├── .gitignore ├── LICENSE ├── README.md ├── README_CN.md ├── config.py ├── data └── rssrai_sense_cls │ └── ClsName2id.txt ├── examples ├── __init__.py ├── accuracy.py ├── multi_gpus_train.py └── submit.py ├── models ├── __init__.py ├── losses.py ├── network.py ├── resnet_utils.py └── run_net.py ├── prepare_data ├── __init__.py ├── gen_data_batch.py └── gen_tf_records.py └── tools └── img_encode.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | *.egg-info/ 24 | .installed.cfg 25 | *.egg 26 | MANIFEST 27 | 28 | # PyInstaller 29 | # Usually these files are written by a python script from a template 30 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 31 | *.manifest 32 | *.spec 33 | 34 | # Installer logs 35 | pip-log.txt 36 | pip-delete-this-directory.txt 37 | 38 | # Unit test / coverage reports 39 | htmlcov/ 40 | .tox/ 41 | .coverage 42 | .coverage.* 43 | .cache 44 | nosetests.xml 45 | coverage.xml 46 | *.cover 47 | .hypothesis/ 48 | .pytest_cache/ 49 | 50 | # Translations 51 | *.mo 52 | *.pot 53 | 54 | # Django stuff: 55 | *.log 56 | local_settings.py 57 | db.sqlite3 58 | 59 | # Flask stuff: 60 | instance/ 61 | .webassets-cache 62 | 63 | # Scrapy stuff: 64 | .scrapy 65 | 66 | # Sphinx documentation 67 | docs/_build/ 68 | 69 | # PyBuilder 70 | target/ 71 | 72 | # Jupyter Notebook 73 | .ipynb_checkpoints 74 | 75 | # pyenv 76 | .python-version 77 | 78 | # celery beat schedule file 79 | celerybeat-schedule 80 | 81 | # SageMath parsed files 82 | *.sage.py 83 | 84 | # Environments 85 | .env 86 | .venv 87 | env/ 88 | venv/ 89 | ENV/ 90 | env.bak/ 91 | venv.bak/ 92 | 93 | # Spyder project settings 94 | .spyderproject 95 | .spyproject 96 | 97 | # Rope project settings 98 | .ropeproject 99 | 100 | # mkdocs documentation 101 | /site 102 | 103 | # mypy 104 | .mypy_cache/ 105 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # sense_classification 2 | 3 | [中文版本](README_CN.md) 4 | 5 | ## Introduction 6 | [RSCUP: 遥感图像场景分类](http://rscup.bjxintong.com.cn/#/theme/1) 7 | 8 | 9 | 10 | This repo is organized as follows: 11 | 12 | ``` 13 | sense_classification/ 14 | |->examples 15 | |->models 16 | |->prepare_data 17 | |->data 18 | | |->rssrai_sense_cls 19 | | | |->train 20 | | | |->val 21 | | | |->test 22 | | |->tf_records 23 | | |->train_list 24 | |->ckpt 25 | |->tools 26 | ``` 27 | 28 | ## Requirements 29 | 1. tensorflow-gpu==1.12.0 (I only test on tensorflow 1.12.0) 30 | 2. python==3.4.3 31 | 3. numpy 32 | 4. easydict 33 | 5. opencv==3.4.1 34 | 6. Python packages might missing. pls fix it according to the error message. 35 | 36 | ## Installation, Prepare data, Training, Val, Generate submit 37 | ### Installation 38 | 1. Clone the sense_classification repository, and we'll call the directory that you cloned sense_classification as `${sense_classification_ROOT}`. 39 | 40 | ``` 41 | git clone https://github.com/vicwer/sense_classification.git 42 | ``` 43 | 44 | ### Prepare data 45 | data should be organized as follows: 46 | 47 | ``` 48 | data/ 49 | |->rssrai_sense_cls 50 | | |->train 51 | | |->val 52 | | |->test 53 | | |->ClsName2id.txt 54 | |->train_list/train.txt 55 | |->tf_records 56 | ``` 57 | 1. Download dataset and unzip: train.zip, val.zip, test.zip, ClsName2id.txt 58 | 59 | 2. Generate tf_records: 60 | 61 | ``` 62 | cd tools 63 | python3 img_encode.py 64 | ``` 65 | 66 | ### Training 67 | 68 | I provide common used config.py in ${sense_classification_ROOT}, which can set hyperparameters. 69 | 70 | e.g. 71 | ``` 72 | cd ${sense_classification_ROOT} 73 | vim config.py 74 | cfg.train.num_gpus = {your gpu nums} 75 | etc. 76 | 77 | cd ${sense_classification_ROOT}/examples/ 78 | python3 multi_gpus_train.py 79 | ``` 80 | 81 | ### Val 82 | 83 | ``` 84 | cd ${sense_classification_ROOT}/examples/ 85 | python3 accuracy.py 86 | ``` 87 | 88 | ### Generate submit 89 | 90 | ``` 91 | cd ${sense_classification_ROOT}/examples/ 92 | python3 submit.py 93 | ``` 94 | 95 | ## Result: 96 | 97 | Val: 0.908+ 98 | Test: 0.90509 99 | -------------------------------------------------------------------------------- /README_CN.md: -------------------------------------------------------------------------------- 1 | # 遥感图像场景分类 2 | 3 | [English Version](README.md) 4 | 5 | ## 简介 6 | [RSCUP: 遥感图像场景分类](http://rscup.bjxintong.com.cn/#/theme/1) 7 | 8 | 9 | 10 | 仓库路径应该组织成如下结构: 11 | 12 | ``` 13 | sense_classification/ 14 | |->examples 15 | |->models 16 | |->prepare_data 17 | |->data 18 | | |->rssrai_sense_cls 19 | | | |->train 20 | | | |->val 21 | | | |->test 22 | | |->tf_records 23 | | |->train_list 24 | |->ckpt 25 | |->tools 26 | ``` 27 | 28 | ## 环境依赖 29 | 1. tensorflow-gpu==1.12.0 (I only test on tensorflow 1.12.0) 30 | 2. python==3.4.3 31 | 3. numpy 32 | 4. easydict 33 | 5. opencv==3.4.1 34 | 6. 有些包可能没列出来,根据错误提示安装 35 | 36 | ## 安装, 准备数据, 训练, 验证, 生成提交文件 37 | ### 安装 38 | 1. 下载代码 39 | 40 | ``` 41 | git clone https://github.com/vicwer/sense_classification.git 42 | ``` 43 | 44 | ### 准备数据 45 | data目录结构: 46 | 47 | ``` 48 | data/ 49 | |->rssrai_sense_cls 50 | | |->train 51 | | |->val 52 | | |->test 53 | | |->ClsName2id.txt 54 | |->train_list/train.txt 55 | |->tf_records 56 | ``` 57 | 1. 下载数据集并解压: train.zip, val.zip, test.zip, ClsName2id.txt 58 | 59 | 2. 生成 tf_records: 60 | 61 | ``` 62 | cd tools 63 | python3 img_encode.py 64 | ``` 65 | 66 | ### 训练 67 | 68 | ${sense_classification_ROOT}目录提供了config.py, 可设置超参数 69 | 70 | 例如 71 | ``` 72 | cd ${sense_classification_ROOT} 73 | vim config.py 74 | cfg.train.num_gpus = {your gpu nums} 75 | etc. 76 | 77 | cd ${sense_classification_ROOT}/examples/ 78 | python3 multi_gpus_train.py 79 | ``` 80 | 81 | ### 验证 82 | 83 | ``` 84 | cd ${sense_classification_ROOT}/examples/ 85 | python3 accuracy.py 86 | ``` 87 | 88 | ### 生成提交文件 89 | 90 | ``` 91 | cd ${sense_classification_ROOT}/examples/ 92 | python3 submit.py 93 | ``` 94 | 95 | ## 结果: 96 | 97 | 验证集: 0.908+ 98 | 测试集：0.90509 99 | -------------------------------------------------------------------------------- /config.py: -------------------------------------------------------------------------------- 1 | from easydict import EasyDict as edict 2 | import numpy as np 3 | import tensorflow as tf 4 | import os 5 | 6 | cfg = edict() 7 | 8 | cfg.classes = 45 9 | cfg.batch_size = 128 10 | # cfg.batch_size = 256 11 | 12 | cfg.data_path = '../data/tf_records/train.records' 13 | cfg.ckpt_path = '../ckpt/' 14 | 15 | # training options 16 | cfg.train = edict() 17 | 18 | cfg.train.ignore_thresh = .5 19 | cfg.train.ohem_ratio = 0.8 20 | cfg.train.momentum = 0.9 21 | cfg.train.bn_training = True 22 | cfg.train.weight_decay = 0.0001 23 | cfg.train.learning_rate = [1e-3, 1e-4, 1e-5] 24 | cfg.train.max_batches = 30000 25 | cfg.train.lr_steps = [10000., 20000.] 26 | cfg.train.lr_scales = [.1, .1] 27 | cfg.train.num_gpus = 1 28 | cfg.train.tower = 'tower' 29 | 30 | cfg.train.learn_rate = 0.001 31 | cfg.train.learn_rate_decay = 0.8 32 | cfg.train.learn_rate_decay_epoch = 2 33 | cfg.train.num_samples = 177408 34 | cfg.epochs = 100 35 | cfg.PRINT_LAYER_LOG = True 36 | -------------------------------------------------------------------------------- /data/rssrai_sense_cls/ClsName2id.txt: -------------------------------------------------------------------------------- 1 | 旱地:dry-field:1 2 | 水田:paddy-field:2 3 | 梯田:terraced-field:3 4 | 草地:meadow:4 5 | 林地:forest:5 6 | 商业区:commercial-area:6 7 | 油田:oil-field:7 8 | 油罐区:storage-tank:8 9 | 工厂:works:9 10 | 矿区:mine:10 11 | 太阳能发电厂:solar-power-plant:11 12 | 风力发电站:wind-turbine:12 13 | 公园:park:13 14 | 游泳池:swimming-pool:14 15 | 教堂:church:15 16 | 墓地:cemetery:16 17 | 棒球场:baseball-field:17 18 | 篮球场:basketball-court:18 19 | 高尔夫球场:golf-course:19 20 | 足球场:soccer-field:20 21 | 温室:greenhouse:21 22 | 网球场:tennis-court:22 23 | 居民区:residential-area:23 24 | 岛屿:island:24 25 | 河流:river:25 26 | 停机坪:apron:26 27 | 直升机场:helipad:27 28 | 机场跑道:runway:28 29 | 桥梁:bridge:29 30 | 停车场:parking-lot:30 31 | 公路:road:31 32 | 路边停车区:roadside-parking-lot:32 33 | 转盘:roundabout:33 34 | 立交桥:viaduct:34 35 | 港口:port:35 36 | 铁路:railway:36 37 | 火车站:train-station:37 38 | 裸地:bare-land:38 39 | 沙漠:desert:39 40 | 冰岛:ice-land:40 41 | 山地:mountain:41 42 | 石质地:rock-land:42 43 | 稀疏灌木地:sparse-shrub-land:43 44 | 海滩:beach:44 45 | 湖泊:lake:45 46 | -------------------------------------------------------------------------------- /examples/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/vicwer/sense_classification/687ae9196d1d89da1d309f764f6993e1523c74c5/examples/__init__.py -------------------------------------------------------------------------------- /examples/accuracy.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # encoding: utf-8 3 | 4 | import numpy as np 5 | import tensorflow as tf 6 | import sys 7 | sys.path.append('..') 8 | from models.run_net import SenseClsNet 9 | from config import cfg 10 | import cv2 11 | import os 12 | from tqdm import tqdm 13 | import zipfile 14 | os.environ["CUDA_VISIBLE_DEVICES"] = "0" 15 | 16 | def accuracy(img_path, label_file, epoch): 17 | forders = os.listdir(img_path) 18 | labels_f = open(label_file, 'r') 19 | labels = labels_f.readlines() 20 | label_dict = dict() 21 | for l in labels: 22 | key_value = l.strip().split(':')[0::2] 23 | label_dict.update({key_value[0] : key_value[1]}) 24 | 25 | is_training = False 26 | cfg.batch_size = 1 27 | ckpt_dir = cfg.ckpt_path 28 | 29 | correct = 0 30 | wrong = 0 31 | all_image = 0 32 | 33 | configer = tf.ConfigProto() 34 | configer.gpu_options.per_process_gpu_memory_fraction = 0.1 35 | with tf.Session(config=configer) as sess: 36 | imgs_holder = tf.placeholder(tf.float32, shape=[1, 224, 224, 3]) 37 | model = SenseClsNet(imgs_holder, None, is_training) 38 | classes, scores = model.predict() 39 | 40 | saver = tf.train.Saver() 41 | sess.run(tf.global_variables_initializer()) 42 | 43 | saver.restore(sess, ckpt_dir + 'senceCls-' + str(epoch)) 44 | sess.run(tf.local_variables_initializer()) 45 | 46 | for f in tqdm(forders): 47 | label = float(label_dict[f]) 48 | imgs = os.listdir(os.path.join(img_path, f)) 49 | for img in imgs: 50 | path = os.path.join(os.path.join(img_path, f), img) 51 | img = cv2.imread(path) 52 | 53 | image = cv2.resize(img, (224, 224)) 54 | img_data = image.astype(np.float32) / 255.0 * 2.0 55 | 56 | all_image += 1 57 | classes_index, scores_0 = sess.run([classes, scores], feed_dict={imgs_holder: np.reshape(img_data, [1, 224, 224, 3])}) 58 | if classes_index[0] + 1 == label: 59 | correct += 1 60 | else: 61 | wrong += 1 62 | 63 | accuracy = float(correct) / float(correct + wrong) 64 | # print('global_step: ', g_step) 65 | print("All images:\n {}".format(int(correct + wrong))) 66 | print("Accuracy: {:.4f}".format(accuracy)) 67 | 68 | tf.reset_default_graph() 69 | 70 | if __name__ == '__main__': 71 | img_path = '../data/rssrai_sense_cls/train' 72 | label_file = '../data/rssrai_sense_cls/ClsName2id.txt' 73 | 74 | epoch = np.arange(12, 1, -1) 75 | print(epoch) 76 | for i in epoch: 77 | print('================{}================'.format(i)) 78 | accuracy(img_path, label_file, i) 79 | -------------------------------------------------------------------------------- /examples/multi_gpus_train.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # encoding: utf-8 3 | 4 | import numpy as np 5 | import tensorflow as tf 6 | import sys 7 | sys.path.append('..') 8 | from models.run_net import SenseClsNet 9 | from prepare_data.gen_data_batch import gen_data_batch 10 | from config import cfg 11 | import os 12 | import re 13 | import tensorflow.contrib.slim as slim 14 | 15 | gpu_list = np.arange(cfg.train.num_gpus) 16 | os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(str(i) for i in gpu_list) 17 | 18 | def get_variables_to_restore(include_vars=[], exclude_global_pool=False): 19 | variables_to_restore = [] 20 | for var in slim.get_model_variables(): 21 | if exclude_global_pool and 'global_pool' in var.op.name: 22 | #print(var) 23 | continue 24 | variables_to_restore.append(var) 25 | for var in slim.get_variables_to_restore(include=include_vars): 26 | if exclude_global_pool and 'global_pool' in var.op.name: 27 | #print(var) 28 | continue 29 | variables_to_restore.append(var) 30 | return variables_to_restore 31 | 32 | def average_gradients(tower_grads): 33 | """Calculate the average gradient for each shared variable across all towers. 34 | 35 | Note that this function provides a synchronization point across all towers. 36 | 37 | Args: 38 | tower_grads: List of lists of (gradient, variable) tuples. The outer list 39 | is over individual gradients. The inner list is over the gradient 40 | calculation for each tower. 41 | Returns: 42 | List of pairs of (gradient, variable) where the gradient has been averaged 43 | across all towers. 44 | """ 45 | average_grads = [] 46 | for grad_and_vars in zip(*tower_grads): 47 | # Note that each grad_and_vars looks like the following: 48 | # ((grad0_gpu0, var0_gpu0), ... , (grad0_gpuN, var0_gpuN)) 49 | grads = [] 50 | for g, _ in grad_and_vars: 51 | # Add 0 dimension to the gradients to represent the tower. 52 | expanded_g = tf.expand_dims(g, 0) 53 | 54 | # Append on a 'tower' dimension which we will average over below. 55 | grads.append(expanded_g) 56 | 57 | # Average over the 'tower' dimension. 58 | grad = tf.concat(axis=0, values=grads) 59 | grad = tf.reduce_mean(grad, 0) 60 | 61 | # Keep in mind that the Variables are redundant because they are shared 62 | # across towers. So .. we will just return the first tower's pointer to 63 | # the Variable. 64 | v = grad_and_vars[0][1] 65 | grad_and_var = (grad, v) 66 | average_grads.append(grad_and_var) 67 | return average_grads 68 | 69 | def train(finetune): 70 | is_training = True 71 | 72 | # data pipeline 73 | imgs, true_boxes = gen_data_batch(cfg.data_path, cfg.batch_size*cfg.train.num_gpus) 74 | imgs_split = tf.split(imgs, cfg.train.num_gpus) 75 | true_boxes_split = tf.split(true_boxes, cfg.train.num_gpus) 76 | 77 | global_step = tf.get_variable('global_step', [], initializer=tf.constant_initializer(0.), trainable=False) 78 | #lr = tf.train.piecewise_constant(global_step, cfg.train.lr_steps, cfg.train.learning_rate) 79 | #optimizer = tf.train.AdamOptimizer(learning_rate=lr) 80 | learn_rate_decay_step = int(cfg.train.num_samples / cfg.batch_size / cfg.train.num_gpus * cfg.train.learn_rate_decay_epoch) 81 | learning_rate = tf.train.exponential_decay(cfg.train.learn_rate, global_step, learn_rate_decay_step, cfg.train.learn_rate_decay, staircase=True) 82 | #optimizer = tf.train.GradientDescentOptimizer(learning_rate) 83 | optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate) 84 | 85 | # Calculate the gradients for each model tower. 86 | tower_grads = [] 87 | with tf.variable_scope(tf.get_variable_scope()): 88 | for i in range(cfg.train.num_gpus): 89 | with tf.device('/gpu:%d' % i): 90 | with tf.name_scope('%s_%d' % (cfg.train.tower, i)) as scope: 91 | model = SenseClsNet(imgs_split[i], true_boxes_split[i], is_training) 92 | loss = model.compute_loss() 93 | tf.get_variable_scope().reuse_variables() 94 | grads = optimizer.compute_gradients(loss) 95 | tower_grads.append(grads) 96 | if i == 0: 97 | current_loss = loss 98 | update_op = tf.get_collection(tf.GraphKeys.UPDATE_OPS) 99 | # print(tf.GraphKeys.UPDATE_OPS) 100 | # print(update_op) 101 | # print(grads) 102 | # vars_det = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="BioRecNet") 103 | grads = average_gradients(tower_grads) 104 | with tf.control_dependencies(update_op): 105 | apply_gradient_op = optimizer.apply_gradients(grads, global_step=global_step) 106 | train_op = tf.group(apply_gradient_op,*update_op) 107 | 108 | # GPU config 109 | config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=False) 110 | config.gpu_options.allow_growth = True 111 | sess = tf.Session(config=config) 112 | 113 | # Create a saver 114 | saver = tf.train.Saver(max_to_keep=1000) 115 | ckpt_dir = cfg.ckpt_path 116 | if not os.path.exists(ckpt_dir): 117 | os.makedirs(ckpt_dir) 118 | 119 | # init 120 | sess.run(tf.global_variables_initializer()) 121 | if finetune: 122 | checkpoint = './pre_train.ckpt' 123 | 124 | # variables_to_restore = slim.get_variables_to_restore() 125 | # init_assign_op, init_feed_dict = slim.assign_from_checkpoint(checkpoint, variables_to_restore, ignore_missing_vars=True) 126 | # sess.run(init_assign_op, init_feed_dict) 127 | 128 | variables_to_restore = get_variables_to_restore(exclude_global_pool=True) 129 | init_assign_op, init_feed_dict = slim.assign_from_checkpoint(checkpoint, variables_to_restore, ignore_missing_vars=True) 130 | sess.run(init_assign_op, init_feed_dict) 131 | 132 | # running 133 | cnt_epoch = 0 134 | 135 | for i in range(1, cfg.train.max_batches): 136 | _, loss_, lr_ = sess.run([train_op, current_loss, learning_rate]) 137 | if(i % 5 == 0): 138 | print(i,': ', loss_, ' lr: ', lr_) 139 | if int(i) % int(cfg.train.num_samples / cfg.train.num_gpus / cfg.batch_size) == 0: 140 | cnt_epoch += 1 141 | saver.save(sess, ckpt_dir+'senceCls', global_step=cnt_epoch, write_meta_graph=True) 142 | 143 | if __name__ == '__main__': 144 | train(finetune=False) 145 | -------------------------------------------------------------------------------- /examples/submit.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # encoding: utf-8 3 | 4 | import numpy as np 5 | import tensorflow as tf 6 | import sys 7 | sys.path.append('..') 8 | from models.run_net import SenseClsNet 9 | from config import cfg 10 | import cv2 11 | import os 12 | from tqdm import tqdm 13 | import zipfile 14 | os.environ["CUDA_VISIBLE_DEVICES"] = "0" 15 | 16 | def submit(img_path, submit_file, epoch): 17 | submit_f = open(submit_file ,'w') 18 | is_training = False 19 | cfg.batch_size = 1 20 | ckpt_dir = cfg.ckpt_path 21 | 22 | configer = tf.ConfigProto() 23 | configer.gpu_options.per_process_gpu_memory_fraction = 0.1 24 | with tf.Session(config=configer) as sess: 25 | imgs_holder = tf.placeholder(tf.float32, shape=[1, 224, 224, 3]) 26 | model = SenseClsNet(imgs_holder, None, is_training) 27 | classes, scores = model.predict() 28 | 29 | saver = tf.train.Saver() 30 | sess.run(tf.global_variables_initializer()) 31 | 32 | saver.restore(sess, ckpt_dir + 'senceCls-' + str(epoch)) 33 | sess.run(tf.local_variables_initializer()) 34 | 35 | for idx in tqdm(range(1, 89234)): 36 | img = cv2.imread(os.path.join(img_path, str(idx).zfill(5) + '.jpg')) 37 | image = cv2.resize(img, (224, 224)) 38 | img_data = image.astype(np.float32) / 255.0 * 2.0 39 | 40 | classes_index, scores_0 = sess.run([classes, scores], feed_dict={imgs_holder: np.reshape(img_data, [1, 224, 224, 3])}) 41 | submit_f.write(str(idx).zfill(5) + '.jpg' + ' ' + str(classes_index[0] + 1) + '\n') 42 | 43 | submit_f.close() 44 | zf = zipfile.ZipFile('./submit/classification.zip', 'w', zipfile.zlib.DEFLATED) 45 | zf.write(submit_file) 46 | zf.close() 47 | 48 | tf.reset_default_graph() 49 | 50 | if __name__ == '__main__': 51 | img_path = '../data/rssrai_sense_cls/test' 52 | if not os.path.exists('./submit'): 53 | os.makedirs('./submit') 54 | submit_file = './submit/classification.txt' 55 | 56 | epoch = [80] 57 | print(epoch) 58 | for i in epoch: 59 | submit(img_path, submit_file, i) 60 | -------------------------------------------------------------------------------- /models/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/vicwer/sense_classification/687ae9196d1d89da1d309f764f6993e1523c74c5/models/__init__.py -------------------------------------------------------------------------------- /models/losses.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # encoding: utf-8 3 | 4 | import tensorflow as tf 5 | import sys 6 | sys.path.append('..') 7 | from config import cfg 8 | 9 | PRINT_LAYER_LOG = cfg.PRINT_LAYER_LOG 10 | 11 | def loss(preds, labels): 12 | labels = tf.cast(labels, tf.int64) 13 | if PRINT_LAYER_LOG: 14 | print('pre labels', labels.get_shape()) 15 | labels = tf.reshape(labels, (cfg.batch_size, -1)) 16 | if PRINT_LAYER_LOG: 17 | print('labels', labels.get_shape()) 18 | labels = tf.one_hot(labels, cfg.classes) 19 | cross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=preds, labels=labels) 20 | pred_loss = tf.reduce_mean(cross_entropy) 21 | return pred_loss 22 | 23 | def loss_ohem(preds, labels): 24 | labels = tf.cast(labels, tf.int64) 25 | labels = tf.reshape(labels, (cfg.batch_size,)) 26 | print('pre labels', labels.get_shape()) 27 | labels = tf.one_hot(labels, cfg.classes) 28 | print('labels', labels.get_shape()) 29 | cross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=preds, labels=labels) 30 | print('cross_entropy', cross_entropy.get_shape()) 31 | keep_num = tf.cast(cfg.batch_size * cfg.train.ohem_ratio, tf.int32) 32 | cross_entropy = tf.reshape(cross_entropy, (cfg.batch_size,)) 33 | print('cross_entropy', cross_entropy.get_shape()) 34 | _, k_index = tf.nn.top_k(cross_entropy, keep_num) 35 | loss = tf.gather(cross_entropy, k_index) 36 | print('ohem loss', loss.get_shape()) 37 | 38 | return tf.reduce_mean(loss) 39 | -------------------------------------------------------------------------------- /models/network.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # encoding: utf-8 3 | 4 | import tensorflow as tf 5 | import tensorflow.contrib.slim as slim 6 | from tensorflow.python.framework import ops 7 | import sys 8 | sys.path.append('..') 9 | import numpy as np 10 | from config import cfg 11 | from models import resnet_utils 12 | from tensorflow.contrib import layers as layers_lib 13 | from tensorflow.contrib.framework.python.ops import add_arg_scope 14 | from tensorflow.contrib.framework.python.ops import arg_scope 15 | from tensorflow.contrib.layers.python.layers import initializers 16 | from tensorflow.contrib.layers.python.layers import layers 17 | from tensorflow.contrib.layers.python.layers import regularizers 18 | from tensorflow.contrib.layers.python.layers import utils 19 | from tensorflow.python.framework import ops 20 | from tensorflow.python.ops import array_ops 21 | from tensorflow.python.ops import nn_ops 22 | from tensorflow.python.ops import variable_scope 23 | 24 | PRINT_LAYER_LOG = cfg.PRINT_LAYER_LOG 25 | 26 | def network_arg_scope(is_training=True, 27 | weight_decay=cfg.train.weight_decay, 28 | batch_norm_decay=0.997, 29 | batch_norm_epsilon=1e-5, 30 | batch_norm_scale=True): 31 | batch_norm_params = { 32 | 'is_training': is_training, 'decay': batch_norm_decay, 33 | 'epsilon': batch_norm_epsilon, 'scale': batch_norm_scale, 34 | 'updates_collections': ops.GraphKeys.UPDATE_OPS, 35 | #'variables_collections': [ tf.GraphKeys.TRAINABLE_VARIABLES ], 36 | 'trainable': cfg.train.bn_training, 37 | } 38 | 39 | with slim.arg_scope( 40 | [slim.conv2d, slim.separable_convolution2d], 41 | weights_regularizer=slim.l2_regularizer(weight_decay), 42 | weights_initializer=slim.variance_scaling_initializer(), 43 | trainable=is_training, 44 | activation_fn=tf.nn.relu6, 45 | #activation_fn=tf.nn.relu, 46 | normalizer_fn=slim.batch_norm, 47 | normalizer_params=batch_norm_params, 48 | padding='SAME'): 49 | with slim.arg_scope([slim.batch_norm], **batch_norm_params) as arg_sc: 50 | return arg_sc 51 | 52 | class Network(object): 53 | def __init__(self): 54 | pass 55 | 56 | def inference(self, mode, inputs, scope='SenseCls'): 57 | is_training = mode 58 | with slim.arg_scope(network_arg_scope(is_training=is_training)): 59 | with tf.variable_scope(scope, reuse=False): 60 | conv0 = slim.conv2d(inputs, 61 | num_outputs=64, 62 | kernel_size=[7,7], 63 | stride=2, 64 | scope='conv0') 65 | if PRINT_LAYER_LOG: 66 | print(conv0.name, conv0.get_shape()) 67 | pool0 = slim.max_pool2d(conv0, kernel_size=[3, 3], stride=2, scope='pool0') 68 | if PRINT_LAYER_LOG: 69 | print('pool0', pool0.get_shape()) 70 | 71 | block0_0 = block(pool0, 64, 1, 'block0_0') 72 | block0_1 = block(block0_0, 64, 1, 'block0_1') 73 | block0_2 = block(block0_1, 64, 1, 'block0_2') 74 | 75 | block1_0 = block(block0_2, 128, 2, 'block1_0') 76 | block1_1 = block(block1_0, 128, 1, 'block1_1') 77 | block1_2 = block(block1_1, 128, 1, 'block1_2') 78 | block1_3 = block(block1_2, 128, 1, 'block1_3') 79 | 80 | block2_0 = block(block1_3, 256, 2, 'block2_0') 81 | block2_1 = block(block2_0, 256, 1, 'block2_1') 82 | block2_2 = block(block2_1, 256, 1, 'block2_2') 83 | block2_3 = block(block2_2, 256, 1, 'block2_3') 84 | block2_4 = block(block2_3, 256, 1, 'block2_4') 85 | block2_5 = block(block2_4, 256, 1, 'block2_5') 86 | 87 | block3_0 = block(block2_5, 512, 2, 'block3_0') 88 | block3_1 = block(block3_0, 512, 1, 'block3_1') 89 | block3_2 = block(block3_1, 512, 1, 'block3_2') 90 | 91 | net = tf.reduce_mean(block3_2, [1, 2], keepdims=True, name='global_pool_v4') 92 | if PRINT_LAYER_LOG: 93 | print('avg_pool', net.get_shape()) 94 | net = slim.flatten(net, scope='PreLogitsFlatten') 95 | net = slim.dropout(net, 0.8, is_training=is_training, scope='dropout') 96 | logits = fully_connected(net, cfg.classes, name='fc') 97 | if PRINT_LAYER_LOG: 98 | print('logits', logits.get_shape()) 99 | if is_training: 100 | l2_loss = tf.add_n(tf.losses.get_regularization_losses()) 101 | return logits, l2_loss 102 | else: 103 | return logits 104 | 105 | def block(inputs, c_outputs, s, name): 106 | se_module = True 107 | out1 = slim.conv2d(inputs, 108 | num_outputs=c_outputs, 109 | kernel_size=[3,3], 110 | stride=s, 111 | scope=name+'_0') 112 | if PRINT_LAYER_LOG: 113 | print(name+'_0', out1.get_shape()) 114 | output = slim.conv2d(out1, 115 | num_outputs=c_outputs, 116 | kernel_size=[3,3], 117 | stride=1, 118 | activation_fn=None, 119 | scope=name+'_1') 120 | if PRINT_LAYER_LOG: 121 | print(name+'_1', output.get_shape()) 122 | if s == 2: 123 | return nn_ops.relu(output) 124 | else: 125 | if se_module: 126 | squeeze = tf.reduce_mean(output, [1, 2], keepdims=True, name='global_pool_v4') 127 | if PRINT_LAYER_LOG: 128 | print('squeeze', squeeze.get_shape()) 129 | fc1 = slim.conv2d(squeeze, 130 | num_outputs=squeeze.get_shape()[-1] // 16, 131 | normalizer_fn=None, 132 | normalizer_params=None, 133 | weights_regularizer=None, 134 | kernel_size=[1,1], 135 | stride=1, 136 | activation_fn=tf.nn.relu, 137 | scope=name+'_fc1') 138 | if PRINT_LAYER_LOG: 139 | print('fc1', fc1.get_shape()) 140 | fc2 = slim.conv2d(fc1, 141 | num_outputs=squeeze.get_shape()[-1], 142 | normalizer_fn=None, 143 | normalizer_params=None, 144 | weights_regularizer=None, 145 | kernel_size=[1,1], 146 | stride=1, 147 | activation_fn=tf.sigmoid, 148 | scope=name+'_fc2') 149 | if PRINT_LAYER_LOG: 150 | print('fc2', fc2.get_shape()) 151 | output = output * fc2 152 | output = nn_ops.relu(inputs + output) 153 | if PRINT_LAYER_LOG: 154 | print(name, output.get_shape()) 155 | return output 156 | 157 | def dense_block(inputs, depth, depth_bottleneck, stride, name, rate=1): 158 | depth_in = inputs.get_shape()[3] 159 | if depth == depth_in: 160 | if stride == 1: 161 | shortcut = inputs 162 | else: 163 | shortcut = layers.max_pool2d(inputs, [1, 1], stride=factor, scope=name+'_shortcut') 164 | else: 165 | shortcut = layers.conv2d( 166 | inputs, 167 | depth, [1, 1], 168 | stride=stride, 169 | activation_fn=None, 170 | scope=name+'_shortcut') 171 | if PRINT_LAYER_LOG: 172 | print(name+'_shortcut', shortcut.get_shape()) 173 | 174 | residual = layers.conv2d( 175 | inputs, depth_bottleneck, [1, 1], stride=1, scope=name+'_conv1') 176 | if PRINT_LAYER_LOG: 177 | print(name+'_conv1', residual.get_shape()) 178 | residual = resnet_utils.conv2d_same( 179 | residual, depth_bottleneck, 3, stride, rate=rate, scope=name+'_conv2') 180 | if PRINT_LAYER_LOG: 181 | print(name+'_conv2', residual.get_shape()) 182 | residual = layers.conv2d( 183 | residual, depth, [1, 1], stride=1, activation_fn=None, scope=name+'_conv3') 184 | if PRINT_LAYER_LOG: 185 | print(name+'_conv3', residual.get_shape()) 186 | output = nn_ops.relu(shortcut + residual) 187 | return output 188 | 189 | def conv2d(inputs, c_outputs, s, name): 190 | output = slim.conv2d(inputs, num_outputs=c_outputs, kernel_size=[3,3], stride=s, scope=name) 191 | if PRINT_LAYER_LOG: 192 | print(name, output.get_shape()) 193 | return output 194 | 195 | def maxpool2x2(input, name): 196 | output = slim.max_pool2d(input, kernel_size=[2, 2], stride=2, scope=name) 197 | if PRINT_LAYER_LOG: 198 | print(name, output.get_shape()) 199 | return output 200 | 201 | def fully_connected(input, c_outputs, name): 202 | output = slim.fully_connected(input, c_outputs, activation_fn=None, scope=name) 203 | if PRINT_LAYER_LOG: 204 | print(name, output.get_shape()) 205 | return output 206 | 207 | def d_p_conv(inputs, c_outputs, s, name): 208 | output = slim.separable_convolution2d(inputs, 209 | num_outputs=None, 210 | stride=s, 211 | depth_multiplier=1, 212 | kernel_size=[3, 3], 213 | normalizer_fn=slim.batch_norm, 214 | scope=name+'_d_conv') 215 | if PRINT_LAYER_LOG: 216 | print(name, output.get_shape()) 217 | 218 | output = slim.conv2d(output, 219 | num_outputs=c_outputs, 220 | kernel_size=[1,1], 221 | stride=1, 222 | scope=name+'_p_conv') 223 | if PRINT_LAYER_LOG: 224 | print(name, output.get_shape()) 225 | return output 226 | 227 | def route(input_list, name): 228 | with tf.name_scope(name): 229 | output = tf.concat(input_list, 3, name='concat') 230 | if PRINT_LAYER_LOG: 231 | print(name, output.get_shape()) 232 | return output 233 | -------------------------------------------------------------------------------- /models/resnet_utils.py: -------------------------------------------------------------------------------- 1 | # Copyright 2016 The TensorFlow Authors. All Rights Reserved. 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | # ============================================================================== 15 | """Contains building blocks for various versions of Residual Networks. 16 | Residual networks (ResNets) were proposed in: 17 | Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun 18 | Deep Residual Learning for Image Recognition. arXiv:1512.03385, 2015 19 | More variants were introduced in: 20 | Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun 21 | Identity Mappings in Deep Residual Networks. arXiv: 1603.05027, 2016 22 | We can obtain different ResNet variants by changing the network depth, width, 23 | and form of residual unit. This module implements the infrastructure for 24 | building them. Concrete ResNet units and full ResNet networks are implemented in 25 | the accompanying resnet_v1.py and resnet_v2.py modules. 26 | Compared to https://github.com/KaimingHe/deep-residual-networks, in the current 27 | implementation we subsample the output activations in the last residual unit of 28 | each block, instead of subsampling the input activations in the first residual 29 | unit of each block. The two implementations give identical results but our 30 | implementation is more memory efficient. 31 | """ 32 | 33 | from __future__ import absolute_import 34 | from __future__ import division 35 | from __future__ import print_function 36 | 37 | import collections 38 | 39 | from tensorflow.contrib import layers as layers_lib 40 | from tensorflow.contrib.framework.python.ops import add_arg_scope 41 | from tensorflow.contrib.framework.python.ops import arg_scope 42 | from tensorflow.contrib.layers.python.layers import initializers 43 | from tensorflow.contrib.layers.python.layers import layers 44 | from tensorflow.contrib.layers.python.layers import regularizers 45 | from tensorflow.contrib.layers.python.layers import utils 46 | from tensorflow.python.framework import ops 47 | from tensorflow.python.ops import array_ops 48 | from tensorflow.python.ops import nn_ops 49 | from tensorflow.python.ops import variable_scope 50 | 51 | def subsample(inputs, factor, scope=None): 52 | """Subsamples the input along the spatial dimensions. 53 | Args: 54 | inputs: A `Tensor` of size [batch, height_in, width_in, channels]. 55 | factor: The subsampling factor. 56 | scope: Optional variable_scope. 57 | Returns: 58 | output: A `Tensor` of size [batch, height_out, width_out, channels] with the 59 | input, either intact (if factor == 1) or subsampled (if factor > 1). 60 | """ 61 | if factor == 1: 62 | return inputs 63 | else: 64 | return layers.max_pool2d(inputs, [1, 1], stride=factor, scope=scope) 65 | 66 | def conv2d_same(inputs, num_outputs, kernel_size, stride, rate=1, scope=None): 67 | """Strided 2-D convolution with 'SAME' padding. 68 | When stride > 1, then we do explicit zero-padding, followed by conv2d with 69 | 'VALID' padding. 70 | Note that 71 | net = conv2d_same(inputs, num_outputs, 3, stride=stride) 72 | is equivalent to 73 | net = tf.contrib.layers.conv2d(inputs, num_outputs, 3, stride=1, 74 | padding='SAME') 75 | net = subsample(net, factor=stride) 76 | whereas 77 | net = tf.contrib.layers.conv2d(inputs, num_outputs, 3, stride=stride, 78 | padding='SAME') 79 | is different when the input's height or width is even, which is why we add the 80 | current function. For more details, see ResnetUtilsTest.testConv2DSameEven(). 81 | Args: 82 | inputs: A 4-D tensor of size [batch, height_in, width_in, channels]. 83 | num_outputs: An integer, the number of output filters. 84 | kernel_size: An int with the kernel_size of the filters. 85 | stride: An integer, the output stride. 86 | rate: An integer, rate for atrous convolution. 87 | scope: Scope. 88 | Returns: 89 | output: A 4-D tensor of size [batch, height_out, width_out, channels] with 90 | the convolution output. 91 | """ 92 | if stride == 1: 93 | return layers_lib.conv2d( 94 | inputs, 95 | num_outputs, 96 | kernel_size, 97 | stride=1, 98 | rate=rate, 99 | padding='SAME', 100 | scope=scope) 101 | else: 102 | kernel_size_effective = kernel_size + (kernel_size - 1) * (rate - 1) 103 | pad_total = kernel_size_effective - 1 104 | pad_beg = pad_total // 2 105 | pad_end = pad_total - pad_beg 106 | inputs = array_ops.pad( 107 | inputs, [[0, 0], [pad_beg, pad_end], [pad_beg, pad_end], [0, 0]]) 108 | return layers_lib.conv2d( 109 | inputs, 110 | num_outputs, 111 | kernel_size, 112 | stride=stride, 113 | rate=rate, 114 | padding='VALID', 115 | scope=scope) 116 | -------------------------------------------------------------------------------- /models/run_net.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # encoding: utf-8 3 | 4 | import tensorflow as tf 5 | import sys 6 | sys.path.append('..') 7 | from models.network import Network 8 | from config import cfg 9 | from models.losses import loss, loss_ohem 10 | 11 | class SenseClsNet: 12 | def __init__(self, img, truth, is_training, batcn_norm_decay=0.997): 13 | self.img = img 14 | self.truth = truth 15 | self.is_training = is_training 16 | self.batch_norm_decay = batcn_norm_decay 17 | self.img_shape = tf.shape(self.img) 18 | backbone = Network() 19 | if is_training: 20 | self.head, self.l2_loss = backbone.inference(self.is_training, self.img) 21 | else: 22 | self.head = backbone.inference(self.is_training, self.img) 23 | 24 | def compute_loss(self): 25 | with tf.name_scope('loss_0'): 26 | cls_loss = loss(self.head, self.truth) 27 | self.all_loss = cls_loss + self.l2_loss 28 | return self.all_loss 29 | 30 | def predict(self): 31 | ''' 32 | only support single image prediction 33 | ''' 34 | pred_score = tf.reshape(self.head, (-1, cfg.classes)) 35 | score = tf.nn.softmax(tf.reshape(self.head, (-1, cfg.classes))) 36 | class_index = tf.argmax(pred_score, 1) 37 | return class_index, score 38 | -------------------------------------------------------------------------------- /prepare_data/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/vicwer/sense_classification/687ae9196d1d89da1d309f764f6993e1523c74c5/prepare_data/__init__.py -------------------------------------------------------------------------------- /prepare_data/gen_data_batch.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # encoding: utf-8 3 | 4 | import tensorflow as tf 5 | import numpy as np 6 | import sys 7 | sys.path.append('..') 8 | from config import cfg 9 | import os 10 | import re 11 | import cv2 12 | import math 13 | 14 | def distorted_bounding_box_crop(image, 15 | bbox, 16 | min_object_covered=0.1, 17 | aspect_ratio_range=(0.75, 1.33), 18 | area_range=(0.9, 1.0), 19 | max_attempts=100, 20 | scope=None): 21 | 22 | with tf.name_scope(scope, 'distort_image', [image, image.shape[0], image.shape[1], bbox]): 23 | if bbox is None: 24 | bbox = tf.constant([0.0, 0.0, 1.0, 1.0], 25 | dtype=tf.float32, 26 | shape=[1, 1, 4]) 27 | if image.dtype != tf.float32: 28 | img = tf.image.convert_image_dtype(image, dtype=tf.float32) 29 | 30 | with tf.name_scope(scope, 'distorted_bounding_box_crop', [image, bbox]): 31 | sample_distorted_bounding_box = tf.image.sample_distorted_bounding_box( 32 | tf.shape(image), 33 | bounding_boxes=bbox, 34 | min_object_covered=min_object_covered, 35 | aspect_ratio_range=aspect_ratio_range, 36 | area_range=area_range, 37 | max_attempts=max_attempts, 38 | use_image_if_no_bounding_boxes=True) 39 | bbox_begin, bbox_size, distort_bbox = sample_distorted_bounding_box 40 | 41 | # Crop the image to the specified bounding box. 42 | cropped_image = tf.slice(image, bbox_begin, bbox_size) 43 | return cropped_image, distort_bbox 44 | 45 | def parser(example): 46 | feats = tf.parse_single_example(example, features={'label' : tf.FixedLenFeature([1], tf.float32), 47 | 'feature': tf.FixedLenFeature([], tf.string)}) 48 | coord = feats['label'] 49 | 50 | img = tf.decode_raw(feats['feature'], tf.uint8) 51 | img = tf.reshape(img, [256, 256, 3]) 52 | img = tf.contrib.image.rotate(img, tf.random_uniform([], minval=-10*math.pi/180., maxval=10*math.pi/180.)) 53 | 54 | rot_img = tf.image.rot90(img) 55 | rot_seed = tf.random_uniform([], maxval=1.0) 56 | img = tf.cond(rot_seed > 0.5, lambda: img, lambda: rot_img) 57 | img = tf.image.random_flip_left_right(img) 58 | img = tf.image.random_flip_up_down(img) 59 | 60 | crop_img = tf.random_crop(img, [224, 224, 3]) 61 | img = tf.image.resize_images(img, [224, 224]) 62 | img = tf.cast(img, tf.uint8) 63 | # img, _ = distorted_bounding_box_crop(img, bbox=None) 64 | # crop_img = tf.image.resize_images(crop_img, [224, 224]) 65 | 66 | rand_seed = tf.random_uniform([], maxval=1.0) 67 | img = tf.cond(rand_seed > 0.5, lambda: img, lambda: crop_img) 68 | 69 | img = tf.cast(img, tf.float32) / 255.0 70 | img = tf.image.random_hue(img, max_delta=0.1) 71 | img = tf.image.random_contrast(img, lower=0.9, upper=1.1) 72 | img = tf.image.random_brightness(img, max_delta=0.05) 73 | img = tf.image.random_saturation(img, lower=0.8, upper=1.2) 74 | img = tf.clip_by_value(img, 0.0, 1.0) 75 | img = tf.multiply(img, 2.0) 76 | return img, coord 77 | 78 | def gen_data_batch(tf_records_filename, batch_size): 79 | dt = tf.data.TFRecordDataset(tf_records_filename) 80 | dt = dt.map(parser, num_parallel_calls=4) 81 | dt = dt.prefetch(batch_size) 82 | dt = dt.shuffle(buffer_size=8*batch_size) 83 | dt = dt.repeat() 84 | dt = dt.batch(batch_size) 85 | iterator = dt.make_one_shot_iterator() 86 | imgs, true_boxes = iterator.get_next() 87 | 88 | return imgs, true_boxes 89 | 90 | if __name__ == '__main__': 91 | os.environ["CUDA_VISIBLE_DEVICES"] = "2,3" 92 | tf_records_filename = cfg.data_path 93 | 94 | imgs, true_boxes = gen_data_batch(tf_records_filename, cfg.batch_size) 95 | imgs_split = tf.split(imgs, cfg.train.num_gpus) 96 | true_boxes_split = tf.split(true_boxes, cfg.train.num_gpus) 97 | configer = tf.ConfigProto() 98 | configer.gpu_options.per_process_gpu_memory_fraction = 0.3 99 | sess=tf.Session(config=configer) 100 | for i in range(2): 101 | for j in range(cfg.train.num_gpus): 102 | imgs_, true_boxes_ = sess.run([imgs_split[j], true_boxes_split[j]]) 103 | print(true_boxes_.shape) 104 | 105 | for k in range(imgs_.shape[0]): 106 | cv2.imshow('img', imgs_[k].astype(np.uint8)) 107 | cv2.waitKey(0) 108 | 109 | -------------------------------------------------------------------------------- /prepare_data/gen_tf_records.py: -------------------------------------------------------------------------------- 1 | # encoding: utf-8 2 | 3 | import numpy as np 4 | import tensorflow as tf 5 | import os 6 | import cv2 7 | from tqdm import tqdm 8 | import re 9 | import sys 10 | sys.path.append('..') 11 | from config import cfg 12 | 13 | def load_file(file_path): 14 | ''' 15 | load imgs_path, classes and labels 16 | ''' 17 | imgs_path = [] 18 | labels = [] 19 | with open(file_path, 'r') as f: 20 | lines = f.readlines() 21 | for line in lines: 22 | img_path = line.strip().split(' ')[0] 23 | label = float(line.strip().split(' ')[1]) - 1 24 | imgs_path.append(img_path) 25 | labels.append(label) 26 | return np.asarray(imgs_path), np.asarray(labels) 27 | 28 | def extract_image(image_path, height, width, is_resize=True): 29 | ''' 30 | get b->g->r image data 31 | ''' 32 | img = cv2.imread(image_path) 33 | if is_resize: 34 | h, w, _ = img.shape 35 | if h == height and w == width: 36 | image = img 37 | else: 38 | image = cv2.resize(img, (height, height)) 39 | # cv2.imshow("img", image) 40 | # cv2.waitKey(0) 41 | else: 42 | image = img 43 | # cv2.imshow('img', image) 44 | # cv2.waitKey(0) 45 | image_data = np.array(image, dtype='uint8') 46 | return image_data 47 | 48 | def run_encode(file_path, tf_records_filename): 49 | ''' 50 | encode func 51 | ''' 52 | print('generate records...') 53 | imgs_path, labels = load_file(file_path) 54 | height, width = 256, 256 55 | imgs = [] 56 | writer = tf.python_io.TFRecordWriter(tf_records_filename) 57 | for i in tqdm(range(imgs_path.shape[0])): 58 | img = extract_image(imgs_path[i], height, width, is_resize=True) 59 | img = img.tostring() 60 | label = labels[i].flatten().tolist() 61 | example = tf.train.Example(features=tf.train.Features(feature={ 62 | 'label' : tf.train.Feature(float_list = tf.train.FloatList(value=label)), 63 | 'feature': tf.train.Feature(bytes_list = tf.train.BytesList(value=[img])) 64 | })) 65 | writer.write(example.SerializeToString()) 66 | writer.close() 67 | 68 | if __name__ == '__main__': 69 | file_path = './train.txt' 70 | tf_records_filename = '/train.records' 71 | 72 | run_encode(file_path, tf_records_filename) 73 | -------------------------------------------------------------------------------- /tools/img_encode.py: -------------------------------------------------------------------------------- 1 | #-*- coding:utf-8 -*- 2 | 3 | import sys 4 | import os 5 | import numpy as np 6 | import numpy.random as npr 7 | sys.path.append('..') 8 | from prepare_data.gen_tf_records import run_encode 9 | import cv2 10 | from tqdm import tqdm 11 | 12 | def gen_img_list(img_path, label_file, img_list_path): 13 | forders = os.listdir(img_path) 14 | print(forders) 15 | labels_f = open(label_file, 'r') 16 | labels = labels_f.readlines() 17 | img_f = open(os.path.join(img_list_path, 'train.txt'), 'w') 18 | 19 | label_dict = dict() 20 | for l in labels: 21 | key_value = l.strip().split(':')[0::2] 22 | label_dict.update({key_value[0] : key_value[1]}) 23 | print(label_dict) 24 | 25 | print('remove dead img') 26 | cnt_path = os.getcwd() 27 | for f in tqdm(forders): 28 | label = label_dict[f] 29 | imgs = os.listdir(os.path.join(img_path, f)) 30 | for img in imgs: 31 | path = os.path.join(cnt_path, os.path.join(os.path.join(img_path, f), img)) 32 | img = cv2.imread(path) 33 | try: 34 | h, w, _ = img.shape 35 | img_f.write(path + ' ' + label + '\n') 36 | except: 37 | print('dead img: {}'.format(path)) 38 | img_f.close() 39 | 40 | def shuffle_list(img_list_path): 41 | with open(os.path.join(img_list_path, 'train.txt'), 'r') as f: 42 | lines = f.readlines() 43 | 44 | with open(os.path.join(img_list_path, 'train.txt'), "w") as f: 45 | num = len(lines) 46 | lines_keep = npr.choice(len(lines), size=int(num),replace=False) 47 | 48 | for i in lines_keep: 49 | f.write(lines[i]) 50 | 51 | def encode(img_path, label_file, img_list_path, records_path): 52 | for dir_path in [img_list_path, records_path]: 53 | if not os.path.exists(dir_path): 54 | os.makedirs(dir_path) 55 | 56 | gen_img_list(img_path, label_file, img_list_path) 57 | for i in range(10): 58 | shuffle_list(img_list_path) 59 | 60 | records_f = os.path.join(records_path, 'train.records') 61 | run_encode(os.path.join(img_list_path, 'train.txt'), records_f) 62 | 63 | if __name__ == '__main__': 64 | img_path = '../data/rssrai_sense_cls/train' 65 | label_file = '../data/rssrai_sense_cls/ClsName2id.txt' 66 | img_list_path = '../data/train_list/' 67 | records_path = '../data/tf_records' 68 | encode(img_path, label_file, img_list_path, records_path) 69 | --------------------------------------------------------------------------------