File size: 6,482 Bytes
5d2fa0b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 | import argparse
import math
import os
import torch
import wandb
from omegaconf import OmegaConf
from timm.optim import create_optimizer_v2
from torch.optim.lr_scheduler import CosineAnnealingLR, ReduceLROnPlateau, StepLR
from src.dataset import get_dataloaders
from src.loss import get_criterion
from src.models import PlantDiseaseModel, get_param_groups
from src.trainer import Trainer
from src.utils import CosineAnnealingWarmupLR, load_config, set_seed
def build_optimizer(model, config):
layer_decay = getattr(config.optimizer, "layer_decay", 1.0)
param_groups = get_param_groups(
model,
base_lr=config.optimizer.backbone_lr,
head_lr=config.optimizer.head_lr,
weight_decay=config.optimizer.weight_decay,
)
if config.optimizer.name.lower() == "adamw":
if layer_decay == 1:
optimizer = torch.optim.AdamW(param_groups)
else:
optimizer = create_optimizer_v2(
model,
opt="adamw",
lr=config.optimizer.head_lr,
layer_decay=layer_decay,
weight_decay=config.optimizer.weight_decay,
)
else:
optimizer = torch.optim.Adam(param_groups)
return optimizer
def build_scheduler(optimizer, config, len_loader):
if config.scheduler.name.lower() == "cosine":
return CosineAnnealingLR(
optimizer, T_max=config.training.epochs, eta_min=config.scheduler.min_lr
)
elif config.scheduler.name.lower() == "step":
return StepLR(optimizer, step_size=3, gamma=0.1)
elif config.scheduler.name.lower() == "plateau":
return ReduceLROnPlateau(
optimizer,
mode="max",
factor=0.1,
patience=3,
min_lr=config.scheduler.min_lr,
)
elif config.scheduler.name.lower() == "cosine_warmup":
return CosineAnnealingWarmupLR(
optimizer,
warmup_steps=config.scheduler.warmup_epochs
* len_loader
/ config.training.gradient_accumulation_steps,
total_steps=config.training.epochs
* len_loader
/ config.training.gradient_accumulation_steps,
min_lr=config.scheduler.min_lr,
)
else:
return None
def main():
parser = argparse.ArgumentParser(
description="Train Plant Disease Classification Baseline"
)
parser.add_argument(
"--config", type=str, default="configs/config.yaml", help="Path to config file"
)
parser.add_argument(
"--resume", type=str, default=None, help="Path to checkpoint to resume from"
)
parser.add_argument(
"--init_weights", type=str, default=None, help="Path to weights for warm start"
)
args = parser.parse_args()
config = load_config(args.config)
set_seed(config.seed, deterministic=True)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Environment: Using device {device}")
train_loader, val_loader, num_classes = get_dataloaders(config)
if num_classes == 0:
print(
"WARNING: No data found. Make sure your datasets are correctly structured."
)
# Fallback to prevent immediate crash if no data is present yet
num_classes = 1
config.model.num_classes = num_classes
model = PlantDiseaseModel(config, num_classes=num_classes)
model.to(device)
if args.init_weights and os.path.exists(args.init_weights):
print(f"Warm starting from weights: {args.init_weights}")
checkpoint = torch.load(args.init_weights, map_location=device)
state_dict = checkpoint.get("state_dict", checkpoint)
model.load_state_dict(state_dict)
optimizer = build_optimizer(model, config)
criterion = get_criterion(config)
scheduler = build_scheduler(optimizer, config, len(train_loader))
# resume Logic
start_epoch = 1
checkpoint = None
run_id = None
if args.resume and os.path.exists(args.resume):
print(f"Resuming experiment from checkpoint: {args.resume}")
checkpoint = torch.load(args.resume, map_location=device)
model.load_state_dict(checkpoint["state_dict"])
optimizer.load_state_dict(checkpoint["optimizer"])
if scheduler and checkpoint["scheduler"]:
scheduler.load_state_dict(checkpoint["scheduler"])
start_epoch = checkpoint["epoch"] + 1
if "rng_states" in checkpoint:
torch.set_rng_state(checkpoint["rng_states"]["torch"].cpu())
if device.type == "cuda" and checkpoint["rng_states"]["cuda"] is not None:
torch.cuda.set_rng_state_all(
[s.cpu() for s in checkpoint["rng_states"]["cuda"]]
)
if config.logging.use_wandb:
run_id = checkpoint.get("wandb_run_id")
if start_epoch > config.training.epochs:
print(
f"Requested to resume at epoch {start_epoch}, but total epochs is {config.training.epochs}. Exiting."
)
return
# Wandb tracking
if config.logging.use_wandb:
wandb_config = OmegaConf.to_container(config, resolve=True)
wandb.init(
project=config.logging.project_name,
name=config.experiment_name,
config=wandb_config,
id=run_id, # Use the loaded ID (or None if brand new)
resume="allow",
)
trainer = Trainer(
model=model,
train_loader=train_loader,
val_loader=val_loader,
criterion=criterion,
optimizer=optimizer,
scheduler=scheduler,
config=config,
device=device,
)
if checkpoint is not None:
if trainer.use_ema and checkpoint.get("state_dict_ema"):
trainer.model_ema.module.load_state_dict(checkpoint["state_dict_ema"])
if args.resume and os.path.exists(args.resume):
if checkpoint["scaler"]:
trainer.scaler.load_state_dict(checkpoint["scaler"])
if checkpoint["early_stopping"]:
trainer.early_stopping.best_score = checkpoint["early_stopping"][
"best_score"
]
trainer.early_stopping.counter = checkpoint["early_stopping"]["counter"]
trainer.early_stopping.early_stop = checkpoint["early_stopping"][
"early_stop"
]
trainer.fit()
if __name__ == "__main__":
main()
|