app.py

import gradio as gr
import numpy as np
from diffusers import UNet2DModel, DDPMPipeline, DDPMScheduler, DiffusionPipeline
import torch
import torch.nn.functional as F
from matplotlib import pyplot as plt
from PIL import Image
import spaces


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
pipeline = DiffusionPipeline.from_pretrained("gjbooth2/Unconditional_A4C_1").to(device)



#try to return dataframe
def image_gen(click,rows = 4,cols = 4):
    images = pipeline(batch_size=16).images
    w, h = images[0].size
    grid = Image.new('L', size=(cols*w, rows*h))
    for i, image in enumerate(images):
        grid.paste(image, box=(i%cols*w, i//cols*h))
    return grid
    #return 'button clicked'
    
@spaces.GPU(duration = 300)
def image_gen_modified(rows=4,cols=4):
  pic_hold = []
  model_output = pipeline(batch_size=16).images
  count = 0
  for i in range(len(model_output)):
    pic = np.array(model_output[i].convert('L'))
    max_val = max([element for row in pic for element in row])
    min_val = min([element for row in pic for element in row])

    if min_val > 55: #for washed out images, set them to all black
      normalized_pic = np.ones((128,128))
      pic_hold.append(Image.fromarray(np.uint8(normalized_pic)))
      
    if min_val < 56:
      def normalize_images(x,min_val,max_val): #normalize pixels to be more homogenous grayscale appearance
          return 200*((x-min_val)/(max_val-min_val))
      vectorized_normalizer = np.vectorize(normalize_images)
      normalized_pic = vectorized_normalizer(pic,min_val,max_val)
      pic_hold.append(Image.fromarray(np.uint8(normalized_pic)))
      count+=1

  w, h = model_output[0].size
  grid = Image.new('L', size=(cols*w, rows*h))
  for i, image in enumerate(pic_hold):
      grid.paste(image, box=(i%cols*w, i//cols*h))

  return grid


with gr.Blocks(theme=gr.themes.Soft()) as demo:
  gr.Markdown('CS 614 Greg Booth Vision Assignment')
  gr.Markdown('This gradio app can be used to generate realistic cardiac ultrasound images.')
  gr.HTML("<a href = "+'https://pocus.sg/topic/subcostal-4-chamber/'+" _target='blank'>" +'Example anatomy'+ "</a>")

  with gr.Tab('Generate a cardiac ultrasound image'):
    playground_btn = gr.Button(value='Push me some images! (may take a couple minutes depending on hardware)')
    playground_out = gr.Image()
    playground_btn.click(image_gen_modified,outputs = playground_out)


demo.launch()