├── LICENSE ├── README.md ├── SketchyDataset_README.md ├── check ├── __init__.py └── feature_vis.py ├── data ├── __init__.py ├── image_input.py └── triplet_input.py ├── feature └── feature_extract.py ├── models ├── TripletEmbedding.py ├── TripletLoss.py ├── __init__.py ├── sketch_resnet.py └── vgg.py ├── record ├── feature_vis │ ├── resnet50_0.png │ ├── resnet50_315.png │ ├── resnet50_64_265.png │ ├── resnet50_940.png │ ├── resnet_150.png │ ├── resnet_90.png │ ├── resnet_imagenet.png │ ├── resnet_separate.png │ ├── vgg16_190epoch.png │ ├── vgg16_5epoch.png │ ├── vgg16_995epoch.png │ └── vgg16_baseline.png └── retrieval_reslut │ └── result.png ├── retrieval_test.py ├── test_img.txt ├── test_sketch.txt ├── train.py └── utils ├── __init__.py ├── extractor.py ├── feature_visualizer.py ├── test.py └── visualize.py /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2019 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. 22 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # SketchyDatabase 2 | 3 | [![LICENSE](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE) ![](https://img.shields.io/badge/python-3.6.5-brightgreen.svg) ![](https://img.shields.io/badge/pytorch-0.4.1-brightgreen.svg) ![](https://img.shields.io/badge/visdom-0.1.8.5-brightgreen.svg) ![](https://img.shields.io/badge/tqdm-4.28.1-brightgreen.svg) 4 | 5 | This project is a repo of [**The Sketchy Database: Learning to Retrieve Badly Drawn Bunnies**](https://www.cc.gatech.edu/~hays/tmp/sketchy-database.pdf). 6 | 7 | The [**homepage**](http://sketchy.eye.gatech.edu/) of the original project. 8 | 9 | Get the dataset via Google Drive [sketchydataset](https://drive.google.com/file/d/0B7ISyeE8QtDdTjE1MG9Gcy1kSkE/view?usp=sharing&resourcekey=0-r6nB4crmdU-LK7H38xnOUw) [SketchyDataset Intro](https://github.com/CDOTAD/SketchyDatabase/blob/master/SketchyDataset_README.md) 10 | 11 | # DataSet 12 | 13 | Sketchy Database 14 | 15 | ### Test Set 16 | 17 | As I didn't notice that the Sketchy Database contained a list of the testing photos, I randomly chose the testing photos and their related sketches myself. The test data set are listed in [TEST_IMG](test_img.txt) and [TEST_SKETCH](test_sketch.txt) 18 | 19 | | category | photo | sketch | 20 | | :---: | :---: | :---: | 21 | | airplane | 10 | 75 | 22 | | alarm_clock | | 52 | 23 | | ant | | 53 | 24 | | . | | . | 25 | | . | | . | 26 | | . | | . | 27 | | window | | 54 | 28 | | wine_bottle | | 52 | 29 | | zebra | | 66 | 30 | | **Total** | 1250 | 7875 | 31 | 32 | ### The Dataset Structure in My Project 33 | 34 | ```Bash 35 | Dataset 36 | ├── photo-train # the training set of photos 37 | ├── sketch-triplet-train # the training set of sketches 38 | ├── photo-test # the testing set of photos 39 | ├── sketch-triplet-test # the testing set of sketches 40 | ``` 41 | # Test 42 | 43 | using [feature_extract.py](https://github.com/CDOTAD/SketchyDatabase/blob/master/feature/feature_extract.py) to get the extracted feature files ('\*.pkl') 44 | 45 | using [retrieval_test.py](https://github.com/CDOTAD/SketchyDatabase/blob/master/retrieval_test.py) to get the testing result. 46 | 47 | # Testing Result 48 | 49 | There is no GoogLeNet, which resulted the best in the original paper, implement in PyTorch, so I used vgg16 instead. 50 | 51 | | model | epoch | recall@1 | recall@5| 52 | | :---: | :---: | :---: | :---: | 53 | | resnet34(pretrained;mixed training set;metric='cosine') | | | | 54 | | | 90 | 8.51% | 18.68% | 55 | | | 150 | 9.31% | 20.44% | 56 | |resnet34(pretrained;mixed training set;metric='euclidean')| | | | 57 | | | 90 | 6.45% | 14.79% | 58 | | | 150 | 6.96% | 16.46% | 59 | |resnet34(150 epoch;triplet loss m=0.02;metric='euclidean';lr=1e-5 batch_size=16)| | | | 60 | | | 85 | 9.87% | 22.37% | 61 | |vgg16(pretrained;triplet loss m=0.3;metric='euclidean';lr=1e-5;batch_size=16) | | | | 62 | | | 0 | 0.17% | 0.72% | 63 | | | 5 | 17.59% | 45.51% | 64 | | | 190 | 31.03% | 67.86% | 65 | | | 275 | 32.22% | 68.48% | 66 | | | 975 | 35.24% | 71.53% | 67 | |vgg16(fine-tune(275epoch);m=0.15;metric='euclidean';lr=1e-7;batch_size=16) | | | | 68 | | | 55 | 33.22% | 70.04% | 69 | | | 625 | 35.78% | 72.44% | 70 | | | 995 | 36.09% | 73.02% | 71 | |resnet50(pretrained; triplet loss m=0.15; metric='euclidean'; lr=1e-7;batch_size=16)|||| 72 | | | 0 | 0.71% | 11.48% | 73 | | | 55 | 10.18% | 29.94% | 74 | | | 940 | 15.17% | 47.61% | 75 | |resnet50(pretrained; triplet loss m=0.1; metric='euclidean'; lr=1e-6 batch_size=32)||| | 76 | | | 315 | 19.58% | 57.19% | 77 | |resnet50(pretrained; triplet loss m=0.3; metric='euclidean'; lr=1e-5 batch_size=48)||| | 78 | | | 20 | 21.56% | 57.50% | 79 | | | 95 | 30.32% | 71.73% | 80 | | | 265 | 40.08% | 78.83% | 81 | | | 930 | 46.04% | 83.30% | 82 | 83 | I have no idea about why the resnet34 got that bad result, while the vgg16 and resnet50 resulted pretty well. 84 | 85 | ### Retrieval Result 86 | 87 | I randomly chose 20 sketches as the query sketch and here is the retrieval result. The model I used is the [resnet50](#resnet)(pretrained; triplet loss m=0.3; metric='euclidean'; lr=1e-5 batch_size=48) after 265 training epoch. 88 | 89 | ![retrieval_result](record/retrieval_reslut/result.png) 90 | 91 | # Feature Visulization via T-SNE 92 | 93 | all the visulizated categories are the first ten categories in alphabetical order. 94 | 95 | **The boxes represent the photos, while the points represent the sketches.** 96 | 97 | | model | vis | 98 | | :---: | :---: | 99 | | resnet34 pretrained on ImageNet |![](record/feature_vis/resnet_imagenet.png)| 100 | | **pretrained; sketch branch& photo branch are trained sparately** | 101 | | resnet34 |![](record/feature_vis/resnet_separate.png)| 102 | | **pretrained; mixed training set** | 103 | | resnet34 after 90 training epoch | ![](record/feature_vis/resnet_90.png) | 104 | | resnet34 after 150 training epoch | ![](record/feature_vis/resnet_150.png) | 105 | | **pretrained; triplet loss m=0.3 lr=1e-5** | 106 | | vgg16 after 0 training epoch | ![](record/feature_vis/vgg16_baseline.png)| 107 | | vgg16 after 5 training epoch | ![](record/feature_vis/vgg16_5epoch.png)| 108 | | vgg16 after 190 training epoch | ![](record/feature_vis/vgg16_190epoch.png)| 109 | | **fine tune; triplet loss m=0.15 lr=1e-7** | 110 | | vgg16(fine tune) after 995 training epoch | ![](record/feature_vis/vgg16_995epoch.png)| 111 | | **pretrained; triplet loss m=0.15 lr=1e-7** | 112 | | resnet50 after 0 training epoch | ![](record/feature_vis/resnet50_0.png) | 113 | | resnet50 after 940 training epoch | ![](record/feature_vis/resnet50_940.png) | 114 | | **pretrained; triplet loss m=0.1 lr=1e-6** | 115 | | resnet50 after 315 training epoch | ![](record/feature_vis/resnet50_315.png) | 116 | | **pretrained; triplet loss m=0.3 lr=1e-5 batch_size=48** | 117 | | resnet50 after 265 training epoch |![](record/feature_vis/resnet50_64_265.png)| 118 | -------------------------------------------------------------------------------- /SketchyDataset_README.md: -------------------------------------------------------------------------------- 1 | Sketchy Database - Rendered sketches and augmented photos 2 | 3 | Contents: 4 | photo - a directory containing two different renderings of 5 | all photographs contained within the Sketchy Database 6 | sketch - a directory containing six different renderings of 7 | all sketches contained within the Sketchy Database 8 | 9 | Photographs 10 | All photographs are rendered in JPEG format. Resizing is 11 | performed via OpenCV Imgproc (typically area interpolation 12 | for full images renderings and cubic for bounding box 13 | renderings). 14 | 15 | Augmentations (directories within 'photo') 16 | 17 | tx_000000000000 : image is non-uniformly scaled to 256x256 18 | tx_000100000000 : image bounding box scaled to 256x256 with 19 | an additional +10% on each edge; note 20 | that due to position within the image, 21 | sometimes the object is not centered 22 | 23 | Sketches 24 | All sketches are rendered in PNG format. The original 25 | sketch canvas size is 640x480. In rendering the sketch to a 26 | 256x256 canvas, we take the original photo aspect ratio 27 | as well as the original sketch canvas aspect ratio into 28 | account. We render sketches such that they are consistent 29 | with the transformation made to the image (non-uniform 30 | scale to 256x256). In order to ensure sketches remain fully 31 | on the canvas, some minor adjustments to scale and/or 32 | location are occasionally necessary. 33 | All sketches are rendered using custom OpenGL code, with 34 | a PNG encoding provided by Java's ImageIO API. 35 | 36 | Augmentations (directories within 'sketch') 37 | 38 | tx_000000000000 : sketch canvas is rendered to 256x256 39 | such that it undergoes the same 40 | scaling as the paired photo 41 | tx_000100000000 : sketch is centered and uniformly scaled 42 | such that its greatest dimension (x or y) 43 | fills 78% of the canvas (roughly the same 44 | as in Eitz 2012 sketch data set) 45 | tx_000000000010 : sketch is translated such that it is 46 | centered on the object bounding box 47 | tx_000000000110 : sketch is centered on bounding box and 48 | is uniformly scaled such that one dimension 49 | (x or y; whichever requires the least amount 50 | of scaling) fits within the bounding box 51 | tx_000000001010 : sketch is centered on bounding box and 52 | is uniformly scaled such that one dimension 53 | (x or y; whichever requires the most amount 54 | of scaling) fits within the bounding box 55 | tx_000000001110 : sketch is centered on bounding box and 56 | is non-uniformly scaled such that it 57 | completely fits within the bounding box 58 | 59 | 60 | -------------------------------------------------------------------------------- /check/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/check/__init__.py -------------------------------------------------------------------------------- /check/feature_vis.py: -------------------------------------------------------------------------------- 1 | from utils.feature_visualizer import FeatureVisualizer 2 | 3 | test_set_root = '/data1/zzl/dataset/sketch-triplet-test' 4 | test_photo_root = '/data1/zzl/dataset/photo-test' 5 | vis = FeatureVisualizer(data_root=test_set_root, feature_root='../feature/sketch-vgg-190epoch.pkl', 6 | env='caffe2torch_featurevis') 7 | # vis.visualize(10, win='training together sketch') 8 | 9 | vis.embedding_vis(em_data_root=test_photo_root, em_feature_root='../feature/photo-vgg-190epoch.pkl', 10 | min_class=0, max_class=10, win='embedding') 11 | -------------------------------------------------------------------------------- /data/__init__.py: -------------------------------------------------------------------------------- 1 | from data.triplet_input import TripleDataset 2 | from data.image_input import ImageDataset 3 | import torch.utils.data 4 | 5 | 6 | class TripleDataLoader(object): 7 | def __init__(self, opt): 8 | self.dataset = TripleDataset(opt.photo_root, opt.sketch_root) 9 | self.dataloader = torch.utils.data.DataLoader( 10 | self.dataset, 11 | shuffle=True, 12 | batch_size=opt.batch_size, 13 | num_workers=4, 14 | drop_last=True 15 | ) 16 | 17 | def load_data(self): 18 | return self 19 | 20 | def __len__(self): 21 | return len(self.dataset) 22 | 23 | def __iter__(self): 24 | for i, data in enumerate(self.dataloader): 25 | yield data 26 | 27 | 28 | class ImageDataLoader(object): 29 | def __init__(self, opt): 30 | self.dataset = ImageDataset(opt.image_root) 31 | self.dataloader = torch.utils.data.DataLoader( 32 | self.dataset, 33 | shuffle=False, 34 | batch_size=opt.batch_size, 35 | num_workers=4, 36 | drop_last=False 37 | ) 38 | 39 | def load_data(self): 40 | return self 41 | 42 | def __len__(self): 43 | return len(self.dataset) 44 | 45 | def __iter__(self): 46 | for i, data in enumerate(self.dataloader): 47 | yield data -------------------------------------------------------------------------------- /data/image_input.py: -------------------------------------------------------------------------------- 1 | import os 2 | import torch.utils.data as data 3 | import torchvision.transforms as transforms 4 | from PIL import Image 5 | 6 | 7 | def make_dataset(root): 8 | images = [] 9 | 10 | cnames = os.listdir(root) 11 | for cname in cnames: 12 | c_path = os.path.join(root, cname) 13 | if os.path.isdir(c_path): 14 | fnames = os.listdir(c_path) 15 | for fname in fnames: 16 | path = os.path.join(c_path, fname) 17 | images.append(path) 18 | 19 | return images 20 | 21 | 22 | class ImageDataset(data.Dataset): 23 | def __init__(self, image_root): 24 | self.tranform = transforms.Compose([ 25 | transforms.Resize(224), 26 | transforms.CenterCrop(224), 27 | transforms.ToTensor(), 28 | transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) 29 | ]) 30 | 31 | self.image_root = image_root 32 | self.image_paths = sorted(make_dataset(self.image_root)) 33 | 34 | self.len = len(self.image_paths) 35 | 36 | def __getitem__(self, index): 37 | 38 | image_path = self.image_paths[index] 39 | 40 | image = Image.open(image_path).convert('RGB') 41 | image = self.tranform(image) 42 | 43 | image_path = image_path.split('/') 44 | cname = image_path[-2] 45 | fname = image_path[-1] 46 | 47 | name = cname + '/' + fname 48 | 49 | return {'I': image, 'N': name} 50 | 51 | def __len__(self): 52 | return self.len 53 | 54 | -------------------------------------------------------------------------------- /data/triplet_input.py: -------------------------------------------------------------------------------- 1 | import os 2 | import torch.utils.data as data 3 | import torchvision.transforms as transforms 4 | import numpy as np 5 | from PIL import Image 6 | 7 | 8 | def find_classes(root): 9 | classes = [d for d in os.listdir(root) if os.path.isdir(os.path.join(root, d))] 10 | classes.sort() 11 | class_to_idex = {classes[i]: i for i in range(len(classes))} 12 | return classes, class_to_idex 13 | 14 | 15 | def make_dataset(root): 16 | images = [] 17 | 18 | cnames = os.listdir(root) 19 | for cname in cnames: 20 | c_path = os.path.join(root, cname) 21 | if os.path.isdir(c_path): 22 | fnames = os.listdir(c_path) 23 | for fname in fnames: 24 | path = os.path.join(c_path, fname) 25 | images.append(path) 26 | 27 | return images 28 | 29 | 30 | class TripleDataset(data.Dataset): 31 | def __init__(self, photo_root, sketch_root): 32 | super(TripleDataset, self).__init__() 33 | self.tranform = transforms.Compose([ 34 | transforms.Resize(224), 35 | transforms.CenterCrop(224), 36 | transforms.ToTensor(), 37 | transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) 38 | ]) 39 | classes, class_to_idx = find_classes(photo_root) 40 | 41 | self.photo_root = photo_root 42 | self.sketch_root = sketch_root 43 | 44 | self.photo_paths = sorted(make_dataset(self.photo_root)) 45 | self.classes = classes 46 | self.class_to_idx = class_to_idx 47 | 48 | self.len = len(self.photo_paths) 49 | 50 | def __getitem__(self, index): 51 | 52 | photo_path = self.photo_paths[index] 53 | sketch_path, label = self._getrelate_sketch(photo_path) 54 | 55 | photo = Image.open(photo_path).convert('RGB') 56 | sketch = Image.open(sketch_path).convert('RGB') 57 | 58 | P = self.tranform(photo) 59 | S = self.tranform(sketch) 60 | L = label 61 | return {'P': P, 'S': S, 'L': L} 62 | 63 | def __len__(self): 64 | return self.len 65 | 66 | def _getrelate_sketch(self, photo_path): 67 | 68 | paths = photo_path.split('/') 69 | fname = paths[-1].split('.')[0] 70 | cname = paths[-2] 71 | 72 | label = self.class_to_idx[cname] 73 | 74 | sketchs = sorted(os.listdir(os.path.join(self.sketch_root, cname))) 75 | 76 | sketch_rel = [] 77 | for sketch_name in sketchs: 78 | if sketch_name.split('-')[0] == fname: 79 | sketch_rel.append(sketch_name) 80 | 81 | rnd = np.random.randint(0, len(sketch_rel)) 82 | 83 | sketch = sketch_rel[rnd] 84 | 85 | return os.path.join(self.sketch_root, cname, sketch), label 86 | -------------------------------------------------------------------------------- /feature/feature_extract.py: -------------------------------------------------------------------------------- 1 | from utils.extractor import Extractor 2 | from models.vgg import vgg16 3 | from models.sketch_resnet import resnet50 4 | import torch as t 5 | from torch import nn 6 | import os 7 | 8 | # The script to extract sketches or photos' features using the trained model 9 | 10 | os.environ['CUDA_VISIBLE_DEVICES'] = '1' 11 | 12 | train_set_root = '/data1/zzl/dataset/sketch-triplet-train' 13 | test_set_root = '/data1/zzl/dataset/sketch-triplet-test' 14 | 15 | train_photo_root = '/data1/zzl/dataset/photo-train' 16 | test_photo_root = '/data1/zzl/dataset/photo-test' 17 | 18 | # The trained model root for resnet 19 | SKETCH_RESNET = '/data1/zzl/model/caffe2torch/mixed_triplet_loss/sketch/sketch_resnet_85.pth' 20 | PHOTO_RESNET = '/data1/zzl/model/caffe2torch/mixed_triplet_loss/photo/photo_resnet_85.pth' 21 | 22 | # The trained model root for vgg 23 | SKETCH_VGG = '/data1/zzl/model/caffe2torch/vgg_triplet_loss/sketch/sketch_vgg_190.pth' 24 | PHOTO_VGG = '/data1/zzl/model/caffe2torch/vgg_triplet_loss/photo/photo_vgg_190.pth' 25 | 26 | FINE_TUNE_RESNET = '/data1/zzl/model/caffe2torch/fine_tune/model_270.pth' 27 | 28 | device = 'cuda:1' 29 | 30 | '''vgg''' 31 | vgg = vgg16(pretrained=False) 32 | vgg.classifier[6] = nn.Linear(in_features=4096, out_features=125, bias=True) 33 | vgg.load_state_dict(t.load(PHOTO_VGG, map_location=t.device('cpu'))) 34 | vgg.cuda() 35 | 36 | ext = Extractor(pretrained=False) 37 | ext.reload_model(vgg) 38 | 39 | photo_feature = ext.extract_with_dataloader(test_photo_root, 'photo-vgg-190epoch.pkl') 40 | 41 | vgg.load_state_dict(t.load(SKETCH_VGG, map_location=t.device('cpu'))) 42 | ext.reload_model(vgg) 43 | 44 | sketch_feature = ext.extract_with_dataloader(test_set_root, 'sketch-vgg-190epoch.pkl') 45 | 46 | 47 | '''resnet''' 48 | resnet = resnet50() 49 | resnet.fc = nn.Linear(in_features=2048, out_features=125) 50 | resnet.load_state_dict(t.load(PHOTO_RESNET, map_location=t.device('cpu'))) 51 | resnet.cuda() 52 | 53 | ext = Extractor(pretrained=False) 54 | ext.reload_model(resnet) 55 | 56 | photo_feature = ext.extract_with_dataloader(test_photo_root, 'photo-resnet-epoch.pkl') 57 | 58 | resnet.load_state_dict(t.load(SKETCH_RESNET, map_location=t.device('cpu'))) 59 | ext.reload_model(resnet) 60 | 61 | sketch_feature = ext.extract_with_dataloader(test_set_root, 'sketch-resnet-epoch.pkl') 62 | 63 | -------------------------------------------------------------------------------- /models/TripletEmbedding.py: -------------------------------------------------------------------------------- 1 | import torch as t 2 | from torch import nn 3 | from data import TripleDataLoader 4 | from models.vgg import vgg16 5 | from models.sketch_resnet import resnet34, resnet50 6 | from utils.visualize import Visualizer 7 | from torchnet.meter import AverageValueMeter 8 | # import tqdm 9 | from utils.extractor import Extractor 10 | from sklearn.neighbors import NearestNeighbors 11 | from torch.nn import DataParallel 12 | from .TripletLoss import TripletLoss 13 | from utils.test import Tester 14 | import os 15 | import numpy as np 16 | 17 | 18 | class Config(object): 19 | def __init__(self): 20 | return 21 | 22 | 23 | class TripletNet(object): 24 | 25 | def __init__(self, opt): 26 | # train config 27 | self.photo_root = opt.photo_root 28 | self.sketch_root = opt.sketch_root 29 | self.batch_size = opt.batch_size 30 | # self.device = opt.device 31 | self.epochs = opt.epochs 32 | self.lr = opt.lr 33 | 34 | # testing config 35 | self.photo_test = opt.photo_test 36 | self.sketch_test = opt.sketch_test 37 | self.test = opt.test 38 | self.test_f = opt.test_f 39 | 40 | self.save_model = opt.save_model 41 | self.save_dir = opt.save_dir 42 | 43 | # vis 44 | self.vis = opt.vis 45 | self.env = opt.env 46 | 47 | # fine_tune 48 | self.fine_tune = opt.fine_tune 49 | self.model_root = opt.model_root 50 | 51 | # dataloader config 52 | data_opt = Config() 53 | data_opt.photo_root = opt.photo_root 54 | data_opt.sketch_root = opt.sketch_root 55 | data_opt.batch_size = opt.batch_size 56 | 57 | self.dataloader_opt = data_opt 58 | 59 | # triplet config 60 | self.margin = opt.margin 61 | self.p = opt.p 62 | 63 | # feature extractor net 64 | self.net = opt.net 65 | self.cat = opt.cat 66 | 67 | def _get_vgg16(self, pretrained=True): 68 | model = vgg16(pretrained=pretrained) 69 | model.classifier[6] = nn.Linear(in_features=4096, out_features=125, bias=True) 70 | 71 | return model 72 | 73 | def _get_resnet34(self, pretrained=True): 74 | model = resnet34(pretrained=pretrained) 75 | model.fc = nn.Linear(in_features=512, out_features=125) 76 | 77 | return model 78 | 79 | def _get_resnet50(self, pretrained=True): 80 | model = resnet50(pretrained=pretrained) 81 | model.fc = nn.Linear(in_features=2048, out_features=125) 82 | 83 | return model 84 | 85 | def train(self): 86 | 87 | if self.net == 'vgg16': 88 | photo_net = DataParallel(self._get_vgg16()).cuda() 89 | sketch_net = DataParallel(self._get_vgg16()).cuda() 90 | elif self.net == 'resnet34': 91 | photo_net = DataParallel(self._get_resnet34()).cuda() 92 | sketch_net = DataParallel(self._get_resnet34()).cuda() 93 | elif self.net == 'resnet50': 94 | photo_net = DataParallel(self._get_resnet50()).cuda() 95 | sketch_net = DataParallel(self._get_resnet50()).cuda() 96 | 97 | if self.fine_tune: 98 | photo_net_root = self.model_root 99 | sketch_net_root = self.model_root.replace('photo', 'sketch') 100 | 101 | photo_net.load_state_dict(t.load(photo_net_root, map_location=t.device('cpu'))) 102 | sketch_net.load_state_dict(t.load(sketch_net_root, map_location=t.device('cpu'))) 103 | 104 | print('net') 105 | print(photo_net) 106 | 107 | # triplet_loss = nn.TripletMarginLoss(margin=self.margin, p=self.p).cuda() 108 | photo_cat_loss = nn.CrossEntropyLoss().cuda() 109 | sketch_cat_loss = nn.CrossEntropyLoss().cuda() 110 | 111 | my_triplet_loss = TripletLoss().cuda() 112 | 113 | # optimizer 114 | photo_optimizer = t.optim.Adam(photo_net.parameters(), lr=self.lr) 115 | sketch_optimizer = t.optim.Adam(sketch_net.parameters(), lr=self.lr) 116 | 117 | if self.vis: 118 | vis = Visualizer(self.env) 119 | 120 | triplet_loss_meter = AverageValueMeter() 121 | sketch_cat_loss_meter = AverageValueMeter() 122 | photo_cat_loss_meter = AverageValueMeter() 123 | 124 | data_loader = TripleDataLoader(self.dataloader_opt) 125 | dataset = data_loader.load_data() 126 | 127 | for epoch in range(self.epochs): 128 | 129 | print('---------------{0}---------------'.format(epoch)) 130 | 131 | if self.test and epoch % self.test_f == 0: 132 | 133 | tester_config = Config() 134 | tester_config.test_bs = 128 135 | tester_config.photo_net = photo_net 136 | tester_config.sketch_net = sketch_net 137 | 138 | tester_config.photo_test = self.photo_test 139 | tester_config.sketch_test = self.sketch_test 140 | 141 | tester = Tester(tester_config) 142 | test_result = tester.test_instance_recall() 143 | 144 | result_key = list(test_result.keys()) 145 | vis.plot('recall', np.array([test_result[result_key[0]], test_result[result_key[1]]]), 146 | legend=[result_key[0], result_key[1]]) 147 | if self.save_model: 148 | t.save(photo_net.state_dict(), self.save_dir + '/photo' + '/photo_' + self.net + '_%s.pth' % epoch) 149 | t.save(sketch_net.state_dict(), self.save_dir + '/sketch' + '/sketch_' + self.net + '_%s.pth' % epoch) 150 | 151 | photo_net.train() 152 | sketch_net.train() 153 | 154 | for ii, data in enumerate(dataset): 155 | 156 | photo_optimizer.zero_grad() 157 | sketch_optimizer.zero_grad() 158 | 159 | photo = data['P'].cuda() 160 | sketch = data['S'].cuda() 161 | label = data['L'].cuda() 162 | 163 | p_cat, p_feature = photo_net(photo) 164 | s_cat, s_feature = sketch_net(sketch) 165 | 166 | # category loss 167 | p_cat_loss = photo_cat_loss(p_cat, label) 168 | s_cat_loss = sketch_cat_loss(s_cat, label) 169 | 170 | photo_cat_loss_meter.add(p_cat_loss.item()) 171 | sketch_cat_loss_meter.add(s_cat_loss.item()) 172 | 173 | # triplet loss 174 | loss = p_cat_loss + s_cat_loss 175 | 176 | # tri_record = 0. 177 | ''' 178 | for i in range(self.batch_size): 179 | # negative 180 | negative_feature = t.cat([p_feature[0:i, :], p_feature[i + 1:, :]], dim=0) 181 | # print('negative_feature.size :', negative_feature.size()) 182 | # photo_feature 183 | anchor_feature = s_feature[i, :] 184 | anchor_feature = anchor_feature.expand_as(negative_feature) 185 | # print('anchor_feature.size :', anchor_feature.size()) 186 | 187 | # positive 188 | positive_feature = p_feature[i, :] 189 | positive_feature = positive_feature.expand_as(negative_feature) 190 | # print('positive_feature.size :', positive_feature.size()) 191 | 192 | tri_loss = triplet_loss(anchor_feature, positive_feature, negative_feature) 193 | 194 | tri_record = tri_record + tri_loss 195 | 196 | # print('tri_loss :', tri_loss) 197 | loss = loss + tri_loss 198 | ''' 199 | # print('tri_record : ', tri_record) 200 | 201 | my_tri_loss = my_triplet_loss(s_feature, p_feature) / (self.batch_size - 1) 202 | triplet_loss_meter.add(my_tri_loss.item()) 203 | # print('my_tri_loss : ', my_tri_loss) 204 | 205 | # print(tri_record - my_tri_loss) 206 | loss = loss + my_tri_loss 207 | # print('loss :', loss) 208 | # loss = loss / opt.batch_size 209 | 210 | loss.backward() 211 | 212 | photo_optimizer.step() 213 | sketch_optimizer.step() 214 | 215 | if self.vis: 216 | vis.plot('triplet_loss', np.array([triplet_loss_meter.value()[0], photo_cat_loss_meter.value()[0], 217 | sketch_cat_loss_meter.value()[0]]), 218 | legend=['triplet_loss', 'photo_cat_loss', 'sketch_cat_loss']) 219 | 220 | triplet_loss_meter.reset() 221 | photo_cat_loss_meter.reset() 222 | sketch_cat_loss_meter.reset() 223 | 224 | 225 | 226 | 227 | 228 | 229 | 230 | 231 | 232 | -------------------------------------------------------------------------------- /models/TripletLoss.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | 6 | class TripletLoss(nn.Module): 7 | def __init__(self, p=2, m=0.3): 8 | super(TripletLoss, self).__init__() 9 | 10 | self.p = p 11 | self.m = m 12 | 13 | def forward(self, sketch, photo): 14 | 15 | bs = sketch.size(0) 16 | 17 | pos_distance = sketch - photo 18 | pos_distance = torch.pow(pos_distance, 2) 19 | pos_distance = torch.sqrt(torch.sum(pos_distance, dim=1)) 20 | 21 | sketch_self = sketch.unsqueeze(0) 22 | photo_T = photo.unsqueeze(1) 23 | 24 | negative_distance = sketch_self - photo_T 25 | negative_distance = torch.pow(negative_distance, 2) 26 | negative_distance = torch.sqrt(torch.sum(negative_distance, dim=2)) 27 | 28 | triplet_loss = pos_distance - negative_distance # bs x bs x num_vec 29 | # print('TripletLoss.forward.triplet_loss', triplet_loss) 30 | triplet_loss = triplet_loss + self.m 31 | eye = torch.eye(bs).cuda() 32 | triplet_loss = triplet_loss * (1 - eye) 33 | triplet_loss = F.relu(triplet_loss) 34 | 35 | triplet_loss = torch.sum(triplet_loss, dim=1) 36 | 37 | return torch.sum(triplet_loss) 38 | -------------------------------------------------------------------------------- /models/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/models/__init__.py -------------------------------------------------------------------------------- /models/sketch_resnet.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import math 3 | import torch.utils.model_zoo as model_zoo 4 | 5 | __all__ = ['ResNet', 'resnet50', 'resnet34'] 6 | 7 | model_urls = { 8 | 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', 9 | 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', 10 | 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 11 | 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', 12 | 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', 13 | } 14 | 15 | 16 | def conv3x3(in_planes, out_planes, stride=1): 17 | """3x3 convolution with padding""" 18 | return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, 19 | padding=1, bias=False) 20 | 21 | 22 | class BasicBlock(nn.Module): 23 | expansion = 1 24 | 25 | def __init__(self, inplanes, planes, stride=1, downsample=None): 26 | super(BasicBlock, self).__init__() 27 | self.conv1 = conv3x3(inplanes, planes, stride) 28 | self.bn1 = nn.BatchNorm2d(planes) 29 | self.relu = nn.ReLU(inplace=True) 30 | self.conv2 = conv3x3(planes, planes) 31 | self.bn2 = nn.BatchNorm2d(planes) 32 | self.downsample = downsample 33 | self.stride = stride 34 | 35 | def forward(self, x): 36 | residual = x 37 | 38 | out = self.conv1(x) 39 | out = self.bn1(out) 40 | out = self.relu(out) 41 | 42 | out = self.conv2(out) 43 | out = self.bn2(out) 44 | 45 | if self.downsample is not None: 46 | residual = self.downsample(x) 47 | 48 | out += residual 49 | out = self.relu(out) 50 | 51 | return out 52 | 53 | 54 | class Bottleneck(nn.Module): 55 | expansion = 4 56 | 57 | def __init__(self, inplanes, planes, stride=1, downsample=None): 58 | super(Bottleneck, self).__init__() 59 | self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) 60 | self.bn1 = nn.BatchNorm2d(planes) 61 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, 62 | padding=1, bias=False) 63 | self.bn2 = nn.BatchNorm2d(planes) 64 | self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False) 65 | self.bn3 = nn.BatchNorm2d(planes * 4) 66 | self.relu = nn.ReLU(inplace=True) 67 | self.downsample = downsample 68 | self.stride = stride 69 | 70 | def forward(self, x): 71 | residual = x 72 | 73 | out = self.conv1(x) 74 | out = self.bn1(out) 75 | out = self.relu(out) 76 | 77 | out = self.conv2(out) 78 | out = self.bn2(out) 79 | out = self.relu(out) 80 | 81 | out = self.conv3(out) 82 | out = self.bn3(out) 83 | 84 | if self.downsample is not None: 85 | residual = self.downsample(x) 86 | 87 | out += residual 88 | out = self.relu(out) 89 | 90 | return out 91 | 92 | 93 | 94 | 95 | class ResNet(nn.Module): 96 | 97 | def __init__(self, block, layers, num_classes=1000): 98 | self.inplanes = 64 99 | super(ResNet, self).__init__() 100 | self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, 101 | bias=False) 102 | self.bn1 = nn.BatchNorm2d(64) 103 | self.relu = nn.ReLU(inplace=True) 104 | self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) 105 | self.layer1 = self._make_layer(block, 64, layers[0]) 106 | self.layer2 = self._make_layer(block, 128, layers[1], stride=2) 107 | self.layer3 = self._make_layer(block, 256, layers[2], stride=2) 108 | self.layer4 = self._make_layer(block, 512, layers[3], stride=2) 109 | self.avgpool = nn.AdaptiveAvgPool2d(1) 110 | 111 | self.fc = nn.Linear(512 * block.expansion, num_classes) 112 | 113 | for m in self.modules(): 114 | if isinstance(m, nn.Conv2d): 115 | n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels 116 | m.weight.data.normal_(0, math.sqrt(2. / n)) 117 | elif isinstance(m, nn.BatchNorm2d): 118 | m.weight.data.fill_(1) 119 | m.bias.data.zero_() 120 | 121 | def _make_layer(self, block, planes, blocks, stride=1): 122 | downsample = None 123 | if stride != 1 or self.inplanes != planes * block.expansion: 124 | downsample = nn.Sequential( 125 | nn.Conv2d(self.inplanes, planes * block.expansion, 126 | kernel_size=1, stride=stride, bias=False), 127 | nn.BatchNorm2d(planes * block.expansion), 128 | ) 129 | 130 | layers = [] 131 | layers.append(block(self.inplanes, planes, stride, downsample)) 132 | self.inplanes = planes * block.expansion 133 | for i in range(1, blocks): 134 | layers.append(block(self.inplanes, planes)) 135 | 136 | return nn.Sequential(*layers) 137 | 138 | def forward(self, x): 139 | x = self.conv1(x) 140 | x = self.bn1(x) 141 | x = self.relu(x) 142 | x = self.maxpool(x) 143 | 144 | x = self.layer1(x) 145 | x = self.layer2(x) 146 | x = self.layer3(x) 147 | x = self.layer4(x) 148 | 149 | x = self.avgpool(x) 150 | x = x.view(x.size(0), -1) 151 | 152 | feature = x 153 | x = self.fc(x) 154 | 155 | return x, feature 156 | 157 | 158 | def resnet34(pretrained=False, **kwargs): 159 | """Constructs a ResNet-34 model. 160 | 161 | Args: 162 | pretrained (bool): If True, returns a model pre-trained on ImageNet 163 | """ 164 | model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs) 165 | if pretrained: 166 | model.load_state_dict(model_zoo.load_url(model_urls['resnet34'])) 167 | return model 168 | 169 | 170 | def resnet50(pretrained=False, **kwargs): 171 | """Constructs a ResNet-50 model. 172 | 173 | Args: 174 | pretrained (bool): If True, returns a model pre-trained on ImageNet 175 | """ 176 | model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs) 177 | if pretrained: 178 | model.load_state_dict(model_zoo.load_url(model_urls['resnet50']), strict=False) 179 | return model 180 | 181 | 182 | -------------------------------------------------------------------------------- /models/vgg.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch.utils.model_zoo as model_zoo 3 | import math 4 | 5 | __all__ = [ 6 | 'VGG', 'vgg16', 7 | ] 8 | 9 | 10 | model_urls = { 11 | 'vgg16': 'https://download.pytorch.org/models/vgg16-397923af.pth', 12 | } 13 | 14 | 15 | class VGG(nn.Module): 16 | 17 | def __init__(self, features, num_classes=1000, init_weights=True): 18 | super(VGG, self).__init__() 19 | self.features = features 20 | self.avgpool = nn.AdaptiveAvgPool2d(1) 21 | 22 | self.classifier = nn.Sequential( 23 | nn.Linear(512 * 7 * 7, 4096), 24 | nn.ReLU(True), 25 | nn.Dropout(), 26 | nn.Linear(4096, 4096), 27 | nn.ReLU(True), 28 | nn.Dropout(), 29 | nn.Linear(4096, num_classes), 30 | ) 31 | if init_weights: 32 | self._initialize_weights() 33 | 34 | def forward(self, x): 35 | x = self.features(x) 36 | # print('feature.size()', x.size()) 37 | feature = self.avgpool(x) 38 | x = x.view(x.size(0), -1) 39 | x = self.classifier(x) 40 | 41 | return x, feature 42 | 43 | def _initialize_weights(self): 44 | for m in self.modules(): 45 | if isinstance(m, nn.Conv2d): 46 | n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels 47 | m.weight.data.normal_(0, math.sqrt(2. / n)) 48 | if m.bias is not None: 49 | m.bias.data.zero_() 50 | elif isinstance(m, nn.BatchNorm2d): 51 | m.weight.data.fill_(1) 52 | m.bias.data.zero_() 53 | elif isinstance(m, nn.Linear): 54 | m.weight.data.normal_(0, 0.01) 55 | m.bias.data.zero_() 56 | 57 | 58 | def make_layers(cfg, batch_norm=False): 59 | layers = [] 60 | in_channels = 3 61 | for v in cfg: 62 | if v == 'M': 63 | layers += [nn.MaxPool2d(kernel_size=2, stride=2)] 64 | else: 65 | conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1) 66 | if batch_norm: 67 | layers += [conv2d, nn.BatchNorm2d(v), nn.ReLU(inplace=True)] 68 | else: 69 | layers += [conv2d, nn.ReLU(inplace=True)] 70 | in_channels = v 71 | return nn.Sequential(*layers) 72 | 73 | 74 | cfg = { 75 | 'A': [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'], 76 | 'B': [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'], 77 | 'D': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'], 78 | 'E': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512, 'M'], 79 | } 80 | 81 | 82 | def vgg16(pretrained=False, **kwargs): 83 | """VGG 16-layer model (configuration "D") 84 | 85 | Args: 86 | pretrained (bool): If True, returns a model pre-trained on ImageNet 87 | """ 88 | if pretrained: 89 | kwargs['init_weights'] = False 90 | model = VGG(make_layers(cfg['D']), **kwargs) 91 | if pretrained: 92 | model.load_state_dict(model_zoo.load_url(model_urls['vgg16'])) 93 | return model 94 | 95 | -------------------------------------------------------------------------------- /record/feature_vis/resnet50_0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/feature_vis/resnet50_0.png -------------------------------------------------------------------------------- /record/feature_vis/resnet50_315.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/feature_vis/resnet50_315.png -------------------------------------------------------------------------------- /record/feature_vis/resnet50_64_265.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/feature_vis/resnet50_64_265.png -------------------------------------------------------------------------------- /record/feature_vis/resnet50_940.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/feature_vis/resnet50_940.png -------------------------------------------------------------------------------- /record/feature_vis/resnet_150.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/feature_vis/resnet_150.png -------------------------------------------------------------------------------- /record/feature_vis/resnet_90.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/feature_vis/resnet_90.png -------------------------------------------------------------------------------- /record/feature_vis/resnet_imagenet.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/feature_vis/resnet_imagenet.png -------------------------------------------------------------------------------- /record/feature_vis/resnet_separate.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/feature_vis/resnet_separate.png -------------------------------------------------------------------------------- /record/feature_vis/vgg16_190epoch.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/feature_vis/vgg16_190epoch.png -------------------------------------------------------------------------------- /record/feature_vis/vgg16_5epoch.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/feature_vis/vgg16_5epoch.png -------------------------------------------------------------------------------- /record/feature_vis/vgg16_995epoch.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/feature_vis/vgg16_995epoch.png -------------------------------------------------------------------------------- /record/feature_vis/vgg16_baseline.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/feature_vis/vgg16_baseline.png -------------------------------------------------------------------------------- /record/retrieval_reslut/result.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/record/retrieval_reslut/result.png -------------------------------------------------------------------------------- /retrieval_test.py: -------------------------------------------------------------------------------- 1 | import pickle 2 | from sklearn.neighbors import NearestNeighbors 3 | import numpy as np 4 | from PIL import Image 5 | import os 6 | import torchvision.transforms as transforms 7 | from utils.visualize import Visualizer 8 | import tqdm 9 | 10 | PHOTO_ROOT = '/data1/zzl/dataset/photo-test' 11 | SKETCH_ROOT = '/data1/zzl/dataset/sketch-triplet-test' 12 | 13 | photo_data = pickle.load(open('feature/photo-resnet50_64-265.pkl', 'rb')) 14 | sketch_data = pickle.load(open('feature/sketch-resnet50_64-265.pkl', 'rb')) 15 | # print(photo_data['name'][0]) 16 | photo_feature = photo_data['feature'] 17 | photo_name = photo_data['name'] 18 | 19 | sketch_feature = sketch_data['feature'] 20 | sketch_name = sketch_data['name'] 21 | # print(np.size(photo_feature, 0)) 22 | 23 | nbrs = NearestNeighbors(n_neighbors=np.size(photo_feature, 0), 24 | algorithm='brute', metric='euclidean').fit(photo_feature) 25 | 26 | s_len = np.size(sketch_feature, 0) 27 | picked = [0] * s_len 28 | 29 | transform = transforms.Compose([ 30 | transforms.ToTensor() 31 | ]) 32 | 33 | vis = Visualizer('caffe2torch_test') 34 | count = 0 35 | for ii in range(20): 36 | index = np.random.randint(0, s_len) 37 | while picked[index]: 38 | index = np.random.randint(0, s_len) 39 | 40 | picked[index] = 1 41 | 42 | query_feature = sketch_feature[index] 43 | query_feature = np.reshape(query_feature, [1, np.shape(query_feature)[0]]) 44 | query_name = sketch_name[index] 45 | # print(query_name) 46 | distances, indices = nbrs.kneighbors(query_feature) 47 | 48 | query_split = query_name.split('/') 49 | query_class = query_split[0] 50 | query_img = query_split[1] 51 | 52 | query_image = np.array(Image.open(os.path.join(SKETCH_ROOT, query_name)).convert('RGB')) 53 | for i, indice in enumerate(indices[0][:5]): 54 | retrievaled_name = photo_name[indice] 55 | retrievaled_im = np.array(Image.open(os.path.join(PHOTO_ROOT, retrievaled_name)).convert('RGB')) 56 | query_image = np.append(query_image, retrievaled_im, axis=1) 57 | 58 | retrievaled_class = retrievaled_name.split('/')[0] 59 | retrievaled_name = retrievaled_name.split('/')[1] 60 | retrievaled_name = retrievaled_name.split('.')[0] 61 | 62 | if retrievaled_class == query_class: 63 | print(ii, 'correct class', query_name, retrievaled_name) 64 | if query_img.find(retrievaled_name) != -1: 65 | print(ii, 'correct item', query_name) 66 | count += 1 67 | 68 | if ii == 0: 69 | result = query_image 70 | else: 71 | result = np.append(result, query_image, axis=0) 72 | 73 | result = transform(result) 74 | vis.images(result.numpy(), win='result') 75 | 76 | print(count) 77 | 78 | 79 | count = 0 80 | count_5 = 0 81 | K = 5 82 | 83 | div = 0 84 | 85 | for ii, (query_sketch, query_name) in tqdm.tqdm(enumerate(zip(sketch_feature, sketch_name))): 86 | query_sketch = np.reshape(query_sketch, [1, np.shape(query_sketch)[0]]) 87 | 88 | query_split = query_name.split('/') 89 | query_class = query_split[0] 90 | query_img = query_split[1] 91 | 92 | distances, indices = nbrs.kneighbors(query_sketch) 93 | 94 | div += distances[0][1] - distances[0][0] 95 | 96 | # top K 97 | 98 | for i, indice in enumerate(indices[0][:K]): 99 | 100 | retrievaled_name = photo_name[indice] 101 | retrievaled_class = retrievaled_name.split('/')[0] 102 | 103 | retrievaled_name = retrievaled_name.split('/')[1] 104 | retrievaled_name = retrievaled_name.split('.')[0] 105 | 106 | if retrievaled_class == query_class: 107 | if query_img.find(retrievaled_name) != -1: 108 | if i == 0: 109 | count += 1 110 | count_5 += 1 111 | break 112 | recall = count / (ii+1) 113 | recall_5 = count_5 / (ii+1) 114 | print('recall@1 :', recall, ' recall@5 :', recall_5, 'div :', div/(ii+1)) 115 | 116 | -------------------------------------------------------------------------------- /test_img.txt: -------------------------------------------------------------------------------- 1 | airplane/n02691156_2701.jpg 2 | airplane/n02691156_8876.jpg 3 | airplane/n02691156_8352.jpg 4 | airplane/n02691156_7027.jpg 5 | airplane/n02691156_7472.jpg 6 | airplane/n02691156_5813.jpg 7 | airplane/n02691156_53586.jpg 8 | airplane/n02691156_4134.jpg 9 | airplane/n02691156_9491.jpg 10 | airplane/n02691156_3791.jpg 11 | alarm_clock/n02694662_7256.jpg 12 | alarm_clock/n02694662_14931.jpg 13 | alarm_clock/n02694662_7591.jpg 14 | alarm_clock/n02694662_1079.jpg 15 | alarm_clock/n02694662_14901.jpg 16 | alarm_clock/n02694662_12296.jpg 17 | alarm_clock/n02694662_17547.jpg 18 | alarm_clock/n02694662_11650.jpg 19 | alarm_clock/n02694662_9154.jpg 20 | alarm_clock/n02694662_4008.jpg 21 | ant/n02219486_677.jpg 22 | ant/n02219486_17589.jpg 23 | ant/n02219486_30822.jpg 24 | ant/n02219486_29662.jpg 25 | ant/n02219486_28415.jpg 26 | ant/n02219486_25482.jpg 27 | ant/n02219486_24244.jpg 28 | ant/n02219486_26134.jpg 29 | ant/n02219486_27825.jpg 30 | ant/n02219486_21712.jpg 31 | ape/n02483362_6484.jpg 32 | ape/n02470325_11452.jpg 33 | ape/n02480495_788.jpg 34 | ape/n02483362_5683.jpg 35 | ape/n02480495_162.jpg 36 | ape/n02483708_6167.jpg 37 | ape/n02480495_25570.jpg 38 | ape/n02481823_3822.jpg 39 | ape/n02481823_3882.jpg 40 | ape/n02481500_7639.jpg 41 | apple/n07739125_10746.jpg 42 | apple/n07739125_5971.jpg 43 | apple/n07739125_7934.jpg 44 | apple/n07739125_10944.jpg 45 | apple/n07739125_4074.jpg 46 | apple/n07739125_9823.jpg 47 | apple/n07739125_1406.jpg 48 | apple/n07739125_6509.jpg 49 | apple/n07739125_6351.jpg 50 | apple/n07739125_12353.jpg 51 | armor/n02895154_34811.jpg 52 | armor/n03146219_14890.jpg 53 | armor/n02895154_14298.jpg 54 | armor/n03000247_55528.jpg 55 | armor/n03000247_68189.jpg 56 | armor/n03000247_17925.jpg 57 | armor/n02895154_36195.jpg 58 | armor/n03146219_4724.jpg 59 | armor/n03000247_9005.jpg 60 | armor/n03000247_71108.jpg 61 | axe/n02764044_13358.jpg 62 | axe/n02764044_8541.jpg 63 | axe/n02764044_20350.jpg 64 | axe/n02764044_25921.jpg 65 | axe/n02764044_56605.jpg 66 | axe/n02764044_33304.jpg 67 | axe/n02764044_30843.jpg 68 | axe/n02764044_23413.jpg 69 | axe/n02764044_17192.jpg 70 | axe/n02764044_64984.jpg 71 | banana/n07753592_6075.jpg 72 | banana/n07753592_2049.jpg 73 | banana/n07753592_7792.jpg 74 | banana/n07753592_20750.jpg 75 | banana/n07753592_1416.jpg 76 | banana/n07753592_10233.jpg 77 | banana/n07753592_1193.jpg 78 | banana/n07753592_11641.jpg 79 | banana/n07753592_12072.jpg 80 | banana/n07753592_6756.jpg 81 | bat/n02139199_12462.jpg 82 | bat/n02139199_11487.jpg 83 | bat/n02139199_8857.jpg 84 | bat/n02147947_6236.jpg 85 | bat/n02149420_2799.jpg 86 | bat/n02147947_3903.jpg 87 | bat/n02147328_2008.jpg 88 | bat/n02147328_642.jpg 89 | bat/n02139199_16145.jpg 90 | bat/n02149420_3768.jpg 91 | bear/n02131653_7966.jpg 92 | bear/n02131653_12462.jpg 93 | bear/n02131653_1550.jpg 94 | bear/n02131653_2786.jpg 95 | bear/n02131653_6170.jpg 96 | bear/n02131653_3012.jpg 97 | bear/n02131653_5922.jpg 98 | bear/n02131653_7163.jpg 99 | bear/n02131653_6207.jpg 100 | bear/n02131653_5813.jpg 101 | bee/n02206856_3121.jpg 102 | bee/n02206856_5850.jpg 103 | bee/n02206856_4261.jpg 104 | bee/n02206856_6361.jpg 105 | bee/n02206856_6903.jpg 106 | bee/n02206856_1815.jpg 107 | bee/n02206856_6552.jpg 108 | bee/n02206856_8544.jpg 109 | bee/n02206856_1139.jpg 110 | bee/n02206856_2419.jpg 111 | beetle/n02169497_3918.jpg 112 | beetle/n02169497_5428.jpg 113 | beetle/n02169497_197.jpg 114 | beetle/n02167151_967.jpg 115 | beetle/n02169497_3809.jpg 116 | beetle/n02168699_7300.jpg 117 | beetle/n02176261_7522.jpg 118 | beetle/n02167151_7464.jpg 119 | beetle/n02169497_2275.jpg 120 | beetle/n02167151_9858.jpg 121 | bell/n02824448_19917.jpg 122 | bell/n02824448_3251.jpg 123 | bell/n02824448_1680.jpg 124 | bell/n02824448_13821.jpg 125 | bell/n03028596_4575.jpg 126 | bell/n02824448_13582.jpg 127 | bell/n02824448_30419.jpg 128 | bell/n02824448_11201.jpg 129 | bell/n03028596_2273.jpg 130 | bell/n02824448_15595.jpg 131 | bench/n03891251_15.jpg 132 | bench/n02828884_4015.jpg 133 | bench/n02828884_2200.jpg 134 | bench/n03891251_7764.jpg 135 | bench/n02828884_7329.jpg 136 | bench/n03891251_1444.jpg 137 | bench/n03891251_3664.jpg 138 | bench/n03891251_2480.jpg 139 | bench/n02828884_10535.jpg 140 | bench/n02828884_1200.jpg 141 | bicycle/n02834778_6942.jpg 142 | bicycle/n02834778_11645.jpg 143 | bicycle/n03792782_4600.jpg 144 | bicycle/n04126066_9549.jpg 145 | bicycle/n02834778_2873.jpg 146 | bicycle/n02834778_12279.jpg 147 | bicycle/n02834778_7767.jpg 148 | bicycle/n02834778_11636.jpg 149 | bicycle/n02834778_6186.jpg 150 | bicycle/n02834778_12174.jpg 151 | blimp/n02850950_12373.jpg 152 | blimp/n02850950_3246.jpg 153 | blimp/n02850950_22283.jpg 154 | blimp/n02850950_3765.jpg 155 | blimp/n02850950_10002.jpg 156 | blimp/n02850950_58.jpg 157 | blimp/n02850950_13395.jpg 158 | blimp/n02850950_1715.jpg 159 | blimp/n02850950_18396.jpg 160 | blimp/n02850950_4074.jpg 161 | bread/n07682316_7608.jpg 162 | bread/n07683786_25458.jpg 163 | bread/n07679356_27111.jpg 164 | bread/n07683786_6427.jpg 165 | bread/n07684084_2909.jpg 166 | bread/n07684084_1579.jpg 167 | bread/n07684084_2352.jpg 168 | bread/n07684084_368.jpg 169 | bread/n07683786_14450.jpg 170 | bread/n07682316_1438.jpg 171 | butterfly/n02274259_21905.jpg 172 | butterfly/n02274259_18681.jpg 173 | butterfly/n02274259_5479.jpg 174 | butterfly/n02274259_20043.jpg 175 | butterfly/n02274259_19471.jpg 176 | butterfly/n02274259_14994.jpg 177 | butterfly/n02274259_23833.jpg 178 | butterfly/n02274259_15828.jpg 179 | butterfly/n02274259_8298.jpg 180 | butterfly/n02274259_19616.jpg 181 | cabin/n02932400_7727.jpg 182 | cabin/n02932400_11404.jpg 183 | cabin/n02932400_4934.jpg 184 | cabin/n02932400_9565.jpg 185 | cabin/n02932400_5733.jpg 186 | cabin/n03686924_4151.jpg 187 | cabin/n03686924_2295.jpg 188 | cabin/n02932400_1945.jpg 189 | cabin/n02932400_1408.jpg 190 | cabin/n03686924_10083.jpg 191 | camel/n02437136_3994.jpg 192 | camel/n02437136_3674.jpg 193 | camel/n02437136_4845.jpg 194 | camel/n02437136_584.jpg 195 | camel/n02437136_3428.jpg 196 | camel/n02437136_4429.jpg 197 | camel/n02437136_8319.jpg 198 | camel/n02437136_14156.jpg 199 | camel/n02437136_257.jpg 200 | camel/n02437136_4752.jpg 201 | candle/n02948072_3987.jpg 202 | candle/n02948072_21856.jpg 203 | candle/n02948072_12086.jpg 204 | candle/n02948072_29247.jpg 205 | candle/n02948072_1666.jpg 206 | candle/n02948072_33885.jpg 207 | candle/n02948072_6033.jpg 208 | candle/n02948072_5309.jpg 209 | candle/n02948072_34562.jpg 210 | candle/n02948072_14341.jpg 211 | cannon/n02950826_10208.jpg 212 | cannon/n02950826_14185.jpg 213 | cannon/n02950826_1353.jpg 214 | cannon/n02950826_18518.jpg 215 | cannon/n02950826_5567.jpg 216 | cannon/n02950826_4158.jpg 217 | cannon/n02950826_17195.jpg 218 | cannon/n02950826_2757.jpg 219 | cannon/n02950826_4653.jpg 220 | cannon/n02950826_12133.jpg 221 | car_(sedan)/n04166281_5551.jpg 222 | car_(sedan)/n02958343_1853.jpg 223 | car_(sedan)/n02958343_520.jpg 224 | car_(sedan)/n04166281_1917.jpg 225 | car_(sedan)/n04166281_263.jpg 226 | car_(sedan)/n02958343_12252.jpg 227 | car_(sedan)/n02958343_4750.jpg 228 | car_(sedan)/n02958343_6586.jpg 229 | car_(sedan)/n04166281_185.jpg 230 | car_(sedan)/n04166281_4002.jpg 231 | castle/n02980441_8875.jpg 232 | castle/n02980441_22886.jpg 233 | castle/n02980441_14457.jpg 234 | castle/n02980441_11297.jpg 235 | castle/n02980441_11789.jpg 236 | castle/n02980441_8524.jpg 237 | castle/n02980441_1533.jpg 238 | castle/n02980441_14009.jpg 239 | castle/n02980441_11128.jpg 240 | castle/n02980441_6317.jpg 241 | cat/n02121620_10527.jpg 242 | cat/n02121620_4954.jpg 243 | cat/n02121620_51.jpg 244 | cat/n02121620_31930.jpg 245 | cat/n02121620_5192.jpg 246 | cat/n02121620_9283.jpg 247 | cat/n02121620_7791.jpg 248 | cat/n02121620_531.jpg 249 | cat/n02121620_20153.jpg 250 | cat/n02121620_13243.jpg 251 | chair/n03001627_3318.jpg 252 | chair/n03001627_6105.jpg 253 | chair/n02738535_904.jpg 254 | chair/n03001627_12243.jpg 255 | chair/n02738535_14114.jpg 256 | chair/n02738535_6285.jpg 257 | chair/n02738535_1312.jpg 258 | chair/n03001627_2245.jpg 259 | chair/n03001627_1776.jpg 260 | chair/n02738535_7205.jpg 261 | chicken/n01791625_8908.jpg 262 | chicken/n01791625_4689.jpg 263 | chicken/n01791625_15751.jpg 264 | chicken/n01791625_16214.jpg 265 | chicken/n01791625_4044.jpg 266 | chicken/n01791625_7354.jpg 267 | chicken/n01791625_1914.jpg 268 | chicken/n01791625_4567.jpg 269 | chicken/n01791625_17466.jpg 270 | chicken/n01791625_1081.jpg 271 | church/n03028079_9810.jpg 272 | church/n03028079_5432.jpg 273 | church/n03028079_2585.jpg 274 | church/n03028079_29412.jpg 275 | church/n03028079_9283.jpg 276 | church/n03028079_2718.jpg 277 | church/n03028079_13935.jpg 278 | church/n03028079_10579.jpg 279 | church/n03028079_11250.jpg 280 | church/n03028079_15095.jpg 281 | couch/n04256520_22692.jpg 282 | couch/n04256520_21394.jpg 283 | couch/n04256520_20048.jpg 284 | couch/n04256520_14377.jpg 285 | couch/n04256520_24708.jpg 286 | couch/n04256520_28871.jpg 287 | couch/n04256520_3251.jpg 288 | couch/n04256520_19834.jpg 289 | couch/n04256520_3986.jpg 290 | couch/n04256520_1299.jpg 291 | cow/n01887787_7620.jpg 292 | cow/n01887787_5589.jpg 293 | cow/n02403454_3113.jpg 294 | cow/n01887787_1.jpg 295 | cow/n02404432_9953.jpg 296 | cow/n01887787_7746.jpg 297 | cow/n01887787_6114.jpg 298 | cow/n01887787_591.jpg 299 | cow/n01887787_6520.jpg 300 | cow/n01887787_5137.jpg 301 | crab/n01980166_9179.jpg 302 | crab/n01978455_795.jpg 303 | crab/n01981276_14823.jpg 304 | crab/n01978455_88.jpg 305 | crab/n01978455_2200.jpg 306 | crab/n01978455_9813.jpg 307 | crab/n01980166_9451.jpg 308 | crab/n01978455_5679.jpg 309 | crab/n01978455_2536.jpg 310 | crab/n01978455_11617.jpg 311 | crocodilian/n01698640_1904.jpg 312 | crocodilian/n01697457_7336.jpg 313 | crocodilian/n01698640_2000.jpg 314 | crocodilian/n01698640_3055.jpg 315 | crocodilian/n01698640_4477.jpg 316 | crocodilian/n01698640_7483.jpg 317 | crocodilian/n01697457_8961.jpg 318 | crocodilian/n01697457_9993.jpg 319 | crocodilian/n01697457_2737.jpg 320 | crocodilian/n01698640_7977.jpg 321 | cup/n03147509_12636.jpg 322 | cup/n03147509_12211.jpg 323 | cup/n03063073_9082.jpg 324 | cup/n03063073_933.jpg 325 | cup/n03147509_21660.jpg 326 | cup/n03063073_3117.jpg 327 | cup/n03147509_13746.jpg 328 | cup/n03147509_20483.jpg 329 | cup/n03063073_3104.jpg 330 | cup/n03147509_20180.jpg 331 | deer/n02430045_2684.jpg 332 | deer/n02431628_9369.jpg 333 | deer/n02432983_12240.jpg 334 | deer/n02433925_15161.jpg 335 | deer/n02431628_3868.jpg 336 | deer/n02432983_8470.jpg 337 | deer/n02433925_22295.jpg 338 | deer/n02432983_7830.jpg 339 | deer/n02433925_12638.jpg 340 | deer/n02430045_5969.jpg 341 | dog/n02106662_13178.jpg 342 | dog/n02106662_18405.jpg 343 | dog/n02103406_7708.jpg 344 | dog/n02103406_5563.jpg 345 | dog/n02106662_15858.jpg 346 | dog/n02106662_2157.jpg 347 | dog/n02103406_2706.jpg 348 | dog/n02106662_8511.jpg 349 | dog/n02103406_2976.jpg 350 | dog/n02109525_13700.jpg 351 | dolphin/n02068974_6910.jpg 352 | dolphin/n02068974_3178.jpg 353 | dolphin/n02068974_613.jpg 354 | dolphin/n02068974_5747.jpg 355 | dolphin/n02068974_4003.jpg 356 | dolphin/n02068974_7259.jpg 357 | dolphin/n02068974_212.jpg 358 | dolphin/n02068974_4210.jpg 359 | dolphin/n02068974_4066.jpg 360 | dolphin/n02072040_156.jpg 361 | door/n03222176_11589.jpg 362 | door/n03222318_13609.jpg 363 | door/n03222176_13242.jpg 364 | door/n03222318_10058.jpg 365 | door/n03222318_10367.jpg 366 | door/n03222176_3296.jpg 367 | door/n03222318_13264.jpg 368 | door/n03222318_7012.jpg 369 | door/n03222176_16166.jpg 370 | door/n03226880_3614.jpg 371 | duck/n01846331_6199.jpg 372 | duck/n01846331_4246.jpg 373 | duck/n01846331_16000.jpg 374 | duck/n01846331_15616.jpg 375 | duck/n01846331_6072.jpg 376 | duck/n01846331_16486.jpg 377 | duck/n01846331_16375.jpg 378 | duck/n01846331_3530.jpg 379 | duck/n01846331_13988.jpg 380 | duck/n01846331_5951.jpg 381 | elephant/n02503517_8222.jpg 382 | elephant/n02503517_5527.jpg 383 | elephant/n02503517_8205.jpg 384 | elephant/n02503517_6880.jpg 385 | elephant/n02503517_3808.jpg 386 | elephant/n02503517_9827.jpg 387 | elephant/n02503517_10193.jpg 388 | elephant/n02503517_3799.jpg 389 | elephant/n02503517_4687.jpg 390 | elephant/n02503517_5117.jpg 391 | eyeglasses/n04272054_17512.jpg 392 | eyeglasses/n04356056_7859.jpg 393 | eyeglasses/n04356056_1773.jpg 394 | eyeglasses/n04356056_1989.jpg 395 | eyeglasses/n04356056_1223.jpg 396 | eyeglasses/n04356056_1937.jpg 397 | eyeglasses/n04272054_6470.jpg 398 | eyeglasses/n04272054_3524.jpg 399 | eyeglasses/n04356056_2972.jpg 400 | eyeglasses/n04272054_17451.jpg 401 | fan/n03271574_2512.jpg 402 | fan/n03271574_14023.jpg 403 | fan/n03271574_304.jpg 404 | fan/n03271574_9446.jpg 405 | fan/n03271574_8678.jpg 406 | fan/n03271574_6313.jpg 407 | fan/n03271574_12669.jpg 408 | fan/n03271574_11720.jpg 409 | fan/n03271574_1965.jpg 410 | fan/n03271574_4669.jpg 411 | fish/n02607072_3344.jpg 412 | fish/n02607072_3296.jpg 413 | fish/n02607072_909.jpg 414 | fish/n02534734_8449.jpg 415 | fish/n02605316_6101.jpg 416 | fish/n02605316_2350.jpg 417 | fish/n02513560_8353.jpg 418 | fish/n02643566_671.jpg 419 | fish/n02606052_1521.jpg 420 | fish/n02514041_5398.jpg 421 | flower/n11669921_22603.jpg 422 | flower/n11669921_22395.jpg 423 | flower/n11939491_3128.jpg 424 | flower/n11939491_19282.jpg 425 | flower/n11939491_8731.jpg 426 | flower/n11939491_55149.jpg 427 | flower/n11669921_21097.jpg 428 | flower/n11669921_23918.jpg 429 | flower/n11939491_44353.jpg 430 | flower/n11939491_35516.jpg 431 | frog/n01639765_19550.jpg 432 | frog/n01639765_14517.jpg 433 | frog/n01639765_20156.jpg 434 | frog/n01639765_7349.jpg 435 | frog/n01641577_13063.jpg 436 | frog/n01641577_8170.jpg 437 | frog/n01641391_19772.jpg 438 | frog/n01641577_218.jpg 439 | frog/n01639765_10652.jpg 440 | frog/n01639765_10465.jpg 441 | geyser/n09288635_11801.jpg 442 | geyser/n09288635_14337.jpg 443 | geyser/n09288635_14555.jpg 444 | geyser/n09288635_25596.jpg 445 | geyser/n09288635_13949.jpg 446 | geyser/n09288635_11393.jpg 447 | geyser/n09288635_11016.jpg 448 | geyser/n09288635_12147.jpg 449 | geyser/n09288635_10966.jpg 450 | geyser/n09288635_29220.jpg 451 | giraffe/n02439033_8875.jpg 452 | giraffe/n02439033_13602.jpg 453 | giraffe/n02439033_11666.jpg 454 | giraffe/n02439033_8846.jpg 455 | giraffe/n02439033_14962.jpg 456 | giraffe/n02439033_13384.jpg 457 | giraffe/n02439033_3944.jpg 458 | giraffe/n02439033_11894.jpg 459 | giraffe/n02439033_221.jpg 460 | giraffe/n02439033_9719.jpg 461 | guitar/n02676566_4605.jpg 462 | guitar/n02676566_8467.jpg 463 | guitar/n03272010_6662.jpg 464 | guitar/n03272010_12151.jpg 465 | guitar/n02676566_9062.jpg 466 | guitar/n02676566_307.jpg 467 | guitar/n03467517_22246.jpg 468 | guitar/n02676566_7385.jpg 469 | guitar/n02676566_6980.jpg 470 | guitar/n02676566_7830.jpg 471 | hamburger/n07697313_13156.jpg 472 | hamburger/n07697313_4367.jpg 473 | hamburger/n07697313_8515.jpg 474 | hamburger/n07697313_2730.jpg 475 | hamburger/n07697313_12432.jpg 476 | hamburger/n07697313_2172.jpg 477 | hamburger/n07697313_9789.jpg 478 | hamburger/n07697313_11264.jpg 479 | hamburger/n07697313_12711.jpg 480 | hamburger/n07697313_14416.jpg 481 | hammer/n03481172_16541.jpg 482 | hammer/n03481172_19147.jpg 483 | hammer/n03481172_4618.jpg 484 | hammer/n03481172_1565.jpg 485 | hammer/n03481172_11589.jpg 486 | hammer/n03481172_12416.jpg 487 | hammer/n03481172_22505.jpg 488 | hammer/n03481172_26915.jpg 489 | hammer/n03481172_15468.jpg 490 | hammer/n03481172_15211.jpg 491 | harp/n03495258_14373.jpg 492 | harp/n03495258_15784.jpg 493 | harp/n03495258_5207.jpg 494 | harp/n03495258_12596.jpg 495 | harp/n03495258_9888.jpg 496 | harp/n03495258_15623.jpg 497 | harp/n03495258_15739.jpg 498 | harp/n03495258_20155.jpg 499 | harp/n03495258_15256.jpg 500 | harp/n03495258_14000.jpg 501 | hat/n02859184_51537.jpg 502 | hat/n02954340_7738.jpg 503 | hat/n03124170_6952.jpg 504 | hat/n02954340_13808.jpg 505 | hat/n02859184_41293.jpg 506 | hat/n04259630_11895.jpg 507 | hat/n02859184_22326.jpg 508 | hat/n03497657_10828.jpg 509 | hat/n02859184_47904.jpg 510 | hat/n03124170_216.jpg 511 | hedgehog/n01872401_10227.jpg 512 | hedgehog/n01872401_12468.jpg 513 | hedgehog/n02346627_785.jpg 514 | hedgehog/n01894207_264.jpg 515 | hedgehog/n02346627_10071.jpg 516 | hedgehog/n01872401_15771.jpg 517 | hedgehog/n02346627_1757.jpg 518 | hedgehog/n02346627_15025.jpg 519 | hedgehog/n01872401_2948.jpg 520 | hedgehog/n01872401_9740.jpg 521 | helicopter/n03512147_1830.jpg 522 | helicopter/n03512147_2013.jpg 523 | helicopter/n03512147_224.jpg 524 | helicopter/n03512147_39770.jpg 525 | helicopter/n03512147_2302.jpg 526 | helicopter/n03512147_2935.jpg 527 | helicopter/n03512147_2669.jpg 528 | helicopter/n03512147_582.jpg 529 | helicopter/n03512147_6451.jpg 530 | helicopter/n03512147_1414.jpg 531 | hermit_crab/n01986214_5645.jpg 532 | hermit_crab/n01986214_10516.jpg 533 | hermit_crab/n01986214_6402.jpg 534 | hermit_crab/n01986214_9627.jpg 535 | hermit_crab/n01986214_4221.jpg 536 | hermit_crab/n01986214_17404.jpg 537 | hermit_crab/n01986214_11695.jpg 538 | hermit_crab/n01986214_15370.jpg 539 | hermit_crab/n01986214_18591.jpg 540 | hermit_crab/n01986214_3659.jpg 541 | horse/n02374451_9225.jpg 542 | horse/n02374451_8269.jpg 543 | horse/n02374451_4818.jpg 544 | horse/n02374451_262.jpg 545 | horse/n02374451_11492.jpg 546 | horse/n02374451_8065.jpg 547 | horse/n02374451_12174.jpg 548 | horse/n02374451_503.jpg 549 | horse/n02374451_11894.jpg 550 | horse/n02374451_17474.jpg 551 | hot-air_balloon/n02782093_5111.jpg 552 | hot-air_balloon/n02782093_1989.jpg 553 | hot-air_balloon/n03541923_1963.jpg 554 | hot-air_balloon/n03541923_3571.jpg 555 | hot-air_balloon/n02782093_5421.jpg 556 | hot-air_balloon/n03541923_7575.jpg 557 | hot-air_balloon/n03541923_3913.jpg 558 | hot-air_balloon/n02782093_4675.jpg 559 | hot-air_balloon/n03541923_1367.jpg 560 | hot-air_balloon/n02782093_622.jpg 561 | hotdog/n07697537_23696.jpg 562 | hotdog/n07697537_32913.jpg 563 | hotdog/n07697537_23589.jpg 564 | hotdog/n07697537_13566.jpg 565 | hotdog/n07697537_1271.jpg 566 | hotdog/n07697537_23978.jpg 567 | hotdog/n07697537_34124.jpg 568 | hotdog/n07697537_14033.jpg 569 | hotdog/n07697537_19413.jpg 570 | hotdog/n07697537_18866.jpg 571 | hourglass/n03544143_8556.jpg 572 | hourglass/n03544143_4358.jpg 573 | hourglass/n03544143_10288.jpg 574 | hourglass/n03544143_577.jpg 575 | hourglass/n03544143_3348.jpg 576 | hourglass/n03544143_4868.jpg 577 | hourglass/n03544143_11550.jpg 578 | hourglass/n03544143_15035.jpg 579 | hourglass/n03544143_17851.jpg 580 | hourglass/n03544143_10028.jpg 581 | jack-o-lantern/n03590841_5679.jpg 582 | jack-o-lantern/n03590841_11444.jpg 583 | jack-o-lantern/n03590841_6382.jpg 584 | jack-o-lantern/n03590841_3947.jpg 585 | jack-o-lantern/n03590841_11344.jpg 586 | jack-o-lantern/n03590841_6963.jpg 587 | jack-o-lantern/n03590841_10464.jpg 588 | jack-o-lantern/n03590841_6798.jpg 589 | jack-o-lantern/n03590841_11395.jpg 590 | jack-o-lantern/n03590841_7100.jpg 591 | jellyfish/n01910747_1552.jpg 592 | jellyfish/n01910747_10023.jpg 593 | jellyfish/n01910747_563.jpg 594 | jellyfish/n01910747_13178.jpg 595 | jellyfish/n01910747_2724.jpg 596 | jellyfish/n01910747_1522.jpg 597 | jellyfish/n01910747_2663.jpg 598 | jellyfish/n01910747_11859.jpg 599 | jellyfish/n01910747_11139.jpg 600 | jellyfish/n01910747_3076.jpg 601 | kangaroo/n01877134_11382.jpg 602 | kangaroo/n01877134_1720.jpg 603 | kangaroo/n01877134_1108.jpg 604 | kangaroo/n01877134_9114.jpg 605 | kangaroo/n01877134_9452.jpg 606 | kangaroo/n01877134_11335.jpg 607 | kangaroo/n01877134_10341.jpg 608 | kangaroo/n01877134_10270.jpg 609 | kangaroo/n01877134_10244.jpg 610 | kangaroo/n01877134_728.jpg 611 | knife/n02976123_8715.jpg 612 | knife/n03624134_13816.jpg 613 | knife/n02976123_3654.jpg 614 | knife/n03624134_17319.jpg 615 | knife/n02976123_4150.jpg 616 | knife/n03624134_15272.jpg 617 | knife/n02976123_1752.jpg 618 | knife/n03624134_12494.jpg 619 | knife/n03624134_15440.jpg 620 | knife/n03624134_12920.jpg 621 | lion/n02129165_13679.jpg 622 | lion/n02129165_325.jpg 623 | lion/n02129165_5082.jpg 624 | lion/n02129165_20590.jpg 625 | lion/n02129165_5079.jpg 626 | lion/n02129165_4588.jpg 627 | lion/n02129165_4058.jpg 628 | lion/n02129165_12630.jpg 629 | lion/n02129165_1901.jpg 630 | lion/n02129165_2496.jpg 631 | lizard/n01674464_1831.jpg 632 | lizard/n01693334_10211.jpg 633 | lizard/n01693334_9741.jpg 634 | lizard/n01674464_1561.jpg 635 | lizard/n01693334_1360.jpg 636 | lizard/n01674464_11126.jpg 637 | lizard/n01674464_2019.jpg 638 | lizard/n01674464_10844.jpg 639 | lizard/n01674464_7290.jpg 640 | lizard/n01683558_5963.jpg 641 | lobster/n01985128_9104.jpg 642 | lobster/n01985128_32662.jpg 643 | lobster/n01985128_20238.jpg 644 | lobster/n01985128_9799.jpg 645 | lobster/n01985128_749.jpg 646 | lobster/n01985128_5607.jpg 647 | lobster/n01985128_27838.jpg 648 | lobster/n01985128_12648.jpg 649 | lobster/n01985128_29355.jpg 650 | lobster/n01985128_52414.jpg 651 | motorcycle/n03790512_12830.jpg 652 | motorcycle/n03790512_1734.jpg 653 | motorcycle/n03790512_2611.jpg 654 | motorcycle/n03790512_345.jpg 655 | motorcycle/n03790512_12653.jpg 656 | motorcycle/n03790512_400.jpg 657 | motorcycle/n03790512_393.jpg 658 | motorcycle/n03790512_10417.jpg 659 | motorcycle/n03790512_13786.jpg 660 | motorcycle/n03790512_13819.jpg 661 | mouse/n02330245_822.jpg 662 | mouse/n02330245_9247.jpg 663 | mouse/n02330245_137.jpg 664 | mouse/n02330245_3252.jpg 665 | mouse/n02330245_15445.jpg 666 | mouse/n02330245_15553.jpg 667 | mouse/n02330245_11922.jpg 668 | mouse/n02330245_12266.jpg 669 | mouse/n02330245_15069.jpg 670 | mouse/n02330245_3938.jpg 671 | mushroom/n12998815_9340.jpg 672 | mushroom/n12998815_5759.jpg 673 | mushroom/n12997919_7227.jpg 674 | mushroom/n12998815_12868.jpg 675 | mushroom/n12998815_6887.jpg 676 | mushroom/n12997919_3128.jpg 677 | mushroom/n12997919_2931.jpg 678 | mushroom/n12997919_7536.jpg 679 | mushroom/n12998815_22994.jpg 680 | mushroom/n12997919_3413.jpg 681 | owl/n01621127_2580.jpg 682 | owl/n01621127_13869.jpg 683 | owl/n01621127_2379.jpg 684 | owl/n01621127_2756.jpg 685 | owl/n01621127_14495.jpg 686 | owl/n01621127_208.jpg 687 | owl/n01621127_18305.jpg 688 | owl/n01621127_2733.jpg 689 | owl/n01621127_12894.jpg 690 | owl/n01621127_2025.jpg 691 | parrot/n01816887_4698.jpg 692 | parrot/n01817953_3417.jpg 693 | parrot/n01816887_224.jpg 694 | parrot/n01816887_8397.jpg 695 | parrot/n01817953_9201.jpg 696 | parrot/n01816887_1320.jpg 697 | parrot/n01816887_9494.jpg 698 | parrot/n01816887_6792.jpg 699 | parrot/n01820546_6753.jpg 700 | parrot/n01816887_9978.jpg 701 | pear/n12651611_7758.jpg 702 | pear/n12651611_10828.jpg 703 | pear/n12651611_7402.jpg 704 | pear/n07768230_3086.jpg 705 | pear/n07768230_13745.jpg 706 | pear/n07767847_4987.jpg 707 | pear/n12651611_6870.jpg 708 | pear/n07767847_5698.jpg 709 | pear/n07767847_10507.jpg 710 | pear/n12651611_3083.jpg 711 | penguin/n02055803_3489.jpg 712 | penguin/n02056570_8253.jpg 713 | penguin/n02055803_7917.jpg 714 | penguin/n02055803_1502.jpg 715 | penguin/n02055803_2199.jpg 716 | penguin/n02056570_8709.jpg 717 | penguin/n02055803_422.jpg 718 | penguin/n02055803_3848.jpg 719 | penguin/n02056570_9810.jpg 720 | penguin/n02055803_6994.jpg 721 | piano/n03452741_8023.jpg 722 | piano/n03452741_6003.jpg 723 | piano/n03452741_6174.jpg 724 | piano/n03452741_4470.jpg 725 | piano/n03452741_10838.jpg 726 | piano/n03452741_6507.jpg 727 | piano/n03452741_10036.jpg 728 | piano/n03452741_16897.jpg 729 | piano/n03452741_15379.jpg 730 | piano/n03452741_8763.jpg 731 | pickup_truck/n03930630_5135.jpg 732 | pickup_truck/n03930630_9324.jpg 733 | pickup_truck/n03930630_6704.jpg 734 | pickup_truck/n03930630_10956.jpg 735 | pickup_truck/n03930630_7422.jpg 736 | pickup_truck/n03930630_5801.jpg 737 | pickup_truck/n03930630_7937.jpg 738 | pickup_truck/n03930630_661.jpg 739 | pickup_truck/n03930630_19292.jpg 740 | pickup_truck/n03930630_11741.jpg 741 | pig/n02395406_33040.jpg 742 | pig/n02395406_5730.jpg 743 | pig/n02395406_7226.jpg 744 | pig/n02395406_11157.jpg 745 | pig/n02395406_33109.jpg 746 | pig/n02395406_2776.jpg 747 | pig/n02395406_27034.jpg 748 | pig/n02395406_32520.jpg 749 | pig/n02395406_2303.jpg 750 | pig/n02395406_8445.jpg 751 | pineapple/n07753275_20080.jpg 752 | pineapple/n07753275_21346.jpg 753 | pineapple/n07753275_28461.jpg 754 | pineapple/n07753275_23767.jpg 755 | pineapple/n07753275_27111.jpg 756 | pineapple/n07753275_17554.jpg 757 | pineapple/n07753275_26165.jpg 758 | pineapple/n07753275_15275.jpg 759 | pineapple/n07753275_26035.jpg 760 | pineapple/n07753275_29605.jpg 761 | pistol/n03948459_12984.jpg 762 | pistol/n03948459_18932.jpg 763 | pistol/n04086273_3414.jpg 764 | pistol/n03948459_8078.jpg 765 | pistol/n03948459_13616.jpg 766 | pistol/n03948459_10942.jpg 767 | pistol/n04086273_5765.jpg 768 | pistol/n03948459_20127.jpg 769 | pistol/n03948459_17330.jpg 770 | pistol/n04086273_23984.jpg 771 | pizza/n07873807_12811.jpg 772 | pizza/n07873807_8370.jpg 773 | pizza/n07873807_1766.jpg 774 | pizza/n07873807_10042.jpg 775 | pizza/n07873807_12794.jpg 776 | pizza/n07873807_9429.jpg 777 | pizza/n07873807_2885.jpg 778 | pizza/n07873807_10932.jpg 779 | pizza/n07873807_3557.jpg 780 | pizza/n07873807_5785.jpg 781 | pretzel/n07695742_8381.jpg 782 | pretzel/n07695742_9706.jpg 783 | pretzel/n07695742_4492.jpg 784 | pretzel/n07695742_6230.jpg 785 | pretzel/n07695742_8444.jpg 786 | pretzel/n07695742_3358.jpg 787 | pretzel/n07695742_6804.jpg 788 | pretzel/n07695742_4716.jpg 789 | pretzel/n07695742_4484.jpg 790 | pretzel/n07695742_1832.jpg 791 | rabbit/n02325366_327.jpg 792 | rabbit/n02325366_6689.jpg 793 | rabbit/n02325366_11692.jpg 794 | rabbit/n02325366_5095.jpg 795 | rabbit/n02325366_6527.jpg 796 | rabbit/n02325366_6132.jpg 797 | rabbit/n02325366_5669.jpg 798 | rabbit/n02325366_6270.jpg 799 | rabbit/n02325366_9233.jpg 800 | rabbit/n02325366_5308.jpg 801 | raccoon/n02508021_10406.jpg 802 | raccoon/n02508021_12493.jpg 803 | raccoon/n02508021_2938.jpg 804 | raccoon/n02508021_1752.jpg 805 | raccoon/n02508021_2458.jpg 806 | raccoon/n02508021_15891.jpg 807 | raccoon/n02508021_4712.jpg 808 | raccoon/n02508021_6078.jpg 809 | raccoon/n02508021_750.jpg 810 | raccoon/n02508021_10788.jpg 811 | racket/n04039381_14533.jpg 812 | racket/n02772700_1761.jpg 813 | racket/n02772700_2240.jpg 814 | racket/n02772700_1967.jpg 815 | racket/n04039381_8327.jpg 816 | racket/n04039381_8917.jpg 817 | racket/n04039381_593.jpg 818 | racket/n04039381_4860.jpg 819 | racket/n04039381_7370.jpg 820 | racket/n04409806_5226.jpg 821 | ray/n01496331_33236.jpg 822 | ray/n01496331_10373.jpg 823 | ray/n01496331_34865.jpg 824 | ray/n01498041_2351.jpg 825 | ray/n01496331_21710.jpg 826 | ray/n01496331_15801.jpg 827 | ray/n01496331_23623.jpg 828 | ray/n01498041_4234.jpg 829 | ray/n01498041_197.jpg 830 | ray/n01496331_13467.jpg 831 | rhinoceros/n02391994_3673.jpg 832 | rhinoceros/n02391994_14936.jpg 833 | rhinoceros/n02391994_18653.jpg 834 | rhinoceros/n02391994_16313.jpg 835 | rhinoceros/n02391994_17073.jpg 836 | rhinoceros/n02391994_15908.jpg 837 | rhinoceros/n02391994_3259.jpg 838 | rhinoceros/n02391994_19014.jpg 839 | rhinoceros/n02391994_530.jpg 840 | rhinoceros/n02391994_12492.jpg 841 | rifle/n03416775_11882.jpg 842 | rifle/n03416775_14703.jpg 843 | rifle/n02749479_7185.jpg 844 | rifle/n02749479_407.jpg 845 | rifle/n02907391_21279.jpg 846 | rifle/n04090263_24512.jpg 847 | rifle/n02749479_1437.jpg 848 | rifle/n02749479_7403.jpg 849 | rifle/n03416775_8292.jpg 850 | rifle/n02749479_4107.jpg 851 | rocket/n04099429_20475.jpg 852 | rocket/n04415663_4998.jpg 853 | rocket/n04415663_11259.jpg 854 | rocket/n04415663_4677.jpg 855 | rocket/n03773504_21983.jpg 856 | rocket/n04099429_8939.jpg 857 | rocket/n03773504_22120.jpg 858 | rocket/n04415663_4099.jpg 859 | rocket/n04099429_1534.jpg 860 | rocket/n03773504_6665.jpg 861 | sailboat/n04128499_6235.jpg 862 | sailboat/n04128499_2799.jpg 863 | sailboat/n04128499_5397.jpg 864 | sailboat/n04128499_6797.jpg 865 | sailboat/n04128499_2893.jpg 866 | sailboat/n04128499_6762.jpg 867 | sailboat/n04128499_7174.jpg 868 | sailboat/n04128499_230.jpg 869 | sailboat/n04128499_3865.jpg 870 | sailboat/n04128499_13658.jpg 871 | saw/n04140064_8460.jpg 872 | saw/n04140064_18814.jpg 873 | saw/n02770585_3240.jpg 874 | saw/n04140064_23555.jpg 875 | saw/n04140064_45049.jpg 876 | saw/n04140064_10178.jpg 877 | saw/n04140064_18657.jpg 878 | saw/n04140064_3739.jpg 879 | saw/n04140064_19771.jpg 880 | saw/n02770585_5006.jpg 881 | saxophone/n04141076_41535.jpg 882 | saxophone/n04141076_4140.jpg 883 | saxophone/n04141076_39197.jpg 884 | saxophone/n04141076_35278.jpg 885 | saxophone/n04141076_10007.jpg 886 | saxophone/n04141076_4403.jpg 887 | saxophone/n04141076_39548.jpg 888 | saxophone/n04141076_12329.jpg 889 | saxophone/n04141076_41723.jpg 890 | saxophone/n04141076_42145.jpg 891 | scissors/n04148054_2603.jpg 892 | scissors/n04148054_13312.jpg 893 | scissors/n04148054_15870.jpg 894 | scissors/n04148054_7900.jpg 895 | scissors/n04148054_20645.jpg 896 | scissors/n04148054_1233.jpg 897 | scissors/n04148054_11540.jpg 898 | scissors/n04148054_370.jpg 899 | scissors/n04148054_17514.jpg 900 | scissors/n04148054_8138.jpg 901 | scorpion/n01770393_3955.jpg 902 | scorpion/n01770393_144.jpg 903 | scorpion/n01770393_2975.jpg 904 | scorpion/n01770393_13358.jpg 905 | scorpion/n01770393_6144.jpg 906 | scorpion/n01770393_3338.jpg 907 | scorpion/n01770393_12256.jpg 908 | scorpion/n01770393_10666.jpg 909 | scorpion/n01770393_27241.jpg 910 | scorpion/n01770393_1283.jpg 911 | sea_turtle/n01664065_9290.jpg 912 | sea_turtle/n01664065_15939.jpg 913 | sea_turtle/n01663401_3035.jpg 914 | sea_turtle/n01663401_8444.jpg 915 | sea_turtle/n01663401_11120.jpg 916 | sea_turtle/n01663401_1997.jpg 917 | sea_turtle/n01663401_8261.jpg 918 | sea_turtle/n01664065_2112.jpg 919 | sea_turtle/n01664065_26514.jpg 920 | sea_turtle/n01664065_4095.jpg 921 | seagull/n02041246_8583.jpg 922 | seagull/n02041246_7834.jpg 923 | seagull/n02041246_1005.jpg 924 | seagull/n02041246_18235.jpg 925 | seagull/n02041246_4512.jpg 926 | seagull/n02041246_18641.jpg 927 | seagull/n02041246_5731.jpg 928 | seagull/n02041246_35326.jpg 929 | seagull/n02041246_4572.jpg 930 | seagull/n02041246_16400.jpg 931 | seal/n02076196_17306.jpg 932 | seal/n02077923_95.jpg 933 | seal/n02077923_1297.jpg 934 | seal/n02077923_1231.jpg 935 | seal/n02077923_12945.jpg 936 | seal/n02077923_3713.jpg 937 | seal/n02077923_521.jpg 938 | seal/n02076196_1919.jpg 939 | seal/n02076196_2469.jpg 940 | seal/n02076196_10596.jpg 941 | shark/n01491361_7484.jpg 942 | shark/n01483021_50.jpg 943 | shark/n01494475_2974.jpg 944 | shark/n01482330_3285.jpg 945 | shark/n01494475_8365.jpg 946 | shark/n01494475_714.jpg 947 | shark/n01491361_4444.jpg 948 | shark/n01491361_4716.jpg 949 | shark/n01482330_1753.jpg 950 | shark/n01489920_9229.jpg 951 | sheep/n02412080_2578.jpg 952 | sheep/n02412080_7720.jpg 953 | sheep/n02411705_13543.jpg 954 | sheep/n02411705_14656.jpg 955 | sheep/n02411705_1556.jpg 956 | sheep/n02412080_7753.jpg 957 | sheep/n02411705_6546.jpg 958 | sheep/n02412080_6209.jpg 959 | sheep/n02413131_1409.jpg 960 | sheep/n02412080_7894.jpg 961 | shoe/n03680355_2585.jpg 962 | shoe/n04120489_3237.jpg 963 | shoe/n04120489_18.jpg 964 | shoe/n03680355_4637.jpg 965 | shoe/n04120489_1990.jpg 966 | shoe/n04120489_3518.jpg 967 | shoe/n04120489_3775.jpg 968 | shoe/n04546081_2665.jpg 969 | shoe/n04120489_4013.jpg 970 | shoe/n03680355_887.jpg 971 | skyscraper/n04233124_4190.jpg 972 | skyscraper/n04233124_3400.jpg 973 | skyscraper/n04233124_8777.jpg 974 | skyscraper/n04233124_1041.jpg 975 | skyscraper/n04233124_11521.jpg 976 | skyscraper/n04233124_18366.jpg 977 | skyscraper/n04233124_15373.jpg 978 | skyscraper/n04233124_11849.jpg 979 | skyscraper/n04233124_15445.jpg 980 | skyscraper/n04233124_3550.jpg 981 | snail/n01944390_3725.jpg 982 | snail/n01944390_5807.jpg 983 | snail/n01944390_10342.jpg 984 | snail/n01944390_3028.jpg 985 | snail/n01944390_4191.jpg 986 | snail/n01944390_6382.jpg 987 | snail/n01944390_4196.jpg 988 | snail/n01944390_10864.jpg 989 | snail/n01944390_7676.jpg 990 | snail/n01944390_11743.jpg 991 | snake/n01748264_14024.jpg 992 | snake/n01726692_2433.jpg 993 | snake/n01726692_5470.jpg 994 | snake/n01726692_80.jpg 995 | snake/n01726692_1615.jpg 996 | snake/n01726692_2164.jpg 997 | snake/n01726692_981.jpg 998 | snake/n01726692_9144.jpg 999 | snake/n01729977_5068.jpg 1000 | snake/n01726692_182.jpg 1001 | songbird/n01527347_331.jpg 1002 | songbird/n01531178_11416.jpg 1003 | songbird/n01594787_5190.jpg 1004 | songbird/n01592084_5207.jpg 1005 | songbird/n01592084_760.jpg 1006 | songbird/n01530575_509.jpg 1007 | songbird/n01531178_6171.jpg 1008 | songbird/n01558993_19761.jpg 1009 | songbird/n01558993_2370.jpg 1010 | songbird/n01527347_17690.jpg 1011 | spider/n01772222_954.jpg 1012 | spider/n01772222_3092.jpg 1013 | spider/n01772222_3438.jpg 1014 | spider/n01772222_10676.jpg 1015 | spider/n01772222_6998.jpg 1016 | spider/n01772222_12437.jpg 1017 | spider/n01772222_8205.jpg 1018 | spider/n01772222_1171.jpg 1019 | spider/n01772222_9346.jpg 1020 | spider/n01772222_8692.jpg 1021 | spoon/n04350769_6642.jpg 1022 | spoon/n04350769_483.jpg 1023 | spoon/n03633091_15236.jpg 1024 | spoon/n04350769_5916.jpg 1025 | spoon/n03633091_1273.jpg 1026 | spoon/n04597913_6066.jpg 1027 | spoon/n04284002_20512.jpg 1028 | spoon/n04350769_9981.jpg 1029 | spoon/n03633091_15372.jpg 1030 | spoon/n04350769_1634.jpg 1031 | squirrel/n02355227_13541.jpg 1032 | squirrel/n02355227_13077.jpg 1033 | squirrel/n02355227_15209.jpg 1034 | squirrel/n02355227_15129.jpg 1035 | squirrel/n02355227_12764.jpg 1036 | squirrel/n02355227_9552.jpg 1037 | squirrel/n02355227_14224.jpg 1038 | squirrel/n02355227_451.jpg 1039 | squirrel/n02355227_14716.jpg 1040 | squirrel/n02355227_15878.jpg 1041 | starfish/n02317335_832.jpg 1042 | starfish/n02317335_16368.jpg 1043 | starfish/n02317335_2072.jpg 1044 | starfish/n02317335_572.jpg 1045 | starfish/n02317335_762.jpg 1046 | starfish/n02317335_4568.jpg 1047 | starfish/n02317335_22895.jpg 1048 | starfish/n02317335_19633.jpg 1049 | starfish/n02317335_26045.jpg 1050 | starfish/n02317335_16276.jpg 1051 | strawberry/n07745940_2642.jpg 1052 | strawberry/n07745940_5035.jpg 1053 | strawberry/n07745940_991.jpg 1054 | strawberry/n07745940_2329.jpg 1055 | strawberry/n07745940_2244.jpg 1056 | strawberry/n07745940_20814.jpg 1057 | strawberry/n07745940_12035.jpg 1058 | strawberry/n07745940_3145.jpg 1059 | strawberry/n07745940_5599.jpg 1060 | strawberry/n07745940_20922.jpg 1061 | swan/n01858441_9652.jpg 1062 | swan/n01858441_2937.jpg 1063 | swan/n01858441_1041.jpg 1064 | swan/n01860187_739.jpg 1065 | swan/n01858441_7162.jpg 1066 | swan/n01858441_10732.jpg 1067 | swan/n01858441_2850.jpg 1068 | swan/n01858441_9626.jpg 1069 | swan/n01858845_12958.jpg 1070 | swan/n01860187_950.jpg 1071 | sword/n04373894_29721.jpg 1072 | sword/n04373894_11246.jpg 1073 | sword/n04373894_37359.jpg 1074 | sword/n04147793_13881.jpg 1075 | sword/n04373894_40106.jpg 1076 | sword/n04373894_35504.jpg 1077 | sword/n04373894_36791.jpg 1078 | sword/n04373894_32014.jpg 1079 | sword/n04373894_47693.jpg 1080 | sword/n04147793_6149.jpg 1081 | table/n04379964_24141.jpg 1082 | table/n03201208_32524.jpg 1083 | table/n03201208_25958.jpg 1084 | table/n04379243_1127.jpg 1085 | table/n04379964_2419.jpg 1086 | table/n04379243_23142.jpg 1087 | table/n04379243_24695.jpg 1088 | table/n03201208_37167.jpg 1089 | table/n03201208_2363.jpg 1090 | table/n04379964_30971.jpg 1091 | tank/n04389033_3055.jpg 1092 | tank/n04389033_24620.jpg 1093 | tank/n04389033_25719.jpg 1094 | tank/n04389033_8657.jpg 1095 | tank/n04389033_19025.jpg 1096 | tank/n04389033_23639.jpg 1097 | tank/n04389033_25541.jpg 1098 | tank/n04389033_12442.jpg 1099 | tank/n04389033_14209.jpg 1100 | tank/n04389033_20803.jpg 1101 | teapot/n04398044_18256.jpg 1102 | teapot/n04398044_19613.jpg 1103 | teapot/n04398044_2917.jpg 1104 | teapot/n04398044_5486.jpg 1105 | teapot/n04398044_14381.jpg 1106 | teapot/n04398044_7880.jpg 1107 | teapot/n04398044_25219.jpg 1108 | teapot/n04398044_35589.jpg 1109 | teapot/n04398044_11626.jpg 1110 | teapot/n04398044_31041.jpg 1111 | teddy_bear/n04399382_22297.jpg 1112 | teddy_bear/n04399382_16584.jpg 1113 | teddy_bear/n04399382_1100.jpg 1114 | teddy_bear/n04399382_26765.jpg 1115 | teddy_bear/n04399382_4704.jpg 1116 | teddy_bear/n04399382_34899.jpg 1117 | teddy_bear/n04399382_2073.jpg 1118 | teddy_bear/n04399382_11606.jpg 1119 | teddy_bear/n04399382_33819.jpg 1120 | teddy_bear/n04399382_28912.jpg 1121 | tiger/n02129604_9840.jpg 1122 | tiger/n02129604_15991.jpg 1123 | tiger/n02129604_1431.jpg 1124 | tiger/n02129604_1675.jpg 1125 | tiger/n02129604_7019.jpg 1126 | tiger/n02129604_8618.jpg 1127 | tiger/n02129604_15877.jpg 1128 | tiger/n02129604_10691.jpg 1129 | tiger/n02129604_17243.jpg 1130 | tiger/n02129604_570.jpg 1131 | tree/n12755727_1126.jpg 1132 | tree/n12726670_6812.jpg 1133 | tree/n12713866_6140.jpg 1134 | tree/n12305089_6316.jpg 1135 | tree/n12593994_8214.jpg 1136 | tree/n11646344_1934.jpg 1137 | tree/n12582231_1696.jpg 1138 | tree/n12587803_33319.jpg 1139 | tree/n12713866_9719.jpg 1140 | tree/n12523475_9706.jpg 1141 | trumpet/n03110669_93834.jpg 1142 | trumpet/n03110669_148863.jpg 1143 | trumpet/n03110669_89645.jpg 1144 | trumpet/n03110669_96018.jpg 1145 | trumpet/n03110669_98252.jpg 1146 | trumpet/n03110669_91100.jpg 1147 | trumpet/n03110669_120627.jpg 1148 | trumpet/n03110669_98315.jpg 1149 | trumpet/n03110669_73253.jpg 1150 | trumpet/n03110669_19514.jpg 1151 | turtle/n01669191_7682.jpg 1152 | turtle/n01669191_3313.jpg 1153 | turtle/n01669191_1717.jpg 1154 | turtle/n01669191_5474.jpg 1155 | turtle/n01669191_10039.jpg 1156 | turtle/n01669191_9556.jpg 1157 | turtle/n01669191_9607.jpg 1158 | turtle/n01669191_6969.jpg 1159 | turtle/n01669191_329.jpg 1160 | turtle/n01669191_9750.jpg 1161 | umbrella/n04507155_8835.jpg 1162 | umbrella/n04507155_1165.jpg 1163 | umbrella/n04507155_15141.jpg 1164 | umbrella/n04507155_17309.jpg 1165 | umbrella/n04507155_3113.jpg 1166 | umbrella/n04507155_4852.jpg 1167 | umbrella/n04507155_2158.jpg 1168 | umbrella/n04507155_9727.jpg 1169 | umbrella/n04507155_5583.jpg 1170 | umbrella/n04507155_17155.jpg 1171 | violin/n04536866_1290.jpg 1172 | violin/n02803934_21947.jpg 1173 | violin/n04536866_1926.jpg 1174 | violin/n04536866_2123.jpg 1175 | violin/n04536866_6286.jpg 1176 | violin/n04536866_10581.jpg 1177 | violin/n04536866_9941.jpg 1178 | violin/n04536866_12600.jpg 1179 | violin/n04536866_7704.jpg 1180 | violin/n02803934_1380.jpg 1181 | volcano/n09472597_38264.jpg 1182 | volcano/n09472597_52209.jpg 1183 | volcano/n09472597_13955.jpg 1184 | volcano/n09472597_35779.jpg 1185 | volcano/n09472597_34930.jpg 1186 | volcano/n09472597_446.jpg 1187 | volcano/n09472597_12664.jpg 1188 | volcano/n09472597_16556.jpg 1189 | volcano/n09472597_39459.jpg 1190 | volcano/n09472597_36414.jpg 1191 | wading_bird/n02007558_4041.jpg 1192 | wading_bird/n02007558_4727.jpg 1193 | wading_bird/n02009912_10484.jpg 1194 | wading_bird/n02009912_2366.jpg 1195 | wading_bird/n02009912_13753.jpg 1196 | wading_bird/n02012849_2562.jpg 1197 | wading_bird/n02009912_6953.jpg 1198 | wading_bird/n02009912_13815.jpg 1199 | wading_bird/n02000954_14335.jpg 1200 | wading_bird/n02009912_13292.jpg 1201 | wheelchair/n04576002_8360.jpg 1202 | wheelchair/n04576002_388.jpg 1203 | wheelchair/n04576002_1665.jpg 1204 | wheelchair/n04576002_6833.jpg 1205 | wheelchair/n04576002_602.jpg 1206 | wheelchair/n04576002_6836.jpg 1207 | wheelchair/n04576002_8668.jpg 1208 | wheelchair/n04576002_7851.jpg 1209 | wheelchair/n04576002_7247.jpg 1210 | wheelchair/n04576002_8774.jpg 1211 | windmill/n04587404_14737.jpg 1212 | windmill/n04587404_18560.jpg 1213 | windmill/n04587559_13032.jpg 1214 | windmill/n04587559_11407.jpg 1215 | windmill/n04587559_9281.jpg 1216 | windmill/n04587559_13370.jpg 1217 | windmill/n04587404_14153.jpg 1218 | windmill/n04587559_11131.jpg 1219 | windmill/n04587559_9715.jpg 1220 | windmill/n04587559_12849.jpg 1221 | window/n04589593_11620.jpg 1222 | window/n04587648_16394.jpg 1223 | window/n04589593_12197.jpg 1224 | window/n04587648_4526.jpg 1225 | window/n04589593_13410.jpg 1226 | window/n04587648_811.jpg 1227 | window/n04589593_12080.jpg 1228 | window/n04587648_3571.jpg 1229 | window/n03227184_3521.jpg 1230 | window/n04587648_16618.jpg 1231 | wine_bottle/n04591713_4821.jpg 1232 | wine_bottle/n04591713_1989.jpg 1233 | wine_bottle/n04591713_3955.jpg 1234 | wine_bottle/n04591713_3642.jpg 1235 | wine_bottle/n04591713_4054.jpg 1236 | wine_bottle/n04591713_6022.jpg 1237 | wine_bottle/n04591713_3570.jpg 1238 | wine_bottle/n04591713_727.jpg 1239 | wine_bottle/n04591713_1407.jpg 1240 | wine_bottle/n04591713_110.jpg 1241 | zebra/n02391049_1950.jpg 1242 | zebra/n02391049_9822.jpg 1243 | zebra/n02391049_1551.jpg 1244 | zebra/n02391049_4069.jpg 1245 | zebra/n02391049_10452.jpg 1246 | zebra/n02391049_1526.jpg 1247 | zebra/n02391049_7077.jpg 1248 | zebra/n02391049_7261.jpg 1249 | zebra/n02391049_6467.jpg 1250 | zebra/n02391049_5695.jpg 1251 | -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import os 3 | from models.TripletEmbedding import TripletNet 4 | 5 | 6 | def str2bool(v): 7 | if v.lower() in ('yes', 'true', 't', 'y', '1'): 8 | return True 9 | elif v.lower() in ('no', 'false', 'f', 'n', '0'): 10 | return False 11 | else: 12 | raise argparse.ArgumentTypeError('Unsupported value encountered.') 13 | 14 | 15 | def parase_args(): 16 | parser = argparse.ArgumentParser() 17 | 18 | parser.add_argument('--photo_root', type=str, default='/data1/zzl/dataset/photo-train', help='Training photo root') 19 | parser.add_argument('--sketch_root', type=str, default='/data1/zzl/dataset/sketch-triplet-train', 20 | help='Training sketch root') 21 | parser.add_argument('--batch_size', type=int, default=16, help='The size of batch (default :16') 22 | parser.add_argument('--device', type=str, default='0', help='The cuda device to be used (default: 0)') 23 | parser.add_argument('--epochs', type=int, default=1000, help='The number of epochs to run (default: 1000)') 24 | parser.add_argument('--lr', type=float, default=1e-5, help='The learning rate of the model') 25 | 26 | parser.add_argument('--test', type=str2bool, nargs='?', default=True) 27 | parser.add_argument('--test_f', type=int, default=5, help='The frequency of testing (default: 5)') 28 | parser.add_argument('--photo_test', type=str, default='/data1/zzl/dataset/photo-test', help='Testing photo root') 29 | parser.add_argument('--sketch_test', type=str, default='/data1/zzl/dataset/sketch-triplet-test', 30 | help='Testing sketch root') 31 | 32 | parser.add_argument('--save_model', type=str2bool, nargs='?', default=False) 33 | parser.add_argument('--save_dir', type=str, default='/data1/zzl/model/caffe2torch/vgg_triplet_loss', 34 | help='The folder to save the model status') 35 | 36 | parser.add_argument('--vis', type=str2bool, nargs='?', default=True, help='Whether to visualize') 37 | parser.add_argument('--env', type=str, default='caffe2torch_tripletloss', help='The visualization environment') 38 | 39 | parser.add_argument('--fine_tune', type=str2bool, nargs='?', default=False, help='Whether to fine tune') 40 | parser.add_argument('--model_root', type=str, default=None, help='The model status files\'s root') 41 | 42 | parser.add_argument('--margin', type=float, default=0.3, help='The margin of the triplet loss') 43 | parser.add_argument('--p', type=int, default=2, help='The p of the triplet loss') 44 | 45 | parser.add_argument('--net', type=str, default='vgg16', help='The model to be used (vgg16, resnet34, resnet50)') 46 | parser.add_argument('--cat', type=str2bool, nargs='?', default=True, help='Whether to use category loss') 47 | 48 | return check_args(parser.parse_args()) 49 | 50 | 51 | def check_args(args): 52 | if args.save_model: 53 | save_photo_dir = os.path.join(args.save_dir, 'photo') 54 | save_sketch_dir = os.path.join(args.save_dir, 'sketch') 55 | if not os.path.exists(args.save_dir): 56 | os.mkdir(args.save_dir) 57 | os.mkdir(save_photo_dir) 58 | os.mkdir(save_sketch_dir) 59 | 60 | try: 61 | assert args.epochs >= 1 62 | except: 63 | print('number of epochs must be larger than or equal to one') 64 | 65 | try: 66 | assert args.batch_size >= 1 67 | except: 68 | print('batch size must be larger than or equal to one') 69 | 70 | try: 71 | assert args.net in ['vgg16', 'resnet34', 'resnet50'] 72 | except: 73 | print('net model must be chose from [\'vgg16\', \'resnet34\', \'resnet50\']') 74 | 75 | if args.fine_tune: 76 | try: 77 | assert not args.model_root 78 | 79 | except: 80 | print('you should specify the model status file') 81 | 82 | return args 83 | 84 | 85 | def main(): 86 | 87 | args = parase_args() 88 | if args is None: 89 | exit() 90 | 91 | os.environ["CUDA_VISIBLE_DEVICES"] = str(args.device) 92 | 93 | tripletNet = TripletNet(args) 94 | tripletNet.train() 95 | 96 | 97 | if __name__ == '__main__': 98 | 99 | main() -------------------------------------------------------------------------------- /utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CDOTAD/SketchyDatabase/0d8f5abf98c8a6d9847e16258493a1e56371b54c/utils/__init__.py -------------------------------------------------------------------------------- /utils/extractor.py: -------------------------------------------------------------------------------- 1 | import os 2 | from PIL import Image 3 | import torch as t 4 | import torchvision as tv 5 | from torch import nn 6 | import pickle 7 | from utils.visualize import Visualizer 8 | from data import ImageDataLoader 9 | import numpy as np 10 | 11 | 12 | class Config(object): 13 | def __init__(self): 14 | return 15 | 16 | 17 | class Extractor(object): 18 | 19 | def __init__(self, e_model, batch_size=128, cat_info=True, 20 | vis=False, dataloader=False): 21 | self.batch_size = batch_size 22 | self.cat_info = cat_info 23 | 24 | self.model = e_model 25 | 26 | if dataloader: 27 | self.dataloader = dataloader 28 | else: 29 | self.transform = tv.transforms.Compose([ 30 | tv.transforms.Resize(224), 31 | tv.transforms.ToTensor(), 32 | tv.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) 33 | ]) 34 | self.vis = vis 35 | if self.vis: 36 | self.viser = Visualizer('caffe2torch_test') 37 | 38 | # extract the inputs' feature via self.model 39 | # the model's output only contains the inputs' feature 40 | @t.no_grad() 41 | def extract(self, data_root, out_root=None): 42 | if self.dataloader: 43 | return self._extract_with_dataloader(data_root=data_root, cat_info=self.cat_info, out_root=out_root) 44 | else: 45 | return self._extract_without_dataloader(data_root=data_root, cat_info=self.cat_info, out_root=out_root) 46 | 47 | # extract the inputs' feature via self.model 48 | # the model's output contains both the inputs' feature and category info 49 | @t.no_grad() 50 | def _extract_without_dataloader(self, data_root, cat_info, out_root): 51 | feature = [] 52 | name = [] 53 | 54 | self.model.eval() 55 | 56 | cnames = sorted(os.listdir(data_root)) 57 | 58 | for cname in cnames: 59 | c_path = os.path.join(data_root, cname) 60 | if os.path.isdir(c_path): 61 | fnames = sorted(os.listdir(c_path)) 62 | for fname in fnames: 63 | path = os.path.join(c_path, fname) 64 | 65 | image = Image.open(path) 66 | image = self.transform(image) 67 | image = image[None] 68 | image = image.cuda() 69 | 70 | if self.vis: 71 | self.viser.images(image.cpu().numpy() * 0.5 + 0.5, win='extractor') 72 | out = self.model(image) 73 | if cat_info: 74 | i_feature = out[1] 75 | else: 76 | i_feature = out 77 | 78 | feature.append(i_feature.cpu().squeeze().numpy()) 79 | name.append(cname + '/' + fname) 80 | 81 | data = {'name': name, 'feature': feature} 82 | if out_root: 83 | out = open(out_root, 'wb') 84 | pickle.dump(data, out) 85 | 86 | out.close() 87 | 88 | return data 89 | 90 | # extract the inputs' feature via self.model 91 | # the model's output contains both the inputs' feature and category info 92 | # the input is loaded by dataloader 93 | @t.no_grad() 94 | def _extract_with_dataloader(self, data_root, cat_info, out_root): 95 | names = [] 96 | 97 | self.model.eval() 98 | 99 | opt = Config() 100 | opt.image_root = data_root 101 | opt.batch_size = 128 102 | 103 | dataloader = ImageDataLoader(opt) 104 | dataset = dataloader.load_data() 105 | 106 | for i, data in enumerate(dataset): 107 | image = data['I'].cuda() 108 | name = data['N'] 109 | 110 | out = self.model(image) 111 | if cat_info: 112 | i_feature = out[1] 113 | else: 114 | i_feature = out 115 | if i == 0: 116 | feature = i_feature.cpu().squeeze().numpy() 117 | 118 | else: 119 | feature = np.append(feature, i_feature.cpu().squeeze().numpy(), axis=0) 120 | 121 | names += name 122 | 123 | data = {'name': names, 'feature': feature} 124 | if out_root: 125 | out = open(out_root, 'wb') 126 | pickle.dump(data, out) 127 | 128 | out.close() 129 | 130 | return data 131 | 132 | # reload model with model file 133 | # the reloaded model contains fully connection layer 134 | def reload_state_dict_with_fc(self, state_file): 135 | temp_model = tv.models.resnet34(pretrained=False) 136 | temp_model.fc = nn.Linear(512, 125) 137 | temp_model.load_state_dict(t.load(state_file)) 138 | 139 | pretrained_dict = temp_model.state_dict() 140 | 141 | model_dict = self.model.state_dict() 142 | 143 | pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict} 144 | 145 | model_dict.update(pretrained_dict) 146 | self.model.load_state_dict(model_dict) 147 | 148 | # reload model with model file 149 | # the reloaded model doesn't contain fully connection layer 150 | def reload_state_dic(self, state_file): 151 | self.model.load_state_dict(t.load(state_file)) 152 | 153 | # reload model with model object directly 154 | def reload_model(self, model): 155 | self.model = model 156 | -------------------------------------------------------------------------------- /utils/feature_visualizer.py: -------------------------------------------------------------------------------- 1 | import visdom 2 | import pickle 3 | import os 4 | import numpy as np 5 | from sklearn.manifold import TSNE 6 | import matplotlib.pyplot as plt 7 | 8 | 9 | class FeatureVisualizer(object): 10 | 11 | def __init__(self, data_root, feature_root, env): 12 | self.data_root = data_root 13 | 14 | self.feature_root = feature_root 15 | self.feature_data = pickle.load(open(self.feature_root, 'rb')) 16 | 17 | self.vis = visdom.Visdom(env=env) 18 | 19 | self.c_index = dict() 20 | cnames = sorted(os.listdir(data_root)) 21 | for i, cname in enumerate(cnames): 22 | self.c_index[cname] = i 23 | 24 | cname_list, item_num = self._get_feature_info(data_root) 25 | self.cname_list = cname_list 26 | self.item_num = item_num 27 | 28 | # for (cname, num) in zip(cname_list, item_num): 29 | # print(cname, ':', num) 30 | 31 | # visualize the self.data_root 32 | def visualize(self, max_class, win='Feature Vis'): 33 | 34 | calculate_num = 0 35 | 36 | for i in range(max_class): 37 | 38 | calculate_num += self.item_num[i] 39 | 40 | feature = self.feature_data['feature'][:calculate_num] 41 | name = self.feature_data['name'][:calculate_num] 42 | 43 | feature_class = [] 44 | for f_name in name: 45 | class_name = f_name.split('/')[0] 46 | feature_class.append(self.c_index[class_name]) 47 | # print(np.shape(feature)) 48 | 49 | tsne = TSNE(n_components=2, init='pca', random_state=0) 50 | X_tsen = tsne.fit_transform(feature) 51 | 52 | # print(np.shape(X_tsen)) 53 | 54 | # print(set(feature_class)) 55 | plt.figure() 56 | plt.scatter(X_tsen[:, 0], X_tsen[:, 1], c=np.array(feature_class), marker='.', cmap=plt.cm.Spectral) 57 | plt.colorbar() 58 | plt.grid(True) 59 | plt.xlabel('1st') 60 | plt.ylabel('2nd') 61 | plt.title('Feature Vis') 62 | 63 | self.vis.matplot(plt, win=win, opts=dict(title=win)) 64 | 65 | # embedding vis 66 | def embedding_vis(self, em_data_root, em_feature_root, max_class, min_class=0, win='embedding vis'): 67 | 68 | cal_a = 0 69 | cal_b = 0 70 | 71 | em_cname_list, em_item_num = self._get_feature_info(em_data_root) 72 | 73 | em_feature_data = pickle.load(open(em_feature_root, 'rb')) 74 | 75 | start_a = 0 76 | start_b = 0 77 | for i in range(min_class): 78 | start_a += self.item_num[i] 79 | start_b += em_item_num[i] 80 | 81 | for i in range(min_class, max_class): 82 | 83 | cal_a += self.item_num[i] 84 | cal_b += em_item_num[i] 85 | 86 | a_feature = self.feature_data['feature'][start_a:cal_a+start_a] 87 | a_name = self.feature_data['name'][start_a:cal_a+start_a] 88 | print('np.shape(a_feature)', np.shape(a_feature)) 89 | 90 | b_feature = em_feature_data['feature'][start_b:cal_b+start_b] 91 | b_name = em_feature_data['name'][start_b:cal_b+start_b] 92 | print('np.shape(b_feature)', np.shape(b_feature)) 93 | 94 | a_class = [] 95 | for name in a_name: 96 | c_name = name.split('/')[0] 97 | a_class.append(self.c_index[c_name]) 98 | 99 | b_class = [] 100 | for name in b_name: 101 | c_name = name.split('/')[0] 102 | b_class.append(self.c_index[c_name]) 103 | 104 | tsne = TSNE(n_components=2, init='pca', random_state=0) 105 | feature = np.append(a_feature, b_feature, axis=0) 106 | 107 | print('np.shape(feature) :', np.shape(feature)) 108 | 109 | X_tsen = tsne.fit_transform(feature) 110 | 111 | X_a = X_tsen[:cal_a] 112 | X_b = X_tsen[cal_a:] 113 | 114 | print('np.shape(X_b) :', np.shape(X_b)) 115 | print('np.shape(b_class) :', np.shape(b_class)) 116 | print('a_class :', a_class) 117 | print('b_class :', b_class) 118 | 119 | plt.figure() 120 | plt.scatter(X_a[:, 0], X_a[:, 1], c=np.array(a_class), marker='.', cmap=plt.cm.Spectral) 121 | plt.scatter(X_b[:, 0], X_b[:, 1], c=np.array(b_class), marker='s', cmap=plt.cm.Spectral) 122 | 123 | plt.colorbar() 124 | plt.grid(True) 125 | plt.xlabel('1st') 126 | plt.ylabel('2nd') 127 | 128 | plt.title('Embedding Vis') 129 | self.vis.matplot(plt, win=win, opts=dict(title=win)) 130 | 131 | def _get_feature_info(self, data_root): 132 | item_num = [] 133 | cname_list = [] 134 | 135 | cnames = sorted(os.listdir(data_root)) 136 | for cname in cnames: 137 | num = len(os.listdir(os.path.join(data_root, cname))) 138 | 139 | item_num.append(num) 140 | cname_list.append(cname) 141 | 142 | return cname_list, item_num 143 | 144 | -------------------------------------------------------------------------------- /utils/test.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from sklearn.neighbors import NearestNeighbors 3 | from utils.extractor import Extractor 4 | from utils.visualize import Visualizer 5 | import torch as t 6 | 7 | 8 | class Tester(object): 9 | 10 | def __init__(self, opt): 11 | 12 | # self.vis = opt.vis 13 | self.test_bs = opt.test_bs 14 | 15 | self.photo_net = opt.photo_net 16 | self.sketch_net = opt.sketch_net 17 | 18 | self.photo_test = opt.photo_test 19 | self.sketch_test = opt.sketch_test 20 | 21 | self.eps = 1e-8 22 | 23 | @t.no_grad() 24 | def _extract_feature(self): 25 | with t.no_grad(): 26 | self.photo_net.eval() 27 | self.sketch_net.eval() 28 | 29 | extractor = Extractor(e_model=self.photo_net, vis=False, dataloader=True) 30 | photo_data = extractor.extract(self.photo_test) 31 | 32 | extractor.reload_model(self.sketch_net) 33 | sketch_data = extractor.extract(self.sketch_test) 34 | 35 | photo_name = photo_data['name'] 36 | photo_feature = photo_data['feature'] 37 | 38 | sketch_name = sketch_data['name'] 39 | sketch_feature = sketch_data['feature'] 40 | 41 | return photo_name, photo_feature, sketch_name, sketch_feature 42 | 43 | @t.no_grad() 44 | def _extract_feature_embedding(self): 45 | with t.no_grad(): 46 | self.photo_net.eval() 47 | self.sketch_net.eval() 48 | 49 | extractor = Extractor(e_model=self.photo_net, cat_info=False, vis=False, dataloader=True) 50 | photo_data = extractor.extract(self.photo_test, batch_size=self.test_bs) 51 | 52 | extractor.reload_model(self.sketch_net) 53 | sketch_data = extractor.extract(self.sketch_test, batch_size=self.test_bs) 54 | 55 | photo_name = photo_data['name'] 56 | photo_feature = photo_data['feature'] 57 | 58 | sketch_name = sketch_data['name'] 59 | sketch_feature = sketch_data['feature'] 60 | 61 | return photo_name, photo_feature, sketch_name, sketch_feature 62 | 63 | @t.no_grad() 64 | def test_category_recall(self): 65 | 66 | photo_name, photo_feature, sketch_name, sketch_feature = self._extract_feature() 67 | 68 | nbrs = NearestNeighbors(n_neighbors=np.size(photo_feature, 0), 69 | algorithm='brute', metric='euclidean').fit(photo_feature) 70 | 71 | count_1 = 0 72 | count_5 = 0 73 | K = 5 74 | for ii, (query_sketch, query_name) in enumerate(zip(sketch_feature, sketch_name)): 75 | query_sketch = np.reshape(query_sketch, [1, np.shape(query_sketch)[0]]) 76 | 77 | query_split = query_name.split('/') 78 | query_class = query_split[0] 79 | 80 | distance, indices = nbrs.kneighbors(query_sketch) 81 | 82 | for i, indice in enumerate(indices[0][:K]): 83 | 84 | retrieved_name = photo_name[indice] 85 | retrieved_class = retrieved_name.split('/')[0] 86 | 87 | if retrieved_class == query_class: 88 | 89 | if i == 0: 90 | count_1 += 1 91 | count_5 += 1 92 | break 93 | 94 | recall_1 = count_1 / (ii + 1) 95 | recall_5 = count_5 / (ii + 1) 96 | 97 | print('recall@1 :', recall_1, ' recall@5 :', recall_5) 98 | return {'recall@1': recall_1, 'recall@5': recall_5} 99 | 100 | @t.no_grad() 101 | def test_instance_recall(self): 102 | photo_name, photo_feature, sketch_name, sketch_feature = self._extract_feature() 103 | 104 | nbrs = NearestNeighbors(n_neighbors=np.size(photo_feature, 0), 105 | algorithm='brute', metric='euclidean').fit(photo_feature) 106 | 107 | count_1 = 0 108 | count_5 = 0 109 | K = 5 110 | for ii, (query_sketch, query_name) in enumerate(zip(sketch_feature, sketch_name)): 111 | query_sketch = np.reshape(query_sketch, [1, np.shape(query_sketch)[0]]) 112 | 113 | query_split = query_name.split('/') 114 | query_class = query_split[0] 115 | query_img = query_split[1] 116 | 117 | distance, indices = nbrs.kneighbors(query_sketch) 118 | 119 | for i, indice in enumerate(indices[0][:K]): 120 | 121 | retrieved_name = photo_name[indice] 122 | retrieved_class = retrieved_name.split('/')[0] 123 | 124 | retrieved_name = retrieved_name.split('/')[1] 125 | retrieved_name = retrieved_name.split('.')[0] 126 | if retrieved_class == query_class: 127 | if query_img.split('-')[0] == retrieved_name: 128 | if i == 0: 129 | count_1 += 1 130 | count_5 += 1 131 | break 132 | 133 | recall_1 = count_1 / (ii + 1) 134 | recall_5 = count_5 / (ii + 1) 135 | 136 | print('recall@1 :', recall_1, ' recall@5 :', recall_5) 137 | return {'recall@1': recall_1, 'recall@5': recall_5} 138 | 139 | 140 | 141 | 142 | 143 | 144 | 145 | 146 | 147 | 148 | 149 | -------------------------------------------------------------------------------- /utils/visualize.py: -------------------------------------------------------------------------------- 1 | # coding:utf8 2 | from itertools import chain 3 | import visdom 4 | import torch 5 | import time 6 | import torchvision as tv 7 | import numpy as np 8 | 9 | 10 | class Visualizer(): 11 | """ 12 | 封装了visdom的基本操作，但是你仍然可以通过`self.vis.function` 13 | 调用原生的visdom接口 14 | """ 15 | 16 | def __init__(self, env='default', **kwargs): 17 | import visdom 18 | self.vis = visdom.Visdom(env=env, use_incoming_socket=False,**kwargs) 19 | 20 | # 画的第几个数，相当于横座标 21 | # 保存（’loss',23） 即loss的第23个点 22 | self.index = {} 23 | self.log_text = '' 24 | 25 | def reinit(self, env='default', **kwargs): 26 | """ 27 | 修改visdom的配置 28 | """ 29 | self.vis = visdom.Visdom(env=env,use_incoming_socket=False, **kwargs) 30 | return self 31 | 32 | def plot_many(self, d): 33 | """ 34 | 一次plot多个 35 | @params d: dict (name,value) i.e. ('loss',0.11) 36 | """ 37 | for k, v in d.items(): 38 | self.plot(k, v) 39 | 40 | def img_many(self, d): 41 | for k, v in d.items(): 42 | self.img(k, v) 43 | 44 | def plot(self, name, y, legend=None): 45 | """ 46 | self.plot('loss',1.00) 47 | """ 48 | x = self.index.get(name, 0) 49 | self.vis.line(Y=np.array([y]), X=np.array([x]), 50 | win=(name), 51 | opts=dict(title=name) if legend is None else dict(title=name, legend=legend), 52 | update=None if x == 0 else 'append' 53 | ) 54 | self.index[name] = x + 1 55 | 56 | def img(self, name, img_): 57 | """ 58 | self.img('input_img',t.Tensor(64,64)) 59 | """ 60 | 61 | if len(img_.size()) < 3: 62 | img_ = img_.cpu().unsqueeze(0) 63 | self.vis.image(img_.cpu(), 64 | win=(name), 65 | opts=dict(title=name) 66 | ) 67 | 68 | def img_grid_many(self, d): 69 | for k, v in d.items(): 70 | self.img_grid(k, v) 71 | 72 | def img_grid(self, name, input_3d): 73 | """ 74 | 一个batch的图片转成一个网格图，i.e. input（36，64，64） 75 | 会变成 6*6 的网格图，每个格子大小64*64 76 | """ 77 | self.img(name, tv.utils.make_grid( 78 | input_3d.cpu()[0].unsqueeze(1).clamp(max=1, min=0))) 79 | 80 | def log(self, info, win='log_text'): 81 | """ 82 | self.log({'loss':1,'lr':0.0001}) 83 | """ 84 | 85 | self.log_text += ('[{time}] {info}
'.format( 86 | time=time.strftime('%m%d_%H%M%S'), 87 | info=info)) 88 | self.vis.text(self.log_text, win=win) 89 | 90 | def __getattr__(self, name): 91 | return getattr(self.vis, name) 92 | --------------------------------------------------------------------------------