|
1
|
|
|
"""Support backcompatibility for configs at v0.1.1.""" |
|
2
|
|
|
|
|
3
|
|
|
from copy import deepcopy |
|
4
|
|
|
from typing import Dict |
|
5
|
|
|
|
|
6
|
|
|
from deepreg.constant import KNOWN_DATA_SPLITS |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
|
|
def parse_v011(old_config: Dict) -> Dict: |
|
10
|
|
|
""" |
|
11
|
|
|
Transform configuration from V0.1.1 format to the latest format. |
|
12
|
|
|
|
|
13
|
|
|
V0.1.1 to latest. |
|
14
|
|
|
|
|
15
|
|
|
:param old_config: |
|
16
|
|
|
:return: transformed config |
|
17
|
|
|
""" |
|
18
|
|
|
|
|
19
|
|
|
new_config = deepcopy(old_config) |
|
20
|
|
|
|
|
21
|
|
|
new_config["dataset"] = parse_data(data_config=new_config["dataset"]) |
|
22
|
|
|
|
|
23
|
|
|
model_config = new_config["train"].pop("model", None) |
|
24
|
|
|
if model_config is not None: |
|
25
|
|
|
model_config = parse_model(model_config=model_config) |
|
26
|
|
|
new_config["train"].update(model_config) |
|
27
|
|
|
|
|
28
|
|
|
new_config["train"]["loss"] = parse_loss(loss_config=new_config["train"]["loss"]) |
|
29
|
|
|
|
|
30
|
|
|
new_config["train"]["preprocess"] = parse_preprocess( |
|
31
|
|
|
preprocess_config=new_config["train"]["preprocess"] |
|
32
|
|
|
) |
|
33
|
|
|
|
|
34
|
|
|
new_config["train"]["optimizer"] = parse_optimizer( |
|
35
|
|
|
opt_config=new_config["train"]["optimizer"] |
|
36
|
|
|
) |
|
37
|
|
|
|
|
38
|
|
|
return new_config |
|
39
|
|
|
|
|
40
|
|
|
|
|
41
|
|
|
def parse_data(data_config: dict) -> Dict: |
|
42
|
|
|
""" |
|
43
|
|
|
Parse the data configuration. |
|
44
|
|
|
|
|
45
|
|
|
:param data_config: potentially outdated config |
|
46
|
|
|
:return: latest config |
|
47
|
|
|
""" |
|
48
|
|
|
if "format" not in data_config: |
|
49
|
|
|
# up-to-date |
|
50
|
|
|
return data_config |
|
51
|
|
|
|
|
52
|
|
|
# define format and labeled in each split |
|
53
|
|
|
dir = data_config.pop("dir") |
|
54
|
|
|
format = data_config.pop("format") |
|
55
|
|
|
labeled = data_config.pop("labeled") |
|
56
|
|
|
# get dir for each split |
|
57
|
|
|
for split in KNOWN_DATA_SPLITS: |
|
58
|
|
|
if split not in dir: |
|
59
|
|
|
continue |
|
60
|
|
|
data_config[split] = { |
|
61
|
|
|
"dir": dir[split], |
|
62
|
|
|
"format": format, |
|
63
|
|
|
"labeled": labeled, |
|
64
|
|
|
} |
|
65
|
|
|
return data_config |
|
66
|
|
|
|
|
67
|
|
|
|
|
68
|
|
|
def parse_model(model_config: Dict) -> Dict: |
|
69
|
|
|
""" |
|
70
|
|
|
Parse the model configuration. |
|
71
|
|
|
|
|
72
|
|
|
:param model_config: potentially outdated config |
|
73
|
|
|
:return: latest config |
|
74
|
|
|
""" |
|
75
|
|
|
# remove model layer |
|
76
|
|
|
if "model" in model_config: |
|
77
|
|
|
model_config = model_config["model"] |
|
78
|
|
|
|
|
79
|
|
|
if isinstance(model_config["backbone"], dict): |
|
80
|
|
|
# up-to-date |
|
81
|
|
|
return model_config |
|
82
|
|
|
|
|
83
|
|
|
backbone_name = model_config["backbone"] |
|
84
|
|
|
|
|
85
|
|
|
# backbone_config is the backbone name |
|
86
|
|
|
backbone_config = {"name": backbone_name, **model_config[backbone_name]} |
|
87
|
|
|
|
|
88
|
|
|
if backbone_name == "global": |
|
89
|
|
|
# global net use depth instead of extract_levels |
|
90
|
|
|
if "extract_levels" in backbone_config: |
|
91
|
|
|
extract_levels = backbone_config.pop("extract_levels") |
|
92
|
|
|
backbone_config["depth"] = max(extract_levels) |
|
93
|
|
|
|
|
94
|
|
|
model_config = {"method": model_config["method"], "backbone": backbone_config} |
|
95
|
|
|
return model_config |
|
96
|
|
|
|
|
97
|
|
|
|
|
98
|
|
|
def parse_image_loss(loss_config: Dict) -> Dict: |
|
99
|
|
|
""" |
|
100
|
|
|
Parse the image loss part in loss configuration. |
|
101
|
|
|
|
|
102
|
|
|
:param loss_config: potentially outdated config |
|
103
|
|
|
:return: latest config |
|
104
|
|
|
""" |
|
105
|
|
|
if "image" not in loss_config: |
|
106
|
|
|
# no image loss |
|
107
|
|
|
return loss_config |
|
108
|
|
|
|
|
109
|
|
|
if isinstance(loss_config["image"], list): |
|
110
|
|
|
# config up-to-date |
|
111
|
|
|
return loss_config |
|
112
|
|
|
|
|
113
|
|
|
image_loss_name = loss_config["image"]["name"] |
|
114
|
|
|
|
|
115
|
|
|
if image_loss_name not in loss_config["image"]: |
|
116
|
|
|
# config up-to-date |
|
117
|
|
|
return loss_config |
|
118
|
|
|
|
|
119
|
|
|
image_loss_config = { |
|
120
|
|
|
"name": image_loss_name, |
|
121
|
|
|
"weight": loss_config["image"].get("weight", 1.0), |
|
122
|
|
|
} |
|
123
|
|
|
image_loss_config.update(loss_config["image"][image_loss_name]) |
|
124
|
|
|
loss_config["image"] = image_loss_config |
|
125
|
|
|
return loss_config |
|
126
|
|
|
|
|
127
|
|
|
|
|
128
|
|
|
def parse_label_loss(loss_config: Dict) -> Dict: |
|
129
|
|
|
""" |
|
130
|
|
|
Parse the label loss part in loss configuration. |
|
131
|
|
|
|
|
132
|
|
|
:param loss_config: potentially outdated config |
|
133
|
|
|
:return: latest config |
|
134
|
|
|
""" |
|
135
|
|
|
if "label" not in loss_config: |
|
136
|
|
|
# no label loss |
|
137
|
|
|
return loss_config |
|
138
|
|
|
|
|
139
|
|
|
if isinstance(loss_config["label"], list): |
|
140
|
|
|
# config up-to-date |
|
141
|
|
|
return loss_config |
|
142
|
|
|
|
|
143
|
|
|
label_loss_name = loss_config["label"]["name"] |
|
144
|
|
|
if label_loss_name == "single_scale": |
|
145
|
|
|
loss_config["label"] = { |
|
146
|
|
|
"name": loss_config["label"]["single_scale"]["loss_type"], |
|
147
|
|
|
"weight": loss_config["label"].get("weight", 1.0), |
|
148
|
|
|
} |
|
149
|
|
|
elif label_loss_name == "multi_scale": |
|
150
|
|
|
loss_config["label"] = { |
|
151
|
|
|
"name": loss_config["label"]["multi_scale"]["loss_type"], |
|
152
|
|
|
"weight": loss_config["label"].get("weight", 1.0), |
|
153
|
|
|
"scales": loss_config["label"]["multi_scale"]["loss_scales"], |
|
154
|
|
|
} |
|
155
|
|
|
|
|
156
|
|
|
# mean-squared renamed to ssd |
|
157
|
|
|
if loss_config["label"]["name"] == "mean-squared": |
|
158
|
|
|
loss_config["label"]["name"] = "ssd" |
|
159
|
|
|
|
|
160
|
|
|
# dice_generalized merged into dice |
|
161
|
|
|
if loss_config["label"]["name"] == "dice_generalized": |
|
162
|
|
|
loss_config["label"]["name"] = "dice" |
|
163
|
|
|
|
|
164
|
|
|
# rename neg_weight to background_weight |
|
165
|
|
|
if "neg_weight" in loss_config["label"]: |
|
166
|
|
|
background_weight = loss_config["label"].pop("neg_weight") |
|
167
|
|
|
loss_config["label"]["background_weight"] = background_weight |
|
168
|
|
|
|
|
169
|
|
|
return loss_config |
|
170
|
|
|
|
|
171
|
|
|
|
|
172
|
|
|
def parse_reg_loss(loss_config: Dict) -> Dict: |
|
173
|
|
|
""" |
|
174
|
|
|
Parse the regularization loss part in loss configuration. |
|
175
|
|
|
|
|
176
|
|
|
:param loss_config: potentially outdated config |
|
177
|
|
|
:return: latest config |
|
178
|
|
|
""" |
|
179
|
|
|
if "regularization" not in loss_config: |
|
180
|
|
|
# no regularization loss |
|
181
|
|
|
return loss_config |
|
182
|
|
|
|
|
183
|
|
|
if isinstance(loss_config["regularization"], list): |
|
184
|
|
|
# config up-to-date |
|
185
|
|
|
return loss_config |
|
186
|
|
|
|
|
187
|
|
|
if "energy_type" not in loss_config["regularization"]: |
|
188
|
|
|
# up-to-date |
|
189
|
|
|
return loss_config |
|
190
|
|
|
|
|
191
|
|
|
energy_type = loss_config["regularization"]["energy_type"] |
|
192
|
|
|
reg_config = {"weight": loss_config["regularization"].get("weight", 1.0)} |
|
193
|
|
|
if energy_type == "bending": |
|
194
|
|
|
reg_config["name"] = "bending" |
|
195
|
|
|
elif energy_type == "gradient-l2": |
|
196
|
|
|
reg_config["name"] = "gradient" |
|
197
|
|
|
reg_config["l1"] = False |
|
198
|
|
|
elif energy_type == "gradient-l1": |
|
199
|
|
|
reg_config["name"] = "gradient" |
|
200
|
|
|
reg_config["l1"] = True |
|
201
|
|
|
loss_config["regularization"] = reg_config |
|
202
|
|
|
|
|
203
|
|
|
return loss_config |
|
204
|
|
|
|
|
205
|
|
|
|
|
206
|
|
|
def parse_loss(loss_config: Dict) -> Dict: |
|
207
|
|
|
""" |
|
208
|
|
|
Parse the loss configuration. |
|
209
|
|
|
|
|
210
|
|
|
:param loss_config: potentially outdated config |
|
211
|
|
|
:return: latest config |
|
212
|
|
|
""" |
|
213
|
|
|
# remove dissimilarity layer |
|
214
|
|
|
if "dissimilarity" in loss_config: |
|
215
|
|
|
dissim_config = loss_config.pop("dissimilarity") |
|
216
|
|
|
loss_config.update(dissim_config) |
|
217
|
|
|
|
|
218
|
|
|
loss_config = parse_image_loss(loss_config=loss_config) |
|
219
|
|
|
loss_config = parse_label_loss(loss_config=loss_config) |
|
220
|
|
|
loss_config = parse_reg_loss(loss_config=loss_config) |
|
221
|
|
|
|
|
222
|
|
|
return loss_config |
|
223
|
|
|
|
|
224
|
|
|
|
|
225
|
|
|
def parse_preprocess(preprocess_config: Dict) -> Dict: |
|
226
|
|
|
""" |
|
227
|
|
|
Parse the preprocess configuration. |
|
228
|
|
|
|
|
229
|
|
|
:param preprocess_config: potentially outdated config |
|
230
|
|
|
:return: latest config |
|
231
|
|
|
""" |
|
232
|
|
|
if "data_augmentation" not in preprocess_config: |
|
233
|
|
|
preprocess_config["data_augmentation"] = {"name": "affine"} |
|
234
|
|
|
return preprocess_config |
|
235
|
|
|
|
|
236
|
|
|
|
|
237
|
|
|
def parse_optimizer(opt_config: Dict) -> Dict: |
|
238
|
|
|
""" |
|
239
|
|
|
Parse the optimizer configuration. |
|
240
|
|
|
|
|
241
|
|
|
:param opt_config: potentially outdated config |
|
242
|
|
|
:return: latest config |
|
243
|
|
|
""" |
|
244
|
|
|
name = opt_config["name"] |
|
245
|
|
|
if name not in opt_config: |
|
246
|
|
|
# up-to-date |
|
247
|
|
|
return opt_config |
|
248
|
|
|
|
|
249
|
|
|
name_dict = dict( |
|
250
|
|
|
adam="Adam", |
|
251
|
|
|
sgd="SGD", |
|
252
|
|
|
rms="RMSprop", |
|
253
|
|
|
) |
|
254
|
|
|
new_name = name_dict[name] |
|
255
|
|
|
return {"name": new_name, **opt_config[name]} |
|
256
|
|
|
|
DeepRegNet /
DeepReg
