app.py

import gradio as gr
import timm
import torch 
import torch.nn as nn


class Net2D(nn.Module):

    def __init__(self, weights):
        super().__init__()
        self.backbone = timm.create_model("tf_efficientnet_b6_ns", pretrained=False, global_pool="", num_classes=0)
        self.pool_layer = nn.AdaptiveAvgPool2d(1)
        self.dropout = nn.Dropout(0.5)
        self.linear = nn.Linear(2304, 5)
        self.load_state_dict(weights)

    def forward(self, x):
        x = self.backbone(x)
        x = self.pool_layer(x).view(x.size(0), -1)
        x = self.dropout(x)
        x = self.linear(x)
        return x[:, 0] if x.size(1) == 1 else x


def rescale(x):
    x = x / 255.0
    x = x - 0.5
    x = x * 2.0
    return x


weights = torch.load("model0.ckpt", map_location=torch.device("cpu"))["state_dict"]
weights = {k.replace("model.", ""): v for k, v in weights.items()}
model = Net2D(weights).eval()


def predict(Image):
    img = torch.from_numpy(Image)
    img = img.permute(2, 0, 1)
    img = img.unsqueeze(0)
    img = rescale(img)
    with torch.no_grad():
        grade = torch.softmax(model(img.float()), dim=1)[0]
    cats = ["None", "Mild", "Moderate", "Severe", "Proliferative"]
    output_dict = {}
    for cat, value in zip(cats, grade):
        output_dict[cat] = value.item()
    return output_dict



image = gr.Image(shape=(512, 512), image_mode="RGB")
label = gr.Label(label="Grade")

demo = gr.Interface(
    fn=predict,
    inputs=image,
    outputs=label,
    examples=["examples/0.png", "examples/1.png", "examples/2.png", "examples/3.png", "examples/4.png"]
    )


if __name__ == "__main__":
    demo.launch(debug=True)