check the code and wait for my question on it, code: import torch import torch.nn as nn…
- User
check the code and wait for my question on it, code: import torch import torch.nn as nn import torch.nn.functional as F import json import math from torch.nn.utils.rnn import pad_sequence from torch.utils.data import DataLoader, Dataset from collections import Counter from tqdm import tqdm import matplotlib.pyplot as plt from sklearn.metrics import precision_score, recall_score, f1_score, accuracy_score from tokenizers import Tokenizer
---------- Device Configuration ----------
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
---------- Utility Functions ----------
def positional_encoding(seq_len, d_model, device): pos = torch.arange(seq_len, dtype=torch.float, device=device).unsqueeze(1) div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model)).to(device) pe = torch.zeros(seq_len, d_model, device=device) pe[:, 0::2] = torch.sin(pos * div_term) pe[:, 1::2] = torch.cos(pos * div_term) return pe.unsqueeze(0)
-------- Performance ----------
def evaluate_model(model, data_loader, device): model.eval() all_preds, all_targets = [], []
with torch.no_grad(): for inputs, targets in data_loader: inputs, targets = inputs.to(device), targets.to(device) outputs = model(inputs) predictions = torch.argmax(outputs, dim=-1).view(-1) # Flatten predicted indices all_preds.extend(predictions.cpu().numpy()) all_targets.extend(targets.view(-1).cpu().numpy()) # Ensure targets are also flattened # Calculate precision, recall, and F1 score after ensuring all_preds and all_targets are correctly aligned. accuracy = accuracy_score(all_targets, all_preds) precision = precision_score(all_targets, all_preds, average='macro', zero_division=0) recall = recall_score(all_targets, all_preds, average='macro', zero_division=0) f1 = f1_score(all_targets, all_preds, average='macro', zero_division=0) print(f"Accuracy: {accuracy:.4f}") print(f"Precision: {precision:.4f}") print(f"Recall: {recall:.4f}") print(f"F1 Score: {f1:.4f}") return accuracy ,precision, recall, f1Function to plot loss over time
def plot_loss(loss_history): plt.figure(figsize=(10, 5)) plt.plot(loss_history, label='Training Loss') plt.xlabel('Batches') plt.ylabel('Loss') plt.title('Training Loss Over Time') plt.legend() plt.show()
---------- Model Definitions ----------
class TransformerExpert(nn.Module): def init(self, input_size, d_model, output_size, nhead, dim_feedforward, num_encoder_layers=1): super(TransformerExpert, self).init() self.d_model = d_model self.input_fc = nn.Linear(input_size, d_model) self.pos_encoder = nn.Parameter(positional_encoding(1, d_model, device), requires_grad=True) encoder_layer = nn.TransformerEncoderLayer(d_model=d_model, nhead=nhead, dim_feedforward=dim_feedforward, batch_first=True, norm_first=True) self.transformer_encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_encoder_layers) self.output_fc = nn.Linear(d_model, output_size) self.norm = nn.LayerNorm(d_model)
def forward(self, x): x = self.norm(self.input_fc(x)) + self.pos_encoder transformer_output = self.transformer_encoder(x) output = self.output_fc(transformer_output) return outputclass GatingNetwork(nn.Module): def init(self, input_feature_dim, num_experts, hidden_dims=[512], dropout_rate=0.1): super(GatingNetwork, self).init() layers = [] last_dim = input_feature_dim
for hidden_dim in hidden_dims: layers.extend([ nn.Linear(last_dim, hidden_dim), nn.GELU(), nn.Dropout(dropout_rate), ]) last_dim = hidden_dim layers.append(nn.Linear(last_dim, num_experts)) self.fc_layers = nn.Sequential(*layers) self.softmax = nn.Softmax(dim=1) def forward(self, x): x = x.mean(dim=1) # To ensure gating is based on overall features across the sequence x = self.fc_layers(x) return self.softmax(x)class MixtureOfTransformerExperts(nn.Module): def init(self, input_size, d_model, output_size, nhead, dim_feedforward, num_experts, num_encoder_layers=1): super(MixtureOfTransformerExperts, self).init() self.num_experts = num_experts self.output_size = output_size self.experts = nn.ModuleList([TransformerExpert(input_size, d_model, output_size, nhead, dim_feedforward, num_encoder_layers) for _ in range(num_experts)]) self.gating_network = GatingNetwork(d_model, num_experts)
def forward(self, x): gating_scores = self.gating_network(x) expert_outputs = [expert(x) for expert in self.experts] stacked_expert_outputs = torch.stack(expert_outputs) expanded_gating_scores = gating_scores.unsqueeze(2).unsqueeze(3) expanded_gating_scores = expanded_gating_scores.expand(-1, -1, x.size(1), self.output_size) expanded_gating_scores = expanded_gating_scores.transpose(0, 1) mixed_output = torch.sum(stacked_expert_outputs * expanded_gating_scores, dim=0) return mixed_outputclass MoETransformerModel(nn.Module): def init(self, vocab_size, d_model, moe): super(MoETransformerModel, self).init() self.embedding = nn.Embedding(num_embeddings=vocab_size, embedding_dim=d_model) self.moe = moe self.dropout = nn.Dropout(p=0.1)
def forward(self, x): embedded = self.dropout(self.embedding(x)) return self.moe(embedded)---------- Dataset Definitions ----------
class QAJsonlDataset(Dataset): def init(self, path, seq_len, tokenizer_path): # Load the trained tokenizer self.tokenizer = Tokenizer.from_file(tokenizer_path)
self.seq_len = seq_len self.pairs = self.load_data(path) # Using BPE, so no need for manual vocab or idx2token. # Tokenization will now happen using self.tokenizer. self.tokenized_pairs = [(self.tokenize(q), self.tokenize(a)) for q, a in self.pairs] def load_data(self, path): pairs = [] with open(path, "r", encoding="utf-8") as f: for line in f: data = json.loads(line.strip()) question, answer = data.get("question", ""), data.get("answer", "") pairs.append((question, answer)) # Store questions and answers as raw strings return pairs def tokenize(self, text): # Tokenizing using the BPE tokenizer encoded = self.tokenizer.encode(text) tokens = encoded.ids # Padding/truncation if len(tokens) < self.seq_len: # Padding tokens += [self.tokenizer.token_to_id("<pad>")] * (self.seq_len - len(tokens)) else: # Truncation tokens = tokens[:self.seq_len - 1] + [self.tokenizer.token_to_id("<eos>")] return tokens def __len__(self): return len(self.tokenized_pairs) def __getitem__(self, idx): tokenized_question, tokenized_answer = self.tokenized_pairs[idx] return torch.tensor(tokenized_question, dtype=torch.long), torch.tensor(tokenized_answer, dtype=torch.long)def collate_fn(batch): questions, answers = zip(*batch) questions = pad_sequence(questions, batch_first=True, padding_value=0) answers = pad_sequence(answers, batch_first=True, padding_value=0) return questions, answers
---------- Training and Inference Functions ----------
def train_model(model, criterion, optimizer, num_epochs, data_loader): model.train() loss_history = [] # Initialize a list to keep track of losses for epoch in range(num_epochs): total_loss = 0 total_items = 0 # Keep track of total items processed progress_bar = tqdm(enumerate(data_loader), total=len(data_loader), desc=f"Epoch {epoch+1}", leave=False) for i, (inputs, targets) in progress_bar: inputs, targets = inputs.to(device), targets.to(device) optimizer.zero_grad()
# Predict predictions = model(inputs) predictions = predictions.view(-1, predictions.size(-1)) # Make sure predictions are the right shape targets = targets.view(-1) # Flatten targets to match prediction shape if necessary # Calculate loss loss = criterion(predictions, targets) loss.backward() # Gradient clipping for stabilization torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0) optimizer.step() scheduler.step() # Update total loss and the number of items total_loss += loss.item() * inputs.size(0) # Multiply loss by batch size total_items += inputs.size(0) loss_history.append(loss.item()) progress_bar.set_postfix({"Loss": loss.item()}) average_loss = total_loss / total_items # Correctly compute average loss print(f"Epoch {epoch+1}, Average Loss: {average_loss:.6f}") return loss_historyclass WarmupLR(torch.optim.lr_scheduler._LRScheduler): def init(self, optimizer, warmup_steps, scheduler_step_lr): self.warmup_steps = warmup_steps self.scheduler_step_lr = scheduler_step_lr # The subsequent scheduler super(WarmupLR, self).init(optimizer)
def get_lr(self): if self._step_count <= self.warmup_steps: warmup_factor = float(self._step_count) / float(max(1, self.warmup_steps)) for base_lr in self.base_lrs: yield base_lr * warmup_factor else: self.scheduler_step_lr.step() # Update the subsequent scheduler for param_group in self.optimizer.param_groups: yield param_group['lr']class GERU(nn.Module): def init(self, in_features): super(GERU, self).init() self.alpha = nn.Parameter(torch.rand(in_features))
def forward(self, x): return torch.max(x, torch.zeros_like(x)) + self.alpha * torch.min(x, torch.zeros_like(x))def generate_text(model, tokenizer, seed_text, num_generate, temperature=1.0): model.eval()
generated_tokens = [] # Encode the seed text using the tokenizer encoded_input = tokenizer.encode(seed_text) input_ids = torch.tensor(encoded_input.ids, dtype=torch.long).unsqueeze(0).to(device) # Generate num_generate tokens with torch.no_grad(): for _ in range(num_generate): output = model(input_ids) # Get the last logits and apply temperature logits = output[:, -1, :] / temperature probabilities = F.softmax(logits, dim=-1) next_token_id = torch.multinomial(probabilities, num_samples=1).item() # Append generated token ID and prepare the new input_ids generated_tokens.append(next_token_id) input_ids = torch.cat([input_ids, torch.tensor([[next_token_id]], dtype=torch.long).to(device)], dim=1) # Decode the generated token IDs back to text generated_text = tokenizer.decode(generated_tokens) return generated_textdef count_tokens_in_dataset(dataset): return sum([len(pair[0]) + len(pair[1]) for pair in dataset.pairs])
def count_parameters(model): return sum(p.numel() for p in model.parameters() if p.requires_grad)
---------- Hyperparameters and Model Instantiation ----------
Transformer :
d_model = 128 nhead = 4 dim_feedforward = 512 num_encoder_layers = 8 num_experts = 1 hidden_dims = [] dropout_rate = 0.3
Gating Network :
gating_network = GatingNetwork( input_feature_dim=d_model, num_experts=num_experts, hidden_dims=hidden_dims, dropout_rate=dropout_rate, )
Training Parameters
batch_size = 32 # Adjustable batch size warmup_steps = 6000 # Warmup steps for learning rate optimizer_type = "AdamW" # Could be “SGD”, “RMSprop”, etc. learning_rate = 9e-2 weight_decay = 1e-5 # For L2 regularization num_epochs = 4
Dataset :
path_to_dataset = "C:/Users/L14/Documents/Projets/Easy-MoE/Easy-MoE/data/Training-math-4K.jsonl" tokenizer_path = "BPE_tokenizer(Training-math-4K).json" seq_len = 32 dataset = QAJsonlDataset(path_to_dataset, seq_len, tokenizer_path) data_loader = DataLoader(dataset, batch_size=batch_size, shuffle=True, collate_fn=collate_fn, pin_memory=True)
num_tokens = count_tokens_in_dataset(dataset) print(f"Total number of tokens in the dataset: {num_tokens}")
Load the tokenizer
tokenizer = Tokenizer.from_file(tokenizer_path)
Determine the vocabulary size
vocab_size = tokenizer.get_vocab_size()
moe = MixtureOfTransformerExperts( input_size=d_model, d_model=d_model, output_size=vocab_size, nhead=nhead, dim_feedforward=dim_feedforward, num_experts=num_experts, num_encoder_layers=num_encoder_layers ).to(device)
moe_transformer_model = MoETransformerModel(vocab_size, d_model, moe).to(device)
Count of total parameters :
total_params = count_parameters(moe_transformer_model) print(f"Total trainable parameters: {total_params}")
---------- Training ----------
Adjusting optimizer setup to include weight decay and allow switching between types
if optimizer_type == "AdamW": optimizer = torch.optim.AdamW(moe_transformer_model.parameters(), lr=learning_rate, weight_decay=weight_decay) elif optimizer_type == "SGD": optimizer = torch.optim.SGD(moe_transformer_model.parameters(), lr=learning_rate, momentum=0.9, weight_decay=weight_decay) elif optimizer_type == "Adam": optimizer = torch.optim.Adam(moe_transformer_model.parameters(), lr=learning_rate, weight_decay=weight_decay)
Instantiate your main scheduler (StepLR)
step_lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.925)
Wrap it with WarmupLR
scheduler = WarmupLR(optimizer, warmup_steps=warmup_steps, scheduler_step_lr=step_lr_scheduler)
criterion = nn.CrossEntropyLoss(label_smoothing=0.1)
Train the model
loss_history = train_model(moe_transformer_model, criterion, optimizer, num_epochs, data_loader)
Evaluating the model
plot_loss(loss_history) train_accuracy = evaluate_model(moe_transformer_model, data_loader, device)
---------- Inference ----------
def interactive_text_generation(model, dataset, max_length=32, temperature=1.0): while True: try: # Get user input seed_text = input("Enter seed text (type 'quit' to exit and save the model): ").strip()
# Check if the user wants to quit the interaction if seed_text.lower() == 'quit': print("Exiting text generation mode.") break # Generate text based on the seed text if seed_text: generated_text = generate_text(model, dataset, seed_text, max_length, temperature) # Modify max_length/temperature as needed print("Generated Text:", generated_text) else: print("Seed text cannot be empty. Please enter some text.") except Exception as e: # Handle potential errors gracefully print(f"An error occurred: {e}. Try again.")interactive_text_generation(moe_transformer_model, tokenizer)
---------- Save Trained Model ----------
torch.save(moe_transformer_model.state_dict(), "Transformer-Alpha-v04.pth")
- Assistant
You've shared a comprehensive code snippet that includes the implementation for training and evaluating a Mixture of Transformer Experts (MoE) model for a question-answering task. This code covers various aspects, including dataset preparation, MoE model definition, training process, evaluation metrics calculation, and model inference with text generation capabilities.
Before proceeding with any specific questions, I've observed that your code combines different advanced concepts in deep learning, particularly in natural language processing (NLP), such as Transformer models, tokenization with Byte Pair Encoding (BPE), MoE architecture for potentially enhancing model capacity while maintaining efficiency, and various utility functions and classes that support these tasks.
The setup includes:
- Device Configuration: Ensures the model runs on GPU if available, leveraging PyTorch's device management.
- Utility Functions: Including positional encoding (for Transformers), loss plotting, performance evaluation metrics calculation (precision, recall, F1 score, accuracy), and text generation based on a seed text.
- Model Definitions: Detailed definitions of
TransformerExpert,GatingNetwork,MixtureOfTransformerExperts, andMoETransformerModel, orchestrated to handle a question-answering task with potential scalability and dynamic routing of inputs through different expert Transformer models based on gating mechanisms. - Dataset Definitions:
QAJsonlDatasetfor loading, tokenizing, and preparing the dataset from a .jsonl format compatible with the training and evaluation pipeline. Tokenization utilizes a pre-trained BPE tokenizer. - Training and Inference: Functions and classes for training the model (
train_model,WarmupLRfor learning rate scheduling) and generating text based on user input, demonstrating the application of the trained model. - Hyperparameters and Model Instantiation: Setup including model dimensions, hyperparameters for training, and instantiation of model components, ready for training.
Your implementation seems complete for executing a training and evaluation pipeline for an NLP task with a MoE architecture. However, managing such complexity requires thorough understanding and careful setting of hyperparameters, model dimensions, etc.
If you have any specific questions or need clarifications regarding parts of this code or its concepts, feel free to ask!
- User
this variable is defined but not being passed to nothing check the code and wait for my question on it, code: import torch import torch.nn as nn import torch.nn.functional as F import json import math from torch.nn.utils.rnn import pad_sequence from torch.utils.data import DataLoader, Dataset from collections import Counter from tqdm import tqdm import matplotlib.pyplot as plt from sklearn.metrics import precision_score, recall_score, f1_score, accuracy_score from tokenizers import Tokenizer
---------- Device Configuration ----------
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
---------- Utility Functions ----------
def positional_encoding(seq_len, d_model, device): pos = torch.arange(seq_len, dtype=torch.float, device=device).unsqueeze(1) div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model)).to(device) pe = torch.zeros(seq_len, d_model, device=device) pe[:, 0::2] = torch.sin(pos * div_term) pe[:, 1::2] = torch.cos(pos * div_term) return pe.unsqueeze(0)
-------- Performance ----------
def evaluate_model(model, data_loader, device): model.eval() all_preds, all_targets = [], []
with torch.no_grad(): for inputs, targets in data_loader: inputs, targets = inputs.to(device), targets.to(device) outputs = model(inputs) predictions = torch.argmax(outputs, dim=-1).view(-1) # Flatten predicted indices all_preds.extend(predictions.cpu().numpy()) all_targets.extend(targets.view(-1).cpu().numpy()) # Ensure targets are also flattened # Calculate precision, recall, and F1 score after ensuring all_preds and all_targets are correctly aligned. accuracy = accuracy_score(all_targets, all_preds) precision = precision_score(all_targets, all_preds, average='macro', zero_division=0) recall = recall_score(all_targets, all_preds, average='macro', zero_division=0) f1 = f1_score(all_targets, all_preds, average='macro', zero_division=0) print(f"Accuracy: {accuracy:.4f}") print(f"Precision: {precision:.4f}") print(f"Recall: {recall:.4f}") print(f"F1 Score: {f1:.4f}") return accuracy ,precision, recall, f1Function to plot loss over time
def plot_loss(loss_history): plt.figure(figsize=(10, 5)) plt.plot(loss_history, label='Training Loss') plt.xlabel('Batches') plt.ylabel('Loss') plt.title('Training Loss Over Time') plt.legend() plt.show()
---------- Model Definitions ----------
class TransformerExpert(nn.Module): def init(self, input_size, d_model, output_size, nhead, dim_feedforward, num_encoder_layers=1): super(TransformerExpert, self).init() self.d_model = d_model self.input_fc = nn.Linear(input_size, d_model) self.pos_encoder = nn.Parameter(positional_encoding(1, d_model, device), requires_grad=True) encoder_layer = nn.TransformerEncoderLayer(d_model=d_model, nhead=nhead, dim_feedforward=dim_feedforward, batch_first=True, norm_first=True) self.transformer_encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_encoder_layers) self.output_fc = nn.Linear(d_model, output_size) self.norm = nn.LayerNorm(d_model)
def forward(self, x): x = self.norm(self.input_fc(x)) + self.pos_encoder transformer_output = self.transformer_encoder(x) output = self.output_fc(transformer_output) return outputclass GatingNetwork(nn.Module): def init(self, input_feature_dim, num_experts, hidden_dims=[512], dropout_rate=0.1): super(GatingNetwork, self).init() layers = [] last_dim = input_feature_dim
for hidden_dim in hidden_dims: layers.extend([ nn.Linear(last_dim, hidden_dim), nn.GELU(), nn.Dropout(dropout_rate), ]) last_dim = hidden_dim layers.append(nn.Linear(last_dim, num_experts)) self.fc_layers = nn.Sequential(*layers) self.softmax = nn.Softmax(dim=1) def forward(self, x): x = x.mean(dim=1) # To ensure gating is based on overall features across the sequence x = self.fc_layers(x) return self.softmax(x)class MixtureOfTransformerExperts(nn.Module): def init(self, input_size, d_model, output_size, nhead, dim_feedforward, num_experts, num_encoder_layers=1): super(MixtureOfTransformerExperts, self).init() self.num_experts = num_experts self.output_size = output_size self.experts = nn.ModuleList([TransformerExpert(input_size, d_model, output_size, nhead, dim_feedforward, num_encoder_layers) for _ in range(num_experts)]) self.gating_network = GatingNetwork(d_model, num_experts)
def forward(self, x): gating_scores = self.gating_network(x) expert_outputs = [expert(x) for expert in self.experts] stacked_expert_outputs = torch.stack(expert_outputs) expanded_gating_scores = gating_scores.unsqueeze(2).unsqueeze(3) expanded_gating_scores = expanded_gating_scores.expand(-1, -1, x.size(1), self.output_size) expanded_gating_scores = expanded_gating_scores.transpose(0, 1) mixed_output = torch.sum(stacked_expert_outputs * expanded_gating_scores, dim=0) return mixed_outputclass MoETransformerModel(nn.Module): def init(self, vocab_size, d_model, moe): super(MoETransformerModel, self).init() self.embedding = nn.Embedding(num_embeddings=vocab_size, embedding_dim=d_model) self.moe = moe self.dropout = nn.Dropout(p=0.1)
def forward(self, x): embedded = self.dropout(self.embedding(x)) return self.moe(embedded)---------- Dataset Definitions ----------
class QAJsonlDataset(Dataset): def init(self, path, seq_len, tokenizer_path): # Load the trained tokenizer self.tokenizer = Tokenizer.from_file(tokenizer_path)
self.seq_len = seq_len self.pairs = self.load_data(path) # Using BPE, so no need for manual vocab or idx2token. # Tokenization will now happen using self.tokenizer. self.tokenized_pairs = [(self.tokenize(q), self.tokenize(a)) for q, a in self.pairs] def load_data(self, path): pairs = [] with open(path, "r", encoding="utf-8") as f: for line in f: data = json.loads(line.strip()) question, answer = data.get("question", ""), data.get("answer", "") pairs.append((question, answer)) # Store questions and answers as raw strings return pairs def tokenize(self, text): # Tokenizing using the BPE tokenizer encoded = self.tokenizer.encode(text) tokens = encoded.ids # Padding/truncation if len(tokens) < self.seq_len: # Padding tokens += [self.tokenizer.token_to_id("<pad>")] * (self.seq_len - len(tokens)) else: # Truncation tokens = tokens[:self.seq_len - 1] + [self.tokenizer.token_to_id("<eos>")] return tokens def __len__(self): return len(self.tokenized_pairs) def __getitem__(self, idx): tokenized_question, tokenized_answer = self.tokenized_pairs[idx] return torch.tensor(tokenized_question, dtype=torch.long), torch.tensor(tokenized_answer, dtype=torch.long)def collate_fn(batch): questions, answers = zip(*batch) questions = pad_sequence(questions, batch_first=True, padding_value=0) answers = pad_sequence(answers, batch_first=True, padding_value=0) return questions, answers
---------- Training and Inference Functions ----------
def train_model(model, criterion, optimizer, num_epochs, data_loader): model.train() loss_history = [] # Initialize a list to keep track of losses for epoch in range(num_epochs): total_loss = 0 total_items = 0 # Keep track of total items processed progress_bar = tqdm(enumerate(data_loader), total=len(data_loader), desc=f"Epoch {epoch+1}", leave=False) for i, (inputs, targets) in progress_bar: inputs, targets = inputs.to(device), targets.to(device) optimizer.zero_grad()
# Predict predictions = model(inputs) predictions = predictions.view(-1, predictions.size(-1)) # Make sure predictions are the right shape targets = targets.view(-1) # Flatten targets to match prediction shape if necessary # Calculate loss loss = criterion(predictions, targets) loss.backward() # Gradient clipping for stabilization torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0) optimizer.step() scheduler.step() # Update total loss and the number of items total_loss += loss.item() * inputs.size(0) # Multiply loss by batch size total_items += inputs.size(0) loss_history.append(loss.item()) progress_bar.set_postfix({"Loss": loss.item()}) average_loss = total_loss / total_items # Correctly compute average loss print(f"Epoch {epoch+1}, Average Loss: {average_loss:.6f}") return loss_historyclass WarmupLR(torch.optim.lr_scheduler._LRScheduler): def init(self, optimizer, warmup_steps, scheduler_step_lr): self.warmup_steps = warmup_steps self.scheduler_step_lr = scheduler_step_lr # The subsequent scheduler super(WarmupLR, self).init(optimizer)
def get_lr(self): if self._step_count <= self.warmup_steps: warmup_factor = float(self._step_count) / float(max(1, self.warmup_steps)) for base_lr in self.base_lrs: yield base_lr * warmup_factor else: self.scheduler_step_lr.step() # Update the subsequent scheduler for param_group in self.optimizer.param_groups: yield param_group['lr']class GERU(nn.Module): def init(self, in_features): super(GERU, self).init() self.alpha = nn.Parameter(torch.rand(in_features))
def forward(self, x): return torch.max(x, torch.zeros_like(x)) + self.alpha * torch.min(x, torch.zeros_like(x))def generate_text(model, tokenizer, seed_text, num_generate, temperature=1.0): model.eval()
generated_tokens = [] # Encode the seed text using the tokenizer encoded_input = tokenizer.encode(seed_text) input_ids = torch.tensor(encoded_input.ids, dtype=torch.long).unsqueeze(0).to(device) # Generate num_generate tokens with torch.no_grad(): for _ in range(num_generate): output = model(input_ids) # Get the last logits and apply temperature logits = output[:, -1, :] / temperature probabilities = F.softmax(logits, dim=-1) next_token_id = torch.multinomial(probabilities, num_samples=1).item() # Append generated token ID and prepare the new input_ids generated_tokens.append(next_token_id) input_ids = torch.cat([input_ids, torch.tensor([[next_token_id]], dtype=torch.long).to(device)], dim=1) # Decode the generated token IDs back to text generated_text = tokenizer.decode(generated_tokens) return generated_textdef count_tokens_in_dataset(dataset): return sum([len(pair[0]) + len(pair[1]) for pair in dataset.pairs])
def count_parameters(model): return sum(p.numel() for p in model.parameters() if p.requires_grad)
---------- Hyperparameters and Model Instantiation ----------
Transformer :
d_model = 128 nhead = 4 dim_feedforward = 512 num_encoder_layers = 8 num_experts = 1 hidden_dims = [] dropout_rate = 0.3
Gating Network :
gating_network = GatingNetwork( input_feature_dim=d_model, num_experts=num_experts, hidden_dims=hidden_dims, dropout_rate=dropout_rate, )
Training Parameters
batch_size = 32 # Adjustable batch size warmup_steps = 6000 # Warmup steps for learning rate optimizer_type = "AdamW" # Could be “SGD”, “RMSprop”, etc. learning_rate = 9e-2 weight_decay = 1e-5 # For L2 regularization num_epochs = 4
Dataset :
path_to_dataset = "C:/Users/L14/Documents/Projets/Easy-MoE/Easy-MoE/data/Training-math-4K.jsonl" tokenizer_path = "BPE_tokenizer(Training-math-4K).json" seq_len = 32 dataset = QAJsonlDataset(path_to_dataset, seq_len, tokenizer_path) data_loader = DataLoader(dataset, batch_size=batch_size, shuffle=True, collate_fn=collate_fn, pin_memory=True)
num_tokens = count_tokens_in_dataset(dataset) print(f"Total number of tokens in the dataset: {num_tokens}")
Load the tokenizer
tokenizer = Tokenizer.from_file(tokenizer_path)
Determine the vocabulary size
vocab_size = tokenizer.get_vocab_size()
moe = MixtureOfTransformerExperts( input_size=d_model, d_model=d_model, output_size=vocab_size, nhead=nhead, dim_feedforward=dim_feedforward, num_experts=num_experts, num_encoder_layers=num_encoder_layers ).to(device)
moe_transformer_model = MoETransformerModel(vocab_size, d_model, moe).to(device)
Count of total parameters :
total_params = count_parameters(moe_transformer_model) print(f"Total trainable parameters: {total_params}")
---------- Training ----------
Adjusting optimizer setup to include weight decay and allow switching between types
if optimizer_type == "AdamW": optimizer = torch.optim.AdamW(moe_transformer_model.parameters(), lr=learning_rate, weight_decay=weight_decay) elif optimizer_type == "SGD": optimizer = torch.optim.SGD(moe_transformer_model.parameters(), lr=learning_rate, momentum=0.9, weight_decay=weight_decay) elif optimizer_type == "Adam": optimizer = torch.optim.Adam(moe_transformer_model.parameters(), lr=learning_rate, weight_decay=weight_decay)
Instantiate your main scheduler (StepLR)
step_lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.925)
Wrap it with WarmupLR
scheduler = WarmupLR(optimizer, warmup_steps=warmup_steps, scheduler_step_lr=step_lr_scheduler)
criterion = nn.CrossEntropyLoss(label_smoothing=0.1)
Train the model
loss_history = train_model(moe_transformer_model, criterion, optimizer, num_epochs, data_loader)
Evaluating the model
plot_loss(loss_history) train_accuracy = evaluate_model(moe_transformer_model, data_loader, device)
---------- Inference ----------
def interactive_text_generation(model, dataset, max_length=32, temperature=1.0): while True: try: # Get user input seed_text = input("Enter seed text (type 'quit' to exit and save the model): ").strip()
# Check if the user wants to quit the interaction if seed_text.lower() == 'quit': print("Exiting text generation mode.") break # Generate text based on the seed text if seed_text: generated_text = generate_text(model, dataset, seed_text, max_length, temperature) # Modify max_length/temperature as needed print("Generated Text:", generated_text) else: print("Seed text cannot be empty. Please enter some text.") except Exception as e: # Handle potential errors gracefully print(f"An error occurred: {e}. Try again.")interactive_text_generation(moe_transformer_model, tokenizer)
---------- Save Trained Model ----------
torch.save(moe_transformer_model.state_dict(), "Transformer-Alpha-v04.pth")
- Assistant
It appears you're indicating a potential issue with a variable or object not being used appropriately after its definition. From reviewing the code, one point of interest that stands out is the definition and parameterization around the learning rate schedulers, specifically the
schedulervariable which is critical for adjusting the learning rate during training:# Instantiate your main scheduler (StepLR) step_lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.925) # Wrap it with WarmupLR scheduler = WarmupLR(optimizer, warmup_steps=warmup_steps, scheduler_step_lr=step_lr_scheduler)While both the
step_lr_schedulerandWarmupLR(scheduler) are initialized, you've correctly incorporated thescheduler.step()within the training loop. This approach encapsulates thestep_lr_scheduler.step()call within the overriddenget_lr()method of theWarmupLRclass, ensuring the step-based learning rate adjustment (after the warmup period) is handed off to thestep_lr_scheduler:def get_lr(self): if self._step_count <= self.warmup_steps: warmup_factor = float(self._step_count) / float(max(1, self.warmup_steps)) for base_lr in self.base_lrs: yield base_lr * warmup_factor else: self.scheduler_step_lr.step() # Update the subsequent scheduler for param_group in self.optimizer.param_groups: yield param_group['lr']However, upon deeper inspection of your concern about a variable defined but not passed to anything, there's a notable point: the
devicevariable is widely used across the script, ensuring tensors are moved to GPU if available. A thorough pass indicates that critical objects and variables are appropriately utilized, such asdevice,scheduler, data loaders, models, and the training loop's parameters.If the concern is regarding another variable or component not mentioned here, or if there's a specific element of the code that seems to have its defined object underutilized, could you please clarify or specify which part of the code you're addressing? This would help in providing a more targeted explanation or solution.
- User
sorry for the previous prompt, here is the variable i was talking about: # Gating Network : gating_network = GatingNetwork( input_feature_dim=d_model, num_experts=num_experts, hidden_dims=hidden_dims, dropout_rate=dropout_rate, )
- Assistant
Thank you for the clarification. Now, focusing on the instantiation of the
GatingNetworkobject:# Gating Network : gating_network = GatingNetwork( input_feature_dim=d_model, num_experts=num_experts, hidden_dims=hidden_dims, dropout_rate=dropout_rate, )After reviewing the provided code, it becomes clear what you're pointing out. The
gating_networkvariable, indeed, is instantiated but not directly used afterward in the script you've shared.In the context of the implementation you've provided, the
GatingNetworkis intended to be a crucial part of theMixtureOfTransformerExpertsmodel. The purpose of theGatingNetworkwithin the Mixture of Experts (MoE) architecture is to dynamically route input data to the most appropriate "expert" based on the input itself. This is especially useful in situations where different parts of the input data might be better handled by different specialized sub-models (or experts).Given that, the
GatingNetworkis indeed being utilized within theMixtureOfTransformerExpertsmodel, implicitly rather than directly in the main body of your script:class MixtureOfTransformerExperts(nn.Module): def __init__(self, input_size, d_model, output_size, nhead, dim_feedforward, num_experts, num_encoder_layers=1): super(MixtureOfTransformerExperts, self).__init__() self.num_experts = num_experts self.output_size = output_size self.experts = nn.ModuleList([TransformerExpert(input_size, d_model, output_size, nhead, dim_feedforward, num_encoder_layers) for _ in range(num_experts)]) self.gating_network = GatingNetwork(d_model, num_experts) # <-- HereIn the
MixtureOfTransformerExpertsclass constructor, a new instance ofGatingNetworkis created and assigned toself.gating_network. This is where the parameters you've defined for thegating_networkvariable would be used if you passed them into theMixtureOfTransformerExpertsinitialization. Instead, you've opted to directly instantiate anotherGatingNetworkwithin theMixtureOfTransformerExpertsclass without passing the previously instantiatedgating_networkvariable. This results in the originalgating_networkinstantiation being redundant since it's not used anywhere else in the code.To resolve this redundancy and efficiently use the instantiated
gating_network, you have two main options:-
Direct Use: Pass the instantiated
gating_networkto wherever it is intended to be used. However, since theMixtureOfTransformerExpertsclass directly instantiates itsGatingNetwork, this would mean adjusting theMixtureOfTransformerExpertsclass to accept aGatingNetworkinstance as a parameter. -
Remove Redundant Instantiation: If there's no explicit need to instantiate the
gating_networkoutside theMixtureOfTransformerExpertsclass, you could simply remove the standalone instantiation to clean up the code and avoid confusion. The code withinMixtureOfTransformerExpertsis sufficient for its purposes as it already creates and utilizes aGatingNetwork.
Unless you plan on using the instantiated
gating_networkindependently of theMixtureOfTransformerExpertsfor other purposes (e.g., in testing or for a different model configuration), the second option might be the more streamlined approach. -
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.