visynx

Run your Visynx computer-vision models from Python.

Train a model in Visynx, grab an API key, and you are three lines from predictions — detection, segmentation, pose, classification, and semantic segmentation, all through one call.

pip install visynx

Quickstart

from visynx import Visynx

vf = Visynx(api_key="vf_...")          # or set VISYNX_API_KEY
model = vf.model("shelf-scanner/3")    # your project name and version

result = model.predict("shelf.jpg", confidence=0.4)

for d in result:
    print(d.class_name, round(d.confidence, 2), d.box)

result.save("annotated.jpg")

bottle 0.94 [412.0, 118.5, 501.2, 388.0]
bottle 0.91 [502.6, 121.0, 590.4, 390.7]
can    0.78 [604.1, 210.3, 664.9, 372.2]

visynx pulls in only requests and Pillow. No OpenCV, no numpy, no torch — the models run on Visynx's GPUs, not yours.

Naming a model

Your project and its training version — the same name the app shows you:

model = vf.model(train_name="shelf-scanner", version="3")
model = vf.model("shelf-scanner/3")          # the same thing, as one handle
model = vf.model(train_name="shelf-scanner") # newest version

Or take the model with you

Professional and Enterprise accounts can download the weights and run them themselves. Every plan can keep using hosted predict() instead — downloading is an option, not a step.

pip install "visynx[local]"

vf = Visynx(api_key="vf_...")
model = vf.model(train_name="shelf-scanner", version="3")

model.download(path="./models/shelf-scanner")                # native weights
model.download(path="./models/shelf-scanner", format="onnx") # or a portable export
model.download(path="./models/shelf-scanner", format="edge") # ~4x smaller, for devices

local = vf.load_model("./models/shelf-scanner")   # or visynx.load_model(...)
result = local.predict("image.jpg", confidence=0.25)
result.save("annotated.jpg")

The check is the server's: an account without the feature gets a PlanRequiredError and no bytes, whether the call comes from here or from curl. Downloading someone else's model is refused the same way.

One local call covers every export and every kind of model — detection, instance segmentation, pose, classification, semantic segmentation, in FP32, FP16 or Edge. visynx reads what the file is, feeds it the way it expects, and returns predictions in your image's own pixel coordinates: boxes, outlines, keypoints, or a blended mask — the same Result the hosted call returns.

local.labels                                 # class names, read from the file
local.runtime                                # 'portable' or 'native'
local.device                                 # where it is running

Choosing where it runs

Any format — a checkpoint or any of the exports — takes a device:

vf.load_model("./models/x")                  # "auto": the GPU when there is one
vf.load_model("./models/x", device="cpu")    # keep it off the GPU entirely
vf.load_model("./models/x", device="gpu")    # insist on the GPU

"gpu" is a requirement rather than a preference, so a machine without one gets a VisynxError saying what is missing — never a silent fall back to the CPU at a tenth of the speed. "auto" is the one that falls back, by design.

Portable (.onnx) exports need nothing but visynx[local]. A native checkpoint also needs the runtime it was trained with; python -m visynx setup installs it.

Find your models

for m in vf.models():
    print(m.slug, m.display_name)
# shelf-scanner/3   Shelf Scanner v3
# defect-seg/1      Defect Seg v1

"shelf-scanner/3" is the model's slug — your project name and its training version, the same name you see in the app. A raw "yolo/<job_id>" also works and skips the lookup.

What you get back

predict() returns a Result: iterate it for detections, or reach for the raw API payload with .json().

result = model.predict("shelf.jpg")

len(result)              # 3
result.detections[0]     # Detection('bottle', 0.94)
result.json()            # the raw API response, untouched

result.plot()            # a PIL image with boxes drawn on
result.save("out.jpg")   # ...written to disk

Each Detection carries class_id, class_name, confidence, and box ([x1, y1, x2, y2] in absolute pixels). Segmentation models add polygons; pose models add keypoints.

Anything image-shaped works

model.predict("shelf.jpg")                  # a path
model.predict("frames/")                    # a directory -> list of Results
model.predict(open("shelf.jpg", "rb"))      # an open file
model.predict(pil_image)                    # a PIL image
model.predict(numpy_array)                  # a numpy array (from OpenCV, say)

Beyond boxes

Two flags the hosted GPUs do for you, with nothing extra to install. Each runs a second model, so each costs extra credits:

# Turn every detected box into a precise instance polygon with SAM 2
result = model.predict("shelf.jpg", segment=True)
result.detections[0].polygons

# Add a monocular depth map, and a relative distance per detection
result = model.predict("shelf.jpg", depth=True)
result.detections[0].depth_median      # 0-1, higher is closer
result.depth_image.save("depth.png")

Semantic-segmentation models return a dense mask instead of boxes:

result = model.predict("road.jpg")
result.semantic["legend"]     # [{'class_id': 1, 'class_name': 'road', ...}]
result.save("mask.jpg")       # the mask, blended over your image

Get an API key

In Visynx, go to Settings → API Keys, create a key, and copy the secret — it is shown once. Keys start with vf_, never expire, and can be locked to specific IPs or CIDR ranges. Keep it out of source control:

vf = Visynx()                      # reads VISYNX_API_KEY

A predict() call spends 1 Vision Credit on the checkpoint or the full-precision graph, and 2 on the optimised runtimes (format="onnx_fp16" or format="edge"); format="auto" is billed for whichever it runs. segment=True and depth=True each add a surcharge for the extra model they run. You are charged only after the inference succeeds — a failed call is free — and a call retried after a 429 is billed once, not twice.

Some endpoints are deliberately not reachable with an API key: batch prediction over a whole dataset, live-camera inference, workflow runs, and the AI-labelling tools (SAM/DINO/box-prompt). Those run from the Visynx app with a signed-in session. Calling them with a key returns PermissionDeniedError.

Errors

Every failure is a typed exception, and it quotes the response's X-Request-ID when the server sent one — include it if you contact support. (Authentication failures are rejected before that header is assigned, so 401s carry none.)

429s are retried automatically, honouring the server's Retry-After. All of these subclass VisynxError, so one except VisynxError catches everything.

Self-hosting

Point the client at your own deployment:

vf = Visynx(api_key="vf_...", api_url="https://visynx.internal/api")

or set VISYNX_API_URL.

License

MIT