-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
196 lines (171 loc) · 11.5 KB
/
train.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
from omegaconf import DictConfig, OmegaConf
import hydra, logging, os, itertools, glob
import numpy as np
import matplotlib.pyplot as plt
import torch
from torch.utils.tensorboard import SummaryWriter
from torch.utils.data import DataLoader
from geomloss import SamplesLoss
from src.data import TriangleSampling, CombinedIterator
from src.utils import TicToc, save_checkpoint
from src.models import MongeNet
# A logger for this file
logger = logging.getLogger(__name__)
@hydra.main(config_path="configs", config_name='train')
def train_app(cfg):
# override configuration with a user defined config file
if cfg.user_config is not None:
user_config = OmegaConf.load(cfg.user_config)
cfg = OmegaConf.merge(cfg, user_config)
logger.info('Training MongeNet\nConfig:\n{}'.format(OmegaConf.to_yaml(cfg)))
os.makedirs(cfg.trainer.output_dir, exist_ok=True)
# load datasets
train_dl, test_dl = [], []
for data_path in glob.glob(cfg.trainer.data_glob_str):
# read files
triplets, ylist = sorted(glob.glob(os.path.join(data_path,'x*'))), sorted(glob.glob(os.path.join(data_path, 'y*')))
# create train set
training_set = TriangleSampling(triplets[:cfg.trainer.num_train_samples], ylist[:cfg.trainer.num_train_samples])
training_generator = DataLoader(training_set, batch_size=cfg.trainer.batch_size, shuffle=True, num_workers=0, pin_memory=True, worker_init_fn=lambda x: np.random.seed())
train_dl.append(training_generator)
#create test set
test_set = TriangleSampling(triplets[cfg.trainer.num_train_samples:cfg.trainer.num_train_samples+cfg.trainer.num_test_samples],ylist[cfg.trainer.num_train_samples:cfg.trainer.num_train_samples+cfg.trainer.num_test_samples])
test_generator = DataLoader(test_set, batch_size=cfg.trainer.batch_size, shuffle=True, num_workers=0, pin_memory=True, worker_init_fn=lambda x: np.random.seed())
test_dl.append(test_generator)
logger.info('{} training data loaders and {} test data loaders read!'.format(len(train_dl), len(test_dl)))
# tensorboard logger
tb_log_folder = os.path.join(cfg.trainer.output_dir, 'tb_logs')
tb_writer = SummaryWriter(tb_log_folder)
logger.info("Tensorboard logs in {}".format(tb_log_folder))
# model, optimizer, and criterion
device_used = cfg.mongenet.device
model = MongeNet(cfg).to(device_used)
model.train()
optimizer = torch.optim.Adam(model.parameters(), lr=cfg.trainer.adam_lr)
criterion = SamplesLoss(loss="sinkhorn", backend='tensorized', reach=None, debias=True, blur=0.00005, scaling=0.9)
logger.info("MongeNet Model: {}".format(model));logger.info("MongeNet Optim: {}".format(optimizer))
# train loop
num_of_reps, tictoc = 2, TicToc()
for i in range(1, cfg.trainer.num_epochs+1):
tictoc.tic('epoch')
train_loss, train_loss_aprox, train_loss_reg = 0.0, 0.0, 0.0
train_combined_iterator = itertools.islice(CombinedIterator(train_dl), cfg.trainer.train_epoch_size)
Yindex, Yidxs = model.output_index, torch.unique(model.output_index)
for X,Y in train_combined_iterator:
# read batch and perform predictions
X, Y = X.float().to(device_used), Y.float().to(device_used)
optimizer.zero_grad()
Ypred = torch.stack([model(X)[0] for _ in range(num_of_reps)], dim=0)
Y = torch.stack([Y for _ in range(num_of_reps)], dim=0)
# approximation
approxvalue = 0.0
Ypred_l = Ypred.view(num_of_reps*cfg.trainer.batch_size, model.num_outputs, 2)
Y_l = Y.view(num_of_reps*cfg.trainer.batch_size, Y.shape[-2], 2)
for idx in Yidxs:
approxvalue += criterion(Ypred_l[:, Yindex == idx], Y_l).sum()
approxvalue = approxvalue / float(len(Yidxs)*cfg.trainer.batch_size*num_of_reps)
train_loss_aprox += approxvalue.item()
# diversity
regvalue = 0.0
for idx in Yidxs:
regvalue += (-1. * criterion(Ypred[0][:, Yindex == idx], Ypred[1][:, Yindex == idx]).sum())
regvalue = regvalue / float(len(Yidxs)*cfg.trainer.batch_size)
train_loss_reg += regvalue.item()
# final loss
lossvalue = approxvalue + cfg.trainer.reg_coef*regvalue
train_loss += lossvalue.item()
lossvalue.backward()
optimizer.step()
# log losses
train_loss = train_loss/float(cfg.trainer.train_epoch_size); tb_writer.add_scalar('Train/loss', train_loss, i)
train_loss_aprox = train_loss_aprox/float(cfg.trainer.train_epoch_size); tb_writer.add_scalar('Train/approx_loss', train_loss_aprox, i)
train_loss_reg = train_loss_reg/float(cfg.trainer.train_epoch_size); tb_writer.add_scalar('Train/diversity_loss', train_loss_reg, i)
tb_writer.add_scalar("Train/lr", optimizer.param_groups[0]['lr'], i)
# log training progress
logger.info('Train Epoch: {} - Training Loss: {:.6f}, Approx Loss: {:.6f}, Diversity Loss: {:.6f}, LR: {:1.2e} in {:3.4f} sec'.format(
i, train_loss, train_loss_aprox, train_loss_reg, optimizer.param_groups[0]['lr'], tictoc.toc('epoch')))
# evaluate iteration
if i % cfg.trainer.test_inteval == 0:
tictoc.tic('test_it')
# obs: turn on the dropout also during inference
model.train()
with torch.no_grad():
# compute test loss
test_loss, test_loss_aprox, test_loss_reg = 0.0, 0.0, 0.0
test_combined_iterator = itertools.islice(CombinedIterator(test_dl), cfg.trainer.test_epoch_size)
for X,Y in test_combined_iterator:
X, Y = X.float().to(device_used), Y.float().to(device_used)
Ypred = torch.stack([model(X)[0] for _ in range(num_of_reps)], dim=0)
Y = torch.stack([Y for _ in range(num_of_reps)], dim=0)
# approximation
approxvalue = 0.0
Ypred_l = Ypred.view(num_of_reps*cfg.trainer.batch_size, model.num_outputs, 2)
Y_l = Y.view(num_of_reps*cfg.trainer.batch_size, Y.shape[-2], 2)
for idx in Yidxs:
approxvalue += criterion(Ypred_l[:, Yindex == idx], Y_l).sum()
approxvalue = approxvalue / float(len(Yidxs)*cfg.trainer.batch_size*num_of_reps)
test_loss_aprox += approxvalue.item()
# diversity
regvalue = 0.0
for idx in Yidxs:
regvalue += (-1.0 * criterion(Ypred[0][:, Yindex == idx], Ypred[1][:, Yindex == idx]).sum())
regvalue = regvalue / float(len(Yidxs)*cfg.trainer.batch_size)
test_loss_reg += regvalue.item()
lossvalue = approxvalue + cfg.trainer.reg_coef*regvalue
test_loss += lossvalue.item()
# tensorboard logger
test_loss = test_loss/float(cfg.trainer.test_epoch_size); tb_writer.add_scalar('Validation/loss', test_loss, i)
test_loss_aprox = test_loss_aprox/float(cfg.trainer.test_epoch_size); tb_writer.add_scalar('Validation/approx_loss', test_loss_aprox, i)
test_loss_reg = test_loss_reg/float(cfg.trainer.test_epoch_size); tb_writer.add_scalar('Validation/diversity_loss', test_loss_reg, i)
# Show predictions for multiple triangles
X_cpu, Y_cpu = X.cpu().numpy(), Y[0].cpu().numpy()
Ypred_cpu, Yindex_cpu = Ypred.cpu().detach().numpy(), Yindex.cpu().numpy()
# plot predictions
num_cols = len(Yidxs) + 1
fig, axis = plt.subplots(4, num_cols, figsize=(10*15,20));
for g in range(4):
pts, ptsGT = Ypred_cpu[0, g], Y_cpu[g]
for s in range(num_cols):
# axis[g, s].plot(X_cpu[g,:,0], X_cpu[g,:,1],'or');
axis[g, s].plot(X_cpu[g,:,0], X_cpu[g,:,1],'or'); axis[g, s].plot(X_cpu[g,[0, 1], 0], X_cpu[g, [0,1], 1], 'r');
axis[g, s].plot(X_cpu[g,[1, 2], 0], X_cpu[g, [1,2], 1], 'r'); axis[g, s].plot(X_cpu[g,[2, 0], 0], X_cpu[g, [2,0], 1], 'r');
if s == 0:
axis[g, s].plot(ptsGT[:, 0], ptsGT[:, 1],'.k');
axis[g, s].set_title("GT_points = {}".format(ptsGT.shape[0]))
else:
axis[g, s].plot(pts[Yindex_cpu == s, 0], pts[Yindex_cpu == s, 1],'.k');
axis[g, s].set_title("Pred Points = {}".format(s))
axis[g, s].set_ylim([0, 1.0]); axis[g, s].set_xlim([0, 1.0]);
tb_writer.add_figure('Predictions', fig, i)
# plot diversity
num_cols = len(Yidxs) + 1
fig, axis = plt.subplots(4, num_cols, figsize=(10*15, 20));
colors = ['tab:blue', 'tab:orange', 'tab:green', 'tab:purple']
for g in range(4):
for c_idx in range(num_cols):
axis[g, c_idx].plot(X_cpu[g,:,0], X_cpu[g,:,1],'or'); axis[g, c_idx].plot(X_cpu[g,[0, 1], 0], X_cpu[g, [0,1], 1], 'r');
axis[g, c_idx].plot(X_cpu[g,[1, 2], 0], X_cpu[g, [1,2], 1], 'r'); axis[g, c_idx].plot(X_cpu[g,[2, 0], 0], X_cpu[g, [2,0], 1], 'r');
axis[g, c_idx].set_ylim([0, 1.0]); axis[g, c_idx].set_xlim([0, 1.0]);
for rep in range(num_of_reps):
pts, ptsGT = Ypred_cpu[rep, g], Y_cpu[g]
if c_idx == 0:
axis[g, c_idx].plot(ptsGT[:, 0], ptsGT[:, 1],'.k');
axis[g, c_idx].set_title("GT_points = {}".format(ptsGT.shape[0]))
break
else:
axis[g, c_idx].scatter(pts[Yindex_cpu == c_idx, 0], pts[Yindex_cpu == c_idx, 1], marker='.', c=colors[rep]);
axis[g, c_idx].set_title("Pred. {} points".format(c_idx))
tb_writer.add_figure('Randomness', fig, i)
# save trained model at every test epoch
save_checkpoint_file = os.path.join(cfg.trainer.output_dir, 'MongeNet_epoch{:04d}_testloss{:.4f}.tar'.format(i, test_loss))
save_checkpoint(model, optimizer, save_checkpoint_file)
logger.info("Checkpoint saved to {}".format(save_checkpoint_file))
# print test progress
logger.info('Test epoch: {} - Test Loss: {:.6f}, Approx Loss: {:.6f}, Diversity Loss: {:.6f} in {:3.4f} sec (including plots and snapshoting)'.format(
i, test_loss, test_loss_aprox, test_loss_reg, tictoc.toc('test_it')))
# save trained model at the final epoch
save_checkpoint_file = os.path.join(cfg.trainer.output_dir, 'MongeNet_epoch{:04d}_testloss{:.4f}.tar'.format(i, test_loss))
save_checkpoint(model, optimizer, save_checkpoint_file)
logger.info("Final Checkpoint saved to {}".format(save_checkpoint_file))
if __name__ == "__main__":
train_app()