Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,451 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
import torch
|
| 3 |
+
import torch.nn as nn
|
| 4 |
+
from transformers import (
|
| 5 |
+
AutoTokenizer, AutoModel, AutoProcessor,
|
| 6 |
+
AutoModelForCausalLM, TrainingArguments, Trainer,
|
| 7 |
+
DataCollatorForLanguageModeling
|
| 8 |
+
)
|
| 9 |
+
from datasets import Dataset, load_dataset, concatenate_datasets
|
| 10 |
+
import json
|
| 11 |
+
import os
|
| 12 |
+
import requests
|
| 13 |
+
from PIL import Image
|
| 14 |
+
import librosa
|
| 15 |
+
import cv2
|
| 16 |
+
import numpy as np
|
| 17 |
+
from pathlib import Path
|
| 18 |
+
import logging
|
| 19 |
+
from typing import Dict, List, Optional, Union
|
| 20 |
+
import time
|
| 21 |
+
from huggingface_hub import HfApi, list_datasets_in_collection
|
| 22 |
+
import tempfile
|
| 23 |
+
import shutil
|
| 24 |
+
|
| 25 |
+
# Configuration du logging
|
| 26 |
+
logging.basicConfig(level=logging.INFO)
|
| 27 |
+
logger = logging.getLogger(__name__)
|
| 28 |
+
|
| 29 |
+
class MultimodalTrainer:
|
| 30 |
+
def __init__(self):
|
| 31 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 32 |
+
self.current_model = None
|
| 33 |
+
self.current_tokenizer = None
|
| 34 |
+
self.current_processor = None
|
| 35 |
+
self.training_data = []
|
| 36 |
+
self.hf_api = HfApi()
|
| 37 |
+
|
| 38 |
+
def load_model(self, model_name: str, model_type: str = "causal"):
|
| 39 |
+
"""Charge un modèle depuis Hugging Face"""
|
| 40 |
+
try:
|
| 41 |
+
logger.info(f"Chargement du modèle: {model_name}")
|
| 42 |
+
|
| 43 |
+
if model_type == "causal":
|
| 44 |
+
self.current_model = AutoModelForCausalLM.from_pretrained(
|
| 45 |
+
model_name,
|
| 46 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
| 47 |
+
device_map="auto" if torch.cuda.is_available() else None,
|
| 48 |
+
trust_remote_code=True
|
| 49 |
+
)
|
| 50 |
+
else:
|
| 51 |
+
self.current_model = AutoModel.from_pretrained(
|
| 52 |
+
model_name,
|
| 53 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
| 54 |
+
device_map="auto" if torch.cuda.is_available() else None,
|
| 55 |
+
trust_remote_code=True
|
| 56 |
+
)
|
| 57 |
+
|
| 58 |
+
# Charge le tokenizer et processor
|
| 59 |
+
try:
|
| 60 |
+
self.current_tokenizer = AutoTokenizer.from_pretrained(
|
| 61 |
+
model_name, trust_remote_code=True
|
| 62 |
+
)
|
| 63 |
+
except:
|
| 64 |
+
logger.warning("Tokenizer non trouvé, utilisation d'un tokenizer par défaut")
|
| 65 |
+
|
| 66 |
+
try:
|
| 67 |
+
self.current_processor = AutoProcessor.from_pretrained(
|
| 68 |
+
model_name, trust_remote_code=True
|
| 69 |
+
)
|
| 70 |
+
except:
|
| 71 |
+
logger.warning("Processor non trouvé")
|
| 72 |
+
|
| 73 |
+
return f"✅ Modèle {model_name} chargé avec succès!"
|
| 74 |
+
|
| 75 |
+
except Exception as e:
|
| 76 |
+
error_msg = f"❌ Erreur lors du chargement: {str(e)}"
|
| 77 |
+
logger.error(error_msg)
|
| 78 |
+
return error_msg
|
| 79 |
+
|
| 80 |
+
def load_collection_datasets(self, collection_url: str):
|
| 81 |
+
"""Charge tous les datasets d'une collection HF"""
|
| 82 |
+
try:
|
| 83 |
+
# Extrait l'ID de la collection depuis l'URL
|
| 84 |
+
collection_id = collection_url.split("/")[-1]
|
| 85 |
+
|
| 86 |
+
# Liste les datasets de la collection
|
| 87 |
+
collection_items = list_datasets_in_collection(collection_id)
|
| 88 |
+
|
| 89 |
+
datasets_info = []
|
| 90 |
+
loaded_datasets = []
|
| 91 |
+
|
| 92 |
+
for item in collection_items:
|
| 93 |
+
try:
|
| 94 |
+
dataset_name = item.id
|
| 95 |
+
dataset = load_dataset(dataset_name, split='train', streaming=False)
|
| 96 |
+
loaded_datasets.append(dataset)
|
| 97 |
+
datasets_info.append(f"✅ {dataset_name}: {len(dataset)} exemples")
|
| 98 |
+
logger.info(f"Dataset chargé: {dataset_name}")
|
| 99 |
+
except Exception as e:
|
| 100 |
+
datasets_info.append(f"❌ {dataset_name}: {str(e)}")
|
| 101 |
+
logger.error(f"Erreur dataset {dataset_name}: {e}")
|
| 102 |
+
|
| 103 |
+
# Combine tous les datasets
|
| 104 |
+
if loaded_datasets:
|
| 105 |
+
combined_dataset = concatenate_datasets(loaded_datasets)
|
| 106 |
+
self.training_data = combined_dataset
|
| 107 |
+
|
| 108 |
+
result = f"📊 Collection chargée!\n" + "\n".join(datasets_info)
|
| 109 |
+
result += f"\n\n🔢 Total combiné: {len(self.training_data)} exemples"
|
| 110 |
+
|
| 111 |
+
return result
|
| 112 |
+
|
| 113 |
+
except Exception as e:
|
| 114 |
+
error_msg = f"❌ Erreur collection: {str(e)}"
|
| 115 |
+
logger.error(error_msg)
|
| 116 |
+
return error_msg
|
| 117 |
+
|
| 118 |
+
def load_single_dataset(self, dataset_name: str, split: str = "train"):
|
| 119 |
+
"""Charge un dataset individuel"""
|
| 120 |
+
try:
|
| 121 |
+
dataset = load_dataset(dataset_name, split=split)
|
| 122 |
+
|
| 123 |
+
if hasattr(self, 'training_data') and self.training_data:
|
| 124 |
+
# Combine avec les données existantes
|
| 125 |
+
self.training_data = concatenate_datasets([self.training_data, dataset])
|
| 126 |
+
else:
|
| 127 |
+
self.training_data = dataset
|
| 128 |
+
|
| 129 |
+
return f"✅ Dataset {dataset_name} ajouté! Total: {len(self.training_data)} exemples"
|
| 130 |
+
|
| 131 |
+
except Exception as e:
|
| 132 |
+
error_msg = f"❌ Erreur dataset: {str(e)}"
|
| 133 |
+
logger.error(error_msg)
|
| 134 |
+
return error_msg
|
| 135 |
+
|
| 136 |
+
def process_multimodal_data(self, example):
|
| 137 |
+
"""Traite les données multimodales pour l'entraînement"""
|
| 138 |
+
processed = {}
|
| 139 |
+
|
| 140 |
+
# Traitement du texte
|
| 141 |
+
if 'text' in example:
|
| 142 |
+
if self.current_tokenizer:
|
| 143 |
+
tokens = self.current_tokenizer(
|
| 144 |
+
example['text'],
|
| 145 |
+
truncation=True,
|
| 146 |
+
padding=True,
|
| 147 |
+
max_length=512,
|
| 148 |
+
return_tensors="pt"
|
| 149 |
+
)
|
| 150 |
+
processed.update(tokens)
|
| 151 |
+
|
| 152 |
+
# Traitement des images
|
| 153 |
+
if 'image' in example:
|
| 154 |
+
try:
|
| 155 |
+
if isinstance(example['image'], str):
|
| 156 |
+
# URL ou chemin
|
| 157 |
+
if example['image'].startswith('http'):
|
| 158 |
+
response = requests.get(example['image'])
|
| 159 |
+
image = Image.open(io.BytesIO(response.content))
|
| 160 |
+
else:
|
| 161 |
+
image = Image.open(example['image'])
|
| 162 |
+
else:
|
| 163 |
+
image = example['image']
|
| 164 |
+
|
| 165 |
+
if self.current_processor:
|
| 166 |
+
image_inputs = self.current_processor(
|
| 167 |
+
images=image, return_tensors="pt"
|
| 168 |
+
)
|
| 169 |
+
processed.update(image_inputs)
|
| 170 |
+
|
| 171 |
+
except Exception as e:
|
| 172 |
+
logger.warning(f"Erreur traitement image: {e}")
|
| 173 |
+
|
| 174 |
+
# Traitement audio
|
| 175 |
+
if 'audio' in example:
|
| 176 |
+
try:
|
| 177 |
+
if isinstance(example['audio'], str):
|
| 178 |
+
audio_data, sr = librosa.load(example['audio'], sr=16000)
|
| 179 |
+
else:
|
| 180 |
+
audio_data = example['audio']
|
| 181 |
+
sr = 16000
|
| 182 |
+
|
| 183 |
+
# Conversion basique pour l'exemple
|
| 184 |
+
processed['audio'] = torch.tensor(audio_data).unsqueeze(0)
|
| 185 |
+
|
| 186 |
+
except Exception as e:
|
| 187 |
+
logger.warning(f"Erreur traitement audio: {e}")
|
| 188 |
+
|
| 189 |
+
return processed
|
| 190 |
+
|
| 191 |
+
def start_training(self,
|
| 192 |
+
output_dir: str,
|
| 193 |
+
num_epochs: int = 3,
|
| 194 |
+
learning_rate: float = 5e-5,
|
| 195 |
+
batch_size: int = 4,
|
| 196 |
+
save_steps: int = 500):
|
| 197 |
+
"""Lance l'entraînement du modèle"""
|
| 198 |
+
|
| 199 |
+
if not self.current_model:
|
| 200 |
+
return "❌ Aucun modèle chargé!"
|
| 201 |
+
|
| 202 |
+
if not self.training_data:
|
| 203 |
+
return "❌ Aucune donnée d'entraînement!"
|
| 204 |
+
|
| 205 |
+
try:
|
| 206 |
+
# Préparation des données
|
| 207 |
+
logger.info("Préparation des données...")
|
| 208 |
+
|
| 209 |
+
# Arguments d'entraînement
|
| 210 |
+
training_args = TrainingArguments(
|
| 211 |
+
output_dir=output_dir,
|
| 212 |
+
num_train_epochs=num_epochs,
|
| 213 |
+
per_device_train_batch_size=batch_size,
|
| 214 |
+
learning_rate=learning_rate,
|
| 215 |
+
logging_steps=50,
|
| 216 |
+
save_steps=save_steps,
|
| 217 |
+
eval_steps=save_steps,
|
| 218 |
+
warmup_steps=100,
|
| 219 |
+
fp16=torch.cuda.is_available(),
|
| 220 |
+
dataloader_num_workers=2,
|
| 221 |
+
remove_unused_columns=False,
|
| 222 |
+
report_to=None # Désactive wandb/tensorboard
|
| 223 |
+
)
|
| 224 |
+
|
| 225 |
+
# Data collator
|
| 226 |
+
data_collator = DataCollatorForLanguageModeling(
|
| 227 |
+
tokenizer=self.current_tokenizer,
|
| 228 |
+
mlm=False
|
| 229 |
+
) if self.current_tokenizer else None
|
| 230 |
+
|
| 231 |
+
# Trainer
|
| 232 |
+
trainer = Trainer(
|
| 233 |
+
model=self.current_model,
|
| 234 |
+
args=training_args,
|
| 235 |
+
train_dataset=self.training_data,
|
| 236 |
+
data_collator=data_collator,
|
| 237 |
+
)
|
| 238 |
+
|
| 239 |
+
# Lance l'entraînement
|
| 240 |
+
logger.info("🚀 Début de l'entraînement...")
|
| 241 |
+
trainer.train()
|
| 242 |
+
|
| 243 |
+
# Sauvegarde
|
| 244 |
+
trainer.save_model()
|
| 245 |
+
if self.current_tokenizer:
|
| 246 |
+
self.current_tokenizer.save_pretrained(output_dir)
|
| 247 |
+
|
| 248 |
+
return f"✅ Entraînement terminé! Modèle sauvegardé dans {output_dir}"
|
| 249 |
+
|
| 250 |
+
except Exception as e:
|
| 251 |
+
error_msg = f"❌ Erreur entraînement: {str(e)}"
|
| 252 |
+
logger.error(error_msg)
|
| 253 |
+
return error_msg
|
| 254 |
+
|
| 255 |
+
def get_model_info(self):
|
| 256 |
+
"""Retourne les informations du modèle actuel"""
|
| 257 |
+
if not self.current_model:
|
| 258 |
+
return "Aucun modèle chargé"
|
| 259 |
+
|
| 260 |
+
info = f"📋 Modèle actuel:\n"
|
| 261 |
+
info += f"Type: {type(self.current_model).__name__}\n"
|
| 262 |
+
info += f"Device: {next(self.current_model.parameters()).device}\n"
|
| 263 |
+
|
| 264 |
+
# Compte les paramètres
|
| 265 |
+
total_params = sum(p.numel() for p in self.current_model.parameters())
|
| 266 |
+
trainable_params = sum(p.numel() for p in self.current_model.parameters() if p.requires_grad)
|
| 267 |
+
|
| 268 |
+
info += f"Paramètres totaux: {total_params:,}\n"
|
| 269 |
+
info += f"Paramètres entraînables: {trainable_params:,}\n"
|
| 270 |
+
|
| 271 |
+
if hasattr(self, 'training_data') and self.training_data:
|
| 272 |
+
info += f"\n📊 Données: {len(self.training_data)} exemples"
|
| 273 |
+
|
| 274 |
+
return info
|
| 275 |
+
|
| 276 |
+
# Initialisation du trainer
|
| 277 |
+
trainer = MultimodalTrainer()
|
| 278 |
+
|
| 279 |
+
# Interface Gradio
|
| 280 |
+
def create_interface():
|
| 281 |
+
with gr.Blocks(title="🔥 Multimodal Training Hub", theme=gr.themes.Soft()) as app:
|
| 282 |
+
|
| 283 |
+
gr.Markdown("""
|
| 284 |
+
# 🔥 Multimodal Training Hub
|
| 285 |
+
### Entraînez vos modèles multimodaux avec facilité!
|
| 286 |
+
|
| 287 |
+
Supporté: Texte 📝 • Images 🖼️ • Audio 🎵 • Vidéo 🎬
|
| 288 |
+
""")
|
| 289 |
+
|
| 290 |
+
with gr.Tab("🤖 Modèle"):
|
| 291 |
+
with gr.Row():
|
| 292 |
+
with gr.Column():
|
| 293 |
+
model_input = gr.Textbox(
|
| 294 |
+
label="Nom du modèle HuggingFace",
|
| 295 |
+
placeholder="kvn420/Tenro_V4.1",
|
| 296 |
+
value="kvn420/Tenro_V4.1"
|
| 297 |
+
)
|
| 298 |
+
model_type = gr.Dropdown(
|
| 299 |
+
label="Type de modèle",
|
| 300 |
+
choices=["causal", "base"],
|
| 301 |
+
value="causal"
|
| 302 |
+
)
|
| 303 |
+
load_model_btn = gr.Button("🔄 Charger le modèle", variant="primary")
|
| 304 |
+
|
| 305 |
+
with gr.Column():
|
| 306 |
+
model_status = gr.Textbox(
|
| 307 |
+
label="Status du modèle",
|
| 308 |
+
interactive=False,
|
| 309 |
+
lines=8
|
| 310 |
+
)
|
| 311 |
+
|
| 312 |
+
load_model_btn.click(
|
| 313 |
+
trainer.load_model,
|
| 314 |
+
inputs=[model_input, model_type],
|
| 315 |
+
outputs=model_status
|
| 316 |
+
)
|
| 317 |
+
|
| 318 |
+
with gr.Tab("📊 Données"):
|
| 319 |
+
with gr.Row():
|
| 320 |
+
with gr.Column():
|
| 321 |
+
gr.Markdown("### 📦 Collection HuggingFace")
|
| 322 |
+
collection_input = gr.Textbox(
|
| 323 |
+
label="URL de la collection",
|
| 324 |
+
placeholder="https://huggingface.co/collections/kvn420/op-67aa4430ba254a4ff0689742"
|
| 325 |
+
)
|
| 326 |
+
load_collection_btn = gr.Button("📥 Charger collection", variant="secondary")
|
| 327 |
+
|
| 328 |
+
gr.Markdown("### 📝 Dataset individuel")
|
| 329 |
+
dataset_input = gr.Textbox(
|
| 330 |
+
label="Nom du dataset",
|
| 331 |
+
placeholder="microsoft/coco"
|
| 332 |
+
)
|
| 333 |
+
dataset_split = gr.Textbox(
|
| 334 |
+
label="Split",
|
| 335 |
+
value="train"
|
| 336 |
+
)
|
| 337 |
+
load_dataset_btn = gr.Button("➕ Ajouter dataset", variant="secondary")
|
| 338 |
+
|
| 339 |
+
with gr.Column():
|
| 340 |
+
data_status = gr.Textbox(
|
| 341 |
+
label="Status des données",
|
| 342 |
+
interactive=False,
|
| 343 |
+
lines=12
|
| 344 |
+
)
|
| 345 |
+
|
| 346 |
+
load_collection_btn.click(
|
| 347 |
+
trainer.load_collection_datasets,
|
| 348 |
+
inputs=collection_input,
|
| 349 |
+
outputs=data_status
|
| 350 |
+
)
|
| 351 |
+
|
| 352 |
+
load_dataset_btn.click(
|
| 353 |
+
trainer.load_single_dataset,
|
| 354 |
+
inputs=[dataset_input, dataset_split],
|
| 355 |
+
outputs=data_status
|
| 356 |
+
)
|
| 357 |
+
|
| 358 |
+
with gr.Tab("🏋️ Entraînement"):
|
| 359 |
+
with gr.Row():
|
| 360 |
+
with gr.Column():
|
| 361 |
+
output_dir = gr.Textbox(
|
| 362 |
+
label="Dossier de sortie",
|
| 363 |
+
value="./trained_model"
|
| 364 |
+
)
|
| 365 |
+
|
| 366 |
+
with gr.Row():
|
| 367 |
+
num_epochs = gr.Number(
|
| 368 |
+
label="Époques",
|
| 369 |
+
value=3,
|
| 370 |
+
minimum=1
|
| 371 |
+
)
|
| 372 |
+
batch_size = gr.Number(
|
| 373 |
+
label="Batch size",
|
| 374 |
+
value=4,
|
| 375 |
+
minimum=1
|
| 376 |
+
)
|
| 377 |
+
|
| 378 |
+
with gr.Row():
|
| 379 |
+
learning_rate = gr.Number(
|
| 380 |
+
label="Learning rate",
|
| 381 |
+
value=5e-5,
|
| 382 |
+
step=1e-6
|
| 383 |
+
)
|
| 384 |
+
save_steps = gr.Number(
|
| 385 |
+
label="Save steps",
|
| 386 |
+
value=500,
|
| 387 |
+
minimum=100
|
| 388 |
+
)
|
| 389 |
+
|
| 390 |
+
train_btn = gr.Button("🚀 Lancer l'entraînement", variant="primary", size="lg")
|
| 391 |
+
|
| 392 |
+
with gr.Column():
|
| 393 |
+
training_status = gr.Textbox(
|
| 394 |
+
label="Status de l'entraînement",
|
| 395 |
+
interactive=False,
|
| 396 |
+
lines=8
|
| 397 |
+
)
|
| 398 |
+
|
| 399 |
+
info_btn = gr.Button("ℹ️ Info modèle")
|
| 400 |
+
model_info = gr.Textbox(
|
| 401 |
+
label="Informations du modèle",
|
| 402 |
+
interactive=False,
|
| 403 |
+
lines=6
|
| 404 |
+
)
|
| 405 |
+
|
| 406 |
+
train_btn.click(
|
| 407 |
+
trainer.start_training,
|
| 408 |
+
inputs=[output_dir, num_epochs, learning_rate, batch_size, save_steps],
|
| 409 |
+
outputs=training_status
|
| 410 |
+
)
|
| 411 |
+
|
| 412 |
+
info_btn.click(
|
| 413 |
+
trainer.get_model_info,
|
| 414 |
+
outputs=model_info
|
| 415 |
+
)
|
| 416 |
+
|
| 417 |
+
with gr.Tab("📚 Aide"):
|
| 418 |
+
gr.Markdown("""
|
| 419 |
+
## 🚀 Guide d'utilisation
|
| 420 |
+
|
| 421 |
+
### 1. Charger un modèle
|
| 422 |
+
- Entrez le nom d'un modèle HuggingFace (ex: `kvn420/Tenro_V4.1`)
|
| 423 |
+
- Choisissez le type (causal pour génération, base pour embedding)
|
| 424 |
+
- Cliquez sur "Charger le modèle"
|
| 425 |
+
|
| 426 |
+
### 2. Ajouter des données
|
| 427 |
+
**Collection:** Chargez tous les datasets d'une collection HF
|
| 428 |
+
**Dataset individuel:** Ajoutez un dataset spécifique
|
| 429 |
+
|
| 430 |
+
### 3. Entraîner
|
| 431 |
+
- Configurez les paramètres d'entraînement
|
| 432 |
+
- Lancez l'entraînement avec "🚀 Lancer l'entraînement"
|
| 433 |
+
|
| 434 |
+
### 📋 Formats supportés
|
| 435 |
+
- **Texte:** Colonnes `text`, `prompt`, `conversation`
|
| 436 |
+
- **Images:** Colonnes `image`, `images` (URLs ou chemins)
|
| 437 |
+
- **Audio:** Colonnes `audio` (fichiers audio)
|
| 438 |
+
- **Vidéo:** Colonnes `video` (fichiers vidéo)
|
| 439 |
+
|
| 440 |
+
### ⚡ Conseils
|
| 441 |
+
- Utilisez un GPU pour l'entraînement (T4, A10G recommandé)
|
| 442 |
+
- Ajustez le batch_size selon votre mémoire GPU
|
| 443 |
+
- Sauvegardez régulièrement avec save_steps
|
| 444 |
+
""")
|
| 445 |
+
|
| 446 |
+
return app
|
| 447 |
+
|
| 448 |
+
# Lancement de l'application
|
| 449 |
+
if __name__ == "__main__":
|
| 450 |
+
app = create_interface()
|
| 451 |
+
app.launch(share=True, server_name="0.0.0.0", server_port=7860)
|