Installation¶

Recommended¶

pip install 'rpx-benchmark[depth]'

This pulls the core library, the HuggingFace downloader, the rich terminal UI, torch, torchvision, and transformers — enough to run the full monocular absolute depth slate from the CLI.

From source¶

git clone https://github.com/IRVLUTD/RPX.git
cd RPX/benchmark
pip install -e '.[depth,dev,docs]'

Optional extras matrix¶

Extra	What it installs	When to use
(core)	`numpy`, `Pillow`	Always — manifest loading, metric computation
`hub`	`huggingface_hub[hf_xet]`	Pulling datasets from HuggingFace
`ui`	`rich`	Pretty terminal output (progress bar, tables)
`depth-hf`	`torch`, `torchvision`, `transformers`, `accelerate`	Any HuggingFace depth model (DA-v2, Depth Pro, ZoeDepth, PromptDA)
`depth-unidepth`	`depth-hf` + `einops`, `timm`, `huggingface_hub`	UniDepth V2
`depth-metric3d`	`depth-hf` + `timm`, `mmcv-lite`, `mmengine`	Metric3D V2 (CUDA only)
`depth`	`depth-hf` + `ui`	Recommended default for monocular depth
`depth-all`	everything above	Monocular depth, full slate
`dev`	`pytest`, `ruff`	Contributing
`docs`	`mkdocs`, `mkdocs-material`, `mkdocstrings[python]`	Building docs locally

Models that need manual installs¶

UniDepth V2¶

Ships via GitHub (not PyPI):

pip install 'rpx-benchmark[depth-unidepth]'
pip install 'unidepth @ git+https://github.com/lpiccinelli-eth/UniDepth.git'

Metric3D V2¶

pip install 'rpx-benchmark[depth-metric3d]'
# Weights are downloaded via torch.hub on first use; no install step.

CUDA-only

Metric3D V2's upstream decoder hardcodes torch.linspace(..., device="cuda"), so CPU inference is not supported. The adapter raises a clean error on CPU rather than failing mid-inference.

Verify your install¶

rpx --help
rpx ls                 # lists tasks + splits
rpx models             # lists runnable + deferred model adapters

All three should exit 0. If you installed [depth] you can also run a synthetic smoke without downloading a real dataset:

python -c "
import rpx_benchmark as rpx
import numpy as np

def fake_depth(rgb):
    return np.full(rgb.shape[:2], 2.0, dtype=np.float32)

bm = rpx.make_numpy_depth_model(fake_depth)
print(f'BenchmarkableModel ready: {bm.name}')
"