Merge pull request #32 from henryaddison/unet-paper

Tidy deterministic (U-Net) code ready for paper publishing

bin/deterministic/main.py

@@ -13,7 +13,7 @@
     "config", None, "Training configuration.", lock_config=True
 )
 flags.DEFINE_string("workdir", None, "Work directory.")
-flags.DEFINE_enum("mode", None, ["train", "eval"], "Running mode: train or eval")
+flags.DEFINE_enum("mode", None, ["train"], "Running mode: train.")
 flags.mark_flags_as_required(["workdir", "config", "mode"])
 
 

src/ml_downscaling_emulator/bin/evaluate.py

@@ -58,7 +58,7 @@ def sample(
     workdir: Path,
     dataset: str = typer.Option(...),
     split: str = "val",
-    epoch: int = typer.Option(...),
+    checkpoint: str = typer.Option(...),
     batch_size: int = None,
     num_samples: int = 1,
     input_transform_dataset: str = None,
@@ -71,16 +71,17 @@ def sample(
 
     if batch_size is not None:
         config.eval.batch_size = batch_size
-    if input_transform_dataset is not None:
-        config.data.input_transform_dataset = input_transform_dataset
-    else:
-        config.data.input_transform_dataset = dataset
+    with config.unlocked():
+        if input_transform_dataset is not None:
+            config.data.input_transform_dataset = input_transform_dataset
+        else:
+            config.data.input_transform_dataset = dataset
     if input_transform_key is not None:
         config.data.input_transform_key = input_transform_key
 
     output_dirpath = samples_path(
         workdir=workdir,
-        checkpoint=f"epoch-{epoch}",
+        checkpoint=checkpoint,
         dataset=dataset,
         input_xfm=f"{config.data.input_transform_dataset}-{config.data.input_transform_key}",
         split=split,
@@ -105,7 +106,7 @@ def sample(
         shuffle=False,
     )
 
-    ckpt_filename = os.path.join(workdir, "checkpoints", f"epoch_{epoch}.pth")
+    ckpt_filename = os.path.join(workdir, "checkpoints", f"{checkpoint}.pth")
     num_predictors = eval_dl.dataset[0][0].shape[0]
     state = load_model(config, num_predictors, ckpt_filename)
 

src/ml_downscaling_emulator/deterministic/configs/default.py → src/ml_downscaling_emulator/deterministic/configs/ukcp_local_12em_pr_unet.py

@@ -16,9 +16,12 @@ def get_config():
     evaluate.batch_size = 64
 
     config.data = data = ml_collections.ConfigDict()
-    data.dataset_name = ""
+    data.dataset_name = (
+        "bham_gcmx-4x_12em_psl-sphum4th-temp4th-vort4th_eqvt_random-season"
+    )
     data.input_transform_key = "stan"
     data.target_transform_key = "sqrturrecen"
+    data.input_transform_dataset = None
     data.time_inputs = False
 
     config.model = model = ml_collections.ConfigDict()

src/ml_downscaling_emulator/deterministic/run_lib.py

@@ -13,7 +13,7 @@
 
 from ..training import log_epoch, track_run
 from .utils import restore_checkpoint, save_checkpoint, create_model
-from ..torch import get_dataloader
+from ..data import get_dataloader
 
 FLAGS = flags.FLAGS
 EXPERIMENT_NAME = os.getenv("WANDB_EXPERIMENT_NAME")
@@ -208,7 +208,6 @@ def train_step_fn(state, batch, cond):
         EXPERIMENT_NAME, run_name, run_config, [config.model.name, "baseline"], tb_dir
     ) as (wandb_run, tb_writer):
         # Fit model
-        wandb_run.watch(model, criterion=criterion, log_freq=100)
 
         logging.info("Starting training loop at epoch %d." % (initial_epoch,))
 

tests/deterministic/smoke-test

@@ -0,0 +1,25 @@
+#! /usr/bin/env bash
+
+set -euo pipefail
+
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
+
+cpm_dataset="bham_gcmx-4x_1em_psl-sphum4th-temp4th-vort4th_eqvt_random-season-historic"
+gcm_dataset="bham_60km-4x_1em_psl-sphum4th-temp4th-vort4th_eqvt_random-season-historic"
+workdir="output/test/unet/test-run"
+
+config_path="src/ml_downscaling_emulator/deterministic/configs/ukcp_local_12em_pr_unet.py"
+
+train_batch_size=2
+epoch=2
+
+rm -rf ${workdir}
+WANDB_DISABLE_SERVICE=True WANDB_EXPERIMENT_NAME="test" python ${SCRIPT_DIR}/../../bin/deterministic/main.py --mode train --workdir ${workdir} --config ${config_path} --config.data.dataset_name=${cpm_dataset} --config.training.batch_size=${train_batch_size} --config.training.n_epochs=${epoch} --config.data.time_inputs=True --config.model.name=debug
+
+num_samples=2
+eval_batchsize=32
+
+rm -rf "${workdir}/samples/epoch_${epoch}/${cpm_dataset}"
+mlde evaluate sample ${workdir} --dataset ${cpm_dataset} --checkpoint epoch_${epoch} --batch-size ${eval_batchsize} --num-samples ${num_samples}
+rm -rf "${workdir}/samples/epoch_${epoch}/${gcm_dataset}"
+mlde evaluate sample ${workdir} --dataset ${gcm_dataset} --checkpoint epoch_${epoch} --batch-size ${eval_batchsize} --num-samples ${num_samples}

tests/smoke-test

@@ -12,11 +12,12 @@ config_path="src/ml_downscaling_emulator/score_sde_pytorch/configs/${sde}/${conf
 loc_spec_channels=0
 train_batch_size=2
 random_crop_size=32
+epoch=2
 
 rm -rf ${workdir}
-WANDB_EXPERIMENT_NAME="test" python bin/main.py --workdir ${workdir} --config ${config_path} --mode train --config.data.dataset_name=${dataset} --config.training.snapshot_freq=5 --config.training.eval_freq=100 --config.training.log_freq=50 --config.training.n_epochs=2 --config.model.loc_spec_channels=${loc_spec_channels} --config.training.random_crop_size=${random_crop_size} --config.training.batch_size=${train_batch_size} --config.data.input_transform_key=stan --config.data.target_transform_key=sqrturrecen --config.data.time_inputs=True
+WANDB_EXPERIMENT_NAME="test" python bin/main.py --workdir ${workdir} --config ${config_path} --mode train --config.data.dataset_name=${dataset} --config.training.snapshot_freq=5 --config.training.eval_freq=100 --config.training.log_freq=50 --config.training.n_epochs=${epoch} --config.model.loc_spec_channels=${loc_spec_channels} --config.training.random_crop_size=${random_crop_size} --config.training.batch_size=${train_batch_size} --config.data.input_transform_key=stan --config.data.target_transform_key=sqrturrecen --config.data.time_inputs=True
+
 
-epoch=2
 num_samples=2
 eval_batchsize=32
 checkpoint="epoch_${epoch}"