File size: 3,357 Bytes
884b8f8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 | import torch
import numpy as np
import mir_eval
from marble.tasks.GTZANBeatTracking.modules import TimeEventFMeasure
from marble.utils.utils import mask_to_times
def test_time_event_fmeasure():
"""
Test suite for TimeEventFMeasure. Uses synthetic masks to validate behavior.
"""
label_freq = 10 # 10 frames per second
tol = 0.07 # 70 ms tolerance
# Helper: builds a mask of length T with events at given times (in seconds)
def make_mask(event_times, T, fps):
mask = np.zeros(T, dtype=np.float32)
frames = np.round(np.array(event_times) * fps).astype(int)
valid = (frames >= 0) & (frames < T)
mask[frames[valid]] = 1.0
return torch.tensor(mask).unsqueeze(0) # shape (1, T)
T = 20 # total frames (2 seconds at 10 fps)
# 1. Both predicted and reference are empty → F1 = 1.0
metric1 = TimeEventFMeasure(label_freq=label_freq, tol=tol)
est1 = torch.zeros((1, T))
ref1 = torch.zeros((1, T))
metric1.update(est1, ref1)
f1_1 = metric1.compute().item()
print(f"Test 1 (both empty): F1 = {f1_1:.2f} (expected 1.00)")
# 2. Single matching event at exactly the same time → F1 = 1.0
metric2 = TimeEventFMeasure(label_freq=label_freq, tol=tol)
# Event at t = 1.0 s → frame index 10
est2 = make_mask([1.0], T, label_freq)
ref2 = make_mask([1.0], T, label_freq)
metric2.update(est2, ref2)
f1_2 = metric2.compute().item()
print(f"Test 2 (perfect match): F1 = {f1_2:.2f} (expected 1.00)")
# 3. Reference has an event, prediction is empty → F1 = 0.0
metric3 = TimeEventFMeasure(label_freq=label_freq, tol=tol)
est3 = torch.zeros((1, T))
ref3 = make_mask([0.5], T, label_freq) # event at 0.5 s → frame 5
metric3.update(est3, ref3)
f1_3 = metric3.compute().item()
print(f"Test 3 (ref only): F1 = {f1_3:.2f} (expected 0.00)")
# 4. Prediction has an event, reference is empty → F1 = 0.0
metric4 = TimeEventFMeasure(label_freq=label_freq, tol=tol)
est4 = make_mask([0.5], T, label_freq) # frame 5
ref4 = torch.zeros((1, T))
metric4.update(est4, ref4)
f1_4 = metric4.compute().item()
print(f"Test 4 (est only): F1 = {f1_4:.2f} (expected 0.00)")
# 5a. Two events on each side; one match, one miss → F1 = 0.50
metric5a = TimeEventFMeasure(label_freq=label_freq, tol=tol)
# References at t = 1.0 s (frame 10) and t = 1.5 s (frame 15)
ref5a = make_mask([1.0, 1.5], T, label_freq)
# Predictions at t = 1.0 s (match) and t = 1.6 s (frame 16, 0.1 s away from 1.5)
est5a = make_mask([1.0, 1.6], T, label_freq)
metric5a.update(est5a, ref5a)
f1_5a = metric5a.compute().item()
print(f"Test 5a (2 ref vs. 2 pred, one match): F1 = {f1_5a:.2f} (expected 0.50)")
# 5b. One reference event, two predicted events; only one match → F1 ≈ 0.67
metric5b = TimeEventFMeasure(label_freq=label_freq, tol=tol)
# Reference only at t = 1.0 s (frame 10)
ref5b = make_mask([1.0], T, label_freq)
# Predictions at t = 1.0 s (match) and t = 1.6 s (non‐match)
est5b = make_mask([1.0, 1.6], T, label_freq)
metric5b.update(est5b, ref5b)
f1_5b = metric5b.compute().item()
print(f"Test 5b (1 ref vs. 2 pred, one match): F1 = {f1_5b:.2f} (expected ~0.67)")
if __name__ == "__main__":
test_time_event_fmeasure()
|