import cv2 as cv import numpy as np import argparse from tqdm import tqdm from functools import partial from copy import deepcopy ## let use write description of the script and general information how to use it ''' This sample proposes experimental inpainting sample using Latent Diffusion Model (LDM) for inpainting. Most of the script is based on the code from the official repository of the LDM model: https://github.com/CompVis/latent-diffusion Current limitations of the script: - Slow diffusion sampling - Not exact reproduction of the results from the original repository (due to issues related deviation in covolution operation. See issue for more details: https://github.com/opencv/opencv/pull/25973) Steps for running the script: 1. Firstly generate ONNX graph of the Latent Diffusion Model. Generate the using this [repo](https://github.com/Abdurrahheem/latent-diffusion/tree/ash/export2onnx) and follow instructions below - git clone https://github.com/Abdurrahheem/latent-diffusion.git - cd latent-diffusion - conda env create -f environment.yaml - conda activate ldm - wget -O models/ldm/inpainting_big/last.ckpt https://heibox.uni-heidelberg.de/f/4d9ac7ea40c64582b7c9/?dl=1 - python -m scripts.inpaint.py --indir data/inpainting_examples/ --outdir outputs/inpainting_results --export=True 2. Build opencv (preferebly with CUDA support enabled 3. Run the script - cd opencv/samples/dnn - python ldm_inpainting.py -e= -d= -df= -i= Right after the last command you will be promted with image. You can click on left mouse botton and starting selection a region you would like to be inpainted (delited). Once you finish marking the region, click on left mouse botton again and press esc botton on your keyboard. The inpainting proccess will start. Note: If you are running it on CPU it might take a large chank of time. Also make sure to have abount 15GB of RAM to make proccess faster (other wise swapping will ckick in and everything will be slower) ''' backends = (cv.dnn.DNN_BACKEND_DEFAULT, cv.dnn.DNN_BACKEND_OPENCV, cv.dnn.DNN_BACKEND_INFERENCE_ENGINE, cv.dnn.DNN_BACKEND_CUDA) targets = (cv.dnn.DNN_TARGET_CPU, cv.dnn.DNN_TARGET_OPENCL, cv.dnn.DNN_TARGET_MYRIAD, cv.dnn.DNN_TARGET_HDDL, cv.dnn.DNN_TARGET_CUDA) parser = argparse.ArgumentParser(description='Use this script to run inpainting using Latent Diffusion Model', formatter_class=argparse.ArgumentDefaultsHelpFormatter) parser.add_argument('--encoder', '-e', type=str, help='Path to encoder network.', default=None) parser.add_argument('--decoder', '-d', type=str, help='Path to decoder network.', default=None) parser.add_argument('--diffusor', '-df', type=str, help='Path to diffusion network.', default=None) parser.add_argument('--image', '-i', type=str, help='Path to input image.', default=None) parser.add_argument('--samples', '-s', type=int, help='Number of times to sample the model.', default=50) parser.add_argument('--backend', choices=backends, default=cv.dnn.DNN_BACKEND_DEFAULT, type=int, help="Choose one of computation backends: " "%d: automatically (by default), " "%d: OpenCV implementation, " "%d: Intel's Deep Learning Inference Engine (https://software.intel.com/openvino-toolkit), " "%d: CUDA, " % backends) parser.add_argument('--target', choices=targets, default=cv.dnn.DNN_TARGET_CPU, type=int, help='Choose one of target computation devices: ' '%d: CPU target (by default), ' '%d: OpenCL, ' '%d: NCS2 VPU, ' '%d: HDDL VPU, ' '%d: CUDA ' % targets) def make_batch(image, mask): image = image.astype(np.float32)/255.0 image = image[np.newaxis, ...].transpose(0,3,1,2) mask = mask.astype(np.float32)/255.0 mask = mask[np.newaxis, np.newaxis, ...] mask[mask < 0.5] = 0 mask[mask >= 0.5] = 1 masked_image = (1-mask)*image batch = {"image": image, "mask": mask, "masked_image": masked_image} for k in batch: batch[k] = batch[k]*2.0-1.0 return batch def noise_like(shape, repeat=False): repeat_noise = lambda: np.random.randn((1, *shape[1:])).repeat(shape[0], *((1,) * (len(shape) - 1))) noise = lambda: np.random.randn(*shape) return repeat_noise() if repeat else noise() def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True): if ddim_discr_method == 'uniform': c = num_ddpm_timesteps // num_ddim_timesteps ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c))) elif ddim_discr_method == 'quad': ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int) else: raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"') # assert ddim_timesteps.shape[0] == num_ddim_timesteps # add one to get the final alpha values right (the ones from first scale to data during sampling) steps_out = ddim_timesteps + 1 if verbose: print(f'Selected timesteps for ddim sampler: {steps_out}') return steps_out def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True): # select alphas for computing the variance schedule alphas = alphacums[ddim_timesteps] alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist()) # according the the formula provided in https://arxiv.org/abs/2010.02502 sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev)) if verbose: print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}') print(f'For the chosen value of eta, which is {eta}, ' f'this results in the following sigma_t schedule for ddim sampler {sigmas}') return sigmas, alphas, alphas_prev def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): if schedule == "linear": betas = ( np.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep).astype(np.float64) ** 2 ) elif schedule == "cosine": timesteps = ( np.arange(n_timestep + 1).astype(np.float64) / n_timestep + cosine_s ) alphas = timesteps / (1 + cosine_s) * np.pi / 2 alphas = np.cos(alphas).pow(2) alphas = alphas / alphas[0] betas = 1 - alphas[1:] / alphas[:-1] betas = np.clip(betas, a_min=0, a_max=0.999) elif schedule == "sqrt_linear": betas = np.linspace(linear_start, linear_end, n_timestep).astype(np.float64) elif schedule == "sqrt": betas = np.linspace(linear_start, linear_end, n_timestep).astype(np.float64) ** 0.5 else: raise ValueError(f"schedule '{schedule}' unknown.") return betas class DDIMSampler(object): def __init__(self, model, schedule="linear", ddpm_num_timesteps=1000): super().__init__() self.model = model self.ddpm_num_timesteps = ddpm_num_timesteps self.schedule = schedule def register_buffer(self, name, attr): setattr(self, name, attr) def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) alphas_cumprod = self.model.alphas_cumprod assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' to_numpy = partial(np.array, copy=True, dtype=np.float32) self.register_buffer('betas', to_numpy(self.model.betas)) self.register_buffer('alphas_cumprod', to_numpy(alphas_cumprod)) self.register_buffer('alphas_cumprod_prev', to_numpy(self.model.alphas_cumprod_prev)) # calculations for diffusion q(x_t | x_{t-1}) and others self.register_buffer('sqrt_alphas_cumprod', to_numpy(np.sqrt(alphas_cumprod))) self.register_buffer('sqrt_one_minus_alphas_cumprod', to_numpy(np.sqrt(1. - alphas_cumprod))) self.register_buffer('log_one_minus_alphas_cumprod', to_numpy(np.log(1. - alphas_cumprod))) self.register_buffer('sqrt_recip_alphas_cumprod', to_numpy(np.sqrt(1. / alphas_cumprod))) self.register_buffer('sqrt_recipm1_alphas_cumprod', to_numpy(np.sqrt(1. / alphas_cumprod - 1))) # ddim sampling parameters ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod, ddim_timesteps=self.ddim_timesteps, eta=ddim_eta,verbose=verbose) self.register_buffer('ddim_sigmas', ddim_sigmas) self.register_buffer('ddim_alphas', ddim_alphas) self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) sigmas_for_original_sampling_steps = ddim_eta * np.sqrt( (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) def sample(self, S, batch_size, shape, conditioning=None, eta=0., temperature=1., verbose=True, x_T=None, log_every_t=100, unconditional_guidance_scale=1., unconditional_conditioning=None, # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... **kwargs ): if conditioning is not None: if isinstance(conditioning, dict): cbs = conditioning[list(conditioning.keys())[0]].shape[0] if cbs != batch_size: print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") else: if conditioning.shape[0] != batch_size: print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) # sampling C, H, W = shape size = (batch_size, C, H, W) print(f'Data shape for DDIM sampling is {size}, eta {eta}') samples, intermediates = self.ddim_sampling(conditioning, size, ddim_use_original_steps=False, temperature=temperature, x_T=x_T, log_every_t=log_every_t, unconditional_guidance_scale=unconditional_guidance_scale, unconditional_conditioning=unconditional_conditioning, ) return samples, intermediates def ddim_sampling(self, cond, shape, x_T=None, ddim_use_original_steps=False, timesteps=None,log_every_t=100, temperature=1., unconditional_guidance_scale=1., unconditional_conditioning=None,): b = shape[0] if x_T is None: img = np.random.randn(*shape) else: img = x_T if timesteps is None: timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps elif timesteps is not None and not ddim_use_original_steps: subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 timesteps = self.ddim_timesteps[:subset_end] intermediates = {'x_inter': [img], 'pred_x0': [img]} time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps) total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] print(f"Running DDIM Sampling with {total_steps} timesteps") iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps) for i, step in enumerate(iterator): index = total_steps - i - 1 ts = np.full((b, ), step, dtype=np.int64) outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, temperature=temperature, unconditional_guidance_scale=unconditional_guidance_scale, unconditional_conditioning=unconditional_conditioning) img, pred_x0 = outs if index % log_every_t == 0 or index == total_steps - 1: intermediates['x_inter'].append(img) intermediates['pred_x0'].append(pred_x0) return img, intermediates def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, temperature=1., unconditional_guidance_scale=1., unconditional_conditioning=None): b = x.shape[0] if unconditional_conditioning is None or unconditional_guidance_scale == 1.: e_t = self.model.apply_model(x, t, c) alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas # select parameters corresponding to the currently considered timestep a_t = np.full((b, 1, 1, 1), alphas[index]) a_prev = np.full((b, 1, 1, 1), alphas_prev[index]) sigma_t = np.full((b, 1, 1, 1), sigmas[index]) sqrt_one_minus_at = np.full((b, 1, 1, 1), sqrt_one_minus_alphas[index]) # current prediction for x_0 pred_x0 = (x - sqrt_one_minus_at * e_t) / np.sqrt(a_t) # direction pointing to x_t dir_xt = np.sqrt(1. - a_prev - sigma_t**2) * e_t noise = sigma_t * noise_like(x.shape, repeat_noise) * temperature x_prev = np.sqrt(a_prev) * pred_x0 + dir_xt + noise return x_prev, pred_x0 class DDIMInpainter(object): def __init__(self, args, v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta parameterization="eps", # all assuming fixed variance schedules linear_start=0.0015, linear_end=0.0205, conditioning_key="concat", ): super().__init__() self.v_posterior = v_posterior self.parameterization = parameterization self.conditioning_key = conditioning_key self.register_schedule(linear_start=linear_start, linear_end=linear_end) self.encoder = cv.dnn.readNet(args.encoder) self.decoder = cv.dnn.readNet(args.decoder) self.diffusor = cv.dnn.readNet(args.diffusor) self.sampler = DDIMSampler(self, ddpm_num_timesteps=self.num_timesteps) self.set_backend(backend=args.backend, target=args.target) def set_backend(self, backend=cv.dnn.DNN_BACKEND_DEFAULT, target=cv.dnn.DNN_TARGET_CPU): self.encoder.setPreferableBackend(backend) self.encoder.setPreferableTarget(target) self.decoder.setPreferableBackend(backend) self.decoder.setPreferableTarget(target) self.diffusor.setPreferableBackend(backend) self.diffusor.setPreferableTarget(target) def apply_diffusor(self, x, timestep, cond): x = np.concatenate([x, cond], axis=1) x = cv.Mat(x.astype(np.float32)) timestep = cv.Mat(timestep.astype(np.int64)) names = ["xc", "t"] self.diffusor.setInputsNames(names) self.diffusor.setInput(x, names[0]) self.diffusor.setInput(timestep, names[1]) output = self.diffusor.forward() return output def register_buffer(self, name, attr): setattr(self, name, attr) def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): if given_betas is not None: betas = given_betas else: betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s) alphas = 1. - betas alphas_cumprod = np.cumprod(alphas, axis=0) alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) timesteps, = betas.shape self.num_timesteps = int(timesteps) self.linear_start = linear_start self.linear_end = linear_end assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' to_numpy = partial(np.array, dtype=np.float32) self.register_buffer('betas', to_numpy(betas)) self.register_buffer('alphas_cumprod', to_numpy(alphas_cumprod)) self.register_buffer('alphas_cumprod_prev', to_numpy(alphas_cumprod_prev)) # calculations for diffusion q(x_t | x_{t-1}) and others self.register_buffer('sqrt_alphas_cumprod', to_numpy(np.sqrt(alphas_cumprod))) self.register_buffer('sqrt_one_minus_alphas_cumprod', to_numpy(np.sqrt(1. - alphas_cumprod))) self.register_buffer('log_one_minus_alphas_cumprod', to_numpy(np.log(1. - alphas_cumprod))) self.register_buffer('sqrt_recip_alphas_cumprod', to_numpy(np.sqrt(1. / alphas_cumprod))) self.register_buffer('sqrt_recipm1_alphas_cumprod', to_numpy(np.sqrt(1. / alphas_cumprod - 1))) # calculations for posterior q(x_{t-1} | x_t, x_0) posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / ( 1. - alphas_cumprod) + self.v_posterior * betas # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t) self.register_buffer('posterior_variance', to_numpy(posterior_variance)) # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain self.register_buffer('posterior_log_variance_clipped', to_numpy(np.log(np.maximum(posterior_variance, 1e-20)))) self.register_buffer('posterior_mean_coef1', to_numpy( betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))) self.register_buffer('posterior_mean_coef2', to_numpy( (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod))) if self.parameterization == "eps": lvlb_weights = self.betas ** 2 / ( 2 * self.posterior_variance * to_numpy(alphas) * (1 - self.alphas_cumprod)) elif self.parameterization == "x0": lvlb_weights = 0.5 * np.sqrt(alphas_cumprod) / (2. * 1 - alphas_cumprod) else: raise NotImplementedError("mu not supported") # TODO how to choose this term lvlb_weights[0] = lvlb_weights[1] self.register_buffer('lvlb_weights', lvlb_weights) assert not np.isnan(self.lvlb_weights).all() def apply_model(self, x_noisy, t, cond, return_ids=False): if isinstance(cond, dict): # hybrid case, cond is exptected to be a dict pass else: # if not isinstance(cond, list): # cond = [cond] key = 'c_concat' if self.conditioning_key == 'concat' else 'c_crossattn' cond = {key: cond} x_recon = self.apply_diffusor(x_noisy, t, cond['c_concat']) if isinstance(x_recon, tuple) and not return_ids: return x_recon[0] else: return x_recon def __call__(self, image : np.ndarray, mask : np.ndarray, S : int = 50) -> np.ndarray: # Encode the image and mask self.encoder.setInput(image) c = self.encoder.forward() cc = cv.resize(np.squeeze(mask), dsize=(c.shape[3], c.shape[2]), interpolation=cv.INTER_NEAREST) #TODO:check for correcteness of intepolation cc = cc[None,None] c = np.concatenate([c, cc], axis=1) shape = (c.shape[1] - 1,) + c.shape[2:] # Sample from the model samples_ddim, _ = self.sampler.sample( S=S, conditioning=c, batch_size=c.shape[0], shape=shape, verbose=False) ## Decode the sample samples_ddim = samples_ddim.astype(np.float32) samples_ddim = cv.Mat(samples_ddim) self.decoder.setInput(samples_ddim) x_samples_ddim = self.decoder.forward() image = np.clip((image + 1.0) / 2.0, a_min=0.0, a_max=1.0) mask = np.clip((mask + 1.0) / 2.0, a_min=0.0, a_max=1.0) predicted_image = np.clip((x_samples_ddim + 1.0) / 2.0, a_min=0.0, a_max=1.0) inpainted = (1 - mask) * image + mask * predicted_image inpainted = np.transpose(inpainted, (0, 2, 3, 1)) * 255 return inpainted def create_mask(img, radius=20): drawing = False # True if the mouse is pressed counter = 0 # Mouse callback function def draw_circle(event, x, y, flags, param): nonlocal drawing, counter, radius if event == cv.EVENT_LBUTTONDOWN: drawing = True if counter % 2 == 0 else False counter += 1 cv.circle(img, (x, y), radius, (255, 255, 255), -1) cv.circle(mask, (x, y), radius, 255, -1) elif event == cv.EVENT_MOUSEMOVE: if drawing: cv.circle(img, (x, y), radius, (255, 255, 255), -1) cv.circle(mask, (x, y), radius, 255, -1) mask = np.zeros((img.shape[0], img.shape[1]), np.uint8) cv.namedWindow('image') cv.setMouseCallback('image', draw_circle) while True: cv.imshow('image', img) if cv.waitKey(1) & 0xFF == 27: # Press 'ESC' to exit break cv.destroyAllWindows() return mask def main(args): image = cv.imread(args.image) mask = create_mask(deepcopy(image)) image = cv.cvtColor(image, cv.COLOR_BGR2RGB) batch = make_batch(image, mask) image, mask, masked_image = batch["image"], batch["mask"], batch["masked_image"] model = DDIMInpainter(args) result = model(masked_image, mask, S=args.samples) result = np.squeeze(result) # save the result in the directore of args.image cv.imwrite(args.image.replace(".png", "_inpainted.png"), result[..., ::-1]) if __name__ == '__main__': args = parser.parse_args() main(args)