-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathsupervised_utils.py
176 lines (141 loc) · 7.29 KB
/
supervised_utils.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
import os
import torch
import torchvision
import torchvision.transforms as TF
from supervised_models import CLIPClassifier
def create_save_path_sl(cfg, root_folder):
# return path to folder where metrics will be saved that includes current datetime
# get current datetime as string
import datetime
now = datetime.datetime.now()
now_str = now.strftime("%Y-%m-%d_%H-%M-%S")
# get scientific notation of lr
lr_str = "{:.0e}".format(cfg["lr"])
path = f"{root_folder}/{now_str}_{cfg['model_name'].replace('/','_')}_{lr_str}_{str(cfg['dataset_size']).replace('.', '_')}_{cfg['seed']}/"
os.makedirs(path, exist_ok=True)
# save cfg there as json
import json
with open(path + "cfg.json", "w") as f:
json.dump(cfg, f)
return path
def apply_training(base_model, data_module, cfg, verbose=True):
base_model.label_names = data_module.label_names
# import here because we need to set the gpu before importing torch
root_folder = "/raid/8wiehe/"
from learning_utils import init_trainer#, init_test_dms
from supervised_models import LitCLIP
lit_model = LitCLIP(base_model, data_module.label_names,
cfg["max_epochs"], cfg["lr"], data_module.steps_per_epoch,
weight_decay=cfg["weight_decay"], use_pos_weight=cfg["use_pos_weight"],
pos_fraction=data_module.pos_fraction, eps=cfg["eps"], beta1=cfg["beta1"], beta2=cfg["beta2"])
#cfg["val_check_interval"] = min(int(cfg["val_check_interval"] * (32 / cfg["batch_size"])), len(data_module.train_dataloader()))
if verbose:
cfg["val_check_interval"] = 0.1 if cfg["dataset_size"] > 0.1 else 0.5
else:
cfg["val_check_interval"] = data_module.steps_per_epoch
trainer = init_trainer(root_folder, cfg, "early_tests", num_sanity_val_steps=0)
trainer.fit(lit_model, data_module)
import wandb
wandb.finish(quiet=not verbose)
return lit_model
def train_sl_model(cfg):
# to fix pytorch issue with "Too many files open"
import torch.multiprocessing
torch.multiprocessing.set_sharing_strategy('file_system')
# seed everything using pytorch lightning
import pytorch_lightning
pytorch_lightning.seed_everything(cfg["seed"])
# init model
#from supervised_utils import init_model
base_model, transform_basic, transform_aug, name = init_model(cfg)
from clip_utils import FinetuneDataModule
data_module = FinetuneDataModule(base_model, transform_basic,
dataset_name=cfg["dataset_name"],
mode=cfg["mode"],
batch_size=cfg["batch_size"],
augs=transform_aug,
use_cl=False,
dataset_size=cfg["dataset_size"],)
lit_model = apply_training(base_model, data_module, cfg)
return lit_model, data_module
def make_224_transforms(rot_aug, shift_aug, scale_aug, normalize, size):
transform_basic = TF.Compose([TF.Resize(size=size,
interpolation=TF.InterpolationMode.BILINEAR),
TF.CenterCrop(size=(size, size)),
TF.ToTensor(),
normalize])
transform_aug = TF.Compose([TF.Resize(size=size,
interpolation=TF.InterpolationMode.BILINEAR),
TF.CenterCrop(size=(size, size)),
TF.RandomAffine(rot_aug,
translate=(shift_aug, shift_aug),
scale=(1.0 - scale_aug, 1.0 + scale_aug)),
TF.ToTensor(),
normalize])
return transform_basic, transform_aug
def init_model(cfg):
pretrained = cfg["pretrained"]
#convert_models_to_fp32(model)
num_labels = 14 #data_module.num_labels
if cfg["model_name"] == "densenet_224_cxr":
name = "densenet_224"
densenet_size = 224
import torchxrayvision as xrv
model = xrv.models.DenseNet(weights="densenet121-res224-mimic_ch")
labels_to_remove = ["No finding", "Support devices", "Pleural other"]
# Use XRV transforms to crop and resize the images
transform_basic = TF.Compose([xrv.datasets.ToPILImage(),
xrv.datasets.XRayCenterCrop(),
xrv.datasets.XRayResizer(densenet_size)])
transform_aug = TF.Compose([
xrv.datasets.ToPILImage(),
TF.RandomAffine(cfg["rot_aug"],
translate=(cfg["shift_aug"], cfg["shift_aug"]),
scale=(1.0 - cfg["scale_aug"], 1.0 + cfg["scale_aug"])),
TF.ToTensor()
])
elif "densenet" in cfg["model_name"]:
if cfg["model_name"] == "densenet_256":
name = "densenet_256"
densenet_size = 256
model = torchvision.models.densenet121(pretrained=pretrained)
elif cfg["model_name"] == "densenet_224":
name = "densenet_224"
densenet_size = 224
import torchxrayvision as xrv
model = xrv.models.DenseNet(weights="densenet121-res224-mimic_ch")
# get model and re-init out layer
model.classifier = torch.nn.Linear(1024, num_labels)
model.num_labels = num_labels
# create transform
normalize = TF.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
transform_basic, transform_aug = make_224_transforms(cfg["rot_aug"],
cfg["shift_aug"],
cfg["scale_aug"],
normalize,
224)
else:
from clip_utils import load_clip
clip_base_model, transform, name = load_clip(cfg["model_name"], cfg["mode"], device="cpu",
down_sample_size=cfg["down_sample_size"],
adapter_flow=cfg["adapter_flow"])
if "ViT" in cfg["model_name"]:
size = 224
elif cfg["model_name"] == "RN50x4":
size = 288
elif cfg["model_name"] == "RN50" or cfg["model_name"] == "RN101":
size = 224
else:
size = 224
if not pretrained:
clip_base_model.initialize_parameters()
model = CLIPClassifier(clip_base_model, cfg["mode"], num_labels)
normalize = transform.transforms[-1]
transform_basic, transform_aug = make_224_transforms(cfg["rot_aug"],
cfg["shift_aug"],
cfg["scale_aug"],
normalize,
size)
model.name = name
return model, transform_basic, transform_aug, name