File size: 7,695 Bytes
1fb3b1b 2c17b0b 263005f 2c17b0b 1fb3b1b 2c17b0b 1fb3b1b 2c17b0b 263005f 2c17b0b 1fb3b1b 2c17b0b 1fb3b1b 2c17b0b 1fb3b1b 263005f 1fb3b1b 2c17b0b 1fb3b1b 2c17b0b 1fb3b1b 263005f 1fb3b1b 2c17b0b 1fb3b1b 2c17b0b 1fb3b1b 263005f 1fb3b1b 2c17b0b 1fb3b1b 2c17b0b 1fb3b1b 263005f 1fb3b1b 2c17b0b 1fb3b1b 2c17b0b 1fb3b1b 263005f 1fb3b1b 2c17b0b 1fb3b1b 2c17b0b 1fb3b1b 263005f 1fb3b1b 263005f 1a480eb 3331842 1a480eb 3331842 1a480eb 3331842 1a480eb 3331842 1a480eb 3331842 1a480eb 1fb3b1b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 | {
"model": "setup_53_2d",
"task": "mito_affinities",
"architecture": "Unet3dIso (2D-collapsed)",
"checkpoint": "student_60000",
"params": 1927907,
"channels": [
"mito_aff_z",
"mito_aff_y",
"mito_aff_x"
],
"input_voxel_size_nm": [
16,
16,
16
],
"output_voxel_size_nm": [
16,
16,
16
],
"output": "raw logits (apply sigmoid / threshold at 0)",
"input_norm": [
{
"name": "MinMaxNormalizer",
"min_value": 0,
"max_value": 255
},
{
"name": "LambdaNormalizer",
"expression": "2*x-1"
}
],
"io": "input 'input' float32 [1,1,1,y,x] (one z-slice); output 'output' float32 [1,C,1,y',x']",
"kind": "2D slice student distilled from the 3D model (train_2d_distill.py)",
"step": 60000,
"distilled_from": "setup_53",
"distilled_from_repo": "cellmap/mito-aff-unetiso-setup-53",
"variants": [
{
"model": "setup_53_2d",
"variant": "fp32",
"input_size": [
1,
300,
300
],
"output_size": [
1,
212,
212
],
"file": "setup_53_mito_affinities_16nm_2d_in300_fp32.onnx",
"file_mb": 7.75,
"max_abs": 2.288818359375e-05,
"flip_frac": 0.0
},
{
"model": "setup_53_2d",
"variant": "fp16",
"input_size": [
1,
300,
300
],
"output_size": [
1,
212,
212
],
"file": "setup_53_mito_affinities_16nm_2d_in300_fp16.onnx",
"file_mb": 3.9,
"max_abs": 0.10926437377929688,
"flip_frac": 4.4499822000712e-05
},
{
"model": "setup_53_2d",
"variant": "fp32",
"input_size": [
1,
556,
556
],
"output_size": [
1,
468,
468
],
"file": "setup_53_mito_affinities_16nm_2d_in556_fp32.onnx",
"file_mb": 7.75,
"max_abs": 2.09808349609375e-05,
"flip_frac": 0.0
},
{
"model": "setup_53_2d",
"variant": "fp16",
"input_size": [
1,
556,
556
],
"output_size": [
1,
468,
468
],
"file": "setup_53_mito_affinities_16nm_2d_in556_fp16.onnx",
"file_mb": 3.9,
"max_abs": 0.10926437377929688,
"flip_frac": 3.652567755131858e-05
},
{
"model": "setup_53_2d",
"variant": "fp32",
"input_size": [
1,
1068,
1068
],
"output_size": [
1,
980,
980
],
"file": "setup_53_mito_affinities_16nm_2d_in1068_fp32.onnx",
"file_mb": 7.75,
"max_abs": 2.574920654296875e-05,
"flip_frac": 0.0
},
{
"model": "setup_53_2d",
"variant": "fp16",
"input_size": [
1,
1068,
1068
],
"output_size": [
1,
980,
980
],
"file": "setup_53_mito_affinities_16nm_2d_in1068_fp16.onnx",
"file_mb": 3.9,
"max_abs": 0.10926437377929688,
"flip_frac": 4.859086491739553e-05
}
],
"val_f1": {
"metric": "affinity F1 (mean of 3 channels), 6 GT val crops, centered 200^3, threshold logit>0",
"student": {
"crop495": 0.8164,
"crop496": 0.796,
"crop1017": 0.9015,
"crop528": 0.9217,
"crop1091": 0.9166,
"crop1081": 0.9481
},
"student_mean": 0.8834,
"teacher_3d": {
"crop495": 0.955,
"crop496": 0.928,
"crop1017": 0.966,
"crop528": 0.9746,
"crop1091": 0.9517,
"crop1081": 0.9456
},
"teacher_3d_mean": 0.9535
},
"torch_version": "2.9.1",
"exporter": "torch.onnx.export (TorchScript, dynamo=False)",
"opset": 18,
"finetune": {
"head": {
"final_layer": "final_conv.c",
"onnx_node": "/final_conv/c/Conv",
"features": "/final_conv/Gather_output_0",
"weight": "final_conv.c.weight",
"bias": "final_conv.c.bias",
"in_channels": 32,
"out_channels": 3,
"activation": "none (logits)",
"files": [
"setup_53_mito_affinities_16nm_2d_in1068_fp16.onnx",
"setup_53_mito_affinities_16nm_2d_in1068_fp32.onnx",
"setup_53_mito_affinities_16nm_2d_in300_fp16.onnx",
"setup_53_mito_affinities_16nm_2d_in300_fp32.onnx",
"setup_53_mito_affinities_16nm_2d_in556_fp16.onnx",
"setup_53_mito_affinities_16nm_2d_in556_fp32.onnx"
],
"initializers": "finetune/initializers.json",
"loss": "masked BCE on logits",
"note": "train the final 1x1 conv on frozen features"
},
"full": {
"library": "tfjs",
"files": [
"finetune/train_graph.json",
"finetune/cellmap_tfjs_train.js"
],
"weight_map": "identity",
"initializers": "finetune/initializers.json",
"write_back": [
"setup_53_mito_affinities_16nm_2d_in1068_fp16.onnx",
"setup_53_mito_affinities_16nm_2d_in1068_fp32.onnx",
"setup_53_mito_affinities_16nm_2d_in300_fp16.onnx",
"setup_53_mito_affinities_16nm_2d_in300_fp32.onnx",
"setup_53_mito_affinities_16nm_2d_in556_fp16.onnx",
"setup_53_mito_affinities_16nm_2d_in556_fp32.onnx"
],
"frozen": [
"encoder0.bn1.running_mean",
"encoder0.bn1.running_var",
"encoder0.bn2.running_mean",
"encoder0.bn2.running_var",
"encoder1.bn1.running_mean",
"encoder1.bn1.running_var",
"encoder1.bn2.running_mean",
"encoder1.bn2.running_var",
"encoder2.bn1.running_mean",
"encoder2.bn1.running_var",
"encoder2.bn2.running_mean",
"encoder2.bn2.running_var",
"bottleneck.bn1.running_mean",
"bottleneck.bn1.running_var",
"bottleneck.bn2.running_mean",
"bottleneck.bn2.running_var",
"decoder2.bn1.running_mean",
"decoder2.bn1.running_var",
"decoder2.bn2.running_mean",
"decoder2.bn2.running_var",
"decoder1.bn1.running_mean",
"decoder1.bn1.running_var",
"decoder1.bn2.running_mean",
"decoder1.bn2.running_var",
"decoder0.bn1.running_mean",
"decoder0.bn1.running_var",
"decoder0.bn2.running_mean",
"decoder0.bn2.running_var"
],
"note": "params named as the ONNX initializers; BatchNorm running stats frozen",
"spike": {
"setup": "TF.js 4.22.0 WebGPU backend, Chrome 154 on macOS, Apple GPU (metal-3, unified memory); batch 1, fp32, masked BCE on logits, Adam lr 1e-4; graph = finetune/train_graph.json, weights read from the served fp32 .onnx; median of 8 steps after 2 warm-up; peak = tf.profile peakBytes above baseline (tracked tensors; the WebGPU buffer pool holds more)",
"date": "2026-10-17",
"params": 1927907,
"full": {
"300": {
"step_ms": 308,
"peak_mb": 980
},
"556": {
"step_ms": 1189,
"peak_mb": 3897
}
},
"head_through_full_graph": {
"300": {
"step_ms": 31
},
"556": {
"step_ms": 133
}
},
"forward_ms": {
"300": 11,
"556": 41
},
"tfjs_webgpu_vs_ort_webgpu_556": {
"max_abs": 1.81e-05,
"flips": 0
},
"note": "2026-10-17 rerun of finetune_spike.html unchanged: peak_mb corrected (the 2026-10-16 values, 320/1307 MB for setup_53_2d, were not reproducible; this rerun matches an independent RTX 2080 Ti run: 980/3897 MB). Peak counts tracked tensors only, the WebGPU buffer pool holds more."
}
}
},
"source_repo": "cellmap/mito-aff-unetiso-setup-53-2d"
} |