app.py

import gradio as gr

import cv2

import torch
import torch.utils.data as data
from torchvision import transforms
from torch import nn
import torch.nn.functional as F

import matplotlib.pyplot as plt
from matplotlib import cm
from matplotlib import colors
from mpl_toolkits.axes_grid1 import ImageGrid

import fire_network

import numpy as np

from PIL import Image

# Possible Scales for multiscale inference
scales = [2.0, 1.414, 1.0, 0.707, 0.5, 0.353, 0.25] 

device = 'cpu'

# Load nets
state = torch.load('fire.pth', map_location='cpu')
state['net_params']['pretrained'] = None # no need for imagenet pretrained model
net_sfm = fire_network.init_network(**state['net_params']).to(device)
net_sfm.load_state_dict(state['state_dict'])
dim_red_params_dict = {}
for name, param in net_sfm.named_parameters():
    if 'dim_reduction' in name:
        dim_red_params_dict[name] = param


state2 = torch.load('fire_imagenet.pth', map_location='cpu')
state2['net_params'] = state['net_params']
state2['state_dict'] = dict(state2['state_dict'], **dim_red_params_dict);
net_imagenet = fire_network.init_network(**state['net_params']).to(device)
net_imagenet.load_state_dict(state2['state_dict'], strict=False)

# ---------------------------------------
transform = transforms.Compose([
        transforms.Resize(1024),
        transforms.ToTensor(), 
        transforms.Normalize(**dict(zip(["mean", "std"], net_sfm.runtime['mean_std'])))
        ])
# ---------------------------------------

# class ImgDataset(data.Dataset):
#     def __init__(self, images, imsize):
#         self.images = images
#         self.imsize = imsize
#         self.transform = transforms.Compose([transforms.ToTensor(), \
#             transforms.Normalize(**dict(zip(["mean", "std"], net.runtime['mean_std'])))])
#     def __getitem__(self, index):
#         img = self.images[index]
#         img.thumbnail((self.imsize, self.imsize), Image.Resampling.LANCZOS)
#         print('after imresize:', img.size)
#         return  self.transform(img)
#     def __len__(self):
#         return len(self.images)

# ---------------------------------------    

def match(query_feat, pos_feat, LoweRatioTh=0.9):
    # first perform reciprocal nn
    dist = torch.cdist(query_feat, pos_feat)
    print('dist.size',dist.size())
    best1 = torch.argmin(dist, dim=1)
    best2 = torch.argmin(dist, dim=0)
    print('best2.size',best2.size())
    arange = torch.arange(best2.size(0))
    reciprocal = best1[best2]==arange
    # check Lowe ratio test
    dist2 = dist.clone()
    dist2[best2,arange] = float('Inf')
    dist2_second2 = torch.argmin(dist2, dim=0)
    ratio1to2 = dist[best2,arange] / dist2_second2
    valid = torch.logical_and(reciprocal, ratio1to2<=LoweRatioTh)
    pindices = torch.where(valid)[0]
    qindices = best2[pindices]
    # keep only the ones with same indices 
    valid = pindices==qindices
    return pindices[valid]
    

# sf_idx_ = [55, 14, 5, 4, 52, 57, 40, 9]
def clear_figures():
    plt.figure().clear()
    plt.close()
    plt.cla()
    plt.clf()

col = plt.get_cmap('tab10')

def generate_matching_superfeatures(im1, im2, Imagenet_model=False, scale_id=6, threshold=50, random_mode=False, sf_ids=''): #, only_matching=True):
    print('im1:', im1.size)
    print('im2:', im2.size)

    clear_figures()

    net = net_sfm
    if Imagenet_model:
        net = net_imagenet

    # dataset_ = ImgDataset(images=[im1, im2], imsize=1024)
    # loader = torch.utils.data.DataLoader(dataset_, shuffle=False, pin_memory=True)


    im1_tensor = transform(im1).unsqueeze(0)
    im2_tensor = transform(im2).unsqueeze(0)

    im1_cv = np.array(im1)[:, :, ::-1].copy() 
    im2_cv = np.array(im2)[:, :, ::-1].copy() 

    # extract features
    with torch.no_grad():
        output1 = net.get_superfeatures(im1_tensor.to(device), scales=[scales[scale_id]])
        feats1 = output1[0][0]
        attns1 = output1[1][0]
        strenghts1 = output1[2][0]

        output2 = net.get_superfeatures(im2_tensor.to(device), scales=[scales[scale_id]])
        feats2 = output2[0][0]
        attns2 = output2[1][0]
        strenghts2 = output2[2][0]

        feats1n = F.normalize(torch.t(torch.squeeze(feats1)), dim=1)
        feats2n = F.normalize(torch.t(torch.squeeze(feats2)), dim=1)
        print('feats1n.shape', feats1n.shape)
        ind_match = match(feats1n, feats2n)
        print('ind', ind_match)
        print('ind.shape', ind_match.shape)
        # outputs = []
        # for im_tensor in loader:
        #     outputs.append(net.get_superfeatures(im_tensor.to(device), scales=[scales[scale_id]]))
        # feats1 = outputs[0][0][0]
        # attns1 = outputs[0][1][0]
        # strenghts1 = outputs[0][2][0]
        # feats2 = outputs[1][0][0]
        # attns2 = outputs[1][1][0]
        # strenghts2 = outputs[1][2][0]
    print(feats1.shape, feats2.shape)
    print(attns1.shape, attns2.shape)
    print(strenghts1.shape, strenghts2.shape)
    
    # which sf 
    sf_idx_ = [55, 14, 5, 4, 52, 57, 40, 9]
    n_sf_ids = 8
    if random_mode or sf_ids == '':
        sf_idx_ = np.random.randint(256, size=n_sf_ids)
    else:
        sf_idx_ = map(int, sf_ids.strip().split(','))
        n_sf_ids = len(sf_idx_)
        
    # if only_matching:
    if random_mode:
        sf_idx_ = [int(jj) for jj in ind_match[np.random.randint(len(list(ind_match)), size=n_sf_ids)].numpy()]
        sf_idx_ = list( dict.fromkeys(sf_idx_) )
    else:
        sf_idx_ = [i for i in sf_idx_ if i in list(ind_match)]

    # Store all binary SF att maps to show them all at once in the end
    all_att_bin1 = []
    all_att_bin2 = []
    for n, i in enumerate(sf_idx_):
        # all_atts[n].append(attn[j][scale_id][0,i,:,:].numpy())
        att_heat = np.array(attns1[0,i,:,:].numpy(), dtype=np.float32)
        att_heat = np.uint8(att_heat / np.max(att_heat[:]) * 255.0)
        att_heat_bin  = np.where(att_heat>threshold, 255, 0)
        # print(att_heat_bin)
        all_att_bin1.append(att_heat_bin)

        att_heat = np.array(attns2[0,i,:,:].numpy(), dtype=np.float32)
        att_heat = np.uint8(att_heat / np.max(att_heat[:]) * 255.0)
        att_heat_bin  = np.where(att_heat>threshold, 255, 0)
        all_att_bin2.append(att_heat_bin)

    
    fin_img = []
    img1rsz = np.copy(im1_cv)
    print('im1:', im1.size)
    print('img1rsz:', img1rsz.shape)
    for j, att in enumerate(all_att_bin1):
        att = cv2.resize(att, im1.size, interpolation=cv2.INTER_NEAREST)
        # att = cv2.resize(att, imgz[i].shape[:2][::-1], interpolation=cv2.INTER_CUBIC)
        # att = cv2.resize(att, imgz[i].shape[:2][::-1])
        # att = att.resize(shape)
        # att = resize(att, im1.size)
        mask2d = zip(*np.where(att==255))
        for m,n in mask2d:
            col_ = col.colors[j] if j < 7 else col.colors[j+1]
            if j == 0: col_ = col.colors[9]
            col_ = 255*np.array(colors.to_rgba(col_))[:3]
            img1rsz[m,n, :] = col_[::-1]   
            
    img2rsz = np.copy(im2_cv)
    print('im2:', im2.size)
    print('img2rsz:', img2rsz.shape)
    for j, att in enumerate(all_att_bin2):
        att = cv2.resize(att, im2.size, interpolation=cv2.INTER_NEAREST)
        # att = cv2.resize(att, imgz[i].shape[:2][::-1], interpolation=cv2.INTER_CUBIC)
        # # att = cv2.resize(att, imgz[i].shape[:2][::-1])
        # att = att.resize(im2.shape)
        # print('att:', att.shape)
        mask2d = zip(*np.where(att==255))
        for m,n in mask2d:
            col_ = col.colors[j]
            # col_ = col.colors[j] if j < 7 else col.colors[j+1]
            # if j == 0: col_ = col.colors[9]
            col_ = 255*np.array(colors.to_rgba(col_))[:3]
            img2rsz[m,n, :] = col_[::-1]   

    fig1 = plt.figure(1)
    plt.imshow(cv2.cvtColor(img1rsz, cv2.COLOR_BGR2RGB))
    ax1 = plt.gca()
    # ax1.axis('scaled')
    ax1.axis('off')
    plt.tight_layout()    
    # fig1.canvas.draw()
    
    fig2 = plt.figure(2)
    plt.imshow(cv2.cvtColor(img2rsz, cv2.COLOR_BGR2RGB))
    ax2 = plt.gca()
    # ax2.axis('scaled')
    ax2.axis('off')
    plt.tight_layout()    
    # fig2.canvas.draw()
    
    f = lambda m,c: plt.plot([],[],marker=m, color=c, ls="none")[0]
    handles = [f("s", col.colors[i]) for i in range(n_sf_ids)]
    fig_leg = plt.figure(3)
    legend = plt.legend(handles, sf_idx_, framealpha=1, frameon=False, facecolor='w',fontsize=15, loc="center")
    # fig_leg  = legend.figure
    # fig_leg.canvas.draw()
    ax3 = plt.gca()
    # ax2.axis('scaled')
    ax3.axis('off')
    plt.tight_layout()    
    # bbox  = legend.get_window_extent().transformed(fig.dpi_scale_trans.inverted())
    

    return fig1, fig2, fig_leg
    # ','.join(map(str, sf_idx_))


# GRADIO APP
title = "Visualizing Super-features"
description = "This is a visualization demo for the ICLR 2022 paper <b><a href='https://github.com/naver/fire' target='_blank'>Learning Super-Features for Image Retrieval</a></p></b>" 
article = "<p style='text-align: center'><a href='https://github.com/naver/fire' target='_blank'>Original Github Repo</a></p>"

iface = gr.Interface(
    fn=generate_matching_superfeatures,
    inputs=[
#        gr.inputs.Image(shape=(1024, 1024), type="pil", label="First Image"),
#        gr.inputs.Image(shape=(1024, 1024), type="pil", label="Second Image"),
        gr.inputs.Image(type="pil", label="First Image"),
        gr.inputs.Image(type="pil", label="Second Image"),
        gr.inputs.Checkbox(default=False, label="Model trained on ImageNet (Default: SfM-120k)"),
        gr.inputs.Slider(minimum=0, maximum=6, step=1, default=2, label="Scale"),
        gr.inputs.Slider(minimum=0, maximum=255, step=25, default=150, label="Binarization Threshold"),
        gr.inputs.Checkbox(default=False, label="Show random (matching) SFs"),
        gr.inputs.Textbox(lines=1, default="", label="...or show specific SF IDs:", optional=True),
        # gr.inputs.Checkbox(default=True, label="Show only matching SFs"),
        ],
    outputs=[
        gr.outputs.Image(type="plot", label="First Image SFs"),
        gr.outputs.Image(type="plot", label="Second Image SFs"),
        gr.outputs.Image(type="plot", label="SF legend")],
        # gr.outputs.Textbox(label="SFs")],
    # outputs=gr.outputs.Image(shape=(1024,2048), type="plot"),
    title=title,
    theme='peach',
    layout="horizontal",
    description=description,
    article=article,
    examples=[
        ["chateau_1.png", "chateau_2.png", False, 3, 150, True, ''],
        ["anafi1.jpeg", "anafi2.jpeg", False, 4, 150, True, ''],
        ["areopoli1.jpeg", "areopoli2.jpeg", False, 4, 150, True, ''],
    ]
)
iface.launch(enable_queue=True)