rn50_supcon_bs4096_e1000

ResNet-50 supervised contrastive run on ImageNet-1K with FastSSL (main_pretrain.py). Implemented as SimLAP with identical pairs and an open gate (SimLAP.type="identical", SimLAP.gate_fn=@OpenGate).

Method	SupCon via SimLAP identical + OpenGate
Backbone	ResNet-50, projector 2048-2048-256, τ=0.1, 1000 classes
Data	ImageNet-1K, 224², two global crops
Optim	LARS, lr=0.2, wd=1e-4, warmup 40, 1000 epochs planned
Batch	4×1024 = 4096
Latest	checkpoint.pth (epoch 840 / 1000)

Periodic snapshots every 50 epochs: checkpoint-0.pth … checkpoint-800.pth. Training configs are config.yml and config.gin.