app.py

import gradio as gr
import torch
from torch import nn
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision.transforms import ToTensor
import torch.nn.functional as F

device = (
    "cuda"
    if torch.cuda.is_available()
    else "mps"
    if torch.backends.mps.is_available()
    else "cpu"
)

class CNN(nn.Module):
    def __init__(self):
        super(CNN, self).__init__()
        # Definimos las capas convolucionales
        self.conv1 = nn.Conv2d(in_channels=1, out_channels=32, kernel_size=3, padding=1)
        self.conv2 = nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, padding=1)
        self.conv3 = nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, padding=1)

        # Definimos capas fully connected
        self.fc1 = nn.Linear(128 * 3 * 3, 256)
        self.fc2 = nn.Linear(256, 10)

        # Definimos un max pooling y dropout
        self.pool = nn.MaxPool2d(2, 2)
        self.dropout = nn.Dropout(0.25)

    def forward(self, x):
        # Pasamos las entradas por las capas convolucionales y el max pooling
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = self.pool(F.relu(self.conv3(x)))
        
        # Aplanamos la salida de las capas convolucionales para pasar a fully connected
        x = x.view(-1, 128 * 3 * 3)
        
        # Pasamos por las capas fully connected
        x = F.relu(self.fc1(x))
        x = self.dropout(x)
        x = self.fc2(x)
        
        return x

model = CNN().to(device)

# Cargar el modelo en la CPU
model = CNN().to(device)
model.load_state_dict(torch.load("model_mnist_cnn.pth", map_location=torch.device('cpu')))

def predict(im):

  imagen = np.array(im["composite"])
  imagen = imagen[:,:,3] #accedo al canal que tiene la iamgen

  print(imagen.shape)
  print(imagen.dtype)

  # parte 2
  # Crea la imagen en escala de grises
  imagen_pil = Image.fromarray(imagen, mode='L')
  img_resize = imagen_pil.resize((28, 28))

  # parte 3: A la imagen resize se convierte numpy
  img_np = np.array(img_resize)

  # parte 4: Normalizar
  img_np = img_np.astype(np.float32) / 255.0

  # Parte 4: Tensor
  img_tensor = torch.from_numpy(img_np)

  # Parte 5: Añadimos el canal al tensor
  # img_tensor = img_tensor.unsqueeze(-1)
  img_tensor = img_tensor.unsqueeze(0)

  print(img_tensor.shape)
  print(img_tensor.dtype)

  classes = [
      "Cero",
      "Uno",
      "Dos",
      "Tres",
      "Cuatro",
      "Cinco",
      "Seis",
      "Siete",
      "Ocho",
      "Nueve",
  ]

  model.eval()

  with torch.no_grad():
      # Mover la imagen al dispositivo (GPU o CPU)
      img_tensor = img_tensor.unsqueeze(0).to(device)  # Agregar una dimensión para el batch
      print("Dentro del grad")
      print("Forma de x: ", img_tensor.shape)
      print("Tipo de datos de x: ", img_tensor.dtype)
      print("\n")

      pred = model(img_tensor)
      print("Dentro del model")
      print(pred)
      print("Forma de pred: ", pred.shape)
      print("Tipo de datos de pred: ", pred.dtype)
      print("\n")

      # Obtener las clases predicha y real
      # predicted, actual = classes[pred[0].argmax(0)], classes[y]
      # print(f'Predicted: "{predicted}", Actual: "{actual}"')
      predicted = classes[pred[0].argmax(0)]
      print(f'Predicción: "{predicted}"')

      # Mover la imagen de vuelta a la CPU si está en la GPU y permutar los ejes
      img_tensor = img_tensor.squeeze(0).cpu()  # Eliminar la dimensión del batch
      img_tensor = img_tensor.permute(1, 2, 0)  # Cambiar el orden de los canales para plt.imshow()
      print("Dentro del squeeze")
      print("Forma de x: ", img_tensor.shape)
      print("Tipo de datos de x: ", img_tensor.dtype)
      print("\n")


  # return im["composite"]
  return im["composite"], predicted


with gr.Blocks() as demo:
    descripcion = """
    # MNIST
    Creado por Gabriel Olmos Leiva
    """.strip()
    gr.Markdown(descripcion)
    with gr.Row():
      with gr.Column():
        im = gr.Sketchpad(type="pil", image_mode='RGBA',)
      with gr.Column():
        # gr.CheckboxGroup()
        prediction_text = gr.Textbox(label="Predicción")
        im_preview = gr.Image()
        # gr
    # im.change(predict, outputs=im_preview, inputs=im, show_progress="full", )
    # im.change(predict, outputs=prediction_text, inputs=im, show_progress="full", )
    im.change(predict, outputs=[im_preview, prediction_text], inputs=im, show_progress="full", )
      

demo.launch(share=True, debug=False)