resnet18.a1_in1k

resnet18.a1_in1k is a lightweight convolutional neural network from the timm library, implementing a ResNet-B variant trained on ImageNet-1K using the improved "ResNet Strikes Back" A1 training recipe. It features ReLU activations, a single 7x7 convolution with pooling, and 1x1 convolutional shortcuts for downsampling. With only 11.7 million parameters, it's designed to be efficient while maintaining strong baseline performance for image classification tasks. The model was optimized using the LAMB optimizer, a cosine learning rate schedule with warmup, and binary cross-entropy (BCE) loss. It achieves 73.16% Top-1 and 91.03% Top-5 accuracy at 288×288 image resolution. It's highly suited for feature extraction and embeddings as it supports classification-free forward passes and intermediate feature map access. This makes it a flexible backbone for real-time computer vision pipelines on resource-constrained devices.