extinsions-builtin

extensions-builtin/LDSR/ldsr_model_arch.py

@@ -0,0 +1,253 @@
+import os
+import gc
+import time
+
+import numpy as np
+import torch
+import torchvision
+from PIL import Image
+from einops import rearrange, repeat
+from omegaconf import OmegaConf
+import safetensors.torch
+
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.util import instantiate_from_config, ismap
+from modules import shared, sd_hijack
+
+cached_ldsr_model: torch.nn.Module = None
+
+
+# Create LDSR Class
+class LDSR:
+    def load_model_from_config(self, half_attention):
+        global cached_ldsr_model
+
+        if shared.opts.ldsr_cached and cached_ldsr_model is not None:
+            print("Loading model from cache")
+            model: torch.nn.Module = cached_ldsr_model
+        else:
+            print(f"Loading model from {self.modelPath}")
+            _, extension = os.path.splitext(self.modelPath)
+            if extension.lower() == ".safetensors":
+                pl_sd = safetensors.torch.load_file(self.modelPath, device="cpu")
+            else:
+                pl_sd = torch.load(self.modelPath, map_location="cpu")
+            sd = pl_sd["state_dict"] if "state_dict" in pl_sd else pl_sd
+            config = OmegaConf.load(self.yamlPath)
+            config.model.target = "ldm.models.diffusion.ddpm.LatentDiffusionV1"
+            model: torch.nn.Module = instantiate_from_config(config.model)
+            model.load_state_dict(sd, strict=False)
+            model = model.to(shared.device)
+            if half_attention:
+                model = model.half()
+            if shared.cmd_opts.opt_channelslast:
+                model = model.to(memory_format=torch.channels_last)
+
+            sd_hijack.model_hijack.hijack(model) # apply optimization
+            model.eval()
+
+            if shared.opts.ldsr_cached:
+                cached_ldsr_model = model
+
+        return {"model": model}
+
+    def __init__(self, model_path, yaml_path):
+        self.modelPath = model_path
+        self.yamlPath = yaml_path
+
+    @staticmethod
+    def run(model, selected_path, custom_steps, eta):
+        example = get_cond(selected_path)
+
+        n_runs = 1
+        guider = None
+        ckwargs = None
+        ddim_use_x0_pred = False
+        temperature = 1.
+        eta = eta
+        custom_shape = None
+
+        height, width = example["image"].shape[1:3]
+        split_input = height >= 128 and width >= 128
+
+        if split_input:
+            ks = 128
+            stride = 64
+            vqf = 4  #
+            model.split_input_params = {"ks": (ks, ks), "stride": (stride, stride),
+                                        "vqf": vqf,
+                                        "patch_distributed_vq": True,
+                                        "tie_braker": False,
+                                        "clip_max_weight": 0.5,
+                                        "clip_min_weight": 0.01,
+                                        "clip_max_tie_weight": 0.5,
+                                        "clip_min_tie_weight": 0.01}
+        else:
+            if hasattr(model, "split_input_params"):
+                delattr(model, "split_input_params")
+
+        x_t = None
+        logs = None
+        for n in range(n_runs):
+            if custom_shape is not None:
+                x_t = torch.randn(1, custom_shape[1], custom_shape[2], custom_shape[3]).to(model.device)
+                x_t = repeat(x_t, '1 c h w -> b c h w', b=custom_shape[0])
+
+            logs = make_convolutional_sample(example, model,
+                                             custom_steps=custom_steps,
+                                             eta=eta, quantize_x0=False,
+                                             custom_shape=custom_shape,
+                                             temperature=temperature, noise_dropout=0.,
+                                             corrector=guider, corrector_kwargs=ckwargs, x_T=x_t,
+                                             ddim_use_x0_pred=ddim_use_x0_pred
+                                             )
+        return logs
+
+    def super_resolution(self, image, steps=100, target_scale=2, half_attention=False):
+        model = self.load_model_from_config(half_attention)
+
+        # Run settings
+        diffusion_steps = int(steps)
+        eta = 1.0
+
+        down_sample_method = 'Lanczos'
+
+        gc.collect()
+        if torch.cuda.is_available:
+            torch.cuda.empty_cache()
+
+        im_og = image
+        width_og, height_og = im_og.size
+        # If we can adjust the max upscale size, then the 4 below should be our variable
+        down_sample_rate = target_scale / 4
+        wd = width_og * down_sample_rate
+        hd = height_og * down_sample_rate
+        width_downsampled_pre = int(np.ceil(wd))
+        height_downsampled_pre = int(np.ceil(hd))
+
+        if down_sample_rate != 1:
+            print(
+                f'Downsampling from [{width_og}, {height_og}] to [{width_downsampled_pre}, {height_downsampled_pre}]')
+            im_og = im_og.resize((width_downsampled_pre, height_downsampled_pre), Image.LANCZOS)
+        else:
+            print(f"Down sample rate is 1 from {target_scale} / 4 (Not downsampling)")
+        
+        # pad width and height to multiples of 64, pads with the edge values of image to avoid artifacts
+        pad_w, pad_h = np.max(((2, 2), np.ceil(np.array(im_og.size) / 64).astype(int)), axis=0) * 64 - im_og.size
+        im_padded = Image.fromarray(np.pad(np.array(im_og), ((0, pad_h), (0, pad_w), (0, 0)), mode='edge'))
+        
+        logs = self.run(model["model"], im_padded, diffusion_steps, eta)
+
+        sample = logs["sample"]
+        sample = sample.detach().cpu()
+        sample = torch.clamp(sample, -1., 1.)
+        sample = (sample + 1.) / 2. * 255
+        sample = sample.numpy().astype(np.uint8)
+        sample = np.transpose(sample, (0, 2, 3, 1))
+        a = Image.fromarray(sample[0])
+
+        # remove padding
+        a = a.crop((0, 0) + tuple(np.array(im_og.size) * 4))
+
+        del model
+        gc.collect()
+        if torch.cuda.is_available:
+            torch.cuda.empty_cache()
+
+        return a
+
+
+def get_cond(selected_path):
+    example = dict()
+    up_f = 4
+    c = selected_path.convert('RGB')
+    c = torch.unsqueeze(torchvision.transforms.ToTensor()(c), 0)
+    c_up = torchvision.transforms.functional.resize(c, size=[up_f * c.shape[2], up_f * c.shape[3]],
+                                                    antialias=True)
+    c_up = rearrange(c_up, '1 c h w -> 1 h w c')
+    c = rearrange(c, '1 c h w -> 1 h w c')
+    c = 2. * c - 1.
+
+    c = c.to(shared.device)
+    example["LR_image"] = c
+    example["image"] = c_up
+
+    return example
+
+
+@torch.no_grad()
+def convsample_ddim(model, cond, steps, shape, eta=1.0, callback=None, normals_sequence=None,
+                    mask=None, x0=None, quantize_x0=False, temperature=1., score_corrector=None,
+                    corrector_kwargs=None, x_t=None
+                    ):
+    ddim = DDIMSampler(model)
+    bs = shape[0]
+    shape = shape[1:]
+    print(f"Sampling with eta = {eta}; steps: {steps}")
+    samples, intermediates = ddim.sample(steps, batch_size=bs, shape=shape, conditioning=cond, callback=callback,
+                                         normals_sequence=normals_sequence, quantize_x0=quantize_x0, eta=eta,
+                                         mask=mask, x0=x0, temperature=temperature, verbose=False,
+                                         score_corrector=score_corrector,
+                                         corrector_kwargs=corrector_kwargs, x_t=x_t)
+
+    return samples, intermediates
+
+
+@torch.no_grad()
+def make_convolutional_sample(batch, model, custom_steps=None, eta=1.0, quantize_x0=False, custom_shape=None, temperature=1., noise_dropout=0., corrector=None,
+                              corrector_kwargs=None, x_T=None, ddim_use_x0_pred=False):
+    log = dict()
+
+    z, c, x, xrec, xc = model.get_input(batch, model.first_stage_key,
+                                        return_first_stage_outputs=True,
+                                        force_c_encode=not (hasattr(model, 'split_input_params')
+                                                            and model.cond_stage_key == 'coordinates_bbox'),
+                                        return_original_cond=True)
+
+    if custom_shape is not None:
+        z = torch.randn(custom_shape)
+        print(f"Generating {custom_shape[0]} samples of shape {custom_shape[1:]}")
+
+    z0 = None
+
+    log["input"] = x
+    log["reconstruction"] = xrec
+
+    if ismap(xc):
+        log["original_conditioning"] = model.to_rgb(xc)
+        if hasattr(model, 'cond_stage_key'):
+            log[model.cond_stage_key] = model.to_rgb(xc)
+
+    else:
+        log["original_conditioning"] = xc if xc is not None else torch.zeros_like(x)
+        if model.cond_stage_model:
+            log[model.cond_stage_key] = xc if xc is not None else torch.zeros_like(x)
+            if model.cond_stage_key == 'class_label':
+                log[model.cond_stage_key] = xc[model.cond_stage_key]
+
+    with model.ema_scope("Plotting"):
+        t0 = time.time()
+
+        sample, intermediates = convsample_ddim(model, c, steps=custom_steps, shape=z.shape,
+                                                eta=eta,
+                                                quantize_x0=quantize_x0, mask=None, x0=z0,
+                                                temperature=temperature, score_corrector=corrector, corrector_kwargs=corrector_kwargs,
+                                                x_t=x_T)
+        t1 = time.time()
+
+        if ddim_use_x0_pred:
+            sample = intermediates['pred_x0'][-1]
+
+    x_sample = model.decode_first_stage(sample)
+
+    try:
+        x_sample_noquant = model.decode_first_stage(sample, force_not_quantize=True)
+        log["sample_noquant"] = x_sample_noquant
+        log["sample_diff"] = torch.abs(x_sample_noquant - x_sample)
+    except:
+        pass
+
+    log["sample"] = x_sample
+    log["time"] = t1 - t0
+
+    return log

extensions-builtin/LDSR/preload.py

@@ -0,0 +1,6 @@
+import os
+from modules import paths
+
+
+def preload(parser):
+    parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with LDSR model file(s).", default=os.path.join(paths.models_path, 'LDSR'))

extensions-builtin/LDSR/scripts/ldsr_model.py

@@ -0,0 +1,69 @@
+import os
+import sys
+import traceback
+
+from basicsr.utils.download_util import load_file_from_url
+
+from modules.upscaler import Upscaler, UpscalerData
+from ldsr_model_arch import LDSR
+from modules import shared, script_callbacks
+import sd_hijack_autoencoder, sd_hijack_ddpm_v1
+
+
+class UpscalerLDSR(Upscaler):
+    def __init__(self, user_path):
+        self.name = "LDSR"
+        self.user_path = user_path
+        self.model_url = "https://heibox.uni-heidelberg.de/f/578df07c8fc04ffbadf3/?dl=1"
+        self.yaml_url = "https://heibox.uni-heidelberg.de/f/31a76b13ea27482981b4/?dl=1"
+        super().__init__()
+        scaler_data = UpscalerData("LDSR", None, self)
+        self.scalers = [scaler_data]
+
+    def load_model(self, path: str):
+        # Remove incorrect project.yaml file if too big
+        yaml_path = os.path.join(self.model_path, "project.yaml")
+        old_model_path = os.path.join(self.model_path, "model.pth")
+        new_model_path = os.path.join(self.model_path, "model.ckpt")
+        safetensors_model_path = os.path.join(self.model_path, "model.safetensors")
+        if os.path.exists(yaml_path):
+            statinfo = os.stat(yaml_path)
+            if statinfo.st_size >= 10485760:
+                print("Removing invalid LDSR YAML file.")
+                os.remove(yaml_path)
+        if os.path.exists(old_model_path):
+            print("Renaming model from model.pth to model.ckpt")
+            os.rename(old_model_path, new_model_path)
+        if os.path.exists(safetensors_model_path):
+            model = safetensors_model_path
+        else:
+            model = load_file_from_url(url=self.model_url, model_dir=self.model_path,
+                                       file_name="model.ckpt", progress=True)
+        yaml = load_file_from_url(url=self.yaml_url, model_dir=self.model_path,
+                                  file_name="project.yaml", progress=True)
+
+        try:
+            return LDSR(model, yaml)
+
+        except Exception:
+            print("Error importing LDSR:", file=sys.stderr)
+            print(traceback.format_exc(), file=sys.stderr)
+        return None
+
+    def do_upscale(self, img, path):
+        ldsr = self.load_model(path)
+        if ldsr is None:
+            print("NO LDSR!")
+            return img
+        ddim_steps = shared.opts.ldsr_steps
+        return ldsr.super_resolution(img, ddim_steps, self.scale)
+
+
+def on_ui_settings():
+    import gradio as gr
+
+    shared.opts.add_option("ldsr_steps", shared.OptionInfo(100, "LDSR processing steps. Lower = faster", gr.Slider, {"minimum": 1, "maximum": 200, "step": 1}, section=('upscaling', "Upscaling")))
+    shared.opts.add_option("ldsr_cached", shared.OptionInfo(False, "Cache LDSR model in memory", gr.Checkbox, {"interactive": True}, section=('upscaling', "Upscaling")))
+
+
+script_callbacks.on_ui_settings(on_ui_settings)

extensions-builtin/LDSR/sd_hijack_autoencoder.py

@@ -0,0 +1,286 @@
+# The content of this file comes from the ldm/models/autoencoder.py file of the compvis/stable-diffusion repo
+# The VQModel & VQModelInterface were subsequently removed from ldm/models/autoencoder.py when we moved to the stability-ai/stablediffusion repo
+# As the LDSR upscaler relies on VQModel & VQModelInterface, the hijack aims to put them back into the ldm.models.autoencoder
+
+import torch
+import pytorch_lightning as pl
+import torch.nn.functional as F
+from contextlib import contextmanager
+from taming.modules.vqvae.quantize import VectorQuantizer2 as VectorQuantizer
+from ldm.modules.diffusionmodules.model import Encoder, Decoder
+from ldm.util import instantiate_from_config
+
+import ldm.models.autoencoder
+
+class VQModel(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 n_embed,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 batch_resize_range=None,
+                 scheduler_config=None,
+                 lr_g_factor=1.0,
+                 remap=None,
+                 sane_index_shape=False, # tell vector quantizer to return indices as bhw
+                 use_ema=False
+                 ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.n_embed = n_embed
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25,
+                                        remap=remap,
+                                        sane_index_shape=sane_index_shape)
+        self.quant_conv = torch.nn.Conv2d(ddconfig["z_channels"], embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+        self.batch_resize_range = batch_resize_range
+        if self.batch_resize_range is not None:
+            print(f"{self.__class__.__name__}: Using per-batch resizing in range {batch_resize_range}.")
+
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self)
+            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+        self.scheduler_config = scheduler_config
+        self.lr_g_factor = lr_g_factor
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.parameters())
+            self.model_ema.copy_to(self)
+            if context is not None:
+                print(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.parameters())
+                if context is not None:
+                    print(f"{context}: Restored training weights")
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+            print(f"Unexpected Keys: {unexpected}")
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self)
+
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        quant, emb_loss, info = self.quantize(h)
+        return quant, emb_loss, info
+
+    def encode_to_prequant(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, quant):
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec
+
+    def decode_code(self, code_b):
+        quant_b = self.quantize.embed_code(code_b)
+        dec = self.decode(quant_b)
+        return dec
+
+    def forward(self, input, return_pred_indices=False):
+        quant, diff, (_,_,ind) = self.encode(input)
+        dec = self.decode(quant)
+        if return_pred_indices:
+            return dec, diff, ind
+        return dec, diff
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
+        if self.batch_resize_range is not None:
+            lower_size = self.batch_resize_range[0]
+            upper_size = self.batch_resize_range[1]
+            if self.global_step <= 4:
+                # do the first few batches with max size to avoid later oom
+                new_resize = upper_size
+            else:
+                new_resize = np.random.choice(np.arange(lower_size, upper_size+16, 16))
+            if new_resize != x.shape[2]:
+                x = F.interpolate(x, size=new_resize, mode="bicubic")
+            x = x.detach()
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        # https://github.com/pytorch/pytorch/issues/37142
+        # try not to fool the heuristics
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+
+        if optimizer_idx == 0:
+            # autoencode
+            aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train",
+                                            predicted_indices=ind)
+
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # discriminator
+            discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        log_dict = self._validation_step(batch, batch_idx)
+        with self.ema_scope():
+            log_dict_ema = self._validation_step(batch, batch_idx, suffix="_ema")
+        return log_dict
+
+    def _validation_step(self, batch, batch_idx, suffix=""):
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+        aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0,
+                                        self.global_step,
+                                        last_layer=self.get_last_layer(),
+                                        split="val"+suffix,
+                                        predicted_indices=ind
+                                        )
+
+        discloss, log_dict_disc = self.loss(qloss, x, xrec, 1,
+                                            self.global_step,
+                                            last_layer=self.get_last_layer(),
+                                            split="val"+suffix,
+                                            predicted_indices=ind
+                                            )
+        rec_loss = log_dict_ae[f"val{suffix}/rec_loss"]
+        self.log(f"val{suffix}/rec_loss", rec_loss,
+                   prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
+        self.log(f"val{suffix}/aeloss", aeloss,
+                   prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
+        if version.parse(pl.__version__) >= version.parse('1.4.0'):
+            del log_dict_ae[f"val{suffix}/rec_loss"]
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+
+    def configure_optimizers(self):
+        lr_d = self.learning_rate
+        lr_g = self.lr_g_factor*self.learning_rate
+        print("lr_d", lr_d)
+        print("lr_g", lr_g)
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quantize.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr_g, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr_d, betas=(0.5, 0.9))
+
+        if self.scheduler_config is not None:
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(opt_ae, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                },
+                {
+                    'scheduler': LambdaLR(opt_disc, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                },
+            ]
+            return [opt_ae, opt_disc], scheduler
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    def log_images(self, batch, only_inputs=False, plot_ema=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if only_inputs:
+            log["inputs"] = x
+            return log
+        xrec, _ = self(x)
+        if x.shape[1] > 3:
+            # colorize with random projection
+            assert xrec.shape[1] > 3
+            x = self.to_rgb(x)
+            xrec = self.to_rgb(xrec)
+        log["inputs"] = x
+        log["reconstructions"] = xrec
+        if plot_ema:
+            with self.ema_scope():
+                xrec_ema, _ = self(x)
+                if x.shape[1] > 3: xrec_ema = self.to_rgb(xrec_ema)
+                log["reconstructions_ema"] = xrec_ema
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+
+class VQModelInterface(VQModel):
+    def __init__(self, embed_dim, *args, **kwargs):
+        super().__init__(embed_dim=embed_dim, *args, **kwargs)
+        self.embed_dim = embed_dim
+
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, h, force_not_quantize=False):
+        # also go through quantization layer
+        if not force_not_quantize:
+            quant, emb_loss, info = self.quantize(h)
+        else:
+            quant = h
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec
+
+setattr(ldm.models.autoencoder, "VQModel", VQModel)
+setattr(ldm.models.autoencoder, "VQModelInterface", VQModelInterface)

extensions-builtin/LDSR/sd_hijack_ddpm_v1.py

@@ -0,0 +1,1449 @@
+# This script is copied from the compvis/stable-diffusion repo (aka the SD V1 repo)
+# Original filename: ldm/models/diffusion/ddpm.py
+# The purpose to reinstate the old DDPM logic which works with VQ, whereas the V2 one doesn't
+# Some models such as LDSR require VQ to work correctly
+# The classes are suffixed with "V1" and added back to the "ldm.models.diffusion.ddpm" module
+
+import torch
+import torch.nn as nn
+import numpy as np
+import pytorch_lightning as pl
+from torch.optim.lr_scheduler import LambdaLR
+from einops import rearrange, repeat
+from contextlib import contextmanager
+from functools import partial
+from tqdm import tqdm
+from torchvision.utils import make_grid
+from pytorch_lightning.utilities.distributed import rank_zero_only
+
+from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config
+from ldm.modules.ema import LitEma
+from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution
+from ldm.models.autoencoder import VQModelInterface, IdentityFirstStage, AutoencoderKL
+from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
+from ldm.models.diffusion.ddim import DDIMSampler
+
+import ldm.models.diffusion.ddpm
+
+__conditioning_keys__ = {'concat': 'c_concat',
+                         'crossattn': 'c_crossattn',
+                         'adm': 'y'}
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+def uniform_on_device(r1, r2, shape, device):
+    return (r1 - r2) * torch.rand(*shape, device=device) + r2
+
+
+class DDPMV1(pl.LightningModule):
+    # classic DDPM with Gaussian diffusion, in image space
+    def __init__(self,
+                 unet_config,
+                 timesteps=1000,
+                 beta_schedule="linear",
+                 loss_type="l2",
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 load_only_unet=False,
+                 monitor="val/loss",
+                 use_ema=True,
+                 first_stage_key="image",
+                 image_size=256,
+                 channels=3,
+                 log_every_t=100,
+                 clip_denoised=True,
+                 linear_start=1e-4,
+                 linear_end=2e-2,
+                 cosine_s=8e-3,
+                 given_betas=None,
+                 original_elbo_weight=0.,
+                 v_posterior=0.,  # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
+                 l_simple_weight=1.,
+                 conditioning_key=None,
+                 parameterization="eps",  # all assuming fixed variance schedules
+                 scheduler_config=None,
+                 use_positional_encodings=False,
+                 learn_logvar=False,
+                 logvar_init=0.,
+                 ):
+        super().__init__()
+        assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"'
+        self.parameterization = parameterization
+        print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
+        self.cond_stage_model = None
+        self.clip_denoised = clip_denoised
+        self.log_every_t = log_every_t
+        self.first_stage_key = first_stage_key
+        self.image_size = image_size  # try conv?
+        self.channels = channels
+        self.use_positional_encodings = use_positional_encodings
+        self.model = DiffusionWrapperV1(unet_config, conditioning_key)
+        count_params(self.model, verbose=True)
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self.model)
+            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        self.use_scheduler = scheduler_config is not None
+        if self.use_scheduler:
+            self.scheduler_config = scheduler_config
+
+        self.v_posterior = v_posterior
+        self.original_elbo_weight = original_elbo_weight
+        self.l_simple_weight = l_simple_weight
+
+        if monitor is not None:
+            self.monitor = monitor
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet)
+
+        self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
+                               linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
+
+        self.loss_type = loss_type
+
+        self.learn_logvar = learn_logvar
+        self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
+        if self.learn_logvar:
+            self.logvar = nn.Parameter(self.logvar, requires_grad=True)
+
+
+    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 exists(given_betas):
+            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_torch = partial(torch.tensor, dtype=torch.float32)
+
+        self.register_buffer('betas', to_torch(betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(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_torch(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_torch(np.log(np.maximum(posterior_variance, 1e-20))))
+        self.register_buffer('posterior_mean_coef1', to_torch(
+            betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
+        self.register_buffer('posterior_mean_coef2', to_torch(
+            (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
+
+        if self.parameterization == "eps":
+            lvlb_weights = self.betas ** 2 / (
+                        2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
+        elif self.parameterization == "x0":
+            lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(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, persistent=False)
+        assert not torch.isnan(self.lvlb_weights).all()
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.model.parameters())
+            self.model_ema.copy_to(self.model)
+            if context is not None:
+                print(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.model.parameters())
+                if context is not None:
+                    print(f"{context}: Restored training weights")
+
+    def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
+        sd = torch.load(path, map_location="cpu")
+        if "state_dict" in list(sd.keys()):
+            sd = sd["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
+            sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            print(f"Unexpected Keys: {unexpected}")
+
+    def q_mean_variance(self, x_start, t):
+        """
+        Get the distribution q(x_t | x_0).
+        :param x_start: the [N x C x ...] tensor of noiseless inputs.
+        :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
+        :return: A tuple (mean, variance, log_variance), all of x_start's shape.
+        """
+        mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
+        variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
+        log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
+        return mean, variance, log_variance
+
+    def predict_start_from_noise(self, x_t, t, noise):
+        return (
+                extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
+                extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
+        )
+
+    def q_posterior(self, x_start, x_t, t):
+        posterior_mean = (
+                extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
+                extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
+        )
+        posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
+        posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
+        return posterior_mean, posterior_variance, posterior_log_variance_clipped
+
+    def p_mean_variance(self, x, t, clip_denoised: bool):
+        model_out = self.model(x, t)
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
+        b, *_, device = *x.shape, x.device
+        model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
+        noise = noise_like(x.shape, device, repeat_noise)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+        return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def p_sample_loop(self, shape, return_intermediates=False):
+        device = self.betas.device
+        b = shape[0]
+        img = torch.randn(shape, device=device)
+        intermediates = [img]
+        for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
+            img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
+                                clip_denoised=self.clip_denoised)
+            if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
+                intermediates.append(img)
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, batch_size=16, return_intermediates=False):
+        image_size = self.image_size
+        channels = self.channels
+        return self.p_sample_loop((batch_size, channels, image_size, image_size),
+                                  return_intermediates=return_intermediates)
+
+    def q_sample(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
+
+    def get_loss(self, pred, target, mean=True):
+        if self.loss_type == 'l1':
+            loss = (target - pred).abs()
+            if mean:
+                loss = loss.mean()
+        elif self.loss_type == 'l2':
+            if mean:
+                loss = torch.nn.functional.mse_loss(target, pred)
+            else:
+                loss = torch.nn.functional.mse_loss(target, pred, reduction='none')
+        else:
+            raise NotImplementedError("unknown loss type '{loss_type}'")
+
+        return loss
+
+    def p_losses(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_out = self.model(x_noisy, t)
+
+        loss_dict = {}
+        if self.parameterization == "eps":
+            target = noise
+        elif self.parameterization == "x0":
+            target = x_start
+        else:
+            raise NotImplementedError(f"Paramterization {self.parameterization} not yet supported")
+
+        loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
+
+        log_prefix = 'train' if self.training else 'val'
+
+        loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()})
+        loss_simple = loss.mean() * self.l_simple_weight
+
+        loss_vlb = (self.lvlb_weights[t] * loss).mean()
+        loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb})
+
+        loss = loss_simple + self.original_elbo_weight * loss_vlb
+
+        loss_dict.update({f'{log_prefix}/loss': loss})
+
+        return loss, loss_dict
+
+    def forward(self, x, *args, **kwargs):
+        # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
+        # assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
+        t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
+        return self.p_losses(x, t, *args, **kwargs)
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = rearrange(x, 'b h w c -> b c h w')
+        x = x.to(memory_format=torch.contiguous_format).float()
+        return x
+
+    def shared_step(self, batch):
+        x = self.get_input(batch, self.first_stage_key)
+        loss, loss_dict = self(x)
+        return loss, loss_dict
+
+    def training_step(self, batch, batch_idx):
+        loss, loss_dict = self.shared_step(batch)
+
+        self.log_dict(loss_dict, prog_bar=True,
+                      logger=True, on_step=True, on_epoch=True)
+
+        self.log("global_step", self.global_step,
+                 prog_bar=True, logger=True, on_step=True, on_epoch=False)
+
+        if self.use_scheduler:
+            lr = self.optimizers().param_groups[0]['lr']
+            self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
+
+        return loss
+
+    @torch.no_grad()
+    def validation_step(self, batch, batch_idx):
+        _, loss_dict_no_ema = self.shared_step(batch)
+        with self.ema_scope():
+            _, loss_dict_ema = self.shared_step(batch)
+            loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema}
+        self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
+        self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self.model)
+
+    def _get_rows_from_list(self, samples):
+        n_imgs_per_row = len(samples)
+        denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
+        denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
+        return denoise_grid
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.first_stage_key)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        x = x.to(self.device)[:N]
+        log["inputs"] = x
+
+        # get diffusion row
+        diffusion_row = list()
+        x_start = x[:n_row]
+
+        for t in range(self.num_timesteps):
+            if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                t = t.to(self.device).long()
+                noise = torch.randn_like(x_start)
+                x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+                diffusion_row.append(x_noisy)
+
+        log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
+
+        if sample:
+            # get denoise row
+            with self.ema_scope("Plotting"):
+                samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
+
+            log["samples"] = samples
+            log["denoise_row"] = self._get_rows_from_list(denoise_row)
+
+        if return_keys:
+            if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
+                return log
+            else:
+                return {key: log[key] for key in return_keys}
+        return log
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        params = list(self.model.parameters())
+        if self.learn_logvar:
+            params = params + [self.logvar]
+        opt = torch.optim.AdamW(params, lr=lr)
+        return opt
+
+
+class LatentDiffusionV1(DDPMV1):
+    """main class"""
+    def __init__(self,
+                 first_stage_config,
+                 cond_stage_config,
+                 num_timesteps_cond=None,
+                 cond_stage_key="image",
+                 cond_stage_trainable=False,
+                 concat_mode=True,
+                 cond_stage_forward=None,
+                 conditioning_key=None,
+                 scale_factor=1.0,
+                 scale_by_std=False,
+                 *args, **kwargs):
+        self.num_timesteps_cond = default(num_timesteps_cond, 1)
+        self.scale_by_std = scale_by_std
+        assert self.num_timesteps_cond <= kwargs['timesteps']
+        # for backwards compatibility after implementation of DiffusionWrapper
+        if conditioning_key is None:
+            conditioning_key = 'concat' if concat_mode else 'crossattn'
+        if cond_stage_config == '__is_unconditional__':
+            conditioning_key = None
+        ckpt_path = kwargs.pop("ckpt_path", None)
+        ignore_keys = kwargs.pop("ignore_keys", [])
+        super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
+        self.concat_mode = concat_mode
+        self.cond_stage_trainable = cond_stage_trainable
+        self.cond_stage_key = cond_stage_key
+        try:
+            self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
+        except:
+            self.num_downs = 0
+        if not scale_by_std:
+            self.scale_factor = scale_factor
+        else:
+            self.register_buffer('scale_factor', torch.tensor(scale_factor))
+        self.instantiate_first_stage(first_stage_config)
+        self.instantiate_cond_stage(cond_stage_config)
+        self.cond_stage_forward = cond_stage_forward
+        self.clip_denoised = False
+        self.bbox_tokenizer = None  
+
+        self.restarted_from_ckpt = False
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys)
+            self.restarted_from_ckpt = True
+
+    def make_cond_schedule(self, ):
+        self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
+        ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
+        self.cond_ids[:self.num_timesteps_cond] = ids
+
+    @rank_zero_only
+    @torch.no_grad()
+    def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
+        # only for very first batch
+        if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
+            assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
+            # set rescale weight to 1./std of encodings
+            print("### USING STD-RESCALING ###")
+            x = super().get_input(batch, self.first_stage_key)
+            x = x.to(self.device)
+            encoder_posterior = self.encode_first_stage(x)
+            z = self.get_first_stage_encoding(encoder_posterior).detach()
+            del self.scale_factor
+            self.register_buffer('scale_factor', 1. / z.flatten().std())
+            print(f"setting self.scale_factor to {self.scale_factor}")
+            print("### USING STD-RESCALING ###")
+
+    def register_schedule(self,
+                          given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
+
+        self.shorten_cond_schedule = self.num_timesteps_cond > 1
+        if self.shorten_cond_schedule:
+            self.make_cond_schedule()
+
+    def instantiate_first_stage(self, config):
+        model = instantiate_from_config(config)
+        self.first_stage_model = model.eval()
+        self.first_stage_model.train = disabled_train
+        for param in self.first_stage_model.parameters():
+            param.requires_grad = False
+
+    def instantiate_cond_stage(self, config):
+        if not self.cond_stage_trainable:
+            if config == "__is_first_stage__":
+                print("Using first stage also as cond stage.")
+                self.cond_stage_model = self.first_stage_model
+            elif config == "__is_unconditional__":
+                print(f"Training {self.__class__.__name__} as an unconditional model.")
+                self.cond_stage_model = None
+                # self.be_unconditional = True
+            else:
+                model = instantiate_from_config(config)
+                self.cond_stage_model = model.eval()
+                self.cond_stage_model.train = disabled_train
+                for param in self.cond_stage_model.parameters():
+                    param.requires_grad = False
+        else:
+            assert config != '__is_first_stage__'
+            assert config != '__is_unconditional__'
+            model = instantiate_from_config(config)
+            self.cond_stage_model = model
+
+    def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False):
+        denoise_row = []
+        for zd in tqdm(samples, desc=desc):
+            denoise_row.append(self.decode_first_stage(zd.to(self.device),
+                                                            force_not_quantize=force_no_decoder_quantization))
+        n_imgs_per_row = len(denoise_row)
+        denoise_row = torch.stack(denoise_row)  # n_log_step, n_row, C, H, W
+        denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
+        denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
+        return denoise_grid
+
+    def get_first_stage_encoding(self, encoder_posterior):
+        if isinstance(encoder_posterior, DiagonalGaussianDistribution):
+            z = encoder_posterior.sample()
+        elif isinstance(encoder_posterior, torch.Tensor):
+            z = encoder_posterior
+        else:
+            raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
+        return self.scale_factor * z
+
+    def get_learned_conditioning(self, c):
+        if self.cond_stage_forward is None:
+            if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
+                c = self.cond_stage_model.encode(c)
+                if isinstance(c, DiagonalGaussianDistribution):
+                    c = c.mode()
+            else:
+                c = self.cond_stage_model(c)
+        else:
+            assert hasattr(self.cond_stage_model, self.cond_stage_forward)
+            c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
+        return c
+
+    def meshgrid(self, h, w):
+        y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
+        x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)
+
+        arr = torch.cat([y, x], dim=-1)
+        return arr
+
+    def delta_border(self, h, w):
+        """
+        :param h: height
+        :param w: width
+        :return: normalized distance to image border,
+         wtith min distance = 0 at border and max dist = 0.5 at image center
+        """
+        lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
+        arr = self.meshgrid(h, w) / lower_right_corner
+        dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
+        dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
+        edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0]
+        return edge_dist
+
+    def get_weighting(self, h, w, Ly, Lx, device):
+        weighting = self.delta_border(h, w)
+        weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"],
+                               self.split_input_params["clip_max_weight"], )
+        weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)
+
+        if self.split_input_params["tie_braker"]:
+            L_weighting = self.delta_border(Ly, Lx)
+            L_weighting = torch.clip(L_weighting,
+                                     self.split_input_params["clip_min_tie_weight"],
+                                     self.split_input_params["clip_max_tie_weight"])
+
+            L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
+            weighting = weighting * L_weighting
+        return weighting
+
+    def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1):  # todo load once not every time, shorten code
+        """
+        :param x: img of size (bs, c, h, w)
+        :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
+        """
+        bs, nc, h, w = x.shape
+
+        # number of crops in image
+        Ly = (h - kernel_size[0]) // stride[0] + 1
+        Lx = (w - kernel_size[1]) // stride[1] + 1
+
+        if uf == 1 and df == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)
+
+            weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h, w)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))
+
+        elif uf > 1 and df == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
+                                dilation=1, padding=0,
+                                stride=(stride[0] * uf, stride[1] * uf))
+            fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2)
+
+            weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h * uf, w * uf)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx))
+
+        elif df > 1 and uf == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
+                                dilation=1, padding=0,
+                                stride=(stride[0] // df, stride[1] // df))
+            fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2)
+
+            weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h // df, w // df)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx))
+
+        else:
+            raise NotImplementedError
+
+        return fold, unfold, normalization, weighting
+
+    @torch.no_grad()
+    def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False,
+                  cond_key=None, return_original_cond=False, bs=None):
+        x = super().get_input(batch, k)
+        if bs is not None:
+            x = x[:bs]
+        x = x.to(self.device)
+        encoder_posterior = self.encode_first_stage(x)
+        z = self.get_first_stage_encoding(encoder_posterior).detach()
+
+        if self.model.conditioning_key is not None:
+            if cond_key is None:
+                cond_key = self.cond_stage_key
+            if cond_key != self.first_stage_key:
+                if cond_key in ['caption', 'coordinates_bbox']:
+                    xc = batch[cond_key]
+                elif cond_key == 'class_label':
+                    xc = batch
+                else:
+                    xc = super().get_input(batch, cond_key).to(self.device)
+            else:
+                xc = x
+            if not self.cond_stage_trainable or force_c_encode:
+                if isinstance(xc, dict) or isinstance(xc, list):
+                    # import pudb; pudb.set_trace()
+                    c = self.get_learned_conditioning(xc)
+                else:
+                    c = self.get_learned_conditioning(xc.to(self.device))
+            else:
+                c = xc
+            if bs is not None:
+                c = c[:bs]
+
+            if self.use_positional_encodings:
+                pos_x, pos_y = self.compute_latent_shifts(batch)
+                ckey = __conditioning_keys__[self.model.conditioning_key]
+                c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y}
+
+        else:
+            c = None
+            xc = None
+            if self.use_positional_encodings:
+                pos_x, pos_y = self.compute_latent_shifts(batch)
+                c = {'pos_x': pos_x, 'pos_y': pos_y}
+        out = [z, c]
+        if return_first_stage_outputs:
+            xrec = self.decode_first_stage(z)
+            out.extend([x, xrec])
+        if return_original_cond:
+            out.append(xc)
+        return out
+
+    @torch.no_grad()
+    def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, 'b h w c -> b c h w').contiguous()
+
+        z = 1. / self.scale_factor * z
+
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                uf = self.split_input_params["vqf"]
+                bs, nc, h, w = z.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
+
+                z = unfold(z)  # (bn, nc * prod(**ks), L)
+                # 1. Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                # 2. apply model loop over last dim
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
+                                                                 force_not_quantize=predict_cids or force_not_quantize)
+                                   for i in range(z.shape[-1])]
+                else:
+
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
+                                   for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)  # # (bn, nc, ks[0], ks[1], L)
+                o = o * weighting
+                # Reverse 1. reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization  # norm is shape (1, 1, h, w)
+                return decoded
+            else:
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+                else:
+                    return self.first_stage_model.decode(z)
+
+        else:
+            if isinstance(self.first_stage_model, VQModelInterface):
+                return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+            else:
+                return self.first_stage_model.decode(z)
+
+    # same as above but without decorator
+    def differentiable_decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, 'b h w c -> b c h w').contiguous()
+
+        z = 1. / self.scale_factor * z
+
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                uf = self.split_input_params["vqf"]
+                bs, nc, h, w = z.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
+
+                z = unfold(z)  # (bn, nc * prod(**ks), L)
+                # 1. Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                # 2. apply model loop over last dim
+                if isinstance(self.first_stage_model, VQModelInterface):  
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
+                                                                 force_not_quantize=predict_cids or force_not_quantize)
+                                   for i in range(z.shape[-1])]
+                else:
+
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
+                                   for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)  # # (bn, nc, ks[0], ks[1], L)
+                o = o * weighting
+                # Reverse 1. reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization  # norm is shape (1, 1, h, w)
+                return decoded
+            else:
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+                else:
+                    return self.first_stage_model.decode(z)
+
+        else:
+            if isinstance(self.first_stage_model, VQModelInterface):
+                return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+            else:
+                return self.first_stage_model.decode(z)
+
+    @torch.no_grad()
+    def encode_first_stage(self, x):
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                df = self.split_input_params["vqf"]
+                self.split_input_params['original_image_size'] = x.shape[-2:]
+                bs, nc, h, w = x.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(x, ks, stride, df=df)
+                z = unfold(x)  # (bn, nc * prod(**ks), L)
+                # Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                output_list = [self.first_stage_model.encode(z[:, :, :, :, i])
+                               for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)
+                o = o * weighting
+
+                # Reverse reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization
+                return decoded
+
+            else:
+                return self.first_stage_model.encode(x)
+        else:
+            return self.first_stage_model.encode(x)
+
+    def shared_step(self, batch, **kwargs):
+        x, c = self.get_input(batch, self.first_stage_key)
+        loss = self(x, c)
+        return loss
+
+    def forward(self, x, c, *args, **kwargs):
+        t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
+        if self.model.conditioning_key is not None:
+            assert c is not None
+            if self.cond_stage_trainable:
+                c = self.get_learned_conditioning(c)
+            if self.shorten_cond_schedule:  # TODO: drop this option
+                tc = self.cond_ids[t].to(self.device)
+                c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float()))
+        return self.p_losses(x, c, t, *args, **kwargs)
+
+    def _rescale_annotations(self, bboxes, crop_coordinates):  # TODO: move to dataset
+        def rescale_bbox(bbox):
+            x0 = clamp((bbox[0] - crop_coordinates[0]) / crop_coordinates[2])
+            y0 = clamp((bbox[1] - crop_coordinates[1]) / crop_coordinates[3])
+            w = min(bbox[2] / crop_coordinates[2], 1 - x0)
+            h = min(bbox[3] / crop_coordinates[3], 1 - y0)
+            return x0, y0, w, h
+
+        return [rescale_bbox(b) for b in bboxes]
+
+    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.model.conditioning_key == 'concat' else 'c_crossattn'
+            cond = {key: cond}
+
+        if hasattr(self, "split_input_params"):
+            assert len(cond) == 1  # todo can only deal with one conditioning atm
+            assert not return_ids  
+            ks = self.split_input_params["ks"]  # eg. (128, 128)
+            stride = self.split_input_params["stride"]  # eg. (64, 64)
+
+            h, w = x_noisy.shape[-2:]
+
+            fold, unfold, normalization, weighting = self.get_fold_unfold(x_noisy, ks, stride)
+
+            z = unfold(x_noisy)  # (bn, nc * prod(**ks), L)
+            # Reshape to img shape
+            z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+            z_list = [z[:, :, :, :, i] for i in range(z.shape[-1])]
+
+            if self.cond_stage_key in ["image", "LR_image", "segmentation",
+                                       'bbox_img'] and self.model.conditioning_key:  # todo check for completeness
+                c_key = next(iter(cond.keys()))  # get key
+                c = next(iter(cond.values()))  # get value
+                assert (len(c) == 1)  # todo extend to list with more than one elem
+                c = c[0]  # get element
+
+                c = unfold(c)
+                c = c.view((c.shape[0], -1, ks[0], ks[1], c.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                cond_list = [{c_key: [c[:, :, :, :, i]]} for i in range(c.shape[-1])]
+
+            elif self.cond_stage_key == 'coordinates_bbox':
+                assert 'original_image_size' in self.split_input_params, 'BoudingBoxRescaling is missing original_image_size'
+
+                # assuming padding of unfold is always 0 and its dilation is always 1
+                n_patches_per_row = int((w - ks[0]) / stride[0] + 1)
+                full_img_h, full_img_w = self.split_input_params['original_image_size']
+                # as we are operating on latents, we need the factor from the original image size to the
+                # spatial latent size to properly rescale the crops for regenerating the bbox annotations
+                num_downs = self.first_stage_model.encoder.num_resolutions - 1
+                rescale_latent = 2 ** (num_downs)
+
+                # get top left postions of patches as conforming for the bbbox tokenizer, therefore we
+                # need to rescale the tl patch coordinates to be in between (0,1)
+                tl_patch_coordinates = [(rescale_latent * stride[0] * (patch_nr % n_patches_per_row) / full_img_w,
+                                         rescale_latent * stride[1] * (patch_nr // n_patches_per_row) / full_img_h)
+                                        for patch_nr in range(z.shape[-1])]
+
+                # patch_limits are tl_coord, width and height coordinates as (x_tl, y_tl, h, w)
+                patch_limits = [(x_tl, y_tl,
+                                 rescale_latent * ks[0] / full_img_w,
+                                 rescale_latent * ks[1] / full_img_h) for x_tl, y_tl in tl_patch_coordinates]
+                # patch_values = [(np.arange(x_tl,min(x_tl+ks, 1.)),np.arange(y_tl,min(y_tl+ks, 1.))) for x_tl, y_tl in tl_patch_coordinates]
+
+                # tokenize crop coordinates for the bounding boxes of the respective patches
+                patch_limits_tknzd = [torch.LongTensor(self.bbox_tokenizer._crop_encoder(bbox))[None].to(self.device)
+                                      for bbox in patch_limits]  # list of length l with tensors of shape (1, 2)
+                print(patch_limits_tknzd[0].shape)
+                # cut tknzd crop position from conditioning
+                assert isinstance(cond, dict), 'cond must be dict to be fed into model'
+                cut_cond = cond['c_crossattn'][0][..., :-2].to(self.device)
+                print(cut_cond.shape)
+
+                adapted_cond = torch.stack([torch.cat([cut_cond, p], dim=1) for p in patch_limits_tknzd])
+                adapted_cond = rearrange(adapted_cond, 'l b n -> (l b) n')
+                print(adapted_cond.shape)
+                adapted_cond = self.get_learned_conditioning(adapted_cond)
+                print(adapted_cond.shape)
+                adapted_cond = rearrange(adapted_cond, '(l b) n d -> l b n d', l=z.shape[-1])
+                print(adapted_cond.shape)
+
+                cond_list = [{'c_crossattn': [e]} for e in adapted_cond]
+
+            else:
+                cond_list = [cond for i in range(z.shape[-1])]  # Todo make this more efficient
+
+            # apply model by loop over crops
+            output_list = [self.model(z_list[i], t, **cond_list[i]) for i in range(z.shape[-1])]
+            assert not isinstance(output_list[0],
+                                  tuple)  # todo cant deal with multiple model outputs check this never happens
+
+            o = torch.stack(output_list, axis=-1)
+            o = o * weighting
+            # Reverse reshape to img shape
+            o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+            # stitch crops together
+            x_recon = fold(o) / normalization
+
+        else:
+            x_recon = self.model(x_noisy, t, **cond)
+
+        if isinstance(x_recon, tuple) and not return_ids:
+            return x_recon[0]
+        else:
+            return x_recon
+
+    def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
+        return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \
+               extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
+
+    def _prior_bpd(self, x_start):
+        """
+        Get the prior KL term for the variational lower-bound, measured in
+        bits-per-dim.
+        This term can't be optimized, as it only depends on the encoder.
+        :param x_start: the [N x C x ...] tensor of inputs.
+        :return: a batch of [N] KL values (in bits), one per batch element.
+        """
+        batch_size = x_start.shape[0]
+        t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
+        qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
+        kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0)
+        return mean_flat(kl_prior) / np.log(2.0)
+
+    def p_losses(self, x_start, cond, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_output = self.apply_model(x_noisy, t, cond)
+
+        loss_dict = {}
+        prefix = 'train' if self.training else 'val'
+
+        if self.parameterization == "x0":
+            target = x_start
+        elif self.parameterization == "eps":
+            target = noise
+        else:
+            raise NotImplementedError()
+
+        loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3])
+        loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
+
+        logvar_t = self.logvar[t].to(self.device)
+        loss = loss_simple / torch.exp(logvar_t) + logvar_t
+        # loss = loss_simple / torch.exp(self.logvar) + self.logvar
+        if self.learn_logvar:
+            loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
+            loss_dict.update({'logvar': self.logvar.data.mean()})
+
+        loss = self.l_simple_weight * loss.mean()
+
+        loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3))
+        loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
+        loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
+        loss += (self.original_elbo_weight * loss_vlb)
+        loss_dict.update({f'{prefix}/loss': loss})
+
+        return loss, loss_dict
+
+    def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False,
+                        return_x0=False, score_corrector=None, corrector_kwargs=None):
+        t_in = t
+        model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
+
+        if score_corrector is not None:
+            assert self.parameterization == "eps"
+            model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
+
+        if return_codebook_ids:
+            model_out, logits = model_out
+
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        else:
+            raise NotImplementedError()
+
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+        if quantize_denoised:
+            x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        if return_codebook_ids:
+            return model_mean, posterior_variance, posterior_log_variance, logits
+        elif return_x0:
+            return model_mean, posterior_variance, posterior_log_variance, x_recon
+        else:
+            return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False,
+                 return_codebook_ids=False, quantize_denoised=False, return_x0=False,
+                 temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None):
+        b, *_, device = *x.shape, x.device
+        outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised,
+                                       return_codebook_ids=return_codebook_ids,
+                                       quantize_denoised=quantize_denoised,
+                                       return_x0=return_x0,
+                                       score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
+        if return_codebook_ids:
+            raise DeprecationWarning("Support dropped.")
+            model_mean, _, model_log_variance, logits = outputs
+        elif return_x0:
+            model_mean, _, model_log_variance, x0 = outputs
+        else:
+            model_mean, _, model_log_variance = outputs
+
+        noise = noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+
+        if return_codebook_ids:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1)
+        if return_x0:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
+        else:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False,
+                              img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0.,
+                              score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None,
+                              log_every_t=None):
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        timesteps = self.num_timesteps
+        if batch_size is not None:
+            b = batch_size if batch_size is not None else shape[0]
+            shape = [batch_size] + list(shape)
+        else:
+            b = batch_size = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=self.device)
+        else:
+            img = x_T
+        intermediates = []
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation',
+                        total=timesteps) if verbose else reversed(
+            range(0, timesteps))
+        if type(temperature) == float:
+            temperature = [temperature] * timesteps
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=self.device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img, x0_partial = self.p_sample(img, cond, ts,
+                                            clip_denoised=self.clip_denoised,
+                                            quantize_denoised=quantize_denoised, return_x0=True,
+                                            temperature=temperature[i], noise_dropout=noise_dropout,
+                                            score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(x0_partial)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_loop(self, cond, shape, return_intermediates=False,
+                      x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, start_T=None,
+                      log_every_t=None):
+
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        device = self.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        intermediates = [img]
+        if timesteps is None:
+            timesteps = self.num_timesteps
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(
+            range(0, timesteps))
+
+        if mask is not None:
+            assert x0 is not None
+            assert x0.shape[2:3] == mask.shape[2:3]  # spatial size has to match
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img = self.p_sample(img, cond, ts,
+                                clip_denoised=self.clip_denoised,
+                                quantize_denoised=quantize_denoised)
+            if mask is not None:
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(img)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None,
+               verbose=True, timesteps=None, quantize_denoised=False,
+               mask=None, x0=None, shape=None,**kwargs):
+        if shape is None:
+            shape = (batch_size, self.channels, self.image_size, self.image_size)
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+        return self.p_sample_loop(cond,
+                                  shape,
+                                  return_intermediates=return_intermediates, x_T=x_T,
+                                  verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised,
+                                  mask=mask, x0=x0)
+
+    @torch.no_grad()
+    def sample_log(self,cond,batch_size,ddim, ddim_steps,**kwargs):
+
+        if ddim:
+            ddim_sampler = DDIMSampler(self)
+            shape = (self.channels, self.image_size, self.image_size)
+            samples, intermediates =ddim_sampler.sample(ddim_steps,batch_size,
+                                                        shape,cond,verbose=False,**kwargs)
+
+        else:
+            samples, intermediates = self.sample(cond=cond, batch_size=batch_size,
+                                                 return_intermediates=True,**kwargs)
+
+        return samples, intermediates
+
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
+                   quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
+                   plot_diffusion_rows=True, **kwargs):
+
+        use_ddim = ddim_steps is not None
+
+        log = dict()
+        z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key,
+                                           return_first_stage_outputs=True,
+                                           force_c_encode=True,
+                                           return_original_cond=True,
+                                           bs=N)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        log["inputs"] = x
+        log["reconstruction"] = xrec
+        if self.model.conditioning_key is not None:
+            if hasattr(self.cond_stage_model, "decode"):
+                xc = self.cond_stage_model.decode(c)
+                log["conditioning"] = xc
+            elif self.cond_stage_key in ["caption"]:
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["caption"])
+                log["conditioning"] = xc
+            elif self.cond_stage_key == 'class_label':
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
+                log['conditioning'] = xc
+            elif isimage(xc):
+                log["conditioning"] = xc
+            if ismap(xc):
+                log["original_conditioning"] = self.to_rgb(xc)
+
+        if plot_diffusion_rows:
+            # get diffusion row
+            diffusion_row = list()
+            z_start = z[:n_row]
+            for t in range(self.num_timesteps):
+                if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                    t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                    t = t.to(self.device).long()
+                    noise = torch.randn_like(z_start)
+                    z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
+                    diffusion_row.append(self.decode_first_stage(z_noisy))
+
+            diffusion_row = torch.stack(diffusion_row)  # n_log_step, n_row, C, H, W
+            diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
+            diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
+            diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
+            log["diffusion_row"] = diffusion_grid
+
+        if sample:
+            # get denoise row
+            with self.ema_scope("Plotting"):
+                samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
+                                                         ddim_steps=ddim_steps,eta=ddim_eta)
+                # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+            if plot_denoise_rows:
+                denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
+                log["denoise_row"] = denoise_grid
+
+            if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance(
+                    self.first_stage_model, IdentityFirstStage):
+                # also display when quantizing x0 while sampling
+                with self.ema_scope("Plotting Quantized Denoised"):
+                    samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
+                                                             ddim_steps=ddim_steps,eta=ddim_eta,
+                                                             quantize_denoised=True)
+                    # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True,
+                    #                                      quantize_denoised=True)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_x0_quantized"] = x_samples
+
+            if inpaint:
+                # make a simple center square
+                b, h, w = z.shape[0], z.shape[2], z.shape[3]
+                mask = torch.ones(N, h, w).to(self.device)
+                # zeros will be filled in
+                mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0.
+                mask = mask[:, None, ...]
+                with self.ema_scope("Plotting Inpaint"):
+
+                    samples, _ = self.sample_log(cond=c,batch_size=N,ddim=use_ddim, eta=ddim_eta,
+                                                ddim_steps=ddim_steps, x0=z[:N], mask=mask)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_inpainting"] = x_samples
+                log["mask"] = mask
+
+                # outpaint
+                with self.ema_scope("Plotting Outpaint"):
+                    samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,eta=ddim_eta,
+                                                ddim_steps=ddim_steps, x0=z[:N], mask=mask)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_outpainting"] = x_samples
+
+        if plot_progressive_rows:
+            with self.ema_scope("Plotting Progressives"):
+                img, progressives = self.progressive_denoising(c,
+                                                               shape=(self.channels, self.image_size, self.image_size),
+                                                               batch_size=N)
+            prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
+            log["progressive_row"] = prog_row
+
+        if return_keys:
+            if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
+                return log
+            else:
+                return {key: log[key] for key in return_keys}
+        return log
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        params = list(self.model.parameters())
+        if self.cond_stage_trainable:
+            print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
+            params = params + list(self.cond_stage_model.parameters())
+        if self.learn_logvar:
+            print('Diffusion model optimizing logvar')
+            params.append(self.logvar)
+        opt = torch.optim.AdamW(params, lr=lr)
+        if self.use_scheduler:
+            assert 'target' in self.scheduler_config
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                }]
+            return [opt], scheduler
+        return opt
+
+    @torch.no_grad()
+    def to_rgb(self, x):
+        x = x.float()
+        if not hasattr(self, "colorize"):
+            self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x)
+        x = nn.functional.conv2d(x, weight=self.colorize)
+        x = 2. * (x - x.min()) / (x.max() - x.min()) - 1.
+        return x
+
+
+class DiffusionWrapperV1(pl.LightningModule):
+    def __init__(self, diff_model_config, conditioning_key):
+        super().__init__()
+        self.diffusion_model = instantiate_from_config(diff_model_config)
+        self.conditioning_key = conditioning_key
+        assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm']
+
+    def forward(self, x, t, c_concat: list = None, c_crossattn: list = None):
+        if self.conditioning_key is None:
+            out = self.diffusion_model(x, t)
+        elif self.conditioning_key == 'concat':
+            xc = torch.cat([x] + c_concat, dim=1)
+            out = self.diffusion_model(xc, t)
+        elif self.conditioning_key == 'crossattn':
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(x, t, context=cc)
+        elif self.conditioning_key == 'hybrid':
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc)
+        elif self.conditioning_key == 'adm':
+            cc = c_crossattn[0]
+            out = self.diffusion_model(x, t, y=cc)
+        else:
+            raise NotImplementedError()
+
+        return out
+
+
+class Layout2ImgDiffusionV1(LatentDiffusionV1):
+    # TODO: move all layout-specific hacks to this class
+    def __init__(self, cond_stage_key, *args, **kwargs):
+        assert cond_stage_key == 'coordinates_bbox', 'Layout2ImgDiffusion only for cond_stage_key="coordinates_bbox"'
+        super().__init__(cond_stage_key=cond_stage_key, *args, **kwargs)
+
+    def log_images(self, batch, N=8, *args, **kwargs):
+        logs = super().log_images(batch=batch, N=N, *args, **kwargs)
+
+        key = 'train' if self.training else 'validation'
+        dset = self.trainer.datamodule.datasets[key]
+        mapper = dset.conditional_builders[self.cond_stage_key]
+
+        bbox_imgs = []
+        map_fn = lambda catno: dset.get_textual_label(dset.get_category_id(catno))
+        for tknzd_bbox in batch[self.cond_stage_key][:N]:
+            bboximg = mapper.plot(tknzd_bbox.detach().cpu(), map_fn, (256, 256))
+            bbox_imgs.append(bboximg)
+
+        cond_img = torch.stack(bbox_imgs, dim=0)
+        logs['bbox_image'] = cond_img
+        return logs
+
+setattr(ldm.models.diffusion.ddpm, "DDPMV1", DDPMV1)
+setattr(ldm.models.diffusion.ddpm, "LatentDiffusionV1", LatentDiffusionV1)
+setattr(ldm.models.diffusion.ddpm, "DiffusionWrapperV1", DiffusionWrapperV1)
+setattr(ldm.models.diffusion.ddpm, "Layout2ImgDiffusionV1", Layout2ImgDiffusionV1)

extensions-builtin/Lora/extra_networks_lora.py

@@ -0,0 +1,26 @@
+from modules import extra_networks, shared
+import lora
+
+class ExtraNetworkLora(extra_networks.ExtraNetwork):
+    def __init__(self):
+        super().__init__('lora')
+
+    def activate(self, p, params_list):
+        additional = shared.opts.sd_lora
+
+        if additional != "" and additional in lora.available_loras and len([x for x in params_list if x.items[0] == additional]) == 0:
+            p.all_prompts = [x + f"<lora:{additional}:{shared.opts.extra_networks_default_multiplier}>" for x in p.all_prompts]
+            params_list.append(extra_networks.ExtraNetworkParams(items=[additional, shared.opts.extra_networks_default_multiplier]))
+
+        names = []
+        multipliers = []
+        for params in params_list:
+            assert len(params.items) > 0
+
+            names.append(params.items[0])
+            multipliers.append(float(params.items[1]) if len(params.items) > 1 else 1.0)
+
+        lora.load_loras(names, multipliers)
+
+    def deactivate(self, p):
+        pass

extensions-builtin/Lora/lora.py

@@ -0,0 +1,362 @@
+import glob
+import os
+import re
+import torch
+from typing import Union
+
+from modules import shared, devices, sd_models, errors
+
+metadata_tags_order = {"ss_sd_model_name": 1, "ss_resolution": 2, "ss_clip_skip": 3, "ss_num_train_images": 10, "ss_tag_frequency": 20}
+
+re_digits = re.compile(r"\d+")
+re_x_proj = re.compile(r"(.*)_([qkv]_proj)$")
+re_compiled = {}
+
+suffix_conversion = {
+    "attentions": {},
+    "resnets": {
+        "conv1": "in_layers_2",
+        "conv2": "out_layers_3",
+        "time_emb_proj": "emb_layers_1",
+        "conv_shortcut": "skip_connection",
+    }
+}
+
+
+def convert_diffusers_name_to_compvis(key, is_sd2):
+    def match(match_list, regex_text):
+        regex = re_compiled.get(regex_text)
+        if regex is None:
+            regex = re.compile(regex_text)
+            re_compiled[regex_text] = regex
+
+        r = re.match(regex, key)
+        if not r:
+            return False
+
+        match_list.clear()
+        match_list.extend([int(x) if re.match(re_digits, x) else x for x in r.groups()])
+        return True
+
+    m = []
+
+    if match(m, r"lora_unet_down_blocks_(\d+)_(attentions|resnets)_(\d+)_(.+)"):
+        suffix = suffix_conversion.get(m[1], {}).get(m[3], m[3])
+        return f"diffusion_model_input_blocks_{1 + m[0] * 3 + m[2]}_{1 if m[1] == 'attentions' else 0}_{suffix}"
+
+    if match(m, r"lora_unet_mid_block_(attentions|resnets)_(\d+)_(.+)"):
+        suffix = suffix_conversion.get(m[0], {}).get(m[2], m[2])
+        return f"diffusion_model_middle_block_{1 if m[0] == 'attentions' else m[1] * 2}_{suffix}"
+
+    if match(m, r"lora_unet_up_blocks_(\d+)_(attentions|resnets)_(\d+)_(.+)"):
+        suffix = suffix_conversion.get(m[1], {}).get(m[3], m[3])
+        return f"diffusion_model_output_blocks_{m[0] * 3 + m[2]}_{1 if m[1] == 'attentions' else 0}_{suffix}"
+
+    if match(m, r"lora_unet_down_blocks_(\d+)_downsamplers_0_conv"):
+        return f"diffusion_model_input_blocks_{3 + m[0] * 3}_0_op"
+
+    if match(m, r"lora_unet_up_blocks_(\d+)_upsamplers_0_conv"):
+        return f"diffusion_model_output_blocks_{2 + m[0] * 3}_{2 if m[0]>0 else 1}_conv"
+
+    if match(m, r"lora_te_text_model_encoder_layers_(\d+)_(.+)"):
+        if is_sd2:
+            if 'mlp_fc1' in m[1]:
+                return f"model_transformer_resblocks_{m[0]}_{m[1].replace('mlp_fc1', 'mlp_c_fc')}"
+            elif 'mlp_fc2' in m[1]:
+                return f"model_transformer_resblocks_{m[0]}_{m[1].replace('mlp_fc2', 'mlp_c_proj')}"
+            else:
+                return f"model_transformer_resblocks_{m[0]}_{m[1].replace('self_attn', 'attn')}"
+
+        return f"transformer_text_model_encoder_layers_{m[0]}_{m[1]}"
+
+    return key
+
+
+class LoraOnDisk:
+    def __init__(self, name, filename):
+        self.name = name
+        self.filename = filename
+        self.metadata = {}
+
+        _, ext = os.path.splitext(filename)
+        if ext.lower() == ".safetensors":
+            try:
+                self.metadata = sd_models.read_metadata_from_safetensors(filename)
+            except Exception as e:
+                errors.display(e, f"reading lora {filename}")
+
+        if self.metadata:
+            m = {}
+            for k, v in sorted(self.metadata.items(), key=lambda x: metadata_tags_order.get(x[0], 999)):
+                m[k] = v
+
+            self.metadata = m
+
+        self.ssmd_cover_images = self.metadata.pop('ssmd_cover_images', None)  # those are cover images and they are too big to display in UI as text
+
+
+class LoraModule:
+    def __init__(self, name):
+        self.name = name
+        self.multiplier = 1.0
+        self.modules = {}
+        self.mtime = None
+
+
+class LoraUpDownModule:
+    def __init__(self):
+        self.up = None
+        self.down = None
+        self.alpha = None
+
+
+def assign_lora_names_to_compvis_modules(sd_model):
+    lora_layer_mapping = {}
+
+    for name, module in shared.sd_model.cond_stage_model.wrapped.named_modules():
+        lora_name = name.replace(".", "_")
+        lora_layer_mapping[lora_name] = module
+        module.lora_layer_name = lora_name
+
+    for name, module in shared.sd_model.model.named_modules():
+        lora_name = name.replace(".", "_")
+        lora_layer_mapping[lora_name] = module
+        module.lora_layer_name = lora_name
+
+    sd_model.lora_layer_mapping = lora_layer_mapping
+
+
+def load_lora(name, filename):
+    lora = LoraModule(name)
+    lora.mtime = os.path.getmtime(filename)
+
+    sd = sd_models.read_state_dict(filename)
+
+    keys_failed_to_match = {}
+    is_sd2 = 'model_transformer_resblocks' in shared.sd_model.lora_layer_mapping
+
+    for key_diffusers, weight in sd.items():
+        key_diffusers_without_lora_parts, lora_key = key_diffusers.split(".", 1)
+        key = convert_diffusers_name_to_compvis(key_diffusers_without_lora_parts, is_sd2)
+
+        sd_module = shared.sd_model.lora_layer_mapping.get(key, None)
+
+        if sd_module is None:
+            m = re_x_proj.match(key)
+            if m:
+                sd_module = shared.sd_model.lora_layer_mapping.get(m.group(1), None)
+
+        if sd_module is None:
+            keys_failed_to_match[key_diffusers] = key
+            continue
+
+        lora_module = lora.modules.get(key, None)
+        if lora_module is None:
+            lora_module = LoraUpDownModule()
+            lora.modules[key] = lora_module
+
+        if lora_key == "alpha":
+            lora_module.alpha = weight.item()
+            continue
+
+        if type(sd_module) == torch.nn.Linear:
+            module = torch.nn.Linear(weight.shape[1], weight.shape[0], bias=False)
+        elif type(sd_module) == torch.nn.modules.linear.NonDynamicallyQuantizableLinear:
+            module = torch.nn.Linear(weight.shape[1], weight.shape[0], bias=False)
+        elif type(sd_module) == torch.nn.MultiheadAttention:
+            module = torch.nn.Linear(weight.shape[1], weight.shape[0], bias=False)
+        elif type(sd_module) == torch.nn.Conv2d:
+            module = torch.nn.Conv2d(weight.shape[1], weight.shape[0], (1, 1), bias=False)
+        else:
+            print(f'Lora layer {key_diffusers} matched a layer with unsupported type: {type(sd_module).__name__}')
+            continue
+            assert False, f'Lora layer {key_diffusers} matched a layer with unsupported type: {type(sd_module).__name__}'
+
+        with torch.no_grad():
+            module.weight.copy_(weight)
+
+        module.to(device=devices.cpu, dtype=devices.dtype)
+
+        if lora_key == "lora_up.weight":
+            lora_module.up = module
+        elif lora_key == "lora_down.weight":
+            lora_module.down = module
+        else:
+            assert False, f'Bad Lora layer name: {key_diffusers} - must end in lora_up.weight, lora_down.weight or alpha'
+
+    if len(keys_failed_to_match) > 0:
+        print(f"Failed to match keys when loading Lora {filename}: {keys_failed_to_match}")
+
+    return lora
+
+
+def load_loras(names, multipliers=None):
+    already_loaded = {}
+
+    for lora in loaded_loras:
+        if lora.name in names:
+            already_loaded[lora.name] = lora
+
+    loaded_loras.clear()
+
+    loras_on_disk = [available_loras.get(name, None) for name in names]
+    if any([x is None for x in loras_on_disk]):
+        list_available_loras()
+
+        loras_on_disk = [available_loras.get(name, None) for name in names]
+
+    for i, name in enumerate(names):
+        lora = already_loaded.get(name, None)
+
+        lora_on_disk = loras_on_disk[i]
+        if lora_on_disk is not None:
+            if lora is None or os.path.getmtime(lora_on_disk.filename) > lora.mtime:
+                lora = load_lora(name, lora_on_disk.filename)
+
+        if lora is None:
+            print(f"Couldn't find Lora with name {name}")
+            continue
+
+        lora.multiplier = multipliers[i] if multipliers else 1.0
+        loaded_loras.append(lora)
+
+
+def lora_calc_updown(lora, module, target):
+    with torch.no_grad():
+        up = module.up.weight.to(target.device, dtype=target.dtype)
+        down = module.down.weight.to(target.device, dtype=target.dtype)
+
+        if up.shape[2:] == (1, 1) and down.shape[2:] == (1, 1):
+            updown = (up.squeeze(2).squeeze(2) @ down.squeeze(2).squeeze(2)).unsqueeze(2).unsqueeze(3)
+        else:
+            updown = up @ down
+
+        updown = updown * lora.multiplier * (module.alpha / module.up.weight.shape[1] if module.alpha else 1.0)
+
+        return updown
+
+
+def lora_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn.MultiheadAttention]):
+    """
+    Applies the currently selected set of Loras to the weights of torch layer self.
+    If weights already have this particular set of loras applied, does nothing.
+    If not, restores orginal weights from backup and alters weights according to loras.
+    """
+
+    lora_layer_name = getattr(self, 'lora_layer_name', None)
+    if lora_layer_name is None:
+        return
+
+    current_names = getattr(self, "lora_current_names", ())
+    wanted_names = tuple((x.name, x.multiplier) for x in loaded_loras)
+
+    weights_backup = getattr(self, "lora_weights_backup", None)
+    if weights_backup is None:
+        if isinstance(self, torch.nn.MultiheadAttention):
+            weights_backup = (self.in_proj_weight.to(devices.cpu, copy=True), self.out_proj.weight.to(devices.cpu, copy=True))
+        else:
+            weights_backup = self.weight.to(devices.cpu, copy=True)
+
+        self.lora_weights_backup = weights_backup
+
+    if current_names != wanted_names:
+        if weights_backup is not None:
+            if isinstance(self, torch.nn.MultiheadAttention):
+                self.in_proj_weight.copy_(weights_backup[0])
+                self.out_proj.weight.copy_(weights_backup[1])
+            else:
+                self.weight.copy_(weights_backup)
+
+        for lora in loaded_loras:
+            module = lora.modules.get(lora_layer_name, None)
+            if module is not None and hasattr(self, 'weight'):
+                self.weight += lora_calc_updown(lora, module, self.weight)
+                continue
+
+            module_q = lora.modules.get(lora_layer_name + "_q_proj", None)
+            module_k = lora.modules.get(lora_layer_name + "_k_proj", None)
+            module_v = lora.modules.get(lora_layer_name + "_v_proj", None)
+            module_out = lora.modules.get(lora_layer_name + "_out_proj", None)
+
+            if isinstance(self, torch.nn.MultiheadAttention) and module_q and module_k and module_v and module_out:
+                updown_q = lora_calc_updown(lora, module_q, self.in_proj_weight)
+                updown_k = lora_calc_updown(lora, module_k, self.in_proj_weight)
+                updown_v = lora_calc_updown(lora, module_v, self.in_proj_weight)
+                updown_qkv = torch.vstack([updown_q, updown_k, updown_v])
+
+                self.in_proj_weight += updown_qkv
+                self.out_proj.weight += lora_calc_updown(lora, module_out, self.out_proj.weight)
+                continue
+
+            if module is None:
+                continue
+
+            print(f'failed to calculate lora weights for layer {lora_layer_name}')
+
+        setattr(self, "lora_current_names", wanted_names)
+
+
+def lora_reset_cached_weight(self: Union[torch.nn.Conv2d, torch.nn.Linear]):
+    setattr(self, "lora_current_names", ())
+    setattr(self, "lora_weights_backup", None)
+
+
+def lora_Linear_forward(self, input):
+    lora_apply_weights(self)
+
+    return torch.nn.Linear_forward_before_lora(self, input)
+
+
+def lora_Linear_load_state_dict(self, *args, **kwargs):
+    lora_reset_cached_weight(self)
+
+    return torch.nn.Linear_load_state_dict_before_lora(self, *args, **kwargs)
+
+
+def lora_Conv2d_forward(self, input):
+    lora_apply_weights(self)
+
+    return torch.nn.Conv2d_forward_before_lora(self, input)
+
+
+def lora_Conv2d_load_state_dict(self, *args, **kwargs):
+    lora_reset_cached_weight(self)
+
+    return torch.nn.Conv2d_load_state_dict_before_lora(self, *args, **kwargs)
+
+
+def lora_MultiheadAttention_forward(self, *args, **kwargs):
+    lora_apply_weights(self)
+
+    return torch.nn.MultiheadAttention_forward_before_lora(self, *args, **kwargs)
+
+
+def lora_MultiheadAttention_load_state_dict(self, *args, **kwargs):
+    lora_reset_cached_weight(self)
+
+    return torch.nn.MultiheadAttention_load_state_dict_before_lora(self, *args, **kwargs)
+
+
+def list_available_loras():
+    available_loras.clear()
+
+    os.makedirs(shared.cmd_opts.lora_dir, exist_ok=True)
+
+    candidates = \
+        glob.glob(os.path.join(shared.cmd_opts.lora_dir, '**/*.pt'), recursive=True) + \
+        glob.glob(os.path.join(shared.cmd_opts.lora_dir, '**/*.safetensors'), recursive=True) + \
+        glob.glob(os.path.join(shared.cmd_opts.lora_dir, '**/*.ckpt'), recursive=True)
+
+    for filename in sorted(candidates, key=str.lower):
+        if os.path.isdir(filename):
+            continue
+
+        name = os.path.splitext(os.path.basename(filename))[0]
+
+        available_loras[name] = LoraOnDisk(name, filename)
+
+
+available_loras = {}
+loaded_loras = []
+
+list_available_loras()

extensions-builtin/Lora/preload.py

@@ -0,0 +1,6 @@
+import os
+from modules import paths
+
+
+def preload(parser):
+    parser.add_argument("--lora-dir", type=str, help="Path to directory with Lora networks.", default=os.path.join(paths.models_path, 'Lora'))

extensions-builtin/Lora/scripts/lora_script.py

@@ -0,0 +1,56 @@
+import torch
+import gradio as gr
+
+import lora
+import extra_networks_lora
+import ui_extra_networks_lora
+from modules import script_callbacks, ui_extra_networks, extra_networks, shared
+
+
+def unload():
+    torch.nn.Linear.forward = torch.nn.Linear_forward_before_lora
+    torch.nn.Linear._load_from_state_dict = torch.nn.Linear_load_state_dict_before_lora
+    torch.nn.Conv2d.forward = torch.nn.Conv2d_forward_before_lora
+    torch.nn.Conv2d._load_from_state_dict = torch.nn.Conv2d_load_state_dict_before_lora
+    torch.nn.MultiheadAttention.forward = torch.nn.MultiheadAttention_forward_before_lora
+    torch.nn.MultiheadAttention._load_from_state_dict = torch.nn.MultiheadAttention_load_state_dict_before_lora
+
+
+def before_ui():
+    ui_extra_networks.register_page(ui_extra_networks_lora.ExtraNetworksPageLora())
+    extra_networks.register_extra_network(extra_networks_lora.ExtraNetworkLora())
+
+
+if not hasattr(torch.nn, 'Linear_forward_before_lora'):
+    torch.nn.Linear_forward_before_lora = torch.nn.Linear.forward
+
+if not hasattr(torch.nn, 'Linear_load_state_dict_before_lora'):
+    torch.nn.Linear_load_state_dict_before_lora = torch.nn.Linear._load_from_state_dict
+
+if not hasattr(torch.nn, 'Conv2d_forward_before_lora'):
+    torch.nn.Conv2d_forward_before_lora = torch.nn.Conv2d.forward
+
+if not hasattr(torch.nn, 'Conv2d_load_state_dict_before_lora'):
+    torch.nn.Conv2d_load_state_dict_before_lora = torch.nn.Conv2d._load_from_state_dict
+
+if not hasattr(torch.nn, 'MultiheadAttention_forward_before_lora'):
+    torch.nn.MultiheadAttention_forward_before_lora = torch.nn.MultiheadAttention.forward
+
+if not hasattr(torch.nn, 'MultiheadAttention_load_state_dict_before_lora'):
+    torch.nn.MultiheadAttention_load_state_dict_before_lora = torch.nn.MultiheadAttention._load_from_state_dict
+
+torch.nn.Linear.forward = lora.lora_Linear_forward
+torch.nn.Linear._load_from_state_dict = lora.lora_Linear_load_state_dict
+torch.nn.Conv2d.forward = lora.lora_Conv2d_forward
+torch.nn.Conv2d._load_from_state_dict = lora.lora_Conv2d_load_state_dict
+torch.nn.MultiheadAttention.forward = lora.lora_MultiheadAttention_forward
+torch.nn.MultiheadAttention._load_from_state_dict = lora.lora_MultiheadAttention_load_state_dict
+
+script_callbacks.on_model_loaded(lora.assign_lora_names_to_compvis_modules)
+script_callbacks.on_script_unloaded(unload)
+script_callbacks.on_before_ui(before_ui)
+
+
+shared.options_templates.update(shared.options_section(('extra_networks', "Extra Networks"), {
+    "sd_lora": shared.OptionInfo("None", "Add Lora to prompt", gr.Dropdown, lambda: {"choices": [""] + [x for x in lora.available_loras]}, refresh=lora.list_available_loras),
+}))

extensions-builtin/Lora/ui_extra_networks_lora.py

@@ -0,0 +1,31 @@
+import json
+import os
+import lora
+
+from modules import shared, ui_extra_networks
+
+
+class ExtraNetworksPageLora(ui_extra_networks.ExtraNetworksPage):
+    def __init__(self):
+        super().__init__('Lora')
+
+    def refresh(self):
+        lora.list_available_loras()
+
+    def list_items(self):
+        for name, lora_on_disk in lora.available_loras.items():
+            path, ext = os.path.splitext(lora_on_disk.filename)
+            yield {
+                "name": name,
+                "filename": path,
+                "preview": self.find_preview(path),
+                "description": self.find_description(path),
+                "search_term": self.search_terms_from_path(lora_on_disk.filename),
+                "prompt": json.dumps(f"<lora:{name}:") + " + opts.extra_networks_default_multiplier + " + json.dumps(">"),
+                "local_preview": f"{path}.{shared.opts.samples_format}",
+                "metadata": json.dumps(lora_on_disk.metadata, indent=4) if lora_on_disk.metadata else None,
+            }
+
+    def allowed_directories_for_previews(self):
+        return [shared.cmd_opts.lora_dir]
+

extensions-builtin/ScuNET/preload.py

@@ -0,0 +1,6 @@
+import os
+from modules import paths
+
+
+def preload(parser):
+    parser.add_argument("--scunet-models-path", type=str, help="Path to directory with ScuNET model file(s).", default=os.path.join(paths.models_path, 'ScuNET'))

extensions-builtin/ScuNET/scripts/scunet_model.py

@@ -0,0 +1,87 @@
+import os.path
+import sys
+import traceback
+
+import PIL.Image
+import numpy as np
+import torch
+from basicsr.utils.download_util import load_file_from_url
+
+import modules.upscaler
+from modules import devices, modelloader
+from scunet_model_arch import SCUNet as net
+
+
+class UpscalerScuNET(modules.upscaler.Upscaler):
+    def __init__(self, dirname):
+        self.name = "ScuNET"
+        self.model_name = "ScuNET GAN"
+        self.model_name2 = "ScuNET PSNR"
+        self.model_url = "https://github.com/cszn/KAIR/releases/download/v1.0/scunet_color_real_gan.pth"
+        self.model_url2 = "https://github.com/cszn/KAIR/releases/download/v1.0/scunet_color_real_psnr.pth"
+        self.user_path = dirname
+        super().__init__()
+        model_paths = self.find_models(ext_filter=[".pth"])
+        scalers = []
+        add_model2 = True
+        for file in model_paths:
+            if "http" in file:
+                name = self.model_name
+            else:
+                name = modelloader.friendly_name(file)
+            if name == self.model_name2 or file == self.model_url2:
+                add_model2 = False
+            try:
+                scaler_data = modules.upscaler.UpscalerData(name, file, self, 4)
+                scalers.append(scaler_data)
+            except Exception:
+                print(f"Error loading ScuNET model: {file}", file=sys.stderr)
+                print(traceback.format_exc(), file=sys.stderr)
+        if add_model2:
+            scaler_data2 = modules.upscaler.UpscalerData(self.model_name2, self.model_url2, self)
+            scalers.append(scaler_data2)
+        self.scalers = scalers
+
+    def do_upscale(self, img: PIL.Image, selected_file):
+        torch.cuda.empty_cache()
+
+        model = self.load_model(selected_file)
+        if model is None:
+            return img
+
+        device = devices.get_device_for('scunet')
+        img = np.array(img)
+        img = img[:, :, ::-1]
+        img = np.moveaxis(img, 2, 0) / 255
+        img = torch.from_numpy(img).float()
+        img = img.unsqueeze(0).to(device)
+
+        with torch.no_grad():
+            output = model(img)
+        output = output.squeeze().float().cpu().clamp_(0, 1).numpy()
+        output = 255. * np.moveaxis(output, 0, 2)
+        output = output.astype(np.uint8)
+        output = output[:, :, ::-1]
+        torch.cuda.empty_cache()
+        return PIL.Image.fromarray(output, 'RGB')
+
+    def load_model(self, path: str):
+        device = devices.get_device_for('scunet')
+        if "http" in path:
+            filename = load_file_from_url(url=self.model_url, model_dir=self.model_path, file_name="%s.pth" % self.name,
+                                          progress=True)
+        else:
+            filename = path
+        if not os.path.exists(os.path.join(self.model_path, filename)) or filename is None:
+            print(f"ScuNET: Unable to load model from {filename}", file=sys.stderr)
+            return None
+
+        model = net(in_nc=3, config=[4, 4, 4, 4, 4, 4, 4], dim=64)
+        model.load_state_dict(torch.load(filename), strict=True)
+        model.eval()
+        for k, v in model.named_parameters():
+            v.requires_grad = False
+        model = model.to(device)
+
+        return model
+

extensions-builtin/ScuNET/scunet_model_arch.py

@@ -0,0 +1,265 @@
+# -*- coding: utf-8 -*-
+import numpy as np
+import torch
+import torch.nn as nn
+from einops import rearrange
+from einops.layers.torch import Rearrange
+from timm.models.layers import trunc_normal_, DropPath
+
+
+class WMSA(nn.Module):
+    """ Self-attention module in Swin Transformer
+    """
+
+    def __init__(self, input_dim, output_dim, head_dim, window_size, type):
+        super(WMSA, self).__init__()
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.head_dim = head_dim
+        self.scale = self.head_dim ** -0.5
+        self.n_heads = input_dim // head_dim
+        self.window_size = window_size
+        self.type = type
+        self.embedding_layer = nn.Linear(self.input_dim, 3 * self.input_dim, bias=True)
+
+        self.relative_position_params = nn.Parameter(
+            torch.zeros((2 * window_size - 1) * (2 * window_size - 1), self.n_heads))
+
+        self.linear = nn.Linear(self.input_dim, self.output_dim)
+
+        trunc_normal_(self.relative_position_params, std=.02)
+        self.relative_position_params = torch.nn.Parameter(
+            self.relative_position_params.view(2 * window_size - 1, 2 * window_size - 1, self.n_heads).transpose(1,
+                                                                                                                 2).transpose(
+                0, 1))
+
+    def generate_mask(self, h, w, p, shift):
+        """ generating the mask of SW-MSA
+        Args:
+            shift: shift parameters in CyclicShift.
+        Returns:
+            attn_mask: should be (1 1 w p p),
+        """
+        # supporting square.
+        attn_mask = torch.zeros(h, w, p, p, p, p, dtype=torch.bool, device=self.relative_position_params.device)
+        if self.type == 'W':
+            return attn_mask
+
+        s = p - shift
+        attn_mask[-1, :, :s, :, s:, :] = True
+        attn_mask[-1, :, s:, :, :s, :] = True
+        attn_mask[:, -1, :, :s, :, s:] = True
+        attn_mask[:, -1, :, s:, :, :s] = True
+        attn_mask = rearrange(attn_mask, 'w1 w2 p1 p2 p3 p4 -> 1 1 (w1 w2) (p1 p2) (p3 p4)')
+        return attn_mask
+
+    def forward(self, x):
+        """ Forward pass of Window Multi-head Self-attention module.
+        Args:
+            x: input tensor with shape of [b h w c];
+            attn_mask: attention mask, fill -inf where the value is True;
+        Returns:
+            output: tensor shape [b h w c]
+        """
+        if self.type != 'W': x = torch.roll(x, shifts=(-(self.window_size // 2), -(self.window_size // 2)), dims=(1, 2))
+        x = rearrange(x, 'b (w1 p1) (w2 p2) c -> b w1 w2 p1 p2 c', p1=self.window_size, p2=self.window_size)
+        h_windows = x.size(1)
+        w_windows = x.size(2)
+        # square validation
+        # assert h_windows == w_windows
+
+        x = rearrange(x, 'b w1 w2 p1 p2 c -> b (w1 w2) (p1 p2) c', p1=self.window_size, p2=self.window_size)
+        qkv = self.embedding_layer(x)
+        q, k, v = rearrange(qkv, 'b nw np (threeh c) -> threeh b nw np c', c=self.head_dim).chunk(3, dim=0)
+        sim = torch.einsum('hbwpc,hbwqc->hbwpq', q, k) * self.scale
+        # Adding learnable relative embedding
+        sim = sim + rearrange(self.relative_embedding(), 'h p q -> h 1 1 p q')
+        # Using Attn Mask to distinguish different subwindows.
+        if self.type != 'W':
+            attn_mask = self.generate_mask(h_windows, w_windows, self.window_size, shift=self.window_size // 2)
+            sim = sim.masked_fill_(attn_mask, float("-inf"))
+
+        probs = nn.functional.softmax(sim, dim=-1)
+        output = torch.einsum('hbwij,hbwjc->hbwic', probs, v)
+        output = rearrange(output, 'h b w p c -> b w p (h c)')
+        output = self.linear(output)
+        output = rearrange(output, 'b (w1 w2) (p1 p2) c -> b (w1 p1) (w2 p2) c', w1=h_windows, p1=self.window_size)
+
+        if self.type != 'W': output = torch.roll(output, shifts=(self.window_size // 2, self.window_size // 2),
+                                                 dims=(1, 2))
+        return output
+
+    def relative_embedding(self):
+        cord = torch.tensor(np.array([[i, j] for i in range(self.window_size) for j in range(self.window_size)]))
+        relation = cord[:, None, :] - cord[None, :, :] + self.window_size - 1
+        # negative is allowed
+        return self.relative_position_params[:, relation[:, :, 0].long(), relation[:, :, 1].long()]
+
+
+class Block(nn.Module):
+    def __init__(self, input_dim, output_dim, head_dim, window_size, drop_path, type='W', input_resolution=None):
+        """ SwinTransformer Block
+        """
+        super(Block, self).__init__()
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        assert type in ['W', 'SW']
+        self.type = type
+        if input_resolution <= window_size:
+            self.type = 'W'
+
+        self.ln1 = nn.LayerNorm(input_dim)
+        self.msa = WMSA(input_dim, input_dim, head_dim, window_size, self.type)
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.ln2 = nn.LayerNorm(input_dim)
+        self.mlp = nn.Sequential(
+            nn.Linear(input_dim, 4 * input_dim),
+            nn.GELU(),
+            nn.Linear(4 * input_dim, output_dim),
+        )
+
+    def forward(self, x):
+        x = x + self.drop_path(self.msa(self.ln1(x)))
+        x = x + self.drop_path(self.mlp(self.ln2(x)))
+        return x
+
+
+class ConvTransBlock(nn.Module):
+    def __init__(self, conv_dim, trans_dim, head_dim, window_size, drop_path, type='W', input_resolution=None):
+        """ SwinTransformer and Conv Block
+        """
+        super(ConvTransBlock, self).__init__()
+        self.conv_dim = conv_dim
+        self.trans_dim = trans_dim
+        self.head_dim = head_dim
+        self.window_size = window_size
+        self.drop_path = drop_path
+        self.type = type
+        self.input_resolution = input_resolution
+
+        assert self.type in ['W', 'SW']
+        if self.input_resolution <= self.window_size:
+            self.type = 'W'
+
+        self.trans_block = Block(self.trans_dim, self.trans_dim, self.head_dim, self.window_size, self.drop_path,
+                                 self.type, self.input_resolution)
+        self.conv1_1 = nn.Conv2d(self.conv_dim + self.trans_dim, self.conv_dim + self.trans_dim, 1, 1, 0, bias=True)
+        self.conv1_2 = nn.Conv2d(self.conv_dim + self.trans_dim, self.conv_dim + self.trans_dim, 1, 1, 0, bias=True)
+
+        self.conv_block = nn.Sequential(
+            nn.Conv2d(self.conv_dim, self.conv_dim, 3, 1, 1, bias=False),
+            nn.ReLU(True),
+            nn.Conv2d(self.conv_dim, self.conv_dim, 3, 1, 1, bias=False)
+        )
+
+    def forward(self, x):
+        conv_x, trans_x = torch.split(self.conv1_1(x), (self.conv_dim, self.trans_dim), dim=1)
+        conv_x = self.conv_block(conv_x) + conv_x
+        trans_x = Rearrange('b c h w -> b h w c')(trans_x)
+        trans_x = self.trans_block(trans_x)
+        trans_x = Rearrange('b h w c -> b c h w')(trans_x)
+        res = self.conv1_2(torch.cat((conv_x, trans_x), dim=1))
+        x = x + res
+
+        return x
+
+
+class SCUNet(nn.Module):
+    # def __init__(self, in_nc=3, config=[2, 2, 2, 2, 2, 2, 2], dim=64, drop_path_rate=0.0, input_resolution=256):
+    def __init__(self, in_nc=3, config=None, dim=64, drop_path_rate=0.0, input_resolution=256):
+        super(SCUNet, self).__init__()
+        if config is None:
+            config = [2, 2, 2, 2, 2, 2, 2]
+        self.config = config
+        self.dim = dim
+        self.head_dim = 32
+        self.window_size = 8
+
+        # drop path rate for each layer
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(config))]
+
+        self.m_head = [nn.Conv2d(in_nc, dim, 3, 1, 1, bias=False)]
+
+        begin = 0
+        self.m_down1 = [ConvTransBlock(dim // 2, dim // 2, self.head_dim, self.window_size, dpr[i + begin],
+                                       'W' if not i % 2 else 'SW', input_resolution)
+                        for i in range(config[0])] + \
+                       [nn.Conv2d(dim, 2 * dim, 2, 2, 0, bias=False)]
+
+        begin += config[0]
+        self.m_down2 = [ConvTransBlock(dim, dim, self.head_dim, self.window_size, dpr[i + begin],
+                                       'W' if not i % 2 else 'SW', input_resolution // 2)
+                        for i in range(config[1])] + \
+                       [nn.Conv2d(2 * dim, 4 * dim, 2, 2, 0, bias=False)]
+
+        begin += config[1]
+        self.m_down3 = [ConvTransBlock(2 * dim, 2 * dim, self.head_dim, self.window_size, dpr[i + begin],
+                                       'W' if not i % 2 else 'SW', input_resolution // 4)
+                        for i in range(config[2])] + \
+                       [nn.Conv2d(4 * dim, 8 * dim, 2, 2, 0, bias=False)]
+
+        begin += config[2]
+        self.m_body = [ConvTransBlock(4 * dim, 4 * dim, self.head_dim, self.window_size, dpr[i + begin],
+                                      'W' if not i % 2 else 'SW', input_resolution // 8)
+                       for i in range(config[3])]
+
+        begin += config[3]
+        self.m_up3 = [nn.ConvTranspose2d(8 * dim, 4 * dim, 2, 2, 0, bias=False), ] + \
+                     [ConvTransBlock(2 * dim, 2 * dim, self.head_dim, self.window_size, dpr[i + begin],
+                                     'W' if not i % 2 else 'SW', input_resolution // 4)
+                      for i in range(config[4])]
+
+        begin += config[4]
+        self.m_up2 = [nn.ConvTranspose2d(4 * dim, 2 * dim, 2, 2, 0, bias=False), ] + \
+                     [ConvTransBlock(dim, dim, self.head_dim, self.window_size, dpr[i + begin],
+                                     'W' if not i % 2 else 'SW', input_resolution // 2)
+                      for i in range(config[5])]
+
+        begin += config[5]
+        self.m_up1 = [nn.ConvTranspose2d(2 * dim, dim, 2, 2, 0, bias=False), ] + \
+                     [ConvTransBlock(dim // 2, dim // 2, self.head_dim, self.window_size, dpr[i + begin],
+                                     'W' if not i % 2 else 'SW', input_resolution)
+                      for i in range(config[6])]
+
+        self.m_tail = [nn.Conv2d(dim, in_nc, 3, 1, 1, bias=False)]
+
+        self.m_head = nn.Sequential(*self.m_head)
+        self.m_down1 = nn.Sequential(*self.m_down1)
+        self.m_down2 = nn.Sequential(*self.m_down2)
+        self.m_down3 = nn.Sequential(*self.m_down3)
+        self.m_body = nn.Sequential(*self.m_body)
+        self.m_up3 = nn.Sequential(*self.m_up3)
+        self.m_up2 = nn.Sequential(*self.m_up2)
+        self.m_up1 = nn.Sequential(*self.m_up1)
+        self.m_tail = nn.Sequential(*self.m_tail)
+        # self.apply(self._init_weights)
+
+    def forward(self, x0):
+
+        h, w = x0.size()[-2:]
+        paddingBottom = int(np.ceil(h / 64) * 64 - h)
+        paddingRight = int(np.ceil(w / 64) * 64 - w)
+        x0 = nn.ReplicationPad2d((0, paddingRight, 0, paddingBottom))(x0)
+
+        x1 = self.m_head(x0)
+        x2 = self.m_down1(x1)
+        x3 = self.m_down2(x2)
+        x4 = self.m_down3(x3)
+        x = self.m_body(x4)
+        x = self.m_up3(x + x4)
+        x = self.m_up2(x + x3)
+        x = self.m_up1(x + x2)
+        x = self.m_tail(x + x1)
+
+        x = x[..., :h, :w]
+
+        return x
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)

extensions-builtin/SwinIR/preload.py

@@ -0,0 +1,6 @@
+import os
+from modules import paths
+
+
+def preload(parser):
+    parser.add_argument("--swinir-models-path", type=str, help="Path to directory with SwinIR model file(s).", default=os.path.join(paths.models_path, 'SwinIR'))

extensions-builtin/SwinIR/scripts/swinir_model.py

@@ -0,0 +1,178 @@
+import contextlib
+import os
+
+import numpy as np
+import torch
+from PIL import Image
+from basicsr.utils.download_util import load_file_from_url
+from tqdm import tqdm
+
+from modules import modelloader, devices, script_callbacks, shared
+from modules.shared import cmd_opts, opts, state
+from swinir_model_arch import SwinIR as net
+from swinir_model_arch_v2 import Swin2SR as net2
+from modules.upscaler import Upscaler, UpscalerData
+
+
+device_swinir = devices.get_device_for('swinir')
+
+
+class UpscalerSwinIR(Upscaler):
+    def __init__(self, dirname):
+        self.name = "SwinIR"
+        self.model_url = "https://github.com/JingyunLiang/SwinIR/releases/download/v0.0" \
+                         "/003_realSR_BSRGAN_DFOWMFC_s64w8_SwinIR" \
+                         "-L_x4_GAN.pth "
+        self.model_name = "SwinIR 4x"
+        self.user_path = dirname
+        super().__init__()
+        scalers = []
+        model_files = self.find_models(ext_filter=[".pt", ".pth"])
+        for model in model_files:
+            if "http" in model:
+                name = self.model_name
+            else:
+                name = modelloader.friendly_name(model)
+            model_data = UpscalerData(name, model, self)
+            scalers.append(model_data)
+        self.scalers = scalers
+
+    def do_upscale(self, img, model_file):
+        model = self.load_model(model_file)
+        if model is None:
+            return img
+        model = model.to(device_swinir, dtype=devices.dtype)
+        img = upscale(img, model)
+        try:
+            torch.cuda.empty_cache()
+        except:
+            pass
+        return img
+
+    def load_model(self, path, scale=4):
+        if "http" in path:
+            dl_name = "%s%s" % (self.model_name.replace(" ", "_"), ".pth")
+            filename = load_file_from_url(url=path, model_dir=self.model_path, file_name=dl_name, progress=True)
+        else:
+            filename = path
+        if filename is None or not os.path.exists(filename):
+            return None
+        if filename.endswith(".v2.pth"):
+            model = net2(
+            upscale=scale,
+            in_chans=3,
+            img_size=64,
+            window_size=8,
+            img_range=1.0,
+            depths=[6, 6, 6, 6, 6, 6],
+            embed_dim=180,
+            num_heads=[6, 6, 6, 6, 6, 6],
+            mlp_ratio=2,
+            upsampler="nearest+conv",
+            resi_connection="1conv",
+            )
+            params = None
+        else:
+            model = net(
+                upscale=scale,
+                in_chans=3,
+                img_size=64,
+                window_size=8,
+                img_range=1.0,
+                depths=[6, 6, 6, 6, 6, 6, 6, 6, 6],
+                embed_dim=240,
+                num_heads=[8, 8, 8, 8, 8, 8, 8, 8, 8],
+                mlp_ratio=2,
+                upsampler="nearest+conv",
+                resi_connection="3conv",
+            )
+            params = "params_ema"
+
+        pretrained_model = torch.load(filename)
+        if params is not None:
+            model.load_state_dict(pretrained_model[params], strict=True)
+        else:
+            model.load_state_dict(pretrained_model, strict=True)
+        return model
+
+
+def upscale(
+        img,
+        model,
+        tile=None,
+        tile_overlap=None,
+        window_size=8,
+        scale=4,
+):
+    tile = tile or opts.SWIN_tile
+    tile_overlap = tile_overlap or opts.SWIN_tile_overlap
+
+
+    img = np.array(img)
+    img = img[:, :, ::-1]
+    img = np.moveaxis(img, 2, 0) / 255
+    img = torch.from_numpy(img).float()
+    img = img.unsqueeze(0).to(device_swinir, dtype=devices.dtype)
+    with torch.no_grad(), devices.autocast():
+        _, _, h_old, w_old = img.size()
+        h_pad = (h_old // window_size + 1) * window_size - h_old
+        w_pad = (w_old // window_size + 1) * window_size - w_old
+        img = torch.cat([img, torch.flip(img, [2])], 2)[:, :, : h_old + h_pad, :]
+        img = torch.cat([img, torch.flip(img, [3])], 3)[:, :, :, : w_old + w_pad]
+        output = inference(img, model, tile, tile_overlap, window_size, scale)
+        output = output[..., : h_old * scale, : w_old * scale]
+        output = output.data.squeeze().float().cpu().clamp_(0, 1).numpy()
+        if output.ndim == 3:
+            output = np.transpose(
+                output[[2, 1, 0], :, :], (1, 2, 0)
+            )  # CHW-RGB to HCW-BGR
+        output = (output * 255.0).round().astype(np.uint8)  # float32 to uint8
+        return Image.fromarray(output, "RGB")
+
+
+def inference(img, model, tile, tile_overlap, window_size, scale):
+    # test the image tile by tile
+    b, c, h, w = img.size()
+    tile = min(tile, h, w)
+    assert tile % window_size == 0, "tile size should be a multiple of window_size"
+    sf = scale
+
+    stride = tile - tile_overlap
+    h_idx_list = list(range(0, h - tile, stride)) + [h - tile]
+    w_idx_list = list(range(0, w - tile, stride)) + [w - tile]
+    E = torch.zeros(b, c, h * sf, w * sf, dtype=devices.dtype, device=device_swinir).type_as(img)
+    W = torch.zeros_like(E, dtype=devices.dtype, device=device_swinir)
+
+    with tqdm(total=len(h_idx_list) * len(w_idx_list), desc="SwinIR tiles") as pbar:
+        for h_idx in h_idx_list:
+            if state.interrupted or state.skipped:
+                break
+
+            for w_idx in w_idx_list:
+                if state.interrupted or state.skipped:
+                    break
+                
+                in_patch = img[..., h_idx: h_idx + tile, w_idx: w_idx + tile]
+                out_patch = model(in_patch)
+                out_patch_mask = torch.ones_like(out_patch)
+
+                E[
+                ..., h_idx * sf: (h_idx + tile) * sf, w_idx * sf: (w_idx + tile) * sf
+                ].add_(out_patch)
+                W[
+                ..., h_idx * sf: (h_idx + tile) * sf, w_idx * sf: (w_idx + tile) * sf
+                ].add_(out_patch_mask)
+                pbar.update(1)
+    output = E.div_(W)
+
+    return output
+
+
+def on_ui_settings():
+    import gradio as gr
+
+    shared.opts.add_option("SWIN_tile", shared.OptionInfo(192, "Tile size for all SwinIR.", gr.Slider, {"minimum": 16, "maximum": 512, "step": 16}, section=('upscaling', "Upscaling")))
+    shared.opts.add_option("SWIN_tile_overlap", shared.OptionInfo(8, "Tile overlap, in pixels for SwinIR. Low values = visible seam.", gr.Slider, {"minimum": 0, "maximum": 48, "step": 1}, section=('upscaling', "Upscaling")))
+
+
+script_callbacks.on_ui_settings(on_ui_settings)

extensions-builtin/SwinIR/swinir_model_arch.py

@@ -0,0 +1,867 @@
+# -----------------------------------------------------------------------------------
+# SwinIR: Image Restoration Using Swin Transformer, https://arxiv.org/abs/2108.10257
+# Originally Written by Ze Liu, Modified by Jingyun Liang.
+# -----------------------------------------------------------------------------------
+
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    r""" Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'
+
+    def flops(self, N):
+        # calculate flops for 1 window with token length of N
+        flops = 0
+        # qkv = self.qkv(x)
+        flops += N * self.dim * 3 * self.dim
+        # attn = (q @ k.transpose(-2, -1))
+        flops += self.num_heads * N * (self.dim // self.num_heads) * N
+        #  x = (attn @ v)
+        flops += self.num_heads * N * N * (self.dim // self.num_heads)
+        # x = self.proj(x)
+        flops += N * self.dim * self.dim
+        return flops
+
+
+class SwinTransformerBlock(nn.Module):
+    r""" Swin Transformer Block.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        if min(self.input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(self.input_resolution)
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        if self.shift_size > 0:
+            attn_mask = self.calculate_mask(self.input_resolution)
+        else:
+            attn_mask = None
+
+        self.register_buffer("attn_mask", attn_mask)
+
+    def calculate_mask(self, x_size):
+        # calculate attention mask for SW-MSA
+        H, W = x_size
+        img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        return attn_mask
+
+    def forward(self, x, x_size):
+        H, W = x_size
+        B, L, C = x.shape
+        # assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
+        if self.input_resolution == x_size:
+            attn_windows = self.attn(x_windows, mask=self.attn_mask)  # nW*B, window_size*window_size, C
+        else:
+            attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device))
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
+               f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
+
+    def flops(self):
+        flops = 0
+        H, W = self.input_resolution
+        # norm1
+        flops += self.dim * H * W
+        # W-MSA/SW-MSA
+        nW = H * W / self.window_size / self.window_size
+        flops += nW * self.attn.flops(self.window_size * self.window_size)
+        # mlp
+        flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
+        # norm2
+        flops += self.dim * H * W
+        return flops
+
+
+class PatchMerging(nn.Module):
+    r""" Patch Merging Layer.
+
+    Args:
+        input_resolution (tuple[int]): Resolution of input feature.
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x):
+        """
+        x: B, H*W, C
+        """
+        H, W = self.input_resolution
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
+
+        x = x.view(B, H, W, C)
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f"input_resolution={self.input_resolution}, dim={self.dim}"
+
+    def flops(self):
+        H, W = self.input_resolution
+        flops = H * W * self.dim
+        flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
+        return flops
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
+                                 num_heads=num_heads, window_size=window_size,
+                                 shift_size=0 if (i % 2 == 0) else window_size // 2,
+                                 mlp_ratio=mlp_ratio,
+                                 qkv_bias=qkv_bias, qk_scale=qk_scale,
+                                 drop=drop, attn_drop=attn_drop,
+                                 drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                                 norm_layer=norm_layer)
+            for i in range(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, x_size):
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, x_size)
+            else:
+                x = blk(x, x_size)
+        if self.downsample is not None:
+            x = self.downsample(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
+
+    def flops(self):
+        flops = 0
+        for blk in self.blocks:
+            flops += blk.flops()
+        if self.downsample is not None:
+            flops += self.downsample.flops()
+        return flops
+
+
+class RSTB(nn.Module):
+    """Residual Swin Transformer Block (RSTB).
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+        img_size: Input image size.
+        patch_size: Patch size.
+        resi_connection: The convolutional block before residual connection.
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,
+                 img_size=224, patch_size=4, resi_connection='1conv'):
+        super(RSTB, self).__init__()
+
+        self.dim = dim
+        self.input_resolution = input_resolution
+
+        self.residual_group = BasicLayer(dim=dim,
+                                         input_resolution=input_resolution,
+                                         depth=depth,
+                                         num_heads=num_heads,
+                                         window_size=window_size,
+                                         mlp_ratio=mlp_ratio,
+                                         qkv_bias=qkv_bias, qk_scale=qk_scale,
+                                         drop=drop, attn_drop=attn_drop,
+                                         drop_path=drop_path,
+                                         norm_layer=norm_layer,
+                                         downsample=downsample,
+                                         use_checkpoint=use_checkpoint)
+
+        if resi_connection == '1conv':
+            self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
+        elif resi_connection == '3conv':
+            # to save parameters and memory
+            self.conv = nn.Sequential(nn.Conv2d(dim, dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                      nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
+                                      nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                      nn.Conv2d(dim // 4, dim, 3, 1, 1))
+
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim,
+            norm_layer=None)
+
+        self.patch_unembed = PatchUnEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim,
+            norm_layer=None)
+
+    def forward(self, x, x_size):
+        return self.patch_embed(self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))) + x
+
+    def flops(self):
+        flops = 0
+        flops += self.residual_group.flops()
+        H, W = self.input_resolution
+        flops += H * W * self.dim * self.dim * 9
+        flops += self.patch_embed.flops()
+        flops += self.patch_unembed.flops()
+
+        return flops
+
+
+class PatchEmbed(nn.Module):
+    r""" Image to Patch Embedding
+
+    Args:
+        img_size (int): Image size.  Default: 224.
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        x = x.flatten(2).transpose(1, 2)  # B Ph*Pw C
+        if self.norm is not None:
+            x = self.norm(x)
+        return x
+
+    def flops(self):
+        flops = 0
+        H, W = self.img_size
+        if self.norm is not None:
+            flops += H * W * self.embed_dim
+        return flops
+
+
+class PatchUnEmbed(nn.Module):
+    r""" Image to Patch Unembedding
+
+    Args:
+        img_size (int): Image size.  Default: 224.
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+    def forward(self, x, x_size):
+        B, HW, C = x.shape
+        x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1])  # B Ph*Pw C
+        return x
+
+    def flops(self):
+        flops = 0
+        return flops
+
+
+class Upsample(nn.Sequential):
+    """Upsample module.
+
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+    """
+
+    def __init__(self, scale, num_feat):
+        m = []
+        if (scale & (scale - 1)) == 0:  # scale = 2^n
+            for _ in range(int(math.log(scale, 2))):
+                m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
+                m.append(nn.PixelShuffle(2))
+        elif scale == 3:
+            m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
+            m.append(nn.PixelShuffle(3))
+        else:
+            raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.')
+        super(Upsample, self).__init__(*m)
+
+
+class UpsampleOneStep(nn.Sequential):
+    """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
+       Used in lightweight SR to save parameters.
+
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+
+    """
+
+    def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
+        self.num_feat = num_feat
+        self.input_resolution = input_resolution
+        m = []
+        m.append(nn.Conv2d(num_feat, (scale ** 2) * num_out_ch, 3, 1, 1))
+        m.append(nn.PixelShuffle(scale))
+        super(UpsampleOneStep, self).__init__(*m)
+
+    def flops(self):
+        H, W = self.input_resolution
+        flops = H * W * self.num_feat * 3 * 9
+        return flops
+
+
+class SwinIR(nn.Module):
+    r""" SwinIR
+        A PyTorch impl of : `SwinIR: Image Restoration Using Swin Transformer`, based on Swin Transformer.
+
+    Args:
+        img_size (int | tuple(int)): Input image size. Default 64
+        patch_size (int | tuple(int)): Patch size. Default: 1
+        in_chans (int): Number of input image channels. Default: 3
+        embed_dim (int): Patch embedding dimension. Default: 96
+        depths (tuple(int)): Depth of each Swin Transformer layer.
+        num_heads (tuple(int)): Number of attention heads in different layers.
+        window_size (int): Window size. Default: 7
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None
+        drop_rate (float): Dropout rate. Default: 0
+        attn_drop_rate (float): Attention dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
+        upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
+        img_range: Image range. 1. or 255.
+        upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
+        resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
+    """
+
+    def __init__(self, img_size=64, patch_size=1, in_chans=3,
+                 embed_dim=96, depths=[6, 6, 6, 6], num_heads=[6, 6, 6, 6],
+                 window_size=7, mlp_ratio=4., qkv_bias=True, qk_scale=None,
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
+                 norm_layer=nn.LayerNorm, ape=False, patch_norm=True,
+                 use_checkpoint=False, upscale=2, img_range=1., upsampler='', resi_connection='1conv',
+                 **kwargs):
+        super(SwinIR, self).__init__()
+        num_in_ch = in_chans
+        num_out_ch = in_chans
+        num_feat = 64
+        self.img_range = img_range
+        if in_chans == 3:
+            rgb_mean = (0.4488, 0.4371, 0.4040)
+            self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
+        else:
+            self.mean = torch.zeros(1, 1, 1, 1)
+        self.upscale = upscale
+        self.upsampler = upsampler
+        self.window_size = window_size
+
+        #####################################################################################################
+        ################################### 1, shallow feature extraction ###################################
+        self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)
+
+        #####################################################################################################
+        ################################### 2, deep feature extraction ######################################
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.num_features = embed_dim
+        self.mlp_ratio = mlp_ratio
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+        num_patches = self.patch_embed.num_patches
+        patches_resolution = self.patch_embed.patches_resolution
+        self.patches_resolution = patches_resolution
+
+        # merge non-overlapping patches into image
+        self.patch_unembed = PatchUnEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        # absolute position embedding
+        if self.ape:
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+
+        # build Residual Swin Transformer blocks (RSTB)
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = RSTB(dim=embed_dim,
+                         input_resolution=(patches_resolution[0],
+                                           patches_resolution[1]),
+                         depth=depths[i_layer],
+                         num_heads=num_heads[i_layer],
+                         window_size=window_size,
+                         mlp_ratio=self.mlp_ratio,
+                         qkv_bias=qkv_bias, qk_scale=qk_scale,
+                         drop=drop_rate, attn_drop=attn_drop_rate,
+                         drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],  # no impact on SR results
+                         norm_layer=norm_layer,
+                         downsample=None,
+                         use_checkpoint=use_checkpoint,
+                         img_size=img_size,
+                         patch_size=patch_size,
+                         resi_connection=resi_connection
+
+                         )
+            self.layers.append(layer)
+        self.norm = norm_layer(self.num_features)
+
+        # build the last conv layer in deep feature extraction
+        if resi_connection == '1conv':
+            self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
+        elif resi_connection == '3conv':
+            # to save parameters and memory
+            self.conv_after_body = nn.Sequential(nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1),
+                                                 nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                                 nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0),
+                                                 nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                                 nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1))
+
+        #####################################################################################################
+        ################################ 3, high quality image reconstruction ################################
+        if self.upsampler == 'pixelshuffle':
+            # for classical SR
+            self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                                                      nn.LeakyReLU(inplace=True))
+            self.upsample = Upsample(upscale, num_feat)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+        elif self.upsampler == 'pixelshuffledirect':
+            # for lightweight SR (to save parameters)
+            self.upsample = UpsampleOneStep(upscale, embed_dim, num_out_ch,
+                                            (patches_resolution[0], patches_resolution[1]))
+        elif self.upsampler == 'nearest+conv':
+            # for real-world SR (less artifacts)
+            self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                                                      nn.LeakyReLU(inplace=True))
+            self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            if self.upscale == 4:
+                self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+        else:
+            # for image denoising and JPEG compression artifact reduction
+            self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'absolute_pos_embed'}
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        return {'relative_position_bias_table'}
+
+    def check_image_size(self, x):
+        _, _, h, w = x.size()
+        mod_pad_h = (self.window_size - h % self.window_size) % self.window_size
+        mod_pad_w = (self.window_size - w % self.window_size) % self.window_size
+        x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect')
+        return x
+
+    def forward_features(self, x):
+        x_size = (x.shape[2], x.shape[3])
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        for layer in self.layers:
+            x = layer(x, x_size)
+
+        x = self.norm(x)  # B L C
+        x = self.patch_unembed(x, x_size)
+
+        return x
+
+    def forward(self, x):
+        H, W = x.shape[2:]
+        x = self.check_image_size(x)
+        
+        self.mean = self.mean.type_as(x)
+        x = (x - self.mean) * self.img_range
+
+        if self.upsampler == 'pixelshuffle':
+            # for classical SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            x = self.conv_last(self.upsample(x))
+        elif self.upsampler == 'pixelshuffledirect':
+            # for lightweight SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.upsample(x)
+        elif self.upsampler == 'nearest+conv':
+            # for real-world SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            x = self.lrelu(self.conv_up1(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
+            if self.upscale == 4:
+                x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
+            x = self.conv_last(self.lrelu(self.conv_hr(x)))
+        else:
+            # for image denoising and JPEG compression artifact reduction
+            x_first = self.conv_first(x)
+            res = self.conv_after_body(self.forward_features(x_first)) + x_first
+            x = x + self.conv_last(res)
+
+        x = x / self.img_range + self.mean
+
+        return x[:, :, :H*self.upscale, :W*self.upscale]
+
+    def flops(self):
+        flops = 0
+        H, W = self.patches_resolution
+        flops += H * W * 3 * self.embed_dim * 9
+        flops += self.patch_embed.flops()
+        for i, layer in enumerate(self.layers):
+            flops += layer.flops()
+        flops += H * W * 3 * self.embed_dim * self.embed_dim
+        flops += self.upsample.flops()
+        return flops
+
+
+if __name__ == '__main__':
+    upscale = 4
+    window_size = 8
+    height = (1024 // upscale // window_size + 1) * window_size
+    width = (720 // upscale // window_size + 1) * window_size
+    model = SwinIR(upscale=2, img_size=(height, width),
+                   window_size=window_size, img_range=1., depths=[6, 6, 6, 6],
+                   embed_dim=60, num_heads=[6, 6, 6, 6], mlp_ratio=2, upsampler='pixelshuffledirect')
+    print(model)
+    print(height, width, model.flops() / 1e9)
+
+    x = torch.randn((1, 3, height, width))
+    x = model(x)
+    print(x.shape)

extensions-builtin/SwinIR/swinir_model_arch_v2.py

@@ -0,0 +1,1017 @@
+# -----------------------------------------------------------------------------------
+# Swin2SR: Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration, https://arxiv.org/abs/
+# Written by Conde and Choi et al.
+# -----------------------------------------------------------------------------------
+
+import math
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+class WindowAttention(nn.Module):
+    r""" Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+        pretrained_window_size (tuple[int]): The height and width of the window in pre-training.
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, attn_drop=0., proj_drop=0.,
+                 pretrained_window_size=[0, 0]):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.pretrained_window_size = pretrained_window_size
+        self.num_heads = num_heads
+
+        self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))), requires_grad=True)
+
+        # mlp to generate continuous relative position bias
+        self.cpb_mlp = nn.Sequential(nn.Linear(2, 512, bias=True),
+                                     nn.ReLU(inplace=True),
+                                     nn.Linear(512, num_heads, bias=False))
+
+        # get relative_coords_table
+        relative_coords_h = torch.arange(-(self.window_size[0] - 1), self.window_size[0], dtype=torch.float32)
+        relative_coords_w = torch.arange(-(self.window_size[1] - 1), self.window_size[1], dtype=torch.float32)
+        relative_coords_table = torch.stack(
+            torch.meshgrid([relative_coords_h,
+                            relative_coords_w])).permute(1, 2, 0).contiguous().unsqueeze(0)  # 1, 2*Wh-1, 2*Ww-1, 2
+        if pretrained_window_size[0] > 0:
+            relative_coords_table[:, :, :, 0] /= (pretrained_window_size[0] - 1)
+            relative_coords_table[:, :, :, 1] /= (pretrained_window_size[1] - 1)
+        else:
+            relative_coords_table[:, :, :, 0] /= (self.window_size[0] - 1)
+            relative_coords_table[:, :, :, 1] /= (self.window_size[1] - 1)
+        relative_coords_table *= 8  # normalize to -8, 8
+        relative_coords_table = torch.sign(relative_coords_table) * torch.log2(
+            torch.abs(relative_coords_table) + 1.0) / np.log2(8)
+
+        self.register_buffer("relative_coords_table", relative_coords_table)
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=False)
+        if qkv_bias:
+            self.q_bias = nn.Parameter(torch.zeros(dim))
+            self.v_bias = nn.Parameter(torch.zeros(dim))
+        else:
+            self.q_bias = None
+            self.v_bias = None
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv_bias = None
+        if self.q_bias is not None:
+            qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))
+        qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
+        qkv = qkv.reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        # cosine attention
+        attn = (F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1))
+        logit_scale = torch.clamp(self.logit_scale, max=torch.log(torch.tensor(1. / 0.01)).to(self.logit_scale.device)).exp()
+        attn = attn * logit_scale
+
+        relative_position_bias_table = self.cpb_mlp(self.relative_coords_table).view(-1, self.num_heads)
+        relative_position_bias = relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        relative_position_bias = 16 * torch.sigmoid(relative_position_bias)
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f'dim={self.dim}, window_size={self.window_size}, ' \
+               f'pretrained_window_size={self.pretrained_window_size}, num_heads={self.num_heads}'
+
+    def flops(self, N):
+        # calculate flops for 1 window with token length of N
+        flops = 0
+        # qkv = self.qkv(x)
+        flops += N * self.dim * 3 * self.dim
+        # attn = (q @ k.transpose(-2, -1))
+        flops += self.num_heads * N * (self.dim // self.num_heads) * N
+        #  x = (attn @ v)
+        flops += self.num_heads * N * N * (self.dim // self.num_heads)
+        # x = self.proj(x)
+        flops += N * self.dim * self.dim
+        return flops
+
+class SwinTransformerBlock(nn.Module):
+    r""" Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resulotion.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+        pretrained_window_size (int): Window size in pre-training.
+    """
+
+    def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm, pretrained_window_size=0):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        if min(self.input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(self.input_resolution)
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop,
+            pretrained_window_size=to_2tuple(pretrained_window_size))
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        if self.shift_size > 0:
+            attn_mask = self.calculate_mask(self.input_resolution)
+        else:
+            attn_mask = None
+
+        self.register_buffer("attn_mask", attn_mask)
+        
+    def calculate_mask(self, x_size):
+        # calculate attention mask for SW-MSA
+        H, W = x_size
+        img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        return attn_mask        
+
+    def forward(self, x, x_size):
+        H, W = x_size
+        B, L, C = x.shape
+        #assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = x.view(B, H, W, C)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
+        if self.input_resolution == x_size:
+            attn_windows = self.attn(x_windows, mask=self.attn_mask)  # nW*B, window_size*window_size, C
+        else:
+            attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device))
+            
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+        x = x.view(B, H * W, C)
+        x = shortcut + self.drop_path(self.norm1(x))
+
+        # FFN
+        x = x + self.drop_path(self.norm2(self.mlp(x)))
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
+               f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
+
+    def flops(self):
+        flops = 0
+        H, W = self.input_resolution
+        # norm1
+        flops += self.dim * H * W
+        # W-MSA/SW-MSA
+        nW = H * W / self.window_size / self.window_size
+        flops += nW * self.attn.flops(self.window_size * self.window_size)
+        # mlp
+        flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
+        # norm2
+        flops += self.dim * H * W
+        return flops
+
+class PatchMerging(nn.Module):
+    r""" Patch Merging Layer.
+    Args:
+        input_resolution (tuple[int]): Resolution of input feature.
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(2 * dim)
+
+    def forward(self, x):
+        """
+        x: B, H*W, C
+        """
+        H, W = self.input_resolution
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
+
+        x = x.view(B, H, W, C)
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.reduction(x)
+        x = self.norm(x)
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f"input_resolution={self.input_resolution}, dim={self.dim}"
+
+    def flops(self):
+        H, W = self.input_resolution
+        flops = (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
+        flops += H * W * self.dim // 2
+        return flops    
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+        pretrained_window_size (int): Local window size in pre-training.
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
+                 mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,
+                 pretrained_window_size=0):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
+                                 num_heads=num_heads, window_size=window_size,
+                                 shift_size=0 if (i % 2 == 0) else window_size // 2,
+                                 mlp_ratio=mlp_ratio,
+                                 qkv_bias=qkv_bias,
+                                 drop=drop, attn_drop=attn_drop,
+                                 drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                                 norm_layer=norm_layer,
+                                 pretrained_window_size=pretrained_window_size)
+            for i in range(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, x_size):
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, x_size)
+            else:
+                x = blk(x, x_size)
+        if self.downsample is not None:
+            x = self.downsample(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
+
+    def flops(self):
+        flops = 0
+        for blk in self.blocks:
+            flops += blk.flops()
+        if self.downsample is not None:
+            flops += self.downsample.flops()
+        return flops
+
+    def _init_respostnorm(self):
+        for blk in self.blocks:
+            nn.init.constant_(blk.norm1.bias, 0)
+            nn.init.constant_(blk.norm1.weight, 0)
+            nn.init.constant_(blk.norm2.bias, 0)
+            nn.init.constant_(blk.norm2.weight, 0)
+            
+class PatchEmbed(nn.Module):
+    r""" Image to Patch Embedding
+    Args:
+        img_size (int): Image size.  Default: 224.
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        # FIXME look at relaxing size constraints
+        # assert H == self.img_size[0] and W == self.img_size[1],
+        #     f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x).flatten(2).transpose(1, 2)  # B Ph*Pw C
+        if self.norm is not None:
+            x = self.norm(x)
+        return x
+
+    def flops(self):
+        Ho, Wo = self.patches_resolution
+        flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
+        if self.norm is not None:
+            flops += Ho * Wo * self.embed_dim
+        return flops           
+
+class RSTB(nn.Module):
+    """Residual Swin Transformer Block (RSTB).
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+        img_size: Input image size.
+        patch_size: Patch size.
+        resi_connection: The convolutional block before residual connection.
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
+                 mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,
+                 img_size=224, patch_size=4, resi_connection='1conv'):
+        super(RSTB, self).__init__()
+
+        self.dim = dim
+        self.input_resolution = input_resolution
+
+        self.residual_group = BasicLayer(dim=dim,
+                                         input_resolution=input_resolution,
+                                         depth=depth,
+                                         num_heads=num_heads,
+                                         window_size=window_size,
+                                         mlp_ratio=mlp_ratio,
+                                         qkv_bias=qkv_bias, 
+                                         drop=drop, attn_drop=attn_drop,
+                                         drop_path=drop_path,
+                                         norm_layer=norm_layer,
+                                         downsample=downsample,
+                                         use_checkpoint=use_checkpoint)
+
+        if resi_connection == '1conv':
+            self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
+        elif resi_connection == '3conv':
+            # to save parameters and memory
+            self.conv = nn.Sequential(nn.Conv2d(dim, dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                      nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
+                                      nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                      nn.Conv2d(dim // 4, dim, 3, 1, 1))
+
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=dim, embed_dim=dim,
+            norm_layer=None)
+
+        self.patch_unembed = PatchUnEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=dim, embed_dim=dim,
+            norm_layer=None)
+
+    def forward(self, x, x_size):
+        return self.patch_embed(self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))) + x
+
+    def flops(self):
+        flops = 0
+        flops += self.residual_group.flops()
+        H, W = self.input_resolution
+        flops += H * W * self.dim * self.dim * 9
+        flops += self.patch_embed.flops()
+        flops += self.patch_unembed.flops()
+
+        return flops
+
+class PatchUnEmbed(nn.Module):
+    r""" Image to Patch Unembedding
+
+    Args:
+        img_size (int): Image size.  Default: 224.
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+    def forward(self, x, x_size):
+        B, HW, C = x.shape
+        x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1])  # B Ph*Pw C
+        return x
+
+    def flops(self):
+        flops = 0
+        return flops
+
+
+class Upsample(nn.Sequential):
+    """Upsample module.
+
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+    """
+
+    def __init__(self, scale, num_feat):
+        m = []
+        if (scale & (scale - 1)) == 0:  # scale = 2^n
+            for _ in range(int(math.log(scale, 2))):
+                m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
+                m.append(nn.PixelShuffle(2))
+        elif scale == 3:
+            m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
+            m.append(nn.PixelShuffle(3))
+        else:
+            raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.')
+        super(Upsample, self).__init__(*m)
+        
+class Upsample_hf(nn.Sequential):
+    """Upsample module.
+
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+    """
+
+    def __init__(self, scale, num_feat):
+        m = []
+        if (scale & (scale - 1)) == 0:  # scale = 2^n
+            for _ in range(int(math.log(scale, 2))):
+                m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
+                m.append(nn.PixelShuffle(2))
+        elif scale == 3:
+            m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
+            m.append(nn.PixelShuffle(3))
+        else:
+            raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.')
+        super(Upsample_hf, self).__init__(*m)        
+
+
+class UpsampleOneStep(nn.Sequential):
+    """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
+       Used in lightweight SR to save parameters.
+
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+
+    """
+
+    def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
+        self.num_feat = num_feat
+        self.input_resolution = input_resolution
+        m = []
+        m.append(nn.Conv2d(num_feat, (scale ** 2) * num_out_ch, 3, 1, 1))
+        m.append(nn.PixelShuffle(scale))
+        super(UpsampleOneStep, self).__init__(*m)
+
+    def flops(self):
+        H, W = self.input_resolution
+        flops = H * W * self.num_feat * 3 * 9
+        return flops
+    
+    
+
+class Swin2SR(nn.Module):
+    r""" Swin2SR
+        A PyTorch impl of : `Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration`.
+
+    Args:
+        img_size (int | tuple(int)): Input image size. Default 64
+        patch_size (int | tuple(int)): Patch size. Default: 1
+        in_chans (int): Number of input image channels. Default: 3
+        embed_dim (int): Patch embedding dimension. Default: 96
+        depths (tuple(int)): Depth of each Swin Transformer layer.
+        num_heads (tuple(int)): Number of attention heads in different layers.
+        window_size (int): Window size. Default: 7
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        drop_rate (float): Dropout rate. Default: 0
+        attn_drop_rate (float): Attention dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
+        upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
+        img_range: Image range. 1. or 255.
+        upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
+        resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
+    """
+
+    def __init__(self, img_size=64, patch_size=1, in_chans=3,
+                 embed_dim=96, depths=[6, 6, 6, 6], num_heads=[6, 6, 6, 6],
+                 window_size=7, mlp_ratio=4., qkv_bias=True, 
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
+                 norm_layer=nn.LayerNorm, ape=False, patch_norm=True,
+                 use_checkpoint=False, upscale=2, img_range=1., upsampler='', resi_connection='1conv',
+                 **kwargs):
+        super(Swin2SR, self).__init__()
+        num_in_ch = in_chans
+        num_out_ch = in_chans
+        num_feat = 64
+        self.img_range = img_range
+        if in_chans == 3:
+            rgb_mean = (0.4488, 0.4371, 0.4040)
+            self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
+        else:
+            self.mean = torch.zeros(1, 1, 1, 1)
+        self.upscale = upscale
+        self.upsampler = upsampler
+        self.window_size = window_size
+
+        #####################################################################################################
+        ################################### 1, shallow feature extraction ###################################
+        self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)
+
+        #####################################################################################################
+        ################################### 2, deep feature extraction ######################################
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.num_features = embed_dim
+        self.mlp_ratio = mlp_ratio
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+        num_patches = self.patch_embed.num_patches
+        patches_resolution = self.patch_embed.patches_resolution
+        self.patches_resolution = patches_resolution
+
+        # merge non-overlapping patches into image
+        self.patch_unembed = PatchUnEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        # absolute position embedding
+        if self.ape:
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+
+        # build Residual Swin Transformer blocks (RSTB)
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = RSTB(dim=embed_dim,
+                         input_resolution=(patches_resolution[0],
+                                           patches_resolution[1]),
+                         depth=depths[i_layer],
+                         num_heads=num_heads[i_layer],
+                         window_size=window_size,
+                         mlp_ratio=self.mlp_ratio,
+                         qkv_bias=qkv_bias, 
+                         drop=drop_rate, attn_drop=attn_drop_rate,
+                         drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],  # no impact on SR results
+                         norm_layer=norm_layer,
+                         downsample=None,
+                         use_checkpoint=use_checkpoint,
+                         img_size=img_size,
+                         patch_size=patch_size,
+                         resi_connection=resi_connection
+
+                         )
+            self.layers.append(layer)
+            
+        if self.upsampler == 'pixelshuffle_hf':
+            self.layers_hf = nn.ModuleList()
+            for i_layer in range(self.num_layers):
+                layer = RSTB(dim=embed_dim,
+                             input_resolution=(patches_resolution[0],
+                                               patches_resolution[1]),
+                             depth=depths[i_layer],
+                             num_heads=num_heads[i_layer],
+                             window_size=window_size,
+                             mlp_ratio=self.mlp_ratio,
+                             qkv_bias=qkv_bias, 
+                             drop=drop_rate, attn_drop=attn_drop_rate,
+                             drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],  # no impact on SR results
+                             norm_layer=norm_layer,
+                             downsample=None,
+                             use_checkpoint=use_checkpoint,
+                             img_size=img_size,
+                             patch_size=patch_size,
+                             resi_connection=resi_connection
+
+                             )
+                self.layers_hf.append(layer)
+                        
+        self.norm = norm_layer(self.num_features)
+
+        # build the last conv layer in deep feature extraction
+        if resi_connection == '1conv':
+            self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
+        elif resi_connection == '3conv':
+            # to save parameters and memory
+            self.conv_after_body = nn.Sequential(nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1),
+                                                 nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                                 nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0),
+                                                 nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                                 nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1))
+
+        #####################################################################################################
+        ################################ 3, high quality image reconstruction ################################
+        if self.upsampler == 'pixelshuffle':
+            # for classical SR
+            self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                                                      nn.LeakyReLU(inplace=True))
+            self.upsample = Upsample(upscale, num_feat)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+        elif self.upsampler == 'pixelshuffle_aux':
+            self.conv_bicubic = nn.Conv2d(num_in_ch, num_feat, 3, 1, 1)
+            self.conv_before_upsample = nn.Sequential(
+                nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                nn.LeakyReLU(inplace=True))
+            self.conv_aux = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            self.conv_after_aux = nn.Sequential(
+                nn.Conv2d(3, num_feat, 3, 1, 1),
+                nn.LeakyReLU(inplace=True))            
+            self.upsample = Upsample(upscale, num_feat)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            
+        elif self.upsampler == 'pixelshuffle_hf':
+            self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                                                      nn.LeakyReLU(inplace=True))
+            self.upsample = Upsample(upscale, num_feat)
+            self.upsample_hf = Upsample_hf(upscale, num_feat)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            self.conv_first_hf = nn.Sequential(nn.Conv2d(num_feat, embed_dim, 3, 1, 1),
+                                                      nn.LeakyReLU(inplace=True))
+            self.conv_after_body_hf = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
+            self.conv_before_upsample_hf = nn.Sequential(
+                nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                nn.LeakyReLU(inplace=True))
+            self.conv_last_hf = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            
+        elif self.upsampler == 'pixelshuffledirect':
+            # for lightweight SR (to save parameters)
+            self.upsample = UpsampleOneStep(upscale, embed_dim, num_out_ch,
+                                            (patches_resolution[0], patches_resolution[1]))
+        elif self.upsampler == 'nearest+conv':
+            # for real-world SR (less artifacts)
+            assert self.upscale == 4, 'only support x4 now.'
+            self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                                                      nn.LeakyReLU(inplace=True))
+            self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+        else:
+            # for image denoising and JPEG compression artifact reduction
+            self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'absolute_pos_embed'}
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        return {'relative_position_bias_table'}
+
+    def check_image_size(self, x):
+        _, _, h, w = x.size()
+        mod_pad_h = (self.window_size - h % self.window_size) % self.window_size
+        mod_pad_w = (self.window_size - w % self.window_size) % self.window_size
+        x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect')
+        return x
+
+    def forward_features(self, x):
+        x_size = (x.shape[2], x.shape[3])
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        for layer in self.layers:
+            x = layer(x, x_size)
+
+        x = self.norm(x)  # B L C
+        x = self.patch_unembed(x, x_size)
+
+        return x
+    
+    def forward_features_hf(self, x):
+        x_size = (x.shape[2], x.shape[3])
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        for layer in self.layers_hf:
+            x = layer(x, x_size)
+
+        x = self.norm(x)  # B L C
+        x = self.patch_unembed(x, x_size)
+
+        return x    
+
+    def forward(self, x):
+        H, W = x.shape[2:]
+        x = self.check_image_size(x)
+
+        self.mean = self.mean.type_as(x)
+        x = (x - self.mean) * self.img_range
+
+        if self.upsampler == 'pixelshuffle':
+            # for classical SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            x = self.conv_last(self.upsample(x))
+        elif self.upsampler == 'pixelshuffle_aux':
+            bicubic = F.interpolate(x, size=(H * self.upscale, W * self.upscale), mode='bicubic', align_corners=False)
+            bicubic = self.conv_bicubic(bicubic)
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            aux = self.conv_aux(x) # b, 3, LR_H, LR_W
+            x = self.conv_after_aux(aux)
+            x = self.upsample(x)[:, :, :H * self.upscale, :W * self.upscale] + bicubic[:, :, :H * self.upscale, :W * self.upscale]
+            x = self.conv_last(x)
+            aux = aux / self.img_range + self.mean
+        elif self.upsampler == 'pixelshuffle_hf':
+            # for classical SR with HF
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x_before = self.conv_before_upsample(x)
+            x_out = self.conv_last(self.upsample(x_before))
+            
+            x_hf = self.conv_first_hf(x_before)
+            x_hf = self.conv_after_body_hf(self.forward_features_hf(x_hf)) + x_hf
+            x_hf = self.conv_before_upsample_hf(x_hf)
+            x_hf = self.conv_last_hf(self.upsample_hf(x_hf))
+            x = x_out + x_hf
+            x_hf = x_hf / self.img_range + self.mean
+
+        elif self.upsampler == 'pixelshuffledirect':
+            # for lightweight SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.upsample(x)
+        elif self.upsampler == 'nearest+conv':
+            # for real-world SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            x = self.lrelu(self.conv_up1(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
+            x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
+            x = self.conv_last(self.lrelu(self.conv_hr(x)))
+        else:
+            # for image denoising and JPEG compression artifact reduction
+            x_first = self.conv_first(x)
+            res = self.conv_after_body(self.forward_features(x_first)) + x_first
+            x = x + self.conv_last(res)
+        
+        x = x / self.img_range + self.mean
+        if self.upsampler == "pixelshuffle_aux":
+            return x[:, :, :H*self.upscale, :W*self.upscale], aux
+        
+        elif self.upsampler == "pixelshuffle_hf":
+            x_out = x_out / self.img_range + self.mean
+            return x_out[:, :, :H*self.upscale, :W*self.upscale], x[:, :, :H*self.upscale, :W*self.upscale], x_hf[:, :, :H*self.upscale, :W*self.upscale]
+        
+        else:
+            return x[:, :, :H*self.upscale, :W*self.upscale]
+
+    def flops(self):
+        flops = 0
+        H, W = self.patches_resolution
+        flops += H * W * 3 * self.embed_dim * 9
+        flops += self.patch_embed.flops()
+        for i, layer in enumerate(self.layers):
+            flops += layer.flops()
+        flops += H * W * 3 * self.embed_dim * self.embed_dim
+        flops += self.upsample.flops()
+        return flops
+
+
+if __name__ == '__main__':
+    upscale = 4
+    window_size = 8
+    height = (1024 // upscale // window_size + 1) * window_size
+    width = (720 // upscale // window_size + 1) * window_size
+    model = Swin2SR(upscale=2, img_size=(height, width),
+                   window_size=window_size, img_range=1., depths=[6, 6, 6, 6],
+                   embed_dim=60, num_heads=[6, 6, 6, 6], mlp_ratio=2, upsampler='pixelshuffledirect')
+    print(model)
+    print(height, width, model.flops() / 1e9)
+
+    x = torch.randn((1, 3, height, width))
+    x = model(x)
+    print(x.shape)

extensions-builtin/prompt-bracket-checker/javascript/prompt-bracket-checker.js

@@ -0,0 +1,103 @@
+// Stable Diffusion WebUI - Bracket checker
+// Version 1.0
+// By Hingashi no Florin/Bwin4L
+// Counts open and closed brackets (round, square, curly) in the prompt and negative prompt text boxes in the txt2img and img2img tabs.
+// If there's a mismatch, the keyword counter turns red and if you hover on it, a tooltip tells you what's wrong.
+
+function checkBrackets(evt, textArea, counterElt) {
+  errorStringParen  = '(...) - Different number of opening and closing parentheses detected.\n';
+  errorStringSquare = '[...] - Different number of opening and closing square brackets detected.\n';
+  errorStringCurly  = '{...} - Different number of opening and closing curly brackets detected.\n';
+
+  openBracketRegExp = /\(/g;
+  closeBracketRegExp = /\)/g;
+
+  openSquareBracketRegExp = /\[/g;
+  closeSquareBracketRegExp = /\]/g;
+
+  openCurlyBracketRegExp = /\{/g;
+  closeCurlyBracketRegExp = /\}/g;
+
+  totalOpenBracketMatches = 0;
+  totalCloseBracketMatches = 0;
+  totalOpenSquareBracketMatches = 0;
+  totalCloseSquareBracketMatches = 0;
+  totalOpenCurlyBracketMatches = 0;
+  totalCloseCurlyBracketMatches = 0;
+
+  openBracketMatches = textArea.value.match(openBracketRegExp);
+  if(openBracketMatches) {
+    totalOpenBracketMatches = openBracketMatches.length;
+  }
+
+  closeBracketMatches = textArea.value.match(closeBracketRegExp);
+  if(closeBracketMatches) {
+    totalCloseBracketMatches = closeBracketMatches.length;
+  }
+
+  openSquareBracketMatches = textArea.value.match(openSquareBracketRegExp);
+  if(openSquareBracketMatches) {
+    totalOpenSquareBracketMatches = openSquareBracketMatches.length;
+  }
+
+  closeSquareBracketMatches = textArea.value.match(closeSquareBracketRegExp);
+  if(closeSquareBracketMatches) {
+    totalCloseSquareBracketMatches = closeSquareBracketMatches.length;
+  }
+
+  openCurlyBracketMatches = textArea.value.match(openCurlyBracketRegExp);
+  if(openCurlyBracketMatches) {
+    totalOpenCurlyBracketMatches = openCurlyBracketMatches.length;
+  }
+
+  closeCurlyBracketMatches = textArea.value.match(closeCurlyBracketRegExp);
+  if(closeCurlyBracketMatches) {
+    totalCloseCurlyBracketMatches = closeCurlyBracketMatches.length;
+  }
+
+  if(totalOpenBracketMatches != totalCloseBracketMatches) {
+    if(!counterElt.title.includes(errorStringParen)) {
+      counterElt.title += errorStringParen;
+    }
+  } else {
+    counterElt.title = counterElt.title.replace(errorStringParen, '');
+  }
+
+  if(totalOpenSquareBracketMatches != totalCloseSquareBracketMatches) {
+    if(!counterElt.title.includes(errorStringSquare)) {
+      counterElt.title += errorStringSquare;
+    }
+  } else {
+    counterElt.title = counterElt.title.replace(errorStringSquare, '');
+  }
+
+  if(totalOpenCurlyBracketMatches != totalCloseCurlyBracketMatches) {
+    if(!counterElt.title.includes(errorStringCurly)) {
+      counterElt.title += errorStringCurly;
+    }
+  } else {
+    counterElt.title = counterElt.title.replace(errorStringCurly, '');
+  }
+
+  if(counterElt.title != '') {
+    counterElt.classList.add('error');
+  } else {
+    counterElt.classList.remove('error');
+  }
+}
+
+function setupBracketChecking(id_prompt, id_counter){
+    var textarea = gradioApp().querySelector("#" + id_prompt + " > label > textarea");
+    var counter = gradioApp().getElementById(id_counter)
+
+    textarea.addEventListener("input", function(evt){
+        checkBrackets(evt, textarea, counter)
+    });
+}
+
+onUiLoaded(function(){
+    setupBracketChecking('txt2img_prompt', 'txt2img_token_counter')
+    setupBracketChecking('txt2img_neg_prompt', 'txt2img_negative_token_counter')
+    setupBracketChecking('img2img_prompt', 'img2img_token_counter')
+    setupBracketChecking('img2img_neg_prompt', 'img2img_negative_token_counter')
+})