Update app.py
Browse files
app.py
CHANGED
|
@@ -3,43 +3,94 @@ import os
|
|
| 3 |
import gradio as gr
|
| 4 |
from preprocess import process_dataset
|
| 5 |
import subprocess
|
|
|
|
| 6 |
import time
|
| 7 |
|
| 8 |
-
def train_lora_interface(
|
| 9 |
-
if not
|
| 10 |
-
return "❌ Por favor, envie um ZIP
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
"
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
"
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
if hub_token:
|
| 35 |
-
os.environ["HF_TOKEN"] = hub_token
|
| 36 |
-
cmd.append("--push_to_hub")
|
| 37 |
-
cmd.append("--hub_model_id")
|
| 38 |
-
cmd.append("my-lora-model")
|
| 39 |
-
|
| 40 |
-
# Etapa 3: Executa treinamento
|
| 41 |
-
yield "🔥 Treinando modelo... Isso pode levar alguns minutos."
|
| 42 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
process = subprocess.Popen(
|
| 44 |
cmd,
|
| 45 |
stdout=subprocess.PIPE,
|
|
@@ -52,28 +103,39 @@ def train_lora_interface(dataset_zip, model_name, lora_rank, learning_rate, num_
|
|
| 52 |
log_output = ""
|
| 53 |
for line in process.stdout:
|
| 54 |
log_output += line
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
yield f"📝 Treinando...\n{log_output[-500:]}"
|
| 58 |
|
| 59 |
process.wait()
|
| 60 |
|
| 61 |
if process.returncode == 0:
|
| 62 |
-
yield f"🎉
|
| 63 |
else:
|
| 64 |
-
yield f"❌
|
| 65 |
-
|
| 66 |
except Exception as e:
|
| 67 |
-
yield f"💥 Erro
|
| 68 |
|
| 69 |
-
# Interface
|
| 70 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 71 |
gr.Markdown("# 🎨 Treinador de LoRA - Hugging Face")
|
| 72 |
-
gr.Markdown("
|
| 73 |
|
| 74 |
with gr.Row():
|
| 75 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 76 |
|
|
|
|
| 77 |
with gr.Row():
|
| 78 |
model_name = gr.Dropdown(
|
| 79 |
["runwayml/stable-diffusion-v1-5"],
|
|
@@ -87,15 +149,15 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
|
| 87 |
hub_token = gr.Textbox(label="🔐 Token do Hugging Face (opcional)", type="password")
|
| 88 |
|
| 89 |
btn = gr.Button("🚀 Iniciar Treinamento", variant="primary")
|
| 90 |
-
output = gr.Textbox(label="
|
| 91 |
|
| 92 |
btn.click(
|
| 93 |
train_lora_interface,
|
| 94 |
-
inputs=[
|
| 95 |
outputs=output
|
| 96 |
)
|
| 97 |
|
| 98 |
-
# Ativa fila para
|
| 99 |
demo.queue()
|
| 100 |
|
| 101 |
if __name__ == "__main__":
|
|
|
|
| 3 |
import gradio as gr
|
| 4 |
from preprocess import process_dataset
|
| 5 |
import subprocess
|
| 6 |
+
import zipfile
|
| 7 |
import time
|
| 8 |
|
| 9 |
+
def train_lora_interface(dataset_input, input_type, model_name, lora_rank, learning_rate, num_epochs, hub_token):
|
| 10 |
+
if not dataset_input:
|
| 11 |
+
return "❌ Por favor, envie um ZIP ou selecione imagens."
|
| 12 |
+
|
| 13 |
+
yield "📁 Preparando dataset..."
|
| 14 |
+
|
| 15 |
+
# Pasta temporária
|
| 16 |
+
os.makedirs("processed_data", exist_ok=True)
|
| 17 |
+
for f in os.listdir("processed_data"):
|
| 18 |
+
fp = os.path.join("processed_data", f)
|
| 19 |
+
try:
|
| 20 |
+
if os.path.isfile(fp) or os.path.islink(fp):
|
| 21 |
+
os.unlink(fp)
|
| 22 |
+
elif os.path.isdir(fp):
|
| 23 |
+
os.rmtree(fp)
|
| 24 |
+
except Exception as e:
|
| 25 |
+
yield f"⚠️ Erro ao limpar: {e}"
|
| 26 |
+
|
| 27 |
+
dataset_dir = "processed_data"
|
| 28 |
+
|
| 29 |
+
if input_type == "Upload de ZIP":
|
| 30 |
+
zip_path = dataset_input
|
| 31 |
+
if not zipfile.is_zipfile(zip_path):
|
| 32 |
+
yield "❌ Arquivo enviado não é um ZIP válido."
|
| 33 |
+
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
|
| 35 |
+
# Descompacta
|
| 36 |
+
with zipfile.ZipFile(zip_path, 'r') as z:
|
| 37 |
+
z.extractall(dataset_dir)
|
| 38 |
+
yield f"✅ ZIP descompactado! {len(z.namelist())} arquivos extraídos."
|
| 39 |
+
|
| 40 |
+
else: # Múltiplas imagens
|
| 41 |
+
image_count = 0
|
| 42 |
+
for img_path in dataset_input:
|
| 43 |
+
dest_path = os.path.join(dataset_dir, os.path.basename(img_path.name))
|
| 44 |
+
os.rename(img_path.name, dest_path) # Move para processed_data
|
| 45 |
+
|
| 46 |
+
# Cria .txt vazio (BLIP vai preencher depois)
|
| 47 |
+
txt_path = os.path.splitext(dest_path)[0] + ".txt"
|
| 48 |
+
if not os.path.exists(txt_path):
|
| 49 |
+
with open(txt_path, "w") as f:
|
| 50 |
+
f.write("person")
|
| 51 |
+
image_count += 1
|
| 52 |
+
yield f"✅ {image_count} imagens copiadas para o dataset."
|
| 53 |
+
|
| 54 |
+
# Conta imagens processadas
|
| 55 |
+
image_files = [f for f in os.listdir(dataset_dir) if f.lower().endswith(('.png', '.jpg', '.jpeg'))]
|
| 56 |
+
if len(image_files) == 0:
|
| 57 |
+
yield "❌ Nenhuma imagem encontrada. Envie JPG, PNG ou ZIP com imagens."
|
| 58 |
+
return
|
| 59 |
+
|
| 60 |
+
yield f"🖼️ Dataset pronto com {len(image_files)} imagens. Iniciando pré-processamento..."
|
| 61 |
+
|
| 62 |
+
# Gera captions com BLIP (reutiliza preprocess.py)
|
| 63 |
+
try:
|
| 64 |
+
from preprocess import process_dataset as run_blip
|
| 65 |
+
run_blip(None, dataset_dir, generate_captions=True) # Já temos as imagens, só gera captions
|
| 66 |
+
except Exception as e:
|
| 67 |
+
yield f"⚠️ Falha ao gerar legendas: {str(e)}. Continuando com captions existentes."
|
| 68 |
+
|
| 69 |
+
yield "🔥 Iniciando treinamento LoRA..."
|
| 70 |
+
|
| 71 |
+
# Comando de treinamento
|
| 72 |
+
output_dir = "lora-output"
|
| 73 |
+
os.makedirs(output_dir, exist_ok=True)
|
| 74 |
+
|
| 75 |
+
cmd = [
|
| 76 |
+
"python", "train_lora.py",
|
| 77 |
+
"--dataset_dir", dataset_dir,
|
| 78 |
+
"--model_name", model_name,
|
| 79 |
+
"--lora_rank", str(lora_rank),
|
| 80 |
+
"--learning_rate", str(learning_rate),
|
| 81 |
+
"--num_epochs", str(num_epochs),
|
| 82 |
+
"--batch_size", "1",
|
| 83 |
+
"--output_dir", output_dir
|
| 84 |
+
]
|
| 85 |
+
|
| 86 |
+
if hub_token:
|
| 87 |
+
os.environ["HF_TOKEN"] = hub_token
|
| 88 |
+
cmd.append("--push_to_hub")
|
| 89 |
+
cmd.append("--hub_model_id")
|
| 90 |
+
cmd.append("my-lora-model")
|
| 91 |
+
|
| 92 |
+
# Executa treinamento
|
| 93 |
+
try:
|
| 94 |
process = subprocess.Popen(
|
| 95 |
cmd,
|
| 96 |
stdout=subprocess.PIPE,
|
|
|
|
| 103 |
log_output = ""
|
| 104 |
for line in process.stdout:
|
| 105 |
log_output += line
|
| 106 |
+
if "loss" in line.lower() or "epoch" in line.lower():
|
| 107 |
+
yield f"📊 {line.strip()}"
|
|
|
|
| 108 |
|
| 109 |
process.wait()
|
| 110 |
|
| 111 |
if process.returncode == 0:
|
| 112 |
+
yield f"🎉 SUCESSO! Modelo LoRA treinado e salvo em `{output_dir}`"
|
| 113 |
else:
|
| 114 |
+
yield f"❌ Treinamento falhou com código {process.returncode}.\nLogs:\n{log_output[-1000:]}"
|
| 115 |
+
|
| 116 |
except Exception as e:
|
| 117 |
+
yield f"💥 Erro ao executar treinamento: {str(e)}"
|
| 118 |
|
| 119 |
+
# Interface com opção de entrada
|
| 120 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 121 |
gr.Markdown("# 🎨 Treinador de LoRA - Hugging Face")
|
| 122 |
+
gr.Markdown("Envie suas imagens via ZIP ou múltiplos arquivos!")
|
| 123 |
|
| 124 |
with gr.Row():
|
| 125 |
+
input_type = gr.Radio(
|
| 126 |
+
["Upload de ZIP", "Selecionar várias imagens"],
|
| 127 |
+
label="Tipo de Entrada",
|
| 128 |
+
value="Upload de ZIP"
|
| 129 |
+
)
|
| 130 |
+
|
| 131 |
+
with gr.Row():
|
| 132 |
+
dataset_input = gr.File(
|
| 133 |
+
label="📤 Envie seu ZIP ou imagens",
|
| 134 |
+
file_types=[".zip", ".jpg", ".jpeg", ".png"],
|
| 135 |
+
file_count="multiple" # Permite múltiplos arquivos!
|
| 136 |
+
)
|
| 137 |
|
| 138 |
+
gr.Markdown("### ⚙️ Configurações do Treinamento")
|
| 139 |
with gr.Row():
|
| 140 |
model_name = gr.Dropdown(
|
| 141 |
["runwayml/stable-diffusion-v1-5"],
|
|
|
|
| 149 |
hub_token = gr.Textbox(label="🔐 Token do Hugging Face (opcional)", type="password")
|
| 150 |
|
| 151 |
btn = gr.Button("🚀 Iniciar Treinamento", variant="primary")
|
| 152 |
+
output = gr.Textbox(label="📦 Logs e Status", lines=12)
|
| 153 |
|
| 154 |
btn.click(
|
| 155 |
train_lora_interface,
|
| 156 |
+
inputs=[dataset_input, input_type, model_name, lora_rank, learning_rate, num_epochs, hub_token],
|
| 157 |
outputs=output
|
| 158 |
)
|
| 159 |
|
| 160 |
+
# Ativa fila para yield
|
| 161 |
demo.queue()
|
| 162 |
|
| 163 |
if __name__ == "__main__":
|