Merge branch 'dev'

.gitignore

@@ -1,14 +1,24 @@
 *.ckpt
+*.env
+*.out
+*.model
 *.csv
 *.mp4
 *.npy
 *.txt
 *.xml
 *.json
+*.pkl
+*.pickle
+*.ipynb
+*.tar
+*.tar.gz
 *__pycache__
 logs/runs/
 logs/wandb/
 multirun/
+data/
+baseline/
 tmp*
 !requirements.txt
 !packages.txt

README.md

@@ -21,6 +21,14 @@ conda activate myenv
 pip install -r requirements.txt
 ```
 
+### .env file setting
+```shell
+PROJECT_DIR={PROJECT_DIR}
+CONNECTED_DIR={CONNECTED_DIR}
+DEVICES={DEVICES}
+HF_HOME={HF_HOME}
+```
+
 ### Model Hyper-Parameters Tuning
 
 * end-to-end
@@ -49,6 +57,18 @@ python main.py mode=test is_tuned={tuned or untuned} num_trials={num_trials} epo
 python main.py mode=predict is_tuned={tuned or untuned} num_trials={num_trials} epoch={ckpt epoch}
 ```
 
+### Examples of shell scipts
+
+* train
+```shell
+bash scripts/train.sh
+```
+
+* test
+```shell
+bash scripts/test.sh
+```
+
 
 __If you want to change main config, use --config-name={config_name}.__
 

configs/architecture/rhythm_architecture.yaml

@@ -10,7 +10,7 @@ model:
 
 strategy: ${strategy}
 lr: ${lr}
-t_max: ${t_max}
+period: ${period}
 eta_min: ${eta_min}
 interval: step
 connected_dir: ${connected_dir}

configs/rhythm.yaml

@@ -9,8 +9,8 @@ defaults:
   - trainer: trainer
 
 package_name: rppg-project
-project_dir: /home/ddang/${package_name}
-connected_dir: /data/${package_name}
+project_dir: ${oc.env:PROJECT_DIR}/${package_name}
+connected_dir: ${oc.env:CONNECTED_DIR}/${package_name}
 
 seed: 2024
 
@@ -31,20 +31,22 @@ rnn_num_layers: 1
 direction: bi
 
 lr: 0.0001
-t_max: 50
-eta_min: 0.000025
+period: 2
+eta_min: 0.00001
 
 monitor: val_rmse_loss
 tracking_direction: min
-patience: 3
+patience: 2
 min_delta: 0
 
-devices: 4
+devices: ${oc.decode:${oc.env:DEVICES}}
 accelerator: gpu
 strategy: ddp
 log_every_n_steps: 10
 precision: 32
 accumulate_grad_batches: 1
+gradient_clip_val: 1
+gradient_clip_algorithm: norm
 epoch: 100
 
 model_name: CustomizedRhythmNet

configs/trainer/trainer.yaml

@@ -5,4 +5,6 @@ strategy: ${strategy}
 log_every_n_steps: ${log_every_n_steps}
 precision: ${precision}
 accumulate_grad_batches: ${accumulate_grad_batches}
+gradient_clip_val: ${gradient_clip_val}
+gradient_clip_algorithm: ${gradient_clip_algorithm}
 max_epochs: ${epoch}

configs/tuner/rhythm_tuner.yaml

@@ -24,9 +24,9 @@ hparams:
     low: 0.00005
     high: 0.0005
     log: False
-  t_max:
-    low: 25
-    high: 100
+  period:
+    low: 1
+    high: 10
     log: False
   eta_min:
     low: 0.00005
@@ -42,6 +42,8 @@ module_params:
   log_every_n_steps: ${log_every_n_steps}
   precision: ${precision}
   accumulate_grad_batches: ${accumulate_grad_batches}
+  gradient_clip_val: ${gradient_clip_val}
+  gradient_clip_algorithm: ${gradient_clip_algorithm}
   max_epochs: ${epoch}
   monitor: ${monitor}
   mode: ${tracking_direction}

main.py

@@ -1,8 +1,14 @@
+import dotenv
+
+dotenv.load_dotenv(
+    override=True,
+)
+
 import os
 import warnings
 
 os.environ["HYDRA_FULL_ERROR"] = "1"
-os.environ["HF_HOME"] = "/data/huggingface"
+os.environ["HF_HOME"] = os.environ.get("HF_HOME")
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
 warnings.filterwarnings("ignore")
 

requirements.txt

@@ -24,6 +24,7 @@ opencv-python==4.8.1.78
 optuna==3.5.0
 pandas==1.5.0
 pillow==10.2.0
+python-dotenv==1.0.1
 pytorch-lightning==2.2.0
 scikit-learn==1.3.2
 scipy==1.10.1

src/architectures/rhythm_architecture.py

@@ -16,7 +16,7 @@ def __init__(
         model: nn.Module,
         strategy: str,
         lr: float,
-        t_max: int,
+        period: int,
         eta_min: float,
         interval: str,
         connected_dir: str,
@@ -25,7 +25,7 @@ def __init__(
         self.model = model
         self.strategy = strategy
         self.lr = lr
-        self.t_max = t_max
+        self.period = period
         self.eta_min = eta_min
         self.interval = interval
         self.connected_dir = connected_dir
@@ -81,9 +81,10 @@ def configure_optimizers(self) -> Dict[str, Any]:
                 self.parameters(),
                 lr=self.lr,
             )
+        t_max = self.period * self.trainer.num_training_batches
         scheduler = optim.lr_scheduler.CosineAnnealingLR(
             optimizer=optimizer,
-            T_max=self.t_max,
+            T_max=t_max,
             eta_min=self.eta_min,
         )
         return {

src/tuners/rhythm_tuner.py

@@ -107,12 +107,12 @@ def optuna_objective(
                 high=self.hparams.lr.high,
                 log=self.hparams.lr.log,
             )
-        if self.hparams.t_max:
-            params["t_max"] = trial.suggest_int(
-                name="t_max",
-                low=self.hparams.t_max.low,
-                high=self.hparams.t_max.high,
-                log=self.hparams.t_max.log,
+        if self.hparams.period:
+            params["period"] = trial.suggest_int(
+                name="period",
+                low=self.hparams.period.low,
+                high=self.hparams.period.high,
+                log=self.hparams.period.log,
             )
         if self.hparams.eta_min:
             params["eta_min"] = trial.suggest_float(
@@ -133,7 +133,7 @@ def optuna_objective(
             model=model,
             strategy=self.module_params.strategy,
             lr=params["lr"],
-            t_max=params["t_max"],
+            period=params["period"],
             eta_min=params["eta_min"],
             interval=self.module_params.interval,
             connected_dir=self.module_params.connected_dir,
@@ -154,6 +154,8 @@ def optuna_objective(
             log_every_n_steps=self.module_params.log_every_n_steps,
             precision=self.module_params.precision,
             accumulate_grad_batches=self.module_params.accumulate_grad_batches,
+            gradient_clip_val=self.module_params.gradient_clip_val,
+            gradient_clip_algorithm=self.module_params.gradient_clip_algorithm,
             max_epochs=self.module_params.max_epochs,
             enable_checkpointing=False,
             callbacks=callbacks,