app.py

from huggingface_hub import from_pretrained_fastai
import gradio as gr
from fastai.vision.all import *
import PIL
import torchvision.transforms as transforms
# from albumentations import (
#     Compose,
#     OneOf,
#     ElasticTransform,
#     GridDistortion, 
#     OpticalDistortion,
#     HorizontalFlip,
#     Rotate,
#     Transpose,
#     CLAHE,
#     ShiftScaleRotate
# )

# def get_y_fn (x):
#     return Path(str(x).replace("Images","Labels").replace("color","gt").replace(".jpg",".png"))

# def ParentSplitter(x):
#     return Path(x).parent.name==test_name

# class SegmentationAlbumentationsTransform(ItemTransform):
#     split_idx = 0
    
#     def __init__(self, aug): 
#         self.aug = aug
        
#     def encodes(self, x):
#         img,mask = x
#         aug = self.aug(image=np.array(img), mask=np.array(mask))
#         return PILImage.create(aug["image"]), PILMask.create(aug["mask"])

# transforms=Compose([HorizontalFlip(p=0.5),
#                     Rotate(p=0.40,limit=10),GridDistortion()
#                     ],p=1)
# transformPipeline=SegmentationAlbumentationsTransform(transforms)

# class TargetMaskConvertTransform(ItemTransform):
#     def __init__(self): 
#         pass
#     def encodes(self, x):
#         img,mask = x
        
#         #Convert to array
#         mask = np.array(mask)
        
#         # Aquí definimos cada clase en la máscara
#         # uva:
#         mask[mask==255]=1
#         # hojas:
#         mask[mask==150]=2
#         # conductores:
#         mask[mask==76]=3
#         mask[mask==74]=3
#         # madera:
#         mask[mask==29]=4
#         mask[mask==25]=4
        
#         # Back to PILMask
#         mask = PILMask.create(mask)
#         return img, mask

# repo_id = "paascorb/practica3_Segmentation"

# learner = from_pretrained_fastai(repo_id)

learn = unet_learner(None,resnet18,cbs=None,metrics=[DiceMulti()]).to_fp16()
learn.load('model')

def transform_image(image):
    my_transforms = transforms.Compose([transforms.ToTensor(),
                                        transforms.Normalize(
                                            [0.485, 0.456, 0.406],
                                            [0.229, 0.224, 0.225])])
    image_aux = image
    return my_transforms(image_aux).unsqueeze(0).to(device)

def predict(img):
    img = PIL.Image.fromarray(img, "RGB")
    image = transforms.Resize((480,640))(img)
    tensor = transform_image(image=image) 

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    learner.to(device)
    with torch.no_grad():
        outputs = learner(tensor)
    
    outputs = torch.argmax(outputs,1)
    mask = np.array(outputs.cpu())
    mask[mask==1]=255
    mask[mask==2]=150
    mask[mask==3]=76
    mask[mask==4]=29
    mask=np.reshape(mask,(480,640))
    return Image.fromarray(mask.astype('uint8'))
    
gr.Interface(fn=predict, inputs=gr.inputs.Image(shape=(128, 128)), outputs=[gr.outputs.Image(type="pil", label="Predicción")], examples=['color_155.jpg','color_154.jpg']).launch(share=False)