1
|
|
|
import os |
2
|
|
|
from typing import Dict, List, Union |
3
|
|
|
|
4
|
|
|
import wandb |
5
|
|
|
import yaml |
6
|
|
|
|
7
|
|
|
from deepreg import log |
8
|
|
|
from deepreg.config.v011 import parse_v011 |
9
|
|
|
|
10
|
|
|
logger = log.get(__name__) |
11
|
|
|
|
12
|
|
|
|
13
|
|
|
def update_nested_dict(d: Dict, u: Dict) -> Dict: |
14
|
|
|
""" |
15
|
|
|
Merge two dicts. |
16
|
|
|
|
17
|
|
|
https://stackoverflow.com/questions/3232943/update-value-of-a-nested-dictionary-of-varying-depth |
18
|
|
|
|
19
|
|
|
:param d: dict to be overwritten in case of conflicts. |
20
|
|
|
:param u: dict to be merged into d. |
21
|
|
|
:return: |
22
|
|
|
""" |
23
|
|
|
|
24
|
|
|
for k, v in u.items(): |
25
|
|
|
if isinstance(v, dict): |
26
|
|
|
d[k] = update_nested_dict(d.get(k, {}), v) |
27
|
|
|
else: |
28
|
|
|
d[k] = v |
29
|
|
|
return d |
30
|
|
|
|
31
|
|
|
|
32
|
|
|
def load_configs(config_path: Union[str, List[str]]) -> Dict: |
33
|
|
|
""" |
34
|
|
|
Load multiple configs and update the nested dictionary. |
35
|
|
|
|
36
|
|
|
:param config_path: list of paths or one path. |
37
|
|
|
:return: the loaded config |
38
|
|
|
""" |
39
|
|
|
if isinstance(config_path, str): |
40
|
|
|
config_path = [config_path] |
41
|
|
|
# replace ~ with user home path |
42
|
|
|
config_path = [os.path.expanduser(x) for x in config_path] |
43
|
|
|
config: Dict = {} |
44
|
|
|
for config_path_i in config_path: |
45
|
|
|
with open(config_path_i) as file: |
46
|
|
|
config_i = yaml.load(file, Loader=yaml.FullLoader) |
47
|
|
|
config = update_nested_dict(d=config, u=config_i) |
48
|
|
|
loaded_config = config_sanity_check(config) |
49
|
|
|
|
50
|
|
|
if loaded_config != config: |
51
|
|
|
# config got updated |
52
|
|
|
head, tail = os.path.split(config_path[0]) |
53
|
|
|
filename = "updated_" + tail |
54
|
|
|
save(config=loaded_config, out_dir=head, filename=filename) |
55
|
|
|
logger.error( |
56
|
|
|
"The provided configuration file is outdated. " |
57
|
|
|
"An updated version has been saved at %s.", |
58
|
|
|
os.path.join(head, filename), |
59
|
|
|
) |
60
|
|
|
|
61
|
|
|
return loaded_config |
62
|
|
|
|
63
|
|
|
|
64
|
|
|
def save(config: dict, out_dir: str, filename: str = "config.yaml"): |
65
|
|
|
""" |
66
|
|
|
Save the config into a yaml file. |
67
|
|
|
|
68
|
|
|
:param config: configuration to be outputed |
69
|
|
|
:param out_dir: directory of the output file |
70
|
|
|
:param filename: name of the output file |
71
|
|
|
""" |
72
|
|
|
assert filename.endswith(".yaml") |
73
|
|
|
with open(os.path.join(out_dir, filename), "w+") as f: |
74
|
|
|
f.write(yaml.dump(config)) |
75
|
|
|
|
76
|
|
|
|
77
|
|
|
def config_sanity_check(config: dict) -> dict: |
78
|
|
|
""" |
79
|
|
|
Check if the given config satisfies the requirements. |
80
|
|
|
|
81
|
|
|
:param config: entire config. |
82
|
|
|
""" |
83
|
|
|
|
84
|
|
|
# check data |
85
|
|
|
data_config = config["dataset"] |
86
|
|
|
|
87
|
|
|
# back compatibility support |
88
|
|
|
config = parse_v011(config) |
89
|
|
|
|
90
|
|
|
# check model |
91
|
|
|
if config["train"]["method"] == "conditional": |
92
|
|
|
if data_config["labeled"] is False: # unlabeled |
93
|
|
|
raise ValueError( |
94
|
|
|
"For conditional model, data have to be labeled, got unlabeled data." |
95
|
|
|
) |
96
|
|
|
|
97
|
|
|
return config |
98
|
|
|
|
99
|
|
|
|
100
|
|
|
def has_wandb_callback(config: dict): |
101
|
|
|
""" |
102
|
|
|
Function that checks if a given config has W&B |
103
|
|
|
keys. |
104
|
|
|
:param config: config dictionary with parameters for run. |
105
|
|
|
:return: bool, whether wandb key in config. |
106
|
|
|
""" |
107
|
|
|
if "wandb" in config: |
108
|
|
|
return True |
109
|
|
|
return False |
110
|
|
|
|
111
|
|
|
|
112
|
|
|
def instantiate_wandb_run(config: dict): |
113
|
|
|
""" |
114
|
|
|
From a config dictionary with wandb keys, |
115
|
|
|
run wandb.init to log training. |
116
|
|
|
|
117
|
|
|
:param config: config dictionary with parameters for run. |
118
|
|
|
:return: N/A. |
119
|
|
|
""" |
120
|
|
|
if "init" not in config["wandb"]: |
121
|
|
|
logging.error("No init field in config. Creating empty init.") |
|
|
|
|
122
|
|
|
wandb.init() |
123
|
|
|
else: |
124
|
|
|
wandb_init = config["wandb"]["init"] |
125
|
|
|
wandb.init(**wandb_init) |
126
|
|
|
|
127
|
|
|
|
128
|
|
|
def instantiate_wandb_callback(config: dict): |
129
|
|
|
""" |
130
|
|
|
From a config dictionary with wandb keys, |
131
|
|
|
generate a run callback to use during training. |
132
|
|
|
|
133
|
|
|
:param config: config dictionary with parameters for run. |
134
|
|
|
:return: tf.keras.callback, see W&B docs for more info. |
135
|
|
|
""" |
136
|
|
|
# If the callback key does not exist, initialise an |
137
|
|
|
# empty run. |
138
|
|
|
if "callback" not in config["wandb"]: |
139
|
|
|
logging.error("No callback field in config. Creating empty callback.") |
|
|
|
|
140
|
|
|
wandb_callback = wandb.keras.WandbCallback() |
141
|
|
|
# Get sub dict that contains the wandb params |
142
|
|
|
else: |
143
|
|
|
wandb_dict = config["wandb"]["callback"] |
144
|
|
|
wandb_callback = wandb.keras.WandbCallback(**wandb_dict) |
145
|
|
|
return wandb_callback |
146
|
|
|
|