ConvNeXt t

Click a block to read its description, or select it with Tab and Enter.

ConvNeXt tClassification, 224 × 224 input, 1,000 classes. Tensor sizes exclude batch.LibreYOLOConvNeXt tClassification, 224 × 224 input, 1,000 classes. Tensor sizes exclude batch.NetworkInput3 × 224 × 224Conv2d 4×4 / 496 × 56 × 56LayerNorm over channels96 × 56 × 56Stage 196 × 56 × 56, n=3Stage 2192 × 28 × 28, n=3Stage 3384 × 14 × 14, n=9Stage 4768 × 7 × 7, n=3AdaptiveAvgPool2d768 × 1 × 1LayerNorm over channels768 × 1 × 1Flatten768Linear classifier768 to 1,000 logitsStage 1Identity downsample96 × 56 × 56ConvNeXtBlock, repeat 3 timesDepthwise Conv2d 7×796 channels, g=96, s=1, p=3Permute NCHW to NHWC56 × 56 × 96LayerNorm96 channels, eps=1e-6Linear96 to 384GELU56 × 56 × 384Linear384 to 96Permute NHWC to NCHW96 × 56 × 56Multiply layer scaleLearned gamma: 96 channels+Output 96 × 56 × 56Stage 2LayerNorm over channels96 × 56 × 56Conv2d 2×2 / 296 to 192, p=0ConvNeXtBlock, repeat 3 timesDepthwise Conv2d 7×7192 channels, g=192, s=1, p=3Permute NCHW to NHWC28 × 28 × 192LayerNorm192 channels, eps=1e-6Linear192 to 768GELU28 × 28 × 768Linear768 to 192Permute NHWC to NCHW192 × 28 × 28Multiply layer scaleLearned gamma: 192 channels+Output 192 × 28 × 28Stage 3LayerNorm over channels192 × 28 × 28Conv2d 2×2 / 2192 to 384, p=0ConvNeXtBlock, repeat 9 timesDepthwise Conv2d 7×7384 channels, g=384, s=1, p=3Permute NCHW to NHWC14 × 14 × 384LayerNorm384 channels, eps=1e-6Linear384 to 1536GELU14 × 14 × 1536Linear1536 to 384Permute NHWC to NCHW384 × 14 × 14Multiply layer scaleLearned gamma: 384 channels+Output 384 × 14 × 14Stage 4LayerNorm over channels384 × 14 × 14Conv2d 2×2 / 2384 to 768, p=0ConvNeXtBlock, repeat 3 timesDepthwise Conv2d 7×7768 channels, g=768, s=1, p=3Permute NCHW to NHWC7 × 7 × 768LayerNorm768 channels, eps=1e-6Linear768 to 3072GELU7 × 7 × 3072Linear3072 to 768Permute NHWC to NCHW768 × 7 × 7Multiply layer scaleLearned gamma: 768 channels+Output 768 × 7 × 7Variant valuesSizeC1, C2, C3, C4: stage channelsn1, n2, n3, n4: block countst96, 192, 384, 7683, 3, 9, 3s96, 192, 384, 7683, 3, 27, 3b128, 256, 512, 10243, 3, 27, 3Source: libreyolo/models/convnext/nn.py. Revision a4d0ecc9e17f.libreyolo.com