Thanks for tuning in to Google I/O. View all sessions on demandWatch on demand

Boundless Colab

View on TensorFlow.org Run in Google Colab View on GitHub Download notebook See TF Hub models

Welcome to the Boundless model Colab! This notebook will take you through the steps of running the model on images and visualize the results.

Overview

Boundless is a model for image extrapolation. This model takes an image, internally masks a portion of it (1/2, 1/4, 3/4) and completes the masked part. For more details refer to Boundless: Generative Adversarial Networks for Image Extension or the model documentation on TensorFlow Hub.

Imports and Setup

Lets start with the base imports.

import tensorflow as tf
import tensorflow_hub as hub
from io import BytesIO
from PIL import Image as PilImage
import numpy as np
from matplotlib import pyplot as plt
from six.moves.urllib.request import urlopen
2022-12-14 13:58:58.625094: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory
2022-12-14 13:58:58.625187: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory
2022-12-14 13:58:58.625196: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.

Reading image for input

Lets create a util method to help load the image and format it for the model (257x257x3). This method will also crop the image to a square to avoid distortion and you can use with local images or from the internet.

def read_image(filename):
    fd = None
    if(filename.startswith('http')):
      fd = urlopen(filename)
    else:
      fd = tf.io.gfile.GFile(filename, 'rb')

    pil_image = PilImage.open(fd)
    width, height = pil_image.size
    # crop to make the image square
    pil_image = pil_image.crop((0, 0, height, height))
    pil_image = pil_image.resize((257,257),PilImage.ANTIALIAS)
    image_unscaled = np.array(pil_image)
    image_np = np.expand_dims(
        image_unscaled.astype(np.float32) / 255., axis=0)
    return image_np

Visualization method

We will also create a visuzalization method to show the original image side by side with the masked version and the "filled" version, both generated by the model.

def visualize_output_comparison(img_original, img_masked, img_filled):
  plt.figure(figsize=(24,12))
  plt.subplot(131)
  plt.imshow((np.squeeze(img_original)))
  plt.title("Original", fontsize=24)
  plt.axis('off')
  plt.subplot(132)
  plt.imshow((np.squeeze(img_masked)))
  plt.title("Masked", fontsize=24)
  plt.axis('off')
  plt.subplot(133)
  plt.imshow((np.squeeze(img_filled)))
  plt.title("Generated", fontsize=24)
  plt.axis('off')
  plt.show()

Loading an Image

We will load a sample image but fell free to upload your own image to the colab and try with it. Remember that the model have some limitations regarding human images.

wikimedia = "https://upload.wikimedia.org/wikipedia/commons/thumb/3/31/Nusfjord_road%2C_2010_09.jpg/800px-Nusfjord_road%2C_2010_09.jpg"
# wikimedia = "https://upload.wikimedia.org/wikipedia/commons/thumb/4/47/Beech_forest_M%C3%A1tra_in_winter.jpg/640px-Beech_forest_M%C3%A1tra_in_winter.jpg"
# wikimedia = "https://upload.wikimedia.org/wikipedia/commons/thumb/b/b2/Marmolada_Sunset.jpg/640px-Marmolada_Sunset.jpg"
# wikimedia = "https://upload.wikimedia.org/wikipedia/commons/thumb/9/9d/Aegina_sunset.jpg/640px-Aegina_sunset.jpg"

input_img = read_image(wikimedia)
/tmpfs/tmp/ipykernel_110065/3295904410.py:12: DeprecationWarning: ANTIALIAS is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.LANCZOS instead.
  pil_image = pil_image.resize((257,257),PilImage.ANTIALIAS)

Selecting a model from TensorFlow Hub

On TensorFlow Hub we have 3 versions of the Boundless model: Half, Quarter and Three Quarters. In the following cell you can chose any of them and try on your image. If you want to try with another one, just chose it and execute the following cells.

Model Selection

Now that we've chosen the model we want, lets load it from TensorFlow Hub.

print("Loading model {} ({})".format(model_name, model_handle))
model = hub.load(model_handle)
Loading model Boundless Quarter (https://tfhub.dev/google/boundless/quarter/1)

Doing Inference

The boundless model have two outputs:

  • The input image with a mask applied
  • The masked image with the extrapolation to complete it

we can use these two images to show a comparisson visualization.

result = model.signatures['default'](tf.constant(input_img))
generated_image =  result['default']
masked_image = result['masked_image']

visualize_output_comparison(input_img, masked_image, generated_image)

png