{ "format": "tensorviz.graph", "version": 3, "exportedAt": "2026-09-17T06:18:49.600088+00:00", "status": "static", "warningCount": 0, "document": { "schemaVersion": 2, "semantic": { "graph_id": "snapshot", "name": "TinyViT", "revision": 0, "nodes": [ { "node_id": "input", "kind": "input", "op": "Input", "params": { "shape": "1,3,16,16" }, "in_ports": [], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "input", "name": "Input", "displayName": "Input", "description": "Model input node", "icon": "ArrowRight", "category": "io", "parameters": [ { "name": "shape", "type": "string", "defaultValue": "1,3,224,224", "required": true, "description": "Input tensor shape (comma-separated, e.g., 1,3,224,224)" }, { "name": "name", "type": "string", "defaultValue": "input", "required": false, "description": "Name for this input" } ] } } }, { "node_id": "patches", "kind": "module", "op": "Module", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "module", "name": "Module", "displayName": "Module", "description": "Custom module group (nn.Module)", "icon": "Package", "category": "module", "parameters": [ { "name": "name", "type": "string", "defaultValue": "CustomModule", "required": true, "description": "Module class name" }, { "name": "description", "type": "string", "defaultValue": "", "required": false, "description": "Module description" } ] }, "moduleInstanceId": "patches", "moduleTemplateId": "model::PatchEmbedding", "isGroup": true, "moduleMembers": [ "patches.project", "flatten@patches#0", "python@patches#0" ], "moduleLayerCount": 3 } }, { "node_id": "class_token", "kind": "layer", "op": "Parameter", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "parameter", "description": "", "category": "misc", "icon": "Box", "parameters": [], "name": "Parameter", "displayName": "Parameter" } } }, { "node_id": "python@TinyViT#0", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] } } }, { "node_id": "cat@TinyViT#0", "kind": "layer", "op": "torch.cat", "params": { "dim": 1 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "cat", "name": "torch.cat", "displayName": "torch.cat", "description": "PyTorch cat operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] } } }, { "node_id": "positions", "kind": "layer", "op": "Parameter", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "parameter", "description": "", "category": "misc", "icon": "Box", "parameters": [], "name": "Parameter", "displayName": "Parameter" } } }, { "node_id": "add@TinyViT#0", "kind": "layer", "op": "torch.add", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "add", "name": "torch.add", "displayName": "torch.add", "description": "PyTorch add operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] } } }, { "node_id": "encoder", "kind": "module", "op": "Module", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "module", "name": "Module", "displayName": "Module", "description": "Custom module group (nn.Module)", "icon": "Package", "category": "module", "parameters": [ { "name": "name", "type": "string", "defaultValue": "CustomModule", "required": true, "description": "Module class name" }, { "name": "description", "type": "string", "defaultValue": "", "required": false, "description": "Module description" } ] }, "moduleInstanceId": "encoder", "moduleTemplateId": "model::Encoder", "isGroup": true, "moduleMembers": [ "encoder.norm1", "encoder.attention", "add@encoder#0", "encoder.norm2", "encoder.up", "encoder.gelu", "encoder.down", "add@encoder#1" ], "moduleLayerCount": 16 } }, { "node_id": "norm", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 16 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "layernorm", "name": "LayerNorm", "displayName": "LayerNorm", "description": "Layer Normalization", "icon": "Layers", "category": "norm", "parameters": [ { "name": "normalized_shape", "type": "string", "defaultValue": "512", "required": true, "description": "Input shape for normalization (comma-separated for multi-dim, e.g., 512)" }, { "name": "eps", "type": "number", "defaultValue": 0.00001, "required": false, "description": "Value for numerical stability", "min": 0 }, { "name": "elementwise_affine", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, learnable affine parameters" } ] } } }, { "node_id": "head", "kind": "layer", "op": "Linear", "params": { "in_features": 16, "out_features": 3 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] } } }, { "node_id": "output", "kind": "output", "op": "Output", "params": {}, "in_ports": [ "x" ], "out_ports": [], "metadata": { "layerType": { "id": "output", "name": "Output", "displayName": "Output", "description": "Model output node", "icon": "ArrowLeft", "category": "io", "parameters": [ { "name": "name", "type": "string", "defaultValue": "output", "required": false, "description": "Name for this output" } ] } } }, { "node_id": "patches.project", "kind": "layer", "op": "Conv2d", "params": { "in_channels": 3, "out_channels": 16, "kernel_size": 4, "stride": 4 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "conv2d", "name": "Conv2d", "displayName": "Conv2D", "description": "2D Convolutional Layer", "icon": "Grid3X3", "category": "conv", "parameters": [ { "name": "in_channels", "type": "number", "defaultValue": 3, "required": true, "description": "Number of input channels", "min": 1 }, { "name": "out_channels", "type": "number", "defaultValue": 64, "required": true, "description": "Number of output channels", "min": 1 }, { "name": "kernel_size", "type": "number", "defaultValue": 3, "required": true, "description": "Size of the convolving kernel", "min": 1 }, { "name": "stride", "type": "number", "defaultValue": 1, "required": false, "description": "Stride of the convolution", "min": 1 }, { "name": "padding", "type": "number", "defaultValue": 0, "required": false, "description": "Zero-padding added to both sides", "min": 0 }, { "name": "dilation", "type": "number", "defaultValue": 1, "required": false, "description": "Spacing between kernel elements", "min": 1 }, { "name": "groups", "type": "number", "defaultValue": 1, "required": false, "description": "Number of blocked connections from input to output channels", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, adds a learnable bias" } ] }, "parentId": "patches" } }, { "node_id": "flatten@patches#0", "kind": "layer", "op": "Flatten", "params": { "start_dim": 2 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "flatten", "name": "Flatten", "displayName": "Flatten", "description": "Flattens input tensor contiguously", "icon": "Maximize2", "category": "tensor", "parameters": [ { "name": "start_dim", "type": "number", "defaultValue": 1, "required": false, "description": "First dim to flatten", "min": 0 }, { "name": "end_dim", "type": "number", "defaultValue": -1, "required": false, "description": "Last dim to flatten" } ] }, "parentId": "patches" } }, { "node_id": "python@patches#0", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] }, "parentId": "patches" } }, { "node_id": "encoder.norm1", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 16 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "layernorm", "name": "LayerNorm", "displayName": "LayerNorm", "description": "Layer Normalization", "icon": "Layers", "category": "norm", "parameters": [ { "name": "normalized_shape", "type": "string", "defaultValue": "512", "required": true, "description": "Input shape for normalization (comma-separated for multi-dim, e.g., 512)" }, { "name": "eps", "type": "number", "defaultValue": 0.00001, "required": false, "description": "Value for numerical stability", "min": 0 }, { "name": "elementwise_affine", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, learnable affine parameters" } ] }, "parentId": "encoder" } }, { "node_id": "encoder.attention", "kind": "module", "op": "Module", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "module", "name": "Module", "displayName": "Module", "description": "Custom module group (nn.Module)", "icon": "Package", "category": "module", "parameters": [ { "name": "name", "type": "string", "defaultValue": "CustomModule", "required": true, "description": "Module class name" }, { "name": "description", "type": "string", "defaultValue": "", "required": false, "description": "Module description" } ] }, "parentId": "encoder", "moduleInstanceId": "encoder.attention", "moduleTemplateId": "model::Attention", "isGroup": true, "moduleMembers": [ "encoder.attention.query", "python@encoder.attention#0", "encoder.attention.key", "python@encoder.attention#1", "div@encoder.attention#0", "encoder.attention.value", "python@encoder.attention#2", "encoder.attention.softmax", "encoder.attention.project" ], "moduleLayerCount": 9 } }, { "node_id": "add@encoder#0", "kind": "layer", "op": "torch.add", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "add", "name": "torch.add", "displayName": "torch.add", "description": "PyTorch add operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "encoder" } }, { "node_id": "encoder.norm2", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 16 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "layernorm", "name": "LayerNorm", "displayName": "LayerNorm", "description": "Layer Normalization", "icon": "Layers", "category": "norm", "parameters": [ { "name": "normalized_shape", "type": "string", "defaultValue": "512", "required": true, "description": "Input shape for normalization (comma-separated for multi-dim, e.g., 512)" }, { "name": "eps", "type": "number", "defaultValue": 0.00001, "required": false, "description": "Value for numerical stability", "min": 0 }, { "name": "elementwise_affine", "type": "boolean", "defaultValue": true, "required": false, "description": "If True, learnable affine parameters" } ] }, "parentId": "encoder" } }, { "node_id": "encoder.up", "kind": "layer", "op": "Linear", "params": { "in_features": 16, "out_features": 32 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "encoder" } }, { "node_id": "encoder.gelu", "kind": "layer", "op": "GELU", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "gelu", "name": "GELU", "displayName": "GELU", "description": "Gaussian Error Linear Unit", "icon": "TrendingUp", "category": "activation", "parameters": [ { "name": "approximate", "type": "select", "defaultValue": "none", "required": false, "description": "The gelu approximation algorithm to use: 'none' or 'tanh'", "options": [ "none", "tanh" ] } ] }, "parentId": "encoder" } }, { "node_id": "encoder.down", "kind": "layer", "op": "Linear", "params": { "in_features": 32, "out_features": 16 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "encoder" } }, { "node_id": "add@encoder#1", "kind": "layer", "op": "torch.add", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "add", "name": "torch.add", "displayName": "torch.add", "description": "PyTorch add operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "encoder" } }, { "node_id": "encoder.attention.query", "kind": "layer", "op": "Linear", "params": { "in_features": 16, "out_features": 16 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "encoder.attention" } }, { "node_id": "python@encoder.attention#0", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] }, "parentId": "encoder.attention" } }, { "node_id": "encoder.attention.key", "kind": "layer", "op": "Linear", "params": { "in_features": 16, "out_features": 16 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "encoder.attention" } }, { "node_id": "python@encoder.attention#1", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] }, "parentId": "encoder.attention" } }, { "node_id": "div@encoder.attention#0", "kind": "layer", "op": "torch.div", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "div", "name": "torch.div", "displayName": "torch.div", "description": "PyTorch div operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "encoder.attention" } }, { "node_id": "encoder.attention.value", "kind": "layer", "op": "Linear", "params": { "in_features": 16, "out_features": 16 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "encoder.attention" } }, { "node_id": "python@encoder.attention#2", "kind": "layer", "op": "Python", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "python", "name": "Python", "displayName": "Python", "description": "A statement from the source file that has no graph equivalent. Its code is kept verbatim and its shape is unknown until the model is traced.", "icon": "Hash", "category": "misc", "parameters": [] }, "parentId": "encoder.attention" } }, { "node_id": "encoder.attention.softmax", "kind": "layer", "op": "Softmax", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "softmax", "name": "Softmax", "displayName": "Softmax", "description": "Softmax over a dimension", "icon": "TrendingUp", "category": "activation", "parameters": [ { "name": "dim", "type": "number", "defaultValue": 1, "required": true, "description": "Dimension along which Softmax is computed" } ] }, "parentId": "encoder.attention" } }, { "node_id": "encoder.attention.project", "kind": "layer", "op": "Linear", "params": { "in_features": 16, "out_features": 16 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "linear", "name": "Linear", "displayName": "Linear", "description": "Fully Connected Layer", "icon": "Minus", "category": "linear", "parameters": [ { "name": "in_features", "type": "number", "defaultValue": 512, "required": true, "description": "Size of input features", "min": 1 }, { "name": "out_features", "type": "number", "defaultValue": 256, "required": true, "description": "Size of output features", "min": 1 }, { "name": "bias", "type": "boolean", "defaultValue": true, "required": false, "description": "If set to False, the layer will not learn an additive bias" } ] }, "parentId": "encoder.attention" } } ], "edges": [ { "edge_id": "edge-1", "source_node_id": "patches.project", "target_node_id": "flatten@patches#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-2", "source_node_id": "flatten@patches#0", "target_node_id": "python@patches#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-3", "source_node_id": "encoder.attention.query", "target_node_id": "python@encoder.attention#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-4", "source_node_id": "encoder.attention.key", "target_node_id": "python@encoder.attention#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-5", "source_node_id": "encoder.attention.value", "target_node_id": "python@encoder.attention#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-6", "source_node_id": "python@encoder.attention#0", "target_node_id": "div@encoder.attention#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-7", "source_node_id": "python@encoder.attention#1", "target_node_id": "div@encoder.attention#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-8", "source_node_id": "div@encoder.attention#0", "target_node_id": "encoder.attention.softmax", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-9", "source_node_id": "encoder.attention.softmax", "target_node_id": "encoder.attention.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-10", "source_node_id": "python@encoder.attention#2", "target_node_id": "encoder.attention.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-11", "source_node_id": "encoder.norm1", "target_node_id": "encoder.attention.query", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-12", "source_node_id": "encoder.norm1", "target_node_id": "python@encoder.attention#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-13", "source_node_id": "encoder.norm1", "target_node_id": "encoder.attention.key", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-14", "source_node_id": "encoder.norm1", "target_node_id": "python@encoder.attention#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-15", "source_node_id": "encoder.norm1", "target_node_id": "encoder.attention.value", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-16", "source_node_id": "encoder.norm1", "target_node_id": "python@encoder.attention#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-17", "source_node_id": "encoder.norm1", "target_node_id": "encoder.attention.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-18", "source_node_id": "encoder.norm1", "target_node_id": "encoder.attention", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-19", "source_node_id": "encoder.attention.project", "target_node_id": "add@encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-20", "source_node_id": "encoder.attention", "target_node_id": "add@encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-21", "source_node_id": "add@encoder#0", "target_node_id": "encoder.norm2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-22", "source_node_id": "encoder.norm2", "target_node_id": "encoder.up", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-23", "source_node_id": "encoder.up", "target_node_id": "encoder.gelu", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-24", "source_node_id": "encoder.gelu", "target_node_id": "encoder.down", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-25", "source_node_id": "add@encoder#0", "target_node_id": "add@encoder#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-26", "source_node_id": "encoder.down", "target_node_id": "add@encoder#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-27", "source_node_id": "input", "target_node_id": "patches.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-28", "source_node_id": "input", "target_node_id": "patches", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-29", "source_node_id": "class_token", "target_node_id": "python@TinyViT#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-30", "source_node_id": "input", "target_node_id": "python@TinyViT#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-31", "source_node_id": "python@TinyViT#0", "target_node_id": "cat@TinyViT#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-32", "source_node_id": "python@patches#0", "target_node_id": "cat@TinyViT#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-33", "source_node_id": "patches", "target_node_id": "cat@TinyViT#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-34", "source_node_id": "cat@TinyViT#0", "target_node_id": "add@TinyViT#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-35", "source_node_id": "positions", "target_node_id": "add@TinyViT#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-36", "source_node_id": "add@TinyViT#0", "target_node_id": "encoder.norm1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-37", "source_node_id": "add@TinyViT#0", "target_node_id": "add@encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-38", "source_node_id": "add@TinyViT#0", "target_node_id": "encoder", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-39", "source_node_id": "add@encoder#1", "target_node_id": "norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-40", "source_node_id": "encoder", "target_node_id": "norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-41", "source_node_id": "norm", "target_node_id": "head", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-42", "source_node_id": "head", "target_node_id": "output", "source_port": "y", "target_port": "x", "metadata": {} } ], "inputs": [], "modules": [], "operations": [], "metadata": {} }, "view": { "graph_id": "snapshot", "node_positions": { "input": { "x": 0, "y": 160 }, "patches": { "x": 340, "y": 160 }, "class_token": { "x": 680, "y": 160 }, "python@TinyViT#0": { "x": 1020, "y": 160 }, "cat@TinyViT#0": { "x": 1360, "y": 160 }, "positions": { "x": 1700, "y": 160 }, "add@TinyViT#0": { "x": 2040, "y": 160 }, "encoder": { "x": 2380, "y": 160 }, "norm": { "x": 2720, "y": 160 }, "head": { "x": 3060, "y": 160 }, "output": { "x": 3400, "y": 160 }, "patches.project": { "x": 24, "y": 214 }, "flatten@patches#0": { "x": 24, "y": 364 }, "python@patches#0": { "x": 24, "y": 514 }, "encoder.norm1": { "x": 24, "y": 214 }, "encoder.attention": { "x": 24, "y": 364 }, "add@encoder#0": { "x": 24, "y": 514 }, "encoder.norm2": { "x": 24, "y": 664 }, "encoder.up": { "x": 24, "y": 814 }, "encoder.gelu": { "x": 24, "y": 964 }, "encoder.down": { "x": 24, "y": 1114 }, "add@encoder#1": { "x": 24, "y": 1264 }, "encoder.attention.query": { "x": 24, "y": 214 }, "python@encoder.attention#0": { "x": 24, "y": 364 }, "encoder.attention.key": { "x": 24, "y": 514 }, "python@encoder.attention#1": { "x": 24, "y": 664 }, "div@encoder.attention#0": { "x": 24, "y": 814 }, "encoder.attention.value": { "x": 24, "y": 964 }, "python@encoder.attention#2": { "x": 24, "y": 1114 }, "encoder.attention.softmax": { "x": 24, "y": 1264 }, "encoder.attention.project": { "x": 24, "y": 1414 } }, "collapsed_modules": [ "patches", "encoder", "encoder.attention" ], "selection": [], "layout": {} } }, "snapshotId": "snapshot-b53e0e366a3d3e8c", "collaboration": { "walkthrough": { "title": "Turn an image into tokens", "steps": [ { "id": "patches", "title": "Project non-overlapping RGB patches", "note": "A 4×4, stride-four convolution implements the same shared linear map as flattening each RGB patch. Sixteen patches each contain 48 input values and become 16-feature tokens.", "view": { "camera": { "x": 308.5, "y": -7.5, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "patches": { "x": 340, "y": 160 }, "class_token": { "x": 680, "y": 160 }, "python@TinyViT#0": { "x": 1020, "y": 160 }, "cat@TinyViT#0": { "x": 1360, "y": 160 }, "positions": { "x": 1700, "y": 160 }, "add@TinyViT#0": { "x": 2040, "y": 160 }, "encoder": { "x": 2380, "y": 160 }, "norm": { "x": 2720, "y": 160 }, "head": { "x": 3060, "y": 160 }, "output": { "x": 3400, "y": 160 }, "patches.project": { "x": 24, "y": 214 }, "flatten@patches#0": { "x": 24, "y": 364 }, "python@patches#0": { "x": 24, "y": 514 }, "encoder.norm1": { "x": 24, "y": 214 }, "encoder.attention": { "x": 24, "y": 364 }, "add@encoder#0": { "x": 24, "y": 514 }, "encoder.norm2": { "x": 24, "y": 664 }, "encoder.up": { "x": 24, "y": 814 }, "encoder.gelu": { "x": 24, "y": 964 }, "encoder.down": { "x": 24, "y": 1114 }, "add@encoder#1": { "x": 24, "y": 1264 }, "encoder.attention.query": { "x": 24, "y": 214 }, "python@encoder.attention#0": { "x": 24, "y": 364 }, "encoder.attention.key": { "x": 24, "y": 514 }, "python@encoder.attention#1": { "x": 24, "y": 664 }, "div@encoder.attention#0": { "x": 24, "y": 814 }, "encoder.attention.value": { "x": 24, "y": 964 }, "python@encoder.attention#2": { "x": 24, "y": 1114 }, "encoder.attention.softmax": { "x": 24, "y": 1264 }, "encoder.attention.project": { "x": 24, "y": 1414 } }, "expandedModuleIds": [ "patches" ], "selectedNodeIds": [ "patches.project" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "patches.project" }, "snapshotId": "snapshot-b53e0e366a3d3e8c", "exportedAt": "2026-09-17T06:18:49.600088+00:00", "modelName": "TinyViT", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"patches.project\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"patches.project\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"flatten@patches#0\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"patches.project\",\"kind\":\"layer\",\"op\":\"Conv2d\",\"params\":{\"in_channels\":3,\"kernel_size\":4,\"out_channels\":16,\"stride\":4},\"parent\":\"patches\"}}" } ] }, { "id": "position", "title": "Add a class token and learned positions", "note": "The learned class token precedes the sixteen image tokens. Seventeen position vectors are added so the encoder can distinguish patch locations. Without positions, reordering patches preserves the class output.", "view": { "camera": { "x": -1367.5, "y": 60, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "patches": { "x": 340, "y": 160 }, "class_token": { "x": 680, "y": 160 }, "python@TinyViT#0": { "x": 1020, "y": 160 }, "cat@TinyViT#0": { "x": 1360, "y": 160 }, "positions": { "x": 1700, "y": 160 }, "add@TinyViT#0": { "x": 2040, "y": 160 }, "encoder": { "x": 2380, "y": 160 }, "norm": { "x": 2720, "y": 160 }, "head": { "x": 3060, "y": 160 }, "output": { "x": 3400, "y": 160 }, "patches.project": { "x": 24, "y": 214 }, "flatten@patches#0": { "x": 24, "y": 364 }, "python@patches#0": { "x": 24, "y": 514 }, "encoder.norm1": { "x": 24, "y": 214 }, "encoder.attention": { "x": 24, "y": 364 }, "add@encoder#0": { "x": 24, "y": 514 }, "encoder.norm2": { "x": 24, "y": 664 }, "encoder.up": { "x": 24, "y": 814 }, "encoder.gelu": { "x": 24, "y": 964 }, "encoder.down": { "x": 24, "y": 1114 }, "add@encoder#1": { "x": 24, "y": 1264 }, "encoder.attention.query": { "x": 24, "y": 214 }, "python@encoder.attention#0": { "x": 24, "y": 364 }, "encoder.attention.key": { "x": 24, "y": 514 }, "python@encoder.attention#1": { "x": 24, "y": 664 }, "div@encoder.attention#0": { "x": 24, "y": 814 }, "encoder.attention.value": { "x": 24, "y": 964 }, "python@encoder.attention#2": { "x": 24, "y": 1114 }, "encoder.attention.softmax": { "x": 24, "y": 1264 }, "encoder.attention.project": { "x": 24, "y": 1414 } }, "expandedModuleIds": [], "selectedNodeIds": [ "positions" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "positions" }, "snapshotId": "snapshot-b53e0e366a3d3e8c", "exportedAt": "2026-09-17T06:18:49.600088+00:00", "modelName": "TinyViT", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"positions\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"add@TinyViT#0\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"positions\",\"kind\":\"layer\",\"op\":\"Parameter\",\"params\":{}}}" } ] }, { "id": "attention", "title": "Let the class token consult every patch", "note": "The two unmasked attention heads each compare all seventeen tokens. The heatmap below shows head 0's class-to-patch weights; its separate self-weight completes a row that sums to one.", "view": { "camera": { "x": 308.5, "y": -1044, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "patches": { "x": 340, "y": 160 }, "class_token": { "x": 680, "y": 160 }, "python@TinyViT#0": { "x": 1020, "y": 160 }, "cat@TinyViT#0": { "x": 1360, "y": 160 }, "positions": { "x": 1700, "y": 160 }, "add@TinyViT#0": { "x": 2040, "y": 160 }, "encoder": { "x": 2380, "y": 160 }, "norm": { "x": 2720, "y": 160 }, "head": { "x": 3060, "y": 160 }, "output": { "x": 3400, "y": 160 }, "patches.project": { "x": 24, "y": 214 }, "flatten@patches#0": { "x": 24, "y": 364 }, "python@patches#0": { "x": 24, "y": 514 }, "encoder.norm1": { "x": 24, "y": 214 }, "encoder.attention": { "x": 24, "y": 364 }, "add@encoder#0": { "x": 24, "y": 514 }, "encoder.norm2": { "x": 24, "y": 664 }, "encoder.up": { "x": 24, "y": 814 }, "encoder.gelu": { "x": 24, "y": 964 }, "encoder.down": { "x": 24, "y": 1114 }, "add@encoder#1": { "x": 24, "y": 1264 }, "encoder.attention.query": { "x": 24, "y": 214 }, "python@encoder.attention#0": { "x": 24, "y": 364 }, "encoder.attention.key": { "x": 24, "y": 514 }, "python@encoder.attention#1": { "x": 24, "y": 664 }, "div@encoder.attention#0": { "x": 24, "y": 814 }, "encoder.attention.value": { "x": 24, "y": 964 }, "python@encoder.attention#2": { "x": 24, "y": 1114 }, "encoder.attention.softmax": { "x": 24, "y": 1264 }, "encoder.attention.project": { "x": 24, "y": 1414 } }, "expandedModuleIds": [ "encoder", "encoder.attention" ], "selectedNodeIds": [ "encoder.attention.softmax" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "encoder.attention.softmax" }, "snapshotId": "snapshot-b53e0e366a3d3e8c", "exportedAt": "2026-09-17T06:18:49.600088+00:00", "modelName": "TinyViT", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"div@encoder.attention#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"encoder.attention.softmax\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"encoder.attention.softmax\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"encoder.attention.project\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"encoder.attention.softmax\",\"kind\":\"layer\",\"op\":\"Softmax\",\"params\":{},\"parent\":\"encoder.attention\"}}" } ] }, { "id": "classify", "title": "Read the final class-token representation", "note": "After pre-normalized attention and GELU MLP residual branches, the class token is normalized and passed to a three-class head. The other tokens support that representation through attention.", "view": { "camera": { "x": -2727.5, "y": 60, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "patches": { "x": 340, "y": 160 }, "class_token": { "x": 680, "y": 160 }, "python@TinyViT#0": { "x": 1020, "y": 160 }, "cat@TinyViT#0": { "x": 1360, "y": 160 }, "positions": { "x": 1700, "y": 160 }, "add@TinyViT#0": { "x": 2040, "y": 160 }, "encoder": { "x": 2380, "y": 160 }, "norm": { "x": 2720, "y": 160 }, "head": { "x": 3060, "y": 160 }, "output": { "x": 3400, "y": 160 }, "patches.project": { "x": 24, "y": 214 }, "flatten@patches#0": { "x": 24, "y": 364 }, "python@patches#0": { "x": 24, "y": 514 }, "encoder.norm1": { "x": 24, "y": 214 }, "encoder.attention": { "x": 24, "y": 364 }, "add@encoder#0": { "x": 24, "y": 514 }, "encoder.norm2": { "x": 24, "y": 664 }, "encoder.up": { "x": 24, "y": 814 }, "encoder.gelu": { "x": 24, "y": 964 }, "encoder.down": { "x": 24, "y": 1114 }, "add@encoder#1": { "x": 24, "y": 1264 }, "encoder.attention.query": { "x": 24, "y": 214 }, "python@encoder.attention#0": { "x": 24, "y": 364 }, "encoder.attention.key": { "x": 24, "y": 514 }, "python@encoder.attention#1": { "x": 24, "y": 664 }, "div@encoder.attention#0": { "x": 24, "y": 814 }, "encoder.attention.value": { "x": 24, "y": 964 }, "python@encoder.attention#2": { "x": 24, "y": 1114 }, "encoder.attention.softmax": { "x": 24, "y": 1264 }, "encoder.attention.project": { "x": 24, "y": 1414 } }, "expandedModuleIds": [], "selectedNodeIds": [ "head" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "head" }, "snapshotId": "snapshot-b53e0e366a3d3e8c", "exportedAt": "2026-09-17T06:18:49.600088+00:00", "modelName": "TinyViT", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"head\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"output\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"norm\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"head\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"head\",\"kind\":\"layer\",\"op\":\"Linear\",\"params\":{\"in_features\":16,\"out_features\":3}}}" } ] } ] }, "threads": [], "savedViews": [ { "id": "patches", "title": "Project non-overlapping RGB patches", "note": "A 4×4, stride-four convolution implements the same shared linear map as flattening each RGB patch. Sixteen patches each contain 48 input values and become 16-feature tokens.", "view": { "camera": { "x": 308.5, "y": -7.5, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "patches": { "x": 340, "y": 160 }, "class_token": { "x": 680, "y": 160 }, "python@TinyViT#0": { "x": 1020, "y": 160 }, "cat@TinyViT#0": { "x": 1360, "y": 160 }, "positions": { "x": 1700, "y": 160 }, "add@TinyViT#0": { "x": 2040, "y": 160 }, "encoder": { "x": 2380, "y": 160 }, "norm": { "x": 2720, "y": 160 }, "head": { "x": 3060, "y": 160 }, "output": { "x": 3400, "y": 160 }, "patches.project": { "x": 24, "y": 214 }, "flatten@patches#0": { "x": 24, "y": 364 }, "python@patches#0": { "x": 24, "y": 514 }, "encoder.norm1": { "x": 24, "y": 214 }, "encoder.attention": { "x": 24, "y": 364 }, "add@encoder#0": { "x": 24, "y": 514 }, "encoder.norm2": { "x": 24, "y": 664 }, "encoder.up": { "x": 24, "y": 814 }, "encoder.gelu": { "x": 24, "y": 964 }, "encoder.down": { "x": 24, "y": 1114 }, "add@encoder#1": { "x": 24, "y": 1264 }, "encoder.attention.query": { "x": 24, "y": 214 }, "python@encoder.attention#0": { "x": 24, "y": 364 }, "encoder.attention.key": { "x": 24, "y": 514 }, "python@encoder.attention#1": { "x": 24, "y": 664 }, "div@encoder.attention#0": { "x": 24, "y": 814 }, "encoder.attention.value": { "x": 24, "y": 964 }, "python@encoder.attention#2": { "x": 24, "y": 1114 }, "encoder.attention.softmax": { "x": 24, "y": 1264 }, "encoder.attention.project": { "x": 24, "y": 1414 } }, "expandedModuleIds": [ "patches" ], "selectedNodeIds": [ "patches.project" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "patches.project" }, "snapshotId": "snapshot-b53e0e366a3d3e8c", "exportedAt": "2026-09-17T06:18:49.600088+00:00", "modelName": "TinyViT", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"patches.project\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"patches.project\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"flatten@patches#0\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"patches.project\",\"kind\":\"layer\",\"op\":\"Conv2d\",\"params\":{\"in_channels\":3,\"kernel_size\":4,\"out_channels\":16,\"stride\":4},\"parent\":\"patches\"}}" } ] } ], "openingViewId": "patches" } }