kaimei

Official Python client for the Kaimei explainability API.

Kaimei (解明) — to clarify, elucidate, reveal through investigation — turns trained models into something you can inspect: architecture, feature-level attributions, global behaviour, learned rules, robustness, and drift.

from kaimei import Kaimei

client = Kaimei(api_key="km_live_...")

explanation = client.explanations.create(
    model="credit-risk-v3",
    prediction_id="pred_8f2a19",
    input={
        "income": 72000,
        "credit_score": 681,
        "utilization": 0.41,
    },
)

print(explanation.output)
print(explanation.top_features)
# 0.81 decline  |  credit_score -0.22

The same call as POST https://api.kaimei.ai/v1/explanations. See the product site for the language-agnostic REST examples.

Install

pip install kaimei

To extract and upload a fitted estimator you also need NumPy (and your training stack):

pip install "kaimei[models]"

For a local API (the kaimei-api service in the KaimeiAI/kaimei repo):

client = Kaimei(api_key="km_test_...", base_url="http://localhost:8000")

KAIMEI_API_KEY and KAIMEI_BASE_URL are read from the environment when arguments are omitted.

Create an API key in the dashboard (session login). Keys cannot create other keys.

Upload a trained model from a notebook

The client reads the live estimator in your process and uploads layers and learned weights. Small tensors travel inline in the JSON spec; anything larger than 65,536 values is streamed as a zlib-compressed float32 blob and referenced from the spec, so models with millions of parameters upload at about 4 bytes per weight. Kaimei does not store the original .h5 / .pt / pickle file, and it does not host a binary artifact.

from kaimei import Kaimei

client = Kaimei(api_key="km_live_...")

model = client.models.upload(
    clf,  # a fitted estimator
    name="Churn net",
    feature_names=["tenure", "spend", "tickets"],  # optional if the estimator knows them
    X=X_train,   # optional reference rows for explanations (capped at 5,000)
    y=y_train,
)
# Open the dashboard — the architecture scene shows every layer and tensor.

What happens next:

• Visualize — always. The dashboard draws the uploaded structure.
• Explain / predict / analyse — when the spec is executable (linear, MLP, decision tree, or a forest of ≤30 trees / ≤4,000 nodes). The API runs that structure in-process.
• Otherwise — pass a live scoring URL, or POST pre-computed attributions:

client.models.upload(
    large_forest,
    name="Churn forest",
    inference={"type": "remote", "url": "https://scoring.example.com/predict"},
)

Attach weights to a model you already registered with client.models.upload_structure(model.id, clf).

Large models

Nothing changes in the call. A PyTorch module with thousands of layers and millions of parameters uploads the same way; the client registers the model from its shapes, streams each large tensor (progress= reports ("tensors", done, total)), then sends the spec that references them. Module names become block paths, so the dashboard folds encoder.layers.0 … encoder.layers.47 into one ×48 block and loads a tensor only when you open it.

model = client.models.upload(
    transformer,                 # torch.nn.Module, 1,200 layers / 110M parameters
    name="Encoder",
    feature_names=["token_id"],
    progress=lambda stage, done, total: print(stage, done, "/", total),
)

Options:

• externalize_above= — values per tensor above which it uploads as binary (default 65,536; None keeps everything inline).
• weights="summary" — upload shapes and statistics only. The architecture is drawable; the weights never leave your environment. Explain / predict then need inference={"type": "remote", ...}.
• extract_structure(estimator, ...) returns the spec and the tensors without uploading, for inspection or a custom upload.
• client.models.tensors(model) lists what was uploaded; structure_parameter_values(model, "fc1.weight") reads a tensor back as float32.

A runnable walkthrough lives in examples/upload_sklearn.py and examples/upload_sklearn.ipynb.

Supported estimators

Unknown Keras / PyTorch layers that keep their input's shape are passed through and marked approximate. Those models visualize but cannot be queried unless you add a remote URL. Operations that change the shape and cannot be replayed (for example ShuffleNet's channel shuffle or an upsampling decoder) raise StructureError(code="unsupported_layers") naming them; register such models with a remote inference URL instead.

Tabular explanations in a notebook

kaimei.tabular turns explanations and insights into sentences and charts in the dashboard's colours (pip install matplotlib for the charts):

from kaimei.tabular import (
    compare_explanations, counterfactual_summary, explanation_summary,
    plot_counterfactual, plot_explanation, plot_insight,
)

model = client.models.retrieve("titanic-forest")
explanation = client.explanations.create(
    model=model.slug, input=passenger, method="shap",
    options={"counterfactual": True, "frozen_features": ["Sex", "Age"]},
)
print(explanation_summary(explanation, model=model))
# The model predicts died: P(survived) = 7.4%, against 41.1% on average over the reference data.
# Pulling P(survived) down: Sex = male (−12.1 pts), Title = Mr (−11.0 pts), Pclass = 3 (−4.5 pts).
plot_explanation(explanation, model=model)     # waterfall from the average to this prediction
print(counterfactual_summary(explanation, model=model))
plot_counterfactual(explanation, model=model)  # the smallest change that flips the decision, step by step
compare_explanations([explanation, other_model_explanation], ["Forest", "Logistic"], model=model)
for insight in client.insights.list(model=model.slug).data:
    plot_insight(insight, model=model)         # importance, rules, behaviour curves, interactions, drift, ...

The examples above come from titanic-kaimei-example. Passing model= adds feature units (Fare = 7.25 GBP) and names the output (P(survived)).

For a model Kaimei runs itself with up to 12 inputs, KernelSHAP enumerates every feature coalition over a fixed reference sample. The answer is therefore exact for that sample, and the same request returns the same numbers.

Image models

Images travel as nested height × width × channels lists, channels last, with pixels in [0, 1] unless the feature says otherwise. A PyTorch CNN needs its input size (PyTorch order); a Keras model reports its own.

import torchvision

net = torchvision.models.resnet18(num_classes=10).eval()
client.models.upload(
    net,
    name="Defect classifier",
    task="multiclass_classification",
    outputs={"labels": ["ok", "scratch", "dent", ...]},
    input_shape=(3, 224, 224),
    # torchvision-style models expect Normalize on [0, 1] pixels; bake it into the graph:
    preprocess={"mean": [0.485, 0.456, 0.406], "std": [0.229, 0.224, 0.225]},
    X=images,  # optional reference images, (n, 3, 224, 224) or (n, 224, 224, 3)
)

# A Keras model with a Rescaling(1/255) layer reads 0–255 pixels:
client.models.upload(keras_model, name="Leaf disease", value_range=[0, 255])

client.explanations.create(model="defect-classifier", input={"image": pixels}, method="gradcam")
client.explanations.create(model="defect-classifier", input={"image": pixels}, method="occlusion",
                           options={"grid": 8, "target_class": "dent"})

gradcam reads the last convolutional feature map of an uploaded structure (options={"layer": ...} picks another). occlusion and lime work for any image model, including a remote endpoint; that endpoint receives images as nested lists by default, or as {"b64": ...} PNG / float32 with inference={"image_encoding": "png"}, channels first with "image_layout": "chw", and normalized with "image_normalize": {"mean": [...], "std": [...]}. Requests are split to stay under max_request_bytes (8 MB).

To explain a new photo, kaimei.images (pip install "kaimei[images]") reads it at the model's input size and draws the result exactly as the dashboard's Explain tab does:

from kaimei.images import load_image, render_explanation, explanation_summary, plot_explanation

model = client.models.retrieve("cats-dogs-photos")
photo = load_image("my_dog.jpg", model=model)  # a path, bytes, a PIL image or an array
explanation = client.explanations.create(model=model.slug, input={"photo": photo}, method="gradcam")
print(explanation_summary(explanation))
# "The model looks mostly at the centre of the photo: half of its evidence for dog sits in 21% of the
#  picture. That region alone is enough to say dog (P(dog) = 97.5%); with it hidden, P(dog) falls
#  from 99.2% to 48.1%."
render_explanation(photo, explanation).save("my_dog_gradcam.png")                   # warm glow + focus outline
render_explanation(photo, explanation, mode="spotlight").save("my_dog_spotlight.png")  # the photo, dimmed outside the focus
plot_explanation(photo, explanation)  # matplotlib: the dashboard's view, with numbered regions (below)

load_image scales the shorter side to the model's size and crops the middle, exactly like torchvision's Resize(n) + CenterCrop(n) (crop="stretch" resizes instead), applies the photo's EXIF rotation, and returns pixels in the feature's value_range. The dashboard's Explain tab prepares uploaded photos the same way. NumPy arrays can also be passed straight into any request's input.

Heatmaps use the dashboard's palette: amber → orange → red where the photo is evidence for the explained class, light blue → indigo where it is evidence against it (the SHAP convention), never more than 66% opaque so the eyes and nose under the hot spot stay visible. A white outline marks the focus region, the smallest part of the photo holding half of the evidence. Grad-CAM maps are smoothed; occlusion and LIME keep one flat value per region.

Image explanations carry, besides options["heatmap"], the regions the attributions refer to (options["regions"]: superpixels named by position, upper left, centre 3, …, each with its attribution, area, centroid and an anchor point for a number badge; options["segments"] maps heatmap cells to regions), the focus region (options["focus"]), an evidence check (options["evidence"]: P(target) for the full photo, with the focus hidden and with only the focus shown) and a plain-English options["summary"]. Photo-sized models hide regions by blurring them (options["background"] == "blur", evidence["fill"] == "blur", with blur_sigma), because flat gray is not neutral to a real model; thumbnails under 64 pixels keep gray.

The helpers render all of it with NumPy and Pillow only, byte for byte like the dashboard: render_explanation(image, explanation, mode=...); overlay_heatmap(image, heatmap, method=..., mode=..., threshold=...) for raw heatmaps (without method it infers one: a map with negative values is a region map, anything else Grad-CAM); evidence_images(image, explanation) for the three evidence pictures, blurred with gaussian_blur (the API's exact blur) or grayed, as the API did; numbered_regions(explanation) for the regions to call out, numbered 1..N: up to five for the class, strongest first, then up to three against it; focus_mask / region_mask(explanation, "upper left", h, w) for pixel masks; and the building blocks upsample_heatmap, heat_color, heat_alpha and focus_threshold. plot_explanation (needs matplotlib) draws the dashboard's view: the photo, the heatmap with those numbers as badges at the regions' anchors, and the spotlight; below, the evidence pictures with their probabilities; on the right, a bar chart of the same numbered regions ("For cat" in orange, "Against cat" in blue; Grad-CAM lists its strongest regions), with the summary as the title. It returns the figure; pass figure= (a figure or subfigure) to draw several in one. Regression records print values with three significant figures instead of probabilities. import kaimei.images loads neither Pillow nor matplotlib.

examples/cats_dogs_gradcam.py trains a MobileNetV2 cat/dog classifier on the Oxford-IIIT Pet photos, uploads it with the photos in examples/photos attached as clickable examples, and explains new photos (explain prints the summary, the evidence check and the numbered regions, and saves the figure). examples/cats_dogs_explain.ipynb asks the model why on those photos and draws every answer as the dashboard does: Grad-CAM on a cat and a dog, "why cat?" and "why dog?" on the cat-and-dog composite (each question lights up its own animal and marks the other one against), Grad-CAM, occlusion and LIME side by side on the beagle, and a gallery in which Grad-CAM finds the face on every photo while LIME singles out the ears and eyes of the cats and the noses of the dogs.

The JSON spec warns above ~2 MB; lower externalize_above if you see it. Each binary tensor may hold up to 250M values (the API's default cap on a single upload is 512 MB compressed).

What you can do

# Register a model (or point Kaimei at a remote inference URL / built-in demo)
client.models.create(
    name="Credit risk v3",
    slug="credit-risk-v3",
    task="binary_classification",
    features=[
        {"name": "income", "type": "numeric"},
        {"name": "credit_score", "type": "integer"},
        {"name": "utilization", "type": "numeric"},
    ],
    inference={"type": "builtin", "name": "credit-risk"},
)

# Log a prediction your own stack already made
client.predictions.create(model="credit-risk-v3", input={...}, output={"score": 0.81})

# Explain one input. method is "shap" (KernelSHAP), "shapley", or "lime".
client.explanations.create(model="credit-risk-v3", input={...}, method="shap")

# Study the model as a whole, then read the insights
analysis = client.analyses.create(model="credit-risk-v3", type="full")
client.analyses.wait(analysis.id)
insights = client.insights.list(model="credit-risk-v3")

# Walk every page
for model in client.models.iterate():
    print(model.slug)

Dashboard sign-in (session token) is also supported via client.auth.login(...) / client.auth.signup(...). Production sign-up needs a Cloudflare Turnstile token (turnstile_token=...). Prefer creating API keys in the dashboard.

Resources

Errors raise kaimei.KaimeiError subclasses (AuthenticationError, PermissionDeniedError, NotFoundError, …) and expose .code, .type, and .status_code. PermissionError remains an alias of PermissionDeniedError.

429 responses are retried (honouring Retry-After). GET requests also retry 502/503/504. Default timeout is 120 seconds.

License

Apache 2.0. Copyright Kaimei AI.