# CLIP ViT-B/32 > Dual-encoder contrastive model — a ViT image tower and a Transformer text tower projected into a shared embedding space (OpenAI 2021) Pick to study cross-modal retrieval / zero-shot classification. Two parallel encoders meet at a contrastive similarity head. - Category: Multimodal - Layers: 38 - Parameters: 151.20M - Input shape (batchless): 3 × 224 × 224 - Output shape: 512 - Verifier verdict: pass - Graph JSON: https://neurarch.com/templates/clip-vit-b32/model.json - Open on the canvas: https://neurarch.com/?template=clip-vit-b32 ## Structure | # | Layer | Type | Parameters | Output shape | | --- | --- | --- | --- | --- | | 1 | image | Input | shape=[3, 224, 224] | 3 × 224 × 224 | | 2 | patch_embed | Patch Embed | embedDim=768, patchSize=32 | 49 × 768 | | 3 | vis_pos | Positional Encoding | embedDim=768, maxLen=50 | 49 × 768 | | 4 | vis_block_1 | Transformer Block | embedDim=768, numHeads=12, ffDim=3072 | 49 × 768 | | 5 | vis_block_2 | Transformer Block | embedDim=768, numHeads=12, ffDim=3072 | 49 × 768 | | 6 | vis_block_3 | Transformer Block | embedDim=768, numHeads=12, ffDim=3072 | 49 × 768 | | 7 | vis_block_4 | Transformer Block | embedDim=768, numHeads=12, ffDim=3072 | 49 × 768 | | 8 | vis_block_5 | Transformer Block | embedDim=768, numHeads=12, ffDim=3072 | 49 × 768 | | 9 | vis_block_6 | Transformer Block | embedDim=768, numHeads=12, ffDim=3072 | 49 × 768 | | 10 | vis_block_7 | Transformer Block | embedDim=768, numHeads=12, ffDim=3072 | 49 × 768 | | 11 | vis_block_8 | Transformer Block | embedDim=768, numHeads=12, ffDim=3072 | 49 × 768 | | 12 | vis_block_9 | Transformer Block | embedDim=768, numHeads=12, ffDim=3072 | 49 × 768 | | 13 | vis_block_10 | Transformer Block | embedDim=768, numHeads=12, ffDim=3072 | 49 × 768 | | 14 | vis_block_11 | Transformer Block | embedDim=768, numHeads=12, ffDim=3072 | 49 × 768 | | 15 | vis_block_12 | Transformer Block | embedDim=768, numHeads=12, ffDim=3072 | 49 × 768 | | 16 | vis_norm | LayerNorm | normalizedShape=768 | 49 × 768 | | 17 | vis_pool | GlobalAvgPool1D | | 49 | | 18 | img_proj | Linear | outFeatures=512, inFeatures=768 | 512 | | 19 | tokens | Input | shape=[1, 77] | 1 × 77 | | 20 | tok_embed | Embedding | | 1 × 77 × 512 | | 21 | txt_pos | Positional Encoding | embedDim=512, maxLen=77 | 1 × 77 × 512 | | 22 | txt_block_1 | Transformer Block | embedDim=512, numHeads=8, ffDim=2048 | 1 × 77 × 512 | | 23 | txt_block_2 | Transformer Block | embedDim=512, numHeads=8, ffDim=2048 | 1 × 77 × 512 | | 24 | txt_block_3 | Transformer Block | embedDim=512, numHeads=8, ffDim=2048 | 1 × 77 × 512 | | 25 | txt_block_4 | Transformer Block | embedDim=512, numHeads=8, ffDim=2048 | 1 × 77 × 512 | | 26 | txt_block_5 | Transformer Block | embedDim=512, numHeads=8, ffDim=2048 | 1 × 77 × 512 | | 27 | txt_block_6 | Transformer Block | embedDim=512, numHeads=8, ffDim=2048 | 1 × 77 × 512 | | 28 | txt_block_7 | Transformer Block | embedDim=512, numHeads=8, ffDim=2048 | 1 × 77 × 512 | | 29 | txt_block_8 | Transformer Block | embedDim=512, numHeads=8, ffDim=2048 | 1 × 77 × 512 | | 30 | txt_block_9 | Transformer Block | embedDim=512, numHeads=8, ffDim=2048 | 1 × 77 × 512 | | 31 | txt_block_10 | Transformer Block | embedDim=512, numHeads=8, ffDim=2048 | 1 × 77 × 512 | | 32 | txt_block_11 | Transformer Block | embedDim=512, numHeads=8, ffDim=2048 | 1 × 77 × 512 | | 33 | txt_block_12 | Transformer Block | embedDim=512, numHeads=8, ffDim=2048 | 1 × 77 × 512 | | 34 | txt_norm | LayerNorm | normalizedShape=512 | 1 × 77 × 512 | | 35 | txt_pool | GlobalAvgPool1D | | 1 | | 36 | txt_proj | Linear | outFeatures=512, inFeatures=512 | 512 | | 37 | similarity | MatMul | | 512 | | 38 | logits | Output | | 512 | ## Verifier findings No finding. Shapes propagate end to end and no advisory rule fires. ## Exported PyTorch (first 46 lines) ```python # Architecture designed with Neurarch: https://neurarch.com # PyTorch: compatible with Python 3.8+ and torch>=1.12 # Colab: pip install torch torchvision (usually pre-installed) # # WARNING: 3 layer(s) below are not yet supported by the PyTorch # exporter and pass their input through UNCHANGED in forward(): # - vis_pool (globalAvgPool1d) # - txt_pool (globalAvgPool1d) # - similarity (matmul) import torch import torch.nn as nn import torch.nn.functional as F from typing import Tuple class CLIPViT_B32(nn.Module): def __init__(self): super().__init__() self.patchEmbed_1 = nn.Conv2d(3, 768, kernel_size=32, stride=32) # Patch embedding (ViT-style) self.transformerBlock_1 = nn.TransformerEncoderLayer(d_model=768, nhead=12, dim_feedforward=3072, batch_first=True) self.transformerBlock_2 = nn.TransformerEncoderLayer(d_model=768, nhead=12, dim_feedforward=3072, batch_first=True) self.transformerBlock_3 = nn.TransformerEncoderLayer(d_model=768, nhead=12, dim_feedforward=3072, batch_first=True) self.transformerBlock_4 = nn.TransformerEncoderLayer(d_model=768, nhead=12, dim_feedforward=3072, batch_first=True) self.transformerBlock_5 = nn.TransformerEncoderLayer(d_model=768, nhead=12, dim_feedforward=3072, batch_first=True) self.transformerBlock_6 = nn.TransformerEncoderLayer(d_model=768, nhead=12, dim_feedforward=3072, batch_first=True) self.transformerBlock_7 = nn.TransformerEncoderLayer(d_model=768, nhead=12, dim_feedforward=3072, batch_first=True) self.transformerBlock_8 = nn.TransformerEncoderLayer(d_model=768, nhead=12, dim_feedforward=3072, batch_first=True) self.transformerBlock_9 = nn.TransformerEncoderLayer(d_model=768, nhead=12, dim_feedforward=3072, batch_first=True) self.transformerBlock_10 = nn.TransformerEncoderLayer(d_model=768, nhead=12, dim_feedforward=3072, batch_first=True) self.transformerBlock_11 = nn.TransformerEncoderLayer(d_model=768, nhead=12, dim_feedforward=3072, batch_first=True) self.transformerBlock_12 = nn.TransformerEncoderLayer(d_model=768, nhead=12, dim_feedforward=3072, batch_first=True) self.layerNorm_1 = nn.LayerNorm(768) self.linear_1 = nn.Linear(768, 512) self.embedding_1 = nn.Embedding(49408, 512) self.transformerBlock_13 = nn.TransformerEncoderLayer(d_model=512, nhead=8, dim_feedforward=2048, batch_first=True) self.transformerBlock_14 = nn.TransformerEncoderLayer(d_model=512, nhead=8, dim_feedforward=2048, batch_first=True) self.transformerBlock_15 = nn.TransformerEncoderLayer(d_model=512, nhead=8, dim_feedforward=2048, batch_first=True) self.transformerBlock_16 = nn.TransformerEncoderLayer(d_model=512, nhead=8, dim_feedforward=2048, batch_first=True) self.transformerBlock_17 = nn.TransformerEncoderLayer(d_model=512, nhead=8, dim_feedforward=2048, batch_first=True) self.transformerBlock_18 = nn.TransformerEncoderLayer(d_model=512, nhead=8, dim_feedforward=2048, batch_first=True) self.transformerBlock_19 = nn.TransformerEncoderLayer(d_model=512, nhead=8, dim_feedforward=2048, batch_first=True) self.transformerBlock_20 = nn.TransformerEncoderLayer(d_model=512, nhead=8, dim_feedforward=2048, batch_first=True) self.transformerBlock_21 = nn.TransformerEncoderLayer(d_model=512, nhead=8, dim_feedforward=2048, batch_first=True) self.transformerBlock_22 = nn.TransformerEncoderLayer(d_model=512, nhead=8, dim_feedforward=2048, batch_first=True) self.transformerBlock_23 = nn.TransformerEncoderLayer(d_model=512, nhead=8, dim_feedforward=2048, batch_first=True) ``` ## Machine access - Every architecture: https://neurarch.com/a/index.json - Verify a graph of your own: `POST https://www.neurarch.com/api/v1/check` (see https://neurarch.com/developer.html) - MCP server, so an agent edits the graph with the checks in the loop: https://neurarch.com/docs/mcp.md