Past Competition Discussion Survey

[osuossu8]

スレッドタイトル

• コンペ名
• スレッドリンク

メモ

• fuga
• piyo

---

ぶらさがってるコメントなど

• foo

---

それらを受けてのアイデア、コメントなど

• hoge

[osuossu8]

1st Place Solution

• Cornell Birdcall Identification
• https://www.kaggle.com/c/birdsong-recognition/discussion/183208

メモ

• Data Augmentation

  • Pink noise
  • Gaussian noise
  • Gaussian SNR
  • Gain (Volume Adjustment)

• Models

  • SED model had over 80 million parameters so I switched all my models to use a pretrained densenet121 model as the cnn feature extractor and reduced the attention block size to 1024
  • I also replaced the clamp on the attention with tanh

• Training

  • Cosine Annealing Scheduler with warmup
  • Mixup
  • 50 epochs for non-mixup models and 100 epochs for mixup models
  • AdamW with weight_decay 0.01
  • SpecAugmentation enabled
  • 30 second audio clips during training and evaluating on 2 30 second clips per audio.

• I used a threshold of 0.3 on the framewise_output and 0.3 on the clipwise_output to reduce the impact of false positives.

• inference I also applied 10 TTA by just adding the same audio sample 10 times in the batch and enabling Spec Augmentation.

• inference 実装

• repo

---

ぶらさがってるコメントなど

• foo

---

それらを受けてのアイデア、コメントなど

• SED densenet121 backbone はやってみたい
• TTA もやる

#TTA の実装部分

    for image in tensors:
        image = image.unsqueeze(0).unsqueeze(0)
        image = image.expand(image.shape[0], TTA, image.shape[2])
        image = image.to(device)

        with torch.no_grad():
            prediction = model((image, None))
            framewise_outputs = prediction["framewise_output"].detach(
                ).cpu().numpy()[0].mean(axis=0)
            clipwise_outputs = prediction["clipwise_output"].detach(
                ).cpu().numpy()[0].mean(axis=0)