Yandex ML CVHN Project by Team Baza

Project Document (RUS)

All Models

---

Models

Model Name	Score* without Chapter 3	Link	Notes
Simple CNN-2	0.8428	link	A simple model using Conv2D, Dropout, Batch Normalization, MaxPooling, Flatten, and Dense layers.
Simple CNN-3	0.8488	link	Differs from Simple CNN-2 only in the order of Batch Normalization and Dropout. However, the Private Score differs by 0.01, as we found out after the competition ended. Therefore, we conclude that it is better to apply Batch Normalization first and then Dropout. There are extensive discussions on this topic: Stack Overflow discussion.
ResNet 1	0.91	link	ResNet implementation.
ResNet 2	0.892	link	Added Dropout(0.5) only inside the residual blocks.
ResNet 3	0.93426 0.9424 with data preprocessing	link	Added Dropout(0.4) everywhere after BatchNormalization().
UNet 1	0.8748	link	U-Net implementation.
UNet 2	0.88693 0.89333 with data preprocessing	link	Modified the U-Net configuration by reducing the number of downsampling and upsampling stages.
Densenet 1	No score, because the result was worse than with simple models.	link	This is actually a ResNeXt implementation. The incorrect name remained because the architecture was divided into layers similarly to the classic DenseNet configuration: 6, 12, 24, 16.
Densenet 2	0.9272	link	Reduced the number of parameters from 7 million to 1.5 million while improving quality.
Densenet 2.2	0.93733	link	Differs from DenseNet 2 in that after the final dense connection block, an additional Batch Normalization, Dropout, and ReLU layer is applied before GlobalAveragePooling2D.