app.py

import gradio as gr
from transformers import TimeSeriesTransformerForPrediction, TimeSeriesTransformerConfig
import torch
import numpy as np

# Carregar configuração
config = TimeSeriesTransformerConfig.from_pretrained("google/timesfm-2.0-500m-pytorch")
config.prediction_length = 3
config.context_length = 20  # Aumentado para acomodar os lags
config.lags_sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]  # Lags menores que context_length

# Carregar modelo com a configuração ajustada
model = TimeSeriesTransformerForPrediction.from_pretrained(
    "google/timesfm-2.0-500m-pytorch",
    config=config,
    torch_dtype="auto"
)

def prever_vendas(historico):
    # Converter entrada em tensor
    historico = [float(x) for x in historico.split(",") if x.strip()]
    if len(historico) != config.context_length:
        raise ValueError(f"Histórico deve ter {config.context_length} valores.")
    
    # Formatar dados
    inputs = torch.tensor(historico).unsqueeze(0)
    
    # Adicionar parâmetros necessários
    past_time_features = torch.zeros(1, config.context_length, 1)  # Características temporais dummy
    past_observed_mask = torch.ones(1, config.context_length)      # Dados observados
    
    # Gerar previsão
    with torch.no_grad():
        outputs = model(
            inputs,
            past_time_features=past_time_features,
            past_observed_mask=past_observed_mask
        )
        forecast = outputs.mean.squeeze().tolist()
    
    return np.round(forecast, 2)

# Interface Gradio
iface = gr.Interface(
    fn=prever_vendas,
    inputs=gr.Textbox(label=f"Histórico de Vendas ({config.context_length} meses, separados por vírgulas)"),
    outputs=gr.Textbox(label=f"Previsão para os Próximos {config.prediction_length} Meses"),
    examples=[
        ["140,155,160,145,150,165,170,160,175,160,155,170,180,190,200,210,220,230,240,250"],  # 20 meses
    ],
    cache_examples=False  # Desativar cache para evitar erros
)

iface.launch()