| --- |
| license: mit |
| tags: |
| - location-encoder |
| - geospatial |
| - earth-observation |
| --- |
| |
| # MIND |
|
|
| MIND (Matryoshka Implicit Neural Distillation) maps a coordinate to a geospatial embedding. The released main |
| checkpoint takes `(lat, lon)` in degrees and returns a 3,072-dimensional trunk without imagery or labels at inference. |
| The first 64 dimensions are the default deployment embedding, and other leading prefixes can be selected without |
| retraining the encoder. |
|
|
| The model is distilled on the MINDSET dataset which is composed of four teacher sources: AlphaEarth Foundations, Climplicit, GeoCLIP, and SINR |
|
|
| ## Files |
|
|
| - `mind.safetensors` — recommended fp16 weights (226,980,656 bytes; safe tensor format). |
| - `mind.pt` — fp32 PyTorch weights (453,967,275 bytes; pickle-based loading). |
| - `mind.onnx` and `mind.onnx.data` — ONNX graph and external weights. |
| - `mind.pt2` — PyTorch `ExportedProgram`. |
|
|
| ## Usage |
|
|
| The public ONNX release can be loaded directly from the Hub. Install `numpy`, `onnxruntime`, and `huggingface_hub`. |
|
|
| ```python |
| import os |
| import numpy as np |
| import onnxruntime as ort |
| |
| session = ort.InferenceSession(os.path.join(folder, "mind.onnx"), providers=["CPUExecutionProvider"]) |
| coordinates = np.array([[37.77, -122.42], [51.51, -0.13]], dtype=np.float32) # (lat, lon) |
| embedding = session.run(None, {"latlon": coordinates})[0] # shape [2, 3072] |
| deploy_embedding = embedding[:, :64] |
| ``` |
|
|
| The ONNX graph expects `latlon` with shape `[N, 2]` in `(lat, lon)` order and returns the full 3,072-dimensional trunk. |
| Use the leading 64 columns as the deployment embedding. The safetensors and PyTorch files are also available for users |
| who have a compatible loader. |
|
|
