Environment:Junyanz Pytorch CycleGAN and pix2pix Python PyTorch Runtime

Knowledge Sources	pytorch-CycleGAN-and-pix2pix PyTorch
Domains	Computer_Vision, Deep_Learning, Infrastructure
Last Updated	2026-02-09 16:00 GMT

Overview

Linux or macOS environment with Python 3.11, PyTorch 2.4+, CUDA 12.1, and visualization dependencies (dominate, wandb, Pillow).

Description

This environment provides the full runtime context for training and testing CycleGAN and pix2pix image-to-image translation models. It is built around a Conda environment specification (pytorch-img2img) that pins Python 3.11, PyTorch 2.4.0, torchvision 0.19.0, and CUDA 12.1. The code supports both GPU-accelerated and CPU-only execution (set --gpu_ids -1 for CPU mode). For GPU execution, any NVIDIA GPU with CUDA support is sufficient; no minimum VRAM is strictly enforced, though CycleGAN requires significantly more memory than pix2pix due to loading four networks simultaneously.

Usage

Use this environment for all CycleGAN and pix2pix workflows: training, testing, inference, and dataset preparation. It is the mandatory prerequisite for every Implementation in this repository. GPU acceleration is strongly recommended for training but optional for inference.

System Requirements

Category	Requirement	Notes
OS	Linux or macOS	Windows not officially supported; Docker available as alternative
Hardware	CPU or NVIDIA GPU + CUDA CuDNN	GPU strongly recommended for training
Python	3.11	Pinned in environment.yml
CUDA	12.1 (via pytorch-cuda)	Only required for GPU execution
Disk	~5GB for environment + dataset storage	Datasets vary in size (e.g., horse2zebra ~111MB)

Dependencies

System Packages

conda (Miniconda or Anaconda)
CUDA toolkit 12.1 (for GPU execution)
CuDNN (bundled with PyTorch conda package)

Python Packages

`torch` = 2.4.0
`torchvision` = 0.19.0
`numpy` = 1.24.3
`scikit-image`
`Pillow` >= 10.0.0
`dominate` >= 2.8.0
`wandb` >= 0.16.0 (optional, for logging)

Credentials

The following environment variables are optional:

`WANDB_API_KEY`: Weights & Biases API key, required only when using --use_wandb flag for training visualization.

Quick Install

# Option 1: Conda (recommended)
conda env create -f environment.yml
conda activate pytorch-img2img

# Option 2: Pip (minimal)
pip install torch==2.4.0 torchvision==0.19.0 numpy==1.24.3 Pillow>=10.0.0 dominate>=2.8.0 wandb>=0.16.0 scikit-image

Code Evidence

Device initialization from `util/util.py:52-69`:

def init_ddp():
    is_ddp = "WORLD_SIZE" in os.environ and int(os.environ["WORLD_SIZE"]) > 1
    if is_ddp:
        if not dist.is_initialized():
            dist.init_process_group(backend="nccl")
        local_rank = int(os.environ["LOCAL_RANK"])
        device = torch.device(f"cuda:{local_rank}")
        torch.cuda.set_device(local_rank)
    elif torch.cuda.is_available():
        device = torch.device("cuda:0")
        torch.cuda.set_device(0)
    else:
        device = torch.device("cpu")
    return device

Test-time device selection from `test.py:47`:

opt.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

Optional wandb import with graceful fallback from `test.py:39-42`:

try:
    import wandb
except ImportError:
    print('Warning: wandb package cannot be found. The option "--use_wandb" will result in error.')

PyTorch version requirement stated in `README.md:19`:

**Note**: The current software works well with PyTorch 2.4+.

Conda environment specification from `environment.yml`:

name: pytorch-img2img
channels:
  - pytorch
  - conda-forge
  - nvidia
dependencies:
  - python=3.11
  - pytorch=2.4.0
  - torchvision=0.19.0
  - pytorch-cuda=12.1
  - numpy=1.24.3
  - scikit-image
  - pip:
      - dominate>=2.8.0
      - Pillow>=10.0.0
      - wandb>=0.16.0

Common Errors

Error Message	Cause	Solution
`TypeError: Object of type 'Tensor' is not JSON serializable`	PyTorch version too old (< 2.4)	Upgrade to PyTorch 2.4+
`Warning: wandb package cannot be found`	wandb not installed	`pip install wandb` or remove `--use_wandb` flag
`NotSupportedError: slicing multiple dimensions`	torchvision/pytorch version mismatch	Reinstall matching torchvision for your PyTorch version
`My PyTorch errors on CUDA related code`	PyTorch not built with CUDA	Reinstall PyTorch with CUDA support, or use `--gpu_ids -1` for CPU

Compatibility Notes

macOS: CPU-only execution; no CUDA support available.
Docker: Pre-built image available at `taesungp/pytorch-cyclegan-and-pix2pix`. Requires nvidia-docker for GPU support.
CPU mode: Set `--gpu_ids -1` to run without GPU. Training will be very slow but functional.
Legacy PyTorch: For PyTorch 0.1-0.3, use the `pytorch0.3.1` branch instead.

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment