Model packaging (.mar files) in PyTorch - Model Pipeline Trace

Model Pipeline - Model packaging (.mar files)

This pipeline shows how a PyTorch model is packaged into a .mar file for easy deployment. The model is trained, saved, then packaged with metadata and dependencies into a .mar archive. This archive can be loaded by serving tools to make predictions.

Data Flow - 4 Stages

1Train Model

1000 rows x 20 features→Train a PyTorch neural network on input data→Model weights saved as .pt file

Input: feature vectors of size 20; Output: trained model weights file model.pt

↓

2Save Model Script

Model weights file (.pt)→Save model architecture and weights in a scripted format→Scripted model file (.pt)

torch.jit.script(model) saved as scripted_model.pt

↓

3Create MAR Archive

Scripted model file + extra files (handler, config)→Package model and metadata into a .mar archive using torch-model-archiver→Model archive file (.mar)

torch-model-archiver --model-name my_model --version 1.0 --serialized-file scripted_model.pt --handler handler.py --export-path model_store

↓

4Deploy MAR File

Model archive file (.mar)→Load .mar file into TorchServe for serving predictions→Model ready to serve predictions

torchserve --start --model-store model_store --models my_model=my_model.mar

Training Trace - Epoch by Epoch

Epoch 1: 0.85 *****
Epoch 2: 0.60 ****
Epoch 3: 0.45 ***
Epoch 4: 0.35 **
Epoch 5: 0.30 *

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.85	0.60	Initial training with high loss and moderate accuracy
2	0.60	0.75	Loss decreased, accuracy improved
3	0.45	0.82	Model learning well, loss dropping
4	0.35	0.88	Good convergence, accuracy nearing 90%
5	0.30	0.90	Training converged with low loss and high accuracy

Prediction Trace - 4 Layers

Layer 1: Input preprocessing

Layer 2: Model forward pass

Layer 3: Softmax activation

Layer 4: Prediction output

Model Quiz - 3 Questions

What is the main purpose of the .mar file in this pipeline?

A To package the model and metadata for deployment

B To train the model faster

C To store raw training data

D To visualize model predictions

Key Insight

Packaging a PyTorch model into a .mar file bundles the model code, weights, and metadata into one deployable archive. This makes it easy to serve the model in production environments with tools like TorchServe, ensuring consistent and efficient predictions.