-
Notifications
You must be signed in to change notification settings - Fork 2
/
trainer.py
185 lines (140 loc) · 9.93 KB
/
trainer.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
import time
import torch, torchvision
import wandb
import numpy as np
from tqdm.auto import tqdm
from cut.cut_trainer import CUTTrainer
import helpers
from models.us_rendering_model import UltrasoundRendering
class Trainer:
def __init__(self, hparams, opt_cut, plotter):
self.hparams = hparams
self.opt_cut = opt_cut
self.plotter = plotter
ModuleClass = helpers.load_module_class(hparams.module)
InnerModelClass = helpers.load_model_class(hparams.inner_model)
self.inner_model = InnerModelClass(params=hparams)
if not hparams.outer_model_monai:
OuterModelClass = helpers.load_model_class(hparams.outer_model)
outer_model = OuterModelClass(hparams=hparams)
self.module = ModuleClass(params=hparams, inner_model=self.inner_model, outer_model=outer_model)
else:
self.module = ModuleClass(params=hparams, inner_model=self.inner_model)
self.real_us_train_loader, dataset_real_us = helpers.load_real_us_training_data(opt_cut)
self.cut_trainer = CUTTrainer(opt_cut, dataset_real_us, self.real_us_train_loader)
self.USRendereDefParams = UltrasoundRendering(params=hparams, default_param=True).to(hparams.device)
# ------------------------------------------------------------------------------------------------------------------------------
# Train US Renderer + SEG NET + CUT
# ------------------------------------------------------------------------------------------------------------------------------
def train_step(self, train_loader_ct_labelmaps, epoch):
self.module.train()
self.module.outer_model.train()
self.inner_model.train()
iter_data_time = time.time() # timer for data loading per iteration
dataloader_real_us_iterator = iter(self.real_us_train_loader)
train_losses = []
for i, batch_data_ct in tqdm(enumerate(train_loader_ct_labelmaps), total=len(train_loader_ct_labelmaps), ncols= 100):
self.module.optimizer.zero_grad()
input, label, _ = self.module.get_data(batch_data_ct)
# if 1 in label:
if (label != 0).any(): # if not empty label
us_rendr = self.module.rendering_forward(input)
us_rendr_cut = us_rendr.clone().detach()
self.cut_trainer.train_cut(self.module, us_rendr_cut, epoch, dataloader_real_us_iterator, iter_data_time)
#if using the identity image from CUT instead of us_rendr directly
if self.hparams.use_idtB:
idt_B = self.cut_trainer.get_idtB(us_rendr)
us_rendr = idt_B
#add augmentations during training
if self.hparams.seg_net_input_augmentations_noise_blur or self.hparams.seg_net_input_augmentations_rand_crop:
us_rendr, label = helpers.add_online_augmentations(self.hparams, self.module, us_rendr, label)
loss, _ = self.module.seg_forward(us_rendr, label)
if self.hparams.inner_model_learning_rate == 0:
self.cut_trainer.cut_model.set_requires_grad(self.module.USRenderingModel, False)
loss.backward()
# Perform gradient clipping
if self.hparams.grad_clipping:
torch.nn.utils.clip_grad_norm_(self.module.USRenderingModel.parameters(), max_norm=1)
torch.nn.utils.clip_grad_norm_(self.module.outer_model.parameters(), max_norm=1)
self.module.optimizer.step()
if self.hparams.logging: wandb.log({"train_loss_step": loss.item()})
train_losses.append(loss.item())
if self.hparams.logging: self.plotter.log_us_rendering_values(self.module.USRenderingModel, i)
if self.hparams.scheduler:
self.module.scheduler.step()
wandb.log({"lr_": self.module.optimizer.param_groups[0]["lr"]})
wandb.log({"lr_2": self.module.optimizer.param_groups[1]["lr"]})
# Plot CUT results during training
if len(self.cut_trainer.cut_plot_figs) > 0:
self.plotter.log_image(torchvision.utils.make_grid(self.cut_trainer.cut_plot_figs), "real_A|fake_B|real_B|idt_B")
self.cut_trainer.cut_plot_figs = []
train_loss = np.average(train_losses)
return train_loss
def evaluate(self, val_loader_ct_labelmaps, epoch):
self.module.eval()
self.module.outer_model.eval()
self.inner_model.eval()
self.cut_trainer.cut_model.eval()
self.USRendereDefParams.eval()
def_renderer_plot_figs = []
print(f"--------------- VALIDATION SEG NET ------------")
valid_losses = []
with torch.no_grad():
for nr, val_batch_data_ct in tqdm(enumerate(val_loader_ct_labelmaps), total=len(val_loader_ct_labelmaps), ncols= 100):
val_input, val_label, filename = self.module.get_data(val_batch_data_ct)
if (val_label != 0).any():
val_input_copy = val_input.clone().detach()
us_rendr_val = self.module.rendering_forward(val_input)
if self.hparams.use_idtB:
idt_B_val = self.cut_trainer.get_idtB(us_rendr_val)
us_rendr_val = idt_B_val
val_loss_step, rendered_seg_pred = self.module.seg_forward(us_rendr_val, val_label)
valid_losses.append(val_loss_step.item())
if self.hparams.log_default_renderer and nr < self.hparams.nr_imgs_to_plot:
us_rendr_def = self.USRendereDefParams(val_input_copy.squeeze())
if not self.hparams.use_idtB: idt_B_val = us_rendr_val
plot_fig = self.plotter.plot_stopp_crit(caption="labelmap|default_renderer|learnedUS|seg_input|seg_pred|gt",
imgs=[val_input, us_rendr_def, us_rendr_val, rendered_seg_pred, val_label],
img_text='', epoch=epoch, plot_single=False)
def_renderer_plot_figs.append(plot_fig)
elif not self.hparams.log_default_renderer:
dict = self.module.plot_val_results(val_input, val_loss_step, filename, val_label, rendered_seg_pred, us_rendr_val, epoch)
self.plotter.validation_batch_end(dict)
if len(def_renderer_plot_figs) > 0:
if self.hparams.logging:
self.plotter.log_image(torchvision.utils.make_grid(def_renderer_plot_figs), "default_renderer|labelmap|defaultUS|learnedUS|seg_pred|gt")
avg_valid_loss = np.average(valid_losses)
print(f"--------------- END SEG NETWORK VALIDATION ------------")
return avg_valid_loss
# ------------------------------------------------------------------------------------------------------------------------------
# (GT STOPPING CRITERION) INFER REAL_US_STOPP_CRIT_TEST_SET_IMGS THROUGH THE CUT+SEG NET for stopping the end2end training
# -------------------------------------------------------------------------------------------------------------------------------
def stopp_crit_check(self, real_us_stopp_crit_dataloader, epoch, early_stopping):
stoppcrit_losses = []
self.cut_trainer.cut_model.eval()
avg_stopp_crit_loss = 1
if self.hparams.stopping_crit_gt_imgs and epoch > self.hparams.epochs_check_stopp_crit:
print(f"--------------- STOPPING CRITERIA US IMGS CHECK ------------")
stopp_crit_imgs_plot_figs = []
with torch.no_grad():
for nr, batch_data_real_us_stopp_crit in tqdm(enumerate(real_us_stopp_crit_dataloader), total=len(real_us_stopp_crit_dataloader), ncols= 100, position=0, leave=True):
real_us_stopp_crit_img, real_us_stopp_crit_label = batch_data_real_us_stopp_crit[0].to(self.hparams.device), batch_data_real_us_stopp_crit[1].to(self.hparams.device).float()
reconstructed_us_stopp_crit = self.cut_trainer.cut_model.netG(real_us_stopp_crit_img)
reconstructed_us_stopp_crit = (reconstructed_us_stopp_crit / 2 ) + 0.5 # from [-1,1] to [0,1]
stoppcrit_loss, seg_pred_stopp_crit = self.module.seg_forward(reconstructed_us_stopp_crit, real_us_stopp_crit_label)
stoppcrit_losses.append(stoppcrit_loss)
if self.hparams.logging:
real_us_stopp_crit_img = (real_us_stopp_crit_img / 2 ) + 0.5 # from [-1,1] to [0,1]
wandb.log({"stoppcrit_loss": stoppcrit_loss.item()})
plot_fig_gt = self.plotter.plot_stopp_crit(caption="stopp_crit|real_us|reconstructed_us|seg_pred|gt_label",
imgs=[real_us_stopp_crit_img, reconstructed_us_stopp_crit, seg_pred_stopp_crit, real_us_stopp_crit_label],
img_text='loss=' + "{:.4f}".format(stoppcrit_loss.item()), epoch=epoch, plot_single=False)
stopp_crit_imgs_plot_figs.append(plot_fig_gt)
if len(stopp_crit_imgs_plot_figs) > 0:
self.plotter.log_image(torchvision.utils.make_grid(stopp_crit_imgs_plot_figs), "stopp_crit|real_us|reconstructed_us|seg_pred|gt_label")
avg_stopp_crit_loss = torch.mean(torch.stack(stoppcrit_losses))
wandb.log({"stoppcrit_loss_epoch": avg_stopp_crit_loss, "epoch": epoch})
stoppcrit_losses = []
stopp_crit_imgs_plot_figs = []
if epoch > self.hparams.min_epochs:
early_stopping(avg_stopp_crit_loss, epoch, self.module, self.cut_trainer.cut_model.netG)