app.py

import torch
from torch import nn
import torchvision
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image
import pandas as pd
import segmentation_models_pytorch as smp
import gradio as gr

num_classes = 2
model_unet_path = "unet_model.pth"
model_fpn_path = "fpn_model.pth"
model_deeplab_path = "deeplabv3_model.pth"
image_path = "leaf11.jpg"

# Get cpu or gpu device for training.
device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
print(f"Using {device} device")

model_unet = smp.Unet(
    encoder_name="resnet18",        # choose encoder, e.g. mobilenet_v2 or efficientnet-b7
    encoder_weights=None,     # use `imagenet` pre-trained weights for encoder initialization
    in_channels=3,                  # model input channels (1 for gray-scale images, 3 for RGB, etc.)
    classes=num_classes,                      # model output channels (number of classes in your dataset)
)

model_fpn = smp.FPN(
    encoder_name="resnet18",        # choose encoder, e.g. mobilenet_v2 or efficientnet-b7
    encoder_weights=None,     # use `imagenet` pre-trained weights for encoder initialization
    in_channels=3,                  # model input channels (1 for gray-scale images, 3 for RGB, etc.)
    classes=num_classes,                      # model output channels (number of classes in your dataset)
)

model_deeplab = smp.DeepLabV3(
    encoder_name="resnet34",        # choose encoder, e.g. mobilenet_v2 or efficientnet-b7
    encoder_weights=None,     # use `imagenet` pre-trained weights for encoder initialization
    in_channels=3,                  # model input channels (1 for gray-scale images, 3 for RGB, etc.)
    classes=num_classes,                      # model output channels (number of classes in your dataset)
)

def pred_one_image(inp,option):
  one_image = np.array(inp.resize((256, 256)).convert("RGB"))
  # convert to other format HWC -> CHW
  one_image = np.moveaxis(one_image, -1, 0)
  # mask = np.expand_dims(mask, 0)
  one_image = torch.tensor(one_image).float()
  one_image = one_image.unsqueeze(0)
  one_image = one_image.to(device)
  if option == "unet":
    model_load = model_unet
  elif option == "fpn":
    model_load = model_fpn
  elif option == "deeplab":
    model_load = model_deeplab
  model_load.eval()
  with torch.no_grad():
    output = model_load(one_image)
    # print(output.shape)
    predictions = torch.argmax(output, dim=1)  # 获取预测的类别标签图像
    pred_array = (predictions[0].cpu().numpy()/2*255).astype(np.uint8)
    # print(pred_array.shape)
    pred_img = Image.fromarray(pred_array)
    # pred_img.save("pred.png")
    # print(predictions.shape)
  return pred_img



model_unet.load_state_dict(torch.load(model_unet_path,map_location=torch.device('cpu')))
model_fpn.load_state_dict(torch.load(model_fpn_path,map_location=torch.device('cpu')))
model_deeplab.load_state_dict(torch.load(model_deeplab_path,map_location=torch.device('cpu')))

dropdown = gr.Dropdown(["unet", "fpn","deeplab"])
interface = gr.Interface(fn=pred_one_image,
            inputs=[gr.Image(type="pil"),dropdown], 
            outputs=gr.Image(type="pil"),
            examples=[["leaf11.jpg",'unet']],)
interface.launch(debug=False)