{ "format": "tensorviz.graph", "version": 3, "exportedAt": "2026-09-17T06:47:49.423016+00:00", "status": "static", "warningCount": 0, "document": { "schemaVersion": 2, "semantic": { "graph_id": "snapshot", "name": "TinyCLIP", "revision": 0, "nodes": [ { "node_id": "input", "kind": "input", "op": "Input", "params": { "shape": "3,3,8,8" }, "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": "image_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": "image_encoder", "moduleTemplateId": "model::ImageEncoder", "isGroup": true, "moduleMembers": [ "image_encoder.patches", "flatten@image_encoder#0", "python@image_encoder#0", "image_encoder.class_token", "cat@image_encoder#0", "image_encoder.positions", "add@image_encoder#0", "image_encoder.pre_norm", "image_encoder.block", "image_encoder.post_norm", "image_encoder.project", "python@image_encoder#1" ], "moduleLayerCount": 29 } }, { "node_id": "text_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": "text_encoder", "moduleTemplateId": "model::TextEncoder", "isGroup": true, "moduleMembers": [ "text_encoder.tokens", "text_encoder.positions", "add@text_encoder#0", "text_encoder.block", "text_encoder.norm", "text_encoder.project", "python@text_encoder#0" ], "moduleLayerCount": 24 } }, { "node_id": "similarity", "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": "similarity", "moduleTemplateId": "model::PairSimilarity", "isGroup": true, "moduleMembers": [ "similarity.log_scale", "mul@similarity#0" ], "moduleLayerCount": 2 } }, { "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": "image_encoder.patches", "kind": "layer", "op": "Conv2d", "params": { "in_channels": 3, "out_channels": 8, "kernel_size": 4, "stride": 4, "bias": false }, "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": "image_encoder" } }, { "node_id": "flatten@image_encoder#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": "image_encoder" } }, { "node_id": "python@image_encoder#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": "image_encoder" } }, { "node_id": "image_encoder.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" }, "parentId": "image_encoder" } }, { "node_id": "cat@image_encoder#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" } ] }, "parentId": "image_encoder" } }, { "node_id": "image_encoder.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" }, "parentId": "image_encoder" } }, { "node_id": "add@image_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": "image_encoder" } }, { "node_id": "image_encoder.pre_norm", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 8 }, "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": "image_encoder" } }, { "node_id": "image_encoder.block", "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": "image_encoder", "moduleInstanceId": "image_encoder.block", "moduleTemplateId": "model::EncoderBlock", "isGroup": true, "moduleMembers": [ "image_encoder.block.norm", "image_encoder.block.query", "python@image_encoder.block#0", "image_encoder.block.key", "python@image_encoder.block#1", "div@image_encoder.block#0", "python@image_encoder.block#3", "image_encoder.block.value", "python@image_encoder.block#2", "image_encoder.block.softmax", "python@image_encoder.block#4", "image_encoder.block.project", "add@image_encoder.block#0", "image_encoder.block.ff_norm", "image_encoder.block.up", "image_encoder.block.gelu", "image_encoder.block.down", "add@image_encoder.block#1" ], "moduleLayerCount": 18 } }, { "node_id": "image_encoder.post_norm", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 8 }, "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": "image_encoder" } }, { "node_id": "image_encoder.project", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 4, "bias": false }, "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": "image_encoder" } }, { "node_id": "python@image_encoder#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": "image_encoder" } }, { "node_id": "image_encoder.block.norm", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 8 }, "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": "image_encoder.block" } }, { "node_id": "image_encoder.block.query", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "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": "image_encoder.block" } }, { "node_id": "python@image_encoder.block#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": "image_encoder.block" } }, { "node_id": "image_encoder.block.key", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "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": "image_encoder.block" } }, { "node_id": "python@image_encoder.block#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": "image_encoder.block" } }, { "node_id": "div@image_encoder.block#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": "image_encoder.block" } }, { "node_id": "python@image_encoder.block#3", "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": "image_encoder.block" } }, { "node_id": "image_encoder.block.value", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "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": "image_encoder.block" } }, { "node_id": "python@image_encoder.block#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": "image_encoder.block" } }, { "node_id": "image_encoder.block.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": "image_encoder.block" } }, { "node_id": "python@image_encoder.block#4", "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": "image_encoder.block" } }, { "node_id": "image_encoder.block.project", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "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": "image_encoder.block" } }, { "node_id": "add@image_encoder.block#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": "image_encoder.block" } }, { "node_id": "image_encoder.block.ff_norm", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 8 }, "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": "image_encoder.block" } }, { "node_id": "image_encoder.block.up", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "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": "image_encoder.block" } }, { "node_id": "image_encoder.block.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": "image_encoder.block" } }, { "node_id": "image_encoder.block.down", "kind": "layer", "op": "Linear", "params": { "in_features": 32, "out_features": 8 }, "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": "image_encoder.block" } }, { "node_id": "add@image_encoder.block#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": "image_encoder.block" } }, { "node_id": "text_encoder.tokens", "kind": "layer", "op": "Embedding", "params": { "num_embeddings": 16, "embedding_dim": 8 }, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "embedding", "name": "Embedding", "displayName": "Embedding", "description": "Lookup table for dense vector embeddings", "icon": "Hash", "category": "embedding", "parameters": [ { "name": "num_embeddings", "type": "number", "defaultValue": 10000, "required": true, "description": "Size of the dictionary of embeddings", "min": 1 }, { "name": "embedding_dim", "type": "number", "defaultValue": 128, "required": true, "description": "Size of each embedding vector", "min": 1 }, { "name": "padding_idx", "type": "number", "defaultValue": 0, "required": false, "description": "Pad output with zeros at this index", "min": 0 }, { "name": "max_norm", "type": "number", "defaultValue": 0, "required": false, "description": "If > 0, re-normalizes embeddings with norm > max_norm", "min": 0 }, { "name": "sparse", "type": "boolean", "defaultValue": false, "required": false, "description": "If True, gradient w.r.t. weight is sparse" } ] }, "parentId": "text_encoder" } }, { "node_id": "text_encoder.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" }, "parentId": "text_encoder" } }, { "node_id": "add@text_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": "text_encoder" } }, { "node_id": "text_encoder.block", "kind": "module", "op": "Module", "params": { "causal": true }, "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": "text_encoder", "moduleInstanceId": "text_encoder.block", "moduleTemplateId": "model::EncoderBlock", "isGroup": true, "moduleMembers": [ "text_encoder.block.norm", "text_encoder.block.query", "python@text_encoder.block#0", "text_encoder.block.key", "python@text_encoder.block#1", "div@text_encoder.block#0", "python@text_encoder.block#3", "text_encoder.block.value", "python@text_encoder.block#2", "text_encoder.block.softmax", "python@text_encoder.block#4", "text_encoder.block.project", "add@text_encoder.block#0", "text_encoder.block.ff_norm", "text_encoder.block.up", "text_encoder.block.gelu", "text_encoder.block.down", "add@text_encoder.block#1" ], "moduleLayerCount": 18 } }, { "node_id": "text_encoder.norm", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 8 }, "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": "text_encoder" } }, { "node_id": "text_encoder.project", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 4, "bias": false }, "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": "text_encoder" } }, { "node_id": "python@text_encoder#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": "text_encoder" } }, { "node_id": "text_encoder.block.norm", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 8 }, "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": "text_encoder.block" } }, { "node_id": "text_encoder.block.query", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "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": "text_encoder.block" } }, { "node_id": "python@text_encoder.block#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": "text_encoder.block" } }, { "node_id": "text_encoder.block.key", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "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": "text_encoder.block" } }, { "node_id": "python@text_encoder.block#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": "text_encoder.block" } }, { "node_id": "div@text_encoder.block#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": "text_encoder.block" } }, { "node_id": "python@text_encoder.block#3", "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": "text_encoder.block" } }, { "node_id": "text_encoder.block.value", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "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": "text_encoder.block" } }, { "node_id": "python@text_encoder.block#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": "text_encoder.block" } }, { "node_id": "text_encoder.block.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": "text_encoder.block" } }, { "node_id": "python@text_encoder.block#4", "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": "text_encoder.block" } }, { "node_id": "text_encoder.block.project", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "out_features": 8 }, "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": "text_encoder.block" } }, { "node_id": "add@text_encoder.block#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": "text_encoder.block" } }, { "node_id": "text_encoder.block.ff_norm", "kind": "layer", "op": "LayerNorm", "params": { "normalized_shape": 8 }, "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": "text_encoder.block" } }, { "node_id": "text_encoder.block.up", "kind": "layer", "op": "Linear", "params": { "in_features": 8, "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": "text_encoder.block" } }, { "node_id": "text_encoder.block.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": "text_encoder.block" } }, { "node_id": "text_encoder.block.down", "kind": "layer", "op": "Linear", "params": { "in_features": 32, "out_features": 8 }, "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": "text_encoder.block" } }, { "node_id": "add@text_encoder.block#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": "text_encoder.block" } }, { "node_id": "similarity.log_scale", "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" }, "parentId": "similarity" } }, { "node_id": "mul@similarity#0", "kind": "layer", "op": "torch.mul", "params": {}, "in_ports": [ "x" ], "out_ports": [ "y" ], "metadata": { "layerType": { "id": "mul", "name": "torch.mul", "displayName": "torch.mul", "description": "PyTorch mul operation", "icon": "Activity", "category": "tensor", "parameters": [ { "name": "dim", "type": "string", "defaultValue": "", "required": false, "description": "Dimension(s) to apply the operation on" } ] }, "parentId": "similarity" } } ], "edges": [ { "edge_id": "edge-1", "source_node_id": "image_encoder.block.norm", "target_node_id": "image_encoder.block.query", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-2", "source_node_id": "image_encoder.block.query", "target_node_id": "python@image_encoder.block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-3", "source_node_id": "image_encoder.block.norm", "target_node_id": "image_encoder.block.key", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-4", "source_node_id": "image_encoder.block.key", "target_node_id": "python@image_encoder.block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-5", "source_node_id": "image_encoder.block.norm", "target_node_id": "image_encoder.block.value", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-6", "source_node_id": "image_encoder.block.value", "target_node_id": "python@image_encoder.block#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-7", "source_node_id": "python@image_encoder.block#0", "target_node_id": "div@image_encoder.block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-8", "source_node_id": "python@image_encoder.block#1", "target_node_id": "div@image_encoder.block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-9", "source_node_id": "div@image_encoder.block#0", "target_node_id": "python@image_encoder.block#3", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-10", "source_node_id": "python@image_encoder.block#3", "target_node_id": "image_encoder.block.softmax", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-11", "source_node_id": "image_encoder.block.softmax", "target_node_id": "python@image_encoder.block#4", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-12", "source_node_id": "python@image_encoder.block#2", "target_node_id": "python@image_encoder.block#4", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-13", "source_node_id": "python@image_encoder.block#4", "target_node_id": "image_encoder.block.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-14", "source_node_id": "image_encoder.block.project", "target_node_id": "add@image_encoder.block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-15", "source_node_id": "add@image_encoder.block#0", "target_node_id": "image_encoder.block.ff_norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-16", "source_node_id": "image_encoder.block.ff_norm", "target_node_id": "image_encoder.block.up", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-17", "source_node_id": "image_encoder.block.up", "target_node_id": "image_encoder.block.gelu", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-18", "source_node_id": "image_encoder.block.gelu", "target_node_id": "image_encoder.block.down", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-19", "source_node_id": "add@image_encoder.block#0", "target_node_id": "add@image_encoder.block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-20", "source_node_id": "image_encoder.block.down", "target_node_id": "add@image_encoder.block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-21", "source_node_id": "image_encoder.patches", "target_node_id": "flatten@image_encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-22", "source_node_id": "flatten@image_encoder#0", "target_node_id": "python@image_encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-23", "source_node_id": "image_encoder.class_token", "target_node_id": "cat@image_encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-24", "source_node_id": "python@image_encoder#0", "target_node_id": "cat@image_encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-25", "source_node_id": "cat@image_encoder#0", "target_node_id": "add@image_encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-26", "source_node_id": "image_encoder.positions", "target_node_id": "add@image_encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-27", "source_node_id": "add@image_encoder#0", "target_node_id": "image_encoder.pre_norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-28", "source_node_id": "image_encoder.pre_norm", "target_node_id": "image_encoder.block.norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-29", "source_node_id": "image_encoder.pre_norm", "target_node_id": "python@image_encoder.block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-30", "source_node_id": "image_encoder.pre_norm", "target_node_id": "python@image_encoder.block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-31", "source_node_id": "image_encoder.pre_norm", "target_node_id": "python@image_encoder.block#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-32", "source_node_id": "image_encoder.pre_norm", "target_node_id": "python@image_encoder.block#3", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-33", "source_node_id": "image_encoder.pre_norm", "target_node_id": "python@image_encoder.block#4", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-34", "source_node_id": "image_encoder.pre_norm", "target_node_id": "add@image_encoder.block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-35", "source_node_id": "image_encoder.pre_norm", "target_node_id": "image_encoder.block", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-36", "source_node_id": "add@image_encoder.block#1", "target_node_id": "image_encoder.post_norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-37", "source_node_id": "image_encoder.block", "target_node_id": "image_encoder.post_norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-38", "source_node_id": "image_encoder.post_norm", "target_node_id": "image_encoder.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-39", "source_node_id": "image_encoder.project", "target_node_id": "python@image_encoder#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-40", "source_node_id": "text_encoder.block.norm", "target_node_id": "text_encoder.block.query", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-41", "source_node_id": "text_encoder.block.query", "target_node_id": "python@text_encoder.block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-42", "source_node_id": "text_encoder.block.norm", "target_node_id": "text_encoder.block.key", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-43", "source_node_id": "text_encoder.block.key", "target_node_id": "python@text_encoder.block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-44", "source_node_id": "text_encoder.block.norm", "target_node_id": "text_encoder.block.value", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-45", "source_node_id": "text_encoder.block.value", "target_node_id": "python@text_encoder.block#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-46", "source_node_id": "python@text_encoder.block#0", "target_node_id": "div@text_encoder.block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-47", "source_node_id": "python@text_encoder.block#1", "target_node_id": "div@text_encoder.block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-48", "source_node_id": "div@text_encoder.block#0", "target_node_id": "python@text_encoder.block#3", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-49", "source_node_id": "python@text_encoder.block#3", "target_node_id": "text_encoder.block.softmax", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-50", "source_node_id": "text_encoder.block.softmax", "target_node_id": "python@text_encoder.block#4", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-51", "source_node_id": "python@text_encoder.block#2", "target_node_id": "python@text_encoder.block#4", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-52", "source_node_id": "python@text_encoder.block#4", "target_node_id": "text_encoder.block.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-53", "source_node_id": "text_encoder.block.project", "target_node_id": "add@text_encoder.block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-54", "source_node_id": "add@text_encoder.block#0", "target_node_id": "text_encoder.block.ff_norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-55", "source_node_id": "text_encoder.block.ff_norm", "target_node_id": "text_encoder.block.up", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-56", "source_node_id": "text_encoder.block.up", "target_node_id": "text_encoder.block.gelu", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-57", "source_node_id": "text_encoder.block.gelu", "target_node_id": "text_encoder.block.down", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-58", "source_node_id": "add@text_encoder.block#0", "target_node_id": "add@text_encoder.block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-59", "source_node_id": "text_encoder.block.down", "target_node_id": "add@text_encoder.block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-60", "source_node_id": "text_encoder.tokens", "target_node_id": "add@text_encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-61", "source_node_id": "text_encoder.positions", "target_node_id": "add@text_encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-62", "source_node_id": "add@text_encoder#0", "target_node_id": "text_encoder.block.norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-63", "source_node_id": "add@text_encoder#0", "target_node_id": "python@text_encoder.block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-64", "source_node_id": "add@text_encoder#0", "target_node_id": "python@text_encoder.block#1", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-65", "source_node_id": "add@text_encoder#0", "target_node_id": "python@text_encoder.block#2", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-66", "source_node_id": "add@text_encoder#0", "target_node_id": "python@text_encoder.block#3", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-67", "source_node_id": "add@text_encoder#0", "target_node_id": "python@text_encoder.block#4", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-68", "source_node_id": "add@text_encoder#0", "target_node_id": "add@text_encoder.block#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-69", "source_node_id": "add@text_encoder#0", "target_node_id": "text_encoder.block", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-70", "source_node_id": "add@text_encoder.block#1", "target_node_id": "text_encoder.norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-71", "source_node_id": "text_encoder.block", "target_node_id": "text_encoder.norm", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-72", "source_node_id": "text_encoder.norm", "target_node_id": "text_encoder.project", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-73", "source_node_id": "text_encoder.project", "target_node_id": "python@text_encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-74", "source_node_id": "similarity.log_scale", "target_node_id": "mul@similarity#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-75", "source_node_id": "input", "target_node_id": "image_encoder.patches", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-76", "source_node_id": "input", "target_node_id": "cat@image_encoder#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-77", "source_node_id": "input", "target_node_id": "image_encoder", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-78", "source_node_id": "input", "target_node_id": "text_encoder.tokens", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-79", "source_node_id": "input", "target_node_id": "text_encoder", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-80", "source_node_id": "python@image_encoder#1", "target_node_id": "mul@similarity#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-81", "source_node_id": "image_encoder", "target_node_id": "similarity", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-82", "source_node_id": "python@text_encoder#0", "target_node_id": "mul@similarity#0", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-83", "source_node_id": "text_encoder", "target_node_id": "similarity", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-84", "source_node_id": "mul@similarity#0", "target_node_id": "output", "source_port": "y", "target_port": "x", "metadata": {} }, { "edge_id": "edge-85", "source_node_id": "similarity", "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 }, "image_encoder": { "x": 340, "y": 160 }, "text_encoder": { "x": 680, "y": 160 }, "similarity": { "x": 1020, "y": 160 }, "output": { "x": 1360, "y": 160 }, "image_encoder.patches": { "x": 24, "y": 214 }, "flatten@image_encoder#0": { "x": 24, "y": 364 }, "python@image_encoder#0": { "x": 24, "y": 514 }, "image_encoder.class_token": { "x": 24, "y": 664 }, "cat@image_encoder#0": { "x": 24, "y": 814 }, "image_encoder.positions": { "x": 24, "y": 964 }, "add@image_encoder#0": { "x": 24, "y": 1114 }, "image_encoder.pre_norm": { "x": 24, "y": 1264 }, "image_encoder.block": { "x": 24, "y": 1414 }, "image_encoder.post_norm": { "x": 24, "y": 1564 }, "image_encoder.project": { "x": 24, "y": 1714 }, "python@image_encoder#1": { "x": 24, "y": 1864 }, "image_encoder.block.norm": { "x": 24, "y": 214 }, "image_encoder.block.query": { "x": 24, "y": 364 }, "python@image_encoder.block#0": { "x": 24, "y": 514 }, "image_encoder.block.key": { "x": 24, "y": 664 }, "python@image_encoder.block#1": { "x": 24, "y": 814 }, "div@image_encoder.block#0": { "x": 24, "y": 964 }, "python@image_encoder.block#3": { "x": 24, "y": 1114 }, "image_encoder.block.value": { "x": 24, "y": 1264 }, "python@image_encoder.block#2": { "x": 24, "y": 1414 }, "image_encoder.block.softmax": { "x": 24, "y": 1564 }, "python@image_encoder.block#4": { "x": 24, "y": 1714 }, "image_encoder.block.project": { "x": 24, "y": 1864 }, "add@image_encoder.block#0": { "x": 24, "y": 2014 }, "image_encoder.block.ff_norm": { "x": 24, "y": 2164 }, "image_encoder.block.up": { "x": 24, "y": 2314 }, "image_encoder.block.gelu": { "x": 24, "y": 2464 }, "image_encoder.block.down": { "x": 24, "y": 2614 }, "add@image_encoder.block#1": { "x": 24, "y": 2764 }, "text_encoder.tokens": { "x": 24, "y": 214 }, "text_encoder.positions": { "x": 24, "y": 364 }, "add@text_encoder#0": { "x": 24, "y": 514 }, "text_encoder.block": { "x": 24, "y": 664 }, "text_encoder.norm": { "x": 24, "y": 814 }, "text_encoder.project": { "x": 24, "y": 964 }, "python@text_encoder#0": { "x": 24, "y": 1114 }, "text_encoder.block.norm": { "x": 24, "y": 214 }, "text_encoder.block.query": { "x": 24, "y": 364 }, "python@text_encoder.block#0": { "x": 24, "y": 514 }, "text_encoder.block.key": { "x": 24, "y": 664 }, "python@text_encoder.block#1": { "x": 24, "y": 814 }, "div@text_encoder.block#0": { "x": 24, "y": 964 }, "python@text_encoder.block#3": { "x": 24, "y": 1114 }, "text_encoder.block.value": { "x": 24, "y": 1264 }, "python@text_encoder.block#2": { "x": 24, "y": 1414 }, "text_encoder.block.softmax": { "x": 24, "y": 1564 }, "python@text_encoder.block#4": { "x": 24, "y": 1714 }, "text_encoder.block.project": { "x": 24, "y": 1864 }, "add@text_encoder.block#0": { "x": 24, "y": 2014 }, "text_encoder.block.ff_norm": { "x": 24, "y": 2164 }, "text_encoder.block.up": { "x": 24, "y": 2314 }, "text_encoder.block.gelu": { "x": 24, "y": 2464 }, "text_encoder.block.down": { "x": 24, "y": 2614 }, "add@text_encoder.block#1": { "x": 24, "y": 2764 }, "similarity.log_scale": { "x": 24, "y": 214 }, "mul@similarity#0": { "x": 24, "y": 364 } }, "collapsed_modules": [ "image_encoder", "text_encoder", "similarity", "image_encoder.block", "text_encoder.block" ], "selection": [], "layout": {} } }, "snapshotId": "snapshot-3befe3beceb08080", "collaboration": { "walkthrough": { "title": "Meet in a shared image–text space", "steps": [ { "id": "image", "title": "Encode the image independently", "note": "Four 4 × 4 RGB patches become width-eight tokens. A class token, learned positions and one bidirectional Transformer produce a pooled image representation.", "view": { "camera": { "x": 308.5, "y": -18, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "image_encoder": { "x": 340, "y": 160 }, "text_encoder": { "x": 680, "y": 160 }, "similarity": { "x": 1020, "y": 160 }, "output": { "x": 1360, "y": 160 }, "image_encoder.patches": { "x": 24, "y": 214 }, "flatten@image_encoder#0": { "x": 24, "y": 364 }, "python@image_encoder#0": { "x": 24, "y": 514 }, "image_encoder.class_token": { "x": 24, "y": 664 }, "cat@image_encoder#0": { "x": 24, "y": 814 }, "image_encoder.positions": { "x": 24, "y": 964 }, "add@image_encoder#0": { "x": 24, "y": 1114 }, "image_encoder.pre_norm": { "x": 24, "y": 1264 }, "image_encoder.block": { "x": 24, "y": 1414 }, "image_encoder.post_norm": { "x": 24, "y": 1564 }, "image_encoder.project": { "x": 24, "y": 1714 }, "python@image_encoder#1": { "x": 24, "y": 1864 }, "image_encoder.block.norm": { "x": 24, "y": 214 }, "image_encoder.block.query": { "x": 24, "y": 364 }, "python@image_encoder.block#0": { "x": 24, "y": 514 }, "image_encoder.block.key": { "x": 24, "y": 664 }, "python@image_encoder.block#1": { "x": 24, "y": 814 }, "div@image_encoder.block#0": { "x": 24, "y": 964 }, "python@image_encoder.block#3": { "x": 24, "y": 1114 }, "image_encoder.block.value": { "x": 24, "y": 1264 }, "python@image_encoder.block#2": { "x": 24, "y": 1414 }, "image_encoder.block.softmax": { "x": 24, "y": 1564 }, "python@image_encoder.block#4": { "x": 24, "y": 1714 }, "image_encoder.block.project": { "x": 24, "y": 1864 }, "add@image_encoder.block#0": { "x": 24, "y": 2014 }, "image_encoder.block.ff_norm": { "x": 24, "y": 2164 }, "image_encoder.block.up": { "x": 24, "y": 2314 }, "image_encoder.block.gelu": { "x": 24, "y": 2464 }, "image_encoder.block.down": { "x": 24, "y": 2614 }, "add@image_encoder.block#1": { "x": 24, "y": 2764 }, "text_encoder.tokens": { "x": 24, "y": 214 }, "text_encoder.positions": { "x": 24, "y": 364 }, "add@text_encoder#0": { "x": 24, "y": 514 }, "text_encoder.block": { "x": 24, "y": 664 }, "text_encoder.norm": { "x": 24, "y": 814 }, "text_encoder.project": { "x": 24, "y": 964 }, "python@text_encoder#0": { "x": 24, "y": 1114 }, "text_encoder.block.norm": { "x": 24, "y": 214 }, "text_encoder.block.query": { "x": 24, "y": 364 }, "python@text_encoder.block#0": { "x": 24, "y": 514 }, "text_encoder.block.key": { "x": 24, "y": 664 }, "python@text_encoder.block#1": { "x": 24, "y": 814 }, "div@text_encoder.block#0": { "x": 24, "y": 964 }, "python@text_encoder.block#3": { "x": 24, "y": 1114 }, "text_encoder.block.value": { "x": 24, "y": 1264 }, "python@text_encoder.block#2": { "x": 24, "y": 1414 }, "text_encoder.block.softmax": { "x": 24, "y": 1564 }, "python@text_encoder.block#4": { "x": 24, "y": 1714 }, "text_encoder.block.project": { "x": 24, "y": 1864 }, "add@text_encoder.block#0": { "x": 24, "y": 2014 }, "text_encoder.block.ff_norm": { "x": 24, "y": 2164 }, "text_encoder.block.up": { "x": 24, "y": 2314 }, "text_encoder.block.gelu": { "x": 24, "y": 2464 }, "text_encoder.block.down": { "x": 24, "y": 2614 }, "add@text_encoder.block#1": { "x": 24, "y": 2764 }, "similarity.log_scale": { "x": 24, "y": 214 }, "mul@similarity#0": { "x": 24, "y": 364 } }, "expandedModuleIds": [ "image_encoder" ], "selectedNodeIds": [ "image_encoder.patches" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "image_encoder.patches" }, "snapshotId": "snapshot-3befe3beceb08080", "exportedAt": "2026-09-17T06:47:49.423016+00:00", "modelName": "TinyCLIP", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"image_encoder.patches\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"flatten@image_encoder#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"image_encoder.patches\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"image_encoder.patches\",\"kind\":\"layer\",\"op\":\"Conv2d\",\"params\":{\"bias\":false,\"in_channels\":3,\"kernel_size\":4,\"out_channels\":8,\"stride\":4},\"parent\":\"image_encoder\"}}" } ] }, { "id": "text", "title": "Encode the caption with causal attention", "note": "A separate text Transformer processes four token IDs with a causal mask. The last position stands for EOS and pools the preceding text. Neither encoder reads the other modality.", "view": { "camera": { "x": 308.5, "y": 6, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "image_encoder": { "x": 340, "y": 160 }, "text_encoder": { "x": 680, "y": 160 }, "similarity": { "x": 1020, "y": 160 }, "output": { "x": 1360, "y": 160 }, "image_encoder.patches": { "x": 24, "y": 214 }, "flatten@image_encoder#0": { "x": 24, "y": 364 }, "python@image_encoder#0": { "x": 24, "y": 514 }, "image_encoder.class_token": { "x": 24, "y": 664 }, "cat@image_encoder#0": { "x": 24, "y": 814 }, "image_encoder.positions": { "x": 24, "y": 964 }, "add@image_encoder#0": { "x": 24, "y": 1114 }, "image_encoder.pre_norm": { "x": 24, "y": 1264 }, "image_encoder.block": { "x": 24, "y": 1414 }, "image_encoder.post_norm": { "x": 24, "y": 1564 }, "image_encoder.project": { "x": 24, "y": 1714 }, "python@image_encoder#1": { "x": 24, "y": 1864 }, "image_encoder.block.norm": { "x": 24, "y": 214 }, "image_encoder.block.query": { "x": 24, "y": 364 }, "python@image_encoder.block#0": { "x": 24, "y": 514 }, "image_encoder.block.key": { "x": 24, "y": 664 }, "python@image_encoder.block#1": { "x": 24, "y": 814 }, "div@image_encoder.block#0": { "x": 24, "y": 964 }, "python@image_encoder.block#3": { "x": 24, "y": 1114 }, "image_encoder.block.value": { "x": 24, "y": 1264 }, "python@image_encoder.block#2": { "x": 24, "y": 1414 }, "image_encoder.block.softmax": { "x": 24, "y": 1564 }, "python@image_encoder.block#4": { "x": 24, "y": 1714 }, "image_encoder.block.project": { "x": 24, "y": 1864 }, "add@image_encoder.block#0": { "x": 24, "y": 2014 }, "image_encoder.block.ff_norm": { "x": 24, "y": 2164 }, "image_encoder.block.up": { "x": 24, "y": 2314 }, "image_encoder.block.gelu": { "x": 24, "y": 2464 }, "image_encoder.block.down": { "x": 24, "y": 2614 }, "add@image_encoder.block#1": { "x": 24, "y": 2764 }, "text_encoder.tokens": { "x": 24, "y": 214 }, "text_encoder.positions": { "x": 24, "y": 364 }, "add@text_encoder#0": { "x": 24, "y": 514 }, "text_encoder.block": { "x": 24, "y": 664 }, "text_encoder.norm": { "x": 24, "y": 814 }, "text_encoder.project": { "x": 24, "y": 964 }, "python@text_encoder#0": { "x": 24, "y": 1114 }, "text_encoder.block.norm": { "x": 24, "y": 214 }, "text_encoder.block.query": { "x": 24, "y": 364 }, "python@text_encoder.block#0": { "x": 24, "y": 514 }, "text_encoder.block.key": { "x": 24, "y": 664 }, "python@text_encoder.block#1": { "x": 24, "y": 814 }, "div@text_encoder.block#0": { "x": 24, "y": 964 }, "python@text_encoder.block#3": { "x": 24, "y": 1114 }, "text_encoder.block.value": { "x": 24, "y": 1264 }, "python@text_encoder.block#2": { "x": 24, "y": 1414 }, "text_encoder.block.softmax": { "x": 24, "y": 1564 }, "python@text_encoder.block#4": { "x": 24, "y": 1714 }, "text_encoder.block.project": { "x": 24, "y": 1864 }, "add@text_encoder.block#0": { "x": 24, "y": 2014 }, "text_encoder.block.ff_norm": { "x": 24, "y": 2164 }, "text_encoder.block.up": { "x": 24, "y": 2314 }, "text_encoder.block.gelu": { "x": 24, "y": 2464 }, "text_encoder.block.down": { "x": 24, "y": 2614 }, "add@text_encoder.block#1": { "x": 24, "y": 2764 }, "similarity.log_scale": { "x": 24, "y": 214 }, "mul@similarity#0": { "x": 24, "y": 364 } }, "expandedModuleIds": [ "text_encoder" ], "selectedNodeIds": [ "text_encoder.tokens" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "text_encoder.tokens" }, "snapshotId": "snapshot-3befe3beceb08080", "exportedAt": "2026-09-17T06:47:49.423016+00:00", "modelName": "TinyCLIP", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"input\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"text_encoder.tokens\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"text_encoder.tokens\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"add@text_encoder#0\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"text_encoder.tokens\",\"kind\":\"layer\",\"op\":\"Embedding\",\"params\":{\"embedding_dim\":8,\"num_embeddings\":16},\"parent\":\"text_encoder\"}}" } ] }, { "id": "compare", "title": "Compare every image with every text", "note": "Each encoder linearly projects to four features and L2-normalizes the result. The similarity block computes all nine pairwise cosines and multiplies them by exp(log_scale), the learned inverse temperature.", "view": { "camera": { "x": -705.3846153846154, "y": -103.60839160839157, "z": 1.1349206349206349 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "image_encoder": { "x": 340, "y": 160 }, "text_encoder": { "x": 680, "y": 160 }, "similarity": { "x": 1020, "y": 160 }, "output": { "x": 1360, "y": 160 }, "image_encoder.patches": { "x": 24, "y": 214 }, "flatten@image_encoder#0": { "x": 24, "y": 364 }, "python@image_encoder#0": { "x": 24, "y": 514 }, "image_encoder.class_token": { "x": 24, "y": 664 }, "cat@image_encoder#0": { "x": 24, "y": 814 }, "image_encoder.positions": { "x": 24, "y": 964 }, "add@image_encoder#0": { "x": 24, "y": 1114 }, "image_encoder.pre_norm": { "x": 24, "y": 1264 }, "image_encoder.block": { "x": 24, "y": 1414 }, "image_encoder.post_norm": { "x": 24, "y": 1564 }, "image_encoder.project": { "x": 24, "y": 1714 }, "python@image_encoder#1": { "x": 24, "y": 1864 }, "image_encoder.block.norm": { "x": 24, "y": 214 }, "image_encoder.block.query": { "x": 24, "y": 364 }, "python@image_encoder.block#0": { "x": 24, "y": 514 }, "image_encoder.block.key": { "x": 24, "y": 664 }, "python@image_encoder.block#1": { "x": 24, "y": 814 }, "div@image_encoder.block#0": { "x": 24, "y": 964 }, "python@image_encoder.block#3": { "x": 24, "y": 1114 }, "image_encoder.block.value": { "x": 24, "y": 1264 }, "python@image_encoder.block#2": { "x": 24, "y": 1414 }, "image_encoder.block.softmax": { "x": 24, "y": 1564 }, "python@image_encoder.block#4": { "x": 24, "y": 1714 }, "image_encoder.block.project": { "x": 24, "y": 1864 }, "add@image_encoder.block#0": { "x": 24, "y": 2014 }, "image_encoder.block.ff_norm": { "x": 24, "y": 2164 }, "image_encoder.block.up": { "x": 24, "y": 2314 }, "image_encoder.block.gelu": { "x": 24, "y": 2464 }, "image_encoder.block.down": { "x": 24, "y": 2614 }, "add@image_encoder.block#1": { "x": 24, "y": 2764 }, "text_encoder.tokens": { "x": 24, "y": 214 }, "text_encoder.positions": { "x": 24, "y": 364 }, "add@text_encoder#0": { "x": 24, "y": 514 }, "text_encoder.block": { "x": 24, "y": 664 }, "text_encoder.norm": { "x": 24, "y": 814 }, "text_encoder.project": { "x": 24, "y": 964 }, "python@text_encoder#0": { "x": 24, "y": 1114 }, "text_encoder.block.norm": { "x": 24, "y": 214 }, "text_encoder.block.query": { "x": 24, "y": 364 }, "python@text_encoder.block#0": { "x": 24, "y": 514 }, "text_encoder.block.key": { "x": 24, "y": 664 }, "python@text_encoder.block#1": { "x": 24, "y": 814 }, "div@text_encoder.block#0": { "x": 24, "y": 964 }, "python@text_encoder.block#3": { "x": 24, "y": 1114 }, "text_encoder.block.value": { "x": 24, "y": 1264 }, "python@text_encoder.block#2": { "x": 24, "y": 1414 }, "text_encoder.block.softmax": { "x": 24, "y": 1564 }, "python@text_encoder.block#4": { "x": 24, "y": 1714 }, "text_encoder.block.project": { "x": 24, "y": 1864 }, "add@text_encoder.block#0": { "x": 24, "y": 2014 }, "text_encoder.block.ff_norm": { "x": 24, "y": 2164 }, "text_encoder.block.up": { "x": 24, "y": 2314 }, "text_encoder.block.gelu": { "x": 24, "y": 2464 }, "text_encoder.block.down": { "x": 24, "y": 2614 }, "add@text_encoder.block#1": { "x": 24, "y": 2764 }, "similarity.log_scale": { "x": 24, "y": 214 }, "mul@similarity#0": { "x": 24, "y": 364 } }, "expandedModuleIds": [ "similarity" ], "selectedNodeIds": [ "similarity" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "similarity" }, "snapshotId": "snapshot-3befe3beceb08080", "exportedAt": "2026-09-17T06:47:49.423016+00:00", "modelName": "TinyCLIP", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"image_encoder\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"similarity\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"mul@similarity#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"output\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"python@image_encoder#1\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@similarity#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"python@text_encoder#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@similarity#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"similarity\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"output\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"similarity.log_scale\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@similarity#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"text_encoder\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"similarity\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[\"{\\\"id\\\":\\\"mul@similarity#0\\\",\\\"kind\\\":\\\"layer\\\",\\\"op\\\":\\\"torch.mul\\\",\\\"params\\\":{},\\\"parent\\\":\\\"similarity\\\"}\",\"{\\\"id\\\":\\\"similarity.log_scale\\\",\\\"kind\\\":\\\"layer\\\",\\\"op\\\":\\\"Parameter\\\",\\\"params\\\":{},\\\"parent\\\":\\\"similarity\\\"}\"],\"node\":{\"id\":\"similarity\",\"kind\":\"module\",\"op\":\"Module\",\"params\":{}}}" } ] }, { "id": "objective", "title": "Train the paired diagonal in both directions", "note": "The local objective averages image-to-text and text-to-image cross entropy. Diagonal targets identify supplied pairs; batch negatives supply alternatives. Gradients reach both projections and both encoders. The panel exposes pairing and temperature effects.", "view": { "camera": { "x": 308.5, "y": -1497.5, "z": 1.25 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "image_encoder": { "x": 340, "y": 160 }, "text_encoder": { "x": 680, "y": 160 }, "similarity": { "x": 1020, "y": 160 }, "output": { "x": 1360, "y": 160 }, "image_encoder.patches": { "x": 24, "y": 214 }, "flatten@image_encoder#0": { "x": 24, "y": 364 }, "python@image_encoder#0": { "x": 24, "y": 514 }, "image_encoder.class_token": { "x": 24, "y": 664 }, "cat@image_encoder#0": { "x": 24, "y": 814 }, "image_encoder.positions": { "x": 24, "y": 964 }, "add@image_encoder#0": { "x": 24, "y": 1114 }, "image_encoder.pre_norm": { "x": 24, "y": 1264 }, "image_encoder.block": { "x": 24, "y": 1414 }, "image_encoder.post_norm": { "x": 24, "y": 1564 }, "image_encoder.project": { "x": 24, "y": 1714 }, "python@image_encoder#1": { "x": 24, "y": 1864 }, "image_encoder.block.norm": { "x": 24, "y": 214 }, "image_encoder.block.query": { "x": 24, "y": 364 }, "python@image_encoder.block#0": { "x": 24, "y": 514 }, "image_encoder.block.key": { "x": 24, "y": 664 }, "python@image_encoder.block#1": { "x": 24, "y": 814 }, "div@image_encoder.block#0": { "x": 24, "y": 964 }, "python@image_encoder.block#3": { "x": 24, "y": 1114 }, "image_encoder.block.value": { "x": 24, "y": 1264 }, "python@image_encoder.block#2": { "x": 24, "y": 1414 }, "image_encoder.block.softmax": { "x": 24, "y": 1564 }, "python@image_encoder.block#4": { "x": 24, "y": 1714 }, "image_encoder.block.project": { "x": 24, "y": 1864 }, "add@image_encoder.block#0": { "x": 24, "y": 2014 }, "image_encoder.block.ff_norm": { "x": 24, "y": 2164 }, "image_encoder.block.up": { "x": 24, "y": 2314 }, "image_encoder.block.gelu": { "x": 24, "y": 2464 }, "image_encoder.block.down": { "x": 24, "y": 2614 }, "add@image_encoder.block#1": { "x": 24, "y": 2764 }, "text_encoder.tokens": { "x": 24, "y": 214 }, "text_encoder.positions": { "x": 24, "y": 364 }, "add@text_encoder#0": { "x": 24, "y": 514 }, "text_encoder.block": { "x": 24, "y": 664 }, "text_encoder.norm": { "x": 24, "y": 814 }, "text_encoder.project": { "x": 24, "y": 964 }, "python@text_encoder#0": { "x": 24, "y": 1114 }, "text_encoder.block.norm": { "x": 24, "y": 214 }, "text_encoder.block.query": { "x": 24, "y": 364 }, "python@text_encoder.block#0": { "x": 24, "y": 514 }, "text_encoder.block.key": { "x": 24, "y": 664 }, "python@text_encoder.block#1": { "x": 24, "y": 814 }, "div@text_encoder.block#0": { "x": 24, "y": 964 }, "python@text_encoder.block#3": { "x": 24, "y": 1114 }, "text_encoder.block.value": { "x": 24, "y": 1264 }, "python@text_encoder.block#2": { "x": 24, "y": 1414 }, "text_encoder.block.softmax": { "x": 24, "y": 1564 }, "python@text_encoder.block#4": { "x": 24, "y": 1714 }, "text_encoder.block.project": { "x": 24, "y": 1864 }, "add@text_encoder.block#0": { "x": 24, "y": 2014 }, "text_encoder.block.ff_norm": { "x": 24, "y": 2164 }, "text_encoder.block.up": { "x": 24, "y": 2314 }, "text_encoder.block.gelu": { "x": 24, "y": 2464 }, "text_encoder.block.down": { "x": 24, "y": 2614 }, "add@text_encoder.block#1": { "x": 24, "y": 2764 }, "similarity.log_scale": { "x": 24, "y": 214 }, "mul@similarity#0": { "x": 24, "y": 364 } }, "expandedModuleIds": [ "image_encoder" ], "selectedNodeIds": [ "image_encoder.project" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "image_encoder.project" }, "snapshotId": "snapshot-3befe3beceb08080", "exportedAt": "2026-09-17T06:47:49.423016+00:00", "modelName": "TinyCLIP", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"image_encoder.post_norm\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"image_encoder.project\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"image_encoder.project\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"python@image_encoder#1\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[],\"node\":{\"id\":\"image_encoder.project\",\"kind\":\"layer\",\"op\":\"Linear\",\"params\":{\"bias\":false,\"in_features\":8,\"out_features\":4},\"parent\":\"image_encoder\"}}" } ] } ] }, "threads": [], "savedViews": [ { "id": "compare", "title": "Compare every image with every text", "note": "Each encoder linearly projects to four features and L2-normalizes the result. The similarity block computes all nine pairwise cosines and multiplies them by exp(log_scale), the learned inverse temperature.", "view": { "camera": { "x": -705.3846153846154, "y": -103.60839160839157, "z": 1.1349206349206349 }, "nodePositions": { "input": { "x": 0, "y": 160 }, "image_encoder": { "x": 340, "y": 160 }, "text_encoder": { "x": 680, "y": 160 }, "similarity": { "x": 1020, "y": 160 }, "output": { "x": 1360, "y": 160 }, "image_encoder.patches": { "x": 24, "y": 214 }, "flatten@image_encoder#0": { "x": 24, "y": 364 }, "python@image_encoder#0": { "x": 24, "y": 514 }, "image_encoder.class_token": { "x": 24, "y": 664 }, "cat@image_encoder#0": { "x": 24, "y": 814 }, "image_encoder.positions": { "x": 24, "y": 964 }, "add@image_encoder#0": { "x": 24, "y": 1114 }, "image_encoder.pre_norm": { "x": 24, "y": 1264 }, "image_encoder.block": { "x": 24, "y": 1414 }, "image_encoder.post_norm": { "x": 24, "y": 1564 }, "image_encoder.project": { "x": 24, "y": 1714 }, "python@image_encoder#1": { "x": 24, "y": 1864 }, "image_encoder.block.norm": { "x": 24, "y": 214 }, "image_encoder.block.query": { "x": 24, "y": 364 }, "python@image_encoder.block#0": { "x": 24, "y": 514 }, "image_encoder.block.key": { "x": 24, "y": 664 }, "python@image_encoder.block#1": { "x": 24, "y": 814 }, "div@image_encoder.block#0": { "x": 24, "y": 964 }, "python@image_encoder.block#3": { "x": 24, "y": 1114 }, "image_encoder.block.value": { "x": 24, "y": 1264 }, "python@image_encoder.block#2": { "x": 24, "y": 1414 }, "image_encoder.block.softmax": { "x": 24, "y": 1564 }, "python@image_encoder.block#4": { "x": 24, "y": 1714 }, "image_encoder.block.project": { "x": 24, "y": 1864 }, "add@image_encoder.block#0": { "x": 24, "y": 2014 }, "image_encoder.block.ff_norm": { "x": 24, "y": 2164 }, "image_encoder.block.up": { "x": 24, "y": 2314 }, "image_encoder.block.gelu": { "x": 24, "y": 2464 }, "image_encoder.block.down": { "x": 24, "y": 2614 }, "add@image_encoder.block#1": { "x": 24, "y": 2764 }, "text_encoder.tokens": { "x": 24, "y": 214 }, "text_encoder.positions": { "x": 24, "y": 364 }, "add@text_encoder#0": { "x": 24, "y": 514 }, "text_encoder.block": { "x": 24, "y": 664 }, "text_encoder.norm": { "x": 24, "y": 814 }, "text_encoder.project": { "x": 24, "y": 964 }, "python@text_encoder#0": { "x": 24, "y": 1114 }, "text_encoder.block.norm": { "x": 24, "y": 214 }, "text_encoder.block.query": { "x": 24, "y": 364 }, "python@text_encoder.block#0": { "x": 24, "y": 514 }, "text_encoder.block.key": { "x": 24, "y": 664 }, "python@text_encoder.block#1": { "x": 24, "y": 814 }, "div@text_encoder.block#0": { "x": 24, "y": 964 }, "python@text_encoder.block#3": { "x": 24, "y": 1114 }, "text_encoder.block.value": { "x": 24, "y": 1264 }, "python@text_encoder.block#2": { "x": 24, "y": 1414 }, "text_encoder.block.softmax": { "x": 24, "y": 1564 }, "python@text_encoder.block#4": { "x": 24, "y": 1714 }, "text_encoder.block.project": { "x": 24, "y": 1864 }, "add@text_encoder.block#0": { "x": 24, "y": 2014 }, "text_encoder.block.ff_norm": { "x": 24, "y": 2164 }, "text_encoder.block.up": { "x": 24, "y": 2314 }, "text_encoder.block.gelu": { "x": 24, "y": 2464 }, "text_encoder.block.down": { "x": 24, "y": 2614 }, "add@text_encoder.block#1": { "x": 24, "y": 2764 }, "similarity.log_scale": { "x": 24, "y": 214 }, "mul@similarity#0": { "x": 24, "y": 364 } }, "expandedModuleIds": [ "similarity" ], "selectedNodeIds": [ "similarity" ], "viewport": { "width": 1100, "height": 700 } }, "anchors": [ { "target": { "kind": "node", "nodeId": "similarity" }, "snapshotId": "snapshot-3befe3beceb08080", "exportedAt": "2026-09-17T06:47:49.423016+00:00", "modelName": "TinyCLIP", "context": "{\"connections\":[\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"image_encoder\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"similarity\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"mul@similarity#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"output\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"python@image_encoder#1\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@similarity#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"python@text_encoder#0\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@similarity#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"similarity\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"output\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"similarity.log_scale\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"mul@similarity#0\\\",\\\"targetPort\\\":\\\"x\\\"}\",\"{\\\"kind\\\":\\\"edge\\\",\\\"source\\\":\\\"text_encoder\\\",\\\"sourcePort\\\":\\\"y\\\",\\\"target\\\":\\\"similarity\\\",\\\"targetPort\\\":\\\"x\\\"}\"],\"contents\":[\"{\\\"id\\\":\\\"mul@similarity#0\\\",\\\"kind\\\":\\\"layer\\\",\\\"op\\\":\\\"torch.mul\\\",\\\"params\\\":{},\\\"parent\\\":\\\"similarity\\\"}\",\"{\\\"id\\\":\\\"similarity.log_scale\\\",\\\"kind\\\":\\\"layer\\\",\\\"op\\\":\\\"Parameter\\\",\\\"params\\\":{},\\\"parent\\\":\\\"similarity\\\"}\"],\"node\":{\"id\":\"similarity\",\"kind\":\"module\",\"op\":\"Module\",\"params\":{}}}" } ] } ], "openingViewId": "compare" } }