# 5th place solution

Thanks to the organizers for an interesting contest and
[@cpmpml](https://www.kaggle.com/cpmpml)
for the motivation.
My decision is based on my public code posted here:
[https://www.kaggle.com/c/birdsong-recognition/discussion/183269](https://www.kaggle.com/c/birdsong-recognition/discussion/183269)
[https://github.com/vlomme/Birdcall-Identification-competition](https://github.com/vlomme/Birdcall-Identification-competition)
The basic solution easily gets silver. What I changed:
Switched to SED +1%.
Lowered threshold +1% .
Left only birds from region of record +0%
Changed ensemble averaging +1%
Pre-processing:
Used log-melspectrograms
n_fft = 1536, sr = 21952, hop_length = 245, n_mels = 224, len_chack 448, image_size = 224 * 448, 5 seconds
Models:
Ensemble of 14 models (sed_resnet50, sed_resnest50, sed _efficientnet-b0)
Augmentations:
For contrast, I raised the image to a power of 0.5 to 3. at 0.5, the background noise is closer to the birds, and at 3, on the contrary, the quiet sounds become even quieter.
Slightly accelerated / slowed down recording
Add a different sound without birds(rain, noise, conversations, etc.)
Added white, pink, and band noise. Increasing the noise level increases recall, but reduces precision.
With a probability of 0.5 lowered the upper frequencies. In the real world, the upper frequencies fade faster with distance
Train:
Used BCEWithLogitsLoss. For the main birds, the label was 1. For birds in the background 0.3.
Used loss:
train_los1 = nn.BCEWithLogitsLoss()(prediction['clipwise_output'], true)
train_los2 = nn.BCEWithLogitsLoss()(prediction["segmentwise_output_max"], true)
train_loss = (train_los1 + train_los2)/2
I didn't look at metrics on training records, but only on validation files (train_soundscapes)
20-40 epochs (24 hours on gtx1060)
Postprocessing
If there was a bird in the segment, I increased the probability of finding it in the entire file.
Model ensemble averaging
proba1 = proba.prod(axis = 0) ** (1.0/len(proba))
proba = proba**2
proba = proba.mean(axis=0)
proba = proba**(1/2)
proba = (proba + proba1)/2
Since the training took a long time and other approaches didn't work, I switched to other competitions and have hardly taught any new models in the last month. It's a shame that a little was not enough, I will try to do better