-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrainer.py
242 lines (196 loc) · 9.63 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
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
import pytorch_lightning as pl
import torchvision
from models.solo_head import SOLOHead
from pytorch_lightning import loggers as pl_loggers
from torch.optim.lr_scheduler import MultiStepLR
import torch
import datasets.solo_dataset as dataset
import matplotlib.pyplot as plt
from models.backbone import Resnet50Backbone
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
class SOLOTrainer(pl.LightningModule):
_default_cfg = {
'num_classes': 4,
'in_channels': 256,
'seg_feat_channels': 256,
'stacked_convs': 7,
'strides': [8, 8, 16, 32, 32],
'scale_ranges': [(1, 96), (48, 192), (96, 384), (192, 768), (384, 2048)],
'epsilon': 0.2,
'num_grids': [40, 36, 24, 16, 12],
'mask_loss_cfg': dict(weight=3),
'cate_loss_cfg': dict(gamma=2, alpha=0.25, weight=1),
'postprocess_cfg': dict(cate_thresh=0.2, mask_thresh=0.5, pre_NMS_num=50, keep_instance=5, IoU_thresh=0.5)
}
def __init__(self, **kwargs):
super().__init__()
for k, v in {**self._default_cfg, **kwargs}.items():
setattr(self, k, v)
pretrained_model = torchvision.models.detection.maskrcnn_resnet50_fpn(pretrained=False, pretrained_backbone=True)
self.backbone = pretrained_model.backbone
self.solo_head = SOLOHead(self.num_classes, device)
self.train_outputs = []
self.train_loss_epoch = []
self.train_cat_loss_epoch = []
self.train_msk_loss_epoch = []
self.validation_outputs = []
self.val_loss_epoch = []
self.val_cat_loss_epoch = []
self.val_msk_loss_epoch = []
# Forward function should calculate across each level of the feature pyramid network.
# Input:
# images: batch_size number of images
# Output:
# if eval = False
# category_predictions: list, len(fpn_levels), each (batch_size, C-1, S, S)
# mask_predictions: list, len(fpn_levels), each (batch_size, S^2, 2*feature_h, 2*feature_w)
# if eval==True
# category_predictions: list, len(fpn_levels), each (batch_size, S, S, C-1)
# / after point_NMS
# mask_predictions: list, len(fpn_levels), each (batch_size, S^2, image_h/4, image_w/4)
# / after upsampling
def forward(self, images, eval=True):
# you can modify this if you want to train the backbone
feature_pyramid = [v.detach() for v in self.backbone(images).values()] # this has strides [4,8,16,32,64]
cate_pred_list, ins_pred_list = self.solo_head(feature_pyramid, eval=eval)
return cate_pred_list, ins_pred_list
def training_step(self, batch, batch_idx):
images, labels, masks, bounding_boxes = batch
images = images.to(device)
cat_pred, msk_pred = self.forward(images, eval=False)
mask_targets, active_masks, category_targets = self.solo_head.target(msk_pred,bounding_boxes, labels, masks)
cat_loss, msk_loss, train_loss = self.solo_head.loss(cat_pred, msk_pred, mask_targets, active_masks, category_targets)
# total_memory = torch.cuda.get_device_properties(0).total_memory
# reserved_memory = torch.cuda.memory_reserved(0)
# allocated_memory = torch.cuda.memory_allocated(0)
# free_memory = reserved_memory - allocated_memory
# Print memory details
# print(f"Device: {device}")
# print(f"Total Memory: {total_memory / (1024 ** 3):.2f} GB")
# print(f"Available Memory: {free_memory / (1024 ** 3):.2f} GB")
# del images, labels, masks, bounding_boxes
# del mask_targets, category_targets, active_masks
torch.cuda.empty_cache()
# total_memory = torch.cuda.get_device_properties(0).total_memory
# reserved_memory = torch.cuda.memory_reserved(0)
# allocated_memory = torch.cuda.memory_allocated(0)
# free_memory = reserved_memory - allocated_memory
# # Print memory details
# print(f"Device: {device}")
# print(f"Total Memory: {total_memory / (1024 ** 3):.2f} GB")
# print(f"Available Memory: {free_memory / (1024 ** 3):.2f} GB")
self.log("train_loss", train_loss.detach().cpu().item(), prog_bar=True, on_epoch=True)
self.log("mask_loss", msk_loss.detach().cpu().item(), prog_bar=True)
self.log("category loss", cat_loss.detach().cpu().item(), prog_bar=True)
self.train_outputs.append({"loss":train_loss.detach().cpu().item(), "msk_loss":msk_loss.detach().cpu().item(), "cat_loss":cat_loss.cpu().item()})
return train_loss
def on_train_epoch_end(self):
avg_train_loss = 0
avg_cat_loss = 0
avg_mask_loss = 0
for i in range(len(self.train_outputs)):
avg_train_loss += self.train_outputs[i]["loss"]
avg_cat_loss += self.train_outputs[i]["cat_loss"]
avg_mask_loss += self.train_outputs[i]["msk_loss"]
self.train_loss_epoch.append(avg_train_loss)
self.train_cat_loss_epoch.append(avg_cat_loss)
self.train_msk_loss_epoch.append(avg_mask_loss)
self.train_outputs = []
def validation_step(self, batch, batch_idx):
images, labels, masks, bounding_boxes = batch
images = images.to(device)
cat_pred, msk_pred = self.forward(images, eval=False)
mask_targets, active_masks, category_targets = self.solo_head.target(msk_pred,bounding_boxes, labels, masks)
cat_loss, msk_loss, val_loss = self.solo_head.loss(cat_pred, msk_pred, mask_targets, active_masks, category_targets)
#del images, labels, masks, bounding_boxes
#del mask_targets, category_targets, active_masks
torch.cuda.empty_cache()
self.log("val_loss", val_loss.detach().cpu().item(),prog_bar=True,on_epoch=True)
self.validation_outputs.append({"loss":val_loss.detach().cpu().item(), "msk_loss":msk_loss.detach().cpu().item(), "cat_loss":cat_loss.detach().cpu().item()})
def on_validation_epoch_end(self):
avg_train_loss = 0
avg_cat_loss = 0
avg_mask_loss = 0
for i in range(len(self.validation_outputs)):
avg_train_loss += self.validation_outputs[i]["loss"]
avg_cat_loss += self.validation_outputs[i]["cat_loss"]
avg_mask_loss += self.validation_outputs[i]["msk_loss"]
self.val_loss_epoch.append(avg_train_loss)
self.val_cat_loss_epoch.append(avg_cat_loss)
self.val_msk_loss_epoch.append(avg_mask_loss)
self.validation_outputs = []
def configure_optimizers(self):
optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.9, weight_decay=0.0001)
scheduler = MultiStepLR(optimizer, milestones=[27,33], gamma=0.1)
return optimizer
if __name__ == "__main__":
imgs_path = './hw3/data/hw3_mycocodata_img_comp_zlib.h5'
masks_path = './hw3/data/hw3_mycocodata_mask_comp_zlib.h5'
labels_path = './hw3/data/hw3_mycocodata_labels_comp_zlib.npy'
bboxes_path = './hw3/data/hw3_mycocodata_bboxes_comp_zlib.npy'
paths = [imgs_path, masks_path, labels_path, bboxes_path]
# load the data into data.Dataset
dataset_solo = dataset.BuildDataset(paths)
## Visualize debugging
# --------------------------------------------
# build the dataloader
# set 20% of the dataset as the training data
full_size = len(dataset_solo)
train_size = int(full_size * 0.8)
test_size = full_size - train_size
# random split the dataset into training and testset
# set seed
torch.random.manual_seed(11)
train_dataset, test_dataset = torch.utils.data.random_split(dataset_solo, [train_size, test_size])
# push the randomized training data into the dataloader
batch_size = 1
train_build_loader = dataset.BuildDataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=3)
train_loader = train_build_loader.loader()
test_build_loader = dataset.BuildDataLoader(test_dataset, batch_size=batch_size, shuffle=False, num_workers=3)
test_loader = test_build_loader.loader()
model = SOLOTrainer()
#model.backbone.load_state_dict(torch.load("./hw3/backbone.pth"))
#model.solo_head.load_state_dict(torch.load("./hw3/solo.pth"))
epochs = 5
logger = pl_loggers.TensorBoardLogger("tb_logs", name="SOLO")
# trainer = pl.Trainer(accelerator='cpu', devices=1, max_epochs=epochs, logger=logger)
trainer = pl.Trainer(accelerator='gpu', devices=1, max_epochs=epochs, logger=logger)
trainer.fit(model, train_loader, test_loader)
torch.save(model.solo_head.state_dict(), "./solo_head.pth")
torch.save(model.backbone.state_dict(), "./backbone.pth")
tot_cat_train_losses = model.train_cat_loss_epoch
plt.plot(tot_cat_train_losses)
plt.title("Training Category loss (Focal loss) per epoch")
plt.xlabel("Epochs")
plt.ylabel("Loss")
plt.show()
tot_msk_train_losses = model.train_msk_loss_epoch
plt.plot(tot_msk_train_losses)
plt.title("Training Mask loss (Dice loss) per epoch")
plt.xlabel("Epochs")
plt.ylabel("Loss")
plt.show()
tot_train_losses = model.train_loss_epoch
plt.plot(tot_train_losses)
plt.title("Training loss per epoch")
plt.xlabel("Epochs")
plt.ylabel("Loss")
plt.show()
tot_cat_val_losses = model.val_cat_loss_epoch
plt.plot(tot_cat_val_losses)
plt.title("Validation Category loss (Focal loss) per epoch")
plt.xlabel("Epochs")
plt.ylabel("Loss")
plt.show()
tot_msk_val_losses = model.val_msk_loss_epoch
plt.plot(tot_msk_val_losses)
plt.title("Validation Mask loss (Dice loss) per epoch")
plt.xlabel("Epochs")
plt.ylabel("Loss")
plt.show()
tot_val_losses = model.val_loss_epoch
plt.plot(tot_val_losses)
plt.title("Validation loss per epoch")
plt.xlabel("Epochs")
plt.ylabel("Loss")
plt.show()