PTHFileTrainer / app.py
WillemVH's picture
Update app.py
cf72181 verified
import gradio as gr
import json
import os
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
import pickle
from pathlib import Path
from datetime import datetime
import threading
import glob
from collections import Counter
import struct
class SimpleTokenizer:
"""A simple tokenizer for faster startup"""
def __init__(self):
self.vocab = {}
self.inverse_vocab = {}
self.vocab_size = 0
self.pad_token = "<pad>"
self.pad_token_id = 0
self.eos_token = "<eos>"
self.eos_token_id = 1
self.unk_token = "<unk>"
self.unk_token_id = 2
# Start with basic tokens
self.add_token(self.pad_token) # ID 0
self.add_token(self.eos_token) # ID 1
self.add_token(self.unk_token) # ID 2
def add_token(self, token):
if token not in self.vocab:
self.vocab[token] = self.vocab_size
self.inverse_vocab[self.vocab_size] = token
self.vocab_size += 1
return True
return False
def build_vocab_from_texts(self, texts, max_vocab_size=10000):
"""Build vocabulary from all training texts"""
print("Building vocabulary from training data...")
# Count all tokens
token_counter = Counter()
for text in texts:
tokens = text.split()
token_counter.update(tokens)
# Add most frequent tokens to vocabulary
for token, _ in token_counter.most_common(max_vocab_size - self.vocab_size):
self.add_token(token)
print(f"Vocabulary built with {self.vocab_size} tokens")
def tokenize(self, text):
# Simple word-level tokenization
tokens = text.split()
token_ids = []
for token in tokens:
if token in self.vocab:
token_ids.append(self.vocab[token])
else:
token_ids.append(self.unk_token_id) # Use UNK token for out-of-vocab words
return token_ids
def encode(self, text, max_length=None, padding=False, truncation=False):
token_ids = self.tokenize(text)
if truncation and max_length and len(token_ids) > max_length:
token_ids = token_ids[:max_length]
if padding and max_length and len(token_ids) < max_length:
token_ids = token_ids + [self.pad_token_id] * (max_length - len(token_ids))
return token_ids
def decode(self, token_ids):
# Remove padding tokens for cleaner output
filtered_ids = [id for id in token_ids if id != self.pad_token_id]
return " ".join([self.inverse_vocab.get(id, self.unk_token) for id in filtered_ids])
class TextDataset(Dataset):
def __init__(self, texts, tokenizer, max_length=512):
self.tokenizer = tokenizer
self.texts = texts
self.max_length = max_length
# Filter out empty texts
self.texts = [text for text in texts if text.strip()]
def __len__(self):
return len(self.texts)
def __getitem__(self, idx):
text = self.texts[idx]
# Ensure text is not empty
if not text.strip():
text = " " # Use space for empty text
token_ids = self.tokenizer.encode(
text,
max_length=self.max_length,
padding=True,
truncation=True
)
# Convert to tensor and ensure all IDs are within valid range
token_ids = [min(id, self.tokenizer.vocab_size - 1) for id in token_ids]
return {
'input_ids': torch.tensor(token_ids, dtype=torch.long),
'labels': torch.tensor(token_ids, dtype=torch.long)
}
class SimpleGPT(nn.Module):
"""A simplified GPT-like model for faster training"""
def __init__(self, vocab_size, d_model=512, n_layers=6, n_heads=8, max_seq_len=512):
super().__init__()
self.d_model = d_model
self.vocab_size = vocab_size
self.max_seq_len = max_seq_len
# Token and position embeddings
self.token_embedding = nn.Embedding(vocab_size, d_model, padding_idx=0) # padding_idx=0 for pad token
self.position_embedding = nn.Embedding(max_seq_len, d_model)
# Transformer layers
encoder_layer = nn.TransformerEncoderLayer(
d_model=d_model,
nhead=n_heads,
dim_feedforward=d_model * 4,
batch_first=True,
dropout=0.1
)
self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=n_layers)
# Output layer with dropout for regularization
self.dropout = nn.Dropout(0.1)
self.output_layer = nn.Linear(d_model, vocab_size)
# Initialize weights properly
self.apply(self._init_weights)
def _init_weights(self, module):
if isinstance(module, nn.Linear):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
torch.nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
def forward(self, input_ids, labels=None):
batch_size, seq_len = input_ids.shape
# Ensure all token IDs are within valid range
input_ids = torch.clamp(input_ids, 0, self.vocab_size - 1)
# Create token embeddings
token_embeds = self.token_embedding(input_ids)
# Create position embeddings
positions = torch.arange(seq_len, device=input_ids.device).unsqueeze(0).expand(batch_size, seq_len)
position_embeds = self.position_embedding(positions)
# Combine embeddings
x = token_embeds + position_embeds
# Create attention mask (ignore padding tokens)
attention_mask = (input_ids != 0).float()
# Transformer with attention mask
x = self.transformer(x, src_key_padding_mask=attention_mask == 0)
# Apply dropout
x = self.dropout(x)
# Output
logits = self.output_layer(x)
# Calculate loss if labels provided
loss = None
if labels is not None:
# Ensure labels are within valid range
labels = torch.clamp(labels, 0, self.vocab_size - 1)
# Create loss mask to ignore padding tokens
loss_mask = (labels != 0).float()
loss_fn = nn.CrossEntropyLoss(ignore_index=0, reduction='none') # ignore padding
losses = loss_fn(logits.view(-1, self.vocab_size), labels.view(-1))
loss = (losses * loss_mask.view(-1)).sum() / loss_mask.sum()
return {'logits': logits, 'loss': loss}
class AITrainerApp:
def __init__(self):
# Use simple tokenizer for faster startup
self.tokenizer = SimpleTokenizer()
self.model = None
self.training_data = []
# Default model configuration
self.model_config = {
"d_model": 512,
"n_layers": 6,
"n_heads": 8,
"max_seq_len": 512
}
# Training control
self.training_thread = None
self.stop_training_flag = False
self.training_status = "Ready - Load training data to begin"
self.output_log = "Training output will appear here...\n"
def get_device(self, device_type="auto"):
"""Get the selected device based on user choice"""
if device_type == "auto":
return torch.device('cuda' if torch.cuda.is_available() else 'cpu')
elif device_type == "cuda":
if torch.cuda.is_available():
return torch.device('cuda')
else:
return torch.device('cpu')
else:
return torch.device('cpu')
def log_output(self, message):
"""Add message to output log"""
self.output_log += message + "\n"
return self.output_log
def verify_model_file(self, file_path):
"""Verify if a model file is valid before loading"""
try:
# Simple file checks
if not os.path.exists(file_path):
return False, "File does not exist"
if os.path.getsize(file_path) < 1024: # Less than 1KB
return False, "File is too small to be a valid model"
return True, "File appears valid"
except Exception as e:
return False, f"Error verifying file: {str(e)}"
def load_training_files(self, files):
"""Load training files from provided file objects"""
if not files:
return "No files selected", self.output_log
total_texts = []
for file_info in files:
try:
# Read the content from the file object
content = file_info.read().decode('utf-8')
# Split into smaller chunks if needed
chunks = self.split_into_chunks(content, 1000)
total_texts.extend(chunks)
self.output_log = self.log_output(f"Loaded {len(chunks)} chunks from {file_info.name}")
except Exception as e:
error_msg = f"Error reading {file_info.name}: {str(e)}"
self.output_log = self.log_output(error_msg)
return error_msg, self.output_log
self.training_data.extend(total_texts)
# Build vocabulary from all training texts
self.tokenizer.build_vocab_from_texts(self.training_data, max_vocab_size=10000)
status_msg = f"Loaded {len(total_texts)} text chunks from {len(files)} files"
self.output_log = self.log_output(status_msg)
self.output_log = self.log_output(f"Vocabulary size: {self.tokenizer.vocab_size}")
return status_msg, self.output_log
def split_into_chunks(self, text, chunk_size):
words = text.split()
chunks = []
for i in range(0, len(words), chunk_size):
chunk = ' '.join(words[i:i+chunk_size])
chunks.append(chunk)
return chunks
def view_training_data(self):
if not self.training_data:
return "No training data loaded"
preview = ""
for i, text in enumerate(self.training_data[:50]): # Show first 50 chunks
preview += f"Chunk {i+1}:\n{text}\n\n{'='*50}\n\n"
return preview
def start_training(self, d_model, n_layers, n_heads, batch_size, learning_rate, epochs, device_type):
if not self.training_data:
error_msg = "Error: No training data loaded!"
self.output_log = self.log_output(error_msg)
return error_msg, self.output_log, gr.update(interactive=True)
self.stop_training_flag = False
self.training_status = "Training started..."
self.output_log = self.log_output("Starting training...")
# Update model config from UI
self.model_config.update({
"d_model": int(d_model),
"n_layers": int(n_layers),
"n_heads": int(n_heads)
})
# Start training in separate thread
self.training_thread = threading.Thread(
target=self.train_model,
args=(int(batch_size), float(learning_rate), int(epochs), device_type)
)
self.training_thread.daemon = True
self.training_thread.start()
return "Training started...", self.output_log, gr.update(interactive=False)
def stop_training(self):
self.stop_training_flag = True
self.training_status = "Stopping training..."
self.output_log = self.log_output("Stopping training...")
return "Stopping training...", self.output_log, gr.update(interactive=True)
def train_model(self, batch_size, learning_rate, epochs, device_type):
try:
# Create dataset and dataloader
dataset = TextDataset(self.training_data, self.tokenizer)
dataloader = DataLoader(
dataset,
batch_size=batch_size,
shuffle=True
)
# Initialize model
self.model = SimpleGPT(
vocab_size=self.tokenizer.vocab_size,
d_model=self.model_config["d_model"],
n_layers=self.model_config["n_layers"],
n_heads=self.model_config["n_heads"],
max_seq_len=self.model_config["max_seq_len"]
)
# Setup optimizer
optimizer = optim.AdamW(
self.model.parameters(),
lr=learning_rate
)
# Training loop
device = self.get_device(device_type)
self.model.to(device)
self.output_log = self.log_output(f"Using device: {device}")
for epoch in range(epochs):
if self.stop_training_flag:
break
self.model.train()
total_loss = 0
total_batches = 0
for batch_idx, batch in enumerate(dataloader):
if self.stop_training_flag:
break
optimizer.zero_grad()
input_ids = batch['input_ids'].to(device)
labels = batch['labels'].to(device)
# Debug: Check for invalid token IDs
max_id = input_ids.max().item()
if max_id >= self.tokenizer.vocab_size:
self.output_log = self.log_output(f"Warning: Found token ID {max_id} but vocab size is {self.tokenizer.vocab_size}")
# Clamp values to valid range
input_ids = torch.clamp(input_ids, 0, self.tokenizer.vocab_size - 1)
labels = torch.clamp(labels, 0, self.tokenizer.vocab_size - 1)
outputs = self.model(input_ids=input_ids, labels=labels)
loss = outputs['loss']
if torch.isnan(loss) or torch.isinf(loss):
self.output_log = self.log_output("Warning: NaN or Inf loss detected, skipping batch")
continue
loss.backward()
# Gradient clipping to prevent explosions
torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
optimizer.step()
total_loss += loss.item()
total_batches += 1
if batch_idx % 10 == 0:
status_msg = f"Epoch {epoch+1}/{epochs}, Batch {batch_idx}/{len(dataloader)}, Loss: {loss.item():.4f}"
self.training_status = status_msg
if batch_idx % 50 == 0: # Log less frequently to avoid UI slowdown
self.output_log = self.log_output(status_msg)
if total_batches > 0:
avg_loss = total_loss / total_batches
epoch_msg = f"Epoch {epoch+1} completed. Average Loss: {avg_loss:.4f}"
self.training_status = epoch_msg
self.output_log = self.log_output(epoch_msg)
if not self.stop_training_flag:
completion_msg = "Training completed successfully!"
self.training_status = completion_msg
self.output_log = self.log_output(completion_msg)
except Exception as e:
error_msg = f"Training error: {str(e)}"
self.training_status = error_msg
self.output_log = self.log_output(error_msg)
import traceback
self.output_log = self.log_output(traceback.format_exc())
finally:
self.stop_training_flag = False
# Re-enable the start training button
return gr.update(interactive=True)
def save_model(self, file_path):
if self.model is None:
self.output_log = self.log_output("Error: No model to save!")
return "Error: No model to save!", self.output_log
try:
torch.save({
'model_state_dict': self.model.state_dict(),
'tokenizer': self.tokenizer,
'config': self.model_config,
'training_data_info': {
'num_chunks': len(self.training_data),
'vocab_size': self.tokenizer.vocab_size
}
}, file_path)
success_msg = f"Model saved to {file_path}"
self.training_status = success_msg
self.output_log = self.log_output(success_msg)
return success_msg, self.output_log
except Exception as e:
error_msg = f"Error saving model: {str(e)}"
self.output_log = self.log_output(error_msg)
return error_msg, self.output_log
def load_model(self, file_path):
if not file_path:
return "No file selected", self.output_log
try:
checkpoint = torch.load(file_path, map_location='cpu')
# Recreate the model architecture
self.model_config = checkpoint['config']
self.model = SimpleGPT(
vocab_size=checkpoint['tokenizer'].vocab_size,
d_model=self.model_config["d_model"],
n_layers=self.model_config["n_layers"],
n_heads=self.model_config["n_heads"],
max_seq_len=self.model_config["max_seq_len"]
)
# Load weights
self.model.load_state_dict(checkpoint['model_state_dict'])
# Load tokenizer
self.tokenizer = checkpoint['tokenizer']
success_msg = f"Model loaded from {file_path}"
self.training_status = success_msg
self.output_log = self.log_output(success_msg)
return success_msg, self.output_log, str(self.model_config['d_model']), str(self.model_config['n_layers']), str(self.model_config['n_heads'])
except Exception as e:
error_msg = f"Error loading model: {str(e)}"
self.output_log = self.log_output(error_msg)
return error_msg, self.output_log, gr.update(), gr.update(), gr.update()
# Create the app instance
app = AITrainerApp()
# Create Gradio interface
with gr.Blocks(title="AI Text Generation Trainer") as demo:
gr.Markdown("# AI Text Generation Trainer")
with gr.Row():
with gr.Column(scale=1):
gr.Markdown("## Controls")
# Data management
gr.Markdown("### Data Management")
file_input = gr.File(file_count="multiple", label="Training Files")
load_btn = gr.Button("Load Text Files")
view_data_btn = gr.Button("View Training Data")
data_preview = gr.Textbox(label="Training Data Preview", lines=10, interactive=False)
# Device selection
gr.Markdown("### Device Selection")
device_type = gr.Radio(
choices=["auto", "cpu", "cuda"],
value="auto",
label="Processing Device"
)
device_info = gr.Textbox(
label="Device Info",
value=f"GPU available: {'Yes' if torch.cuda.is_available() else 'No'}",
interactive=False
)
# Model configuration
gr.Markdown("### Model Configuration")
d_model = gr.Number(value=512, label="Embedding Size")
n_layers = gr.Number(value=6, label="Number of Layers")
n_heads = gr.Number(value=8, label="Number of Heads")
# Training parameters
gr.Markdown("### Training Parameters")
batch_size = gr.Number(value=4, label="Batch Size")
learning_rate = gr.Number(value=0.001, label="Learning Rate")
epochs = gr.Number(value=3, label="Epochs")
# Training controls
gr.Markdown("### Training Control")
start_btn = gr.Button("Start Training", variant="primary")
stop_btn = gr.Button("Stop Training")
# Export buttons
gr.Markdown("### Export Model")
save_path = gr.Textbox(label="Save Path", value="model.pth")
save_btn = gr.Button("Save Model")
load_path = gr.Textbox(label="Load Path", value="model.pth")
load_btn = gr.Button("Load Model")
with gr.Column(scale=2):
gr.Markdown("## Status & Output")
status = gr.Textbox(label="Status", value=app.training_status, interactive=False)
output = gr.Textbox(label="Output Log", value=app.output_log, lines=20, interactive=False)
# Define event handlers
load_btn.click(
app.load_training_files,
inputs=[file_input],
outputs=[status, output]
)
view_data_btn.click(
app.view_training_data,
inputs=[],
outputs=[data_preview]
)
start_btn.click(
app.start_training,
inputs=[d_model, n_layers, n_heads, batch_size, learning_rate, epochs, device_type],
outputs=[status, output, start_btn]
)
stop_btn.click(
app.stop_training,
inputs=[],
outputs=[status, output, start_btn]
)
save_btn.click(
app.save_model,
inputs=[save_path],
outputs=[status, output]
)
load_btn.click(
app.load_model,
inputs=[load_path],
outputs=[status, output, d_model, n_layers, n_heads]
)
if __name__ == "__main__":
demo.launch()