Principle:LaurentMazare Tch rs Pretrained Model Instantiation
| Knowledge Sources | |
|---|---|
| Domains | Computer_Vision, Model_Architecture |
| Last Updated | 2026-02-08 14:00 GMT |
Overview
Pattern for instantiating a predefined vision model architecture by registering its parameters into a VarStore with a specified number of output classes.
Description
Pretrained model instantiation creates the full computational graph of a known architecture (ResNet, VGG, EfficientNet, etc.) and registers all learnable parameters in a VarStore. The model is constructed with randomly initialized weights, which are then overwritten by loading pretrained weights via VarStore::load. The constructor takes a VarStore path for parameter namespacing and a class count (typically 1000 for ImageNet). The returned model implements ModuleT for training-aware forward passes.
Usage
Use this when performing image classification with a known architecture. After instantiation, load pretrained weights to enable inference or use as a starting point for fine-tuning.
Theoretical Basis
ResNet-18 architecture:
Input [3, 224, 224]
-> conv1 (7x7, stride 2) -> bn1 -> relu -> maxpool (3x3, stride 2)
-> layer1 (2 BasicBlocks, 64 channels)
-> layer2 (2 BasicBlocks, 128 channels, stride 2)
-> layer3 (2 BasicBlocks, 256 channels, stride 2)
-> layer4 (2 BasicBlocks, 512 channels, stride 2)
-> adaptive_avg_pool2d -> flatten -> fc (512 -> num_classes)
Output [num_classes]