New metric here: Look at uploaded features that triggered okay_sense, and run through nets
uploading net: model_aug30_lstm_med_stateful_okay_sense_stop_snooze_tiny_end0_plus_1115_ep050
597 counts of okay_sense from Jan 14 to end of 17 (4 days)
86 / 597 false positives from this net
379 / 597 false positives from the aug30 net
Not apples to apples, since the thresholds are the same for each net... but roughly 1/5 the false positives and a much higher success rate. Let's see what the real world brings.
benejoseph
commented
7 years ago
27/30 in our anechoic chamber at 4 second delay,
21/30 at 2 second delay
New metric here: Look at uploaded features that triggered okay_sense, and run through nets uploading net: model_aug30_lstm_med_stateful_okay_sense_stop_snooze_tiny_end0_plus_1115_ep050 597 counts of okay_sense from Jan 14 to end of 17 (4 days)
86 / 597 false positives from this net 379 / 597 false positives from the aug30 net
Not apples to apples, since the thresholds are the same for each net... but roughly 1/5 the false positives and a much higher success rate. Let's see what the real world brings.