{ "schemaVersion": 1, "slug": "unet", "title": "Bring spatial detail back into the decoder", "className": "TinyUNet", "question": "How does a segmentation decoder recover features lost during downsampling?", "description": "Trace a U-shaped network through its bottleneck, then inspect the cropped feature maps that rejoin its expanding path.", "category": "Vision milestones", "tags": [ "U-Net", "Segmentation", "Skip connections", "Convolution" ], "curator": "TensorViz", "attribution": "Original TensorViz teaching example. PyTorch provides the underlying operators.", "license": "No separate redistribution license has been declared for these project examples.", "limitations": "Untrained reduced network with two levels instead of the original four. No biomedical dataset, elastic-augmentation recipe, overlap-tile inference or segmentation accuracy. Heatmaps summarize one feature channel with 8×8 average-pooled thumbnails.", "openingStep": "skip", "provenance": { "capturedAt": "2026-09-17T06:18:45.955001+00:00", "productRevision": "cf0ab1c9c58c1dc640235b742fa032cade00b43c", "sourceSha256": "27af400b501694077500e40027cc19b66d217ee26708e30e805bf0e12b373dbc", "python": "3.13.13", "torch": "2.7.1", "execution": { "inputs": { "args": [ { "shape": [ 1, 1, 96, 96 ], "dtype": "float32", "fill": "randn" } ] }, "device": "cpu", "mode": "eval", "seed": 0, "deterministic": true }, "outcome": "passed", "outputShape": [ 1, 2, 56, 56 ], "method": "TensorViz multifile projection and isolated execution worker", "scope": "Reduced U-Net CPU forward, exact center-crop/concatenation checks and skip-gradient support. Feature thumbnails are pooled recorded activations, not trained segmentation results.", "numericalChecks": [ "Valid convolutions map 96×96 inputs to two-channel 56×56 logits", "Upsampling to 34×34 joins eight decoder and eight centrally cropped skip channels", "The crop preserves exact skip values, rather than interpolating them", "A skip-only loss sends gradients through the cropped region and not through the decoder branch" ] }, "layers": [ { "id": "input", "label": "input", "op": "Input", "kind": "input", "parent": null, "params": { "shape": "1,1,96,96" } }, { "id": "encoder1", "label": "encoder1", "op": "Module", "kind": "module", "parent": null, "params": { "arg_0": 1, "arg_1": 4 }, "source": { "line": 37, "endLine": 37 } }, { "id": "pool1", "label": "pool1", "op": "MaxPool2d", "kind": "layer", "parent": null, "params": { "kernel_size": 2 }, "source": { "line": 38, "endLine": 38 } }, { "id": "encoder2", "label": "encoder2", "op": "Module", "kind": "module", "parent": null, "params": { "arg_0": 4, "arg_1": 8 }, "source": { "line": 39, "endLine": 39 } }, { "id": "pool2", "label": "pool2", "op": "MaxPool2d", "kind": "layer", "parent": null, "params": { "kernel_size": 2 }, "source": { "line": 40, "endLine": 40 } }, { "id": "bottleneck", "label": "bottleneck", "op": "Module", "kind": "module", "parent": null, "params": { "arg_0": 8, "arg_1": 16 }, "source": { "line": 41, "endLine": 41 } }, { "id": "up2", "label": "up2", "op": "ConvTranspose2d", "kind": "layer", "parent": null, "params": { "in_channels": 16, "out_channels": 8, "kernel_size": 2, "stride": 2 }, "source": { "line": 42, "endLine": 42 } }, { "id": "join2", "label": "join2", "op": "Module", "kind": "module", "parent": null, "params": {}, "source": { "line": 43, "endLine": 43 } }, { "id": "decoder2", "label": "decoder2", "op": "Module", "kind": "module", "parent": null, "params": { "arg_0": 16, "arg_1": 8 }, "source": { "line": 44, "endLine": 44 } }, { "id": "up1", "label": "up1", "op": "ConvTranspose2d", "kind": "layer", "parent": null, "params": { "in_channels": 8, "out_channels": 4, "kernel_size": 2, "stride": 2 }, "source": { "line": 45, "endLine": 45 } }, { "id": "join1", "label": "join1", "op": "Module", "kind": "module", "parent": null, "params": {}, "source": { "line": 46, "endLine": 46 } }, { "id": "decoder1", "label": "decoder1", "op": "Module", "kind": "module", "parent": null, "params": { "arg_0": 8, "arg_1": 4 }, "source": { "line": 47, "endLine": 47 } }, { "id": "mask_head", "label": "mask_head", "op": "Conv2d", "kind": "layer", "parent": null, "params": { "in_channels": 4, "out_channels": 2, "kernel_size": 1 }, "source": { "line": 48, "endLine": 48 } }, { "id": "output", "label": "output", "op": "Output", "kind": "output", "parent": null, "params": {} }, { "id": "encoder1.conv1", "label": "encoder1.conv1", "op": "Conv2d", "kind": "layer", "parent": "encoder1", "params": { "kernel_size": 3 }, "source": { "line": 14, "endLine": 14 } }, { "id": "encoder1.relu1", "label": "encoder1.relu1", "op": "ReLU", "kind": "layer", "parent": "encoder1", "params": {}, "source": { "line": 15, "endLine": 15 } }, { "id": "encoder1.conv2", "label": "encoder1.conv2", "op": "Conv2d", "kind": "layer", "parent": "encoder1", "params": { "kernel_size": 3 }, "source": { "line": 16, "endLine": 16 } }, { "id": "encoder1.relu2", "label": "encoder1.relu2", "op": "ReLU", "kind": "layer", "parent": "encoder1", "params": {}, "source": { "line": 17, "endLine": 17 } }, { "id": "encoder2.conv1", "label": "encoder2.conv1", "op": "Conv2d", "kind": "layer", "parent": "encoder2", "params": { "kernel_size": 3 }, "source": { "line": 14, "endLine": 14 } }, { "id": "encoder2.relu1", "label": "encoder2.relu1", "op": "ReLU", "kind": "layer", "parent": "encoder2", "params": {}, "source": { "line": 15, "endLine": 15 } }, { "id": "encoder2.conv2", "label": "encoder2.conv2", "op": "Conv2d", "kind": "layer", "parent": "encoder2", "params": { "kernel_size": 3 }, "source": { "line": 16, "endLine": 16 } }, { "id": "encoder2.relu2", "label": "encoder2.relu2", "op": "ReLU", "kind": "layer", "parent": "encoder2", "params": {}, "source": { "line": 17, "endLine": 17 } }, { "id": "bottleneck.conv1", "label": "bottleneck.conv1", "op": "Conv2d", "kind": "layer", "parent": "bottleneck", "params": { "kernel_size": 3 }, "source": { "line": 14, "endLine": 14 } }, { "id": "bottleneck.relu1", "label": "bottleneck.relu1", "op": "ReLU", "kind": "layer", "parent": "bottleneck", "params": {}, "source": { "line": 15, "endLine": 15 } }, { "id": "bottleneck.conv2", "label": "bottleneck.conv2", "op": "Conv2d", "kind": "layer", "parent": "bottleneck", "params": { "kernel_size": 3 }, "source": { "line": 16, "endLine": 16 } }, { "id": "bottleneck.relu2", "label": "bottleneck.relu2", "op": "ReLU", "kind": "layer", "parent": "bottleneck", "params": {}, "source": { "line": 17, "endLine": 17 } }, { "id": "cat@join2#0", "label": "cat@join2#0", "op": "torch.cat", "kind": "layer", "parent": "join2", "params": { "dim": 1 }, "source": { "line": 30, "endLine": 30 } }, { "id": "decoder2.conv1", "label": "decoder2.conv1", "op": "Conv2d", "kind": "layer", "parent": "decoder2", "params": { "kernel_size": 3 }, "source": { "line": 14, "endLine": 14 } }, { "id": "decoder2.relu1", "label": "decoder2.relu1", "op": "ReLU", "kind": "layer", "parent": "decoder2", "params": {}, "source": { "line": 15, "endLine": 15 } }, { "id": "decoder2.conv2", "label": "decoder2.conv2", "op": "Conv2d", "kind": "layer", "parent": "decoder2", "params": { "kernel_size": 3 }, "source": { "line": 16, "endLine": 16 } }, { "id": "decoder2.relu2", "label": "decoder2.relu2", "op": "ReLU", "kind": "layer", "parent": "decoder2", "params": {}, "source": { "line": 17, "endLine": 17 } }, { "id": "cat@join1#0", "label": "cat@join1#0", "op": "torch.cat", "kind": "layer", "parent": "join1", "params": { "dim": 1 }, "source": { "line": 30, "endLine": 30 } }, { "id": "decoder1.conv1", "label": "decoder1.conv1", "op": "Conv2d", "kind": "layer", "parent": "decoder1", "params": { "kernel_size": 3 }, "source": { "line": 14, "endLine": 14 } }, { "id": "decoder1.relu1", "label": "decoder1.relu1", "op": "ReLU", "kind": "layer", "parent": "decoder1", "params": {}, "source": { "line": 15, "endLine": 15 } }, { "id": "decoder1.conv2", "label": "decoder1.conv2", "op": "Conv2d", "kind": "layer", "parent": "decoder1", "params": { "kernel_size": 3 }, "source": { "line": 16, "endLine": 16 } }, { "id": "decoder1.relu2", "label": "decoder1.relu2", "op": "ReLU", "kind": "layer", "parent": "decoder1", "params": {}, "source": { "line": 17, "endLine": 17 } } ], "connections": [ { "id": "edge-1", "source": "encoder1.conv1", "target": "encoder1.relu1", "sourcePort": null, "targetPort": null }, { "id": "edge-2", "source": "encoder1.relu1", "target": "encoder1.conv2", "sourcePort": null, "targetPort": null }, { "id": "edge-3", "source": "encoder1.conv2", "target": "encoder1.relu2", "sourcePort": null, "targetPort": null }, { "id": "edge-4", "source": "encoder2.conv1", "target": "encoder2.relu1", "sourcePort": null, "targetPort": null }, { "id": "edge-5", "source": "encoder2.relu1", "target": "encoder2.conv2", "sourcePort": null, "targetPort": null }, { "id": "edge-6", "source": "encoder2.conv2", "target": "encoder2.relu2", "sourcePort": null, "targetPort": null }, { "id": "edge-7", "source": "bottleneck.conv1", "target": "bottleneck.relu1", "sourcePort": null, "targetPort": null }, { "id": "edge-8", "source": "bottleneck.relu1", "target": "bottleneck.conv2", "sourcePort": null, "targetPort": null }, { "id": "edge-9", "source": "bottleneck.conv2", "target": "bottleneck.relu2", "sourcePort": null, "targetPort": null }, { "id": "edge-10", "source": "decoder2.conv1", "target": "decoder2.relu1", "sourcePort": null, "targetPort": null }, { "id": "edge-11", "source": "decoder2.relu1", "target": "decoder2.conv2", "sourcePort": null, "targetPort": null }, { "id": "edge-12", "source": "decoder2.conv2", "target": "decoder2.relu2", "sourcePort": null, "targetPort": null }, { "id": "edge-13", "source": "decoder1.conv1", "target": "decoder1.relu1", "sourcePort": null, "targetPort": null }, { "id": "edge-14", "source": "decoder1.relu1", "target": "decoder1.conv2", "sourcePort": null, "targetPort": null }, { "id": "edge-15", "source": "decoder1.conv2", "target": "decoder1.relu2", "sourcePort": null, "targetPort": null }, { "id": "edge-16", "source": "input", "target": "encoder1.conv1", "sourcePort": null, "targetPort": null }, { "id": "edge-17", "source": "input", "target": "encoder1", "sourcePort": null, "targetPort": null }, { "id": "edge-18", "source": "encoder1.relu2", "target": "pool1", "sourcePort": null, "targetPort": null }, { "id": "edge-19", "source": "encoder1", "target": "pool1", "sourcePort": null, "targetPort": null }, { "id": "edge-20", "source": "pool1", "target": "encoder2.conv1", "sourcePort": null, "targetPort": null }, { "id": "edge-21", "source": "pool1", "target": "encoder2", "sourcePort": null, "targetPort": null }, { "id": "edge-22", "source": "encoder2.relu2", "target": "pool2", "sourcePort": null, "targetPort": null }, { "id": "edge-23", "source": "encoder2", "target": "pool2", "sourcePort": null, "targetPort": null }, { "id": "edge-24", "source": "pool2", "target": "bottleneck.conv1", "sourcePort": null, "targetPort": null }, { "id": "edge-25", "source": "pool2", "target": "bottleneck", "sourcePort": null, "targetPort": null }, { "id": "edge-26", "source": "bottleneck.relu2", "target": "up2", "sourcePort": null, "targetPort": null }, { "id": "edge-27", "source": "bottleneck", "target": "up2", "sourcePort": null, "targetPort": null }, { "id": "edge-28", "source": "up2", "target": "cat@join2#0", "sourcePort": null, "targetPort": null }, { "id": "edge-29", "source": "up2", "target": "join2", "sourcePort": null, "targetPort": null }, { "id": "edge-30", "source": "encoder2.relu2", "target": "cat@join2#0", "sourcePort": null, "targetPort": null }, { "id": "edge-31", "source": "encoder2", "target": "join2", "sourcePort": null, "targetPort": null }, { "id": "edge-32", "source": "cat@join2#0", "target": "decoder2.conv1", "sourcePort": null, "targetPort": null }, { "id": "edge-33", "source": "join2", "target": "decoder2", "sourcePort": null, "targetPort": null }, { "id": "edge-34", "source": "decoder2.relu2", "target": "up1", "sourcePort": null, "targetPort": null }, { "id": "edge-35", "source": "decoder2", "target": "up1", "sourcePort": null, "targetPort": null }, { "id": "edge-36", "source": "up1", "target": "cat@join1#0", "sourcePort": null, "targetPort": null }, { "id": "edge-37", "source": "up1", "target": "join1", "sourcePort": null, "targetPort": null }, { "id": "edge-38", "source": "encoder1.relu2", "target": "cat@join1#0", "sourcePort": null, "targetPort": null }, { "id": "edge-39", "source": "encoder1", "target": "join1", "sourcePort": null, "targetPort": null }, { "id": "edge-40", "source": "cat@join1#0", "target": "decoder1.conv1", "sourcePort": null, "targetPort": null }, { "id": "edge-41", "source": "join1", "target": "decoder1", "sourcePort": null, "targetPort": null }, { "id": "edge-42", "source": "decoder1.relu2", "target": "mask_head", "sourcePort": null, "targetPort": null }, { "id": "edge-43", "source": "decoder1", "target": "mask_head", "sourcePort": null, "targetPort": null }, { "id": "edge-44", "source": "mask_head", "target": "output", "sourcePort": null, "targetPort": null } ], "steps": [ { "id": "contract", "target": "pool1", "title": "Contract spatial resolution", "note": "Two valid convolutions first shrink the input from 96 to 92 pixels. A 2×2 max pool halves that width to 46. The encoder retains its 92×92 features for the later high-resolution skip." }, { "id": "bridge", "target": "bottleneck", "title": "Build context at the lowest resolution", "note": "After another encoder and pooling stage, the bottleneck has 16 channels at 17×17. Upsampling doubles the width to 34 before it meets the first cropped skip." }, { "id": "skip", "target": "join2", "title": "Crop and concatenate the encoder features", "note": "The encoder's 42×42 skip is cropped to its central 34×34 region. Concatenating eight skip and eight decoder channels gives 16 channels. This preserves two feature sets; it is not residual addition." }, { "id": "mask", "target": "mask_head", "title": "Make a decision at every output pixel", "note": "A second upsample/cropped-skip stage leads to four 56×56 feature maps. A 1×1 convolution produces two logits per output pixel. Valid convolutions explain why the output is smaller than the input." } ], "milestone": { "name": "U-Net", "year": 2015, "kind": "Architecture", "fidelity": "Two down/up levels with widths 4/8/16, valid 3×3 convolutions, transposed-convolution upsampling and center-cropped skips. A 96×96 input produces 56×56 two-class logits.", "papers": [ { "title": "U-Net: Convolutional Networks for Biomedical Image Segmentation", "url": "https://arxiv.org/abs/1505.04597" } ] }, "experiment": { "kind": "matrices", "title": "Recover detail through cropped skip features.", "description": "Recorded activations from a synthetic rectangle image. U-Net joins high-resolution encoder features to the expanding decoder along the channel axis. Thumbnails show spatial structure; exact dimensions and channel counts are separate.", "controlLabel": "Skip connection", "cases": [ { "id": "coarse", "label": "Coarse skip: 42 → 34 pixels", "target": "join2", "note": "Each square below is an average-pooled thumbnail of one recorded feature channel. The skip is center-cropped to match the upsampled decoder; concatenation keeps both channel sets. These randomly initialized features are not a learned segmentation or an explanation of pixel importance.", "matrices": [ { "label": "Decoder channel 0 · 8×8 pooled thumbnail", "values": [ [ -0.013429127142834346, -0.013536024295974905, -0.013418184550175969, -0.013406500731399835, -0.014262445963016231, -0.014207848888822988, -0.01405626569927489, -0.014116453217500474 ], [ -0.013415241815980689, -0.013434444580830759, -0.013025640862074866, -0.013060382397805062, -0.013992352457591267, -0.014158063502163226, -0.014092162248570277, -0.014122541859838352 ], [ -0.013440478791012378, -0.013469101997431104, -0.012999177546854753, -0.012990059747584432, -0.014001583472266937, -0.01419508071738917, -0.014066341134214474, -0.014127972816328122 ], [ -0.013438767347175634, -0.01347258806635793, -0.013026043630246259, -0.013030131839825713, -0.014030261881937445, -0.014206413097904111, -0.01405487459820324, -0.014125231280197334 ], [ -0.007715828153148316, -0.0077604293541854805, -0.007485593128520011, -0.007390365392029098, -0.009470200031674172, -0.009587378367936855, -0.0095279931793356, -0.009547666753311985 ], [ -0.007713560467315199, -0.007723377872033592, -0.007476780128631273, -0.007394126285615539, -0.009329225943257219, -0.009469825340543688, -0.009479506636742742, -0.00954106584906543 ], [ -0.007722755033099224, -0.00790255179206383, -0.00785133524935638, -0.007755761763944281, -0.009723172831560125, -0.009808642838130868, -0.00968157566093742, -0.00958266465174346 ], [ -0.00770135339019488, -0.007882420953450781, -0.00799645359467557, -0.007965798652595772, -0.009891444140189855, -0.00981309184492735, -0.009650327389602847, -0.009546036716367761 ] ] }, { "label": "Skip before crop · 8×8 pooled thumbnail", "values": [ [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ] ] }, { "label": "Skip after crop · 8×8 pooled thumbnail", "values": [ [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0 ] ] } ], "vectors": [ { "label": "Decoder / skip / joined channel counts", "values": [ 8, 8, 16 ] } ], "metrics": [ { "label": "Skip spatial width before crop", "value": 42 }, { "label": "Shared width after crop", "value": 34 }, { "label": "Pixels cropped from each side", "value": 4 } ] }, { "id": "fine", "label": "Fine skip: 92 → 60 pixels", "target": "join1", "note": "Each square below is an average-pooled thumbnail of one recorded feature channel. The skip is center-cropped to match the upsampled decoder; concatenation keeps both channel sets. These randomly initialized features are not a learned segmentation or an explanation of pixel importance.", "matrices": [ { "label": "Decoder channel 0 · 8×8 pooled thumbnail", "values": [ [ 0.17855490938319776, 0.17862867322908835, 0.17877741245572118, 0.17891897716326094, 0.17891674652303538, 0.1788405409860767, 0.17855004847066688, 0.17856694187221525 ], [ 0.1785475278776386, 0.17857678064202318, 0.17873269215196766, 0.17888360984527826, 0.17889819309088995, 0.17887769579715054, 0.1786224265106748, 0.17856552958481897 ], [ 0.17855604034725686, 0.17857003106352273, 0.1786542552439345, 0.1787913186044501, 0.17880770919376657, 0.17880847209102932, 0.17860934155919295, 0.17857416869622314 ], [ 0.17855691143299904, 0.1785693215115575, 0.17864851988657449, 0.17878732802943995, 0.1788037251436974, 0.1788062622471854, 0.17860768405976601, 0.1785740949018832 ], [ 0.17855688795004504, 0.17856920873219018, 0.17864816871483488, 0.17878699916626026, 0.1788033835717817, 0.17880586547346083, 0.17860748967018317, 0.17857409850244155 ], [ 0.17855657544885167, 0.17856651782637736, 0.17863950706243825, 0.1787739855957343, 0.17878999556589867, 0.1787941316924029, 0.17859963593039282, 0.1785726446282669 ], [ 0.17855562459607646, 0.17848879345600505, 0.1785000922000058, 0.17860517546199378, 0.17862591387565271, 0.17865371103433372, 0.17856138879003033, 0.17857916632445006 ], [ 0.17857508903875985, 0.17846130480371206, 0.17819195133458685, 0.1781870930233484, 0.17819033119006325, 0.17828244645676633, 0.17846021895452344, 0.1785751925955184 ] ] }, { "label": "Skip before crop · 8×8 pooled thumbnail", "values": [ [ 0.1107112281180097, 0.1107112281180097, 0.1107112281180097, 0.1107112281180097, 0.1107112281180097, 0.1107112281180097, 0.1107112281180097, 0.1107112281180097 ], [ 0.1107112281180097, 0.1107112281180097, 0.09826085618198895, 0.08932911870134341, 0.08932911870134341, 0.10395272751760752, 0.1107112281180097, 0.1107112281180097 ], [ 0.1107112281180097, 0.1107112281180097, 0.05762941691009827, 0.0008492707392610185, 0.0008492707392610185, 0.06737896264593758, 0.1107112281180097, 0.1107112281180097 ], [ 0.1107112281180097, 0.1107112281180097, 0.0569851982871242, 0, 0, 0.06793282639138991, 0.1107112281180097, 0.1107112281180097 ], [ 0.1107112281180097, 0.1107112281180097, 0.0569851982871242, 0, 0, 0.06793282639138991, 0.1107112281180097, 0.1107112281180097 ], [ 0.1107112281180097, 0.1107112281180097, 0.05733041221338578, 0.0001847638359969555, 0.0001847638359969555, 0.06740951825904999, 0.1107112281180097, 0.1107112281180097 ], [ 0.1107112281180097, 0.1107112281180097, 0.10681662632697053, 0.1017694508869189, 0.1017694508869189, 0.10667967875384358, 0.1107112281180097, 0.1107112281180097 ], [ 0.1107112281180097, 0.1107112281180097, 0.1107112281180097, 0.1107112281180097, 0.1107112281180097, 0.1107112281180097, 0.1107112281180097, 0.1107112281180097 ] ] }, { "label": "Skip after crop · 8×8 pooled thumbnail", "values": [ [ 0.11071122811800935, 0.09713112573407252, 0.06533477418887111, 0.0660730665871505, 0.0660730665871505, 0.06643312953305144, 0.10819647583562428, 0.11071122811800935 ], [ 0.11071122811800935, 0.08615197170004343, 0.0014122706867803246, 0.0012739061088915277, 0.0012739061088915277, 0.0011736366241140161, 0.1001163625901488, 0.11071122811800935 ], [ 0.11071122811800935, 0.08547779743068634, 0, 0, 0, 0, 0.10189923958708479, 0.11071122811800935 ], [ 0.11071122811800935, 0.08547779743068634, 0, 0, 0, 0, 0.10189923958708479, 0.11071122811800935 ], [ 0.11071122811800935, 0.08547779743068634, 0, 0, 0, 0, 0.10189923958708479, 0.11071122811800935 ], [ 0.11071122811800935, 0.08547779743068634, 0, 0, 0, 0, 0.10189923958708479, 0.11071122811800935 ], [ 0.11071122811800935, 0.08565557898120817, 0.0007375226605644443, 0.00027714575399543327, 0.00027714575399543327, 0.0002425025347460041, 0.1006178666315717, 0.11071122811800935 ], [ 0.11071122811800935, 0.10581578509272167, 0.08405970605056069, 0.08373680451061755, 0.08373680451061755, 0.08326625642520591, 0.10639960831623728, 0.11071122811800935 ] ] } ], "vectors": [ { "label": "Decoder / skip / joined channel counts", "values": [ 4, 4, 8 ] } ], "metrics": [ { "label": "Skip spatial width before crop", "value": 92 }, { "label": "Shared width after crop", "value": 60 }, { "label": "Pixels cropped from each side", "value": 16 } ] } ] }, "revision": "3a861b8cdaf42ce4", "files": { "graph.tensorviz.json": { "url": "/models/unet/3a861b8cdaf42ce4/graph.tensorviz.json", "sha256": "50e343b593ea679ee60525b8463e249fe390657d84216a6739fc07073c96483d", "bytes": 105297 }, "model.py": { "url": "/models/unet/3a861b8cdaf42ce4/model.py", "sha256": "27af400b501694077500e40027cc19b66d217ee26708e30e805bf0e12b373dbc", "bytes": 6725 }, "run.py": { "url": "/models/unet/3a861b8cdaf42ce4/run.py", "sha256": "560c1ffd305fca1acc5be953b26417c97affdcdbaec0e235e19541d4dfabe549", "bytes": 649 }, "requirements.txt": { "url": "/models/unet/3a861b8cdaf42ce4/requirements.txt", "sha256": "edb86016b42cb3cc6cdc0c490aab929d3bb098ced4515394f0b55766033e0daf", "bytes": 91 }, "preview.svg": { "url": "/models/unet/3a861b8cdaf42ce4/preview.svg", "sha256": "507148d7a56b371342fec72bbff4f3369b5e0c6d0b901f8fe7bb805be7667758", "bytes": 2543 } } }