Skip to content
TheChatArchiveThe Chat Archive
GLMchatglm-6b· Zhipu AIHugging Face Arena Human Preference 55K

how compare two embedding in python

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Claudeclaude-1’s answer.
  1. User

    how compare two embedding in python

  2. GLM Assistant

    To compare two embeddings in Python, you can use the torch.nn.functional module and the torch.nn.utils.functional module.

    Here's an example of how you can do this:

    import torch
    from torch.nn.functional import torch.nn.Conv2d, Conv2dX
    from torch.nn.utils.functional import pad_sequences
    
    # Load the embeddings and create a list of their dimensions
    embedding1 = torch.tensor([[0, 0, 0], [0, 1, 0], [0, 0, 1]], dtype=torch.float32)
    embedding2 = torch.tensor([[0, 0, 0], [0, 1, 0], [0, 0, 1]], dtype=torch.float32)
    
    # Create a list of the dimensions of the embeddings
    dim1 = torch.tensor([embedding1.size(0), embedding1.size(1), embedding1.size(2)])
    dim2 = torch.tensor([embedding2.size(0), embedding2.size(1), embedding2.size(2)])
    
    # Create a list of the embeddings and pad them with zeros
    embeddings_padded = pad_sequences(dim1, padding='post',truncation=True, length_max=100)
    embeddings_un padded = pad_sequences(dim2, padding='post',truncation=True, length_max=100)
    
    # Convert the padded and un padded embeddings to tensors of the same shape
    padded_embeddings = embeddings_padded.view(-1, dim1.size(0))
    un padded_embeddings = embeddings_un padded.view(-1, dim1.size(0))
    
    # Compute the similarities
    similarity = torch.nn.functional.cosine_similarity(padded_embeddings,un padded_embeddings)
    
    print(similarity.item())
    

    This will print the cosine similarity between the padded and un padded embeddings, which should be close to 1.

    Note that this example assumes that the embeddings are of the same length and have the same shape. If the embeddings have different shapes, you will need to reshape them before computing the similarity.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.

Identical first prompt