-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
180 lines (159 loc) · 7.43 KB
/
main.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
import os
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import glob
import librosa
import torch
import torchaudio
import torchaudio.functional as F
import torchaudio.transforms as transforms
from torch.utils.data import Dataset, DataLoader, ConcatDataset, random_split
from librosa import feature
import torch.nn as nn
from torchsummary import summary
from tqdm import tqdm
import torchvision.transforms as vtransforms
import dcase_util
from sed_eval.audio_tag import AudioTaggingMetrics as get_metrics_sedeval
from collections import OrderedDict
from torchlibrosa.augmentation import SpecAugmentation
import torch.nn.functional as Ft
from constraints_code import Constraints
from model import SED
from get_data import getdata
from utils import make_one_hot, step, evaluate
os.environ["CUDA_VISIBLE_DEVICES"]="0"
device = torch.device('cuda')
num_epochs = 130
batch_size = 80
epsilon = 0.1
save_path = 'inference_results_settings_final12'
train = getdata(mode = 'train', device = device,
load_existing = False,save_path = save_path)
val = getdata(mode = 'val', device = device,
load_existing = False,save_path = save_path)
sampler = torch.utils.data.RandomSampler(data_source=train,
replacement=False,
num_samples=None,
generator=None)
train_dl = DataLoader(train,batch_size=batch_size, drop_last=True,sampler=sampler)
dd_train_dl = DataLoader(train,batch_size=batch_size, shuffle=False, drop_last=True,sampler=sampler)
val_dl = DataLoader(val,batch_size=batch_size, shuffle=False, drop_last=True)
num_batches = [len(train_dl),len(dd_train_dl),len(val_dl)]
warmup_iters = int(10*num_batches[0])#8epochs
initial_lr_w = 0.003
max_lr_w = 2.5*initial_lr_w
min_lr_w = 1e-2*initial_lr_w
initial_lr_dd = 0.0035
beta = 0.01
initial_loss,current_loss = 0.0,0.0
t = 1
t1 = 1
num_iter = 0
lambda_iters = 0
total_iters = int(num_batches[0]*num_epochs)#120 epochs
evaluate_after = int(num_batches[0]*3) #6epochs
l = 2 # the number of iterations (l) to wait before dd are updated
d = 50 # the increment in number of iterations (l) to wait before dd are updated
last_lambda_iter = 0
lambda_batch_idx = 0
lambda_iter = 0
last_logged_eval = 0
last_lambda_update = 0
w_weight_bce,w_weight_ce,dd_constraint_wt = 1.1,1.6,0.85
model = SED().to(device)
constraints = Constraints(ontology = train.ontology).to(device)
criterion_bce = nn.BCEWithLogitsLoss(weight=None,
reduction='mean',
pos_weight=torch.Tensor([1.5,2.1,1.8,1.5,1.5,1.5,1.5,
1.5,1.5,2.1,1.1,1.1,1.1,1.1])).to(device)
criterion_ce = nn.CrossEntropyLoss(reduction='mean',
label_smoothing=0.0)
optimizer_w = torch.optim.Adam(model.parameters(),
lr=initial_lr_w)
optimizer_dd = torch.optim.Adam(constraints.parameters(),
lr=initial_lr_dd)
scheduler_w = torch.optim.lr_scheduler.OneCycleLR(optimizer = optimizer_w,
max_lr = max_lr_w,
epochs = num_epochs,
steps_per_epoch = num_batches[0],
anneal_strategy='cos',
pct_start=0.3,
cycle_momentum=True,
base_momentum=0.85,
max_momentum=0.95,
div_factor = max_lr_w/initial_lr_w,
final_div_factor = initial_lr_w/min_lr_w,
three_phase=False,
last_epoch=-1,
verbose=False)
scheduler_dd = torch.optim.lr_scheduler.LambdaLR(optimizer = optimizer_dd,
lr_lambda = lambda t: 1/(1+(beta*t)),
last_epoch=-1,
verbose=False)
h_k = pd.DataFrame(np.zeros((evaluate_after,len(train.ontology.keys()))),columns=[f'constraint{k}' for k in range(len(train.ontology.keys()))])
training_loss = pd.DataFrame(np.zeros((num_epochs,4)),columns=['bce','ce','cons','total'])
torch.cuda.empty_cache()
epoch = 0
while epsilon*initial_loss<=current_loss and num_iter<total_iters:
w_loss_bce = 0.0
w_loss_ce = 0.0
c_loss = 0.0
tot_loss = 0.0
for mel_spec, target_c in tqdm(train_dl):
model.train()
constraints.train()
mel_spec,target_c,target_p = mel_spec.to(device),target_c[0].to(device),target_c[1].to(device)
logits = model(mel_spec)
# logits_children,logits_parents
loss_w_bce = criterion_bce(logits,make_one_hot(target_c,target_p))
loss_w_ce = criterion_ce(logits[:,:10],target_c.squeeze(1).long())
for lambda_param in constraints.parameters():
lambda_param.requires_grad = False
closs,hk = constraints(torch.sigmoid(logits))
train_loss = w_weight_bce*loss_w_bce + w_weight_ce*loss_w_ce + dd_constraint_wt*closs
w_loss_bce += loss_w_bce.item()
w_loss_ce += loss_w_ce.item()
c_loss += closs.item()
tot_loss += train_loss.item()
h_k.iloc[num_iter-int(last_logged_eval*evaluate_after),:] = hk.tolist()
step(model = model,
loss = train_loss,
optimizer = optimizer_w,
scheduler = scheduler_w)
num_iter += 1
if ((num_iter+1) % evaluate_after == 0):
last_logged_eval += 1
# print(f'Evaluating-{last_logged_eval}th time. Total iters so far-{num_iter+1}')
evaluate(model = model,
constraints = constraints,
dataset = val,
label2id = val.unique_labels,
thresh=0.5,
save_path = save_path,
iteration_name = f'eval-{last_logged_eval}-totaliters-{num_iter+1}')
training_loss.to_csv(os.path.join(save_path,'train_loss.csv'))
h_k.to_csv(os.path.join(save_path,f'h_k_after_{num_iter+1}iters.csv'))
# print('Evaluation done.')
if (num_iter >= warmup_iters) and (num_iter-last_lambda_iter >= l):
# print('updating lambda')
model.eval()
constraints.train()
for lambda_param in constraints.parameters():
lambda_param.requires_grad = True
for idx,minibatch in enumerate(dd_train_dl):
with torch.no_grad():
logits= model(minibatch[0])
closs,_ = constraints(torch.sigmoid(logits).detach())
step(model = constraints,
loss = -1.0*dd_constraint_wt*closs,
optimizer = optimizer_dd,
scheduler = scheduler_dd)
t+=1
last_lambda_iter = num_iter
l += d
break
model.train()
training_loss.iloc[epoch] = [w_loss_bce/num_batches[0],w_loss_ce/num_batches[0],c_loss/num_batches[0],tot_loss/num_batches[0]]
epoch += 1