PyTorch Model Parts

PyTorch model parts provide an alternative deep learning framework to Keras for building autonomous driving models.

Overview

Donkeycar supports PyTorch models through a modular architecture similar to Keras parts. PyTorch models can be used interchangeably with Keras models in your vehicle. Location: donkeycar/parts/pytorch/

Getting Started

Installation

# Install PyTorch (CPU version)
pip install torch torchvision

# For NVIDIA Jetson (GPU)
pip install torch torchvision --extra-index-url https://developer.download.nvidia.com/compute/redist/jp/v461 tensorflow

# For Raspberry Pi (CPU optimized)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

Basic Usage

from donkeycar.parts.pytorch.torch_utils import get_model_by_type

# Create model
model = get_model_by_type(
    model_type='resnet18',
    cfg=cfg,
    checkpoint_path='./models/my_model.pth'
)

# Add to vehicle
V.add(model,
      inputs=['cam/image_array'],
      outputs=['pilot/angle', 'pilot/throttle'],
      run_condition='run_pilot')

Available Models

ResNet18

Transfer learning model based on ImageNet pre-trained ResNet18. Location: donkeycar/parts/pytorch/ResNet18.py

from donkeycar.parts.pytorch.ResNet18 import ResNet18

model = ResNet18(
    input_shape=(cfg.BATCH_SIZE, 3, 224, 224),  # BatchSize, Channels, Height, Width
    num_outputs=2  # angle, throttle
)

Architecture:

class ResNet18(nn.Module):
    def __init__(self, input_shape, num_outputs=2):
        super().__init__()
        
        # Load pre-trained ResNet18
        self.backbone = models.resnet18(pretrained=True)
        
        # Replace final layer
        num_features = self.backbone.fc.in_features
        self.backbone.fc = nn.Sequential(
            nn.Linear(num_features, 512),
            nn.ReLU(),
            nn.Dropout(0.5),
            nn.Linear(512, num_outputs)
        )
    
    def forward(self, x):
        return self.backbone(x)

Features:

Transfer Learning: Leverages ImageNet pre-trained weights
Fixed Input Size: Requires 224x224x3 images
High Accuracy: Better feature extraction than training from scratch
GPU Recommended: Computationally intensive

Configuration: In myconfig.py:

# PyTorch settings
DEFAULT_MODEL_TYPE = 'resnet18'
MODEL_PATH = './models/resnet18.pth'

# Image size (ResNet18 requires 224x224)
IMAGE_W = 224
IMAGE_H = 224
IMAGE_DEPTH = 3

# Training
BATCH_SIZE = 16  # Adjust based on GPU memory
LEARNING_RATE = 0.0001
EPOCHS = 10

Model Utilities

get_model_by_type

Location: donkeycar/parts/pytorch/torch_utils.py:4

from donkeycar.parts.pytorch.torch_utils import get_model_by_type

model = get_model_by_type(
    model_type='resnet18',  # Model architecture
    cfg=cfg,                 # Configuration object
    checkpoint_path=None     # Optional: path to saved weights
)

Supported Model Types:

'resnet18' - ResNet18 transfer learning model
(Additional models can be added)

Training PyTorch Models

Training Script

Location: donkeycar/parts/pytorch/torch_train.py

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from donkeycar.parts.pytorch.torch_data import DonkeyDataset
from donkeycar.parts.pytorch.ResNet18 import ResNet18

# Load configuration
import donkeycar as dk
cfg = dk.load_config()

# Create datasets
train_dataset = DonkeyDataset(
    tub_paths=['./data/tub_1', './data/tub_2'],
    config=cfg,
    mode='train'
)

val_dataset = DonkeyDataset(
    tub_paths=['./data/tub_3'],
    config=cfg,
    mode='val'
)

# Create data loaders
train_loader = DataLoader(
    train_dataset,
    batch_size=cfg.BATCH_SIZE,
    shuffle=True,
    num_workers=4
)

val_loader = DataLoader(
    val_dataset,
    batch_size=cfg.BATCH_SIZE,
    shuffle=False,
    num_workers=4
)

# Create model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = ResNet18(input_shape=(cfg.BATCH_SIZE, 3, 224, 224))
model = model.to(device)

# Loss and optimizer
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=cfg.LEARNING_RATE)

# Training loop
for epoch in range(cfg.EPOCHS):
    model.train()
    train_loss = 0.0
    
    for images, labels in train_loader:
        images = images.to(device)
        labels = labels.to(device)
        
        # Forward pass
        outputs = model(images)
        loss = criterion(outputs, labels)
        
        # Backward pass
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        train_loss += loss.item()
    
    # Validation
    model.eval()
    val_loss = 0.0
    
    with torch.no_grad():
        for images, labels in val_loader:
            images = images.to(device)
            labels = labels.to(device)
            outputs = model(images)
            loss = criterion(outputs, labels)
            val_loss += loss.item()
    
    print(f'Epoch {epoch+1}/{cfg.EPOCHS}')
    print(f'Train Loss: {train_loss/len(train_loader):.4f}')
    print(f'Val Loss: {val_loss/len(val_loader):.4f}')
    
    # Save checkpoint
    torch.save({
        'epoch': epoch,
        'model_state_dict': model.state_dict(),
        'optimizer_state_dict': optimizer.state_dict(),
        'loss': train_loss,
    }, f'./models/checkpoint_epoch_{epoch}.pth')

# Save final model
torch.save(model.state_dict(), './models/resnet18_final.pth')

Training Command

cd ~/mycar
python -m donkeycar.parts.pytorch.torch_train --tub ./data --model ./models/resnet18.pth

Dataset Handling

DonkeyDataset

Location: donkeycar/parts/pytorch/torch_data.py PyTorch Dataset class for loading Donkeycar tub data.

from donkeycar.parts.pytorch.torch_data import DonkeyDataset
import torch
from torch.utils.data import DataLoader

# Create dataset
dataset = DonkeyDataset(
    tub_paths=['./data/tub_1', './data/tub_2'],
    config=cfg,
    mode='train',  # 'train' or 'val'
    transform=None  # Optional transforms
)

# Create data loader
loader = DataLoader(
    dataset,
    batch_size=32,
    shuffle=True,
    num_workers=4,
    pin_memory=True  # For GPU training
)

# Iterate
for images, labels in loader:
    # images: (batch_size, channels, height, width)
    # labels: (batch_size, 2) - angle, throttle
    pass

Data Augmentation:

import torchvision.transforms as transforms

transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.RandomHorizontalFlip(p=0.5),
    transforms.ColorJitter(
        brightness=0.2,
        contrast=0.2,
        saturation=0.2,
        hue=0.1
    ),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485, 0.456, 0.406],  # ImageNet stats
        std=[0.229, 0.224, 0.225]
    )
])

dataset = DonkeyDataset(
    tub_paths=['./data'],
    config=cfg,
    transform=transform
)

Model Inference

Loading Saved Models

import torch
from donkeycar.parts.pytorch.ResNet18 import ResNet18

# Create model
model = ResNet18(input_shape=(1, 3, 224, 224))

# Load weights
checkpoint = torch.load('./models/resnet18.pth')
model.load_state_dict(checkpoint['model_state_dict'])

# Set to evaluation mode
model.eval()

# Use GPU if available
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = model.to(device)

Running Inference

import torch
import numpy as np
from PIL import Image
import torchvision.transforms as transforms

# Preprocessing
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
    )
])

# Load and preprocess image
img = Image.open('test_image.jpg')
img_tensor = transform(img).unsqueeze(0)  # Add batch dimension
img_tensor = img_tensor.to(device)

# Inference
with torch.no_grad():
    output = model(img_tensor)
    angle, throttle = output[0].cpu().numpy()

print(f"Predicted - Angle: {angle:.3f}, Throttle: {throttle:.3f}")

In Vehicle Loop

class PyTorchPilot:
    def __init__(self, model_path, device='cpu'):
        self.device = torch.device(device)
        self.model = ResNet18(input_shape=(1, 3, 224, 224))
        self.model.load_state_dict(torch.load(model_path))
        self.model.to(self.device)
        self.model.eval()
        
        self.transform = transforms.Compose([
            transforms.ToPILImage(),
            transforms.Resize((224, 224)),
            transforms.ToTensor(),
            transforms.Normalize(
                mean=[0.485, 0.456, 0.406],
                std=[0.229, 0.224, 0.225]
            )
        ])
    
    def run(self, img_arr):
        # Preprocess
        img_tensor = self.transform(img_arr).unsqueeze(0)
        img_tensor = img_tensor.to(self.device)
        
        # Inference
        with torch.no_grad():
            output = self.model(img_tensor)
            angle, throttle = output[0].cpu().numpy()
        
        return angle, throttle
    
    def shutdown(self):
        pass

# In manage.py
pilot = PyTorchPilot('./models/resnet18.pth', device='cuda')
V.add(pilot,
      inputs=['cam/image_array'],
      outputs=['pilot/angle', 'pilot/throttle'],
      run_condition='run_pilot')

Model Export

TorchScript (JIT Compilation)

import torch
from donkeycar.parts.pytorch.ResNet18 import ResNet18

# Load model
model = ResNet18(input_shape=(1, 3, 224, 224))
model.load_state_dict(torch.load('./models/resnet18.pth'))
model.eval()

# Create example input
example_input = torch.rand(1, 3, 224, 224)

# Trace and export
traced_model = torch.jit.trace(model, example_input)
traced_model.save('./models/resnet18_traced.pt')

# Load traced model
loaded_model = torch.jit.load('./models/resnet18_traced.pt')

ONNX Export

import torch
import torch.onnx

# Load model
model = ResNet18(input_shape=(1, 3, 224, 224))
model.load_state_dict(torch.load('./models/resnet18.pth'))
model.eval()

# Create example input
dummy_input = torch.randn(1, 3, 224, 224)

# Export to ONNX
torch.onnx.export(
    model,
    dummy_input,
    './models/resnet18.onnx',
    export_params=True,
    opset_version=11,
    input_names=['input'],
    output_names=['output'],
    dynamic_axes={
        'input': {0: 'batch_size'},
        'output': {0: 'batch_size'}
    }
)

Configuration

In myconfig.py:

# PyTorch Model
DEFAULT_MODEL_TYPE = 'resnet18'
MODEL_PATH = './models/resnet18.pth'

# Device
USE_CUDA = True  # Use GPU if available
DEVICE = 'cuda' if USE_CUDA and torch.cuda.is_available() else 'cpu'

# Training
BATCH_SIZE = 16
LEARNING_RATE = 0.0001
WEIGHT_DECAY = 1e-5
EPOCHS = 10

# Image preprocessing (ResNet18 requirements)
IMAGE_W = 224
IMAGE_H = 224
IMAGE_DEPTH = 3

# Data augmentation
AUG_FLIP_HORIZONTAL = True
AUG_BRIGHTNESS = 0.2
AUG_CONTRAST = 0.2
AUG_SATURATION = 0.2

# Training/validation split
TRAIN_TEST_SPLIT = 0.8

Creating Custom PyTorch Models

import torch
import torch.nn as nn

class CustomPilot(nn.Module):
    def __init__(self, input_shape=(1, 3, 120, 160)):
        super().__init__()
        
        self.conv_layers = nn.Sequential(
            nn.Conv2d(3, 24, kernel_size=5, stride=2),
            nn.ReLU(),
            nn.Conv2d(24, 32, kernel_size=5, stride=2),
            nn.ReLU(),
            nn.Conv2d(32, 64, kernel_size=3, stride=1),
            nn.ReLU(),
            nn.Flatten()
        )
        
        # Calculate flattened size
        with torch.no_grad():
            dummy_input = torch.zeros(input_shape)
            conv_output_size = self.conv_layers(dummy_input).shape[1]
        
        self.fc_layers = nn.Sequential(
            nn.Linear(conv_output_size, 100),
            nn.ReLU(),
            nn.Dropout(0.5),
            nn.Linear(100, 50),
            nn.ReLU(),
            nn.Dropout(0.5),
            nn.Linear(50, 2)  # angle, throttle
        )
    
    def forward(self, x):
        x = self.conv_layers(x)
        x = self.fc_layers(x)
        return x

# Usage
model = CustomPilot(input_shape=(1, 3, 120, 160))
print(model)

Performance Optimization

Mixed Precision Training

from torch.cuda.amp import autocast, GradScaler

scaler = GradScaler()

for images, labels in train_loader:
    images = images.to(device)
    labels = labels.to(device)
    
    optimizer.zero_grad()
    
    # Mixed precision
    with autocast():
        outputs = model(images)
        loss = criterion(outputs, labels)
    
    # Scaled backpropagation
    scaler.scale(loss).backward()
    scaler.step(optimizer)
    scaler.update()

Model Pruning

import torch.nn.utils.prune as prune

# Prune 20% of weights
for name, module in model.named_modules():
    if isinstance(module, nn.Conv2d) or isinstance(module, nn.Linear):
        prune.l1_unstructured(module, name='weight', amount=0.2)
        prune.remove(module, 'weight')  # Make pruning permanent

Common Issues

CUDA Out of Memory

Reduce batch size
Use gradient accumulation
Enable mixed precision training
Clear cache: torch.cuda.empty_cache()

Model Not Learning

Check learning rate (try 1e-4 to 1e-3)
Verify data preprocessing
Check for label errors
Monitor gradients for vanishing/exploding

Inference Too Slow

Convert to TorchScript
Use GPU if available
Batch multiple frames
Optimize image preprocessing

Next Steps

Keras Models - Alternative framework
Data Stores - Prepare training data
Custom Parts - Create custom models

Getting Started

Core Concepts

Building Your Car

Training & Autopilots

Parts Reference

Advanced Topics

PyTorch Model Parts

Overview

Getting Started

Installation

Basic Usage

Available Models

ResNet18

Model Utilities

get_model_by_type

Training PyTorch Models

Training Script

Training Command

Dataset Handling

DonkeyDataset

Model Inference

Loading Saved Models

Running Inference

In Vehicle Loop

Model Export

TorchScript (JIT Compilation)

ONNX Export

Configuration

Creating Custom PyTorch Models

Performance Optimization

Mixed Precision Training

Model Pruning

Common Issues

CUDA Out of Memory

Model Not Learning

Inference Too Slow

Next Steps

Build docs developers (and LLMs) love

Getting Started

Core Concepts

Building Your Car

Training & Autopilots

Parts Reference

Advanced Topics

​Overview

​Getting Started

​Installation

​Basic Usage

​Available Models

​ResNet18

​Model Utilities

​get_model_by_type

​Training PyTorch Models

​Training Script

​Training Command

​Dataset Handling

​DonkeyDataset

​Model Inference

​Loading Saved Models

​Running Inference

​In Vehicle Loop

​Model Export

​TorchScript (JIT Compilation)

​ONNX Export

​Configuration

​Creating Custom PyTorch Models

​Performance Optimization

​Mixed Precision Training

​Model Pruning

​Common Issues

​CUDA Out of Memory

​Model Not Learning

​Inference Too Slow

​Next Steps

Build docs developers (and LLMs) love

Overview

Getting Started

Installation

Basic Usage

Available Models

ResNet18

Model Utilities

get_model_by_type

Training PyTorch Models

Training Script

Training Command

Dataset Handling

DonkeyDataset

Model Inference

Loading Saved Models

Running Inference

In Vehicle Loop

Model Export

TorchScript (JIT Compilation)

ONNX Export

Configuration

Creating Custom PyTorch Models

Performance Optimization

Mixed Precision Training

Model Pruning

Common Issues

CUDA Out of Memory

Model Not Learning

Inference Too Slow

Next Steps