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 | def train(config) -> bool:
"""Train the model
Returns:
success (bool): True if training was successful
"""
pl.seed_everything(42)
log_dir = os.path.join(config.data.output)
os.makedirs(log_dir, exist_ok=True)
csv_logger = CSVLogger(save_dir=log_dir, name="logs")
tb_logger = TensorBoardLogger(
save_dir=log_dir, name="logs", version=csv_logger.version
)
lr_monitor = LearningRateMonitor(logging_interval="step")
# Removing stats monitor because it clutters logs and
# doesn't seem to be particularly useful.
# stats_monitor = DeviceStatsMonitor(cpu_stats=True)
logger.info(f"Logging to: {csv_logger.log_dir}")
os.makedirs(csv_logger.log_dir, exist_ok=True)
from omegaconf import OmegaConf
OmegaConf.save(config, os.path.join(csv_logger.log_dir, "pipeline_config.yaml"))
# For convenience
ckpt = config.model.checkpoint
if ckpt == "last":
logger.info("Attempting to find most recent checkpoint")
from glob import glob
checkpoints = glob(os.path.join(log_dir, "*", "*", "checkpoints", "last.ckpt"))
checkpoints = sorted(checkpoints, key=lambda x: os.stat(x).st_ctime)
if len(checkpoints) > 0:
ckpt = checkpoints[-1]
else:
raise FileNotFoundError("No checkpoints were found in the output directory")
if ckpt is not None:
logger.info(f"Attempting to resume from {ckpt}")
if config.model.name == "segformer":
model = SegformerModule(
model=config.model.name,
backbone=config.model.backbone,
ignore_index=None,
id2label=os.path.join(
os.path.dirname(__file__), "index_to_name_binary.json"
),
learning_rate=float(config.model.learning_rate),
learning_rate_schedule_patience=int(
config.model.learning_rate_schedule_patience
),
)
else:
model = SMPModule(
model=config.model.name,
backbone=config.model.backbone,
weights=config.model.pretrained,
in_channels=int(config.model.in_channels),
num_classes=int(config.model.num_classes),
loss=config.model.loss,
ignore_index=None,
learning_rate=float(config.model.learning_rate),
learning_rate_schedule_patience=int(
config.model.learning_rate_schedule_patience
),
)
# Common setup, don't need to do this if only evaluating
model.configure_models(init_pretrained=True if not ckpt else False)
model.configure_losses()
model.configure_metrics()
# load data
datamodule_config = config.model.datamodule
data_module = COCODataModule(
config.data.root,
train_path=config.data.train,
val_path=config.data.validation,
test_path=config.data.validation,
augment=config.model.augment == "on",
batch_size=int(config.model.batch_size),
num_workers=int(config.model.num_workers),
tile_size=int(config.data.tile_size),
)
# checkpoints and loggers
checkpoint_callback = ModelCheckpoint(
monitor="val/multiclassf1score_tree",
mode="max",
auto_insert_metric_name=False,
save_top_k=1,
filename="{epoch}-f1tree:{val/multiclassf1score_tree:.2f}-loss:{val/loss:.2f}",
save_last=True,
verbose=True,
)
batch_size = int(config.model.batch_size)
if batch_size > 32:
accumulate = 1
else:
accumulate = max(1, int(32 / batch_size))
loggers = [tb_logger, csv_logger]
matmul_precision = "medium"
torch.set_float32_matmul_precision(matmul_precision)
trainer_config = config.model.trainer
trainer = pl.Trainer(
callbacks=[checkpoint_callback, lr_monitor],
logger=loggers,
accelerator="gpu" if torch.cuda.is_available() else "cpu",
max_epochs=int(trainer_config.max_epochs),
accumulate_grad_batches=accumulate,
fast_dev_run=trainer_config.debug_run,
devices=1,
)
try:
logger.info("Starting trainer")
trainer.fit(model=model, datamodule=data_module, ckpt_path=ckpt)
# pylint: disable=broad-except
except Exception as e:
logger.error("Training failed")
logger.error(e)
logger.error(traceback.print_exc())
wandb.finish()
exit(1)
try:
logger.info("Train complete, starting test")
trainer.test(model=model, datamodule=data_module, verbose=True)
# pylint: disable=broad-except
except Exception as e:
logger.error("Training failed at test time")
logger.error(e)
logger.error(traceback.print_exc())
wandb.finish()
exit(1)
if config.model.name == "segformer":
# Dump initial config/model so we can load checkpoints later.
if os.path.exists(checkpoint_callback.best_model_path):
logger.info("Saving model state dictionary")
model = SegformerModule.load_from_checkpoint(
checkpoint_callback.best_model_path
)
torch.save(
model.model.state_dict(),
os.path.join(
os.path.dirname(checkpoint_callback.best_model_path), "best.pt"
),
)
wandb.finish()
return True
|