optim
def print_stats(acc_mat, name, epoch, tensorboard):
classes = acc_mat.shape[0]
ones = np.ones((classes, 1)).squeeze(-1)
corrects = np.diag(acc_mat)
acc = corrects.sum()/acc_mat.sum()
recall = (corrects/acc_mat.dot(ones)).round(4)
precision = (corrects/ones.dot(acc_mat)).round(4)
f1 = (2*recall*precision/(recall+precision)).round(4)
print(f"Epoch {epoch} on {name} dataset")
print(f"Accuracy: {acc}")
if tensorboard:
tensorboard.add_scalar(f"accuracy/{name}", acc, epoch)
print(f"\t\tRecall\tPrecision\tF1")
for c in range(classes):
print(f"Class {c}\t\t{recall[c]}\t{precision[c]}\t\t{f1[c]}")
#if tensorboard:
# tensorboard.add_scalar(f"recall_{c}/{name}", recall[c], epoch)
# tensorboard.add_scalar(f"precision_{c}/{name}", precision[c], epoch)
# tensorboard.add_scalar(f"f1_{c}/{name}", f1[c], epoch)
# tensorboard.flush()
def tensorboard_parse(tensorboard):
'''
tensorboard: a string with comma separated <key>=<value> substrings, each of
them mapping to a tensorboard.SummaryWriter constructor parameter.
E.g.,
log_dir='./runs',comment='',purge_step=None,max_queue=10,flush_secs=120,filename_suffix=''
'''
writer = None
if tensorboard:
conf = {}
for tok in tensorboard.split(','):
kv = tok.split('=')
if len(kv) == 2:
if kv[1] == 'None':
kv[1] = None
conf[kv[0]] = kv[1]
writer = SummaryWriter(**conf)
return writer
class EarlyExitException(Exception):
def __str__(self):
return "Received termination signal"
class CharmTrainer(object):
def __init__(self, id_gpu="0", data_folder=".", batch_size=64, chunk_size=200000, sample_stride=0, loaders=8, dg_coverage=0.999, tensorboard=None):
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = id_gpu
self.device = (torch.device('cuda') if torch.cuda.is_available()
else torch.device('cpu'))
print(f"Training on {self.device}")
signal.signal(signal.SIGINT, self.exit_gracefully)
signal.signal(signal.SIGTERM, self.exit_gracefully)
self.chunk_size = chunk_size
self.loss_fn = nn.CrossEntropyLoss() #dg.GamblerLoss(3)
self.dg_coverage = dg_coverage
self.train_data = riq.IQDataset(data_folder=data_folder, chunk_size=chunk_size, stride=sample_stride)
self.train_data.normalize(torch.tensor([-2.7671e-06, -7.3102e-07]), torch.tensor([0.0002, 0.0002]))
self.train_loader = torch.utils.data.DataLoader(self.train_data, batch_size=batch_size, shuffle=True, num_workers=loaders, pin_memory=True)
self.val_data = riq.IQDataset(data_folder=data_folder, chunk_size=chunk_size, stride=sample_stride, subset='validation')
self.val_data.normalize(torch.tensor([-2.7671e-06, -7.3102e-07]), torch.tensor([0.0002, 0.0002]))
self.val_loader = torch.utils.data.DataLoader(self.val_data, batch_size=batch_size, shuffle=False, num_workers=loaders, pin_memory=True)
self.running = False
self.best_val_accuracy = 0.0
self.tensorboard = tensorboard_parse(tensorboard)
def init(self):
self.model = rn_model.CharmBrain(self.chunk_size).to(self.device)
self.optimizer = optim.Adam(self.model.parameters())
self.best_val_accuracy = 0.0
def training_loop(self, n_epochs):
for self.loss_fn.o in [1.7]:
self.init()
self.model.train()
for epoch in range(n_epochs):
loss_train = 0.0
for chunks, labels in self.train_loader:
if not self.running:
raise EarlyExitException
chunks = chunks.to(self.device, non_blocking=True)
labels = labels.to(self.device, non_blocking=True)
output = self.model(chunks)
loss = self.loss_fn(output, labels)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
loss_train += loss.item()
if self.tensorboard:
self.tensorboard.add_scalar("Loss/train", loss_train/len(self.train_loader), epoch)
if True:
print(f"{datetime.datetime.now()} Epoch {epoch}, loss {loss_train/len(self.train_loader)}")
print(f"Coverage: {self.dg_coverage}, o-parameter {self.loss_fn.o}")
self.validate(epoch, train=True)
self.model.train()
def validate(self, epoch, train=True):
loaders = [('val', self.val_loader)]
if train:
loaders.append(('train', self.train_loader))
self.model.eval()
for name, loader in loaders:
correct = 0
total = 0
acc_mat = np.zeros((len(self.train_data.label), len(self.train_data.label)))
with torch.no_grad():
for chunks, labels in loader:
if not self.running:
raise EarlyExitException
chunks = chunks.to(self.device, non_blocking=True)
labels = labels.to(self.device, non_blocking=True)
output = self.model(chunks)
#predicted = dg.output2class(output, self.dg_coverage, 3)
_, predicted = torch.max(output, dim=1)
total += labels.shape[0]
correct += int((predicted == labels).sum())
for i in range(labels.shape[0]):
acc_mat[labels[i]][predicted[i]] += 1
accuracy = correct/total
print(f"{name} accuracy: {accuracy}")
if name == 'val' and accuracy>self.best_val_accuracy:
self.save_model(f"charm_{self.dg_coverage}_{self.loss_fn.o}_{round(accuracy, 2)}.pt")
self.best_val_accuracy = accuracy
print_stats(acc_mat, name, epoch, self.tensorboard)
def save_model(self, filename='charm.pt'):
'''
load your model with:
>>> model = brain.CharmBrain()
>>> model.load_state_dict(torch.load(filename))
'''
torch.save(self.model.state_dict(), filename)
def execute(self, n_epochs):
self.running = True
try:
self.training_loop(n_epochs)
self.validate(n_epochs-1, train=True)
except EarlyExitException:
pass
if self.tensorboard:
self.tensorboard.close()
print("[Done]")
def exit_gracefully(self, signum, frame):
self.running = False
@autocommand(__name__)
def charm_trainer(id_gpu="0", data_folder="./oran_dataset", n_epochs=25, batch_size=512, chunk_size=20000, sample_stride=0, loaders=6, dg_coverage=0.75, tensorboard=None):
ct = CharmTrainer(id_gpu=id_gpu, data_folder=data_folder, batch_size=batch_size, chunk_size=chunk_size, sample_stride=sample_stride,
loaders=loaders, dg_coverage=dg_coverage, tensorboard=tensorboard)
ct.execute(n_epochs=n_epochs)