---

Eurecat datathon: From 0 to Data Scientist

---

Implementation of the solutions for the Eurecat challenge during Hack EPS 2022 edition. On Applied AI, mainly DL for CV.

Below you can find the same information as in the DevPost project, but before a list of the different README.md for each mission is presented:

• Yellow mission
• Red mission
• Orange mission
• Purple mission
• Green mission
• Black mission

Relevant links

Also I attach some links that may prove useful:

• Local notebooks in the repo (I pushed them in this repo even though I developed them in Kaggle):
  • Orange mission: local notebook
  • Purple mission: local notebook
  • Green mission: local notebook
• Kaggle notebooks where I developed the code (I needed GPU):
  • Orange mission: Kaggle notebook
  • Purple mission: Kaggle notebook
  • Green mission: Kaggle notebook

Finally, in this link you can find the serialization of (large pretrained) models both for the orange and purple missions: link

Inspiration

Most of the proposed challeges are DL tasks, for they require CV approaches. I was eager to understand how the implementations of several techniques (such as transformers or capsum) worked. That's why I took profit and I used this challenge to learn them

What it does

• Solves most of the proposed missions by Eurecat
• Provides a summary of the proposed implementations

How we built it

Using Python, specifically:

• Tensorflow, mainly through the Keras API
• Numpy and pandas For the supervised learning part I built data pipelines, which were also responsible for the data augmentation process. Then both pre-trained models (CNN such as
• EfficientNet & transformers such as ViT16) using transfer learning, as well as CNN from scratch, were built, trained and used.

Challenges we ran into

• Finite computational power & time-consuming trainings
• I worked in parallel implementing the different solutions, and so need more accounts for the Kaggle part
• Hard to serialize and to store pre-trained CNN and transformers

Accomplishments that we're proud of

• Solve most of the challenges, having into account I could not join the hackaton before 4pm and I was alone

What we learned

• How to perform transfer learning with Visual Transformers (ViT).
• How to explain DL models for CV task using Captum

What's next for From 0 to Data Scientist - Eurecat

There are many improvements to be made, mostly regarding the fine-tuning part. Currently, the training process is so basic that just prevents over-fitting. However, te following could be controled:

• Dropout tuning
• Different optimizers with different learning rates schedulers
• Mainly, changes to the backbone
• Finally, more explainability job (even though is not required) in the sense it helps during the model training. For instance, salency maps for the differents channels could have been constructed.
• Oversampling the images corresponding to under-represented classes
• Improve even more the data augmentation process