Applied Deep Learning 2025 - Lecture 12 - Serving, Optimizing, and Practical Aspects

Alexander Pacha · Beginner ·📐 ML Fundamentals ·7mo ago
Skills: Optimisation53%

About this lesson

What happens after you've trained a machine learning model? It's pretty worthless, unless you can run in and serve it properly to some users. This is what we're looking at today. How we can take our amazing models and bring it all the way to our customers through releases, docker images, or platforms that do almost all the work for us. Apart from this, we're also discussing a few more practical aspects and how we can optimize our models even further. Complete Playlist: https://www.youtube.com/watch?v=vlTnIjhhmzA&list=PLNsFwZQ_pkE8H1o874cZbiwnNRJ6hCDJI 00:00:00 - Start 00:00:44 - Machine Learning in Practice 00:03:50 - Simple way to serve a model 00:07:30 - A better way to serve a model 00:11:04 - Serving Models with Tensorflow 00:17:03 - Serving in the Cloud 00:18:14 - Paperspace 00:18:48 - Streamlit 00:19:23 - Gradio 00:20:04 - Hugging Face Spaces 00:20:25 - Serving your model in a Container 00:21:49 - Tensorflow Serving Docker Container 00:23:22 - ONNX 00:27:47 - Publishing your dataset - FAIR 00:32:21 - Deep Learning in the Browser 00:43:13 - Overview and Survey 00:43:52 - Deep Learning on mobile devices 00:45:18 - Benefits of running a model on device 00:46:47 - Boosting Performance 00:49:14 - Optimizing Data Loading 00:53:52 - Floating point quantization 00:59:11 - The effect of minibatch sizes 01:01:22 - Warmup 01:03:08 - How to reduce the costs of running models? 01:04:09 - Pruning neural networks 01:07:35 - Lottery Ticket Hypothesis 01:10:59 - Summary == Literature == 1. Seide et al. 1-Bit Stochastic Gradient Descent and Application to Data-Parallel Distributed Training of Speech DNNs, 2014. 2. Micikevicius, Mixed-Precision Training of Deep Neural Networks, 2017. 3. Addair, What is the difference between FP16 and FP32, 2018. 4. Tensor Processing Unit on Wikipedia. 5. Goyal et al. Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour, 2018. 6. Kim et al. PyTorch Gradual Warmup LR on Github. 7. TensorFlow Data Performance Guide. 8. Prakash, Serving a

Original Description

What happens after you've trained a machine learning model? It's pretty worthless, unless you can run in and serve it properly to some users. This is what we're looking at today. How we can take our amazing models and bring it all the way to our customers through releases, docker images, or platforms that do almost all the work for us. Apart from this, we're also discussing a few more practical aspects and how we can optimize our models even further. Complete Playlist: https://www.youtube.com/watch?v=vlTnIjhhmzA&list=PLNsFwZQ_pkE8H1o874cZbiwnNRJ6hCDJI 00:00:00 - Start 00:00:44 - Machine Learning in Practice 00:03:50 - Simple way to serve a model 00:07:30 - A better way to serve a model 00:11:04 - Serving Models with Tensorflow 00:17:03 - Serving in the Cloud 00:18:14 - Paperspace 00:18:48 - Streamlit 00:19:23 - Gradio 00:20:04 - Hugging Face Spaces 00:20:25 - Serving your model in a Container 00:21:49 - Tensorflow Serving Docker Container 00:23:22 - ONNX 00:27:47 - Publishing your dataset - FAIR 00:32:21 - Deep Learning in the Browser 00:43:13 - Overview and Survey 00:43:52 - Deep Learning on mobile devices 00:45:18 - Benefits of running a model on device 00:46:47 - Boosting Performance 00:49:14 - Optimizing Data Loading 00:53:52 - Floating point quantization 00:59:11 - The effect of minibatch sizes 01:01:22 - Warmup 01:03:08 - How to reduce the costs of running models? 01:04:09 - Pruning neural networks 01:07:35 - Lottery Ticket Hypothesis 01:10:59 - Summary == Literature == 1. Seide et al. 1-Bit Stochastic Gradient Descent and Application to Data-Parallel Distributed Training of Speech DNNs, 2014. 2. Micikevicius, Mixed-Precision Training of Deep Neural Networks, 2017. 3. Addair, What is the difference between FP16 and FP32, 2018. 4. Tensor Processing Unit on Wikipedia. 5. Goyal et al. Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour, 2018. 6. Kim et al. PyTorch Gradual Warmup LR on Github. 7. TensorFlow Data Performance Guide. 8. Prakash, Serving a
Watch on YouTube ↗ (saves to browser)
Sign in to unlock AI tutor explanation · ⚡30

Related Reads

📰
A diagram is data, not a drawing
Learn to think of diagrams as data, not just drawings, and explore the implications for architecture and modeling
Dev.to · Siva Teja
 📰
Monitor Amazon SageMaker Pipelines cross-account with custom Amazon CloudWatch dashboards
Monitor SageMaker Pipelines across AWS accounts with custom CloudWatch dashboards and learn how to centralize monitoring for improved visibility and control
AWS Machine Learning
 📰
🛠️ Day 3 – Python Functions & Functional Power
Learn to create reusable and modular Python functions with flexible parameters and variable scope
Dev.to AI
 📰
Your SAE Passed the Cosine Similarity Bar. That Doesn't Mean It
Learn why passing the cosine similarity bar doesn't guarantee a good sparse autoencoder (SAE) and how to audit its features
Dev.to · Mohamed Bal

Chapters (27)

Start
0:44 Machine Learning in Practice
3:50 Simple way to serve a model
7:30 A better way to serve a model
11:04 Serving Models with Tensorflow
17:03 Serving in the Cloud
18:14 Paperspace
18:48 Streamlit
19:23 Gradio
20:04 Hugging Face Spaces
20:25 Serving your model in a Container
21:49 Tensorflow Serving Docker Container
23:22 ONNX
27:47 Publishing your dataset - FAIR
32:21 Deep Learning in the Browser
43:13 Overview and Survey
43:52 Deep Learning on mobile devices
45:18 Benefits of running a model on device
46:47 Boosting Performance
49:14 Optimizing Data Loading
53:52 Floating point quantization
59:11 The effect of minibatch sizes
1:01:22 Warmup
1:03:08 How to reduce the costs of running models?
1:04:09 Pruning neural networks
1:07:35 Lottery Ticket Hypothesis
1:10:59 Summary
Up next
Arrays vs Lists: What AI Actually Prefers | Common Tech Interview Questions
SCALER
Watch →