ResNet34 was used as a encoder for the U-net network architecture. The network is trained on images of size 384. In the future, it can be trained to a larger size.
This approach was chosen because the ratio of pixels of ships to non-ships is very unbalanced, and it makes no sense to train the network on large images at once, as this will only increase the training time

task

• python version: Python 3.6.4 :: Anaconda custom (64-bit)
• install all the libraries from requirements.txt
• Download weights and put in directory notebooks/Resnet34_lable_384_1.h5
• Download dataset from task and put it in directory ./input/airbus-ship-detection/ <files>

It is also important to note these notebooks that helped in completing this task and understanding it.

• https://www.kaggle.com/code/iafoss/unet34-dice-0-87
• https://www.kaggle.com/code/iafoss/unet34-submission-tta-0-699-new-public-lb
• https://www.kaggle.com/code/julian3833/4-exploring-public-models
• https://www.kaggle.com/code/inversion/run-length-decoding-quick-start
• https://www.kaggle.com/code/iafoss/fine-tuning-resnet34-on-ship-detection
• https://www.kaggle.com/code/julian3833/3-basic-exploratory-analysis