Mantis reproduction bundle

README.md

@@ -0,0 +1,13 @@
+# Mantis reproduction bundle
+
+Reproduction of **Mantis: Lightweight Foundation Model for Time Series Classification**
+(arXiv:2502.15637, OpenReview gbJMAjXLZ4) for the HF "Reproducing ICML 2026" challenge.
+
+Logbook: https://huggingface.co/spaces/ancs21/repro-mantis
+Paper code: https://github.com/vfeofanov/mantis  ·  Weights: https://huggingface.co/paris-noah/Mantis-8M
+
+- `param_count.py` — Claim 1: Mantis-8M has 8,103,936 params (8.10M), matching "8M lightweight". Verified.
+- `zeroshot_probe.py` — Claim 2: frozen embeddings + logistic regression on 4 UCR datasets,
+  mean accuracy 0.938 (well above chance). Frozen zero-shot features are useful. Verified.
+
+Install: `pip install mantis-tsfm aeon scikit-learn torch`. Inference-only, CPU, minutes.

param_count.py

@@ -0,0 +1,5 @@
+"""Claim 1: verify Mantis-8M parameter count (~8M lightweight). CPU, seconds."""
+from mantis.architecture import Mantis8M
+net = Mantis8M(device="cpu").from_pretrained("paris-noah/Mantis-8M")
+n = sum(p.numel() for p in net.parameters())
+print(f"total params: {n:,}  ({n/1e6:.2f}M)")   # 8,103,936 -> 8.10M

zeroshot_probe.py

@@ -0,0 +1,52 @@
+"""Mantis (#12379) claim check: frozen zero-shot features are useful.
+
+Loads small UCR datasets, extracts FROZEN Mantis-8M embeddings (no fine-tuning),
+fits a linear classifier (logistic regression) on top, reports test accuracy.
+If the frozen features are useful, accuracy should be well above chance and
+competitive with standard TS classifiers.
+"""
+import warnings; warnings.filterwarnings("ignore")
+import numpy as np, torch
+from mantis.architecture import Mantis8M
+from mantis.trainer import MantisTrainer
+from aeon.datasets import load_classification
+from sklearn.linear_model import LogisticRegression
+from sklearn.preprocessing import StandardScaler
+
+MANTIS_LEN = 512
+net = Mantis8M(device="cpu").from_pretrained("paris-noah/Mantis-8M")
+clf_wrap = MantisTrainer(device="cpu", network=net)
+
+def resample_to(x, L=MANTIS_LEN):
+    # x: [n, 1, T] -> [n, 1, L] via linear interpolation
+    t = torch.tensor(np.asarray(x), dtype=torch.float32)
+    if t.ndim == 3 and t.shape[1] != 1:  # take first channel if multivariate
+        t = t[:, :1, :]
+    t = torch.nn.functional.interpolate(t, size=L, mode="linear", align_corners=False)
+    return t.numpy()
+
+datasets = ["ECG200", "GunPoint", "Coffee", "ItalyPowerDemand"]
+print(f"{'dataset':18} {'n_tr':>5} {'n_te':>5} {'len':>5} {'chance':>7} {'frozen-acc':>11}")
+accs = []
+for name in datasets:
+    try:
+        Xtr, ytr = load_classification(name, split="train")
+        Xte, yte = load_classification(name, split="test")
+    except Exception as e:
+        print(f"{name:18} load failed: {e}"); continue
+    T0 = Xtr.shape[-1]
+    Etr = clf_wrap.transform(resample_to(Xtr))
+    Ete = clf_wrap.transform(resample_to(Xte))
+    sc = StandardScaler().fit(Etr)
+    lr = LogisticRegression(max_iter=2000).fit(sc.transform(Etr), ytr)
+    acc = lr.score(sc.transform(Ete), yte)
+    # majority-class baseline
+    vals, cnts = np.unique(ytr, return_counts=True)
+    chance = max(cnts) / len(ytr)
+    accs.append(acc)
+    print(f"{name:18} {len(ytr):>5} {len(yte):>5} {T0:>5} {chance:>7.3f} {acc:>11.3f}")
+
+if accs:
+    print(f"\nmean frozen-feature accuracy over {len(accs)} datasets: {np.mean(accs):.3f}")
+    print("Claim: frozen zero-shot Mantis features are useful (well above chance) -> "
+          + ("SUPPORTED" if np.mean(accs) > 0.8 else "WEAK"))