app.py

import gradio as gr
import lightning
import numpy as np
import os
import pandas as pd
import timm
import torch
import torch.nn as nn

from torch.utils.data import Dataset, DataLoader

BACKBONE = "resnet18d"
IMAGE_HEIGHT, IMAGE_WIDTH = 512, 512

trained_weights_path = "epoch=9-step=520.ckpt"
trained_weights = torch.load(trained_weights_path, map_location=torch.device('cpu'))["state_dict"]

# recreate the model

class Net(nn.Module):

  def __init__(self, backbone):
    super().__init__()
    self.backbone = timm.create_model(backbone, pretrained=True, in_chans=1, num_classes=0)
    dim_feats = self.backbone(torch.randn((2, 1, 64, 64))).size(1)
    self.embed = nn.Embedding(2, 32)
    self.regressor = nn.Linear(dim_feats + 32, 1)

  def forward(self, x, female):
    feat = self.backbone(x)
    feat = torch.cat([feat, self.embed(female.long())], dim=1)
    return self.regressor(feat)

class BoneAgeModel(lightning.LightningModule):

  def __init__(self, net, optimizer, scheduler, loss_fn):
    super().__init__()
    self.net = net
    self.optimizer = optimizer
    self.scheduler = scheduler
    self.loss_fn = loss_fn

    self.val_losses = []

  def training_step(self, batch, batch_index):
    out = self.net(batch["x"])
    loss = self.loss_fn(out, batch["y"])
    return loss

  def validation_step(self, batch, batch_index):
    out = self.net(batch["x"])
    loss = self.loss_fn(out, batch["y"])
    self.val_losses.append(loss.item())

  def on_validation_epoch_end(self, *args, **kwargs):
    val_loss = np.mean(self.val_losses)
    self.val_losses = []
    print(f"Validation Loss : {val_loss:0.3f}")

  def configure_optimizers(self):
    lr_scheduler = {"scheduler": self.scheduler, "interval": "step"}
    return {"optimizer": self.optimizer, "lr_scheduler": lr_scheduler}

class BoneAgeModelV2(BoneAgeModel):

  def training_step(self, batch, batch_index):
    out = self.net(batch["x"], batch["female"])
    loss = self.loss_fn(out, batch["y"])
    return loss

  def validation_step(self, batch, batch_index):
    out = self.net(batch["x"], batch["female"])
    loss = self.loss_fn(out, batch["y"])
    self.val_losses.append(loss.item())


net = Net(BACKBONE)
trained_model = BoneAgeModelV2(net, None, None, None)
trained_model.load_state_dict(trained_weights)
trained_model.eval()


def predict_bone_age(Radiograph, Sex):
  img = torch.from_numpy(Radiograph)
  img = img.unsqueeze(0).unsqueeze(0) # add channel and batch dimensions
  img = img / 255. # use same normalization as in the PyTorch dataset
  binary_sex = torch.tensor(Sex == "Female").unsqueeze(0)
  with torch.inference_mode():
    bone_age = trained_model.net(img, binary_sex)[0].item()
  years = int(bone_age)
  months = round((bone_age - years) * 12)
  return f"Predicted Bone Age: {years} years, {months} months"


image = gr.Image(height=IMAGE_HEIGHT, width=IMAGE_WIDTH, image_mode="L") # L for grayscale
# additional input
sex = gr.Radio(["Male", "Female"], type="index")
label = gr.Label(show_label=True, label="Bone Age Prediction")

demo = gr.Interface(fn=predict_bone_age,
                    inputs=[image, sex], # <- adding sex as an input
                    outputs=label)

demo.launch(debug=True)