app.py

import streamlit as st
from PIL import Image
import jax
import jax.numpy as jnp  # JAX NumPy
import numpy as np
from huggingface_hub import HfFileSystem
from flax.serialization import msgpack_restore, from_state_dict
import time
from generator import Generator, LATENT_DIM
import math

generator = Generator()
variables = generator.init(jax.random.PRNGKey(0), jnp.zeros([1, LATENT_DIM]), training=False)

fs = HfFileSystem()
with fs.open("PrakhAI/AIPlane2/g_checkpoint.msgpack", "rb") as f:
  g_state = from_state_dict(variables, msgpack_restore(f.read()))

def sample_latent(batch, key):
  return jax.random.normal(key, shape=(batch, LATENT_DIM))

def gridify(images): # num x image_width x image_height x channels
  # Every num can be padded to make a grid of size floor(sqrt(num)) x ceil(sqrt(num)) or ceil(sqrt(num)) x ceil(sqrt(num))
  num = images.shape[0]
  image_width = images.shape[1]
  image_height = images.shape[2]
  channels = images.shape[3]
  width = math.floor(math.sqrt(num))
  height = math.ceil(math.sqrt(num))
  if width * height < num:
    width += 1
  padded = np.concatenate([images, np.zeros((width*height-num, image_width, image_height, channels))], axis=0)
  return padded.reshape((width, height, image_width, image_height, -1)).transpose((0, 2, 1, 3, 4)).reshape((width * image_width, height * image_height, -1))

st.write("The model and its details are at https://huggingface.co/PrakhAI/AIPlane2")
num_images = st.number_input(label="Number of images to generate", min_value=1, max_value=256, value=16)
if st.button('Generate Planes'):
  latents = sample_latent(num_images, jax.random.PRNGKey(int(1_000_000 * time.time())))
  (g_out128, _, _, _, _, _) = generator.apply({'params': g_state['params'], 'batch_stats': g_state['batch_stats']}, latents, training=False)
  img = (np.array(gridify(g_out128)+1)*255./2.).astype(np.uint8)
  st.image(Image.fromarray(img))