483. Happywhale - Whale and Dolphin Identification | happy-whale-and-dolphin
| 模型 | 训练图像尺寸 | 推理图像尺寸 |
|---|---|---|
| B5 | 960 | 1056 |
| B6 | 768 | 840 |
| B7 | 600 | 660 |
image = tf.image.random_flip_left_right(image)
image = tf.image.random_hue(image, 0.1)
image = tf.image.random_saturation(image, 0.80, 1.20)
image = tf.image.random_contrast(image, 0.80, 1.20)
image = tf.image.random_brightness(image, 0.10)
编码器 {B5/B6/B7} -> GAP {全局平均池化} -> 批归一化 -> 多样本Dropout (Multi-SampleDropout)
多样本Dropout的输出被送入两个Arcface分类头:
individual_margin = head(n_classes = config.N_CLASSES, s = 30, m = 0.3, name=f'head_individual/{config.head}', dtype='float32')
species_margin = head(n_classes = config.N_SPECIES, s = 30, m = 0.3, name=f'head_species/{config.head}', dtype='float32')
使用多个模型,找出大多数模型一致同意的top-1样本,确保第1类和第2类的置信度之间有0.2的差距。生成了12157个伪标签。队友:
@lextoumbourou
@zekunn
@poorneshwaran
由于队友们的鼎力支持和辛勤工作,我们才取得了不错的排名。非常感谢。让我们在其他挑战中再接再厉,争取更好的成绩。
查看 @lextoumbourou 分享的笔记本:
https://www.kaggle.com/code/lextoumbourou/happywhale-tpu-baseline-to-0-804-elasticface/notebook