judsonupchurch/Inverted-Pendulum-Neural-Network-Controller
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Finding best learning rates from the sweeps

Browse Source
This commit is contained in:
judsonupchurch 2025-03-29 02:07:34 +00:00
parent aa34bfac8c
commit e238bed91e
11 changed files with 757 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
analysis/PendulumDynamics.py Normal file
View File

@ -0,0 +1,26 @@
import torch
import torch.nn as nn
class PendulumDynamics(nn.Module):
def __init__(self, controller, m:'float'=1, R:'float'=1, g:'float'=9.81):
super().__init__()
self.controller = controller
self.m: 'float' = m
self.R: 'float' = R
self.g: 'float' = g
def forward(self, t, state):
# Get the current values from the state
theta, omega, alpha, desired_theta = state[:, 0], state[:, 1], state[:, 2], state[:, 3]
# Make the input stack for the controller
input = torch.stack([theta, omega, alpha, desired_theta], dim=1)
# Get the torque (the output of the neural network)
tau = self.controller(input).squeeze(-1)
# Relax alpha
alpha_desired = (self.g / self.R) * torch.sin(theta) + tau / (self.m * self.R**2)
dalpha = alpha_desired - alpha
return torch.stack([omega, alpha, dalpha, torch.zeros_like(desired_theta)], dim=1)

BIN
analysis/__pycache__/PendulumDynamics.cpython-310.pyc Normal file
View File

Binary file not shown.

BIN
analysis/__pycache__/initial_conditions.cpython-310.pyc Normal file
View File

Binary file not shown.

62
analysis/best_base_loss_learning_rate_sweep.py Normal file
View File

@ -0,0 +1,62 @@
data = {
'one': {
'csv': {
'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/one/lr_0.100',
'csv_loss': 0.07867201417684555,
'constant_loss': 2.5390186309814453
},
'constant': {
'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/one/lr_0.100',
'csv_loss': 0.07867201417684555,
'constant_loss': 2.5390186309814453
}
},
'one_fourth': {
'csv': {
'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/one_fourth/lr_0.300',
'csv_loss': 0.08876045793294907,
'constant_loss': 2.5319466590881348
},
'constant': {
'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/one_fourth/lr_0.250',
'csv_loss': 0.09172269701957703,
'constant_loss': 2.5288496017456055
}
},
'four': {
'csv': {
'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/four/lr_0.200',
'csv_loss': 0.1293140947818756,
'constant_loss': 2.9976892471313477
},
'constant': {
'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/four/lr_0.200',
'csv_loss': 0.1293140947818756,
'constant_loss': 2.9976892471313477
}
},
'two': {
'csv': {
'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/two/lr_0.100',
'csv_loss': 0.07678339630365372,
'constant_loss': 2.5585412979125977
},
'constant': {
'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/two/lr_0.100',
'csv_loss': 0.07678339630365372,
'constant_loss': 2.5585412979125977
}
},
'one_half': {
'csv': {
'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/one_half/lr_0.200',
'csv_loss': 0.08620432019233704,
'constant_loss': 2.541421890258789
},
'constant': {
'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/one_half/lr_0.200',
'csv_loss': 0.08620432019233704,
'constant_loss': 2.541421890258789
}
}
}

15
analysis/best_base_loss_learning_rate_sweep.txt Normal file
View File

@ -0,0 +1,15 @@
Final best results (dictionary):
{'one': {'csv': {'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/one/lr_0.100', 'csv_loss': 0.07867201417684555, 'constant_loss': 2.5390186309814453}, 'constant': {'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/one/lr_0.100', 'csv_loss': 0.07867201417684555, 'constant_loss': 2.5390186309814453}}, 'one_fourth': {'csv': {'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/one_fourth/lr_0.300', 'csv_loss': 0.08876045793294907, 'constant_loss': 2.5319466590881348}, 'constant': {'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/one_fourth/lr_0.250', 'csv_loss': 0.09172269701957703, 'constant_loss': 2.5288496017456055}}, 'four': {'csv': {'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/four/lr_0.200', 'csv_loss': 0.1293140947818756, 'constant_loss': 2.9976892471313477}, 'constant': {'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/four/lr_0.200', 'csv_loss': 0.1293140947818756, 'constant_loss': 2.9976892471313477}}, 'two': {'csv': {'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/two/lr_0.100', 'csv_loss': 0.07678339630365372, 'constant_loss': 2.5585412979125977}, 'constant': {'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/two/lr_0.100', 'csv_loss': 0.07678339630365372, 'constant_loss': 2.5585412979125977}}, 'one_half': {'csv': {'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/one_half/lr_0.200', 'csv_loss': 0.08620432019233704, 'constant_loss': 2.541421890258789}, 'constant': {'path': '/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep/one_half/lr_0.200', 'csv_loss': 0.08620432019233704, 'constant_loss': 2.541421890258789}}}
Summary Table:
Function Name Candidate Learning Rate CSV Loss Constant Loss
one CSV 0.100 0.078672 2.539019
Constant 0.100 0.078672 2.539019
one_fourth CSV 0.300 0.088760 2.531947
Constant 0.250 0.091723 2.528850
four CSV 0.200 0.129314 2.997689
Constant 0.200 0.129314 2.997689
two CSV 0.100 0.076783 2.558541
Constant 0.100 0.076783 2.558541
one_half CSV 0.200 0.086204 2.541422
Constant 0.200 0.086204 2.541422

206
analysis/best_time_weighting_learning_rate_sweep.py Normal file
View File

@ -0,0 +1,206 @@
data = {
"inverse": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/inverse/lr_0.250",
"csv_loss": 0.531498372554779,
"constant_loss": 2.5503664016723633
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/inverse/lr_0.250",
"csv_loss": 0.531498372554779,
"constant_loss": 2.5503664016723633
}
},
"linear_mirrored": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/linear_mirrored/lr_0.125",
"csv_loss": 2.3770766258239746,
"constant_loss": 2.552375078201294
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/linear_mirrored/lr_0.125",
"csv_loss": 2.3770766258239746,
"constant_loss": 2.552375078201294
}
},
"inverse_squared_mirrored": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/inverse_squared_mirrored/lr_0.160",
"csv_loss": 0.033845994621515274,
"constant_loss": 2.7603342533111572
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/inverse_squared_mirrored/lr_0.160",
"csv_loss": 0.033845994621515274,
"constant_loss": 2.7603342533111572
}
},
"cubic_mirrored": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/cubic_mirrored/lr_0.080",
"csv_loss": 2.0769901275634766,
"constant_loss": 2.563471555709839
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/cubic_mirrored/lr_0.080",
"csv_loss": 2.0769901275634766,
"constant_loss": 2.563471555709839
}
},
"quadratic": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/quadratic/lr_0.200",
"csv_loss": 0.06192325800657272,
"constant_loss": 3.025479316711426
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/quadratic/lr_0.080",
"csv_loss": 0.14040324091911316,
"constant_loss": 2.982274055480957
}
},
"inverse_squared": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/inverse_squared/lr_0.200",
"csv_loss": 1.1794205904006958,
"constant_loss": 2.5662319660186768
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/inverse_squared/lr_0.200",
"csv_loss": 1.1794205904006958,
"constant_loss": 2.5662319660186768
}
},
"quadratic_mirrored": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/quadratic_mirrored/lr_0.125",
"csv_loss": 2.218207836151123,
"constant_loss": 2.555176258087158
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/quadratic_mirrored/lr_0.125",
"csv_loss": 2.218207836151123,
"constant_loss": 2.555176258087158
}
},
"square_root": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/square_root/lr_0.250",
"csv_loss": 0.6526519656181335,
"constant_loss": 2.5856597423553467
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/square_root/lr_0.250",
"csv_loss": 0.6526519656181335,
"constant_loss": 2.5856597423553467
}
},
"inverse_cubed_mirrored": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/inverse_cubed_mirrored/lr_0.200",
"csv_loss": 0.03754603490233421,
"constant_loss": 2.9996697902679443
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/inverse_cubed_mirrored/lr_0.200",
"csv_loss": 0.03754603490233421,
"constant_loss": 2.9996697902679443
}
},
"cubic_root_mirrored": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/cubic_root_mirrored/lr_0.250",
"csv_loss": 2.47979474067688,
"constant_loss": 2.5389654636383057
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/cubic_root_mirrored/lr_0.250",
"csv_loss": 2.47979474067688,
"constant_loss": 2.5389654636383057
}
},
"inverse_mirrored": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/inverse_mirrored/lr_0.160",
"csv_loss": 0.032234687358140945,
"constant_loss": 2.942859649658203
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/inverse_mirrored/lr_0.160",
"csv_loss": 0.032234687358140945,
"constant_loss": 2.942859649658203
}
},
"inverse_cubed": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/inverse_cubed/lr_0.200",
"csv_loss": 1.4481265544891357,
"constant_loss": 2.557009696960449
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/inverse_cubed/lr_0.200",
"csv_loss": 1.4481265544891357,
"constant_loss": 2.557009696960449
}
},
"cubic_root": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/cubic_root/lr_0.250",
"csv_loss": 1.0203485488891602,
"constant_loss": 2.609311819076538
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/cubic_root/lr_0.250",
"csv_loss": 1.0203485488891602,
"constant_loss": 2.609311819076538
}
},
"square_root_mirrored": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/square_root_mirrored/lr_0.160",
"csv_loss": 2.4792795181274414,
"constant_loss": 2.5693373680114746
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/square_root_mirrored/lr_0.160",
"csv_loss": 2.4792795181274414,
"constant_loss": 2.5693373680114746
}
},
"linear": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/linear/lr_0.125",
"csv_loss": 0.2883843183517456,
"constant_loss": 3.05281400680542
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/linear/lr_0.080",
"csv_loss": 0.28867313265800476,
"constant_loss": 2.9585072994232178
}
},
"constant": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/constant/lr_0.160",
"csv_loss": 2.608083486557007,
"constant_loss": 2.606748342514038
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/constant/lr_0.160",
"csv_loss": 2.608083486557007,
"constant_loss": 2.606748342514038
}
},
"cubic": {
"csv": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/cubic/lr_0.160",
"csv_loss": 0.04065453261137009,
"constant_loss": 3.101959228515625
},
"constant": {
"path": "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep/cubic/lr_0.300",
"csv_loss": 0.049555618315935135,
"constant_loss": 3.0432639122009277
}
}
}

39
analysis/best_time_weighting_learning_rate_sweep.txt Normal file
View File

File diff suppressed because one or more lines are too long

179
analysis/find_best_base_loss_learning_rate_sweep.py Normal file
View File

@ -0,0 +1,179 @@
import os
import csv
import torch
from torchdiffeq import odeint
import pandas as pd
from PendulumController import PendulumController
from PendulumDynamics import PendulumDynamics
from initial_conditions import initial_conditions
# Device and initial conditions setup
device = torch.device("cpu")
state_0 = torch.tensor(initial_conditions, dtype=torch.float32, device=device)
# Constants (same as in your training code)
m = 10.0
g = 9.81
R = 1.0
t_start, t_end, t_points = 0, 10, 1000
t_span = torch.linspace(t_start, t_end, t_points, device=device)
# Base path containing the base loss function directories
base_path = "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/base_loss_learning_rate_sweep"
def compute_constant_loss(controller_path):
"""
Loads a controller from the given path, sets up the dynamics using the constant weighting function,
simulates the system, and returns the computed loss.
"""
controller = PendulumController().to(device)
controller.load_state_dict(torch.load(controller_path, map_location=device))
pendulum_dynamics = PendulumDynamics(controller, m, R, g).to(device)
with torch.no_grad():
state_traj = odeint(pendulum_dynamics, state_0, t_span, method='rk4')
theta = state_traj[:, :, 0]
desired_theta = state_traj[:, :, 3]
loss = torch.mean((theta - desired_theta) ** 2)
return loss.item()
# Dictionary to store the best results for each base loss function.
# Each key maps to a dictionary with keys "csv" and "constant".
# Each candidate dictionary contains:
# "path": best lr directory path,
# "csv_loss": loss from the training log,
# "constant_loss": loss computed via the constant method.
best_results = {}
# Process each base loss function directory
for function_name in os.listdir(base_path):
function_path = os.path.join(base_path, function_name)
if not os.path.isdir(function_path):
continue
print(f"Processing base loss function: {function_name}")
# Initialize best candidate variables for CSV-based best
best_csv_csv_loss = float('inf')
best_csv_constant_loss = float('inf')
best_csv_path = None
# Initialize best candidate variables for constant-based best
best_constant_constant_loss = float('inf')
best_constant_csv_loss = float('inf')
best_constant_path = None
# Loop through each learning rate directory (directories named "lr_*")
for lr_dir in os.listdir(function_path):
if not lr_dir.startswith("lr_"):
continue
lr_path = os.path.join(function_path, lr_dir)
if not os.path.isdir(lr_path):
continue
# --- Compute CSV loss candidate ---
current_csv_loss = None
csv_file = os.path.join(lr_path, "training_log.csv")
if os.path.exists(csv_file):
try:
with open(csv_file, 'r') as f:
reader = csv.DictReader(f)
losses = []
for row in reader:
try:
loss_value = float(row['Loss'])
losses.append(loss_value)
except ValueError:
continue
if losses:
current_csv_loss = min(losses)
except Exception as e:
print(f"Error reading CSV {csv_file}: {e}")
# --- Compute constant loss candidate ---
current_constant_loss = None
controllers_dir = os.path.join(lr_path, "controllers")
controller_file = os.path.join(controllers_dir, "controller_200.pth")
if os.path.exists(controller_file):
try:
current_constant_loss = compute_constant_loss(controller_file)
except Exception as e:
print(f"Error computing constant loss for {controller_file}: {e}")
# Update best CSV candidate (based on CSV loss)
if current_csv_loss is not None:
csv_const_loss_val = current_constant_loss if current_constant_loss is not None else float('inf')
if current_csv_loss < best_csv_csv_loss:
best_csv_csv_loss = current_csv_loss
best_csv_constant_loss = csv_const_loss_val
best_csv_path = lr_path
# Update best Constant candidate (based on constant loss)
if current_constant_loss is not None:
csv_loss_val = current_csv_loss if current_csv_loss is not None else float('inf')
if current_constant_loss < best_constant_constant_loss:
best_constant_constant_loss = current_constant_loss
best_constant_csv_loss = csv_loss_val
best_constant_path = lr_path
best_results[function_name] = {
"csv": {"path": best_csv_path, "csv_loss": best_csv_csv_loss, "constant_loss": best_csv_constant_loss},
"constant": {"path": best_constant_path, "csv_loss": best_constant_csv_loss, "constant_loss": best_constant_constant_loss},
}
print(f"Finished {function_name}:")
print(f" Best CSV candidate - Path: {best_csv_path}, CSV Loss: {best_csv_csv_loss}, Constant Loss: {best_csv_constant_loss}")
print(f" Best Constant candidate - Path: {best_constant_path}, CSV Loss: {best_constant_csv_loss}, Constant Loss: {best_constant_constant_loss}")
print("Final best results:")
print(best_results)
# Build summary table rows using pandas.
# Extract only the learning rate (e.g., from "lr_0.250" get "0.250") rather than the full path.
def extract_lr(path):
if path is None:
return "N/A"
base = os.path.basename(path)
if base.startswith("lr_"):
return base[3:]
return base
table_rows = []
for function_name, results in best_results.items():
csv_info = results.get("csv", {})
constant_info = results.get("constant", {})
csv_lr = extract_lr(csv_info.get("path"))
constant_lr = extract_lr(constant_info.get("path"))
table_rows.append({
"Function Name": function_name,
"Candidate": "CSV",
"Learning Rate": csv_lr,
"CSV Loss": csv_info.get("csv_loss", float('inf')),
"Constant Loss": csv_info.get("constant_loss", float('inf'))
})
table_rows.append({
"Function Name": "", # Leave blank for the second row
"Candidate": "Constant",
"Learning Rate": constant_lr,
"CSV Loss": constant_info.get("csv_loss", float('inf')),
"Constant Loss": constant_info.get("constant_loss", float('inf'))
})
df = pd.DataFrame(table_rows, columns=["Function Name", "Candidate", "Learning Rate", "CSV Loss", "Constant Loss"])
# Get the table as a formatted string
table_str = df.to_string(index=False)
print("\n" + table_str)
# Write the dictionary and table to a file
output_file = "best_base_loss_learning_rate_sweep.txt"
with open(output_file, "w") as f:
f.write("Final best results (dictionary):\n")
f.write(str(best_results) + "\n\n")
f.write("Summary Table:\n")
f.write(table_str)
print(f"\nResults have been written to {output_file}")

179
analysis/find_best_time_weighting_learning_rate_sweep.py Normal file
View File

@ -0,0 +1,179 @@
import os
import csv
import torch
from torchdiffeq import odeint
import pandas as pd
from PendulumController import PendulumController
from PendulumDynamics import PendulumDynamics
from initial_conditions import initial_conditions
# Device and initial conditions setup
device = torch.device("cpu")
state_0 = torch.tensor(initial_conditions, dtype=torch.float32, device=device)
# Constants (same as in your training code)
m = 10.0
g = 9.81
R = 1.0
t_start, t_end, t_points = 0, 10, 1000
t_span = torch.linspace(t_start, t_end, t_points, device=device)
# Base path containing the time_weighting_function directories
base_path = "/home/judson/Neural-Networks-in-GNC/inverted_pendulum/training/time_weighting_learning_rate_sweep"
def compute_constant_loss(controller_path):
"""
Loads a controller from the given path, sets up the dynamics using the constant weighting function,
simulates the system, and returns the computed loss.
"""
controller = PendulumController().to(device)
controller.load_state_dict(torch.load(controller_path, map_location=device))
pendulum_dynamics = PendulumDynamics(controller, m, R, g).to(device)
with torch.no_grad():
state_traj = odeint(pendulum_dynamics, state_0, t_span, method='rk4')
theta = state_traj[:, :, 0]
desired_theta = state_traj[:, :, 3]
loss = torch.mean((theta - desired_theta) ** 2)
return loss.item()
# Dictionary to store the best results for each time weighting function.
# Each key maps to a dictionary with keys "csv" and "constant".
# Each candidate dictionary contains:
# "path": best lr directory path,
# "csv_loss": loss from the training log,
# "constant_loss": loss computed via the constant method.
best_results = {}
# Process each time weighting function directory
for function_name in os.listdir(base_path):
function_path = os.path.join(base_path, function_name)
if not os.path.isdir(function_path):
continue
print(f"Processing weighting function: {function_name}")
# Initialize best candidate variables for CSV-based best
best_csv_csv_loss = float('inf')
best_csv_constant_loss = float('inf')
best_csv_path = None
# Initialize best candidate variables for constant-based best
best_constant_constant_loss = float('inf')
best_constant_csv_loss = float('inf')
best_constant_path = None
# Loop through each learning rate directory (directories named "lr_*")
for lr_dir in os.listdir(function_path):
if not lr_dir.startswith("lr_"):
continue
lr_path = os.path.join(function_path, lr_dir)
if not os.path.isdir(lr_path):
continue
# --- Compute CSV loss candidate ---
current_csv_loss = None
csv_file = os.path.join(lr_path, "training_log.csv")
if os.path.exists(csv_file):
try:
with open(csv_file, 'r') as f:
reader = csv.DictReader(f)
losses = []
for row in reader:
try:
loss_value = float(row['Loss'])
losses.append(loss_value)
except ValueError:
continue
if losses:
current_csv_loss = min(losses)
except Exception as e:
print(f"Error reading CSV {csv_file}: {e}")
# --- Compute constant loss candidate ---
current_constant_loss = None
controllers_dir = os.path.join(lr_path, "controllers")
controller_file = os.path.join(controllers_dir, "controller_200.pth")
if os.path.exists(controller_file):
try:
current_constant_loss = compute_constant_loss(controller_file)
except Exception as e:
print(f"Error computing constant loss for {controller_file}: {e}")
# Update best CSV candidate (based on CSV loss)
if current_csv_loss is not None:
csv_const_loss_val = current_constant_loss if current_constant_loss is not None else float('inf')
if current_csv_loss < best_csv_csv_loss:
best_csv_csv_loss = current_csv_loss
best_csv_constant_loss = csv_const_loss_val
best_csv_path = lr_path
# Update best Constant candidate (based on constant loss)
if current_constant_loss is not None:
csv_loss_val = current_csv_loss if current_csv_loss is not None else float('inf')
if current_constant_loss < best_constant_constant_loss:
best_constant_constant_loss = current_constant_loss
best_constant_csv_loss = csv_loss_val
best_constant_path = lr_path
best_results[function_name] = {
"csv": {"path": best_csv_path, "csv_loss": best_csv_csv_loss, "constant_loss": best_csv_constant_loss},
"constant": {"path": best_constant_path, "csv_loss": best_constant_csv_loss, "constant_loss": best_constant_constant_loss},
}
print(f"Finished {function_name}:")
print(f" Best CSV candidate - Path: {best_csv_path}, CSV Loss: {best_csv_csv_loss}, Constant Loss: {best_csv_constant_loss}")
print(f" Best Constant candidate - Path: {best_constant_path}, CSV Loss: {best_constant_csv_loss}, Constant Loss: {best_constant_constant_loss}")
print("Final best results:")
print(best_results)
# Build summary table rows using pandas.
# Extract only the learning rate (e.g., from "lr_0.250" get "0.250") rather than the full path.
def extract_lr(path):
if path is None:
return "N/A"
base = os.path.basename(path)
if base.startswith("lr_"):
return base[3:]
return base
table_rows = []
for function_name, results in best_results.items():
csv_info = results.get("csv", {})
constant_info = results.get("constant", {})
csv_lr = extract_lr(csv_info.get("path"))
constant_lr = extract_lr(constant_info.get("path"))
table_rows.append({
"Function Name": function_name,
"Candidate": "CSV",
"Learning Rate": csv_lr,
"CSV Loss": csv_info.get("csv_loss", float('inf')),
"Constant Loss": csv_info.get("constant_loss", float('inf'))
})
table_rows.append({
"Function Name": "", # Leave blank for the second row
"Candidate": "Constant",
"Learning Rate": constant_lr,
"CSV Loss": constant_info.get("csv_loss", float('inf')),
"Constant Loss": constant_info.get("constant_loss", float('inf'))
})
df = pd.DataFrame(table_rows, columns=["Function Name", "Candidate", "Learning Rate", "CSV Loss", "Constant Loss"])
# Get the table as a formatted string
table_str = df.to_string(index=False)
print("\n" + table_str)
# Write the dictionary and table to a file
output_file = "best_time_weighting_learning_rate_sweep.txt"
with open(output_file, "w") as f:
f.write("Final best results (dictionary):\n")
f.write(str(best_results) + "\n\n")
f.write("Summary Table:\n")
f.write(table_str)
print(f"\nResults have been written to {output_file}")

26
analysis/initial_conditions.py Normal file
View File

@ -0,0 +1,26 @@
from torch import pi
initial_conditions = [
[1/6 * pi, 0.0, 0.0, 0.0],
[-1/6 * pi, 0.0, 0.0, 0.0],
[2/3 * pi, 0.0, 0.0, 0.0],
[-2/3 * pi, 0.0, 0.0, 0.0],
[0.0, 1/3 * pi, 0.0, 0.0],
[0.0, -1/3 * pi, 0.0, 0.0],
[0.0, 2 * pi, 0.0, 0.0],
[0.0, -2 * pi, 0.0, 0.0],
[0.0, 0.0, 0.0, 2 * pi],
[0.0, 0.0, 0.0, -2 * pi],
[0.0, 0.0, 0.0, 1/2 * pi],
[0.0, 0.0, 0.0, -1/2 * pi],
[0.0, 0.0, 0.0, 1/3 * pi],
[0.0, 0.0, 0.0, -1/3 * pi],
[1/4 * pi, 1 * pi, 0.0, 0.0],
[-1/4 * pi, -1 * pi, 0.0, 0.0],
[1/2 * pi, -1 * pi, 0.0, 1/3 * pi],
[-1/2 * pi, 1 * pi, 0.0, -1/3 * pi],
[1/4 * pi, 1 * pi, 0.0, 2 * pi],
[-1/4 * pi, -1 * pi, 0.0, 2 * pi],
[1/2 * pi, -1 * pi, 0.0, 4 * pi],
[-1/2 * pi, 1 * pi, 0.0, -4 * pi],
]

32
training/base_loss_functions.py
View File

@ -22,7 +22,7 @@ def normalized_loss(theta: torch.Tensor, desired_theta: torch.Tensor, exponent:
denominator = (2 * math.pi + delta) ** exponent - delta ** exponent
return min_val + (1 - min_val) * (numerator / denominator)
# Specific loss functions with exponents 1/4, 1/2, 1, 2, 4
# Existing loss functions
def one_fourth_loss(theta: torch.Tensor, desired_theta: torch.Tensor, min_val: float = 0.01) -> torch.Tensor:
return normalized_loss(theta, desired_theta, exponent=1/4, min_val=min_val)
@ -38,13 +38,31 @@ def square_loss(theta: torch.Tensor, desired_theta: torch.Tensor, min_val: float
def fourth_loss(theta: torch.Tensor, desired_theta: torch.Tensor, min_val: float = 0.01) -> torch.Tensor:
return normalized_loss(theta, desired_theta, exponent=4, min_val=min_val)
# New loss functions
def one_third_loss(theta: torch.Tensor, desired_theta: torch.Tensor, min_val: float = 0.01) -> torch.Tensor:
return normalized_loss(theta, desired_theta, exponent=1/3, min_val=min_val)
def one_fifth_loss(theta: torch.Tensor, desired_theta: torch.Tensor, min_val: float = 0.01) -> torch.Tensor:
return normalized_loss(theta, desired_theta, exponent=1/5, min_val=min_val)
def three_loss(theta: torch.Tensor, desired_theta: torch.Tensor, min_val: float = 0.01) -> torch.Tensor:
return normalized_loss(theta, desired_theta, exponent=3, min_val=min_val)
def five_loss(theta: torch.Tensor, desired_theta: torch.Tensor, min_val: float = 0.01) -> torch.Tensor:
return normalized_loss(theta, desired_theta, exponent=5, min_val=min_val)
# Dictionary mapping function names to a tuple of (exponent, function)
base_loss_functions = {
'one_fourth': (1/4, one_fourth_loss),
'one_half': (1/2, one_half_loss),
'one': (1, abs_loss),
'two': (2, square_loss),
'four': (4, fourth_loss)
# 'one_fourth': (1/4, one_fourth_loss),
# 'one_half': (1/2, one_half_loss),
# 'one': (1, abs_loss),
# 'two': (2, square_loss),
# 'four': (4, fourth_loss),
# New entries:
'one_third': (1/3, one_third_loss),
'one_fifth': (1/5, one_fifth_loss),
'three': (3, three_loss),
'five': (5, five_loss),
}
if __name__ == "__main__":
@ -54,7 +72,7 @@ if __name__ == "__main__":
plt.figure(figsize=(10, 6))
for name, (exponent, loss_fn) in base_loss_functions.items():
# Compute loss for each error value (|theta - 0| = error)
# Compute loss for each error value
losses = loss_fn(errors, desired, min_val=0.01)
plt.plot(errors.numpy(), losses.numpy(), label=f"{name} (exp={exponent})")