import gc import io from collections import namedtuple from typing import Tuple import gradio as gr import matplotlib.pyplot as plt import numpy as np import scipy.sparse import torch import torch.nn.functional as F import torchvision.transforms.functional as TF from gradio.inputs import Image as GradioInputImage from gradio.outputs import Image as GradioOutputImage from matplotlib.pyplot import get_cmap from PIL import Image from scipy.sparse.linalg import eigsh from torch.utils.hooks import RemovableHandle from torchvision import transforms from torchvision.utils import make_grid def get_model(name: str): if 'dino' in name: model = torch.hub.load('facebookresearch/dino:main', name) model.fc = torch.nn.Identity() val_transform = get_transform(name) patch_size = model.patch_embed.patch_size num_heads = model.blocks[0].attn.num_heads elif name in ['mocov3_vits16', 'mocov3_vitb16']: model = torch.hub.load('facebookresearch/dino:main', name.replace('mocov3', 'dino')) checkpoint_file, size_char = { 'mocov3_vits16': ('vit-s-300ep-timm-format.pth', 's'), 'mocov3_vitb16': ('vit-b-300ep-timm-format.pth', 'b'), }[name] url = f'https://dl.fbaipublicfiles.com/moco-v3/vit-{size_char}-300ep/vit-{size_char}-300ep.pth.tar' checkpoint = torch.hub.load_state_dict_from_url(url) model.load_state_dict(checkpoint['model']) model.fc = torch.nn.Identity() val_transform = get_transform(name) patch_size = model.patch_embed.patch_size num_heads = model.blocks[0].attn.num_heads else: raise ValueError(f'Unsupported model: {name}') model = model.eval() return model, val_transform, patch_size, num_heads def get_transform(name: str): if any(x in name for x in ('dino', 'mocov3', 'convnext', )): normalize = transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)) transform = transforms.Compose([ transforms.Resize(size=512, interpolation=TF.InterpolationMode.BICUBIC, max_size=1024), transforms.ToTensor(), normalize ]) else: raise NotImplementedError() return transform def get_diagonal(W: scipy.sparse.csr_matrix, threshold: float = 1e-12): D = W.dot(np.ones(W.shape[1], W.dtype)) D[D < threshold] = 1.0 # Prevent division by zero. D = scipy.sparse.diags(D) return D # Cache torch.cuda.empty_cache() # Parameters model_name = 'dino_vitb16' # TODO: Figure out how to make this user-editable K = 5 # Load model model, val_transform, patch_size, num_heads = get_model(model_name) # Add hook which_block = -1 if 'dino' in model_name or 'mocov3' in model_name: feat_out = {} def hook_fn_forward_qkv(module, input, output): feat_out["qkv"] = output handle: RemovableHandle = model._modules["blocks"][which_block]._modules["attn"]._modules["qkv"].register_forward_hook( hook_fn_forward_qkv ) else: raise ValueError(model_name) # GPU if torch.cuda.is_available(): print("CUDA is available, using GPU.") device = torch.device("cuda") model.to(device) else: print("CUDA is not available, using CPU.") device = torch.device("cpu") @torch.no_grad() def segment(inp: Image): # NOTE: The image is already resized to the desired size. # Preprocess image images: torch.Tensor = val_transform(inp) images = images.unsqueeze(0).to(device) # Reshape image P = patch_size B, C, H, W = images.shape H_patch, W_patch = H // P, W // P H_pad, W_pad = H_patch * P, W_patch * P T = H_patch * W_patch + 1 # number of tokens, add 1 for [CLS] # Crop image to be a multiple of the patch size images = images[:, :, :H_pad, :W_pad] # Extract features if 'dino' in model_name or 'mocov3' in model_name: model.get_intermediate_layers(images)[0].squeeze(0) output_qkv = feat_out["qkv"].reshape(B, T, 3, num_heads, -1 // num_heads).permute(2, 0, 3, 1, 4) feats = output_qkv[1].transpose(1, 2).reshape(B, T, -1)[:, 1:, :].squeeze(0) else: raise ValueError(model_name) # Normalize features normalize = True if normalize: feats = F.normalize(feats, p=2, dim=-1) # Compute affinity matrix W_feat = (feats @ feats.T) # Feature affinities threshold_at_zero = True if threshold_at_zero: W_feat = (W_feat * (W_feat > 0)) W_feat = W_feat / W_feat.max() # NOTE: If features are normalized, this naturally does nothing W_feat = W_feat.cpu().numpy() # # NOTE: Here is where we would add the color information. For simplicity, we will not add it here. # W_comb = W_feat + W_color * image_color_lambda # combination # D_comb = np.array(get_diagonal(W_comb).todense()) # is dense or sparse faster? not sure, should check # Diagonal W_comb = W_feat D_comb = np.array(get_diagonal(W_comb).todense()) # is dense or sparse faster? not sure, should check # Compute eigenvectors try: eigenvalues, eigenvectors = eigsh(D_comb - W_comb, k=(K + 1), sigma=0, which='LM', M=D_comb) except: eigenvalues, eigenvectors = eigsh(D_comb - W_comb, k=(K + 1), which='SM', M=D_comb) eigenvalues = torch.from_numpy(eigenvalues) eigenvectors = torch.from_numpy(eigenvectors.T).float() # Resolve sign ambiguity for k in range(eigenvectors.shape[0]): if 0.5 < torch.mean((eigenvectors[k] > 0).float()).item() < 1.0: # reverse segment eigenvectors[k] = 0 - eigenvectors[k] # Arrange eigenvectors into grid # cmap = get_cmap('viridis') output_images = [] # eigenvectors_upscaled = [] for i in range(1, K + 1): eigenvector = eigenvectors[i].reshape(1, 1, H_patch, W_patch) # .reshape(1, 1, H_pad, W_pad) eigenvector: torch.Tensor = F.interpolate(eigenvector, size=(H_pad, W_pad), mode='bilinear', align_corners=False) # slightly off, but for visualizations this is okay buffer = io.BytesIO() plt.imsave(buffer, eigenvector.squeeze().numpy(), format='png') # save to a temporary location buffer.seek(0) eigenvector_vis = Image.open(buffer).convert('RGB') # eigenvector_vis = TF.to_tensor(eigenvector_vis).unsqueeze(0) eigenvector_vis = np.array(eigenvector_vis) # eigenvectors_upscaled.append(eigenvector) output_images.append(eigenvector_vis) # output_images = torch.cat(output_images, dim=0) # output_images = make_grid(output_images, nrow=8, pad_value=1) # Also add CRF if False: # Imports import denseCRF # Parameters ParamsCRF = namedtuple('ParamsCRF', 'w1 alpha beta w2 gamma it') DEFAULT_CRF_PARAMS = ParamsCRF( w1 = 6, # weight of bilateral term # 10.0, alpha = 40, # spatial std # 80, beta = 13, # rgb std # 13, w2 = 3, # weight of spatial term # 3.0, gamma = 3, # spatial std # 3, it = 5.0, # iteration # 5.0, ) # Get unary potentials unary_potentials = eigenvectors_upscaled[0].squeeze(1).squeeze(0) unary_potentials = (unary_potentials - unary_potentials.min()) / (unary_potentials.max() - unary_potentials.min()) unary_potentials_np = torch.stack((1 - unary_potentials, unary_potentials), dim=-1).cpu().numpy() img_np = images.cpu().numpy().transpose(0, 2, 3, 1) img_np = (img_np * 255).astype(np.uint8)[0] # Return result of CRF out = denseCRF.densecrf(img_np, unary_potentials_np, DEFAULT_CRF_PARAMS) out = out * 255 output_images.append(out) # # Postprocess for Gradio # output_images = np.array(TF.to_pil_image(output_images)) print(f'{len(output_images)=}') # Garbage collection and other memory-related things gc.collect() del eigenvector, eigenvector_vis, eigenvectors, W_comb, D_comb return output_images # Placeholders input_placeholders = GradioInputImage(source="upload", tool="editor", type="pil") # output_placeholders = GradioOutputImage(type="numpy", label=f"Eigenvectors") output_placeholders = [GradioOutputImage(type="numpy", label=(f"Eigenvector {i}")) for i in range(K)] # Metadata examples = [f"examples/{stem}.jpg" for stem in [ '2008_000099', '2008_000499', '2007_009446', '2007_001586', '2010_001256', '2008_000764', '2008_000705', # '2007_000039' ]] title = "Demo: Deep Spectral Methods for Unsupervised Localization and Segmentation" description = """ This is a demo of Deep Spectral Methods: A Surprisingly Strong Baseline for Unsupervised Semantic Segmentation and Localization (CVPR 2022 Oral). Our method decomposes an image into a set of soft segments in a completely unsupervised manner. Specifically, we extract the Laplacian eigenvectors of a feature affinity matrix from a large self-supervised network, and we find that we find that these eigenvectors can be readily used to localize and segment objects. Below, you can upload an image (or select one of the examples) and see how our method decomposes it into soft segments. Hopefully it will localize some of the objects or semantic segments in your image! Note: Due to memory constraints on Huggingface, we have to resize the image to a maximum size length of 512px after upload. For best-quality results at full resolution, use our GitHub repository. """ thumbnail = "https://raw.githubusercontent.com/gradio-app/hub-echonet/master/thumbnail.png" # Gradio gr.Interface( segment, input_placeholders, output_placeholders, examples=examples, allow_flagging=False, analytics_enabled=False, title=title, description=description, thumbnail=thumbnail ).launch()