What I learn from Learning Machine Learning Development - ML Course at Dicoding

Learning Machine Learning Development

Learning Machine Learning Development - Belajar Pengembangan Machine Learning, I very like this course. What I get from this course mostly what Tensorflow Certification test. It's about NLP (Natural Language Processing), Time Series, and Image Classification. How great the course! In case you want to know the course, please visit on here, this course taught at Bahasa Indonesia, so if you use English, better to try another course that related, but this course very worth if you use Bahasa Indonesia and this is the best.

Belajar Pengembangan Machine Learning

I want to tell, what the criteria that you need to achieve to finish this course. This course have name "Belajar Pengembangan Machine Learning", in English we can say Learning Machine Learning Development. I will publish in here what I submit on that course. In this course, need 3 types of submission. The first submission about NLP, the second submission about Time Series, and the third submission about Image Classification. Each submission have rate criteria, the highest rate will have high criteria to achieve and what you should know, the highest rate will bring you to achieve what Tensorflow Certification test. I think this course really worth to get and learn. Learning Machine Learning Development really make me know more how to use Machine Learning and give me some good technical "hack" that you can use to give you better result. I have learned some Machine Learning at my Undergraduate Courses but this is really different, give you more technical into the Tensorflow itself, but I'm sure this also not only about Tensorflow.

Mostly my submission I host at my Github. By default I stored at Google Collaboration, but I also try to mirror or host it on Github.

berviantoleo / machine-learning-experiment

Machine Learning Experiment

Machine Learning Experiment

Machine Learning Experiment

Please Read This First

Please not do any copy paste as it is. You can learn from my code and improve it, but you can't just copy paste as for your submition as example.

This code mostly from my collaboration, I will not accept any pull request. I can receive your issues Question/Improvement.

Happy Open Source! I just want to sharing knowledge about machine learning on this repo.

License

Code license is MIT.

View on GitHub

First Submission - NLP

You can see my first submission here. The criteria to get highest rate in this submission are,

• The dataset to be used is free, but has at least 2000 samples, 3 classes.
• Must use LSTM in model architecture.
• Must use a sequential model.
• The validation set is 20% of the total dataset.
• Must use Embedding.
• Must use tokenizer function.
• The accuracy at training set and validation set of the model is more than 90%.
• Using callback
• Draw plot loss and accuracy at training and validation.

You can try to compare with Tensorflow Certification criteria on this NLP, mostly of the criteria at Learning Machine Learning Development also at Tensorflow Certification. My submission is about News Categorization, I know that very common case, but I sure the dataset have good quantity of it. As you should know, I have good review too, they bring me some advice to give better result.

Second Submission - Time Series

Here the second submission. I'm sure, this is also bring me safe case. In this submission, I use Multivariate Time Series as my dataset, but actually the submission only need the Univariate. Seems I must learn more how to process the Multivariate, I know my solution not really good.

The criteria like these,

• The dataset to be used is free, but has at least 10000 samples and MAE < 10% of data scale.
• Must use LSTM in model architecture.
• The validation set is 20% of the total dataset.
• The model must use a sequential model.
• Must use Learning Rate in Optimizer.
• Have to implement callback
• Give plot of loss and accuracy at training and validation.

Third Submission - Image Classification & Deployment

This third submission is awesome. How you can train 10000 more image fastly? I have a challenge to train them. Hemm... I bit thinking, why so slow when I'm not set anything, I just realize, I need to change to GPU processing, I'm new at Tensorflow. It makes the training faster than before. Just in some minutes, not like before, about 30 minutes, so far differences.

Here the criteria for highest grade,

• The dataset to be used is free, but has at least 20000 image samples, minimal 3 classes, image resolution have differences.
• The validation set is 20% of the total dataset.
• Must use a sequential model.
• Models must use the Conv2D Maxpooling Layer.
• The accuracy at training set and validation set of the model is more than 92%.
• Have to implement callback
• Give plot of loss and accuracy at training and validation.
• Writes code to save the model in TF-Lite format.

For me, the 92% accuracy in validation bit hard to achieve, this is looks like you process the real world image. How hard it is. I almost forget to bring the link. Here my submission.

TLDR

Maybe you don't want to read the long thread. I want to bring you a summarization. Mostly this course have a good start to know what Tensorflow Certification test. I see this bring you about 80% criteria on the Certification. But, is it enough? No, You should read more, at least at Tensorflow documentation, you can read through it. For me, this course is good start for you to achieve Tensorflow Certification and have little bit good knowledge to use it. I'm new at this, I'm new to use Tensorflow. In this course, I can know how to use it and some tricks. It's worth for you when you need Bahasa Indonesia references about Tensorflow. Great job, Dicoding!