hackathonf23-Stacks/rmv_discount_local.py

"""
This is the second ablation study, of removing the discount factor during testing.
"""
import numpy as np
import pandas as pd
import tensorflow as tf
from keras.layers import Dense
from sklearn.preprocessing import MinMaxScaler
import dagshub
import mlflow
import mlflow.keras
import pickle

mlflow.set_tracking_uri('https://dagshub.com/ML-Purdue/hackathonf23-Stacks.mlflow')
dagshub.init(repo_owner='ML-Purdue', repo_name='hackathonf23-Stacks', mlflow=True)

def get_or_create_experiment_id(name):
    exp = mlflow.get_experiment_by_name(name)
    if exp is None:
        exp_id = mlflow.create_experiment(name)
        return exp_id
    return exp.experiment_id

class MaxEntIRL:
    def __init__(self, state_dim):
        self.state_dim = state_dim
        self.model = self._create_irl_model()

    def _create_irl_model(self):
        model = tf.keras.Sequential([
            Dense(self.state_dim, input_shape=(self.state_dim,), activation='relu'),
            Dense(4096, activation='relu'),
            Dense(2048, activation='relu'),
            Dense(self.state_dim, activation='linear')
        ])
        return model

    def generateHumanTrajectories(self, num_trajectories, trajectory_length):
        human_trajectories = []
        for _ in range(num_trajectories):
            trajectory = []
            state = np.zeros(self.state_dim)

            for _ in range(trajectory_length):
                direction_probabilities = self._generate_direction_probabilities()  # Get direction probabilities
                action_coefficients = np.random.choice([-1, 0, 1], p=direction_probabilities)
                action = action_coefficients * 0.1
                new_state = state + action
                trajectory.append((state, action))
                state = new_state

            human_trajectories.append(trajectory)

        return human_trajectories

    def _generate_direction_probabilities(self):
        probabilities = np.random.dirichlet(np.ones(self.state_dim) * 0.1)
        return probabilities

    def loadDataset(self, file_path):
        data = pd.read_csv(file_path)  # Load CSV data
        scaler = MinMaxScaler()
        columns_to_normalize = ['position x [mm]', 'position y [mm]', 'position z (height) [mm]', 'velocity [mm/s]']
        data[columns_to_normalize] = scaler.fit_transform(data[columns_to_normalize])
        return data

    def train_irl_with_dataset(self, data, lr=0.001, epochs=3):
        state_dim = self.state_dim
        optimizer = tf.keras.optimizers.Adam(learning_rate=lr)

        positions = data[['position x [mm]', 'position y [mm]', 'position z (height) [mm]']].values
        velocities = data['velocity [mm/s]'].values
        mlflow.tensorflow.autolog()
        with mlflow.start_run(experiment_id=get_or_create_experiment_id("Ablation Study 2: Removed Discount Factor")):
            for epoch in range(epochs):
                total_loss = 0
                state_frequencies = self._calculate_state_frequencies(positions)

                for idx in range(len(positions)):
                    state = positions[idx]
                    velocity = velocities[idx]

                    with tf.GradientTape() as tape:
                        preferences = self.model(state[np.newaxis, :])
                        prob_human = tf.nn.softmax(preferences)

                        # Define losses excluding the discount factor
                        alignment_loss = -tf.reduce_sum(state_frequencies * tf.math.log(prob_human + 1e-8), axis=1)

                        # Compute the gradients
                        grads = tape.gradient(alignment_loss, self.model.trainable_variables)
                        optimizer.apply_gradients(zip(grads, self.model.trainable_variables))
                        total_loss += tf.reduce_sum(alignment_loss)  # Accumulate the total loss

                avg_loss = total_loss / len(positions)
                mlflow.log_metric(f"loss", avg_loss, step=epoch)
                print(f"Epoch {epoch + 1}/{epochs}, IRL Loss without Discount Factor: {avg_loss}")

    def train_irl(self, human_trajectories=None, data=None, use_dataset=False, lr=0.001, epochs=3):
        if use_dataset and data is not None:
            # Train using the loaded dataset
            self.train_irl_with_dataset(data, lr=lr, epochs=epochs)
        else:
            # Train using the generative function
            if human_trajectories is None:
                human_trajectories = self.generateHumanTrajectories(num_trajectories, trajectory_length)

            self._train_irl_generative(human_trajectories, lr=lr, epochs=epochs)

    def _train_irl_generative(self, human_trajectories, lr=0.001, epochs=3):
        trajectory_length = len(human_trajectories[0])
        optimizer = tf.keras.optimizers.Adam(learning_rate=lr)

        for epoch in range(epochs):
            total_loss = 0
            state_frequencies = self._calculate_state_frequencies(human_trajectories, trajectory_length)

            for trajectory in human_trajectories:
                for state, _ in trajectory:
                    with tf.GradientTape() as tape:
                        preferences = self.model(state[np.newaxis, :])
                        prob_human = tf.nn.softmax(preferences)

                        # Inside the training loop:
                        max_entropy_loss = -tf.reduce_sum(prob_human * tf.math.log(prob_human + 1e-8), axis=1)
                        alignment_loss = -tf.reduce_sum(state_frequencies * tf.math.log(prob_human + 1e-8), axis=1)
                        maxent_irl_objective = max_entropy_loss + alignment_loss

                    grads = tape.gradient(maxent_irl_objective, self.model.trainable_variables)
                    optimizer.apply_gradients(zip(grads, self.model.trainable_variables))
                    total_loss += maxent_irl_objective

            avg_loss = total_loss / (len(human_trajectories) * trajectory_length)
            print(f"Epoch {epoch + 1}/{epochs}, MaxEnt IRL Loss: {avg_loss}")

    def _calculate_state_frequencies(self, positions):
        state_counts = np.sum(positions, axis=0)
        state_frequencies = state_counts / (len(positions) * self.state_dim)
        return state_frequencies
    def save_model(self, file_path):
            model_config = self.model.get_config()
            with open(file_path, 'wb') as f:
                pickle.dump(model_config, f)

    @classmethod
    def load_model(cls, file_path, state_dim):
        with open(file_path, 'rb') as f:
            model_config = pickle.load(f)
            irl_instance = cls(state_dim)
            irl_instance.model = tf.keras.Sequential.from_config(model_config)
        return irl_instance
    # Indicate test completion status
    
state_dim = 3  # Dimension of the state space
irl = MaxEntIRL(state_dim)
num_trajectories = 100
trajectory_length = 20

# Load the dataset
file_path = '/Users/vinay/Desktop/Computer_Science_Projects/ReScience/hackathonf23-Stacks/data/train.csv'  # Replace with the actual file path
data = irl.loadDataset(file_path)

irl.train_irl(data=data, use_dataset=True, lr=0.001, epochs=3)
irl.save_model('/Users/vinay/Desktop/Computer_Science_Projects/ReScience/hackathonf23-Stacks/models/rmv_discount_model.pkl')