import time from transformers import TFAutoModel, AutoTokenizer import tensorflow as tf import numpy as np from typing import List, Tuple, Dict, Optional, Union, Any import math from dataclasses import dataclass import json from pathlib import Path import datetime import faiss import gc from tf_data_pipeline import TFDataPipeline from response_quality_checker import ResponseQualityChecker from cross_encoder_reranker import CrossEncoderReranker from conversation_summarizer import DeviceAwareModel, Summarizer from gpu_monitor import GPUMemoryMonitor import absl.logging from logger_config import config_logger from tqdm.auto import tqdm absl.logging.set_verbosity(absl.logging.WARNING) logger = config_logger(__name__) @dataclass class ChatbotConfig: """Configuration for the RetrievalChatbot.""" max_context_token_limit: int = 512 embedding_dim: int = 768 encoder_units: int = 256 num_attention_heads: int = 8 dropout_rate: float = 0.2 l2_reg_weight: float = 0.001 learning_rate: float = 0.001 min_text_length: int = 3 max_context_turns: int = 5 warmup_steps: int = 200 pretrained_model: str = 'distilbert-base-uncased' dtype: str = 'float32' freeze_embeddings: bool = False embedding_batch_size: int = 64 search_batch_size: int = 64 max_batch_size: int = 64 neg_samples: int = 3 max_retries: int = 3 def to_dict(self) -> Dict: """Convert config to dictionary.""" return {k: (str(v) if isinstance(v, Path) else v) for k, v in self.__dict__.items()} @classmethod def from_dict(cls, config_dict: Dict) -> 'ChatbotConfig': """Create config from dictionary.""" return cls(**{k: v for k, v in config_dict.items() if k in cls.__dataclass_fields__}) class EncoderModel(tf.keras.Model): """Dual encoder model with pretrained embeddings.""" def __init__( self, config: ChatbotConfig, name: str = "encoder", **kwargs ): super().__init__(name=name, **kwargs) self.config = config # Load pretrained model self.pretrained = TFAutoModel.from_pretrained(config.pretrained_model) # Freeze layers based on config self._freeze_layers() # Pooling layer (Global Average Pooling) self.pooler = tf.keras.layers.GlobalAveragePooling1D() # Projection layer self.projection = tf.keras.layers.Dense( config.embedding_dim, activation='tanh', name="projection" ) # Dropout and normalization self.dropout = tf.keras.layers.Dropout(config.dropout_rate) self.normalize = tf.keras.layers.Lambda( lambda x: tf.nn.l2_normalize(x, axis=1), name="l2_normalize" ) def _freeze_layers(self): """Freeze layers of the pretrained model based on configuration.""" if self.config.freeze_embeddings: self.pretrained.trainable = False logger.info("All pretrained layers frozen.") else: # Freeze only the first 'n' transformer layers for i, layer in enumerate(self.pretrained.layers): if isinstance(layer, tf.keras.layers.Layer): if hasattr(layer, 'trainable'): # Freeze the first transformer block if i < 1: layer.trainable = False logger.info(f"Layer {i} frozen.") else: layer.trainable = True def call(self, inputs: tf.Tensor, training: bool = False) -> tf.Tensor: """Forward pass.""" # Get pretrained embeddings pretrained_outputs = self.pretrained(inputs, training=training) x = pretrained_outputs.last_hidden_state # Shape: [batch_size, seq_len, embedding_dim] # Apply pooling, projection, dropout, and normalization x = self.pooler(x) # Shape: [batch_size, 768] x = self.projection(x) # Shape: [batch_size, 768] x = self.dropout(x, training=training) # Apply dropout x = self.normalize(x) # Shape: [batch_size, 768] return x def get_config(self) -> dict: """Return the config of the model.""" config = super().get_config() config.update({ "config": self.config.to_dict(), "name": self.name }) return config class RetrievalChatbot(DeviceAwareModel): """Retrieval-based chatbot using pretrained embeddings and FAISS for similarity search.""" def __init__( self, config: ChatbotConfig, dialogues: List[dict] = [], device: str = None, strategy=None, reranker: Optional[CrossEncoderReranker] = None, summarizer: Optional[Summarizer] = None ): super().__init__() self.config = config self.strategy = strategy self.device = device or self._setup_default_device() # Initialize reranker, summarizer, tokenizer, and memory monitor self.reranker = reranker or self._initialize_reranker() self.summarizer = summarizer or self._initialize_summarizer() self.tokenizer = self._initialize_tokenizer() self.memory_monitor = GPUMemoryMonitor() # Initialize models self.min_batch_size = 8 self.max_batch_size = 128 self.current_batch_size = 32 # Collect unique responses from dialogues self.response_pool, self.unique_responses = self._collect_responses(dialogues) # Initialize training history self.history = { "train_loss": [], "val_loss": [], "train_metrics": {}, "val_metrics": {} } def _setup_default_device(self) -> str: """Set up default device if none is provided.""" if tf.config.list_physical_devices('GPU'): return 'GPU' else: return 'CPU' def _initialize_reranker(self) -> CrossEncoderReranker: """Initialize the CrossEncoderReranker.""" logger.info("Initializing default CrossEncoderReranker...") return CrossEncoderReranker(model_name="cross-encoder/ms-marco-MiniLM-L-12-v2") def _initialize_summarizer(self) -> Summarizer: """Initialize the Summarizer.""" logger.info("Initializing default Summarizer...") return Summarizer(device=self.device) def _initialize_tokenizer(self) -> AutoTokenizer: """Initialize the tokenizer and add special tokens.""" logger.info("Initializing tokenizer and adding special tokens...") tokenizer = AutoTokenizer.from_pretrained(self.config.pretrained_model) special_tokens = { "user": "", "assistant": "", "context": "", "sep": "" } tokenizer.add_special_tokens( {'additional_special_tokens': list(special_tokens.values())} ) return tokenizer def _collect_responses(self, dialogues: List[dict]) -> Tuple[List[str], List[str]]: """ Collect unique responses from dialogues. Returns: response_pool: List of all possible responses. unique_responses: List of unique responses. """ logger.info("Collecting unique responses from dialogues...") responses = set() for dialogue in dialogues: turns = dialogue.get('turns', []) for turn in turns: if turn.get('speaker') == 'assistant' and 'text' in turn: response = turn['text'].strip() if len(response) >= self.config.min_text_length: responses.add(response) response_pool = list(responses) unique_responses = list(responses) # Assuming uniqueness logger.info(f"Collected {len(response_pool)} unique responses.") return response_pool, unique_responses def build_models(self): """Initialize the shared encoder and FAISS index.""" logger.info("Building encoder model...") tf.keras.backend.clear_session() # Shared encoder for both queries and responses self.encoder = EncoderModel( self.config, name="shared_encoder", ) # Resize token embeddings after adding special tokens new_vocab_size = len(self.tokenizer) self.encoder.pretrained.resize_token_embeddings(new_vocab_size) logger.info(f"Token embeddings resized to: {new_vocab_size}") # Initialize FAISS index self._initialize_faiss() # Compute and index embeddings self._compute_and_index_embeddings() # Retrieve embedding dimension from encoder embedding_dim = self.config.embedding_dim vocab_size = len(self.tokenizer) logger.info(f"Encoder Embedding Dimension: {embedding_dim}") logger.info(f"Encoder Embedding Vocabulary Size: {vocab_size}") if vocab_size >= embedding_dim: logger.info("Encoder model built and embeddings resized successfully.") else: logger.error("Vocabulary size is less than embedding dimension.") raise ValueError("Vocabulary size is less than embedding dimension.") def _adjust_batch_size(self) -> None: """Dynamically adjust batch size based on GPU memory usage.""" if self.memory_monitor.should_reduce_batch_size(): new_size = max(self.min_batch_size, self.current_batch_size // 2) if new_size != self.current_batch_size: logger.info(f"Reducing batch size to {new_size} due to high memory usage") self.current_batch_size = new_size gc.collect() if tf.config.list_physical_devices('GPU'): tf.keras.backend.clear_session() elif self.memory_monitor.can_increase_batch_size(): new_size = min(self.max_batch_size, self.current_batch_size * 2) if new_size != self.current_batch_size: logger.info(f"Increasing batch size to {new_size}") self.current_batch_size = new_size def _initialize_faiss(self): """Initialize FAISS with safer GPU handling and memory monitoring.""" logger.info("Initializing FAISS index...") # First, detect if we have GPU-enabled FAISS self.faiss_gpu = False self.gpu_resources = [] try: if hasattr(faiss, 'get_num_gpus'): ngpus = faiss.get_num_gpus() if ngpus > 0: # Configure GPU resources with memory limit for i in range(ngpus): res = faiss.StandardGpuResources() # Set temp memory to 1/4 of total memory to avoid OOM if self.memory_monitor.has_gpu: stats = self.memory_monitor.get_memory_stats() if stats: temp_memory = int(stats.total * 0.25) # 25% of total memory res.setTempMemory(temp_memory) self.gpu_resources.append(res) self.faiss_gpu = True logger.info(f"FAISS GPU resources initialized on {ngpus} GPUs") else: logger.info("Using CPU-only FAISS build") except Exception as e: logger.warning(f"Using CPU due to GPU initialization error: {e}") # TODO: figure out buf with faiss-gpu # TODO: consider IndexIVFFlat in the future (speed). try: # Create appropriate index based on dataset size if len(self.unique_responses) < 1000: logger.info("Small dataset detected, using simple FlatIP index") self.index = faiss.IndexFlatIP(self.config.embedding_dim) else: # Use IVF index with dynamic number of clusters # nlist = min( # 25, # max clusters # max(int(math.sqrt(len(self.unique_responses))), 1) # min 1 cluster # ) # logger.info(f"Using IVF index with {nlist} clusters") # quantizer = faiss.IndexFlatIP(self.config.embedding_dim) # self.index = faiss.IndexIVFFlat( # quantizer, # self.config.embedding_dim, # nlist, # faiss.METRIC_INNER_PRODUCT # ) self.index = faiss.IndexFlatIP(self.config.embedding_dim) # # Move to GPU(s) if available # if self.faiss_gpu and self.gpu_resources: # try: # if len(self.gpu_resources) > 1: # self.index = faiss.index_cpu_to_gpus_list(self.index, self.gpu_resources) # logger.info("FAISS index distributed across multiple GPUs") # else: # self.index = faiss.index_cpu_to_gpu(self.gpu_resources[0], 0, self.index) # logger.info("FAISS index moved to single GPU") # except Exception as e: # logger.warning(f"Failed to move index to GPU: {e}. Falling back to CPU") # self.faiss_gpu = False # # Set search parameters for IVF index # if isinstance(self.index, faiss.IndexIVFFlat): # self.index.nprobe = min(10, nlist) except Exception as e: logger.error(f"Error initializing FAISS: {e}") raise def encode_responses( self, responses: List[str], batch_size: int = 64 ) -> tf.Tensor: """ Encodes responses with more conservative memory management. """ all_embeddings = [] self.current_batch_size = batch_size if self.memory_monitor.has_gpu: batch_size = 128 # Memory stats # if self.memory_monitor.has_gpu: # initial_stats = self.memory_monitor.get_memory_stats() # if initial_stats: # logger.info("Initial GPU memory state:") # logger.info(f"Total: {initial_stats.total / 1e9:.2f}GB") # logger.info(f"Used: {initial_stats.used / 1e9:.2f}GB") # logger.info(f"Free: {initial_stats.free / 1e9:.2f}GB") total_processed = 0 with tqdm(total=len(responses), desc="Encoding responses") as pbar: while total_processed < len(responses): # Monitor memory and adjust batch size if self.memory_monitor.has_gpu: gpu_usage = self.memory_monitor.get_memory_usage() if gpu_usage > 0.8: # Over 80% usage self.current_batch_size = max(128, self.current_batch_size // 2) logger.info(f"High GPU memory usage ({gpu_usage:.1%}), reducing batch size to {self.current_batch_size}") gc.collect() tf.keras.backend.clear_session() # Get batch end_idx = min(total_processed + self.current_batch_size, len(responses)) batch_texts = responses[total_processed:end_idx] try: # Tokenize encodings = self.tokenizer( batch_texts, padding='max_length', truncation=True, max_length=self.config.max_context_token_limit, return_tensors='tf' ) # Encode embeddings_batch = self.encoder(encodings['input_ids'], training=False) # Cast to float32 if embeddings_batch.dtype != tf.float32: embeddings_batch = tf.cast(embeddings_batch, tf.float32) # Store all_embeddings.append(embeddings_batch) # Update progress batch_processed = len(batch_texts) total_processed += batch_processed # Update progress bar if self.memory_monitor.has_gpu: gpu_usage = self.memory_monitor.get_memory_usage() pbar.set_postfix({ 'GPU mem': f'{gpu_usage:.1%}', 'batch_size': self.current_batch_size }) pbar.update(batch_processed) # Memory cleanup every 1000 samples if total_processed % 1000 == 0: gc.collect() if tf.config.list_physical_devices('GPU'): tf.keras.backend.clear_session() except tf.errors.ResourceExhaustedError: logger.warning("GPU memory exhausted during encoding, reducing batch size") self.current_batch_size = max(8, self.current_batch_size // 2) continue except Exception as e: logger.error(f"Error during encoding: {str(e)}") raise # Concatenate results #logger.info("Concatenating embeddings...") if len(all_embeddings) == 1: final_embeddings = all_embeddings[0] else: final_embeddings = tf.concat(all_embeddings, axis=0) return final_embeddings def _train_faiss_index(self, response_embeddings: np.ndarray) -> None: """Train FAISS index with better memory management and robust fallback mechanisms.""" if self.index.is_trained: logger.info("Index already trained, skipping training phase") return logger.info("Starting FAISS index training...") try: # First attempt: Try training with smaller subset subset_size = min(5000, len(response_embeddings)) # Reduced from 10000 logger.info(f"Using {subset_size} samples for initial training attempt") subset_idx = np.random.choice(len(response_embeddings), subset_size, replace=False) training_embeddings = response_embeddings[subset_idx].copy() # Make a copy # Ensure contiguous memory layout training_embeddings = np.ascontiguousarray(training_embeddings) # Force cleanup before training gc.collect() if tf.config.list_physical_devices('GPU'): tf.keras.backend.clear_session() # Verify data properties logger.info(f"FAISS training data shape: {training_embeddings.shape}") logger.info(f"FAISS training data dtype: {training_embeddings.dtype}") logger.info("Starting initial training attempt...") self.index.train(training_embeddings) logger.info("Training completed successfully") except (RuntimeError, Exception) as e: logger.warning(f"Initial training attempt failed: {str(e)}") logger.info("Attempting fallback strategy...") try: # Move to CPU for more stable training if self.faiss_gpu: logger.info("Moving index to CPU for fallback training") cpu_index = faiss.index_gpu_to_cpu(self.index) else: cpu_index = self.index # Create simpler index type if needed if isinstance(cpu_index, faiss.IndexIVFFlat): logger.info("Creating simpler FlatL2 index for fallback") cpu_index = faiss.IndexFlatL2(self.config.embedding_dim) # Use even smaller subset for CPU training subset_size = min(2000, len(response_embeddings)) subset_idx = np.random.choice(len(response_embeddings), subset_size, replace=False) fallback_embeddings = response_embeddings[subset_idx].copy() # Ensure data is properly formatted if not fallback_embeddings.flags['C_CONTIGUOUS']: fallback_embeddings = np.ascontiguousarray(fallback_embeddings) if fallback_embeddings.dtype != np.float32: fallback_embeddings = fallback_embeddings.astype(np.float32) # Train on CPU logger.info("Training fallback index on CPU...") cpu_index.train(fallback_embeddings) # Move back to GPU if needed if self.faiss_gpu: logger.info("Moving trained index back to GPU...") if len(self.gpu_resources) > 1: self.index = faiss.index_cpu_to_gpus_list(cpu_index, self.gpu_resources) else: self.index = faiss.index_cpu_to_gpu(self.gpu_resources[0], 0, cpu_index) else: self.index = cpu_index logger.info("Fallback training completed successfully") except Exception as e2: logger.error(f"Fallback training also failed: {str(e2)}") logger.warning("Creating basic brute-force index as last resort") try: # Create basic brute-force index as last resort dim = response_embeddings.shape[1] basic_index = faiss.IndexFlatL2(dim) if self.faiss_gpu: if len(self.gpu_resources) > 1: self.index = faiss.index_cpu_to_gpus_list(basic_index, self.gpu_resources) else: self.index = faiss.index_cpu_to_gpu(self.gpu_resources[0], 0, basic_index) else: self.index = basic_index logger.info("Basic index created as fallback") except Exception as e3: logger.error(f"All training attempts failed: {str(e3)}") raise RuntimeError("Unable to create working FAISS index") def _add_vectors_to_index(self, response_embeddings: np.ndarray) -> None: """Add vectors to FAISS index with enhanced memory management.""" logger.info("Starting vector addition process...") # Even smaller batches initial_batch_size = 128 min_batch_size = 32 max_batch_size = 1024 total_added = 0 retry_count = 0 max_retries = 5 while total_added < len(response_embeddings): try: # Monitor memory if self.memory_monitor.has_gpu: gpu_usage = self.memory_monitor.get_memory_usage() #logger.info(f"GPU memory usage before batch: {gpu_usage:.1%}") # Force cleanup if memory usage is high if gpu_usage > 0.7: # Lower threshold to 70% logger.info("High memory usage detected, forcing cleanup") gc.collect() tf.keras.backend.clear_session() # Get batch end_idx = min(total_added + initial_batch_size, len(response_embeddings)) batch = response_embeddings[total_added:end_idx] # Add batch self.index.add(batch) # Update progress batch_size = len(batch) total_added += batch_size # Memory cleanup every few batches if total_added % (initial_batch_size * 5) == 0: gc.collect() if tf.config.list_physical_devices('GPU'): tf.keras.backend.clear_session() # Gradually increase batch size if initial_batch_size < max_batch_size: initial_batch_size = min(initial_batch_size + 25, max_batch_size) except Exception as e: logger.warning(f"Error adding batch: {str(e)}") retry_count += 1 if retry_count > max_retries: logger.error("Max retries exceeded.") raise # Reduce batch size initial_batch_size = max(min_batch_size, initial_batch_size // 2) logger.info(f"Reducing batch size to {initial_batch_size} and retrying...") # Cleanup gc.collect() if tf.config.list_physical_devices('GPU'): tf.keras.backend.clear_session() time.sleep(1) # Brief pause before retry logger.info(f"Successfully added all {total_added} vectors to index") def _add_vectors_cpu_fallback(self, remaining_embeddings: np.ndarray, already_added: int = 0) -> None: """CPU fallback with extra safeguards and progress tracking.""" logger.info(f"CPU Fallback: Adding {len(remaining_embeddings)} remaining vectors...") try: # Move index to CPU if self.faiss_gpu: logger.info("Moving index to CPU...") cpu_index = faiss.index_gpu_to_cpu(self.index) else: cpu_index = self.index # Add remaining vectors on CPU with very small batches batch_size = 128 total_added = already_added for i in range(0, len(remaining_embeddings), batch_size): end_idx = min(i + batch_size, len(remaining_embeddings)) batch = remaining_embeddings[i:end_idx] # Add batch cpu_index.add(batch) # Update progress total_added += len(batch) if i % (batch_size * 10) == 0: logger.info(f"Added {total_added} vectors total " f"({i}/{len(remaining_embeddings)} in current phase)") # Periodic cleanup if i % (batch_size * 20) == 0: gc.collect() # Move back to GPU if needed if self.faiss_gpu: logger.info("Moving index back to GPU...") if len(self.gpu_resources) > 1: self.index = faiss.index_cpu_to_gpus_list(cpu_index, self.gpu_resources) else: self.index = faiss.index_cpu_to_gpu(self.gpu_resources[0], 0, cpu_index) else: self.index = cpu_index logger.info("CPU fallback completed successfully") except Exception as e: logger.error(f"Error during CPU fallback: {str(e)}") raise def _compute_and_index_embeddings(self): """Compute embeddings and build FAISS index with simpler handling.""" logger.info("Computing embeddings and indexing with FAISS...") try: # Encode responses with memory monitoring logger.info("Encoding unique responses") response_embeddings = self.encode_responses(self.unique_responses) response_embeddings = response_embeddings.numpy() # Memory cleanup after encoding gc.collect() if tf.config.list_physical_devices('GPU'): tf.keras.backend.clear_session() # Ensure float32 and memory contiguous response_embeddings = response_embeddings.astype('float32') response_embeddings = np.ascontiguousarray(response_embeddings) # Log memory state before normalization if self.memory_monitor.has_gpu: stats = self.memory_monitor.get_memory_stats() if stats: logger.info(f"GPU memory before normalization: {stats.used/1e9:.2f}GB used") # Normalize embeddings logger.info("Normalizing embeddings with FAISS") faiss.normalize_L2(response_embeddings) # Create and initialize simple FlatIP index dim = response_embeddings.shape[1] if self.faiss_gpu: cpu_index = faiss.IndexFlatIP(dim) if len(self.gpu_resources) > 1: self.index = faiss.index_cpu_to_gpus_list(cpu_index, self.gpu_resources) else: self.index = faiss.index_cpu_to_gpu(self.gpu_resources[0], 0, cpu_index) else: self.index = faiss.IndexFlatIP(dim) # Add vectors to index self._add_vectors_to_index(response_embeddings) # Store responses and embeddings self.response_pool = self.unique_responses self.response_embeddings = response_embeddings # Final memory cleanup gc.collect() if tf.config.list_physical_devices('GPU'): tf.keras.backend.clear_session() # Log final state logger.info(f"Successfully indexed {self.index.ntotal} responses") if self.memory_monitor.has_gpu: stats = self.memory_monitor.get_memory_stats() if stats: logger.info(f"Final GPU memory usage: {stats.used/1e9:.2f}GB used") logger.info("Indexing completed successfully") except Exception as e: logger.error(f"Error during indexing: {e}") # Ensure cleanup even on error gc.collect() if tf.config.list_physical_devices('GPU'): tf.keras.backend.clear_session() raise def verify_faiss_index(self): """Verify that FAISS index matches the response pool.""" indexed_size = self.index.ntotal pool_size = len(self.response_pool) logger.info(f"FAISS index size: {indexed_size}") logger.info(f"Response pool size: {pool_size}") if indexed_size != pool_size: logger.warning("Mismatch between FAISS index size and response pool size.") else: logger.info("FAISS index correctly matches the response pool.") def encode_query(self, query: str, context: Optional[List[Tuple[str, str]]] = None) -> tf.Tensor: """Encode a query with optional conversation context.""" # Prepare query with context if context: context_str = ' '.join([ f"{self.special_tokens['user']} {q} " f"{self.special_tokens['assistant']} {r}" for q, r in context[-self.config.max_context_turns:] ]) query = f"{context_str} {self.special_tokens['user']} {query}" else: query = f"{self.special_tokens['user']} {query}" # Tokenize and encode encodings = self.tokenizer( [query], padding='max_length', truncation=True, max_length=self.config.max_context_token_limit, return_tensors='tf' ) input_ids = encodings['input_ids'] # Verify token IDs max_id = tf.reduce_max(input_ids).numpy() new_vocab_size = len(self.tokenizer) if max_id >= new_vocab_size: logger.error(f"Token ID {max_id} exceeds the vocabulary size {new_vocab_size}.") raise ValueError("Token ID exceeds vocabulary size.") # Get embeddings from the shared encoder return self.encoder(input_ids, training=False) def retrieve_responses_cross_encoder( self, query: str, top_k: int, reranker: Optional[CrossEncoderReranker] = None, summarizer: Optional[Summarizer] = None, summarize_threshold: int = 512 # Summarize over 512 tokens ) -> List[Tuple[str, float]]: """ Retrieve top-k from FAISS, then re-rank them with a cross-encoder. Optionally summarize the user query if it's too long. """ if reranker is None: reranker = self.reranker if summarizer is None: summarizer = self.summarizer # Optional summarization if summarizer and len(query.split()) > summarize_threshold: logger.info(f"Query is long. Summarizing before cross-encoder. Original length: {len(query.split())}") query = summarizer.summarize_text(query) logger.info(f"Summarized query: {query}") # 2) Dense retrieval dense_topk = self.retrieve_responses_faiss(query, top_k=top_k) # [(resp, dense_score), ...] if not dense_topk: return [] # 3) Cross-encoder rerank candidate_texts = [pair[0] for pair in dense_topk] cross_scores = reranker.rerank(query, candidate_texts, max_length=256) # Combine combined = [(text, score) for (text, _), score in zip(dense_topk, cross_scores)] # Sort descending by cross-encoder score combined.sort(key=lambda x: x[1], reverse=True) return combined def retrieve_responses_faiss(self, query: str, top_k: int = 5) -> List[Tuple[str, float]]: """Retrieve top-k responses using FAISS.""" # Encode the query q_emb = self.encode_query(query) # Shape: [1, embedding_dim] q_emb_np = q_emb.numpy().astype('float32') # Ensure type match # Normalize the query embedding for cosine similarity faiss.normalize_L2(q_emb_np) # Search the FAISS index distances, indices = self.index.search(q_emb_np, top_k) # Map indices to responses and distances to similarities top_responses = [] for i, idx in enumerate(indices[0]): if idx < len(self.response_pool): top_responses.append((self.response_pool[idx], float(distances[0][i]))) else: logger.warning(f"FAISS returned invalid index {idx}. Skipping.") return top_responses def save_models(self, save_dir: Union[str, Path]): """Save models and configuration.""" save_dir = Path(save_dir) save_dir.mkdir(parents=True, exist_ok=True) # Save config with open(save_dir / "config.json", "w") as f: json.dump(self.config.to_dict(), f, indent=2) # Save models self.encoder.pretrained.save_pretrained(save_dir / "shared_encoder") # Save tokenizer self.tokenizer.save_pretrained(save_dir / "tokenizer") logger.info(f"Models and tokenizer saved to {save_dir}.") @classmethod def load_models(cls, load_dir: Union[str, Path]) -> 'RetrievalChatbot': """Load saved models and configuration.""" load_dir = Path(load_dir) # Load config with open(load_dir / "config.json", "r") as f: config = ChatbotConfig.from_dict(json.load(f)) # Initialize chatbot chatbot = cls(config) # Load models chatbot.encoder.pretrained = TFAutoModel.from_pretrained( load_dir / "shared_encoder", config=config ) # Load tokenizer chatbot.tokenizer = AutoTokenizer.from_pretrained(load_dir / "tokenizer") logger.info(f"Models and tokenizer loaded from {load_dir}.") return chatbot def parse_tfrecord_fn(example_proto, max_length, neg_samples): """ Parses a single TFRecord example. Args: example_proto: A serialized TFRecord example. max_length: The maximum sequence length for tokenization. neg_samples: The number of hard negatives per query. Returns: A tuple of (query_ids, positive_ids, negative_ids). """ feature_description = { 'query_ids': tf.io.FixedLenFeature([max_length], tf.int64), 'positive_ids': tf.io.FixedLenFeature([max_length], tf.int64), 'negative_ids': tf.io.FixedLenFeature([neg_samples * max_length], tf.int64), } parsed_features = tf.io.parse_single_example(example_proto, feature_description) query_ids = tf.cast(parsed_features['query_ids'], tf.int32) positive_ids = tf.cast(parsed_features['positive_ids'], tf.int32) negative_ids = tf.cast(parsed_features['negative_ids'], tf.int32) negative_ids = tf.reshape(negative_ids, [neg_samples, max_length]) return query_ids, positive_ids, negative_ids def train_streaming( self, tfrecord_file_path: str, epochs: int = 20, batch_size: int = 16, validation_split: float = 0.2, checkpoint_dir: str = "checkpoints/", use_lr_schedule: bool = True, peak_lr: float = 2e-5, warmup_steps_ratio: float = 0.1, early_stopping_patience: int = 3, min_delta: float = 1e-4, ) -> None: """Training using a pre-prepared TFRecord dataset.""" logger.info("Starting training with pre-prepared TFRecord dataset...") # Calculate total steps for learning rate schedule # Estimate total pairs by counting the number of records in the TFRecord # Assuming each record corresponds to one pair raw_dataset = tf.data.TFRecordDataset(tfrecord_file_path) total_pairs = sum(1 for _ in raw_dataset) logger.info(f"Total pairs in TFRecord: {total_pairs}") train_size = int(total_pairs * (1 - validation_split)) val_size = total_pairs - train_size steps_per_epoch = math.ceil(train_size / batch_size) val_steps = math.ceil(val_size / batch_size) total_steps = steps_per_epoch * epochs logger.info(f"Training pairs: {train_size}") logger.info(f"Validation pairs: {val_size}") logger.info(f"Steps per epoch: {steps_per_epoch}") logger.info(f"Validation steps: {val_steps}") logger.info(f"Total steps: {total_steps}") # Set up optimizer with learning rate schedule if use_lr_schedule: warmup_steps = int(total_steps * warmup_steps_ratio) lr_schedule = self._get_lr_schedule( total_steps=total_steps, peak_lr=peak_lr, warmup_steps=warmup_steps ) self.optimizer = tf.keras.optimizers.Adam(learning_rate=lr_schedule) logger.info("Using custom learning rate schedule.") else: self.optimizer = tf.keras.optimizers.Adam(learning_rate=peak_lr) logger.info("Using fixed learning rate.") # Initialize checkpoint manager checkpoint = tf.train.Checkpoint(optimizer=self.optimizer, model=self.encoder) manager = tf.train.CheckpointManager(checkpoint, checkpoint_dir, max_to_keep=3) # Setup TensorBoard log_dir = Path(checkpoint_dir) / "tensorboard_logs" log_dir.mkdir(parents=True, exist_ok=True) current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") train_log_dir = str(log_dir / f"train_{current_time}") val_log_dir = str(log_dir / f"val_{current_time}") train_summary_writer = tf.summary.create_file_writer(train_log_dir) val_summary_writer = tf.summary.create_file_writer(val_log_dir) logger.info(f"TensorBoard logs will be saved in {log_dir}") # Define the parsing function with the appropriate max_length and neg_samples parse_fn = lambda x: self.parse_tfrecord_fn(x, self.config.max_context_token_limit, self.config.neg_samples) # Create the full dataset dataset = tf.data.TFRecordDataset(tfrecord_file_path) dataset = dataset.map(parse_fn, num_parallel_calls=tf.data.AUTOTUNE) dataset = dataset.shuffle(buffer_size=10000) # Adjust buffer size as needed TODO: what is this? dataset = dataset.batch(batch_size, drop_remainder=True) dataset = dataset.prefetch(tf.data.AUTOTUNE) # Split into training and validation train_dataset = dataset.take(train_size) val_dataset = dataset.skip(train_size).take(val_size) # Training loop best_val_loss = float("inf") epochs_no_improve = 0 for epoch in range(1, epochs + 1): # --- Training Phase --- epoch_loss_avg = tf.keras.metrics.Mean() batches_processed = 0 try: train_pbar = tqdm(total=steps_per_epoch, desc=f"Training Epoch {epoch}", unit="batch") is_tqdm_train = True except ImportError: train_pbar = None is_tqdm_train = False logger.info("Training progress bar disabled") for q_batch, p_batch, n_batch in train_dataset: loss = self.train_step(q_batch, p_batch, n_batch) epoch_loss_avg(loss) batches_processed += 1 # Log to TensorBoard with train_summary_writer.as_default(): tf.summary.scalar("loss", loss, step=(epoch - 1) * steps_per_epoch + batches_processed) # Update progress bar if use_lr_schedule: current_lr = float(lr_schedule(self.optimizer.iterations)) else: current_lr = float(self.optimizer.learning_rate.numpy()) if is_tqdm_train: train_pbar.update(1) train_pbar.set_postfix({ "loss": f"{loss.numpy():.4f}", "lr": f"{current_lr:.2e}", "batches": f"{batches_processed}/{steps_per_epoch}" }) # Memory cleanup gc.collect() if batches_processed >= steps_per_epoch: break if is_tqdm_train and train_pbar: train_pbar.close() # --- Validation Phase --- val_loss_avg = tf.keras.metrics.Mean() val_batches_processed = 0 try: val_pbar = tqdm(total=val_steps, desc="Validation", unit="batch") is_tqdm_val = True except ImportError: val_pbar = None is_tqdm_val = False logger.info("Validation progress bar disabled") for q_batch, p_batch, n_batch in val_dataset: val_loss = self.validation_step(q_batch, p_batch, n_batch) val_loss_avg(val_loss) val_batches_processed += 1 if is_tqdm_val: val_pbar.update(1) val_pbar.set_postfix({ "val_loss": f"{val_loss.numpy():.4f}", "batches": f"{val_batches_processed}/{val_steps}" }) # Memory cleanup gc.collect() if val_batches_processed >= val_steps: break if is_tqdm_val and val_pbar: val_pbar.close() # End of epoch: compute final epoch stats, log, and save checkpoint train_loss = epoch_loss_avg.result().numpy() val_loss = val_loss_avg.result().numpy() logger.info(f"Epoch {epoch} Complete: Train Loss={train_loss:.4f}, Val Loss={val_loss:.4f}") # Log epoch metrics with train_summary_writer.as_default(): tf.summary.scalar("epoch_loss", train_loss, step=epoch) with val_summary_writer.as_default(): tf.summary.scalar("val_loss", val_loss, step=epoch) # Save checkpoint manager.save() # Store metrics in history self.history['train_loss'].append(train_loss) self.history['val_loss'].append(val_loss) if use_lr_schedule: current_lr = float(lr_schedule(self.optimizer.iterations)) else: current_lr = float(self.optimizer.learning_rate.numpy()) self.history.setdefault('learning_rate', []).append(current_lr) # Early stopping logic if val_loss < best_val_loss - min_delta: best_val_loss = val_loss epochs_no_improve = 0 logger.info(f"Validation loss improved to {val_loss:.4f}. Reset patience.") else: epochs_no_improve += 1 logger.info(f"No improvement this epoch. Patience: {epochs_no_improve}/{early_stopping_patience}") if epochs_no_improve >= early_stopping_patience: logger.info("Early stopping triggered.") break logger.info("Training completed!") @tf.function def train_step( self, q_batch: tf.Tensor, p_batch: tf.Tensor, n_batch: tf.Tensor ) -> tf.Tensor: """ Single training step using queries, positives, and hard negatives. """ with tf.GradientTape() as tape: # Encode queries q_enc = self.encoder(q_batch, training=True) # [batch_size, embed_dim] # Encode positives p_enc = self.encoder(p_batch, training=True) # [batch_size, embed_dim] # Encode negatives # n_batch: [batch_size, neg_samples, max_length] shape = tf.shape(n_batch) bs = shape[0] neg_samples = shape[1] # Flatten negatives to feed them in one pass: # => [batch_size * neg_samples, max_length] n_batch_flat = tf.reshape(n_batch, [bs * neg_samples, shape[2]]) n_enc_flat = self.encoder(n_batch_flat, training=True) # [bs*neg_samples, embed_dim] # Reshape back => [batch_size, neg_samples, embed_dim] n_enc = tf.reshape(n_enc_flat, [bs, neg_samples, -1]) # Combine the positive embedding and negative embeddings along dim=1 # => shape [batch_size, 1 + neg_samples, embed_dim] # The first column is the positive; subsequent columns are negatives combined_p_n = tf.concat( [tf.expand_dims(p_enc, axis=1), n_enc], axis=1 ) # [bs, (1+neg_samples), embed_dim] # Now compute scores: dot product of q_enc with each column in combined_p_n # We'll use `tf.einsum` to handle the batch dimension properly # dot_products => shape [batch_size, (1+neg_samples)] dot_products = tf.einsum('bd,bkd->bk', q_enc, combined_p_n) # The label for each row is 0 (the first column is the correct/positive) labels = tf.zeros([bs], dtype=tf.int32) # Cross-entropy over the [batch_size, 1+neg_samples] scores loss = tf.nn.sparse_softmax_cross_entropy_with_logits( labels=labels, logits=dot_products ) loss = tf.reduce_mean(loss) # Apply gradients gradients = tape.gradient(loss, self.encoder.trainable_variables) self.optimizer.apply_gradients(zip(gradients, self.encoder.trainable_variables)) return loss @tf.function def validation_step( self, q_batch: tf.Tensor, p_batch: tf.Tensor, n_batch: tf.Tensor ) -> tf.Tensor: """ Single validation step using queries, positives, and hard negatives. """ q_enc = self.encoder(q_batch, training=False) p_enc = self.encoder(p_batch, training=False) shape = tf.shape(n_batch) bs = shape[0] neg_samples = shape[1] n_batch_flat = tf.reshape(n_batch, [bs * neg_samples, shape[2]]) n_enc_flat = self.encoder(n_batch_flat, training=False) n_enc = tf.reshape(n_enc_flat, [bs, neg_samples, -1]) combined_p_n = tf.concat( [tf.expand_dims(p_enc, axis=1), n_enc], axis=1 ) dot_products = tf.einsum('bd,bkd->bk', q_enc, combined_p_n) labels = tf.zeros([bs], dtype=tf.int32) loss = tf.nn.sparse_softmax_cross_entropy_with_logits( labels=labels, logits=dot_products ) loss = tf.reduce_mean(loss) return loss # def train_streaming( # self, # dialogues: List[dict], # epochs: int = 20, # batch_size: int = 16, # validation_split: float = 0.2, # checkpoint_dir: str = "checkpoints/", # use_lr_schedule: bool = True, # peak_lr: float = 2e-5, # warmup_steps_ratio: float = 0.1, # early_stopping_patience: int = 3, # min_delta: float = 1e-4, # neg_samples: int = 1 # ) -> None: # """Streaming training with tf.data pipeline.""" # logger.info("Starting streaming training pipeline with tf.data...") # # Initialize TFDataPipeline (replaces StreamingDataPipeline) # dataset_preparer = TFDataPipeline( # embedding_batch_size=self.config.embedding_batch_size, # tokenizer=self.tokenizer, # encoder=self.encoder, # index=self.index, # Pass CPU version of FAISS index # response_pool=self.response_pool, # max_length=self.config.max_context_token_limit, # neg_samples=neg_samples # ) # # Calculate total steps for learning rate schedule # total_pairs = dataset_preparer.estimate_total_pairs(dialogues) # train_size = int(total_pairs * (1 - validation_split)) # val_size = int(total_pairs * validation_split) # steps_per_epoch = int(math.ceil(train_size / batch_size)) # val_steps = int(math.ceil(val_size / batch_size)) # total_steps = steps_per_epoch * epochs # logger.info(f"Total pairs: {total_pairs}") # logger.info(f"Training pairs: {train_size}") # logger.info(f"Validation pairs: {val_size}") # logger.info(f"Steps per epoch: {steps_per_epoch}") # logger.info(f"Validation steps: {val_steps}") # logger.info(f"Total steps: {total_steps}") # # Set up optimizer with learning rate schedule # if use_lr_schedule: # warmup_steps = int(total_steps * warmup_steps_ratio) # lr_schedule = self._get_lr_schedule( # total_steps=total_steps, # peak_lr=peak_lr, # warmup_steps=warmup_steps # ) # self.optimizer = tf.keras.optimizers.Adam(learning_rate=lr_schedule) # logger.info("Using custom learning rate schedule.") # else: # self.optimizer = tf.keras.optimizers.Adam(learning_rate=peak_lr) # logger.info("Using fixed learning rate.") # # Initialize checkpoint manager # checkpoint = tf.train.Checkpoint(optimizer=self.optimizer, model=self.encoder) # manager = tf.train.CheckpointManager(checkpoint, checkpoint_dir, max_to_keep=3) # # Setup TensorBoard # log_dir = Path(checkpoint_dir) / "tensorboard_logs" # log_dir.mkdir(parents=True, exist_ok=True) # current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # train_log_dir = str(log_dir / f"train_{current_time}") # val_log_dir = str(log_dir / f"val_{current_time}") # train_summary_writer = tf.summary.create_file_writer(train_log_dir) # val_summary_writer = tf.summary.create_file_writer(val_log_dir) # logger.info(f"TensorBoard logs will be saved in {log_dir}") # # Create training and validation datasets # train_dataset = dataset_preparer.get_tf_dataset(dialogues, batch_size).take(train_size) # val_dataset = dataset_preparer.get_tf_dataset(dialogues, batch_size).skip(train_size).take(val_size) # # Training loop # best_val_loss = float("inf") # epochs_no_improve = 0 # for epoch in range(1, epochs + 1): # # --- Training Phase --- # epoch_loss_avg = tf.keras.metrics.Mean() # batches_processed = 0 # try: # train_pbar = tqdm(total=steps_per_epoch, desc=f"Training Epoch {epoch}", unit="batch") # is_tqdm_train = True # except ImportError: # train_pbar = None # is_tqdm_train = False # logger.info("Training progress bar disabled") # for q_batch, p_batch, n_batch in train_dataset: # #p_batch = p_n_batch[:, 0, :] # Extract positive from (positive, negative) pair # loss = self.train_step(q_batch, p_batch, n_batch) # epoch_loss_avg(loss) # batches_processed += 1 # # Log to TensorBoard # with train_summary_writer.as_default(): # tf.summary.scalar("loss", loss, step=(epoch - 1) * steps_per_epoch + batches_processed) # # Update progress bar # if use_lr_schedule: # current_lr = float(lr_schedule(self.optimizer.iterations)) # else: # current_lr = float(self.optimizer.learning_rate.numpy()) # if is_tqdm_train: # train_pbar.update(1) # train_pbar.set_postfix({ # "loss": f"{loss.numpy():.4f}", # "lr": f"{current_lr:.2e}", # "batches": f"{batches_processed}/{steps_per_epoch}" # }) # # Memory cleanup # gc.collect() # if batches_processed >= steps_per_epoch: # break # if is_tqdm_train and train_pbar: # train_pbar.close() # # --- Validation Phase --- # val_loss_avg = tf.keras.metrics.Mean() # val_batches_processed = 0 # try: # val_pbar = tqdm(total=val_steps, desc="Validation", unit="batch") # is_tqdm_val = True # except ImportError: # val_pbar = None # is_tqdm_val = False # logger.info("Validation progress bar disabled") # for q_batch, p_batch, n_batch in val_dataset: # #p_batch = p_n_batch[:, 0, :] # Extract positive from (positive, negative) pair # val_loss = self.validation_step(q_batch, p_batch, n_batch) # val_loss_avg(val_loss) # val_batches_processed += 1 # if is_tqdm_val: # val_pbar.update(1) # val_pbar.set_postfix({ # "val_loss": f"{val_loss.numpy():.4f}", # "batches": f"{val_batches_processed}/{val_steps}" # }) # # Memory cleanup # gc.collect() # if val_batches_processed >= val_steps: # break # if is_tqdm_val and val_pbar: # val_pbar.close() # # End of epoch: compute final epoch stats, log, and save checkpoint # train_loss = epoch_loss_avg.result().numpy() # val_loss = val_loss_avg.result().numpy() # logger.info(f"Epoch {epoch} Complete: Train Loss={train_loss:.4f}, Val Loss={val_loss:.4f}") # # Log epoch metrics # with train_summary_writer.as_default(): # tf.summary.scalar("epoch_loss", train_loss, step=epoch) # with val_summary_writer.as_default(): # tf.summary.scalar("val_loss", val_loss, step=epoch) # # Save checkpoint # manager.save() # # Store metrics in history # self.history['train_loss'].append(train_loss) # self.history['val_loss'].append(val_loss) # if use_lr_schedule: # current_lr = float(lr_schedule(self.optimizer.iterations)) # else: # current_lr = float(self.optimizer.learning_rate.numpy()) # self.history.setdefault('learning_rate', []).append(current_lr) # # Early stopping logic # if val_loss < best_val_loss - min_delta: # best_val_loss = val_loss # epochs_no_improve = 0 # logger.info(f"Validation loss improved to {val_loss:.4f}. Reset patience.") # else: # epochs_no_improve += 1 # logger.info(f"No improvement this epoch. Patience: {epochs_no_improve}/{early_stopping_patience}") # if epochs_no_improve >= early_stopping_patience: # logger.info("Early stopping triggered.") # break # logger.info("Streaming training completed!") # @tf.function # def train_step( # self, # q_batch: tf.Tensor, # p_batch: tf.Tensor, # n_batch: tf.Tensor, # attention_mask: Optional[tf.Tensor] = None # ) -> tf.Tensor: # """ # Single training step that uses queries, positives, and negatives in a # contrastive/InfoNCE style. The label is always 0 (the positive) vs. # the negative alternatives. # """ # with tf.GradientTape() as tape: # # Encode queries # q_enc = self.encoder(q_batch, training=True) # [batch_size, embed_dim] # # Encode positives # p_enc = self.encoder(p_batch, training=True) # [batch_size, embed_dim] # # Encode negatives # # n_batch: [batch_size, neg_samples, max_length] # shape = tf.shape(n_batch) # bs = shape[0] # neg_samples = shape[1] # # Flatten negatives to feed them in one pass: # # => [batch_size * neg_samples, max_length] # n_batch_flat = tf.reshape(n_batch, [bs * neg_samples, shape[2]]) # n_enc_flat = self.encoder(n_batch_flat, training=True) # [bs*neg_samples, embed_dim] # # Reshape back => [batch_size, neg_samples, embed_dim] # n_enc = tf.reshape(n_enc_flat, [bs, neg_samples, -1]) # # Combine the positive embedding and negative embeddings along dim=1 # # => shape [batch_size, 1 + neg_samples, embed_dim] # # The first column is the positive; subsequent columns are negatives # combined_p_n = tf.concat( # [tf.expand_dims(p_enc, axis=1), n_enc], # axis=1 # ) # [bs, (1+neg_samples), embed_dim] # # Now compute scores: dot product of q_enc with each column in combined_p_n # # We'll use `tf.einsum` to handle the batch dimension properly # # dot_products => shape [batch_size, (1+neg_samples)] # dot_products = tf.einsum('bd,bkd->bk', q_enc, combined_p_n) # # The label for each row is 0 (the first column is the correct/positive) # labels = tf.zeros([bs], dtype=tf.int32) # # Cross-entropy over the [batch_size, 1+neg_samples] scores # loss = tf.nn.sparse_softmax_cross_entropy_with_logits( # labels=labels, # logits=dot_products # ) # loss = tf.reduce_mean(loss) # # If there's an attention_mask you want to apply (less common in this scenario), # # you could do something like: # if attention_mask is not None: # loss = loss * attention_mask # loss = tf.reduce_sum(loss) / tf.reduce_sum(attention_mask) # # Apply gradients # gradients = tape.gradient(loss, self.encoder.trainable_variables) # self.optimizer.apply_gradients(zip(gradients, self.encoder.trainable_variables)) # return loss # @tf.function # def validation_step( # self, # q_batch: tf.Tensor, # p_batch: tf.Tensor, # n_batch: tf.Tensor, # attention_mask: Optional[tf.Tensor] = None # ) -> tf.Tensor: # """ # Single validation step with queries, positives, and negatives. # Uses the same loss calculation as train_step, but `training=False`. # """ # q_enc = self.encoder(q_batch, training=False) # p_enc = self.encoder(p_batch, training=False) # shape = tf.shape(n_batch) # bs = shape[0] # neg_samples = shape[1] # n_batch_flat = tf.reshape(n_batch, [bs * neg_samples, shape[2]]) # n_enc_flat = self.encoder(n_batch_flat, training=False) # n_enc = tf.reshape(n_enc_flat, [bs, neg_samples, -1]) # combined_p_n = tf.concat( # [tf.expand_dims(p_enc, axis=1), n_enc], # axis=1 # ) # dot_products = tf.einsum('bd,bkd->bk', q_enc, combined_p_n) # labels = tf.zeros([bs], dtype=tf.int32) # loss = tf.nn.sparse_softmax_cross_entropy_with_logits( # labels=labels, # logits=dot_products # ) # loss = tf.reduce_mean(loss) # if attention_mask is not None: # loss = loss * attention_mask # loss = tf.reduce_sum(loss) / tf.reduce_sum(attention_mask) # return loss def _get_lr_schedule( self, total_steps: int, peak_lr: float, warmup_steps: int ) -> tf.keras.optimizers.schedules.LearningRateSchedule: """Create a custom learning rate schedule with warmup and cosine decay.""" class CustomSchedule(tf.keras.optimizers.schedules.LearningRateSchedule): def __init__( self, total_steps: int, peak_lr: float, warmup_steps: int ): super().__init__() self.total_steps = tf.cast(total_steps, tf.float32) self.peak_lr = tf.cast(peak_lr, tf.float32) # Adjust warmup_steps to not exceed half of total_steps adjusted_warmup_steps = min(warmup_steps, max(1, total_steps // 10)) self.warmup_steps = tf.cast(adjusted_warmup_steps, tf.float32) # Calculate and store constants self.initial_lr = self.peak_lr * 0.1 # Start at 10% of peak self.min_lr = self.peak_lr * 0.01 # Minimum 1% of peak logger.info(f"Learning rate schedule initialized:") logger.info(f" Initial LR: {float(self.initial_lr):.6f}") logger.info(f" Peak LR: {float(self.peak_lr):.6f}") logger.info(f" Min LR: {float(self.min_lr):.6f}") logger.info(f" Warmup steps: {int(self.warmup_steps)}") logger.info(f" Total steps: {int(self.total_steps)}") def __call__(self, step): step = tf.cast(step, tf.float32) # Warmup phase warmup_factor = tf.minimum(1.0, step / self.warmup_steps) warmup_lr = self.initial_lr + (self.peak_lr - self.initial_lr) * warmup_factor # Decay phase decay_steps = tf.maximum(1.0, self.total_steps - self.warmup_steps) decay_factor = (step - self.warmup_steps) / decay_steps decay_factor = tf.minimum(tf.maximum(0.0, decay_factor), 1.0) # Clip to [0,1] cosine_decay = 0.5 * (1.0 + tf.cos(tf.constant(math.pi) * decay_factor)) decay_lr = self.min_lr + (self.peak_lr - self.min_lr) * cosine_decay # Choose between warmup and decay final_lr = tf.where(step < self.warmup_steps, warmup_lr, decay_lr) # Ensure learning rate is valid final_lr = tf.maximum(self.min_lr, final_lr) final_lr = tf.where(tf.math.is_finite(final_lr), final_lr, self.min_lr) return final_lr def get_config(self): return { "total_steps": self.total_steps, "peak_lr": self.peak_lr, "warmup_steps": self.warmup_steps, } return CustomSchedule(total_steps, peak_lr, warmup_steps) def _cosine_similarity(self, emb1: np.ndarray, emb2: np.ndarray) -> np.ndarray: """Compute cosine similarity between two numpy arrays.""" normalized_emb1 = emb1 / np.linalg.norm(emb1, axis=1, keepdims=True) normalized_emb2 = emb2 / np.linalg.norm(emb2, axis=1, keepdims=True) return np.dot(normalized_emb1, normalized_emb2.T) def chat( self, query: str, conversation_history: Optional[List[Tuple[str, str]]] = None, quality_checker: Optional['ResponseQualityChecker'] = None, top_k: int = 5, ) -> Tuple[str, List[Tuple[str, float]], Dict[str, Any]]: """ Example chat method that always uses cross-encoder re-ranking if self.reranker is available. """ @self.run_on_device def get_response(self_arg, query_arg): # Add parameters that match decorator's expectations # 1) Build conversation context string conversation_str = self_arg._build_conversation_context(query_arg, conversation_history) # 2) Retrieve + cross-encoder re-rank results = self_arg.retrieve_responses_cross_encoder( query=conversation_str, top_k=top_k, reranker=self_arg.reranker, summarizer=self_arg.summarizer, summarize_threshold=512 ) # 3) Handle empty or confidence if not results: return ( "I'm sorry, but I couldn't find a relevant response.", [], {} ) if quality_checker: metrics = quality_checker.check_response_quality(query_arg, results) if not metrics.get('is_confident', False): return ( "I need more information to provide a good answer. Could you please clarify?", results, metrics ) return results[0][0], results, metrics return results[0][0], results, {} return get_response(self, query) def _build_conversation_context( self, query: str, conversation_history: Optional[List[Tuple[str, str]]] ) -> str: """Build conversation context with better memory management.""" if not conversation_history: return f"{self.special_tokens['user']} {query}" conversation_parts = [] for user_txt, assistant_txt in conversation_history: conversation_parts.extend([ f"{self.special_tokens['user']} {user_txt}", f"{self.special_tokens['assistant']} {assistant_txt}" ]) conversation_parts.append(f"{self.special_tokens['user']} {query}") return "\n".join(conversation_parts)