Face recognition l: ONNX iResNet100
Fit diagram
Read at 100%
Clear selection
Download SVG
Download PNG
Click a block to read its description, or select it with Tab and Enter.
Face recognition l: ONNX iResNet100
Aligned RGB input 112 × 112, batch 1, 512-dimensional embedding. Actual exported ONNX operators, with fused convolution parameters.
LibreYOLO
Face recognition l: ONNX iResNet100
Aligned RGB input 112 × 112, batch 1, 512-dimensional embedding. Actual exported ONNX operators, with fused convolution parameters.
Recognition network
Aligned RGB face
3 × 112 × 112
Normalize (x - 127.5) / 127.5
3 × 112 × 112
Conv 3×3 / 1
3 to 64, p=1; fused bias
PReLU
64 × 112 × 112
Layer 1
64 × 56 × 56, n=3
Layer 2
128 × 28 × 28, n=13
Layer 3
256 × 14 × 14, n=30
Layer 4
512 × 7 × 7, n=3
BatchNormalization
512 × 7 × 7
Flatten
25,088
Gemm
25,088 to 512, bias=True
BatchNormalization
512
L2 normalize
512-dimensional identity vector
Layer 1
Downsample block, n=1
Input 64 × 112 × 112
BatchNormalization
64 × 112 × 112
Conv 3×3 / 1
64 × 112 × 112, p=1
PRelu
64 × 112 × 112
Conv 3×3 / 2
64 × 56 × 56, p=1
+
Conv 1×1 / 2
64 × 56 × 56
Output 64 × 56 × 56
Identity block, n=2
Input 64 × 56 × 56
BatchNormalization
64 × 56 × 56
Conv 3×3 / 1
64 × 56 × 56, p=1
PRelu
64 × 56 × 56
Conv 3×3 / 1
64 × 56 × 56, p=1
+
Output 64 × 56 × 56
Layer 2
Downsample block, n=1
Input 64 × 56 × 56
BatchNormalization
64 × 56 × 56
Conv 3×3 / 1
128 × 56 × 56, p=1
PRelu
128 × 56 × 56
Conv 3×3 / 2
128 × 28 × 28, p=1
+
Conv 1×1 / 2
128 × 28 × 28
Output 128 × 28 × 28
Identity block, n=12
Input 128 × 28 × 28
BatchNormalization
128 × 28 × 28
Conv 3×3 / 1
128 × 28 × 28, p=1
PRelu
128 × 28 × 28
Conv 3×3 / 1
128 × 28 × 28, p=1
+
Output 128 × 28 × 28
Layer 3
Downsample block, n=1
Input 128 × 28 × 28
BatchNormalization
128 × 28 × 28
Conv 3×3 / 1
256 × 28 × 28, p=1
PRelu
256 × 28 × 28
Conv 3×3 / 2
256 × 14 × 14, p=1
+
Conv 1×1 / 2
256 × 14 × 14
Output 256 × 14 × 14
Identity block, n=29
Input 256 × 14 × 14
BatchNormalization
256 × 14 × 14
Conv 3×3 / 1
256 × 14 × 14, p=1
PRelu
256 × 14 × 14
Conv 3×3 / 1
256 × 14 × 14, p=1
+
Output 256 × 14 × 14
Layer 4
Downsample block, n=1
Input 256 × 14 × 14
BatchNormalization
256 × 14 × 14
Conv 3×3 / 1
512 × 14 × 14, p=1
PRelu
512 × 14 × 14
Conv 3×3 / 2
512 × 7 × 7, p=1
+
Conv 1×1 / 2
512 × 7 × 7
Output 512 × 7 × 7
Identity block, n=2
Input 512 × 7 × 7
BatchNormalization
512 × 7 × 7
Conv 3×3 / 1
512 × 7 × 7, p=1
PRelu
512 × 7 × 7
Conv 3×3 / 1
512 × 7 × 7, p=1
+
Output 512 × 7 × 7
Artifact verification
librefacerec-l.onnx: 103 Conv, 51 BatchNormalization, 50 PRelu, 49 Add, one Flatten and one Gemm.
SHA-256 a7933ea5330113b01c9b60351d8f4c33003f145d8470ac5f0e52ee2effe25c60. Shape inference pinned batch 1.
Source: libreyolo/models/facerec/model.py. Revision a4d0ecc9e17f.
libreyolo.com