app.py

# -*- coding: utf-8 -*-
"""Copy of compose_glide.ipynb

Automatically generated by Colaboratory.

Original file is located at
    https://colab.research.google.com/drive/19xx6Nu4FeiGj-TzTUFxBf-15IkeuFx_F
"""

# from PIL import Image
# from IPython.display import display
import torch as th
import numpy as np

from glide_text2im.download import load_checkpoint
from glide_text2im.model_creation import (
    create_model_and_diffusion,
    model_and_diffusion_defaults,
    model_and_diffusion_defaults_upsampler
)

from composable_diffusion.download import download_model
from composable_diffusion.model_creation import create_model_and_diffusion as create_model_and_diffusion_for_clevr
from composable_diffusion.model_creation import model_and_diffusion_defaults as model_and_diffusion_defaults_for_clevr


from PIL import Image
# This notebook supports both CPU and GPU.
# On CPU, generating one sample may take on the order of 20 minutes.
# On a GPU, it should be under a minute.

has_cuda = th.cuda.is_available()
device = th.device('cpu' if not has_cuda else 'cuda')
print(device)

# Create base model.
timestep_respacing = 100  # @param{type: 'number'}
options = model_and_diffusion_defaults()
options['use_fp16'] = has_cuda
options['timestep_respacing'] = str(timestep_respacing)  # use 100 diffusion steps for fast sampling
model, diffusion = create_model_and_diffusion(**options)
model.eval()
if has_cuda:
    model.convert_to_fp16()
model.to(device)
model.load_state_dict(load_checkpoint('base', device))
print('total base parameters', sum(x.numel() for x in model.parameters()))

# Create upsampler model.
options_up = model_and_diffusion_defaults_upsampler()
options_up['use_fp16'] = has_cuda
options_up['timestep_respacing'] = 'fast27'  # use 27 diffusion steps for very fast sampling
model_up, diffusion_up = create_model_and_diffusion(**options_up)
model_up.eval()
if has_cuda:
    model_up.convert_to_fp16()
model_up.to(device)
model_up.load_state_dict(load_checkpoint('upsample', device))
print('total upsampler parameters', sum(x.numel() for x in model_up.parameters()))


def show_images(batch: th.Tensor):
    """ Display a batch of images inline. """
    scaled = ((batch + 1) * 127.5).round().clamp(0, 255).to(th.uint8).cpu()
    reshaped = scaled.permute(2, 0, 3, 1).reshape([batch.shape[2], -1, 3])
    display(Image.fromarray(reshaped.numpy()))


def compose_language_descriptions(prompt, guidance_scale):
    # @markdown `prompt`: when composing  multiple sentences, using `|` as the delimiter.
    prompts = [x.strip() for x in prompt.split('|')]

    batch_size = 1
    # Tune this parameter to control the sharpness of 256x256 images.
    # A value of 1.0 is sharper, but sometimes results in grainy artifacts.
    upsample_temp = 0.980  # @param{type: 'number'}

    masks = [True] * len(prompts) + [False]
    # coefficients = th.tensor([0.5, 0.5], device=device).reshape(-1, 1, 1, 1)
    masks = th.tensor(masks, dtype=th.bool, device=device)

    # sampling function
    def model_fn(x_t, ts, **kwargs):
        half = x_t[:1]
        combined = th.cat([half] * x_t.size(0), dim=0)
        model_out = model(combined, ts, **kwargs)
        eps, rest = model_out[:, :3], model_out[:, 3:]
        cond_eps = eps[masks].mean(dim=0, keepdim=True)
        # cond_eps = (coefficients * eps[masks]).sum(dim=0)[None]
        uncond_eps = eps[~masks].mean(dim=0, keepdim=True)
        half_eps = uncond_eps + guidance_scale * (cond_eps - uncond_eps)
        eps = th.cat([half_eps] * x_t.size(0), dim=0)
        return th.cat([eps, rest], dim=1)

    ##############################
    # Sample from the base model #
    ##############################

    # Create the text tokens to feed to the model.
    def sample_64(prompts):
        tokens_list = [model.tokenizer.encode(prompt) for prompt in prompts]
        outputs = [model.tokenizer.padded_tokens_and_mask(
            tokens, options['text_ctx']
        ) for tokens in tokens_list]

        cond_tokens, cond_masks = zip(*outputs)
        cond_tokens, cond_masks = list(cond_tokens), list(cond_masks)

        full_batch_size = batch_size * (len(prompts) + 1)
        uncond_tokens, uncond_mask = model.tokenizer.padded_tokens_and_mask(
            [], options['text_ctx']
        )

        # Pack the tokens together into model kwargs.
        model_kwargs = dict(
            tokens=th.tensor(
                cond_tokens + [uncond_tokens], device=device
            ),
            mask=th.tensor(
                cond_masks + [uncond_mask],
                dtype=th.bool,
                device=device,
            ),
        )

        # Sample from the base model.
        model.del_cache()
        samples = diffusion.p_sample_loop(
            model_fn,
            (full_batch_size, 3, options["image_size"], options["image_size"]),
            device=device,
            clip_denoised=True,
            progress=True,
            model_kwargs=model_kwargs,
            cond_fn=None,
        )[:batch_size]
        model.del_cache()

        # Show the output
        return samples

    ##############################
    # Upsample the 64x64 samples #
    ##############################

    def upsampling_256(prompts, samples):
        tokens = model_up.tokenizer.encode("".join(prompts))
        tokens, mask = model_up.tokenizer.padded_tokens_and_mask(
            tokens, options_up['text_ctx']
        )

        # Create the model conditioning dict.
        model_kwargs = dict(
            # Low-res image to upsample.
            low_res=((samples + 1) * 127.5).round() / 127.5 - 1,

            # Text tokens
            tokens=th.tensor(
                [tokens] * batch_size, device=device
            ),
            mask=th.tensor(
                [mask] * batch_size,
                dtype=th.bool,
                device=device,
            ),
        )

        # Sample from the base model.
        model_up.del_cache()
        up_shape = (batch_size, 3, options_up["image_size"], options_up["image_size"])
        up_samples = diffusion_up.ddim_sample_loop(
            model_up,
            up_shape,
            noise=th.randn(up_shape, device=device) * upsample_temp,
            device=device,
            clip_denoised=True,
            progress=True,
            model_kwargs=model_kwargs,
            cond_fn=None,
        )[:batch_size]
        model_up.del_cache()

        # Show the output
        return up_samples

    # sampling 64x64 images
    samples = sample_64(prompts)
    # show_images(samples)

    # upsample from 64x64 to 256x256
    upsamples = upsampling_256(prompts, samples)
    # show_images(upsamples)

    out_img = upsamples[0].permute(1, 2, 0)
    out_img = (out_img + 1) / 2
    out_img = (out_img.detach().cpu() * 255.).to(th.uint8)
    out_img = out_img.numpy()
    return out_img


# create model for CLEVR Objects
clevr_options = model_and_diffusion_defaults_for_clevr()

flags = {
    "image_size": 128,
    "num_channels": 192,
    "num_res_blocks": 2,
    "learn_sigma": True,
    "use_scale_shift_norm": False,
    "raw_unet": True,
    "noise_schedule": "squaredcos_cap_v2",
    "rescale_learned_sigmas": False,
    "rescale_timesteps": False,
    "num_classes": '2',
    "dataset": "clevr_pos",
    "use_fp16": has_cuda,
    "timestep_respacing": '100'
}

for key, val in flags.items():
    clevr_options[key] = val

clevr_model, clevr_diffusion = create_model_and_diffusion_for_clevr(**clevr_options)
clevr_model.eval()
if has_cuda:
    clevr_model.convert_to_fp16()

clevr_model.to(device)
clevr_model.load_state_dict(th.load(download_model('clevr_pos'), device))
print('total clevr_pos parameters', sum(x.numel() for x in clevr_model.parameters()))


def compose_clevr_objects(prompt, guidance_scale):
    coordinates = [[float(x.split(',')[0].strip()), float(x.split(',')[1].strip())]
                   for x in prompt.split('|')]
    coordinates += [[-1, -1]]  # add unconditional score label
    batch_size = 1

    def model_fn(x_t, ts, **kwargs):
        half = x_t[:1]
        combined = th.cat([half] * kwargs['y'].size(0), dim=0)
        model_out = clevr_model(combined, ts, **kwargs)
        eps, rest = model_out[:, :3], model_out[:, 3:]
        masks = kwargs.get('masks')
        cond_eps = eps[masks].mean(dim=0, keepdim=True)
        uncond_eps = eps[~masks].mean(dim=0, keepdim=True)
        half_eps = uncond_eps + guidance_scale * (cond_eps - uncond_eps)
        eps = th.cat([half_eps] * x_t.size(0), dim=0)
        return th.cat([eps, rest], dim=1)

    def sample(coordinates):
        masks = [True] * (len(coordinates) - 1) + [False]
        model_kwargs = dict(
            y=th.tensor(coordinates, dtype=th.float, device=device),
            masks=th.tensor(masks, dtype=th.bool, device=device)
        )
        samples = clevr_diffusion.p_sample_loop(
            model_fn,
            (len(coordinates), 3, clevr_options["image_size"], clevr_options["image_size"]),
            device=device,
            clip_denoised=True,
            progress=True,
            model_kwargs=model_kwargs,
            cond_fn=None,
        )[:batch_size]

        return samples

    samples = sample(coordinates)
    out_img = samples[0].permute(1, 2, 0)
    out_img = (out_img + 1) / 2
    out_img = (out_img.detach().cpu() * 255.).to(th.uint8)
    out_img = out_img.numpy()
    Image.fromarray(out_img).convert('RGB').save('test.png')

    return out_img


def compose(prompt, version, guidance_scale):
    if version == 'GLIDE':
        return compose_language_descriptions(prompt, guidance_scale)
    else:
        return compose_clevr_objects(prompt, guidance_scale)


examples_1 = 'a camel | a forest'
examples_2 = 'A cloudy blue sky  | A mountain in the horizon | Cherry Blossoms in front of the mountain'
examples_3 = '0.1, 0.5 | 0.3, 0.5 | 0.5, 0.5 | 0.7, 0.5 | 0.9, 0.5'
examples = [[examples_1, 'GLIDE', 10], [examples_2, 'GLIDE', 10], [examples_3, 'CLEVR Objects', 10]]

import gradio as gr

title = 'Compositional Visual Generation with Composable Diffusion Models'
description = '<p>Demo for Composable Diffusion<ul><li>~30s per GLIDE example</li><li>~10s per CLEVR Object example</li>(<b>Note</b>: time is measured by per example if gpu is used, otherwise it will take quite a bit of time.)</ul></p><p>See more information from our <a href="https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/">Project Page</a>.</p><ul><li>One version is based on the released <a href="https://github.com/openai/glide-text2im">GLIDE</a> for composing natural language description.</li><li>Another is based on our pre-trained CLEVR Object Model for composing objects. <br>(<b>Note</b>: We recommend using <b><i>x</i></b> in range <b><i>[0.1, 0.9]</i></b> and <b><i>y</i></b> in range <b><i>[0.25, 0.7]</i></b>, since the training dataset labels are in given ranges.)</li></ul><p>When composing  multiple sentences, use `|` as the delimiter, see given examples below.</p>'

iface = gr.Interface(compose, inputs=["text", gr.Radio(['GLIDE', 'CLEVR Objects'], type="value", label='version'), gr.Slider(1, 20)], outputs='image',
                     title=title, description=description, examples=examples)

iface.launch()