Hyperparameter Optimization with Ray Tune and Weights and Biases (#76)

.gitignore

@@ -149,3 +149,5 @@ OUT/
 
 examples/experiments/grounded_program_synthesis/dataset
 ckpts/
+
+ray_results/

configs/ppo_config.yml

@@ -6,7 +6,7 @@ model:
 
 train:
   seq_length: 48  # Size of LM context
-  epochs: 1000  # Train for max(epochs, total_steps)
+  epochs: 1000 # Train for max(epochs, total_steps)
   total_steps: 10000  # Train for max(epochs, total_steps)
   batch_size: 128  # batch size
 

configs/ray_tune_configs/ppo_config.yml

@@ -0,0 +1,68 @@
+tune_config:
+  mode: "max"
+  metric: "mean_reward"
+  search_alg: "bohb" # random
+  scheduler: "hyperbandforbohb" # fifo
+  num_samples: 15
+  max_concurrent_trials: null
+  time_budget_s: null
+  reuse_actors: null
+
+model:
+  model_path: "lvwerra/gpt2-imdb"  # Name of hf model to load
+  tokenizer_path: "gpt2"  # Name of hf tokenizer to load
+  model_type: "AcceleratePPOModel"  # Name of accelerate model type to load
+  num_layers_unfrozen:  # Number of bottom layers to freeze during training
+    strategy: "choice"
+    values: [2, 3]
+
+train:
+  seq_length:  # Size of LM context
+    strategy: "choice"
+    values: [36, 48, 52]
+  epochs:  # Train for max(epochs, total_steps)
+    strategy: "choice"
+    values: [80, 100, 120]
+  total_steps: 10000  # Train for max(epochs, total_steps)
+  batch_size: 128  # batch size
+
+  lr_init: 1.412e-4  # init learning rate
+  lr_target: 1.412e-4  # target final learning rate
+  opt_betas: [0.9, 0.95] # adam betas
+  opt_eps: 1.0e-8  # adam eps
+  weight_decay: 1.0e-6  # weight decay param
+
+  checkpoint_interval: 10000  # checkpoint interval
+  eval_interval: 4  # eval interval
+
+  pipeline: "PPOPipeline"  # prompt pipeline to load
+  orchestrator: "PPOOrchestrator"  # orchestrator to load
+  project_name: "trlx-hyperopt-bohb"
+
+method:
+  name: 'ppoconfig'  # Name of RL method config
+  num_rollouts:  # Number of rollouts to collect per epoch
+    strategy: "choice"
+    values: [96, 128]
+  chunk_size: 128  # Number of rollouts to collect in one loop of orchestrator
+  ppo_epochs:  # Number of ppo epochs
+    strategy: "randint"
+    values: [3, 6] # 3 is inclusive, 6 is exclusive
+  init_kl_coef: # init kl coefficient
+    strategy: "quniform"
+    values: [0.1, 0.3, 0.1]
+  target: 6  # target kl coefficient, set None for fixed kl coef
+  horizon: 10000  # PPO horizon
+  gamma: 1  # PPO discount
+  lam: # PPO lambda
+    strategy: "uniform"
+    values: [0.93, 0.98]
+  cliprange: 0.2  # clip range
+  cliprange_value: 0.2  # clip range
+  vf_coef: 2.3  # value term weight
+  gen_kwargs:
+    max_length: 48  # LM max sample gen length
+    min_length: 48  # LM min sample gen length
+    top_k: 0.0  # top k
+    top_p: 1.0  # top p
+    do_sample: True  # sample

setup.cfg

@@ -20,6 +20,8 @@ install_requires =
     transformers>=4.21.2
     tqdm
     wandb
+    ray>=2.0.1
+    tabulate>=0.9.0
 
 [options.extras_require]
 dev =

train_sweep.py

@@ -0,0 +1,96 @@
+# Usage: python train_sweep.py --config configs/ray_tune_configs/ppo_config.yml --example-name ppo_sentiments
+import wandb
+import argparse
+from pathlib import Path
+
+import ray
+from ray.air import session
+from ray import tune
+
+import trlx
+from trlx.ray_tune import load_ray_yaml
+from trlx.ray_tune import get_param_space
+from trlx.ray_tune import get_tune_config
+from trlx.ray_tune import get_train_function
+from trlx.ray_tune.wandb import log_trials, create_report
+
+from ray.tune.logger import JsonLoggerCallback
+from ray.tune.logger import CSVLoggerCallback
+
+
+def tune_function(train_function, param_space: dict, tune_config: dict, resources: dict):
+    tuner = tune.Tuner(
+        tune.with_resources(train_function, resources=resources),
+        param_space=param_space,
+        tune_config=tune.TuneConfig(**tune_config),
+        run_config = ray.air.RunConfig(
+            local_dir="ray_results",
+            callbacks=[CSVLoggerCallback()]
+        ),
+    )
+
+    results = tuner.fit()
+
+    log_trials(
+        tuner._local_tuner.get_experiment_checkpoint_dir(),
+        param_space["train"]["project_name"]
+    )
+
+    create_report(
+        param_space,
+        tune_config,
+        Path(tuner._local_tuner.get_experiment_checkpoint_dir()).stem,
+        results.get_best_result().config
+    )
+
+    print("Best hyperparameters found were: ", results.get_best_result().config)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--example-name", type=str, default="ppo_sentiments", help="Name of the example"
+    )
+    parser.add_argument(
+        "--config", type=str, default=None, required=True, help="The config file defining the param_space."
+    )
+    parser.add_argument(
+        "--num-cpus", type=int, default=4, help="Number of CPUs to use per exp."
+    )
+    parser.add_argument(
+        "--num-gpus", type=int, default=1, help="Number of GPUs to use per exp."
+    )
+    parser.add_argument(
+        "--server-address",
+        type=str,
+        default=None,
+        required=False,
+        help="The address of server to connect to if using Ray Client.",
+    )
+
+    args, _ = parser.parse_known_args()
+
+    # Read config and parse it
+    config = load_ray_yaml(args.config)
+    tune_config = get_tune_config(config)
+    param_space = get_param_space(config)
+
+    # Initialize Ray.
+    if args.server_address:
+        ray.init(address=f"ray://{args.server_address}")
+    else:
+        ray.init()
+
+    resources = {
+        "cpu": args.num_cpus,
+        "gpu": args.num_gpus,
+    }
+
+    # Register the training function that will be used for training the model.
+    train_function = get_train_function(args.example_name)
+    tune.register_trainable("train_function", train_function)
+
+    tune_function(train_function, param_space, tune_config, resources)
+
+    # Shut down Ray.
+    ray.shutdown()

trlx/data/configs.py

@@ -142,7 +142,21 @@ def to_dict(self):
         """
         Convert TRLConfig to dictionary.
         """
-        data = self.model.__dict__.copy()
-        data.update(self.train.__dict__)
-        data.update(self.method.__dict__)
+        data = {
+            "model": self.model.__dict__,
+            "train": self.train.__dict__,
+            "method": self.method.__dict__,
+        }
+
         return data
+
+    @classmethod
+    def from_dict(cls, config_dict: dict):
+        """
+        Convert dictionary to TRLConfig.
+        """
+        return cls(
+            ModelConfig.from_dict(config_dict["model"]),
+            TrainConfig.from_dict(config_dict["train"]),
+            get_method(config_dict["method"]["name"]).from_dict(config_dict["method"]),
+        )

trlx/model/accelerate_base_model.py

@@ -18,6 +18,17 @@
 else:
     from tqdm import tqdm
 
+import ray
+from ray.air import session
+
+
+def parse_results_for_session(results: dict):
+    for k, v in results.items():
+        if isinstance(v, torch.Tensor):
+            results[k] = float(v)
+
+    return results
+
 
 @register_model
 class AccelerateRLModel(BaseRLModel):
@@ -63,7 +74,7 @@ def __init__(self, config, train_mode=True):
         for m in gpt_blocks_to_freeze:
             m.requires_grad_(False)
 
-        if self.accelerator.is_main_process:
+        if self.accelerator.is_main_process and not ray.is_initialized():
             self.accelerator.init_trackers(
                 project_name=self.config.train.project_name,
                 config=self.config.to_dict(),
@@ -199,9 +210,8 @@ def evaluate(self):
                     columns_data.append(values)
 
             rows = list(zip(*columns_data))
-            stats["samples"] = wandb.Table(columns=columns, rows=rows)
-
-            print(rows[0])
+            if not ray.is_initialized():
+                stats["samples"] = wandb.Table(columns=columns, rows=rows)
 
         return stats
 
@@ -246,7 +256,14 @@ def learn(self):
                                 "backward_time": backward_time,
                             }
                         )
-                        self.accelerator.log(results)
+
+                        if not ray.is_initialized():
+                            self.accelerator.log(results)
+
+                        # Report the metrics to Ray Tune.
+                        if ray.is_initialized():
+                            tmp_results = parse_results_for_session(results)
+                            session.report(tmp_results)
 
                     desc = ", ".join(f"{k}: {v:.2f}" for k, v in stats.items())
                     tbar.set_description(desc)

trlx/orchestrator/ppo_orchestrator.py

@@ -10,6 +10,9 @@
 from trlx.utils import Clock
 from trlx.utils.modeling import logprobs_from_logits
 
+import ray
+from ray.air import session
+
 
 @register_orchestrator
 class PPOOrchestrator(Orchestrator):
@@ -124,7 +127,9 @@ def make_experience(self, num_rollouts: int = 1024, iter_count: int = 0):
             ppo_rl_elements += new_ppo_rl_elements
 
         stats = {"exp_time": exp_time}
-        self.rl_model.accelerator.log(stats, step=iter_count)
+
+        if not ray.is_initialized():
+            self.rl_model.accelerator.log(stats, step=iter_count)
 
         # Push samples and rewards to model's rollout storage
         self.rl_model.push_to_store(ppo_rl_elements)