Added preliminary experiment results and updated README

Author: dkapur17

ExperimentLogs.md

@@ -15,7 +15,14 @@ Train without noise and denoiser. Evaluate without noise and denoiser.
 * Input Model: [`base/model1`](./SBAgent/models/base/model1.zip)
 
 #### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 73.50%  |
+| Collision Rate             | 26.50%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -20.25  |
+| Mean Episode Length        | 261.744 |
 
 ---
 ## [Experiment 2A](./experimentConfigs/experiment2a.json)
@@ -33,7 +40,14 @@ Train without noise and denoiser. Evaluate with noise (0.01) but no denoiser.
 * Input Model: [`base/model1`](./SBAgent/models/base/model1.zip)
 
 #### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 69.90%  |
+| Collision Rate             | 30.10%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -86.17  |
+| Mean Episode Length        | 257.42  |
 
 ---
 ## [Experiment 2B](./experimentConfigs/experiment2b.json)
@@ -51,7 +65,14 @@ Train without noise and denoiser. Evaluate with noise (0.05) but no denoiser.
 * Input Model: [`base/model1`](./SBAgent/models/base/model1.zip)
 
 #### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 71.30%  |
+| Collision Rate             | 28.70%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -65.85  |
+| Mean Episode Length        | 262.001 |
 
 ---
 ## [Experiment 2C](./experimentConfigs/experiment2c.json)
@@ -69,7 +90,14 @@ Train without noise and denoiser. Evaluate with noise (0.1) but no denoiser.
 * Input Model: [`base/model1`](./SBAgent/models/base/model1.zip)
 
 #### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 71.00%  |
+| Collision Rate             | 29.00%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -66.10  |
+| Mean Episode Length        | 258.97  |
 
 ---
 
@@ -88,11 +116,18 @@ Train without noise and denoiser. Evaluate with both noise (0.01) and denoiser(L
 * Input Model: [`base/model1`](./SBAgent/models/base/model1.zip)
 
 #### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 67.70%  |
+| Collision Rate             | 32.30%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -133.16 |
+| Mean Episode Length        | 257.793 |
 
 ---
 
-## Ex[periment 3B](./experimentConfigs/experiment3b.json)
+## [Experiment 3B](./experimentConfigs/experiment3b.json)
 
 Train without noise and denoiser. Evaluate with both noise (0.05) and denoiser (LPF).
 ### Train Parameters
@@ -107,7 +142,14 @@ Train without noise and denoiser. Evaluate with both noise (0.05) and denoiser (
 * Input Model: [`base/model1`](./SBAgent/models/base/model1.zip)
 
 #### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 70.80%  |
+| Collision Rate             | 29.20%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -74.76  |
+| Mean Episode Length        | 260.713 |
 
 ---
 
@@ -126,7 +168,14 @@ Train without noise and denoiser. Evaluate with both noise (0.1) and denoiser (L
 * Input Model: [`base/model1`](./SBAgent/models/base/model1.zip)
 
 #### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 69.90%  |
+| Collision Rate             | 30.10%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -95.42  |
+| Mean Episode Length        | 262.544 |
 
 ---
 
@@ -145,7 +194,14 @@ Train without noise and denoiser. Evaluate with both noise (0.01) and denoiser (
 * Input Model: [`base/model1`](./SBAgent/models/base/model1.zip)
 
 #### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 71.40%  |
+| Collision Rate             | 28.60%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -63.36  |
+| Mean Episode Length        | 262.162 |
 
 ---
 
@@ -164,7 +220,14 @@ Train without noise and denoiser. Evaluate with both noise (0.05) and denoiser (
 * Input Model: [`base/model1`](./SBAgent/models/base/model1.zip)
 
 #### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 70.10%  |
+| Collision Rate             | 29.90%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -79.07  |
+| Mean Episode Length        | 256.643 |
 
 ---
 
@@ -183,7 +246,14 @@ Train without noise and denoiser. Evaluate with both noise (0.1) and denoiser (K
 * Input Model: [`base/model1`](./SBAgent/models/base/model1.zip)
 
 #### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 68.60%  |
+| Collision Rate             | 31.40%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -108.65 |
+| Mean Episode Length        | 254.936 |
 
 ---
 
@@ -202,7 +272,14 @@ Train with noise (0.01) but no denoiser. Evaluate with noise (0.01) but no denoi
 * Input Model: [`base/model2_0.01`](./SBAgent/models/base/model2_0.01.zip)
 
 #### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 68.80%  |
+| Collision Rate             | 31.20%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -114.47 |
+| Mean Episode Length        | 258.475 |
 
 ---
 
@@ -221,7 +298,14 @@ Train with noise (0.05) but no denoiser. Evaluate with noise (0.05) but no denoi
 * Input Model: [`base/model2_0.05`](./SBAgent/models/base/model2_0.05.zip)
 
 #### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 69.90%  |
+| Collision Rate             | 30.10%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -120.60 |
+| Mean Episode Length        | 278.681 |
 
 ---
 
@@ -239,8 +323,14 @@ Train with noise (0.1) but no denoiser. Evaluate with noise (0.1) but no denoise
 * Environment Config: [`v2_0.1.json`](./configs/v2_0.1.json)
 * Input Model: [`base/model2_0.1`](./SBAgent/models/base/model2_0.1.zip)
 
-#### Evaluation Results
-*TBD*
+
+| Metric                     | Value   |
+|----------------------------|---------|
+| Success Rate               | 67.20%  |
+| Collision Rate             | 32.80%  |
+| Mean Incompletion Distance | N/A     |
+| Mean Reward                | -183.71 |
+| Mean Episode Length        | 284.602 |
 
 ---
 

README.md

@@ -57,7 +57,7 @@ Check out [experiment1.json](./experimentConfigs/experiment1.json) for reference
 
 From here, you do one of two things:
 
-1. Train Step: Run `TrainDispatcher.py` to the model on the given environment. The script is written to dispatch a batch job on IIIT-H's HPC cluster, so modify it to run on your machine as needed.
+1. Train Step: Run `TrainDispatcher.py` to the model on the given environment. The script is written to dispatch a batch job on IIIT-H's HPC cluster, or on your local machine with the `--local` flag.
 
 2. Evaluation Step: Run `SBAgent/EvaluateExperiment.py` to evaluate the model in the given environment.
 
@@ -141,4 +141,69 @@ To use a Kalman Filter as the Denoise Engine, set `"denoiseEngine"` to the follo
 }
 ```
 
-* `processNoise`: Standard Deviation of the process noise.
+* `processNoise`: Standard Deviation of the process noise.
+
+
+## The Scripts
+
+### `TrainDispatcher.py`
+
+Used to perform the train phase of a given experiment. To run it, run:
+
+```bash
+python TrainDispatcher.py [-h] [-s STEPS] [--local] experimentConfigFile
+```
+
+* `experimentConfigFile` is path to the experiment configuration file.
+* `-s` to specify the number of timesteps to train for. Defaults to 2,000,000.
+* Add the `--local` flag to run the training on your local machine. If this flag is omitted, the script generates a batch job configuration and dispatches it on one of the nodes on IIIT-H's HPC.
+
+### `SBAgent/EvaluateExperiment.py`
+
+Used to perform the evaluation phase of a given experiment. To run it, run:
+
+```bash
+cd SBAgent
+python EvaluateExperiment.py [-h] [-t TRIALS] [--gui] [--no-gui] experimentConfigFile
+```
+
+* `experimentConfigFile` is path to the experiment configuration file
+* `-t` to specify the number of episodes used for evaluation. Defaults to 10.
+* `--gui` or `--no-gui` to specify whether or not to render the simulation. If no flag is provided, then GUI is NOT launched by default.
+
+### `SBAgent/EvaluateModel.py`
+
+Used to evaluate a given model on a given environment. Allows flexibility outside of a defined experiment. To run it, run:
+
+```bash
+cd SBAgent
+python EvaluateModel.py [-h] [-t TRIALS] [--gui] [--no-gui] configFileName inputModelPath
+```
+* `configFileName` is the name of the environment configuration file inside the `config` directory.
+* `inputModelPath` is the path to the model to be evaluated.
+* `-t` to specify the number of episodes used for evaluation. Defaults to 10.
+* `--gui` or `--no-gui` to specify whether or not to render the simulation. If no flag is provided, then GUI is NOT launched by default.
+
+### `SBAgent/TrainModel.py`
+
+Used to train a model on a given environment. Allows flexibity outside of a defined experiment. Used internally by `TrainDispatcher.py`.
+
+
+```bash
+cd SBAgent
+python TrainModel.py [-h] [-s STEPS] configFileName outputModelName
+```
+
+* `configFileName` is the name of the environment configuration file inside the `config` directory.
+* `outputModelName` is the name to save the output model as. This must begin with either `base/` or `finetuned/` (for the current script it is should ideally always be `base/`) followed by the name of the model. The trained model is then saved in `SBAgent/models/<outputModelName>`.
+* `-s` to specify the number of timesteps to train for. Default is 2,000,000.
+
+## Evironment Configurations
+
+Check out [ConfigLogs.md](./configs/ConfigLogs.md) for the description for every environement.
+
+## Models
+
+Check out [base/ModelLogs.md](./SBAgent/models/base/ModelLogs.md) for the description of every model trained from scratch.
+
+Check out [finetuned/ModelLogs.md](./SBAgent/models/finetuned/ModelLogs.md) for the description of every model finetuned from an existing model.

SBAgent/EvaluateExperiment.py

@@ -18,20 +18,29 @@
 parser.add_argument('--no-gui', action='store_false', dest='gui', help='Disable GUI')
 args = parser.parse_args()
 
-
 np.random.seed(42)
 random.seed(42)
 
 with open(args.experimentConfigFile, 'r') as f:
     experimentConfig = json.load(f)
 
 experimentName = experimentConfig["name"]
-configFileName = experimentConfig["trainParameters"]["config"]
-modelName = experimentConfig["trainParameters"]["outputModelName"]
+trainEnvConfig = experimentConfig["trainParameters"]["config"]
+evaluationEnvConfig = experimentConfig["evaluationParameters"]["config"]
+modelName = experimentConfig["evaluationParameters"]["inputModelName"]
 
 print(f"Running Evaluation on {experimentName}")
 
-env = EnvBuilder.buildEnvFromConfig(os.path.join('..', 'configs', configFileName), gui=args.gui)
+trainEnv = EnvBuilder.buildEnvFromConfig(os.path.join('..', 'configs', trainEnvConfig), gui=False)
+print("Model trained on")
+print(trainEnv)
+trainEnv.close()
+del trainEnv
+
+print("Evaluating Model on")
+env = EnvBuilder.buildEnvFromConfig(os.path.join('..', 'configs', evaluationEnvConfig), gui=args.gui)
+print(env)
+
 agent = PPO.load(os.path.join('models', modelName, 'best_model'))
 
 totalTrials = args.trials
@@ -73,4 +82,5 @@
 }
 
 evaluationTable = [[k, v] for k,v in evaluationResults.items()]
+print()
 print(tabulate(evaluationTable, headers=["Metric", "Value"], tablefmt='github'))

SBAgent/TrainModel.py

@@ -28,5 +28,5 @@
 
 agent = PPO("MlpPolicy", env, verbose=1, tensorboard_log=os.path.join('logs', modelName))
 agent.learn(n_steps, callback=eval_callback, tb_log_name="train_logs")
-
+agent.save(os.path.join('models', modelName, 'final_model'))
 

SBAgent/logs/base/model1/train_logs_1/events.out.tfevents.1674991165.gnode015.7020.0

@@ -0,0 +1 @@
+# Model Logs

SBAgent/logs/base/model2_0.01/train_logs_1/events.out.tfevents.1674916563.gnode022.10637.0

@@ -387,4 +387,4 @@ def _randomizeDroneSpawnLocation(self):
                         (self.geoFence.ymin + self.geoFence.ymax) + np.random.uniform(-y_scale/2 + ObstacleAviary.COLLISION_BOUND_RADIUS*2, y_scale/2 - ObstacleAviary.COLLISION_BOUND_RADIUS*2),
                         (self.geoFence.zmin + self.geoFence.zmax)/2])
 
-        self.INIT_XYZS = np.array([self.initPos])
+        self.INIT_XYZS = np.array([self.initPos])

SBAgent/logs/base/model2_0.05/train_logs_1/events.out.tfevents.1674916567.gnode022.10722.0

@@ -8,7 +8,9 @@ class LPFDenoiseEngine:
     def __init__(self, order:int, criticalFreq:float, ftype:str, freq:float) -> None:
 
         self.order = order
-        self.b, self.a = scipy.signal.iirfilter(order, criticalFreq, fs=freq, btype="low", ftype=ftype)
+        self.criticalFreq = criticalFreq
+        self.ftype = ftype
+        self.b, self.a = scipy.signal.iirfilter(self.order, self.criticalFreq, fs=freq, btype="low", ftype=self.ftype)
 
         self._xs = deque([], maxlen=len(self.b))
         self._ys = deque([], maxlen=len(self.a)-1)
@@ -42,11 +44,17 @@ def reset(self, _:np.ndarray) -> None:
         self.observedHistory.clear()
         self.denoisedHistory.clear()
 
+    def __str__(self) -> str:
+
+        return f"LPFDenoiseEngine({self.order}, {self.criticalFreq}, {self.ftype})"
+
 class KFDenoiseEngine:
 
     def __init__(self, measurementNoise:float, dt:float, fixedAltitude:bool, initPos:np.ndarray, processNoise:float=0) -> None:
 
         self.fixedAltitude = fixedAltitude
+        self.processNoise = processNoise
+        
         if self.fixedAltitude:
 
             # state = [x, y, vx, vy]
@@ -138,6 +146,8 @@ def reset(self, initPos) -> None:
             x0 = np.array([initPos[0], initPos[1], initPos[2], 0, 0, 0])
 
         self.initFilter(A, C, x0, P0, Q, R)
-        
+    
+    def __str__(self) -> str:
+        return f"KFDenoiseEngine({self.processNoise})"