Components
-
IAM role: IAM (Identity Access Management) is a set of permissions that define what actions an entity (services, user, aplications) can do. A role uses temporary creds while a user uses more permanent creds.
-
Debugger (At least compatible with TensorFlow). Video Explanations
- can add rules, e.g., vanishing gradient rule (your model is not updating anymore), overfit (can see in trial details about warnings)
- Save the model on S3 automatically
- github amazon sagemaker examples
- Pre-req: load Amazon SageMaker Studio
-
Pre-Installed Frameworks
- XGBoost
- Autogluon:
- Deployment
- these three all have a backbone (encoder), and a decoder
- Models:
- FCN
- PSP
- Deep Lab V3
- Process: 1. Put Data on S3 2. Org data into: train (jpg), validation (jpg), train_annotation(png), validation_annotation (single channel uint8 png), label_map (json, mapping between uint8 -> label names)
- Deployment
- Pytorch
I like Julien Simmon’s videos on Amazon SageMaker.
EC2 GPU & CPU Instances
- Free tier 50 hours per month of m4.xlarge or m5.xlarge for training, 125 hours per month of m4.xlarge or m5.xlarge for hosting
- m5.xlarge, g5.xlarge pricing (Source: Reddit Post), there is a performance and price chart as of Amazon EC2 types:
| Instance Type | Cores | RAM | GPU | Prompt Eval Rate (tokens/s) | Eval Rate (tokens/s) | Price (per hr) | Price (per mo) |
|---|---|---|---|---|---|---|---|
| c7g.8xlarge (CPU) | 32 | 64 GB | N/A | 38.38 | 25.07 | $1.27 | $941.16 |
| r6g.4xlarge (CPU) | 16 | 128 GB | N/A | 10.15 | 8.29 | $0.88 | $657.10 |
| g4dn.xlarge (GPU) | 4 | 16 GB | 16 GB | 222.23 | 41.71 | $0.58 | $434.50 |
| g4dn.2xlarge (GPU) | 8 | 32 GB | 32 GB | 214.25 | 41.74 | $0.84 | $621.24 |
| g5.xlarge (GPU) | 4 | 16 GB | 24 GB | 624.29 | 68.08 | $1.12 | $831.05 |
| g5.2xlarge (GPU) | 8 | 32 GB | 24 GB | 624.48 | 66.67 | $1.35 | $1,000.96 |
