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
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380 | class TrainConfig(ConfigBase):
"""The training configuration."""
# data params
data_dir: Path
metadata_fp: Path
dataset_stats_fp: Path
cv_split: int = 0
bands: List[str]
spectral_indices: List[str]
sahi: bool = False
image_size: int = 352
resize_strategy: Literal["pad", "resize"] = "pad"
interpolation: Literal["nearest", "nearest-exact", "bilinear", "bicubic"] = "nearest"
batch_size: int = 32
num_workers: int = 4
normalization_strategy: Literal[
"min-max",
"z-score",
"quantile",
"per-sample-quantile",
"per-sample-min-max",
] = "quantile"
fill_missing_pixels_with_torch_nan: bool = False
mask_using_qa: bool = False
mask_using_water_mask: bool = False
use_weighted_sampler: bool = False
samples_per_epoch: int = 10240
has_kelp_importance_factor: float = 3.0
kelp_pixels_pct_importance_factor: float = 0.2
qa_ok_importance_factor: float = 0.0
qa_corrupted_pixels_pct_importance_factor: float = -1.0
almost_all_water_importance_factor: float = 0.5
dem_nan_pixels_pct_importance_factor: float = 0.25
dem_zero_pixels_pct_importance_factor: float = -1.0
# model params
architecture: Literal[
"deeplabv3",
"deeplabv3+",
"efficientunet++",
"fcn",
"fpn",
"linknet",
"manet",
"pan",
"pspnet",
"resunet",
"resunet++",
"unet",
"unet++",
] = "unet"
encoder: str = "tu-efficientnet_b5"
encoder_weights: Optional[str] = None
encoder_depth: int = 5
decoder_channels: List[int] = [256, 128, 64, 32, 16]
decoder_attention_type: Optional[str] = None
pretrained: bool = False
num_classes: int = 2
ignore_index: Optional[int] = None
# optimizer params
optimizer: Literal["adam", "adamw", "sgd"] = "adamw"
weight_decay: float = 1e-4
# lr scheduler params
lr_scheduler: Optional[
Literal[
"onecycle",
"cosine",
"cosine_with_warm_restarts",
"cyclic",
"reduce_lr_on_plateau",
"none",
]
] = None
lr: float = 3e-4
onecycle_pct_start: float = 0.1
onecycle_div_factor: float = 2.0
onecycle_final_div_factor: float = 1e2
cyclic_base_lr: float = 1e-5
cyclic_mode: Literal["triangular", "triangular2", "exp_range"] = "exp_range"
cosine_eta_min: float = 1e-7
cosine_T_mult: int = 2
reduce_lr_on_plateau_factor: float = 0.95
reduce_lr_on_plateau_patience: int = 2
reduce_lr_on_plateau_threshold: float = 1e-4
reduce_lr_on_plateau_min_lr: float = 1e-6
# loss params
objective: Literal["binary", "multiclass"] = "binary"
loss: Literal[
"ce",
"jaccard",
"dice",
"tversky",
"focal",
"lovasz",
"soft_ce",
"xedice",
"focal_tversky",
"log_cosh_dice",
"hausdorff",
"t_loss",
"combo",
"exp_log_loss",
"soft_dice",
"batch_soft_dice",
] = "dice"
ce_smooth_factor: float = 0.0
ce_class_weights: Optional[List[float]] = None
# compile/ort params
compile: bool = False
compile_mode: Literal["default", "reduce-overhead", "max-autotune", "max-autotune-no-cudagraphs"] = "default"
compile_dynamic: Optional[bool] = None
ort: bool = False
# eval loop extra params
plot_n_batches: int = 3
tta: bool = False
tta_merge_mode: Literal["min", "max", "mean", "gmean", "sum", "tsharpen"] = "max"
decision_threshold: Optional[float] = None
# callback params
save_top_k: int = 1
monitor_metric: str = "val/dice"
monitor_mode: Literal["min", "max"] = "max"
early_stopping_patience: int = 10
swa: bool = False
swa_epoch_start: float = 0.75
swa_annealing_epochs: int = 10
swa_lr: float = 3e-5
# trainer params
precision: Literal[
"16-true",
"16-mixed",
"bf16-true",
"bf16-mixed",
"32-true",
] = "bf16-mixed"
fast_dev_run: bool = False
epochs: int = 1
limit_train_batches: Optional[Union[int, float]] = None
limit_val_batches: Optional[Union[int, float]] = None
limit_test_batches: Optional[Union[int, float]] = None
log_every_n_steps: int = 50
accumulate_grad_batches: int = 1
val_check_interval: Optional[float] = None
benchmark: bool = False
# misc
experiment: str = "kelp-seg-training-exp"
output_dir: Path
seed: int = 42
@model_validator(mode="before")
def validate_encoder(cls, values: Dict[str, Any]) -> Dict[str, Any]:
if values["pretrained"] and values["encoder"].startswith("tu-"):
import timm
encoder = values["encoder"].replace("tu-", "")
if not any(e.startswith(encoder) for e in timm.list_pretrained()):
_logger.warning(f"No pretrained weights exist for tu-{encoder}. Forcing training with random init.")
values["pretrained"] = False
values["encoder_weights"] = None
for img_size, bs in zip([224, 256, 336, 384, 448, 512], [32, 32, 32, 32, 4, 4]):
if f"{img_size}" in values["encoder"] and values["image_size"] != img_size:
_logger.warning(f"Encoder requires image_size={img_size}. Forcing training with adjusted image size.")
values["image_size"] = img_size
values["resize_strategy"] = "resize"
values["batch_size"] = min(bs, values["batch_size"])
values["accumulate_grad_batches"] = max(1, 32 // bs)
_logger.info(
f"Adjusted image_size={values['image_size']}, "
f"resize_strategy={values['resize_strategy']}, "
f"batch_size={values['batch_size']}, "
f"accumulate_grad_batches={values['accumulate_grad_batches']}"
)
break
channels_item = values.get("decoder_channels", "256,128,64,32,16")
channels = (
channels_item
if isinstance(channels_item, list)
else [int(index.strip()) for index in channels_item.split(",")]
)
values["decoder_channels"] = channels
values["encoder_depth"] = len(channels)
return values
@field_validator("bands", mode="before")
def validate_bands(cls, value: Optional[Union[str, List[str]]] = None) -> List[str]:
all_bands = list(consts.data.ORIGINAL_BANDS)
if value is None:
return all_bands
bands = value if isinstance(value, list) else [band.strip() for band in value.split(",")]
if set(bands).issubset(all_bands):
return bands
raise ValueError(f"{bands=} should be a subset of {all_bands=}")
@field_validator("spectral_indices", mode="before")
def validate_spectral_indices(cls, value: Union[str, Optional[List[str]]] = None) -> List[str]:
if not value:
return []
indices = value if isinstance(value, list) else [index.strip() for index in value.split(",")]
unknown_indices = set(indices).difference(list(SPECTRAL_INDEX_LOOKUP.keys()))
if unknown_indices:
raise ValueError(
f"Unknown spectral indices were provided: {', '.join(unknown_indices)}. "
f"Please provide at most 5 comma separated indices: {', '.join(SPECTRAL_INDEX_LOOKUP.keys())}."
)
return indices
@field_validator("ce_class_weights", mode="before")
def validate_ce_class_weights(
cls,
value: Union[Optional[str], Optional[List[float]]] = None,
) -> Optional[List[float]]:
if not value:
return None
weights = value if isinstance(value, list) else [float(index.strip()) for index in value.split(",")]
if len(weights) != consts.data.NUM_CLASSES:
raise ValueError(
f"Please provide provide per-class weights! There should be {consts.data.NUM_CLASSES} "
f"floating point numbers. You provided {len(weights)}"
)
return weights
@field_validator("lr_scheduler", mode="before")
def validate_lr_scheduler(cls, value: Optional[str] = None) -> Optional[str]:
return None if value is None or value == "none" else value
@property
def resolved_experiment_name(self) -> str:
return os.environ.get("MLFLOW_EXPERIMENT_NAME", self.experiment)
@property
def run_id_from_context(self) -> Optional[str]:
return os.environ.get("MLFLOW_RUN_ID", None)
@property
def tags(self) -> Dict[str, Any]:
return {"trained_at": datetime.utcnow().isoformat()}
@property
def fill_value(self) -> float:
return torch.nan if self.fill_missing_pixels_with_torch_nan else 0.0 # type: ignore[no-any-return]
@property
def dataset_stats(self) -> Dict[str, Dict[str, float]]:
return json.loads(self.dataset_stats_fp.read_text()) # type: ignore[no-any-return]
@property
def data_module_kwargs(self) -> Dict[str, Any]:
return {
"data_dir": self.data_dir,
"metadata_fp": self.metadata_fp,
"dataset_stats": self.dataset_stats,
"cv_split": self.cv_split,
"bands": self.bands,
"spectral_indices": self.spectral_indices,
"image_size": self.image_size,
"resize_strategy": self.resize_strategy,
"sahi": self.sahi,
"interpolation": self.interpolation,
"batch_size": self.batch_size,
"num_workers": self.num_workers,
"normalization_strategy": self.normalization_strategy,
"missing_pixels_fill_value": self.fill_value,
"mask_using_qa": self.mask_using_qa,
"mask_using_water_mask": self.mask_using_water_mask,
"use_weighted_sampler": self.use_weighted_sampler,
"samples_per_epoch": self.samples_per_epoch,
"has_kelp_importance_factor": self.has_kelp_importance_factor,
"kelp_pixels_pct_importance_factor": self.kelp_pixels_pct_importance_factor,
"qa_ok_importance_factor": self.qa_ok_importance_factor,
"qa_corrupted_pixels_pct_importance_factor": self.qa_corrupted_pixels_pct_importance_factor,
"almost_all_water_importance_factor": self.almost_all_water_importance_factor,
"dem_nan_pixels_pct_importance_factor": self.dem_nan_pixels_pct_importance_factor,
"dem_zero_pixels_pct_importance_factor": self.dem_zero_pixels_pct_importance_factor,
}
@property
def callbacks_kwargs(self) -> Dict[str, Any]:
return {
"save_top_k": self.save_top_k,
"monitor_metric": self.monitor_metric,
"monitor_mode": self.monitor_mode,
"early_stopping_patience": self.early_stopping_patience,
"swa": self.swa,
"swa_epoch_start": self.swa_epoch_start,
"swa_annealing_epochs": self.swa_annealing_epochs,
"swa_lr": self.swa_lr,
}
@property
def model_kwargs(self) -> Dict[str, Any]:
return {
"architecture": self.architecture,
"encoder": self.encoder,
"pretrained": self.pretrained,
"encoder_weights": self.encoder_weights,
"encoder_depth": self.encoder_depth,
"decoder_channels": self.decoder_channels,
"decoder_attention_type": self.decoder_attention_type,
"ignore_index": self.ignore_index,
"num_classes": self.num_classes,
"optimizer": self.optimizer,
"weight_decay": self.weight_decay,
"lr_scheduler": self.lr_scheduler,
"lr": self.lr,
"epochs": self.epochs,
"onecycle_pct_start": self.onecycle_pct_start,
"onecycle_div_factor": self.onecycle_div_factor,
"onecycle_final_div_factor": self.onecycle_final_div_factor,
"cyclic_base_lr": self.cyclic_base_lr,
"cyclic_mode": self.cyclic_mode,
"cosine_eta_min": self.cosine_eta_min,
"cosine_T_mult": self.cosine_T_mult,
"reduce_lr_on_plateau_factor": self.reduce_lr_on_plateau_factor,
"reduce_lr_on_plateau_patience": self.reduce_lr_on_plateau_patience,
"reduce_lr_on_plateau_threshold": self.reduce_lr_on_plateau_threshold,
"reduce_lr_on_plateau_min_lr": self.reduce_lr_on_plateau_min_lr,
"objective": self.objective,
"loss": self.loss,
"ce_class_weights": self.ce_class_weights,
"ce_smooth_factor": self.ce_smooth_factor,
"compile": self.compile,
"compile_mode": self.compile_mode,
"compile_dynamic": self.compile_dynamic,
"ort": self.ort,
"plot_n_batches": self.plot_n_batches,
"tta": self.tta,
"tta_merge_mode": self.tta_merge_mode,
"decision_threshold": self.decision_threshold,
}
@property
def trainer_kwargs(self) -> Dict[str, Any]:
return {
"precision": self.precision,
"fast_dev_run": self.fast_dev_run,
"max_epochs": self.epochs,
"limit_train_batches": self.limit_train_batches,
"limit_val_batches": self.limit_val_batches,
"limit_test_batches": self.limit_test_batches,
"log_every_n_steps": self.log_every_n_steps,
"accumulate_grad_batches": self.accumulate_grad_batches,
"val_check_interval": self.val_check_interval,
"benchmark": self.benchmark,
}
|