Simple baseline with simple transformer

For task 3, it is 0.538.

PS: I am happy with this score given the fact that I only trained for 2 epochs with only limited number of rows.

View post 1

Simple baseline with simple transformer

Looking at the provided baseline, I found it hard to follow. Therefore I decided to publish my simple baseline.

Not all data were used nor any fine-tune has been done.

It just took more than 1 hour for training and 15 minutes for inference.

Hope it helps newcomers to understand the data and the problem.

View post 3

Purchase with anomaly detection

Not so sure why the discussion is somehow quiet in round 2. I am sharing my first success.

I love anomaly detection so I tried to apply it here. It has been quite surprise that few tries have even worse result than the random purchase.

However, I finally manage to beat the random purchase. Insight could be found here
AIcrowd | Purchase with anomaly detection | Posts.

Please give me a vote for my hard work

View post 3

:aicrowd: [Update] Round 2 of Data Purchasing Challenge is now live!

Many thanks for your quick reply.

“You will still have to train your models from scratch” => Could we use pre-trained weights as in round 1 ?

View post 0

:aicrowd: [Update] Round 2 of Data Purchasing Challenge is now live!

@snehananavati : Many thanks for your update. As far I understood, we actually don’t need to provide code for the training phase and the prediction phase. The system will use the same training pipeline - B4 with 10 epochs and then generate predictions and the scores by itself.

In other words, we only need to focus on the purchase phase. Do I understand correctly ?

View post 0

0.9+ Baseline Solution for Part 1 of Challenge

Oh, I missed this thread. Many thanks for your sharing @mark_haoxiang . It is quite interesting that the simple approach works well.

View post 1

[Notebooks & Resource] Community Contributions and Baselines

I would like to mention 2 baselines.

My own baseline - score 0.84 AIcrowd Submission Received #174462 - v0p5 (#4) · Issues · moto / data-purchasing-optimization · GitLab.

My new baseline based on Leocd’s code - score 0.885

View post 2

Experiments with “unlabelled” data

@sergey_zlobin : Thanks for your information.

I am wondering if you tried to purchase all 10K then what the score could be.

View post 0

Tensorflow/Keras folks, you are not being left behind in this competition

@huynhngoc : Many thanks. I did not know that we could access labels_df.

View post 0

Pseudo-labeling

Indeed. I don’t see why we can’t use pseudo-labels. One of the naive approach for the purchase policy is to predict all images. 1) If the confidence is high => use the pseudo-label 2) if the confidence is low => purchase the label.

View post 4

Baseline 0.84

In order to support my request (Request to have the same baseline for everyone!), I shared my baseline with densenet. You could access the submission at AIcrowd Submission Received #174462 - v0p5 (#4) · Issues · moto / data-purchasing-optimization · GitLab.

If my submission can access the pre-trained weights for efficient net B6 from https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/, it could score 0.87+.

View post 0

Request to have the same baseline for everyone!

Dear organizer, all,

This challenge is a data-centric competition and the main purpose is to optimize the purchase. However, given our experiments the upper bound for the purchase policy is less than 3%. It means that we need to have a good baseline before we could work on the purchase policy.

Therefore, in my opinion, we should share the same training procedure. Our customer DL pipeline is only to optimize the purchase. I know it would be unfair for the current top teams but it would drive the competition in the way it should be.

Well, I am sharing my baseline with densetnet. It scores 0.84. The same pipeline with B6 could get 0.87+ (however, the pre-trained weights are not from the official site). The code is here AIcrowd

If it is impossible to force all participants, please allow us to use weights from other popular sources! I will try to publish a baseline 0.88+ so that people could focus more on the purchase optimization.

Many thanks.
M

View post 5

The mystery of 0.489 and how to beat 2 deep-learning baselines with a single line of code

If you look at my notebook AIcrowd | Baseline + Exploration: random purchase vs full purchase | Posts you could see that the zero-prediction solution got a score of 0.478 locally.

And that solution will score 0.489 in the LB, beating 2 public baselines.

• Video tutorial of challenge and how to submit first baseline with 0.44 on LB
• Baseline submission - #3 by moto

How to do that, just replace

by

np.zeros(4).astype(int)

That’s it.

View post 0

Baseline submission

Update: The code is now using pre-trained weights. It has a better score now.

View post 1

Baseline submission

I am very happy to share my baseline Files · submission-v0p1p5 · moto / data-purchasing-hello · GitLab

The most challenging job is to use both 2 datasets (5K+3K) to train a model.

Note that the low score might be due to the fact that I haven’t included the pre-trained weights. Just submitted that one.

Right now, you guys could focus on

• DL techniques such as different augmentations, network archs, schedulers …
• Optimize purchase

Enjoy.

View post 7

What did you get so far?

Guys in the LB are fast but not me. I still need to understand how to submit.

View post 0

What did you get so far?

I guess you guys haven’t been able to defeat the random purchase

My experiment showed that 10K purchase is only a little bit better than 3K random purchase (https://www.aicrowd.com/showcase/exploration-random-purchase-vs-full-purchase). So, it is super hard to beat the random purchase.

Updated:

1. The new link is https://www.aicrowd.com/showcase/baseline-exploration-random-purchase-vs-full-purchase
2. The same code is included in my baseline

Files · submission-v0p1p5 · moto / data-purchasing-hello · GitLab

View post 5

🚨 Baseline released, SiamMOT: Siamese Multi-Object Tracking

@shivam: any way to run the baseline in colab for local validation ?

View post 0

Max Number of Submissions?

@shivam: I am unable to submit today after 1 failed submission.

View post 0

🚨 Clarification: Which are valid submissions?

@shivam: Many thanks for your quick reply.

Does it also mean using historical frames is accepted ?

View post 0