import gradio as gr from transformers import AutoTokenizer, AutoModelForMaskedLM, AutoModel import torch # --- Model Loading --- tokenizer_splade = None model_splade = None tokenizer_unicoil = None model_unicoil = None # Load SPLADE v3 model try: tokenizer_splade = AutoTokenizer.from_pretrained("naver/splade-cocondenser-selfdistil") model_splade = AutoModelForMaskedLM.from_pretrained("naver/splade-cocondenser-selfdistil") model_splade.eval() # Set to evaluation mode for inference print("SPLADE v3 model loaded successfully!") except Exception as e: print(f"Error loading SPLADE model: {e}") print("Please ensure you have accepted any user access agreements on the Hugging Face Hub page for 'naver/splade-cocondenser-selfdistil'.") # Load UNICOIL model for binary sparse encoding # Load UNICOIL model for binary sparse encoding try: unicoil_model_name = "castorini/unicoil-msmarco-passage" tokenizer_unicoil = AutoTokenizer.from_pretrained(unicoil_model_name) # --- FIX IS HERE --- model_unicoil = AutoModelForMaskedLM.from_pretrained(unicoil_model_name) # ------------------- model_unicoil.eval() # Set to evaluation mode for inference print(f"UNICOIL model '{unicoil_model_name}' loaded successfully!") except Exception as e: print(f"Error loading UNICOIL model: {e}") print(f"Please ensure '{unicoil_model_name}' is accessible (check Hugging Face Hub for potential agreements).") # --- Core Representation Functions --- def get_splade_representation(text): if tokenizer_splade is None or model_splade is None: return "SPLADE model is not loaded. Please check the console for loading errors." inputs = tokenizer_splade(text, return_tensors="pt", padding=True, truncation=True) inputs = {k: v.to(model_splade.device) for k, v in inputs.items()} with torch.no_grad(): output = model_splade(**inputs) if hasattr(output, 'logits'): splade_vector = torch.max(torch.log(1 + torch.relu(output.logits)) * inputs['attention_mask'].unsqueeze(-1), dim=1)[0].squeeze() else: return "Model output structure not as expected for SPLADE. 'logits' not found." indices = torch.nonzero(splade_vector).squeeze().cpu().tolist() if not isinstance(indices, list): indices = [indices] values = splade_vector[indices].cpu().tolist() token_weights = dict(zip(indices, values)) meaningful_tokens = {} for token_id, weight in token_weights.items(): decoded_token = tokenizer_splade.decode([token_id]) if decoded_token not in ["[CLS]", "[SEP]", "[PAD]", "[UNK]"] and len(decoded_token.strip()) > 0: meaningful_tokens[decoded_token] = weight sorted_representation = sorted(meaningful_tokens.items(), key=lambda item: item[1], reverse=True) formatted_output = "SPLADE Representation (Top 20 Terms):\n" if not sorted_representation: formatted_output += "No significant terms found for this input.\n" else: for i, (term, weight) in enumerate(sorted_representation): if i >= 20: break formatted_output += f"- **{term}**: {weight:.4f}\n" formatted_output += "\n--- Raw SPLADE Vector Info ---\n" formatted_output += f"Total non-zero terms in vector: {len(indices)}\n" formatted_output += f"Sparsity: {1 - (len(indices) / tokenizer_splade.vocab_size):.2%}\n" return formatted_output def get_unicoil_binary_representation(text): if tokenizer_unicoil is None or model_unicoil is None: return "UNICOIL model is not loaded. Please check the console for loading errors." inputs = tokenizer_unicoil(text, return_tensors="pt", padding=True, truncation=True) inputs = {k: v.to(model_unicoil.device) for k, v in inputs.items()} with torch.no_grad(): # UNICOIL models often output a dictionary where 'token_scores' or similar # contain the learned weights for each token. The structure can vary. # For 'castorini/unicoil-msmarco-passage', the token scores are typically # the last hidden state of the model, which is then mapped by a linear layer # into the sparse weights. We might need to manually extract those, # or the model itself might be set up to produce the weights directly. # Based on typical UNICOIL implementations, we usually take the output # from the last layer and map it to vocabulary size. # In many UNICOIL variations, the model itself is designed to output # the "re-weight" scores for each token. Let's assume for simplicity # that the model's forward pass returns something that can be interpreted # as per-token scores, perhaps in `output.last_hidden_state` # and then a simple linear layer (not part of the AutoModel usually) # would project this to vocab size. # For simplicity and to fit `AutoModel`, we'll treat the last hidden state # directly as the basis for term importance for now, which is common in similar models, # or if the model already has a head, we use it. # A more robust UNICOIL implementation would involve a specific head # if not using AutoModelForMaskedLM. However, AutoModel gives us the # last hidden states from which we can infer. # For UNICOIL, we're interested in the weighted token scores. # `model(**inputs)` will typically return a `BaseModelOutput` # or `MaskedLMOutput` if it's based on an MLM. # Let's assume the model's output provides the token importance. # A common way to get UNICOIL scores if not explicitly provided as logits: # It's usually a linear layer on top of the last hidden state. # Since AutoModel just gives the base model, we'll mimic the output # as a direct mapping if the model doesn't have a specific head for scores. # However, looking at `castorini/unicoil-msmarco-passage` # its `config.json` might give hints or the model itself is structured. # Often, it uses `BertForMaskedLM` and then applies `log(1+relu)` to the logits. # Let's assume it behaves similar to SPLADE for simplicity of extraction for now, # or we might need to load it as `AutoModelForMaskedLM` if its internal structure # is indeed like that, and then apply a binarization. # Re-evaluating: UNICOIL typically *learns* explicit token weights. # The common approach for UNICOIL with Hugging Face is indeed to load it # as `AutoModelForMaskedLM` and use its `logits` output, similar to SPLADE, # but with a different aggregation strategy. # Let's verify the model type for 'castorini/unicoil-msmarco-passage'. # Its config.json and architecture implies it's a BertForMaskedLM variant. output = model_unicoil(**inputs) # This should be a BaseModelOutputWithPooling or similar if not hasattr(output, 'logits'): # If `model_unicoil` is an `AutoModel` without a classification head, # we need to add a way to get per-token scores. # This is where a custom model head or a specific model class would be needed. # For `castorini/unicoil-msmarco-passage`, it *is* an MLM variant. # So, `output.logits` *should* be available. return "UNICOIL model output structure not as expected. 'logits' not found." # UNICOIL's output is also typically per-token scores from the MLM head. # For UNICOIL, the weights are often taken directly from the logits after pooling. # Unlike SPLADE's log(1+ReLU), UNICOIL's approach can be simpler, # sometimes just taking the maximum of logits (or similar pooling). # A common binarization for UNICOIL is based on the sign of the re-weighted scores. # Let's mimic a common UNICOIL interpretation for obtaining sparse weights # from the logits. The weights are usually sparse and positive. # We can apply a threshold for binarization. # This is a simplification; actual UNICOIL might have specific layers. # For `castorini/unicoil-msmarco-passage`, it uses the `log(1+exp(logits))` formulation # followed by max pooling, then often binarization based on a threshold. # Applying a common interpretation of UNICOIL-like score generation for sparse weights: # Instead of `log(1+ReLU(logits))`, it often uses `torch.log(1 + torch.exp(output.logits))`. # This is essentially the softplus function, which makes values positive and sparse. # Get the sparse weights using the UNICOIL-like transformation sparse_weights = torch.max(torch.log(1 + torch.exp(output.logits)) * inputs['attention_mask'].unsqueeze(-1), dim=1)[0].squeeze() # --- Binarization Step for UNICOIL --- # For true "binary sparse", we threshold these sparse weights. # A common approach is to simply take any non-zero value as 1, and zero as 0. # Or, define a small threshold for binarization if values are very small but non-zero. # For simplicity, let's treat anything above a very small epsilon as 1. # Convert to binary: 1 if weight > epsilon, else 0 threshold = 1e-6 # Define a small threshold for binarization binary_sparse_vector = (sparse_weights > threshold).int() # Get indices of the '1's in the binary vector binary_indices = torch.nonzero(binary_sparse_vector).squeeze().cpu().tolist() if not isinstance(binary_indices, list): binary_indices = [binary_indices] if binary_indices.numel() > 0 else [] # Map token IDs back to terms for the binary representation binary_terms = {} for token_id in binary_indices: decoded_token = tokenizer_unicoil.decode([token_id]) if decoded_token not in ["[CLS]", "[SEP]", "[PAD]", "[UNK]"] and len(decoded_token.strip()) > 0: binary_terms[decoded_token] = 1 # Value is always 1 for binary sorted_binary_terms = sorted(binary_terms.items(), key=lambda item: item[0]) # Sort by term for consistent display formatted_output = "UNICOIL Binary Sparse Representation (Activated Terms):\n" if not sorted_binary_terms: formatted_output += "No significant terms activated for this input.\n" else: # Display up to 50 activated terms for readability for i, (term, _) in enumerate(sorted_binary_terms): if i >= 50: break formatted_output += f"- **{term}**\n" # Only show term, as weight is always 1 if len(sorted_binary_terms) > 50: formatted_output += f"...and {len(sorted_binary_terms) - 50} more terms.\n" formatted_output += "\n--- Raw Binary Sparse Vector Info ---\n" formatted_output += f"Total activated terms: {len(binary_indices)}\n" # Calculate sparsity based on the number of '1's vs. total vocabulary size formatted_output += f"Sparsity: {1 - (len(binary_indices) / tokenizer_unicoil.vocab_size):.2%}\n" return formatted_output # --- Unified Prediction Function for Gradio --- def predict_representation(model_choice, text): if model_choice == "SPLADE": return get_splade_representation(text) elif model_choice == "UNICOIL (Binary Sparse)": return get_unicoil_binary_representation(text) else: return "Please select a model." # --- Gradio Interface Setup --- demo = gr.Interface( fn=predict_representation, inputs=[ gr.Radio( ["SPLADE", "UNICOIL (Binary Sparse)"], # Added UNICOIL option label="Choose Representation Model", value="SPLADE" # Default selection ), gr.Textbox( lines=5, label="Enter your query or document text here:", placeholder="e.g., Why is Padua the nicest city in Italy?" ) ], outputs=gr.Markdown(), title="🌌 Sparse and Binary Sparse Representation Generator", description="Enter any text to see its SPLADE sparse vector or UNICOIL binary sparse representation.", allow_flagging="never" ) # Launch the Gradio app demo.launch()