simulacra_glide_sample.py

import argparse, os, sys, glob
import random
import torch
from torch import nn
from torch.nn import functional as F
from torchvision.transforms import functional as TF
import numpy as np
from omegaconf import OmegaConf
from PIL import Image
from tqdm import tqdm, trange
from einops import rearrange
from torchvision.utils import make_grid

from ldm.util import instantiate_from_config
from ldm.models.diffusion.ddim import DDIMSampler
from ldm.models.diffusion.plms import PLMSSampler

from CLIP import clip

sys.path.append('./cloob-training')
from cloob_training import model_pt, pretrained

class MakeCutouts(nn.Module):
    def __init__(self, cut_size, cutn, cut_pow=1.):
        super().__init__()
        self.cut_size = cut_size
        self.cutn = cutn
        self.cut_pow = cut_pow

    def forward(self, input):
        sideY, sideX = input.shape[2:4]
        max_size = min(sideX, sideY)
        min_size = min(sideX, sideY, self.cut_size)
        cutouts = []
        for _ in range(self.cutn):
            size = int(torch.rand([])**self.cut_pow * (max_size - min_size) + min_size)
            offsetx = torch.randint(0, sideX - size + 1, ())
            offsety = torch.randint(0, sideY - size + 1, ())
            cutout = input[:, :, offsety:offsety + size, offsetx:offsetx + size]
            cutout = F.adaptive_avg_pool2d(cutout, self.cut_size)
            cutouts.append(cutout)
        return torch.cat(cutouts)


def spherical_dist_loss(x, y):
    x = F.normalize(x, dim=-1)
    y = F.normalize(y, dim=-1)
    return (x - y).norm(dim=-1).div(2).arcsin().pow(2).mul(2)    


def load_model_from_config(config, ckpt, verbose=False):
    print(f"Loading model from {ckpt}")
    pl_sd = torch.load(ckpt, map_location="cpu")
    sd = pl_sd["state_dict"]
    model = instantiate_from_config(config.model)
    m, u = model.load_state_dict(sd, strict=False)
    if len(m) > 0 and verbose:
        print("missing keys:")
        print(m)
    if len(u) > 0 and verbose:
        print("unexpected keys:")
        print(u)

    model.cuda()
    model.eval()
    return model


def main(opt):
    parser = argparse.ArgumentParser()

    parser.add_argument(
        "--prompt",
        type=str,
        nargs="?",
        default="a painting of a virus monster playing guitar",
        help="the prompt to render"
    )

    parser.add_argument(
        "--outdir",
        type=str,
        nargs="?",
        help="dir to write results to",
        default="outputs/txt2img-samples"
    )
    parser.add_argument(
        "--ddim_steps",
        type=int,
        default=200,
        help="number of ddim sampling steps",
    )

    parser.add_argument(
        "--plms",
        action='store_true',
        help="use plms sampling",
    )

    parser.add_argument(
        "--ddim_eta",
        type=float,
        default=0.0,
        help="ddim eta (eta=0.0 corresponds to deterministic sampling",
    )
    parser.add_argument(
        "--n_iter",
        type=int,
        default=1,
        help="sample this often",
    )

    parser.add_argument(
        "--H",
        type=int,
        default=256,
        help="image height, in pixel space",
    )

    parser.add_argument(
        "--W",
        type=int,
        default=256,
        help="image width, in pixel space",
    )

    parser.add_argument(
        "--n_samples",
        type=int,
        default=4,
        help="how many samples to produce for the given prompt",
    )

    parser.add_argument(
        "--scale",
        type=float,
        default=5.0,
        help="unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty))",
    )
    # opt = parser.parse_args()

    config = OmegaConf.load("configs/latent-diffusion/txt2img-1p4B-eval.yaml")  # TODO: Optionally download from same location as ckpt and chnage this logic
    model = load_model_from_config(config, "models/ldm/text2img-large/model.ckpt")  # TODO: check path

    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    model = model.to(device)

    cloob_config = pretrained.get_config(opt.cloob_checkpoint)
    cloob = model_pt.get_pt_model(cloob_config)
    checkpoint = pretrained.download_checkpoint(cloob_config)
    cloob.load_state_dict(model_pt.get_pt_params(cloob_config, checkpoint))
    cloob.eval().requires_grad_(False).to(device)

    clip_model = clip.load('ViT-L/14')[0]
    clip_model.eval().requires_grad_(False).to(device)
    clip_target_embed = clip_model.encode_text(clip.tokenize(opt.prompt).to(device))
    make_cutouts = MakeCutouts(224, opt.cutn)    
    
    if opt.plms:
        sampler = PLMSSampler(model)
    else:
        sampler = DDIMSampler(model)


    prompt = opt.prompt

    all_samples=list()
    with torch.no_grad():
        with model.ema_scope():
            uc = None
            if opt.scale != 1.0:
                uc = model.get_learned_conditioning(opt.n_samples * [""])
            for n in trange(opt.n_iter, desc="Sampling"):
                c = model.get_learned_conditioning(opt.n_samples * [prompt])
                shape = [4, opt.H//8, opt.W//8]
                samples_ddim, _ = sampler.sample(S=opt.ddim_steps,
                                                 conditioning=c,
                                                 batch_size=opt.n_samples,
                                                 shape=shape,
                                                 verbose=False,
                                                 unconditional_guidance_scale=opt.scale,
                                                 unconditional_conditioning=uc,
                                                 eta=opt.ddim_eta)

                x_samples_ddim = model.decode_first_stage(samples_ddim)
                x_samples_ddim = torch.clamp((x_samples_ddim+1.0)/2.0, min=0.0, max=1.0)
                all_samples += x_samples_ddim


    # CLIP rank our outputs
    #target_embed = cloob.text_encoder(cloob.tokenize(opt.prompt).to(device))
    outs = torch.stack(all_samples)
    #clip_in = cloob.normalize(make_cutouts((outs + 1) / 2))
    #image_embeds = clip_model.encode_image(clip_in).view([opt.cutn, outs.shape[0], -1])
    #losses = spherical_dist_loss(image_embeds, clip_target_embed)
    #losses = losses.mean(1)
    #top_ranked = min([lp for lp in zip(outs, losses)], key=lambda lp: lp[1])[0]
    for out in enumerate(outs):
        index, out = out
        outpath = str(opt.seed) + "_" + opt.prompt.replace(" ", "_").replace("/","_") + "_" + str(index + 1) + ".png"
        out = 255. * rearrange(out.cpu().numpy(), 'c h w -> h w c')
        Image.fromarray(out.astype(np.uint8)).save(outpath)

    gridpath = str(opt.seed) + "_" + opt.prompt.replace(" ", "_").replace("/","_") + "_grid" + ".png"
    grid = torch.stack(all_samples, 0)
    grid = rearrange(grid, 'b c h w -> (b) c h w')
    grid = make_grid(grid, nrow=opt.n_samples)

    # to image
    grid = 255. * rearrange(grid, 'c h w -> h w c').cpu().numpy()
    Image.fromarray(grid.astype(np.uint8)).save(gridpath)