File size: 6,247 Bytes
bd3493c | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 | import os
import sys
import pickle
import argparse
import torch
from torch.utils.data import DataLoader
import torch.distributed as dist
import torch.utils.data.distributed
import torch.multiprocessing as mp
from torch.utils.data.distributed import DistributedSampler
from torch.distributed import init_process_group, destroy_process_group
from misc_downscaling_functionality import DownscalingRmseLoss
from e2e_model import *
from loader import *
from models import *
from unet_wrap_padding import *
from trainer import DDPTrainerE2E
from loss_functions import WeightedRmseLoss, PressureWeightedRmseLoss, RmseLoss
torch.set_float32_matmul_precision("medium")
def ddp_setup(rank, world_size, master_port):
"""
Setup DDP
"""
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = master_port
init_process_group(
backend="nccl",
rank=rank,
world_size=world_size,
)
torch.cuda.set_device(rank)
def start_date(name):
"""
Set split start dates
"""
if name == "train":
return "2007-01-02"
elif name == "val":
return "2019-01-01"
elif name == "test":
return "2018-01-01"
else:
raise Exception(f"Unrecognised split name {name}")
def end_date(name):
"""
Set split end dates
"""
if name == "train":
return "2017-12-31"
elif name == "val":
return "2019-11-01"
elif name == "test":
return "2018-12-21"
else:
raise Exception(f"Unrecognised split name {name}")
def main(rank, world_size, output_dir, args):
"""
Run end to end training
"""
master_port = args.master_port
lead_time = args.lead_time
era5_mode = args.era5_mode
ddp_setup(rank, world_size, master_port)
# Setup loss function
if args.loss == "lw_rmse":
lf = WeightedRmseLoss(
args.res,
start_ind=0,
end_ind=24,
weight_per_variable=False,
)
elif args.loss == "lw_rmse_pressure_weighted":
lf = PressureWeightedRmseLoss(args.res, era5_mode)
elif args.loss == "rmse":
lf = RmseLoss()
elif args.loss == "downscaling_rmse":
lf = DownscalingRmseLoss()
# Instantiate model
model = ConvCNPWeatherE2E(
device="cuda",
lead_time=lead_time,
se_model_path=args.se_model_path,
forecast_model_path=args.forecast_model_path,
sf_model_path=args.sf_model_path,
)
dist.barrier()
# Instantiate loaders
train_dataset = WeatherDatasetE2E(
device="cuda",
hadisd_mode="train",
start_date="2007-01-02",
end_date="2017-12-31",
lead_time=lead_time,
era5_mode="4u",
mode="train",
res=args.res,
var_start=0,
var_end=24,
diff=bool(0),
two_frames=bool(0),
region=args.region,
hadisd_var=args.var,
max_steps_per_epoch=args.max_steps_per_epoch,
)
val_dataset = WeatherDatasetE2E(
device="cuda",
hadisd_mode="train",
start_date="2019-01-01",
end_date="2019-12-21",
lead_time=lead_time,
era5_mode="4u",
mode="train",
res=args.res,
var_start=0,
var_end=24,
diff=bool(0),
two_frames=bool(0),
region=args.region,
hadisd_var=args.var,
)
test_dataset = WeatherDatasetE2E(
device="cuda",
hadisd_mode="train",
start_date="2018-01-01",
end_date="2018-12-21",
lead_time=lead_time,
era5_mode="4u",
mode="train",
res=args.res,
var_start=0,
var_end=24,
diff=bool(0),
two_frames=bool(0),
region=args.region,
hadisd_var=args.var,
)
train_sampler = DistributedSampler(train_dataset)
val_sampler = DistributedSampler(val_dataset)
test_sampler = DistributedSampler(test_dataset)
train_loader = DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=False,
sampler=train_sampler,
)
val_loader = DataLoader(
val_dataset,
batch_size=args.batch_size,
shuffle=False,
sampler=val_sampler,
)
test_loader = DataLoader(
test_dataset,
batch_size=args.batch_size,
shuffle=False,
sampler=test_sampler,
)
# Instantiate trainer
trainer = DDPTrainerE2E(
model,
rank,
train_loader,
val_loader,
lf,
output_dir,
args.lr,
train_sampler,
hadisd_variable_name=args.var,
weight_decay=args.weight_decay,
weights_path=None,
tune_film=0,
test_loader=test_loader,
)
# Train model
trainer.train(n_epochs=args.epoch)
destroy_process_group()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--output_dir")
parser.add_argument("--loss")
parser.add_argument("--batch_size", type=int, default=3)
parser.add_argument("--epoch", type=int, default=10)
parser.add_argument("--master_port", default="12345")
parser.add_argument("--lr", type=float, default=5e-5)
parser.add_argument("--lead_time", type=int)
parser.add_argument("--era5_mode", default="4u")
parser.add_argument("--sf_model_path")
parser.add_argument("--se_model_path")
parser.add_argument("--forecast_model_path")
parser.add_argument("--weight_decay", type=float, default=1e-6)
parser.add_argument("--max_steps_per_epoch", type=int)
parser.add_argument("--res", type=int, default=1)
parser.add_argument("--frequency", type=int, default=6)
parser.add_argument("--region", default="global")
parser.add_argument("--var", choices=["tas", "u", "v", "psl", "ws"])
args = parser.parse_args()
torch.device("cuda")
# Make results directory
output_dir = args.output_dir
try:
os.mkdir(output_dir)
except:
pass
# Save config
with open(output_dir + "/config.pkl", "wb") as f:
pickle.dump(vars(args), f)
world_size = torch.cuda.device_count()
mp.spawn(main, args=[world_size, output_dir, args], nprocs=world_size)
|