├── .gitignore ├── README.md ├── data ├── KTHDataset │ └── KTHDateset └── mnist │ └── mnist ├── dataloader ├── KTHDataset │ ├── KTHDataset.py │ └── MakeDataset.py ├── MovingMNIST │ ├── MovingMNIST.py │ └── demo.py └── README.md ├── experiments ├── PredRNN │ └── custom.py ├── README.md └── video_prediction_demo │ └── custom.py ├── logs ├── PredRNN.log └── video_prediction_demo.log ├── model ├── README.md ├── STconvLSTM.py ├── convRNN.py └── loss │ ├── L1_L2_Loss.py │ └── SSIM_Loss.py ├── requirements.txt ├── tools ├── README.md ├── train_config.json ├── train_config_PredRNN.json ├── train_video_prediction_KTHDataset.py └── train_video_prediction_MNIST.py └── utils ├── README.md ├── __init__.py ├── average_meter_helper.py ├── config_helper.py ├── log_helper.py └── memory ├── README.md ├── gpu_mem_track.py └── modelsize_estimate.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 | local_settings.py 56 | db.sqlite3 57 | 58 | # Flask stuff: 59 | instance/ 60 | .webassets-cache 61 | 62 | # Scrapy stuff: 63 | .scrapy 64 | 65 | # Sphinx documentation 66 | docs/_build/ 67 | 68 | # PyBuilder 69 | target/ 70 | 71 | # Jupyter Notebook 72 | .ipynb_checkpoints 73 | 74 | # pyenv 75 | .python-version 76 | 77 | # celery beat schedule file 78 | celerybeat-schedule 79 | 80 | # SageMath parsed files 81 | *.sage.py 82 | 83 | # Environments 84 | .env 85 | .venv 86 | env/ 87 | venv/ 88 | ENV/ 89 | env.bak/ 90 | venv.bak/ 91 | 92 | # Spyder project settings 93 | .spyderproject 94 | .spyproject 95 | 96 | # Rope project settings 97 | .ropeproject 98 | 99 | # mkdocs documentation 100 | /site 101 | 102 | # mypy 103 | .mypy_cache/ 104 | 105 | .DS_Store 106 | data/mnist/raw 107 | data/mnist/processed 108 | data/KTHDataset/raw 109 | data/KTHDataset/processed 110 | .vscode/settings.json 111 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Frame_Video_Prediction_Pytorch 2 | frame video prediction algorithm framework used pytorch. 3 | This project is supposed to be a example of dl to help my frined finish their surf project. 4 | Hope this can help you too~~ 5 | 6 | # How to start 7 | Walkthrough the folders ReadMe.md file for a quick look. 8 | 9 | for the first time, you may want to build up a conda environment use: 10 | 11 | `conda create -n torchenv python=3.6` 12 | 13 | `source activate torchenv` 14 | 15 | `pip install -r requirements.txt` 16 | 17 | Then run our demo derectly: 18 | 19 | `python3 ./tools/train_video_prediction.py` 20 | 21 | The data will be download to ./data and the result will be stored in ./board using tensorboardX and ./logs. 22 | 23 | You can use: 24 | 25 | `tensorboard --logdir ./board/video_prediction_demo/` 26 | 27 | to see the results. 28 | 29 | PS: we assume our running root path is current level. 30 | -------------------------------------------------------------------------------- /data/KTHDataset/KTHDateset: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JaMesLiMers/Frame_Video_Prediction_Pytorch/fec5870a2d2ce8f91085c38c46f86e3b58ee0385/data/KTHDataset/KTHDateset -------------------------------------------------------------------------------- /data/mnist/mnist: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JaMesLiMers/Frame_Video_Prediction_Pytorch/fec5870a2d2ce8f91085c38c46f86e3b58ee0385/data/mnist/mnist -------------------------------------------------------------------------------- /dataloader/KTHDataset/KTHDataset.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | import torch.utils.data as data 3 | from PIL import Image 4 | import sys 5 | import os 6 | import os.path 7 | import errno 8 | import numpy as np 9 | import torch 10 | import codecs 11 | import joblib 12 | from torchvision import transforms 13 | 14 | 15 | parent_path = os.path.dirname(os.path.dirname(os.getcwd())) 16 | if 'KTHDataset' in os.getcwd(): 17 | os.chdir(parent_path) 18 | sys.path.insert(0, os.getcwd()) 19 | 20 | 21 | from dataloader.KTHDataset.MakeDataset import make_data 22 | 23 | 24 | class KTHDataset(data.Dataset): 25 | """`KTHDataset `_ Dataset. 26 | 27 | Args: 28 | root (string): Root directory of dataset where ``processed/training.pt`` 29 | ``processed/validate.pt`` and ``processed/test.pt`` exist. 30 | train (bool or string, optional): If True, creates dataset from ``training.pt``, 31 | otherwise from ``test.pt``, If string in 'test/train/validate' then create 32 | according dataset. 33 | data_length (int or None): number of data per epoch 34 | download (bool, optional): If true, downloads the dataset from the internet and 35 | puts it in root directory. If dataset is already downloaded, it is not 36 | downloaded again. 37 | transform (callable, optional): A function/transform that takes in an PIL image 38 | and returns a transformed version. E.g, ``transforms.RandomCrop`` 39 | target_transform (callable, optional): A function/transform that takes in an PIL 40 | image and returns a transformed version. E.g, ``transforms.RandomCrop`` 41 | """ 42 | urls = [ 43 | 'http://www.nada.kth.se/cvap/actions/boxing.zip', 44 | 'http://www.nada.kth.se/cvap/actions/handclapping.zip', 45 | 'http://www.nada.kth.se/cvap/actions/handwaving.zip', 46 | 'http://www.nada.kth.se/cvap/actions/jogging.zip', 47 | 'http://www.nada.kth.se/cvap/actions/running.zip', 48 | 'http://www.nada.kth.se/cvap/actions/walking.zip' 49 | ] 50 | sequence_url = 'http://www.nada.kth.se/cvap/actions/00sequences.txt' 51 | raw_folder = 'raw' 52 | processed_folder = 'processed' 53 | training_file = 'train.pkl' 54 | test_file = 'test.pkl' 55 | validate_file = 'validate.pkl' 56 | sequence_name = '00sequences.txt' 57 | 58 | def __init__(self, root, train=True, transform=None, target_transform=None, download=False, data_length=None): 59 | self.root = os.path.expanduser(root) 60 | self.transform = transform 61 | self.target_transform = target_transform 62 | self.train = train # training set or test set 63 | self.data_length = data_length 64 | 65 | if download: 66 | self.download() 67 | 68 | if not self._check_exists(): 69 | raise RuntimeError('Dataset not found.' + 70 | ' You can use download=True to download it') 71 | 72 | if self.train: 73 | self.data = joblib.load( 74 | os.path.join(self.root, self.processed_folder, self.training_file)) 75 | self.data += joblib.load( 76 | os.path.join(self.root, self.processed_folder, self.validate_file)) 77 | elif not self.train: 78 | self.data = joblib.load( 79 | os.path.join(self.root, self.processed_folder, self.test_file)) 80 | elif self.train is 'train': 81 | self.data = joblib.load( 82 | os.path.join(self.root, self.processed_folder, self.training_file)) 83 | elif self.train is 'test': 84 | self.data = joblib.load( 85 | os.path.join(self.root, self.processed_folder, self.test_file)) 86 | elif self.train is 'validate': 87 | self.data = joblib.load( 88 | os.path.join(self.root, self.processed_folder, self.validate_file)) 89 | else: 90 | raise NotImplementedError("invalied string input for train, need 'train','test' or 'validate'") 91 | 92 | # self.test_data() 93 | 94 | # init index list 95 | # index_list = [] 96 | # while len(index_list) <= self.__len__(): 97 | # index_list += list(range(len(self.data))) 98 | # self.index_list = index_list[0:self.__len__()] 99 | # or random 100 | self.index_list = np.random.randint(low=0, high=len(self.data), size=self.__len__()) 101 | 102 | def test_data(self): 103 | for i in self.data: 104 | sequence = i['sequence'] 105 | for j in range(len(sequence)): 106 | if j%2 == 0: 107 | if sequence[j+1] - sequence[j] <= 20: 108 | print(i["filename"] + 'error') 109 | 110 | def __len__(self): 111 | if self.data_length is not int: 112 | length = 0 113 | for i in self.data: 114 | length += i['sequence'][-1] 115 | self.data_length = length 116 | return self.data_length 117 | else: 118 | return self.data_length 119 | 120 | def __getitem__(self, index): 121 | """ 122 | Args: 123 | index (int): Index 124 | 125 | Returns: 126 | tuple: (seq, target) where sampled sequences are splitted into a seq 127 | and target part 128 | """ 129 | 130 | sequence = self.data[self.index_list[index]]["sequence"] 131 | choice = np.random.randint(low=0, high=len(sequence)//2)*2 132 | frames = np.random.randint(low=sequence[choice]-1, high=sequence[choice+1] - 20 -1) 133 | train_frames = self.data[self.index_list[index]]["frames"][frames:frames+10] 134 | gt_frames = self.data[self.index_list[index]]["frames"][frames+10:frames+20] 135 | 136 | 137 | train_frames = [Image.fromarray(train_frames[i], mode='L') for i in range(10)] 138 | gt_frames = [Image.fromarray(gt_frames[i], mode='L') for i in range(10)] 139 | 140 | if self.transform is not None: 141 | train_frames = torch.stack([self.transform(train_frames[i]) for i in range(10)]) 142 | 143 | if self.target_transform is not None: 144 | gt_frames = torch.stack([self.target_transform(gt_frames[i]) for i in range(10)]) 145 | 146 | return train_frames, gt_frames 147 | 148 | def _check_exists(self): 149 | return os.path.exists(os.path.join(self.root, self.processed_folder, self.training_file)) and \ 150 | os.path.exists(os.path.join(self.root, self.processed_folder, self.test_file)) and \ 151 | os.path.exists(os.path.join(self.root, self.processed_folder, self.validate_file)) 152 | 153 | 154 | def download(self): 155 | """Download the KTH data if it doesn't exist in processed_folder already.""" 156 | from six.moves import urllib 157 | import zipfile 158 | 159 | if self._check_exists(): 160 | return 161 | 162 | # download files 163 | try: 164 | os.makedirs(os.path.join(self.root, self.raw_folder)) 165 | os.makedirs(os.path.join(self.root, self.processed_folder)) 166 | except OSError as e: 167 | if e.errno == errno.EEXIST: 168 | pass 169 | else: 170 | raise 171 | 172 | for url in self.urls: 173 | print('Downloading ' + url) 174 | data = urllib.request.urlopen(url) 175 | filename = url.rpartition('/')[2] 176 | file_dir = os.path.join(self.root, self.raw_folder) 177 | file_path = os.path.join(self.root, self.raw_folder, filename) 178 | with open(file_path, 'wb') as f: 179 | f.write(data.read()) 180 | 181 | try: 182 | os.makedirs(file_path.replace('.zip', '')) 183 | except OSError as e: 184 | if e.errno == errno.EEXIST: 185 | pass 186 | else: 187 | raise 188 | 189 | with zipfile.ZipFile(file_path) as zip_f: 190 | for fileM in zip_f.namelist(): 191 | zip_f.extract(fileM, file_path.replace('.zip', '')) 192 | os.unlink(file_path) 193 | 194 | print('downloading sequence file...') 195 | data = urllib.request.urlopen(self.sequence_url) 196 | sequence_name = self.sequence_url.rpartition('/')[2] 197 | file_dir = os.path.join(self.root, self.raw_folder) 198 | file_path = os.path.join(self.root, self.raw_folder, sequence_name) 199 | with open(file_path, 'wb') as f: 200 | f.write(data.read()) 201 | 202 | # process and save as torch files 203 | print('Processing...') 204 | train_data, test_data, validate_data = make_data(self.root, self.raw_folder, self.processed_folder, self.sequence_name) 205 | print(len(train_data)) 206 | print(len(test_data)) 207 | print(len(train_data)) 208 | 209 | # Dump data to file. 210 | print('Dumping train data to file...') 211 | with open(os.path.join(self.root, self.processed_folder, self.training_file), 'wb') as f: 212 | joblib.dump(train_data, f) 213 | print('Dumping test data to file...') 214 | with open(os.path.join(self.root, self.processed_folder, self.test_file), 'wb') as f: 215 | joblib.dump(test_data, f) 216 | print('Dumping validate data to file...') 217 | with open(os.path.join(self.root, self.processed_folder, self.validate_file), 'wb') as f: 218 | joblib.dump(validate_data, f) 219 | 220 | print('Done!') 221 | 222 | def __repr__(self): 223 | fmt_str = 'Dataset ' + self.__class__.__name__ + '\n' 224 | fmt_str += ' Number of datapoints: {}\n'.format(self.__len__()) 225 | fmt_str += ' Number of data: {}\n'.format(len(self.data)) 226 | tmp = 'train' if self.train is True else 'test' 227 | fmt_str += ' Train/test: {}\n'.format(tmp) 228 | fmt_str += ' Root Location: {}\n'.format(self.root) 229 | tmp = ' Transforms (if any): ' 230 | fmt_str += '{0}{1}\n'.format(tmp, self.transform.__repr__().replace('\n', '\n' + ' ' * len(tmp))) 231 | tmp = ' Target Transforms (if any): ' 232 | fmt_str += '{0}{1}'.format(tmp, self.target_transform.__repr__().replace('\n', '\n' + ' ' * len(tmp))) 233 | return fmt_str 234 | 235 | if __name__ == "__main__": 236 | a = KTHDataset('./data/KTHDataset', download=True) 237 | -------------------------------------------------------------------------------- /dataloader/KTHDataset/MakeDataset.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import os 3 | import joblib 4 | import re 5 | import cv2 6 | 7 | CATEGORIES = ["boxing", "handclapping", "handwaving", "jogging", "running", 8 | "walking"] 9 | 10 | # Dataset are divided according to the instruction at: 11 | # http://www.nada.kth.se/cvap/actions/00sequences.txt 12 | TRAIN_PEOPLE_ID = [11, 12, 13, 14, 15, 16, 17, 18] 13 | DEV_PEOPLE_ID = [19, 20, 21, 23, 24, 25, 1, 4] 14 | TEST_PEOPLE_ID = [22, 2, 3, 5, 6, 7, 8, 9, 10] 15 | 16 | def make_data(data_path, raw_folder, processed_folder, sequence_name='00sequences.txt'): 17 | 18 | train = [] 19 | dev = [] 20 | test = [] 21 | sequence = match_sequence(os.path.join(data_path, raw_folder, sequence_name)) 22 | 23 | frames = [] 24 | 25 | n_processed_files = 0 26 | 27 | for category in CATEGORIES: 28 | print("Processing category %s" % category) 29 | 30 | # Get all files in current category's folder. 31 | folder_path = os.path.join(data_path, raw_folder, category) 32 | filenames = os.listdir(folder_path) 33 | 34 | for filename in filenames: 35 | filepath = os.path.join(folder_path, filename) 36 | vid = cv2.VideoCapture(filepath) 37 | 38 | # Store features in current file. 39 | frames_current_file = [] 40 | 41 | while vid.isOpened(): 42 | ret, frame = vid.read() 43 | 44 | if not ret: 45 | break 46 | 47 | frame = cv2.cvtColor(frame,cv2.COLOR_BGR2GRAY) 48 | 49 | frames_current_file.append(frame) 50 | 51 | if 'person13_handclapping_d3' in sequence[filename[:-11]]: 52 | continue 53 | 54 | if len(frames_current_file) < sequence[filename[:-11]][-1]: 55 | print('problem: {}, seq {}/c {}, fixed'.format(filename[:-11], sequence[filename[:-11]][-1], len(frames_current_file))) 56 | sequence[filename[:-11]][-1] = len(frames_current_file) 57 | 58 | 59 | 60 | frames.append({ 61 | 'filename': filename, 62 | 'category': category, 63 | 'frames': frames_current_file, 64 | 'sequence': sequence[filename[:-11]], 65 | }) 66 | 67 | n_processed_files += 1 68 | if n_processed_files % 100 == 0: 69 | print("Done %d files. (total: 600)" % n_processed_files) 70 | 71 | return split_data(data_path, processed_folder, frames) 72 | 73 | def split_data(data_path, processed_folder, all_data): 74 | 75 | train_id = list(map(lambda x: str(x).zfill(2), TRAIN_PEOPLE_ID)) 76 | test_id = list(map(lambda x: str(x).zfill(2), TEST_PEOPLE_ID)) 77 | validate_id = list(map(lambda x: str(x).zfill(2), DEV_PEOPLE_ID)) 78 | 79 | train_data = [] 80 | test_data = [] 81 | validate_data = [] 82 | 83 | def find_id(file_name): 84 | return re.findall(r"([0-9]{2})", file_name)[0] 85 | 86 | # 生成 train的 87 | for notes in all_data: 88 | notes_id = find_id(notes['filename']) 89 | 90 | if notes_id in train_id: 91 | train_data.append(notes) 92 | if notes_id in test_id: 93 | test_data.append(notes) 94 | if notes_id in validate_id: 95 | validate_data.append(notes) 96 | 97 | # Dump data to file. 98 | # print('Dumping train data to file...') 99 | # joblib.dump(all_data, open(os.path.join(data_path, processed_folder, 'train.pkl'), "wb")) 100 | # print('Dumping test data to file...') 101 | # joblib.dump(all_data, open(os.path.join(data_path, processed_folder, 'test.pkl'), "wb")) 102 | # print('Dumping validate data to file...') 103 | # joblib.dump(all_data, open(os.path.join(data_path, processed_folder, 'validate.pkl'), "wb")) 104 | return train_data, test_data, validate_data 105 | 106 | def match_sequence(path): 107 | with open(path, 'r') as f: 108 | data = f.read() 109 | h = re.findall(r"((?:person\d\d_\w{1,}_\w{1,})|(?:\d{1,4}))", data) 110 | 111 | start = False 112 | list_ = [] 113 | all_list = [] 114 | for note in h: 115 | if 'person' in note and start == False: 116 | start = True 117 | list_.append(note) 118 | elif 'person' not in note and start == True: 119 | list_.append(note) 120 | elif 'person' in note and start == True: 121 | all_list.append(list_) 122 | list_ = [] 123 | list_.append(note) 124 | all_list.append(list_) 125 | 126 | dic = {} 127 | 128 | for note in all_list: 129 | if 'person13_handclapping_d3' in note[0]: 130 | continue 131 | to_add = [1,] 132 | for i in range(2,len(note)-1): 133 | if int(note[i])+1 == int(note[i+1]) or int(note[i])-1 == int(note[i-1]): 134 | continue 135 | else: 136 | to_add.append(int(note[i])) 137 | to_add.append(int(note[-1])) 138 | dic[note[0]] = to_add 139 | 140 | return dic 141 | 142 | if __name__ == "__main__": 143 | # print(os.getcwd()) 144 | # match_sequence('./data/KTHDataset/raw/00sequences.txt') 145 | make_data(data_path='./data/KTHDataset', raw_folder='raw', processed_folder='processed', sequence_name='00sequences.txt') 146 | pass 147 | -------------------------------------------------------------------------------- /dataloader/MovingMNIST/MovingMNIST.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | import torch.utils.data as data 3 | from PIL import Image 4 | import os 5 | import os.path 6 | import errno 7 | import numpy as np 8 | import torch 9 | import codecs 10 | from torchvision import transforms 11 | 12 | 13 | class MovingMNIST(data.Dataset): 14 | """`MovingMNIST `_ Dataset. 15 | 16 | Args: 17 | root (string): Root directory of dataset where ``processed/training.pt`` 18 | and ``processed/test.pt`` exist. 19 | train (bool, optional): If True, creates dataset from ``training.pt``, 20 | otherwise from ``test.pt``. 21 | split (int, optional): Train/test split size. Number defines how many samples 22 | belong to test set. 23 | download (bool, optional): If true, downloads the dataset from the internet and 24 | puts it in root directory. If dataset is already downloaded, it is not 25 | downloaded again. 26 | transform (callable, optional): A function/transform that takes in an PIL image 27 | and returns a transformed version. E.g, ``transforms.RandomCrop`` 28 | target_transform (callable, optional): A function/transform that takes in an PIL 29 | image and returns a transformed version. E.g, ``transforms.RandomCrop`` 30 | """ 31 | urls = [ 32 | 'https://github.com/JaMesLiMers/MovingMNIST/raw/master/mnist_test_seq.npy.gz' 33 | ] 34 | raw_folder = 'raw' 35 | processed_folder = 'processed' 36 | training_file = 'moving_mnist_train.pt' 37 | test_file = 'moving_mnist_test.pt' 38 | 39 | def __init__(self, root, train=True, split=1000, transform=None, target_transform=None, download=False): 40 | self.root = os.path.expanduser(root) 41 | self.transform = transform 42 | self.target_transform = target_transform 43 | self.split = split 44 | self.train = train # training set or test set 45 | 46 | 47 | if download: 48 | self.download() 49 | 50 | if not self._check_exists(): 51 | raise RuntimeError('Dataset not found.' + 52 | ' You can use download=True to download it') 53 | 54 | if self.train: 55 | self.train_data = torch.load( 56 | os.path.join(self.root, self.processed_folder, self.training_file)) 57 | else: 58 | self.test_data = torch.load( 59 | os.path.join(self.root, self.processed_folder, self.test_file)) 60 | 61 | def __getitem__(self, index): 62 | """ 63 | Args: 64 | index (int): Index 65 | 66 | Returns: 67 | tuple: (seq, target) where sampled sequences are splitted into a seq 68 | and target part 69 | """ 70 | if self.train: 71 | seq, target = self.train_data[index, :10], self.train_data[index, 10:] 72 | else: 73 | seq, target = self.test_data[index, :10], self.test_data[index, 10:] 74 | 75 | # doing this so that it is consistent with all other datasets 76 | # to return a PIL Image 77 | seq = [Image.fromarray(seq.numpy()[i, :, :], mode='L') for i in range(10)] 78 | target = [Image.fromarray(target.numpy()[i, :, :], mode='L') for i in range(10)] 79 | 80 | if self.transform is not None: 81 | seq = torch.stack([self.transform(seq[i]) for i in range(10)]) 82 | 83 | if self.target_transform is not None: 84 | target = torch.stack([self.target_transform(target[i]) for i in range(10)]) 85 | 86 | return seq, target 87 | 88 | def __len__(self): 89 | if self.train: 90 | return len(self.train_data) 91 | else: 92 | return len(self.test_data) 93 | 94 | def _check_exists(self): 95 | return os.path.exists(os.path.join(self.root, self.processed_folder, self.training_file)) and \ 96 | os.path.exists(os.path.join(self.root, self.processed_folder, self.test_file)) 97 | 98 | def download(self): 99 | """Download the Moving MNIST data if it doesn't exist in processed_folder already.""" 100 | from six.moves import urllib 101 | import gzip 102 | 103 | if self._check_exists(): 104 | return 105 | 106 | # download files 107 | try: 108 | os.makedirs(os.path.join(self.root, self.raw_folder)) 109 | os.makedirs(os.path.join(self.root, self.processed_folder)) 110 | except OSError as e: 111 | if e.errno == errno.EEXIST: 112 | pass 113 | else: 114 | raise 115 | 116 | for url in self.urls: 117 | print('Downloading ' + url) 118 | data = urllib.request.urlopen(url) 119 | filename = url.rpartition('/')[2] 120 | file_path = os.path.join(self.root, self.raw_folder, filename) 121 | with open(file_path, 'wb') as f: 122 | f.write(data.read()) 123 | with open(file_path.replace('.gz', ''), 'wb') as out_f, \ 124 | gzip.GzipFile(file_path) as zip_f: 125 | out_f.write(zip_f.read()) 126 | os.unlink(file_path) 127 | 128 | # process and save as torch files 129 | print('Processing...') 130 | 131 | training_set = torch.from_numpy( 132 | np.load(os.path.join(self.root, self.raw_folder, 'mnist_test_seq.npy')).swapaxes(0, 1)[:-self.split] 133 | ) 134 | test_set = torch.from_numpy( 135 | np.load(os.path.join(self.root, self.raw_folder, 'mnist_test_seq.npy')).swapaxes(0, 1)[-self.split:] 136 | ) 137 | 138 | with open(os.path.join(self.root, self.processed_folder, self.training_file), 'wb') as f: 139 | torch.save(training_set, f) 140 | with open(os.path.join(self.root, self.processed_folder, self.test_file), 'wb') as f: 141 | torch.save(test_set, f) 142 | 143 | print('Done!') 144 | 145 | def __repr__(self): 146 | fmt_str = 'Dataset ' + self.__class__.__name__ + '\n' 147 | fmt_str += ' Number of datapoints: {}\n'.format(self.__len__()) 148 | tmp = 'train' if self.train is True else 'test' 149 | fmt_str += ' Train/test: {}\n'.format(tmp) 150 | fmt_str += ' Root Location: {}\n'.format(self.root) 151 | tmp = ' Transforms (if any): ' 152 | fmt_str += '{0}{1}\n'.format(tmp, self.transform.__repr__().replace('\n', '\n' + ' ' * len(tmp))) 153 | tmp = ' Target Transforms (if any): ' 154 | fmt_str += '{0}{1}'.format(tmp, self.target_transform.__repr__().replace('\n', '\n' + ' ' * len(tmp))) 155 | return fmt_str 156 | 157 | 158 | 159 | -------------------------------------------------------------------------------- /dataloader/MovingMNIST/demo.py: -------------------------------------------------------------------------------- 1 | import os 2 | import sys 3 | import numpy as np 4 | import matplotlib.pyplot as plt 5 | 6 | parent_path = os.path.dirname(os.path.dirname(os.getcwd())) 7 | print(os.getcwd()) 8 | if 'MovingMNIST' in os.getcwd(): 9 | os.chdir(parent_path) 10 | sys.path.insert(0, os.getcwd()) 11 | 12 | import torch 13 | import torch.nn as nn 14 | from torch.autograd import Variable 15 | import torchvision.datasets as datasets 16 | import torchvision.transforms as transforms 17 | import torch.nn.functional as F 18 | import torch.optim as optim 19 | 20 | from dataloader.MovingMNIST.MovingMNIST import MovingMNIST 21 | 22 | root = './data' 23 | if not os.path.exists(root): 24 | os.mkdir(root) 25 | 26 | 27 | train_set = MovingMNIST(root=os.path.join('.', 'data','mnist'), train=True, download=True, 28 | transform=transforms.Compose([transforms.ToTensor(),]), 29 | target_transform=transforms.Compose([transforms.ToTensor(),])) 30 | test_set = MovingMNIST(root=os.path.join('.', 'data','mnist'), train=False, download=True, 31 | transform=transforms.Compose([transforms.ToTensor(),]), 32 | target_transform=transforms.Compose([transforms.ToTensor(),])) 33 | 34 | batch_size = 100 35 | 36 | train_loader = torch.utils.data.DataLoader( 37 | dataset=train_set, 38 | batch_size=batch_size, 39 | shuffle=True) 40 | test_loader = torch.utils.data.DataLoader( 41 | dataset=test_set, 42 | batch_size=batch_size, 43 | shuffle=False) 44 | 45 | print('==>>> total trainning batch number: {}'.format(len(train_loader))) 46 | print('==>>> total testing batch number: {}'.format(len(test_loader))) 47 | 48 | 49 | for seq, seq_target in train_loader: 50 | print('--- Sample') 51 | print('Input: ', seq.shape) 52 | print('Target: ', seq_target.shape) 53 | break -------------------------------------------------------------------------------- /dataloader/README.md: -------------------------------------------------------------------------------- 1 | # Dataloader 2 | This file is to store dataloaders. 3 | 4 | Every dataset will have one dataloader and a demo, you can try demo.py to test the datasets. -------------------------------------------------------------------------------- /experiments/PredRNN/custom.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from model.STconvLSTM import STConvLSTMCell 4 | 5 | class Custom(nn.Module): 6 | def __init__(self, cfg): 7 | """ 8 | All we need to use is in cfg 9 | cfg.somepart["par_name"] 10 | """ 11 | super(Custom, self).__init__() 12 | self.input_size=cfg['input_size'] 13 | self.hidden_dim=cfg['hidden_dim'] 14 | self.input_dim=cfg['input_dim'] 15 | self.kernel_size=tuple(cfg['kernel_size']) 16 | 17 | self.stlstm_1 = STConvLSTMCell(input_size=self.input_size, 18 | input_dim=self.input_dim, 19 | hidden_dim=self.hidden_dim, 20 | kernel_size=self.kernel_size, 21 | bias=True) 22 | 23 | self.stlstm_2 = STConvLSTMCell(input_size=self.input_size, 24 | input_dim=self.hidden_dim, 25 | hidden_dim=self.hidden_dim, 26 | kernel_size=self.kernel_size, 27 | bias=True) 28 | 29 | self.stlstm_3 = STConvLSTMCell(input_size=self.input_size, 30 | input_dim=self.hidden_dim, 31 | hidden_dim=self.hidden_dim, 32 | kernel_size=self.kernel_size, 33 | bias=True) 34 | 35 | self.stlstm_4 = STConvLSTMCell(input_size=self.input_size, 36 | input_dim=self.hidden_dim, 37 | hidden_dim=self.hidden_dim, 38 | kernel_size=self.kernel_size, 39 | bias=True) 40 | 41 | self.head = nn.Conv2d(in_channels=self.hidden_dim, 42 | out_channels=self.input_dim, 43 | kernel_size=(1,1)) 44 | 45 | 46 | def forward(self, input, hidden=None, future=10): 47 | """ 48 | input: (b,t,c,h,w) 49 | hidden: hidden of last time (b, c_hidden, h, w) 50 | future: number of future frame to predict 51 | """ 52 | # Init hidden 53 | if hidden is None: 54 | h_t1, c_t1, m_t1 = self.stlstm_1.init_hidden(input.size(0)) 55 | h_t2, c_t2, _ = self.stlstm_2.init_hidden(input.size(0)) 56 | h_t3, c_t3, _ = self.stlstm_3.init_hidden(input.size(0)) 57 | h_t4, c_t4, _ = self.stlstm_4.init_hidden(input.size(0)) 58 | else: 59 | # TODO: build a stateful model 60 | raise NotImplementedError 61 | 62 | outputs = [] 63 | 64 | seq_len = input.size(1) 65 | 66 | for t in range(seq_len): 67 | if t is not 0: 68 | m_t1 = m_t4 69 | 70 | h_t1, c_t1, m_t1 = self.stlstm_1(input_tensor=input[:,t,:,:,:], 71 | cur_state=[h_t1, c_t1, m_t1]) 72 | 73 | h_t2, c_t2, m_t2 = self.stlstm_2(input_tensor=h_t1, 74 | cur_state=[h_t2, c_t2, m_t1]) 75 | 76 | h_t3, c_t3, m_t3 = self.stlstm_3(input_tensor=h_t2, 77 | cur_state=[h_t3, c_t3, m_t2]) 78 | 79 | h_t4, c_t4, m_t4 = self.stlstm_4(input_tensor=h_t3, 80 | cur_state=[h_t4, c_t4, m_t3]) 81 | 82 | output = self.head(h_t4) 83 | output = torch.sigmoid(output) 84 | outputs += [output] 85 | 86 | 87 | for t in range(future): 88 | m_t1 = m_t4 89 | 90 | h_t1, c_t1, m_t1 = self.stlstm_1(input_tensor=outputs[-1], 91 | cur_state=[h_t1, c_t1, m_t1]) 92 | 93 | h_t2, c_t2, m_t2 = self.stlstm_2(input_tensor=h_t1, 94 | cur_state=[h_t2, c_t2, m_t1]) 95 | 96 | h_t3, c_t3, m_t3 = self.stlstm_3(input_tensor=h_t2, 97 | cur_state=[h_t3, c_t3, m_t2]) 98 | 99 | h_t4, c_t4, m_t4 = self.stlstm_4(input_tensor=h_t3, 100 | cur_state=[h_t4, c_t4, m_t3]) 101 | 102 | output = self.head(h_t4) 103 | output = torch.sigmoid(output) 104 | outputs += [output] 105 | 106 | 107 | outputs = torch.stack(outputs, 1) 108 | 109 | return outputs 110 | 111 | def test(): 112 | pass 113 | 114 | if __name__ == "__main__": 115 | test() 116 | 117 | 118 | -------------------------------------------------------------------------------- /experiments/README.md: -------------------------------------------------------------------------------- 1 | # Experiments 2 | This folder is to store the experiments models 3 | 4 | You should specify the 'experiment_name' tag as your own folder in the config file when you're using your own models to train. 5 | 6 | If you didn't specify the 'arch' tag in the config file, the training program will automatically find the Custom class in $ROOT/experiments/'experiment_name'/Custom.py as the implemented models. 7 | 8 | If you want to use other arch, you may want to change the training code for your own arch. 9 | -------------------------------------------------------------------------------- /experiments/video_prediction_demo/custom.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from model.convRNN import ConvGRUCell, ConvLSTMCell 4 | 5 | class Custom(nn.Module): 6 | def __init__(self, cfg): 7 | """ 8 | All we need to use is in cfg 9 | cfg.somepart["par_name"] 10 | """ 11 | super(Custom, self).__init__() 12 | self.input_size=cfg['input_size'] 13 | self.hidden_dim=cfg['hidden_dim'] 14 | self.input_dim=cfg['input_dim'] 15 | 16 | self.convlstm_1 = ConvLSTMCell(input_size=self.input_size, 17 | input_dim=self.input_dim, 18 | hidden_dim=self.hidden_dim, 19 | kernel_size=(3, 3), 20 | bias=True) 21 | 22 | self.convlstm_2 = ConvLSTMCell(input_size=self.input_size, 23 | input_dim=self.hidden_dim, 24 | hidden_dim=self.hidden_dim, 25 | kernel_size=(3, 3), 26 | bias=True) 27 | 28 | self.conv2d = nn.Conv2d(in_channels=self.hidden_dim, 29 | out_channels=self.input_dim, 30 | bias=False, 31 | kernel_size=(3, 3), 32 | padding=1) 33 | 34 | self.conv3d = nn.Conv3d(in_channels=self.hidden_dim, 35 | out_channels=self.input_dim, 36 | bias=False, 37 | kernel_size=(3, 3, 3), 38 | padding=1) 39 | 40 | 41 | def forward(self, input, hidden=None, future=10): 42 | """ 43 | input: (b,t,c,h,w) 44 | hidden: hidden of last time (b, c_hidden, h, w) 45 | future: number of future frame to predict 46 | """ 47 | # Init hidden 48 | if hidden is None: 49 | h_t, c_t = self.convlstm_1.init_hidden(input.size(0)) 50 | h_t2, c_t2 = self.convlstm_1.init_hidden(input.size(0)) 51 | else: 52 | # TODO: build a stateful model 53 | raise NotImplementedError 54 | 55 | outputs = [] 56 | 57 | seq_len = input.size(1) 58 | 59 | for t in range(seq_len): 60 | 61 | h_t, c_t = self.convlstm_1(input_tensor=input[:,t,:,:,:], 62 | cur_state=[h_t, c_t]) 63 | 64 | h_t2, c_t2 = self.convlstm_2(input_tensor=h_t, 65 | cur_state=[h_t2, c_t2]) 66 | 67 | output = self.conv2d(h_t2) 68 | output = nn.Sigmoid()(output) 69 | outputs += [output] 70 | 71 | for i in range(future): 72 | 73 | h_t, c_t = self.convlstm_1(input_tensor=output, 74 | cur_state=[h_t, c_t]) 75 | 76 | h_t2, c_t2 = self.convlstm_2(input_tensor=h_t, 77 | cur_state=[h_t2, c_t2]) 78 | 79 | output = self.conv2d(h_t2) 80 | output = nn.Sigmoid()(output) 81 | outputs += [output] 82 | 83 | outputs = torch.stack(outputs, 1) 84 | 85 | return outputs 86 | 87 | -------------------------------------------------------------------------------- /logs/PredRNN.log: -------------------------------------------------------------------------------- 1 | [2019-09-18 14:53:23,691-config_helper.py# 61] [DEBUG] Used config: 2 | {'meta': {'experiment_path': 'experiments', 'arch': 'Custom', 'board_path': 'board', 'experiment_name': 'PredRNN'}, 'train': {'epoches': 50, 'batch_size': 1, 'lr': 0.001, 'print_freq': 20}, 'model': {'input_size': [64, 64], 'input_dim': 1, 'hidden_dim': 64, 'kernel_size': [3, 3], 'input_num': 10}} 3 | [2019-09-18 14:54:32,258-train_video_prediction_KTHDataset.py# 91] [DEBUG] ==>>> total trainning batch number: 186884 4 | [2019-09-18 14:54:32,258-train_video_prediction_KTHDataset.py# 92] [DEBUG] ==>>> total testing batch number: 102124 5 | [2019-09-18 14:54:34,147-train_video_prediction_KTHDataset.py#127] [DEBUG] Input: torch.Size([1, 10, 1, 64, 64]) 6 | [2019-09-18 14:54:34,147-train_video_prediction_KTHDataset.py#128] [DEBUG] --- Sample 7 | [2019-09-18 14:54:34,147-train_video_prediction_KTHDataset.py#129] [DEBUG] Target: torch.Size([1, 10, 1, 64, 64]) 8 | [2019-09-18 14:54:43,285-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][20/186884] step_time: 0.348360 (0.453101) train_loss: 0.517230 (0.639866) test_loss: 0.726505 (0.766993) train_metric: 0.634010 (0.601406) 9 | [2019-09-18 14:54:43,285-log_helper.py# 77] [INFO] Progress: 20 / 9344200 [0%], Speed: 0.453 s/iter, ETA 49:00:04 (D:H:M) 10 | 11 | [2019-09-18 14:54:49,939-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][40/186884] step_time: 0.325024 (0.390925) train_loss: 0.705032 (0.650462) test_loss: 0.703754 (0.754136) train_metric: 0.632030 (0.652420) 12 | [2019-09-18 14:54:49,939-log_helper.py# 77] [INFO] Progress: 40 / 9344200 [0%], Speed: 0.391 s/iter, ETA 42:06:41 (D:H:M) 13 | 14 | [2019-09-18 14:54:56,857-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][60/186884] step_time: 0.365928 (0.374611) train_loss: 0.749587 (0.650766) test_loss: 0.757015 (0.746311) train_metric: 0.698102 (0.654956) 15 | [2019-09-18 14:54:56,857-log_helper.py# 77] [INFO] Progress: 60 / 9344200 [0%], Speed: 0.375 s/iter, ETA 40:12:20 (D:H:M) 16 | 17 | [2019-09-18 14:55:03,531-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][80/186884] step_time: 0.342730 (0.363407) train_loss: 0.698763 (0.652654) test_loss: 0.704605 (0.737109) train_metric: 0.719438 (0.666852) 18 | [2019-09-18 14:55:03,531-log_helper.py# 77] [INFO] Progress: 80 / 9344200 [0%], Speed: 0.363 s/iter, ETA 39:07:15 (D:H:M) 19 | 20 | [2019-09-18 14:55:10,395-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][100/186884] step_time: 0.322581 (0.358561) train_loss: 0.609662 (0.657957) test_loss: 0.693214 (0.729694) train_metric: 0.631378 (0.665684) 21 | [2019-09-18 14:55:10,395-log_helper.py# 77] [INFO] Progress: 100 / 9344200 [0%], Speed: 0.359 s/iter, ETA 38:18:40 (D:H:M) 22 | 23 | [2019-09-18 14:55:17,484-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][120/186884] step_time: 0.320231 (0.357220) train_loss: 0.711436 (0.655228) test_loss: 0.736374 (0.727588) train_metric: 0.701539 (0.668283) 24 | [2019-09-18 14:55:17,484-log_helper.py# 77] [INFO] Progress: 120 / 9344200 [0%], Speed: 0.357 s/iter, ETA 38:15:11 (D:H:M) 25 | 26 | [2019-09-18 14:55:24,316-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][140/186884] step_time: 0.326212 (0.354441) train_loss: 0.702185 (0.656728) test_loss: 0.676167 (0.725431) train_metric: 0.767893 (0.676842) 27 | [2019-09-18 14:55:24,316-log_helper.py# 77] [INFO] Progress: 140 / 9344200 [0%], Speed: 0.354 s/iter, ETA 38:07:58 (D:H:M) 28 | 29 | [2019-09-18 14:55:31,197-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][160/186884] step_time: 0.338116 (0.352668) train_loss: 0.689022 (0.659467) test_loss: 0.701217 (0.721752) train_metric: 0.686003 (0.674363) 30 | [2019-09-18 14:55:31,197-log_helper.py# 77] [INFO] Progress: 160 / 9344200 [0%], Speed: 0.353 s/iter, ETA 38:03:22 (D:H:M) 31 | 32 | [2019-09-18 14:55:37,944-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][180/186884] step_time: 0.328868 (0.350562) train_loss: 0.693971 (0.659177) test_loss: 0.709595 (0.719115) train_metric: 0.755272 (0.679103) 33 | [2019-09-18 14:55:37,944-log_helper.py# 77] [INFO] Progress: 180 / 9344200 [0%], Speed: 0.351 s/iter, ETA 37:21:54 (D:H:M) 34 | 35 | [2019-09-18 14:55:46,409-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][200/186884] step_time: 0.362261 (0.349478) train_loss: 0.724213 (0.658869) test_loss: 0.729188 (0.718298) train_metric: 0.801228 (0.681950) 36 | [2019-09-18 14:55:46,420-log_helper.py# 77] [INFO] Progress: 200 / 9344200 [0%], Speed: 0.349 s/iter, ETA 37:19:05 (D:H:M) 37 | 38 | [2019-09-18 14:55:53,454-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][220/186884] step_time: 0.360313 (0.349261) train_loss: 0.642676 (0.659909) test_loss: 0.688910 (0.716734) train_metric: 0.673523 (0.684866) 39 | [2019-09-18 14:55:53,454-log_helper.py# 77] [INFO] Progress: 220 / 9344200 [0%], Speed: 0.349 s/iter, ETA 37:18:31 (D:H:M) 40 | 41 | [2019-09-18 14:56:00,186-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][240/186884] step_time: 0.328118 (0.347891) train_loss: 0.704590 (0.660106) test_loss: 0.658484 (0.715130) train_metric: 0.650349 (0.688695) 42 | [2019-09-18 14:56:00,186-log_helper.py# 77] [INFO] Progress: 240 / 9344200 [0%], Speed: 0.348 s/iter, ETA 37:14:57 (D:H:M) 43 | 44 | [2019-09-18 14:56:06,844-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][260/186884] step_time: 0.327543 (0.346452) train_loss: 0.829270 (0.651603) test_loss: 0.804557 (0.718573) train_metric: 0.526103 (0.689696) 45 | [2019-09-18 14:56:06,845-log_helper.py# 77] [INFO] Progress: 260 / 9344200 [0%], Speed: 0.346 s/iter, ETA 37:11:13 (D:H:M) 46 | 47 | [2019-09-18 14:56:13,492-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][280/186884] step_time: 0.329523 (0.345174) train_loss: 0.735278 (0.651227) test_loss: 0.701229 (0.722291) train_metric: 0.648790 (0.687422) 48 | [2019-09-18 14:56:13,492-log_helper.py# 77] [INFO] Progress: 280 / 9344200 [0%], Speed: 0.345 s/iter, ETA 37:07:54 (D:H:M) 49 | 50 | [2019-09-18 14:56:20,162-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][300/186884] step_time: 0.327708 (0.344143) train_loss: 0.696382 (0.651353) test_loss: 0.785641 (0.722203) train_metric: 0.621929 (0.687887) 51 | [2019-09-18 14:56:20,163-log_helper.py# 77] [INFO] Progress: 300 / 9344200 [0%], Speed: 0.344 s/iter, ETA 37:05:14 (D:H:M) 52 | 53 | [2019-09-18 14:56:26,823-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][320/186884] step_time: 0.326851 (0.343209) train_loss: 0.694889 (0.647206) test_loss: 0.781595 (0.722631) train_metric: 0.757142 (0.691552) 54 | [2019-09-18 14:56:26,824-log_helper.py# 77] [INFO] Progress: 320 / 9344200 [0%], Speed: 0.343 s/iter, ETA 37:02:48 (D:H:M) 55 | 56 | [2019-09-18 14:56:33,527-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][340/186884] step_time: 0.331775 (0.342504) train_loss: 0.758644 (0.649206) test_loss: 0.691193 (0.721021) train_metric: 0.619610 (0.689754) 57 | [2019-09-18 14:56:33,528-log_helper.py# 77] [INFO] Progress: 340 / 9344200 [0%], Speed: 0.343 s/iter, ETA 37:00:58 (D:H:M) 58 | 59 | [2019-09-18 14:56:40,402-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][360/186884] step_time: 0.341495 (0.342348) train_loss: 0.695488 (0.650099) test_loss: 0.701947 (0.720252) train_metric: 0.708636 (0.689841) 60 | [2019-09-18 14:56:40,402-log_helper.py# 77] [INFO] Progress: 360 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:34 (D:H:M) 61 | 62 | [2019-09-18 14:56:47,256-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][380/186884] step_time: 0.342752 (0.342159) train_loss: 0.459075 (0.649123) test_loss: 0.681666 (0.719431) train_metric: 0.741584 (0.691283) 63 | [2019-09-18 14:56:47,257-log_helper.py# 77] [INFO] Progress: 380 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:04 (D:H:M) 64 | 65 | [2019-09-18 14:56:54,312-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][400/186884] step_time: 0.331604 (0.342029) train_loss: 0.554359 (0.648302) test_loss: 0.660817 (0.718903) train_metric: 0.545888 (0.691420) 66 | [2019-09-18 14:56:54,312-log_helper.py# 77] [INFO] Progress: 400 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:23:44 (D:H:M) 67 | 68 | [2019-09-18 14:57:01,194-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][420/186884] step_time: 0.335888 (0.341942) train_loss: 0.544910 (0.649137) test_loss: 0.691596 (0.718279) train_metric: 0.750356 (0.693156) 69 | [2019-09-18 14:57:01,194-log_helper.py# 77] [INFO] Progress: 420 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:23:30 (D:H:M) 70 | 71 | [2019-09-18 14:57:08,113-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][440/186884] step_time: 0.357398 (0.341952) train_loss: 0.623607 (0.648168) test_loss: 0.781761 (0.718552) train_metric: 0.722056 (0.694793) 72 | [2019-09-18 14:57:08,114-log_helper.py# 77] [INFO] Progress: 440 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:23:31 (D:H:M) 73 | 74 | [2019-09-18 14:57:15,241-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][460/186884] step_time: 0.326790 (0.342401) train_loss: 0.604105 (0.648317) test_loss: 0.752652 (0.718479) train_metric: 0.590202 (0.696774) 75 | [2019-09-18 14:57:15,241-log_helper.py# 77] [INFO] Progress: 460 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:41 (D:H:M) 76 | 77 | [2019-09-18 14:57:22,112-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][480/186884] step_time: 0.344495 (0.342278) train_loss: 0.425265 (0.648289) test_loss: 0.756242 (0.718228) train_metric: 0.761326 (0.697923) 78 | [2019-09-18 14:57:22,112-log_helper.py# 77] [INFO] Progress: 480 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:22 (D:H:M) 79 | 80 | [2019-09-18 14:57:28,910-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][500/186884] step_time: 0.330012 (0.342030) train_loss: 0.376890 (0.645566) test_loss: 0.669748 (0.718448) train_metric: 0.746557 (0.700582) 81 | [2019-09-18 14:57:28,911-log_helper.py# 77] [INFO] Progress: 500 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:23:43 (D:H:M) 82 | 83 | [2019-09-18 14:57:35,588-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][520/186884] step_time: 0.332534 (0.341573) train_loss: 0.489950 (0.643250) test_loss: 0.758088 (0.719873) train_metric: 0.697546 (0.702059) 84 | [2019-09-18 14:57:35,588-log_helper.py# 77] [INFO] Progress: 520 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:22:32 (D:H:M) 85 | 86 | [2019-09-18 14:57:42,264-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][540/186884] step_time: 0.332781 (0.341147) train_loss: 0.733589 (0.644587) test_loss: 0.689811 (0.721462) train_metric: 0.687096 (0.701822) 87 | [2019-09-18 14:57:42,264-log_helper.py# 77] [INFO] Progress: 540 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:25 (D:H:M) 88 | 89 | [2019-09-18 14:57:49,106-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][560/186884] step_time: 0.325249 (0.341042) train_loss: 0.701638 (0.644935) test_loss: 0.744144 (0.720296) train_metric: 0.613428 (0.703259) 90 | [2019-09-18 14:57:49,106-log_helper.py# 77] [INFO] Progress: 560 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:09 (D:H:M) 91 | 92 | [2019-09-18 14:57:56,151-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][580/186884] step_time: 0.351255 (0.341286) train_loss: 0.498550 (0.644607) test_loss: 0.710355 (0.719491) train_metric: 0.714513 (0.704282) 93 | [2019-09-18 14:57:56,151-log_helper.py# 77] [INFO] Progress: 580 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:47 (D:H:M) 94 | 95 | [2019-09-18 14:58:03,279-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][600/186884] step_time: 0.330402 (0.341356) train_loss: 0.642661 (0.644646) test_loss: 0.682568 (0.718929) train_metric: 0.423996 (0.704366) 96 | [2019-09-18 14:58:03,279-log_helper.py# 77] [INFO] Progress: 600 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:58 (D:H:M) 97 | 98 | [2019-09-18 14:58:10,190-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][620/186884] step_time: 0.350517 (0.341366) train_loss: 0.732954 (0.644604) test_loss: 0.710484 (0.718348) train_metric: 0.764885 (0.704025) 99 | [2019-09-18 14:58:10,190-log_helper.py# 77] [INFO] Progress: 620 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:59 (D:H:M) 100 | 101 | [2019-09-18 14:58:17,084-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][640/186884] step_time: 0.327558 (0.341352) train_loss: 0.672510 (0.643831) test_loss: 0.740274 (0.718899) train_metric: 0.816277 (0.705024) 102 | [2019-09-18 14:58:17,084-log_helper.py# 77] [INFO] Progress: 640 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:57 (D:H:M) 103 | 104 | [2019-09-18 14:58:23,815-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][660/186884] step_time: 0.334197 (0.341089) train_loss: 0.698336 (0.644496) test_loss: 0.692642 (0.718993) train_metric: 0.766523 (0.705627) 105 | [2019-09-18 14:58:23,815-log_helper.py# 77] [INFO] Progress: 660 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:16 (D:H:M) 106 | 107 | [2019-09-18 14:58:30,654-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][680/186884] step_time: 0.341605 (0.341003) train_loss: 0.714170 (0.645157) test_loss: 0.686007 (0.718547) train_metric: 0.733276 (0.706032) 108 | [2019-09-18 14:58:30,655-log_helper.py# 77] [INFO] Progress: 680 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:02 (D:H:M) 109 | 110 | [2019-09-18 14:58:37,676-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][700/186884] step_time: 0.342468 (0.341181) train_loss: 0.526295 (0.645492) test_loss: 0.714966 (0.717509) train_metric: 0.641062 (0.705742) 111 | [2019-09-18 14:58:37,676-log_helper.py# 77] [INFO] Progress: 700 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:30 (D:H:M) 112 | 113 | [2019-09-18 14:58:44,363-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][720/186884] step_time: 0.351441 (0.340888) train_loss: 0.737629 (0.646042) test_loss: 0.697693 (0.717112) train_metric: 0.673805 (0.705640) 114 | [2019-09-18 14:58:44,363-log_helper.py# 77] [INFO] Progress: 720 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:20:44 (D:H:M) 115 | 116 | [2019-09-18 14:58:51,013-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][740/186884] step_time: 0.320224 (0.340550) train_loss: 0.527757 (0.646331) test_loss: 0.693957 (0.716380) train_metric: 0.702426 (0.706620) 117 | [2019-09-18 14:58:51,013-log_helper.py# 77] [INFO] Progress: 740 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:19:51 (D:H:M) 118 | 119 | [2019-09-18 14:58:57,664-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][760/186884] step_time: 0.325731 (0.340239) train_loss: 0.720300 (0.646139) test_loss: 0.693440 (0.716716) train_metric: 0.679519 (0.707454) 120 | [2019-09-18 14:58:57,664-log_helper.py# 77] [INFO] Progress: 760 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:19:03 (D:H:M) 121 | 122 | [2019-09-18 14:59:04,280-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][780/186884] step_time: 0.338799 (0.339903) train_loss: 0.752751 (0.644377) test_loss: 0.825015 (0.717005) train_metric: 0.753655 (0.708552) 123 | [2019-09-18 14:59:04,281-log_helper.py# 77] [INFO] Progress: 780 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:18:10 (D:H:M) 124 | 125 | [2019-09-18 14:59:11,211-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][800/186884] step_time: 0.332133 (0.339719) train_loss: 0.560865 (0.642237) test_loss: 0.770177 (0.719286) train_metric: 0.739026 (0.709522) 126 | [2019-09-18 14:59:11,211-log_helper.py# 77] [INFO] Progress: 800 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:42 (D:H:M) 127 | 128 | [2019-09-18 14:59:17,909-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][820/186884] step_time: 0.321504 (0.339511) train_loss: 0.704266 (0.641766) test_loss: 0.709862 (0.719648) train_metric: 0.643010 (0.709938) 129 | [2019-09-18 14:59:17,909-log_helper.py# 77] [INFO] Progress: 820 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:09 (D:H:M) 130 | 131 | [2019-09-18 14:59:24,527-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][840/186884] step_time: 0.320519 (0.339218) train_loss: 0.689155 (0.642408) test_loss: 0.675148 (0.719547) train_metric: 0.645189 (0.710892) 132 | [2019-09-18 14:59:24,528-log_helper.py# 77] [INFO] Progress: 840 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:23 (D:H:M) 133 | 134 | [2019-09-18 14:59:31,113-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][860/186884] step_time: 0.323160 (0.338903) train_loss: 0.726251 (0.641878) test_loss: 0.751375 (0.719551) train_metric: 0.667097 (0.711387) 135 | [2019-09-18 14:59:31,114-log_helper.py# 77] [INFO] Progress: 860 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:15:34 (D:H:M) 136 | 137 | [2019-09-18 14:59:37,721-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][880/186884] step_time: 0.322946 (0.338622) train_loss: 0.806910 (0.640650) test_loss: 0.837398 (0.720516) train_metric: 0.612515 (0.711691) 138 | [2019-09-18 14:59:37,721-log_helper.py# 77] [INFO] Progress: 880 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:14:50 (D:H:M) 139 | 140 | [2019-09-18 14:59:44,405-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][900/186884] step_time: 0.361146 (0.338439) train_loss: 0.600179 (0.639960) test_loss: 0.623951 (0.720080) train_metric: 0.688135 (0.711709) 141 | [2019-09-18 14:59:44,629-log_helper.py# 77] [INFO] Progress: 900 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:14:22 (D:H:M) 142 | 143 | [2019-09-18 14:59:51,389-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][920/186884] step_time: 0.326705 (0.338348) train_loss: 0.367926 (0.638655) test_loss: 0.696554 (0.719794) train_metric: 0.794559 (0.712247) 144 | [2019-09-18 14:59:51,389-log_helper.py# 77] [INFO] Progress: 920 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:14:07 (D:H:M) 145 | 146 | [2019-09-18 14:59:58,221-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][940/186884] step_time: 0.328056 (0.338338) train_loss: 0.460637 (0.637110) test_loss: 0.706048 (0.719162) train_metric: 0.783932 (0.713101) 147 | [2019-09-18 14:59:58,221-log_helper.py# 77] [INFO] Progress: 940 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:14:06 (D:H:M) 148 | 149 | [2019-09-18 15:00:05,211-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][960/186884] step_time: 0.333569 (0.338488) train_loss: 0.544254 (0.636465) test_loss: 0.684148 (0.718836) train_metric: 0.602092 (0.712778) 150 | [2019-09-18 15:00:05,211-log_helper.py# 77] [INFO] Progress: 960 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:14:29 (D:H:M) 151 | 152 | [2019-09-18 15:00:11,890-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][980/186884] step_time: 0.334288 (0.338314) train_loss: 0.724455 (0.636037) test_loss: 0.693153 (0.718335) train_metric: 0.692899 (0.713573) 153 | [2019-09-18 15:00:11,890-log_helper.py# 77] [INFO] Progress: 980 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:14:02 (D:H:M) 154 | 155 | [2019-09-18 15:00:18,839-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1000/186884] step_time: 0.336447 (0.338246) train_loss: 0.779174 (0.635582) test_loss: 0.946605 (0.718537) train_metric: 0.587165 (0.714475) 156 | [2019-09-18 15:00:18,839-log_helper.py# 77] [INFO] Progress: 1000 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:13:51 (D:H:M) 157 | 158 | [2019-09-18 15:00:25,975-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1020/186884] step_time: 0.360772 (0.338534) train_loss: 0.694328 (0.634908) test_loss: 0.656792 (0.718328) train_metric: 0.680472 (0.714927) 159 | [2019-09-18 15:00:25,976-log_helper.py# 77] [INFO] Progress: 1020 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:14:36 (D:H:M) 160 | 161 | [2019-09-18 15:00:32,795-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1040/186884] step_time: 0.330965 (0.338509) train_loss: 0.695144 (0.633879) test_loss: 0.683907 (0.717665) train_metric: 0.728950 (0.716356) 162 | [2019-09-18 15:00:32,796-log_helper.py# 77] [INFO] Progress: 1040 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:14:32 (D:H:M) 163 | 164 | [2019-09-18 15:00:39,550-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1060/186884] step_time: 0.336396 (0.338425) train_loss: 0.674261 (0.633220) test_loss: 0.691664 (0.717206) train_metric: 0.872526 (0.717470) 165 | [2019-09-18 15:00:39,550-log_helper.py# 77] [INFO] Progress: 1060 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:14:19 (D:H:M) 166 | 167 | [2019-09-18 15:00:46,327-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1080/186884] step_time: 0.344196 (0.338364) train_loss: 0.663008 (0.632017) test_loss: 0.695228 (0.716542) train_metric: 0.850566 (0.718616) 168 | [2019-09-18 15:00:46,327-log_helper.py# 77] [INFO] Progress: 1080 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:14:09 (D:H:M) 169 | 170 | [2019-09-18 15:00:53,110-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1100/186884] step_time: 0.340802 (0.338309) train_loss: 0.658749 (0.631997) test_loss: 0.664278 (0.717305) train_metric: 0.775733 (0.719373) 171 | [2019-09-18 15:00:53,110-log_helper.py# 77] [INFO] Progress: 1100 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:14:00 (D:H:M) 172 | 173 | [2019-09-18 15:00:59,860-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1120/186884] step_time: 0.337368 (0.338229) train_loss: 0.695592 (0.632056) test_loss: 0.716024 (0.716920) train_metric: 0.663405 (0.719508) 174 | [2019-09-18 15:00:59,860-log_helper.py# 77] [INFO] Progress: 1120 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:13:48 (D:H:M) 175 | 176 | [2019-09-18 15:01:06,615-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1140/186884] step_time: 0.336736 (0.338152) train_loss: 0.678107 (0.632063) test_loss: 0.694422 (0.716378) train_metric: 0.731002 (0.719923) 177 | [2019-09-18 15:01:06,615-log_helper.py# 77] [INFO] Progress: 1140 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:13:36 (D:H:M) 178 | 179 | [2019-09-18 15:01:13,631-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1160/186884] step_time: 0.325226 (0.338303) train_loss: 0.439417 (0.630789) test_loss: 0.908915 (0.716867) train_metric: 0.696818 (0.720913) 180 | [2019-09-18 15:01:13,631-log_helper.py# 77] [INFO] Progress: 1160 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:13:59 (D:H:M) 181 | 182 | [2019-09-18 15:01:20,667-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1180/186884] step_time: 0.349917 (0.338465) train_loss: 0.670865 (0.630503) test_loss: 0.686611 (0.716239) train_metric: 0.783299 (0.721519) 183 | [2019-09-18 15:01:20,667-log_helper.py# 77] [INFO] Progress: 1180 / 9344200 [0%], Speed: 0.338 s/iter, ETA 36:14:24 (D:H:M) 184 | 185 | [2019-09-18 15:01:27,970-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1200/186884] step_time: 0.362407 (0.338706) train_loss: 0.676892 (0.629714) test_loss: 0.686181 (0.715627) train_metric: 0.840240 (0.722423) 186 | [2019-09-18 15:01:27,981-log_helper.py# 77] [INFO] Progress: 1200 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:15:02 (D:H:M) 187 | 188 | [2019-09-18 15:01:35,004-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1220/186884] step_time: 0.345468 (0.338837) train_loss: 0.390380 (0.629406) test_loss: 0.688224 (0.715045) train_metric: 0.767983 (0.723250) 189 | [2019-09-18 15:01:35,004-log_helper.py# 77] [INFO] Progress: 1220 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:15:22 (D:H:M) 190 | 191 | [2019-09-18 15:01:42,263-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1240/186884] step_time: 0.382801 (0.339156) train_loss: 0.679508 (0.628850) test_loss: 0.687383 (0.714595) train_metric: 0.758646 (0.723941) 192 | [2019-09-18 15:01:42,263-log_helper.py# 77] [INFO] Progress: 1240 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:12 (D:H:M) 193 | 194 | [2019-09-18 15:01:49,432-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1260/186884] step_time: 0.348366 (0.339395) train_loss: 0.667521 (0.628094) test_loss: 0.690587 (0.714084) train_metric: 0.875762 (0.725135) 195 | [2019-09-18 15:01:49,432-log_helper.py# 77] [INFO] Progress: 1260 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:49 (D:H:M) 196 | 197 | [2019-09-18 15:01:56,514-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1280/186884] step_time: 0.363514 (0.339564) train_loss: 0.677752 (0.628134) test_loss: 0.682729 (0.713557) train_metric: 0.821317 (0.726152) 198 | [2019-09-18 15:01:56,515-log_helper.py# 77] [INFO] Progress: 1280 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:15 (D:H:M) 199 | 200 | [2019-09-18 15:02:03,596-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1300/186884] step_time: 0.358074 (0.339724) train_loss: 0.429435 (0.627243) test_loss: 0.689146 (0.713027) train_metric: 0.839271 (0.727396) 201 | [2019-09-18 15:02:03,597-log_helper.py# 77] [INFO] Progress: 1300 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:40 (D:H:M) 202 | 203 | [2019-09-18 15:02:10,607-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1320/186884] step_time: 0.325676 (0.339829) train_loss: 0.663904 (0.627961) test_loss: 0.660934 (0.712561) train_metric: 0.833197 (0.728431) 204 | [2019-09-18 15:02:10,607-log_helper.py# 77] [INFO] Progress: 1320 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:56 (D:H:M) 205 | 206 | [2019-09-18 15:02:17,279-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1340/186884] step_time: 0.326099 (0.339680) train_loss: 0.672247 (0.626640) test_loss: 0.690298 (0.712100) train_metric: 0.900709 (0.729749) 207 | [2019-09-18 15:02:17,279-log_helper.py# 77] [INFO] Progress: 1340 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:33 (D:H:M) 208 | 209 | [2019-09-18 15:02:23,990-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1360/186884] step_time: 0.331624 (0.339563) train_loss: 0.675758 (0.625801) test_loss: 0.697068 (0.711678) train_metric: 0.858995 (0.730934) 210 | [2019-09-18 15:02:23,990-log_helper.py# 77] [INFO] Progress: 1360 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:14 (D:H:M) 211 | 212 | [2019-09-18 15:02:30,869-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1380/186884] step_time: 0.326340 (0.339570) train_loss: 0.680005 (0.625467) test_loss: 0.688249 (0.711285) train_metric: 0.876767 (0.731401) 213 | [2019-09-18 15:02:30,869-log_helper.py# 77] [INFO] Progress: 1380 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:15 (D:H:M) 214 | 215 | [2019-09-18 15:02:38,108-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1400/186884] step_time: 0.326065 (0.339700) train_loss: 0.681127 (0.625200) test_loss: 0.656816 (0.710938) train_metric: 0.657701 (0.732169) 216 | [2019-09-18 15:02:38,109-log_helper.py# 77] [INFO] Progress: 1400 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:35 (D:H:M) 217 | 218 | [2019-09-18 15:02:45,164-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1420/186884] step_time: 0.334243 (0.339827) train_loss: 0.682113 (0.624529) test_loss: 0.689582 (0.710516) train_metric: 0.752131 (0.733044) 219 | [2019-09-18 15:02:45,165-log_helper.py# 77] [INFO] Progress: 1420 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:55 (D:H:M) 220 | 221 | [2019-09-18 15:02:52,303-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1440/186884] step_time: 0.340621 (0.340008) train_loss: 0.687102 (0.624059) test_loss: 0.710522 (0.710150) train_metric: 0.753485 (0.733835) 222 | [2019-09-18 15:02:52,304-log_helper.py# 77] [INFO] Progress: 1440 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:18:23 (D:H:M) 223 | 224 | [2019-09-18 15:02:59,411-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1460/186884] step_time: 0.372243 (0.340167) train_loss: 0.332800 (0.622785) test_loss: 0.664642 (0.709726) train_metric: 0.887718 (0.734974) 225 | [2019-09-18 15:02:59,412-log_helper.py# 77] [INFO] Progress: 1460 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:18:48 (D:H:M) 226 | 227 | [2019-09-18 15:03:06,598-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1480/186884] step_time: 0.357926 (0.340370) train_loss: 0.689430 (0.622402) test_loss: 0.680774 (0.709453) train_metric: 0.778945 (0.735702) 228 | [2019-09-18 15:03:06,598-log_helper.py# 77] [INFO] Progress: 1480 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:19:19 (D:H:M) 229 | 230 | [2019-09-18 15:03:13,909-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1500/186884] step_time: 0.339111 (0.340653) train_loss: 0.675773 (0.621979) test_loss: 0.667480 (0.709063) train_metric: 0.880763 (0.736269) 231 | [2019-09-18 15:03:13,909-log_helper.py# 77] [INFO] Progress: 1500 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:20:03 (D:H:M) 232 | 233 | [2019-09-18 15:03:21,057-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1520/186884] step_time: 0.364453 (0.340820) train_loss: 0.672631 (0.622101) test_loss: 0.674089 (0.708602) train_metric: 0.831880 (0.736987) 234 | [2019-09-18 15:03:21,057-log_helper.py# 77] [INFO] Progress: 1520 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:20:29 (D:H:M) 235 | 236 | [2019-09-18 15:03:28,037-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1540/186884] step_time: 0.352865 (0.340879) train_loss: 0.689946 (0.621799) test_loss: 0.699086 (0.708267) train_metric: 0.653978 (0.737975) 237 | [2019-09-18 15:03:28,037-log_helper.py# 77] [INFO] Progress: 1540 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:20:38 (D:H:M) 238 | 239 | [2019-09-18 15:03:35,172-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1560/186884] step_time: 0.331213 (0.341029) train_loss: 0.421284 (0.620587) test_loss: 0.663021 (0.707866) train_metric: 0.763609 (0.738657) 240 | [2019-09-18 15:03:35,173-log_helper.py# 77] [INFO] Progress: 1560 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:01 (D:H:M) 241 | 242 | [2019-09-18 15:03:41,732-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1580/186884] step_time: 0.354648 (0.340815) train_loss: 0.668576 (0.619944) test_loss: 0.687252 (0.707630) train_metric: 0.873309 (0.739442) 243 | [2019-09-18 15:03:41,732-log_helper.py# 77] [INFO] Progress: 1580 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:20:28 (D:H:M) 244 | 245 | [2019-09-18 15:03:48,438-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1600/186884] step_time: 0.320349 (0.340596) train_loss: 0.297780 (0.619581) test_loss: 0.655914 (0.707292) train_metric: 0.912116 (0.740484) 246 | [2019-09-18 15:03:48,438-log_helper.py# 77] [INFO] Progress: 1600 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:19:54 (D:H:M) 247 | 248 | [2019-09-18 15:03:55,007-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1620/186884] step_time: 0.320436 (0.340397) train_loss: 0.323908 (0.619266) test_loss: 0.692299 (0.706987) train_metric: 0.837286 (0.741165) 249 | [2019-09-18 15:03:55,008-log_helper.py# 77] [INFO] Progress: 1620 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:19:23 (D:H:M) 250 | 251 | [2019-09-18 15:04:01,695-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1640/186884] step_time: 0.346338 (0.340280) train_loss: 0.308891 (0.619155) test_loss: 0.683858 (0.706657) train_metric: 0.870512 (0.741726) 252 | [2019-09-18 15:04:01,695-log_helper.py# 77] [INFO] Progress: 1640 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:19:04 (D:H:M) 253 | 254 | [2019-09-18 15:04:08,740-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1660/186884] step_time: 0.330658 (0.340381) train_loss: 0.421766 (0.618985) test_loss: 0.689872 (0.706273) train_metric: 0.856727 (0.742567) 255 | [2019-09-18 15:04:08,741-log_helper.py# 77] [INFO] Progress: 1660 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:19:20 (D:H:M) 256 | 257 | [2019-09-18 15:04:15,354-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1680/186884] step_time: 0.321047 (0.340220) train_loss: 0.653995 (0.618826) test_loss: 0.682240 (0.705931) train_metric: 0.851771 (0.743323) 258 | [2019-09-18 15:04:15,354-log_helper.py# 77] [INFO] Progress: 1680 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:18:55 (D:H:M) 259 | 260 | [2019-09-18 15:04:21,918-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1700/186884] step_time: 0.320628 (0.340035) train_loss: 0.642566 (0.617922) test_loss: 0.700911 (0.705664) train_metric: 0.871956 (0.744413) 261 | [2019-09-18 15:04:21,918-log_helper.py# 77] [INFO] Progress: 1700 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:18:26 (D:H:M) 262 | 263 | [2019-09-18 15:04:28,429-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1720/186884] step_time: 0.337839 (0.339823) train_loss: 0.660722 (0.617243) test_loss: 0.662977 (0.705349) train_metric: 0.861413 (0.745253) 264 | [2019-09-18 15:04:28,429-log_helper.py# 77] [INFO] Progress: 1720 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:53 (D:H:M) 265 | 266 | [2019-09-18 15:04:35,478-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1740/186884] step_time: 0.360383 (0.339923) train_loss: 0.691947 (0.616715) test_loss: 0.666526 (0.705011) train_metric: 0.654306 (0.745684) 267 | [2019-09-18 15:04:35,478-log_helper.py# 77] [INFO] Progress: 1740 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:18:08 (D:H:M) 268 | 269 | [2019-09-18 15:04:42,389-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1760/186884] step_time: 0.337389 (0.339943) train_loss: 0.684209 (0.616016) test_loss: 0.667292 (0.704714) train_metric: 0.769740 (0.746238) 270 | [2019-09-18 15:04:42,389-log_helper.py# 77] [INFO] Progress: 1760 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:18:11 (D:H:M) 271 | 272 | [2019-09-18 15:04:49,033-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1780/186884] step_time: 0.323096 (0.339814) train_loss: 0.428084 (0.615242) test_loss: 0.686182 (0.704434) train_metric: 0.861977 (0.747294) 273 | [2019-09-18 15:04:49,033-log_helper.py# 77] [INFO] Progress: 1780 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:51 (D:H:M) 274 | 275 | [2019-09-18 15:04:56,130-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1800/186884] step_time: 0.399143 (0.339826) train_loss: 0.433885 (0.614590) test_loss: 0.699783 (0.704149) train_metric: 0.835313 (0.747975) 276 | [2019-09-18 15:04:56,131-log_helper.py# 77] [INFO] Progress: 1800 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:53 (D:H:M) 277 | 278 | [2019-09-18 15:05:03,284-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1820/186884] step_time: 0.352691 (0.339977) train_loss: 0.686760 (0.614240) test_loss: 0.693115 (0.703939) train_metric: 0.753715 (0.748748) 279 | [2019-09-18 15:05:03,284-log_helper.py# 77] [INFO] Progress: 1820 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:18:16 (D:H:M) 280 | 281 | [2019-09-18 15:05:09,916-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1840/186884] step_time: 0.320302 (0.339843) train_loss: 0.320792 (0.613419) test_loss: 0.664572 (0.703682) train_metric: 0.801132 (0.749438) 282 | [2019-09-18 15:05:09,916-log_helper.py# 77] [INFO] Progress: 1840 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:55 (D:H:M) 283 | 284 | [2019-09-18 15:05:16,468-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1860/186884] step_time: 0.324326 (0.339670) train_loss: 0.691821 (0.612435) test_loss: 0.693085 (0.703513) train_metric: 0.733716 (0.750076) 285 | [2019-09-18 15:05:16,468-log_helper.py# 77] [INFO] Progress: 1860 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:28 (D:H:M) 286 | 287 | [2019-09-18 15:05:23,029-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1880/186884] step_time: 0.320946 (0.339505) train_loss: 0.663650 (0.612264) test_loss: 0.654045 (0.703229) train_metric: 0.852761 (0.750837) 288 | [2019-09-18 15:05:23,029-log_helper.py# 77] [INFO] Progress: 1880 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:02 (D:H:M) 289 | 290 | [2019-09-18 15:05:29,588-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1900/186884] step_time: 0.320462 (0.339342) train_loss: 0.658305 (0.611881) test_loss: 0.702369 (0.702931) train_metric: 0.797013 (0.751270) 291 | [2019-09-18 15:05:29,589-log_helper.py# 77] [INFO] Progress: 1900 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:37 (D:H:M) 292 | 293 | [2019-09-18 15:05:36,487-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1920/186884] step_time: 0.337426 (0.339359) train_loss: 0.710683 (0.611742) test_loss: 0.692885 (0.703190) train_metric: 0.743645 (0.751824) 294 | [2019-09-18 15:05:36,487-log_helper.py# 77] [INFO] Progress: 1920 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:39 (D:H:M) 295 | 296 | [2019-09-18 15:05:43,239-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1940/186884] step_time: 0.334291 (0.339302) train_loss: 0.398263 (0.612076) test_loss: 0.708589 (0.702965) train_metric: 0.812421 (0.752197) 297 | [2019-09-18 15:05:43,240-log_helper.py# 77] [INFO] Progress: 1940 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:30 (D:H:M) 298 | 299 | [2019-09-18 15:05:50,051-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1960/186884] step_time: 0.330553 (0.339276) train_loss: 0.676461 (0.612478) test_loss: 0.666192 (0.702931) train_metric: 0.723193 (0.751513) 300 | [2019-09-18 15:05:50,051-log_helper.py# 77] [INFO] Progress: 1960 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:26 (D:H:M) 301 | 302 | [2019-09-18 15:05:56,853-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][1980/186884] step_time: 0.321968 (0.339248) train_loss: 0.681834 (0.612052) test_loss: 1.068004 (0.703063) train_metric: 0.803224 (0.751407) 303 | [2019-09-18 15:05:56,854-log_helper.py# 77] [INFO] Progress: 1980 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:22 (D:H:M) 304 | 305 | [2019-09-18 15:06:04,026-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2000/186884] step_time: 0.382112 (0.339322) train_loss: 0.473564 (0.612679) test_loss: 0.763582 (0.703435) train_metric: 0.779059 (0.751138) 306 | [2019-09-18 15:06:04,026-log_helper.py# 77] [INFO] Progress: 2000 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:33 (D:H:M) 307 | 308 | [2019-09-18 15:06:10,906-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2020/186884] step_time: 0.330272 (0.339332) train_loss: 0.724166 (0.613190) test_loss: 0.745659 (0.703486) train_metric: 0.752029 (0.750356) 309 | [2019-09-18 15:06:10,907-log_helper.py# 77] [INFO] Progress: 2020 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:35 (D:H:M) 310 | 311 | [2019-09-18 15:06:17,700-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2040/186884] step_time: 0.328058 (0.339296) train_loss: 0.733606 (0.613360) test_loss: 0.718025 (0.703615) train_metric: 0.641657 (0.749719) 312 | [2019-09-18 15:06:17,700-log_helper.py# 77] [INFO] Progress: 2040 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:29 (D:H:M) 313 | 314 | [2019-09-18 15:06:24,404-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2060/186884] step_time: 0.327595 (0.339220) train_loss: 0.670299 (0.613747) test_loss: 0.679352 (0.703701) train_metric: 0.797192 (0.749082) 315 | [2019-09-18 15:06:24,404-log_helper.py# 77] [INFO] Progress: 2060 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:17 (D:H:M) 316 | 317 | [2019-09-18 15:06:31,134-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2080/186884] step_time: 0.330226 (0.339156) train_loss: 0.331636 (0.613466) test_loss: 0.696010 (0.703749) train_metric: 0.756392 (0.748847) 318 | [2019-09-18 15:06:31,135-log_helper.py# 77] [INFO] Progress: 2080 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:07 (D:H:M) 319 | 320 | [2019-09-18 15:06:37,851-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2100/186884] step_time: 0.333279 (0.339089) train_loss: 0.679610 (0.613522) test_loss: 0.671792 (0.703519) train_metric: 0.714956 (0.748960) 321 | [2019-09-18 15:06:37,851-log_helper.py# 77] [INFO] Progress: 2100 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:15:56 (D:H:M) 322 | 323 | [2019-09-18 15:06:44,587-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2120/186884] step_time: 0.331695 (0.339032) train_loss: 0.637338 (0.613062) test_loss: 0.691025 (0.703553) train_metric: 0.728809 (0.749325) 324 | [2019-09-18 15:06:44,587-log_helper.py# 77] [INFO] Progress: 2120 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:15:47 (D:H:M) 325 | 326 | [2019-09-18 15:06:51,503-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2140/186884] step_time: 0.358788 (0.339058) train_loss: 0.667919 (0.612593) test_loss: 0.700078 (0.703625) train_metric: 0.773856 (0.749447) 327 | [2019-09-18 15:06:51,504-log_helper.py# 77] [INFO] Progress: 2140 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:15:51 (D:H:M) 328 | 329 | [2019-09-18 15:06:58,600-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2160/186884] step_time: 0.349786 (0.339165) train_loss: 0.696982 (0.613101) test_loss: 0.693890 (0.703417) train_metric: 0.766265 (0.749118) 330 | [2019-09-18 15:06:58,600-log_helper.py# 77] [INFO] Progress: 2160 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:08 (D:H:M) 331 | 332 | [2019-09-18 15:07:05,616-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2180/186884] step_time: 0.331141 (0.339235) train_loss: 0.677303 (0.613621) test_loss: 0.685683 (0.703517) train_metric: 0.818615 (0.748990) 333 | [2019-09-18 15:07:05,617-log_helper.py# 77] [INFO] Progress: 2180 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:18 (D:H:M) 334 | 335 | [2019-09-18 15:07:12,825-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2200/186884] step_time: 0.356738 (0.339341) train_loss: 0.711069 (0.613922) test_loss: 0.700337 (0.703496) train_metric: 0.720510 (0.749076) 336 | [2019-09-18 15:07:13,117-log_helper.py# 77] [INFO] Progress: 2200 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:35 (D:H:M) 337 | 338 | [2019-09-18 15:07:20,139-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2220/186884] step_time: 0.368811 (0.339377) train_loss: 0.709340 (0.614340) test_loss: 0.703405 (0.703536) train_metric: 0.599064 (0.749145) 339 | [2019-09-18 15:07:20,140-log_helper.py# 77] [INFO] Progress: 2220 / 9344200 [0%], Speed: 0.339 s/iter, ETA 36:16:40 (D:H:M) 340 | 341 | [2019-09-18 15:07:27,441-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2240/186884] step_time: 0.367907 (0.339569) train_loss: 0.677962 (0.614681) test_loss: 0.668346 (0.703465) train_metric: 0.734473 (0.748942) 342 | [2019-09-18 15:07:27,441-log_helper.py# 77] [INFO] Progress: 2240 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:10 (D:H:M) 343 | 344 | [2019-09-18 15:07:34,756-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2260/186884] step_time: 0.365262 (0.339764) train_loss: 0.491446 (0.614614) test_loss: 0.759972 (0.703481) train_metric: 0.642830 (0.748549) 345 | [2019-09-18 15:07:34,756-log_helper.py# 77] [INFO] Progress: 2260 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:17:40 (D:H:M) 346 | 347 | [2019-09-18 15:07:41,996-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2280/186884] step_time: 0.346244 (0.339922) train_loss: 0.546154 (0.614915) test_loss: 0.710608 (0.703990) train_metric: 0.577262 (0.747637) 348 | [2019-09-18 15:07:41,996-log_helper.py# 77] [INFO] Progress: 2280 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:18:05 (D:H:M) 349 | 350 | [2019-09-18 15:07:48,968-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2300/186884] step_time: 0.347959 (0.339963) train_loss: 0.366056 (0.614655) test_loss: 0.774280 (0.704509) train_metric: 0.771952 (0.747273) 351 | [2019-09-18 15:07:48,969-log_helper.py# 77] [INFO] Progress: 2300 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:18:11 (D:H:M) 352 | 353 | [2019-09-18 15:07:55,979-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2320/186884] step_time: 0.349692 (0.340017) train_loss: 0.691247 (0.615118) test_loss: 0.689029 (0.704589) train_metric: 0.766853 (0.746675) 354 | [2019-09-18 15:07:55,979-log_helper.py# 77] [INFO] Progress: 2320 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:18:19 (D:H:M) 355 | 356 | [2019-09-18 15:08:03,107-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2340/186884] step_time: 0.354638 (0.340124) train_loss: 0.683198 (0.615275) test_loss: 0.669649 (0.704653) train_metric: 0.774161 (0.746473) 357 | [2019-09-18 15:08:03,107-log_helper.py# 77] [INFO] Progress: 2340 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:18:36 (D:H:M) 358 | 359 | [2019-09-18 15:08:10,421-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2360/186884] step_time: 0.350204 (0.340305) train_loss: 0.678708 (0.615543) test_loss: 0.690328 (0.705077) train_metric: 0.756995 (0.746159) 360 | [2019-09-18 15:08:10,421-log_helper.py# 77] [INFO] Progress: 2360 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:19:04 (D:H:M) 361 | 362 | [2019-09-18 15:08:17,782-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2380/186884] step_time: 0.357844 (0.340499) train_loss: 0.697740 (0.615646) test_loss: 0.702461 (0.705154) train_metric: 0.680055 (0.745879) 363 | [2019-09-18 15:08:17,782-log_helper.py# 77] [INFO] Progress: 2380 / 9344200 [0%], Speed: 0.340 s/iter, ETA 36:19:34 (D:H:M) 364 | 365 | [2019-09-18 15:08:25,197-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2400/186884] step_time: 0.346995 (0.340639) train_loss: 0.672206 (0.615764) test_loss: 0.669343 (0.705044) train_metric: 0.814767 (0.745863) 366 | [2019-09-18 15:08:25,198-log_helper.py# 77] [INFO] Progress: 2400 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:19:56 (D:H:M) 367 | 368 | [2019-09-18 15:08:32,176-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2420/186884] step_time: 0.349961 (0.340675) train_loss: 0.554206 (0.615945) test_loss: 0.691742 (0.705315) train_metric: 0.636597 (0.745688) 369 | [2019-09-18 15:08:32,176-log_helper.py# 77] [INFO] Progress: 2420 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:20:01 (D:H:M) 370 | 371 | [2019-09-18 15:08:39,431-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2440/186884] step_time: 0.367389 (0.340824) train_loss: 0.659442 (0.615977) test_loss: 0.665414 (0.705454) train_metric: 0.765432 (0.745627) 372 | [2019-09-18 15:08:39,431-log_helper.py# 77] [INFO] Progress: 2440 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:20:24 (D:H:M) 373 | 374 | [2019-09-18 15:08:46,602-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2460/186884] step_time: 0.351975 (0.340936) train_loss: 0.695660 (0.616146) test_loss: 0.708749 (0.705628) train_metric: 0.702448 (0.745503) 375 | [2019-09-18 15:08:46,602-log_helper.py# 77] [INFO] Progress: 2460 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:20:42 (D:H:M) 376 | 377 | [2019-09-18 15:08:53,788-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2480/186884] step_time: 0.353857 (0.341053) train_loss: 0.678941 (0.616026) test_loss: 0.738833 (0.705920) train_metric: 0.688962 (0.745352) 378 | [2019-09-18 15:08:53,788-log_helper.py# 77] [INFO] Progress: 2480 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:00 (D:H:M) 379 | 380 | [2019-09-18 15:09:00,935-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2500/186884] step_time: 0.349749 (0.341154) train_loss: 0.698166 (0.615331) test_loss: 0.745251 (0.706057) train_metric: 0.807063 (0.745577) 381 | [2019-09-18 15:09:00,935-log_helper.py# 77] [INFO] Progress: 2500 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:15 (D:H:M) 382 | 383 | [2019-09-18 15:09:08,045-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2520/186884] step_time: 0.350798 (0.341237) train_loss: 0.652172 (0.615233) test_loss: 0.655154 (0.706085) train_metric: 0.809484 (0.745856) 384 | [2019-09-18 15:09:08,045-log_helper.py# 77] [INFO] Progress: 2520 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:28 (D:H:M) 385 | 386 | [2019-09-18 15:09:15,190-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2540/186884] step_time: 0.349326 (0.341333) train_loss: 0.652715 (0.615375) test_loss: 0.739884 (0.705992) train_metric: 0.794064 (0.745953) 387 | [2019-09-18 15:09:15,190-log_helper.py# 77] [INFO] Progress: 2540 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:43 (D:H:M) 388 | 389 | [2019-09-18 15:09:22,290-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2560/186884] step_time: 0.350563 (0.341409) train_loss: 0.691042 (0.615441) test_loss: 0.689507 (0.705977) train_metric: 0.713689 (0.745936) 390 | [2019-09-18 15:09:22,290-log_helper.py# 77] [INFO] Progress: 2560 / 9344200 [0%], Speed: 0.341 s/iter, ETA 36:21:55 (D:H:M) 391 | 392 | [2019-09-18 15:09:29,441-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2580/186884] step_time: 0.354080 (0.341505) train_loss: 0.510097 (0.615288) test_loss: 0.674782 (0.705948) train_metric: 0.678373 (0.745923) 393 | [2019-09-18 15:09:29,442-log_helper.py# 77] [INFO] Progress: 2580 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:22:10 (D:H:M) 394 | 395 | [2019-09-18 15:09:36,771-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2600/186884] step_time: 0.352740 (0.341617) train_loss: 0.650520 (0.615459) test_loss: 0.685542 (0.705881) train_metric: 0.789629 (0.745972) 396 | [2019-09-18 15:09:37,023-log_helper.py# 77] [INFO] Progress: 2600 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:22:27 (D:H:M) 397 | 398 | [2019-09-18 15:09:44,190-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2620/186884] step_time: 0.349205 (0.341681) train_loss: 0.286523 (0.615136) test_loss: 0.672264 (0.705753) train_metric: 0.789198 (0.746104) 399 | [2019-09-18 15:09:44,190-log_helper.py# 77] [INFO] Progress: 2620 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:22:37 (D:H:M) 400 | 401 | [2019-09-18 15:09:51,279-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2640/186884] step_time: 0.339761 (0.341748) train_loss: 0.662067 (0.615319) test_loss: 0.688089 (0.705590) train_metric: 0.869910 (0.746229) 402 | [2019-09-18 15:09:51,279-log_helper.py# 77] [INFO] Progress: 2640 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:22:47 (D:H:M) 403 | 404 | [2019-09-18 15:09:58,353-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2660/186884] step_time: 0.348954 (0.341809) train_loss: 0.693683 (0.614916) test_loss: 0.664491 (0.705599) train_metric: 0.740675 (0.746525) 405 | [2019-09-18 15:09:58,353-log_helper.py# 77] [INFO] Progress: 2660 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:22:57 (D:H:M) 406 | 407 | [2019-09-18 15:10:05,556-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2680/186884] step_time: 0.340809 (0.341915) train_loss: 0.688855 (0.614764) test_loss: 0.654023 (0.705928) train_metric: 0.741052 (0.746741) 408 | [2019-09-18 15:10:05,557-log_helper.py# 77] [INFO] Progress: 2680 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:23:13 (D:H:M) 409 | 410 | [2019-09-18 15:10:12,493-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2700/186884] step_time: 0.336707 (0.341923) train_loss: 0.676316 (0.614594) test_loss: 0.660201 (0.706302) train_metric: 0.717188 (0.746982) 411 | [2019-09-18 15:10:12,494-log_helper.py# 77] [INFO] Progress: 2700 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:23:14 (D:H:M) 412 | 413 | [2019-09-18 15:10:19,717-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2720/186884] step_time: 0.353641 (0.342035) train_loss: 0.669305 (0.614595) test_loss: 0.660251 (0.706158) train_metric: 0.866923 (0.747219) 414 | [2019-09-18 15:10:19,717-log_helper.py# 77] [INFO] Progress: 2720 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:23:31 (D:H:M) 415 | 416 | [2019-09-18 15:10:26,876-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2740/186884] step_time: 0.352218 (0.342123) train_loss: 0.669822 (0.614284) test_loss: 0.694881 (0.705930) train_metric: 0.853308 (0.747622) 417 | [2019-09-18 15:10:26,876-log_helper.py# 77] [INFO] Progress: 2740 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:23:45 (D:H:M) 418 | 419 | [2019-09-18 15:10:33,931-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2760/186884] step_time: 0.349059 (0.342172) train_loss: 0.693378 (0.614601) test_loss: 0.689324 (0.705747) train_metric: 0.729847 (0.747723) 420 | [2019-09-18 15:10:33,931-log_helper.py# 77] [INFO] Progress: 2760 / 9344200 [0%], Speed: 0.342 s/iter, ETA 36:23:52 (D:H:M) 421 | 422 | [2019-09-18 15:10:40,994-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2780/186884] step_time: 0.359815 (0.342223) train_loss: 0.458687 (0.614783) test_loss: 0.685552 (0.705643) train_metric: 0.700083 (0.747831) 423 | [2019-09-18 15:10:40,995-log_helper.py# 77] [INFO] Progress: 2780 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:00 (D:H:M) 424 | 425 | [2019-09-18 15:10:48,134-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2800/186884] step_time: 0.339121 (0.342246) train_loss: 0.700776 (0.614791) test_loss: 0.668141 (0.705569) train_metric: 0.679171 (0.748081) 426 | [2019-09-18 15:10:48,145-log_helper.py# 77] [INFO] Progress: 2800 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:04 (D:H:M) 427 | 428 | [2019-09-18 15:10:55,116-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2820/186884] step_time: 0.346200 (0.342260) train_loss: 0.413118 (0.614871) test_loss: 0.699926 (0.705413) train_metric: 0.766840 (0.748272) 429 | [2019-09-18 15:10:55,116-log_helper.py# 77] [INFO] Progress: 2820 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:06 (D:H:M) 430 | 431 | [2019-09-18 15:11:02,385-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2840/186884] step_time: 0.341805 (0.342382) train_loss: 0.696310 (0.614687) test_loss: 0.711335 (0.705352) train_metric: 0.682346 (0.748679) 432 | [2019-09-18 15:11:02,385-log_helper.py# 77] [INFO] Progress: 2840 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:25 (D:H:M) 433 | 434 | [2019-09-18 15:11:09,309-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2860/186884] step_time: 0.340345 (0.342380) train_loss: 0.437327 (0.614568) test_loss: 0.688568 (0.705255) train_metric: 0.783514 (0.748995) 435 | [2019-09-18 15:11:09,309-log_helper.py# 77] [INFO] Progress: 2860 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:24 (D:H:M) 436 | 437 | [2019-09-18 15:11:16,226-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2880/186884] step_time: 0.339849 (0.342378) train_loss: 0.663737 (0.614427) test_loss: 0.711036 (0.705127) train_metric: 0.877715 (0.749470) 438 | [2019-09-18 15:11:16,226-log_helper.py# 77] [INFO] Progress: 2880 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:24 (D:H:M) 439 | 440 | [2019-09-18 15:11:23,179-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2900/186884] step_time: 0.341311 (0.342386) train_loss: 0.661277 (0.614378) test_loss: 0.669528 (0.704976) train_metric: 0.802307 (0.749824) 441 | [2019-09-18 15:11:23,180-log_helper.py# 77] [INFO] Progress: 2900 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:25 (D:H:M) 442 | 443 | [2019-09-18 15:11:30,154-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2920/186884] step_time: 0.344871 (0.342400) train_loss: 0.689093 (0.614099) test_loss: 0.646674 (0.704772) train_metric: 0.792014 (0.750070) 444 | [2019-09-18 15:11:30,155-log_helper.py# 77] [INFO] Progress: 2920 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:27 (D:H:M) 445 | 446 | [2019-09-18 15:11:37,073-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2940/186884] step_time: 0.342510 (0.342397) train_loss: 0.680970 (0.613873) test_loss: 0.692124 (0.704620) train_metric: 0.780262 (0.750484) 447 | [2019-09-18 15:11:37,073-log_helper.py# 77] [INFO] Progress: 2940 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:26 (D:H:M) 448 | 449 | [2019-09-18 15:11:44,069-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2960/186884] step_time: 0.344117 (0.342418) train_loss: 0.669855 (0.613792) test_loss: 0.688311 (0.704455) train_metric: 0.882797 (0.750667) 450 | [2019-09-18 15:11:44,070-log_helper.py# 77] [INFO] Progress: 2960 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:30 (D:H:M) 451 | 452 | [2019-09-18 15:11:51,008-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][2980/186884] step_time: 0.342722 (0.342421) train_loss: 0.682214 (0.613727) test_loss: 0.700667 (0.704310) train_metric: 0.840276 (0.751151) 453 | [2019-09-18 15:11:51,008-log_helper.py# 77] [INFO] Progress: 2980 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:30 (D:H:M) 454 | 455 | [2019-09-18 15:11:59,134-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][3000/186884] step_time: 0.356227 (0.342476) train_loss: 0.636109 (0.613467) test_loss: 0.691391 (0.704145) train_metric: 0.794771 (0.751446) 456 | [2019-09-18 15:11:59,134-log_helper.py# 77] [INFO] Progress: 3000 / 9344200 [0%], Speed: 0.342 s/iter, ETA 37:00:38 (D:H:M) 457 | 458 | [2019-09-18 15:12:06,461-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][3020/186884] step_time: 0.347317 (0.342602) train_loss: 0.705074 (0.612945) test_loss: 0.686337 (0.704042) train_metric: 0.559816 (0.751756) 459 | [2019-09-18 15:12:06,461-log_helper.py# 77] [INFO] Progress: 3020 / 9344200 [0%], Speed: 0.343 s/iter, ETA 37:00:58 (D:H:M) 460 | 461 | [2019-09-18 15:12:13,634-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][3040/186884] step_time: 0.356799 (0.342682) train_loss: 0.661920 (0.613174) test_loss: 0.691357 (0.703928) train_metric: 0.911168 (0.752024) 462 | [2019-09-18 15:12:13,634-log_helper.py# 77] [INFO] Progress: 3040 / 9344200 [0%], Speed: 0.343 s/iter, ETA 37:01:10 (D:H:M) 463 | 464 | [2019-09-18 15:12:20,811-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][3060/186884] step_time: 0.352016 (0.342759) train_loss: 0.686521 (0.612864) test_loss: 0.707303 (0.703814) train_metric: 0.866792 (0.752292) 465 | [2019-09-18 15:12:20,811-log_helper.py# 77] [INFO] Progress: 3060 / 9344200 [0%], Speed: 0.343 s/iter, ETA 37:01:22 (D:H:M) 466 | 467 | [2019-09-18 15:12:27,998-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][3080/186884] step_time: 0.360009 (0.342837) train_loss: 0.283406 (0.612936) test_loss: 0.692716 (0.703722) train_metric: 0.867087 (0.752478) 468 | [2019-09-18 15:12:27,998-log_helper.py# 77] [INFO] Progress: 3080 / 9344200 [0%], Speed: 0.343 s/iter, ETA 37:01:34 (D:H:M) 469 | 470 | [2019-09-18 15:12:35,198-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][3100/186884] step_time: 0.371403 (0.342920) train_loss: 0.683332 (0.612833) test_loss: 0.692144 (0.703560) train_metric: 0.790639 (0.752734) 471 | [2019-09-18 15:12:35,198-log_helper.py# 77] [INFO] Progress: 3100 / 9344200 [0%], Speed: 0.343 s/iter, ETA 37:01:47 (D:H:M) 472 | 473 | [2019-09-18 15:12:42,497-train_video_prediction_KTHDataset.py#186] [INFO] Epoch: [1][3120/186884] step_time: 0.352279 (0.343028) train_loss: 0.669642 (0.612649) test_loss: 0.682211 (0.703420) train_metric: 0.828880 (0.752999) 474 | [2019-09-18 15:12:42,498-log_helper.py# 77] [INFO] Progress: 3120 / 9344200 [0%], Speed: 0.343 s/iter, ETA 37:02:04 (D:H:M) 475 | 476 | -------------------------------------------------------------------------------- /logs/video_prediction_demo.log: -------------------------------------------------------------------------------- 1 | [2019-09-16 22:51:45,904-config_helper.py# 61] [DEBUG] Used config: 2 | {'meta': {'experiment_path': 'experiments', 'arch': 'Custom', 'board_path': 'board', 'experiment_name': 'video_prediction_demo'}, 'train': {'epoches': 50, 'batch_size': 2, 'lr': 1e-05, 'print_freq': 20}, 'model': {'input_size': [64, 64], 'input_dim': 1, 'hidden_dim': 64, 'kernel_size': [3, 3], 'input_num': 10}} 3 | [2019-09-16 22:51:46,405-train_video_prediction_MNIST.py# 86] [DEBUG] ==>>> total trainning batch number: 4500 4 | [2019-09-16 22:51:46,405-train_video_prediction_MNIST.py# 87] [DEBUG] ==>>> total testing batch number: 500 5 | [2019-09-16 22:51:47,746-train_video_prediction_MNIST.py#122] [DEBUG] Input: torch.Size([2, 10, 1, 64, 64]) 6 | [2019-09-16 22:51:47,746-train_video_prediction_MNIST.py#123] [DEBUG] --- Sample 7 | [2019-09-16 22:51:47,746-train_video_prediction_MNIST.py#124] [DEBUG] Target: torch.Size([2, 10, 1, 64, 64]) 8 | [2019-09-16 22:51:53,019-train_video_prediction_MNIST.py#175] [INFO] Epoch: [1][20/4500] step_time: 0.085659 (0.176182) train_loss: 0.001326 (0.001210) test_loss: 0.722058 (0.660619) 9 | [2019-09-16 22:51:53,019-log_helper.py# 77] [INFO] Progress: 20 / 225000 [0%], Speed: 0.176 s/iter, ETA 0:11:00 (D:H:M) 10 | 11 | [2019-09-16 22:51:55,188-train_video_prediction_MNIST.py#175] [INFO] Epoch: [1][40/4500] step_time: 0.083421 (0.132536) train_loss: 0.000929 (0.001134) test_loss: 0.651724 (0.660121) 12 | [2019-09-16 22:51:55,198-log_helper.py# 77] [INFO] Progress: 40 / 225000 [0%], Speed: 0.133 s/iter, ETA 0:08:16 (D:H:M) 13 | 14 | [2019-09-16 22:51:57,431-train_video_prediction_MNIST.py#175] [INFO] Epoch: [1][60/4500] step_time: 0.091918 (0.118670) train_loss: 0.000937 (0.001083) test_loss: 0.643997 (0.663288) 15 | [2019-09-16 22:51:57,441-log_helper.py# 77] [INFO] Progress: 60 / 225000 [0%], Speed: 0.119 s/iter, ETA 0:07:24 (D:H:M) 16 | 17 | [2019-09-16 22:51:59,710-train_video_prediction_MNIST.py#175] [INFO] Epoch: [1][80/4500] step_time: 0.092299 (0.111957) train_loss: 0.001022 (0.001052) test_loss: 0.651355 (0.661283) 18 | [2019-09-16 22:51:59,720-log_helper.py# 77] [INFO] Progress: 80 / 225000 [0%], Speed: 0.112 s/iter, ETA 0:06:59 (D:H:M) 19 | 20 | [2019-09-16 22:52:01,975-train_video_prediction_MNIST.py#175] [INFO] Epoch: [1][100/4500] step_time: 0.091466 (0.107946) train_loss: 0.001237 (0.001037) test_loss: 0.650562 (0.663106) 21 | [2019-09-16 22:52:01,986-log_helper.py# 77] [INFO] Progress: 100 / 225000 [0%], Speed: 0.108 s/iter, ETA 0:06:44 (D:H:M) 22 | 23 | [2019-09-16 22:52:04,247-train_video_prediction_MNIST.py#175] [INFO] Epoch: [1][120/4500] step_time: 0.090912 (0.105267) train_loss: 0.000851 (0.001020) test_loss: 0.644908 (0.663379) 24 | [2019-09-16 22:52:04,247-log_helper.py# 77] [INFO] Progress: 120 / 225000 [0%], Speed: 0.105 s/iter, ETA 0:06:34 (D:H:M) 25 | 26 | [2019-09-16 22:52:06,537-train_video_prediction_MNIST.py#175] [INFO] Epoch: [1][140/4500] step_time: 0.093379 (0.103381) train_loss: 0.000965 (0.001004) test_loss: 0.654391 (0.661926) 27 | [2019-09-16 22:52:06,537-log_helper.py# 77] [INFO] Progress: 140 / 225000 [0%], Speed: 0.103 s/iter, ETA 0:06:27 (D:H:M) 28 | 29 | [2019-09-16 22:52:08,817-train_video_prediction_MNIST.py#175] [INFO] Epoch: [1][160/4500] step_time: 0.090414 (0.102013) train_loss: 0.000988 (0.000988) test_loss: 0.690861 (0.662392) 30 | [2019-09-16 22:52:08,817-log_helper.py# 77] [INFO] Progress: 160 / 225000 [0%], Speed: 0.102 s/iter, ETA 0:06:22 (D:H:M) 31 | 32 | -------------------------------------------------------------------------------- /model/README.md: -------------------------------------------------------------------------------- 1 | # Model 2 | This folder is meant to store the layer for your models or some abstract classes for your models. 3 | 4 | You may include the layer or models you defined in this folder at your experiments, the default class convLSTM and convGRU are implemented in convRNN.py. -------------------------------------------------------------------------------- /model/STconvLSTM.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch 3 | import torch.nn.functional as F 4 | 5 | use_cuda = torch.cuda.is_available() 6 | device = torch.device("cuda" if use_cuda else "cpu") 7 | 8 | class STConvLSTMCell(nn.Module): 9 | def __init__(self, input_size, input_dim, hidden_dim, kernel_size, bias, forget_bias=1.0, layer_norm=True): 10 | """ 11 | Initialize ConvLSTM cell. 12 | 13 | Parameters 14 | ---------- 15 | input_size: (int, int) 16 | Height and width of input tensor as (height, width). 17 | input_dim: int 18 | Number of channels of input tensor. 19 | hidden_dim: int 20 | Number of channels of hidden state. 21 | m_dim: int 22 | Number of channels of M state. 23 | kernel_size: (int, int) 24 | Size of the convolutional kernel. 25 | bias: bool 26 | Whether or not to add the bias. 27 | """ 28 | 29 | super(STConvLSTMCell, self).__init__() 30 | 31 | # init parameters 32 | self.height, self.width = input_size # 初始化高和宽 33 | self.input_dim = input_dim # 初始化输入的维度 34 | self.hidden_dim = hidden_dim # 初始化输出的维度 35 | 36 | self.kernel_size = kernel_size # 初始化核的大小 37 | self.padding = kernel_size[0] // 2, kernel_size[1] // 2 # 自动算padding的大小 38 | self.bias = bias 39 | self.forget_bias = forget_bias 40 | self.layer_norm = layer_norm 41 | 42 | 43 | 44 | # split the conv gate layers 45 | # for W * X_t 46 | self.conv_wx = nn.Conv2d(in_channels=self.input_dim, 47 | out_channels=7 * self.hidden_dim, 48 | kernel_size=self.kernel_size, 49 | padding=self.padding, 50 | bias=self.bias) 51 | 52 | # for W * H^t_1 53 | self.conv_wht_1 = nn.Conv2d(in_channels=self.hidden_dim, 54 | out_channels=4 * self.hidden_dim, 55 | kernel_size=self.kernel_size, 56 | padding=self.padding, 57 | bias=self.bias) 58 | 59 | # for W * M^l_1 60 | self.conv_wml_1 = nn.Conv2d(in_channels=self.hidden_dim, 61 | out_channels=3 * self.hidden_dim, 62 | kernel_size=self.kernel_size, 63 | padding=self.padding, 64 | bias=self.bias) 65 | 66 | # for W * M^l 67 | self.conv_wml = nn.Conv2d(in_channels=self.hidden_dim, 68 | out_channels= self.hidden_dim, 69 | kernel_size=self.kernel_size, 70 | padding=self.padding, 71 | bias=self.bias) 72 | 73 | # for W * C^l 74 | self.conv_wcl = nn.Conv2d(in_channels=self.hidden_dim, 75 | out_channels=self.hidden_dim, 76 | kernel_size=self.kernel_size, 77 | padding=self.padding, 78 | bias=self.bias) 79 | 80 | 81 | # for generate H^l 82 | self.conv_h = nn.Conv2d(in_channels=self.hidden_dim + self.hidden_dim, 83 | out_channels=self.hidden_dim, 84 | kernel_size=[1,1], 85 | padding=0, 86 | bias=self.bias) 87 | 88 | # init parameters 89 | nn.init.orthogonal(self.conv_wx.weight) 90 | nn.init.orthogonal(self.conv_wht_1.weight) 91 | nn.init.orthogonal(self.conv_wml_1.weight) 92 | nn.init.orthogonal(self.conv_wml.weight) 93 | nn.init.orthogonal(self.conv_wcl.weight) 94 | nn.init.orthogonal(self.conv_h.weight) 95 | 96 | # for layerNorm 97 | if self.layer_norm: 98 | self.conv_wx_norm = nn.BatchNorm2d(7 * self.hidden_dim) 99 | self.conv_wht_1_norm = nn.BatchNorm2d(4 * self.hidden_dim) 100 | self.conv_wml_1_norm = nn.BatchNorm2d(3 * self.hidden_dim) 101 | self.conv_wml_norm = nn.BatchNorm2d(self.hidden_dim) 102 | self.conv_wcl_norm = nn.BatchNorm2d(self.hidden_dim) 103 | self.conv_h_norm = nn.BatchNorm2d(self.hidden_dim) 104 | 105 | 106 | # for bias 107 | self.forget_bias_h = torch.nn.Parameter(torch.tensor(self.forget_bias)) 108 | self.forget_bias_m = torch.nn.Parameter(torch.tensor(self.forget_bias)) 109 | 110 | def forward(self, input_tensor, cur_state): 111 | """ 112 | Forward of Conv LSTM Cell 113 | Inputs: 114 | --------------------------------------- 115 | input_tensor: (b, c, h, w) 116 | cur_state: [ H, C, M ] 117 | cur_state: [(b, c_hidden, h, w), (b, c_hidden, h, w), (b, c_hidden, h, w)] 118 | --------------------------------------- 119 | Returns: 120 | --------------------------------------- 121 | h_next, c_next, m_next : ((b, c_hidden, h, w), (b, c_hidden, h, w), (b, c_hidden, h, w)) 122 | next hidden state 123 | """ 124 | # state input 125 | h_cur, c_cur, m_cur = cur_state 126 | # conv gate result 127 | conved_wx = self.conv_wx(input_tensor) 128 | conved_wht_1 = self.conv_wht_1(h_cur) 129 | conved_wml_1 = self.conv_wml_1(m_cur) 130 | # for bn 131 | if self.layer_norm: 132 | conved_wx = self.conv_wx_norm(conved_wx) 133 | conved_wht_1 = self.conv_wht_1_norm(conved_wht_1) 134 | conved_wml_1 = self.conv_wml_1_norm(conved_wml_1) 135 | # split gate result 136 | wxg, wxi, wxf, wxg_, wxi_, wxf_, wxo = torch.split(conved_wx, self.hidden_dim, dim=1) 137 | whg, whi, whf, who = torch.split(conved_wht_1, self.hidden_dim, dim=1) 138 | wmg, wmi, wmf = torch.split(conved_wml_1, self.hidden_dim, dim=1) 139 | # for c_next 140 | g_t = torch.tanh(wxg + whg) 141 | i_t = torch.sigmoid(wxi + whi) 142 | f_t = torch.sigmoid(wxf + whf + self.forget_bias_h) 143 | c_next = f_t * c_cur + i_t * g_t 144 | # for m_next 145 | g_t_ = torch.tanh(wxg_ + wmg) 146 | i_t_ = torch.sigmoid(wxi_ + wmi) 147 | f_t_ = torch.sigmoid(wxf_ + wmf + self.forget_bias_m) 148 | m_next = f_t_ * m_cur + i_t_ * g_t_ 149 | # for wco, wmo 150 | wco = self.conv_wcl(c_next) 151 | wmo = self.conv_wml(m_next) 152 | # for bn 153 | if self.layer_norm: 154 | wco = self.conv_wcl_norm(wco) 155 | wmo = self.conv_wml_norm(wmo) 156 | # for output gate 157 | o_t = torch.sigmoid(wxo + who + wco + wmo) 158 | # for h_next 159 | combined_cmn = torch.cat([c_next, m_next], dim=1) 160 | h_next = o_t * torch.tanh(self.conv_h(combined_cmn)) 161 | 162 | return h_next, c_next, m_next 163 | 164 | 165 | def init_hidden(self, batch_size): 166 | return (torch.zeros(batch_size, self.hidden_dim, self.height, self.width).to(device), 167 | torch.zeros(batch_size, self.hidden_dim, self.height, self.width).to(device), 168 | torch.zeros(batch_size, self.hidden_dim, self.height, self.width).to(device)) 169 | 170 | 171 | 172 | def test(): 173 | pass 174 | 175 | 176 | 177 | if __name__ == "__main__": 178 | test() 179 | 180 | -------------------------------------------------------------------------------- /model/convRNN.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | from torch.autograd import Variable 3 | import torch 4 | import torch.nn.functional as F 5 | 6 | 7 | use_cuda = torch.cuda.is_available() 8 | device = torch.device("cuda" if use_cuda else "cpu") 9 | 10 | class ConvLSTMCell(nn.Module): 11 | 12 | def __init__(self, input_size, input_dim, hidden_dim, kernel_size, bias): 13 | """ 14 | Initialize ConvLSTM cell. 15 | 16 | Parameters 17 | ---------- 18 | input_size: (int, int) 19 | Height and width of input tensor as (height, width). 20 | input_dim: int 21 | Number of channels of input tensor. 22 | hidden_dim: int 23 | Number of channels of hidden state. 24 | kernel_size: (int, int) 25 | Size of the convolutional kernel. 26 | bias: bool 27 | Whether or not to add the bias. 28 | """ 29 | 30 | super(ConvLSTMCell, self).__init__() 31 | 32 | # init parameters 33 | self.height, self.width = input_size # init hight and width 34 | self.input_dim = input_dim # init input dim 35 | self.hidden_dim = hidden_dim # init output dim 36 | 37 | self.kernel_size = kernel_size # init kernel size 38 | self.padding = kernel_size[0] // 2, kernel_size[1] // 2 # caculate padding size automatically 39 | self.bias = bias # init bias 40 | 41 | # all gate conv operation can be include in a big conv layer 42 | self.conv = nn.Conv2d(in_channels=self.input_dim + self.hidden_dim, 43 | out_channels=4 * self.hidden_dim, 44 | kernel_size=self.kernel_size, 45 | padding=self.padding, 46 | bias=self.bias) 47 | 48 | # init parameters 49 | nn.init.orthogonal(self.conv.weight) 50 | if self.bias is True: 51 | nn.init.ones_(self.conv.bias) 52 | 53 | def forward(self, input_tensor, cur_state): 54 | """ 55 | Forward of Conv LSTM Cell 56 | Inputs: 57 | --------------------------------------- 58 | input_tensor: (b, c, h, w) 59 | cur_state: [(b, c_hidden, h, w), (b, c_hidden, h, w)] 60 | --------------------------------------- 61 | Returns: 62 | --------------------------------------- 63 | h_next, c_next : ((b, c_hidden, h, w), (b, c_hidden, h, w)) 64 | next hidden state 65 | """ 66 | 67 | h_cur, c_cur = cur_state 68 | 69 | combined = torch.cat([input_tensor, h_cur], dim=1) # concatenate along channel axis 70 | 71 | combined_conv = self.conv(combined) 72 | cc_i, cc_f, cc_o, cc_g = torch.split(combined_conv, self.hidden_dim, dim=1) 73 | i = torch.sigmoid(cc_i) 74 | f = torch.sigmoid(cc_f) 75 | o = torch.sigmoid(cc_o) 76 | g = F.relu(cc_g) 77 | 78 | c_next = f * c_cur + i * g 79 | h_next = o * torch.tanh(c_next) 80 | 81 | return h_next, c_next 82 | 83 | def init_hidden(self, batch_size): 84 | return (torch.zeros(batch_size, self.hidden_dim, self.height, self.width).to(device), 85 | torch.zeros(batch_size, self.hidden_dim, self.height, self.width).to(device)) 86 | 87 | 88 | class ConvGRUCell(nn.Module): 89 | def __init__(self, input_size, input_dim, hidden_dim, kernel_size, bias): 90 | """ 91 | Initialize the ConvLSTM cell 92 | :param input_size: (int, int) 93 | Height and width of input tensor as (height, width). 94 | :param input_dim: int 95 | Number of channels of input tensor. 96 | :param hidden_dim: int 97 | Number of channels of hidden state. 98 | :param kernel_size: (int, int) 99 | Size of the convolutional kernel. 100 | :param bias: bool 101 | Whether or not to add the bias. 102 | """ 103 | super(ConvGRUCell, self).__init__() 104 | 105 | # init all parameter 106 | self.height, self.width = input_size # init hight and width 107 | self.input_dim = input_dim # init input dim 108 | self.hidden_dim = hidden_dim # init output dim 109 | 110 | self.kernel_size = kernel_size # init kernel size 111 | self.padding = kernel_size[0] // 2, kernel_size[1] // 2 # caculate padding size automatically 112 | self.bias = bias # init bias 113 | 114 | # all gate conv operation can be include in a big conv layer 115 | self.conv_gates = nn.Conv2d(in_channels=self.input_dim + self.hidden_dim, 116 | out_channels=2*self.hidden_dim, # for update_gate,reset_gate respectively 117 | kernel_size=self.kernel_size, 118 | padding=self.padding, 119 | bias=self.bias) 120 | 121 | self.conv_can = nn.Conv2d(in_channels=self.input_dim+self.hidden_dim, 122 | out_channels=self.hidden_dim, # for candidate neural memory 123 | kernel_size=self.kernel_size, 124 | padding=self.padding, 125 | bias=self.bias) 126 | 127 | # init conv parameter and bias parameter 128 | nn.init.orthogonal(self.conv_can.weight) 129 | nn.init.orthogonal(self.conv_gates.weight) 130 | if self.bias is True: 131 | nn.init.ones_(self.conv_can.bias) 132 | nn.init.ones_(self.conv_gates.bias) 133 | 134 | def init_hidden(self, batch_size): 135 | return torch.zeros(batch_size, self.hidden_dim, self.height, self.width).to(device) 136 | 137 | def forward(self, input_tensor, cur_state): 138 | """ 139 | Forward of Conv GRU Cell 140 | Inputs: 141 | --------------------------------------- 142 | input_tensor: (b, c, h, w) 143 | cur_state: (b, c_hidden, h, w) 144 | --------------------------------------- 145 | Returns: 146 | --------------------------------------- 147 | h_next : ((b, c_hidden, h, w)) 148 | next hidden state 149 | """ 150 | combined = torch.cat([input_tensor, cur_state], dim=1) 151 | combined_conv = self.conv_gates(combined) 152 | 153 | gamma, beta = torch.split(combined_conv, self.hidden_dim, dim=1) 154 | reset_gate = torch.sigmoid(gamma) 155 | update_gate = torch.sigmoid(beta) 156 | 157 | combined = torch.cat([input_tensor, reset_gate*cur_state], dim=1) 158 | cc_cnm = self.conv_can(combined) 159 | cnm = F.tanh(cc_cnm) 160 | 161 | h_next = (1 - update_gate) * cur_state + update_gate * cnm 162 | return h_next 163 | 164 | 165 | def test(): 166 | pass 167 | 168 | 169 | 170 | if __name__ == "__main__": 171 | test() 172 | 173 | -------------------------------------------------------------------------------- /model/loss/L1_L2_Loss.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | 4 | class L1_L2_Loss(torch.nn.Module): 5 | def __init__(self): 6 | super(L1_L2_Loss, self).__init__() 7 | 8 | def forward(self, target, pred): 9 | diff = target - pred 10 | loss_ = torch.pow(diff, 2) + torch.abs(diff) # L2 + L1 11 | return loss_.mean() 12 | -------------------------------------------------------------------------------- /model/loss/SSIM_Loss.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn.functional as F 3 | from math import exp 4 | import numpy as np 5 | 6 | 7 | # 计算一维的高斯分布向量 8 | def gaussian(window_size, sigma): 9 | gauss = torch.Tensor([exp(-(x - window_size//2)**2/float(2*sigma**2)) for x in range(window_size)]) 10 | return gauss/gauss.sum() 11 | 12 | # 创建高斯核，通过两个一维高斯分布向量进行矩阵乘法得到 13 | # 可以设定channel参数拓展为3通道 14 | def create_window(window_size, channel): 15 | _1D_window = gaussian(window_size, 1.5).unsqueeze(1) 16 | _2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0) # 矩阵乘法 17 | window = _2D_window.expand(channel, 1, window_size, window_size).contiguous() 18 | return window 19 | 20 | # 计算SSIM 21 | # 直接使用SSIM的公式，但是在计算均值时，不是直接求像素平均值，而是采用归一化的高斯核卷积来代替。 22 | # 在计算方差和协方差时用到了公式Var(X)=E[X^2]-E[X]^2, cov(X,Y)=E[XY]-E[X]E[Y]. 23 | # 正如前面提到的，上面求期望的操作采用高斯核卷积代替。 24 | def _ssim(img1, img2, window, window_size, channel, size_average = True): 25 | # 用生成的窗函数对两个图片做卷积 26 | mu1 = F.conv2d(img1, window, padding = window_size//2, groups = channel) 27 | mu2 = F.conv2d(img2, window, padding = window_size//2, groups = channel) 28 | 29 | # 平方相乘 30 | mu1_sq = mu1.pow(2) 31 | mu2_sq = mu2.pow(2) 32 | mu1_mu2 = mu1*mu2 33 | 34 | # 用窗函数求sigma 35 | sigma1_sq = F.conv2d(img1*img1, window, padding = window_size//2, groups = channel) - mu1_sq 36 | sigma2_sq = F.conv2d(img2*img2, window, padding = window_size//2, groups = channel) - mu2_sq 37 | sigma12 = F.conv2d(img1*img2, window, padding = window_size//2, groups = channel) - mu1_mu2 38 | 39 | # 通过C1, C2 求ssim的图 40 | C1 = 0.01**2 41 | C2 = 0.03**2 42 | 43 | ssim_map = ((2*mu1_mu2 + C1)*(2*sigma12 + C2))/((mu1_sq + mu2_sq + C1)*(sigma1_sq + sigma2_sq + C2)) 44 | 45 | # 看是不是要size平均, 如果要的话就整体平均或者分channel平均 46 | if size_average: 47 | return ssim_map.mean() 48 | else: 49 | return ssim_map.mean(1).mean(1).mean(1) 50 | 51 | # loss SSIM 52 | class SSIM(torch.nn.Module): 53 | def __init__(self, window_size = 11, size_average = True): 54 | super(SSIM, self).__init__() 55 | self.window_size = window_size 56 | self.size_average = size_average 57 | # 一开始先做一个初始化 58 | self.channel = 1 59 | self.window = create_window(window_size, self.channel) 60 | 61 | def forward(self, img1, img2): 62 | if len(img1.size()) is 5: 63 | (batch, frames, channel, _, _) = img1.size() 64 | else: 65 | (_, channel, _, _) = img1.size() 66 | 67 | if channel == self.channel and self.window.data.type() == img1.data.type(): 68 | window = self.window 69 | else: 70 | window = create_window(self.window_size, channel) 71 | 72 | if img1.is_cuda: 73 | window = window.cuda(img1.get_device()) 74 | window = window.type_as(img1) 75 | 76 | self.window = window 77 | self.channel = channel 78 | 79 | if len(img1.size()) is 5: 80 | mean = [] 81 | for i in range(frames): 82 | mean.append(_ssim(img1[:,i,:,:,:], img2[:,i,:,:,:], window, self.window_size, channel, self.size_average)) 83 | mean = torch.stack(mean, dim=0) 84 | return mean.mean() 85 | else: 86 | return _ssim(img1, img2, window, self.window_size, channel, self.size_average) 87 | 88 | 89 | # 一次性的测试函数 90 | def ssim(img1, img2, window_size = 11, size_average = True): 91 | if len(img1.size()) is 5: 92 | (batch, frames, channel, _, _) = img1.size() 93 | else: 94 | (_, channel, _, _) = img1.size() 95 | 96 | window = create_window(window_size, channel) 97 | 98 | if img1.is_cuda: 99 | window = window.cuda(img1.get_device()) 100 | window = window.type_as(img1) 101 | 102 | if len(img1.size()) is 5: 103 | mean = [] 104 | for i in range(frames): 105 | mean.append(_ssim(img1[:,i,:,:,:], img2[:,i,:,:,:], window, window_size, channel, size_average)) 106 | mean = torch.stack(mean, dim=0) 107 | return mean.mean() 108 | else: 109 | return _ssim(img1, img2, window, window_size, channel, size_average) 110 | 111 | 112 | 113 | if __name__ == "__main__": 114 | # how to use 115 | img1 = torch.rand(10, 10, 3, 256, 256) 116 | img2 = torch.rand(10, 10, 3, 256, 256) 117 | 118 | if torch.cuda.is_available(): 119 | img1 = img1.cuda() 120 | img2 = img2.cuda() 121 | 122 | print(ssim(img1, img2)) 123 | 124 | ssim_loss = SSIM(window_size = 11) 125 | 126 | print(ssim_loss(img1, img2)) -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | Cython==0.29.4 2 | colorama==0.3.9 3 | numpy==1.15.4 4 | requests==2.21.0 5 | fire==0.1.3 6 | matplotlib==2.2.3 7 | numba==0.39.0 8 | scipy==1.1.0 9 | h5py==2.8.0 10 | pandas==0.23.4 11 | tqdm==4.29.1 12 | tensorboardX==1.6 13 | opencv_python==3.4.3.18 14 | torch==1.1.0 15 | torchvision==0.2.1 16 | joblib==0.13.0 17 | moviepy==1.0.0 18 | nvidia-ml-py3==7.352.0 -------------------------------------------------------------------------------- /tools/README.md: -------------------------------------------------------------------------------- 1 | # Tools 2 | This folder include the training/testing/demo files, you may want to use these code with the config when training/testing your models. 3 | 4 | # Train 5 | 6 | The first thing you need to do is build the layers and abstract class in model folder, implement your model in experiments/'experiment_name'/Custom.py file, and the model class name should be the 'Custom'. 7 | 8 | After build your own Custom class, you can specify your experiment folder name ('expriment_name') in config file. Then run the train_video_prediction.py file. 9 | 10 | if you want to customize your class name, you can change the training file code to adjust your model arch. 11 | 12 | A config file template is also included in this folder. 13 | -------------------------------------------------------------------------------- /tools/train_config.json: -------------------------------------------------------------------------------- 1 | { 2 | "meta":{ 3 | "experiment_path": "experiments", 4 | "arch": "Custom", 5 | "board_path": "board", 6 | "experiment_name":"video_prediction_demo" 7 | }, 8 | "train":{ 9 | "epoches": 50, 10 | "batch_size": 2, 11 | "lr": 1e-5, 12 | "print_freq": 20 13 | }, 14 | "model":{ 15 | "input_size":[64,64], 16 | "input_dim":1, 17 | "hidden_dim":64, 18 | "kernel_size": [3,3], 19 | "input_num": 10 20 | } 21 | } -------------------------------------------------------------------------------- /tools/train_config_PredRNN.json: -------------------------------------------------------------------------------- 1 | { 2 | "meta":{ 3 | "experiment_path": "experiments", 4 | "arch": "Custom", 5 | "board_path": "board", 6 | "experiment_name":"PredRNN" 7 | }, 8 | "train":{ 9 | "epoches": 50, 10 | "batch_size": 1, 11 | "lr": 1e-3, 12 | "print_freq": 20 13 | }, 14 | "model":{ 15 | "input_size":[64,64], 16 | "input_dim":1, 17 | "hidden_dim":64, 18 | "kernel_size": [3,3], 19 | "input_num": 10 20 | } 21 | } -------------------------------------------------------------------------------- /tools/train_video_prediction_KTHDataset.py: -------------------------------------------------------------------------------- 1 | import os 2 | import sys 3 | import time 4 | parent_path = os.path.dirname(os.getcwd()) 5 | if 'tools' in os.getcwd(): 6 | os.chdir(parent_path) 7 | sys.path.insert(0, os.getcwd()) 8 | 9 | import json 10 | import torch 11 | import torch.nn as nn 12 | from torchvision import transforms 13 | from dataloader.KTHDataset.KTHDataset import KTHDataset 14 | 15 | import argparse 16 | 17 | import logging 18 | from utils.log_helper import init_log, add_file_handler, print_speed 19 | from utils.config_helper import Configs 20 | from utils.average_meter_helper import AverageMeter 21 | from model.loss.SSIM_Loss import SSIM 22 | from model.loss.L1_L2_Loss import L1_L2_Loss 23 | 24 | import inspect 25 | from utils.memory.gpu_mem_track import MemTracker 26 | frame = inspect.currentframe() 27 | gpu_tracker = MemTracker(frame) 28 | 29 | # 生成命令行的参数 30 | parser = argparse.ArgumentParser(description='Train moving mnist video prediction algorithm') 31 | parser.add_argument('-c', '--cfg', default=os.path.join(os.getcwd(), "tools", "train_config_PredRNN.json"), type=str, required=False, help='training config file path') 32 | 33 | args = parser.parse_args() 34 | 35 | # 初始化一些变量 36 | cfg = Configs(args.cfg) 37 | # board的路径 38 | board_path = cfg.meta["board_path"] 39 | experiment_path = cfg.meta["experiment_path"] 40 | experiment_name = cfg.meta["experiment_name"] 41 | arch = cfg.meta["arch"] 42 | # 训练时候的一些参数 43 | batch_size = cfg.train['batch_size'] 44 | epoches = cfg.train['epoches'] 45 | lr = cfg.train['lr'] 46 | # 初始化未来帧的数量 47 | num_frame = cfg.model['input_num'] 48 | # print freq 49 | print_freq = cfg.train['print_freq'] 50 | 51 | # 初始化logger 52 | global_logger = init_log('global', level=logging.INFO) 53 | add_file_handler("global", os.path.join(os.getcwd(), 'logs', '{}.log'.format(experiment_name)), level=logging.DEBUG) 54 | 55 | # 打印cfg信息 56 | cfg.log_dict() 57 | 58 | # 初始化avrager 59 | avg = AverageMeter() 60 | 61 | # cuda 62 | use_cuda = torch.cuda.is_available() 63 | device = torch.device("cuda" if use_cuda else "cpu") 64 | torch.backends.cudnn.benchmark = True 65 | 66 | # 准备数据集 67 | train_set = KTHDataset(root='./data/KTHDataset', train=True, download=True, 68 | transform=transforms.Compose([transforms.Resize(cfg.model["input_size"]), transforms.ToTensor(),]), 69 | target_transform=transforms.Compose([transforms.Resize(cfg.model["input_size"]), transforms.ToTensor(),])) 70 | test_set = KTHDataset(root='./data/KTHDataset', train=False, download=True, 71 | transform=transforms.Compose([transforms.Resize(cfg.model["input_size"]), transforms.ToTensor(),]), 72 | target_transform=transforms.Compose([transforms.Resize(cfg.model["input_size"]), transforms.ToTensor(),])) 73 | 74 | # 建立dataloader 75 | train_loader = torch.utils.data.DataLoader( 76 | dataset=train_set, 77 | batch_size=batch_size, 78 | num_workers=16, 79 | shuffle=True) 80 | test_loader = torch.utils.data.DataLoader( 81 | dataset=test_set, 82 | batch_size=batch_size, 83 | num_workers=16, 84 | shuffle=True) 85 | 86 | # 建立test的iter 87 | test_iter = iter(test_loader) 88 | 89 | # 确定数据集长度 90 | train_lenth = len(train_loader) 91 | test_lenth = len(test_loader) 92 | 93 | global_logger.debug('==>>> total trainning batch number: {}'.format(train_lenth)) 94 | global_logger.debug('==>>> total testing batch number: {}'.format(test_lenth)) 95 | 96 | # 加载模型 97 | sys.path.append(os.path.join(".", experiment_path, experiment_name)) 98 | if arch == "Custom": 99 | from custom import Custom 100 | model = Custom(cfg=cfg.model) 101 | model = model.to(device) 102 | model = torch.nn.DataParallel(model, list(range(torch.cuda.device_count()))).to(device) 103 | else: 104 | raise NotImplementedError 105 | 106 | # 建立tensorboard的实例 107 | from tensorboardX import SummaryWriter 108 | writer = SummaryWriter(os.path.join(".", board_path, experiment_name)) 109 | 110 | # 建立优化器 111 | optimizer = torch.optim.Adam(model.parameters(), lr=lr, betas=(0.9,0.999)) 112 | scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer=optimizer, milestones=[10,20,30,40], gamma=0.5) 113 | 114 | # 建立loss 115 | loss_L1 = nn.MSELoss().to(device) 116 | loss_L2 = nn.L1Loss().to(device) 117 | # crossEntropy 118 | loss_BCE = nn.BCELoss().to(device) 119 | # SSIM 120 | loss_SSIM = SSIM(window_size=11, size_average=True) 121 | 122 | loss_L1_L2 = L1_L2_Loss().to(device) 123 | 124 | # 训练的部分 125 | for epoch in range(epoches): 126 | for step, [seq, seq_target] in enumerate(train_loader): 127 | step_time = time.time() 128 | # 打印测试信息 129 | if epoch is 0 and step is 0: 130 | global_logger.debug('Input: {}'.format(seq.shape)) 131 | global_logger.debug('--- Sample') 132 | global_logger.debug('Target: {}'.format(seq_target.shape)) 133 | 134 | # 放到cuda中 135 | seq, seq_target = seq.to(device), seq_target.to(device) 136 | 137 | # 优化器归零 138 | optimizer.zero_grad() 139 | 140 | # 送入模型进行推断 141 | layer_output = model(seq, future=num_frame) 142 | 143 | # loss计算 144 | train_loss = loss_BCE(layer_output[:, -num_frame:, :, :, :], seq_target[:, -num_frame:, :, :, :]) 145 | with torch.no_grad(): 146 | train_metric = loss_SSIM(layer_output[:, -num_frame:, :, :, :], seq_target[:, -num_frame:, :, :, :]) 147 | train_loss.backward() 148 | 149 | # 优化器更新 150 | optimizer.step() 151 | 152 | # validate 153 | 154 | with torch.no_grad(): 155 | # load random test set 156 | try: 157 | seq_test, gt_seq_test = next(test_iter) 158 | except StopIteration: 159 | test_iter = iter(test_loader) 160 | seq_test, gt_seq_test = next(test_iter) 161 | 162 | seq_test, gt_seq_test = seq_test.to(device), gt_seq_test.to(device) 163 | 164 | # 送入模型进行推断 165 | test_output = model(seq_test, future=num_frame) 166 | 167 | # loss计算 168 | test_loss = loss_BCE(test_output[:, -num_frame:, :, :, :], gt_seq_test[:, -num_frame:, :, :, :]) 169 | test_metric = loss_SSIM(test_output[:, -num_frame:, :, :, :], gt_seq_test[:, -num_frame:, :, :, :]) 170 | 171 | step_time = time.time() - step_time 172 | 173 | # 将有用的信息存进tensorboard中 174 | if (step+1) % 200 == 0: 175 | writer.add_video('train_seq/feed_seq', seq, epoch*train_lenth + step + 1) 176 | writer.add_video('train_seq/gt_seq', seq_target, epoch*train_lenth + step + 1) 177 | writer.add_video('train_seq/pred_seq', layer_output, epoch*train_lenth + step + 1) 178 | writer.add_video('test_seq/feed_seq', seq_test, epoch*train_lenth + step + 1) 179 | writer.add_video('test_seq/gt_seq', gt_seq_test, epoch*train_lenth + step + 1) 180 | writer.add_video('test_seq/pred_seq', test_output, epoch*train_lenth + step + 1) 181 | writer.add_scalars('loss/merge', {"train_loss": train_loss,"test_loss":test_loss, "train_metric":train_metric, "test_metric":test_metric}, epoch*train_lenth + step + 1) 182 | 183 | # 更新avrager 184 | avg.update(step_time=step_time, train_loss=train_loss, test_loss=test_loss, train_metric=train_metric) # 算平均值 185 | 186 | # 打印结果 187 | if (step+1) % print_freq == 0: 188 | global_logger.info('Epoch: [{0}][{1}/{2}] {step_time:s}\t{train_loss:s}\t{test_loss:s}\t{train_metric:s}'.format( 189 | epoch+1, (step + 1) % train_lenth, train_lenth, step_time=avg.step_time, train_loss=avg.train_loss, test_loss=avg.test_loss, train_metric=avg.train_metric)) 190 | print_speed(epoch*train_lenth + step + 1, avg.step_time.avg, epoches * train_lenth) 191 | 192 | # scheduler更新 193 | scheduler.step() 194 | 195 | 196 | 197 | -------------------------------------------------------------------------------- /tools/train_video_prediction_MNIST.py: -------------------------------------------------------------------------------- 1 | import os 2 | import sys 3 | import time 4 | parent_path = os.path.dirname(os.getcwd()) 5 | print(os.getcwd()) 6 | if 'tools' in os.getcwd(): 7 | os.chdir(parent_path) 8 | sys.path.insert(0, os.getcwd()) 9 | 10 | import json 11 | import torch 12 | import torch.nn as nn 13 | from torchvision import transforms 14 | from dataloader.MovingMNIST.MovingMNIST import MovingMNIST 15 | 16 | import argparse 17 | 18 | import logging 19 | from utils.log_helper import init_log, add_file_handler, print_speed 20 | from utils.config_helper import Configs 21 | from utils.average_meter_helper import AverageMeter 22 | from model.loss.SSIM_Loss import SSIM 23 | from model.loss.L1_L2_Loss import L1_L2_Loss 24 | 25 | import inspect 26 | from utils.memory.gpu_mem_track import MemTracker 27 | frame = inspect.currentframe() 28 | gpu_tracker = MemTracker(frame) 29 | 30 | # 生成命令行的参数 31 | parser = argparse.ArgumentParser(description='Train moving mnist video prediction algorithm') 32 | parser.add_argument('-c', '--cfg', default=os.path.join(os.getcwd(), "tools", "train_config_PredRNN.json"), type=str, required=False, help='training config file path') 33 | 34 | args = parser.parse_args() 35 | 36 | # 初始化一些变量 37 | cfg = Configs(args.cfg) 38 | # board的路径 39 | board_path = cfg.meta["board_path"] 40 | experiment_path = cfg.meta["experiment_path"] 41 | experiment_name = cfg.meta["experiment_name"] 42 | arch = cfg.meta["arch"] 43 | # 训练时候的一些参数 44 | batch_size = cfg.train['batch_size'] 45 | epoches = cfg.train['epoches'] 46 | lr = cfg.train['lr'] 47 | # 初始化未来帧的数量 48 | num_frame = cfg.model['input_num'] 49 | # print freq 50 | print_freq = cfg.train['print_freq'] 51 | 52 | # 初始化logger 53 | global_logger = init_log('global', level=logging.INFO) 54 | add_file_handler("global", os.path.join(os.getcwd(), 'logs', '{}.log'.format(experiment_name)), level=logging.DEBUG) 55 | 56 | # 打印cfg信息 57 | cfg.log_dict() 58 | 59 | # 初始化avrager 60 | avg = AverageMeter() 61 | 62 | # cuda 63 | use_cuda = torch.cuda.is_available() 64 | device = torch.device("cuda" if use_cuda else "cpu") 65 | torch.backends.cudnn.benchmark = True 66 | 67 | # 准备数据集 68 | train_set = MovingMNIST(root='./data/mnist', train=True, download=True, 69 | transform=transforms.Compose([transforms.ToTensor(),]), 70 | target_transform=transforms.Compose([transforms.ToTensor(),])) 71 | test_set = MovingMNIST(root='./data/mnist', train=False, download=True, 72 | transform=transforms.Compose([transforms.ToTensor(),]), 73 | target_transform=transforms.Compose([transforms.ToTensor(),])) 74 | 75 | # 建立dataloader 76 | train_loader = torch.utils.data.DataLoader( 77 | dataset=train_set, 78 | batch_size=batch_size, 79 | num_workers=16, 80 | shuffle=True) 81 | test_loader = torch.utils.data.DataLoader( 82 | dataset=test_set, 83 | batch_size=batch_size, 84 | num_workers=16, 85 | shuffle=False) 86 | 87 | # 建立test的iter 88 | test_iter = iter(test_loader) 89 | 90 | # 确定数据集长度 91 | train_lenth = len(train_loader) 92 | test_lenth = len(test_loader) 93 | 94 | global_logger.debug('==>>> total trainning batch number: {}'.format(train_lenth)) 95 | global_logger.debug('==>>> total testing batch number: {}'.format(test_lenth)) 96 | 97 | # 加载模型 98 | sys.path.append(os.path.join(".", experiment_path, experiment_name)) 99 | if arch == "Custom": 100 | from custom import Custom 101 | model = Custom(cfg=cfg.model) 102 | model = torch.nn.DataParallel(model, list(range(torch.cuda.device_count()))).to(device) 103 | else: 104 | raise NotImplementedError 105 | 106 | # 建立tensorboard的实例 107 | from tensorboardX import SummaryWriter 108 | writer = SummaryWriter(os.path.join(".", board_path, experiment_name)) 109 | 110 | # 建立优化器 111 | optimizer = torch.optim.Adam(model.parameters(), lr=lr) 112 | scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer=optimizer, milestones=[10,20,30,40], gamma=0.5) 113 | 114 | # 建立loss 115 | loss_L1 = nn.MSELoss().to(device) 116 | loss_L2 = nn.L1Loss().to(device) 117 | # crossEntropy 118 | loss_BCE = nn.BCELoss().to(device) 119 | # SSIM 120 | loss_SSIM = SSIM(window_size=11, size_average=True) 121 | # L1 + L2 mean 122 | loss_L1_L2 = L1_L2_Loss().to(device) 123 | 124 | # 训练的部分 125 | for epoch in range(epoches): 126 | for step, [seq, seq_target] in enumerate(train_loader): 127 | step_time = time.time() 128 | # 打印测试信息 129 | if epoch is 0 and step is 0: 130 | global_logger.debug('Input: {}'.format(seq.shape)) 131 | global_logger.debug('--- Sample') 132 | global_logger.debug('Target: {}'.format(seq_target.shape)) 133 | 134 | # 放到cuda中 135 | seq, seq_target = seq.to(device), seq_target.to(device) 136 | 137 | # 优化器归零 138 | optimizer.zero_grad() 139 | 140 | # 送入模型进行推断 141 | layer_output = model(seq, future=num_frame) 142 | 143 | # loss计算 144 | train_loss = loss_L1_L2(layer_output[:, -num_frame:, :, :, :], seq_target[:, -num_frame:, :, :, :]) 145 | with torch.no_grad(): 146 | train_metric = loss_SSIM(layer_output[:, -num_frame:, :, :, :], seq_target[:, -num_frame:, :, :, :]) 147 | train_loss.backward() 148 | 149 | # 优化器更新 150 | optimizer.step() 151 | 152 | # validate 153 | 154 | with torch.no_grad(): 155 | # load random test set 156 | try: 157 | seq_test, gt_seq_test = next(test_iter) 158 | except StopIteration: 159 | test_iter = iter(test_loader) 160 | seq_test, gt_seq_test = next(test_iter) 161 | 162 | seq_test, gt_seq_test = seq_test.to(device), gt_seq_test.to(device) 163 | 164 | # 送入模型进行推断 165 | test_output = model(seq_test, future=num_frame) 166 | 167 | # loss计算 168 | test_loss = loss_L1_L2(test_output[:, -num_frame:, :, :, :], gt_seq_test[:, -num_frame:, :, :, :]) 169 | test_metric = loss_SSIM(test_output[:, -num_frame:, :, :, :], gt_seq_test[:, -num_frame:, :, :, :]) 170 | 171 | step_time = time.time() - step_time 172 | 173 | # 将有用的信息存进tensorboard中 174 | if (step+1) % print_freq == 0: 175 | writer.add_video('train_seq/feed_seq', seq, epoch*train_lenth + step + 1) 176 | writer.add_video('train_seq/gt_seq', seq_target, epoch*train_lenth + step + 1) 177 | writer.add_video('train_seq/pred_seq', layer_output, epoch*train_lenth + step + 1) 178 | writer.add_video('test_seq/feed_seq', seq_test, epoch*train_lenth + step + 1) 179 | writer.add_video('test_seq/gt_seq', gt_seq_test, epoch*train_lenth + step + 1) 180 | writer.add_video('test_seq/pred_seq', test_output, epoch*train_lenth + step + 1) 181 | writer.add_scalars('loss/merge', {"train_loss": train_loss,"test_loss":test_loss, "train_metric":train_metric, "test_metric":test_metric}, epoch*train_lenth + step + 1) 182 | 183 | # 更新avrager 184 | avg.update(step_time=step_time, train_loss=train_loss, test_loss=test_loss, train_metric=train_metric) # 算平均值 185 | 186 | # 打印结果 187 | if (step+1) % print_freq == 0: 188 | global_logger.info('Epoch: [{0}][{1}/{2}] {step_time:s}\t{train_loss:s}\t{test_loss:s}\t{train_metric:s}'.format( 189 | epoch+1, (step + 1) % train_lenth, train_lenth, step_time=avg.step_time, train_loss=avg.train_loss, test_loss=avg.test_loss, train_metric=avg.train_metric)) 190 | print_speed(epoch*train_lenth + step + 1, avg.step_time.avg, epoches * train_lenth) 191 | 192 | # scheduler更新 193 | scheduler.step() 194 | 195 | 196 | 197 | -------------------------------------------------------------------------------- /utils/README.md: -------------------------------------------------------------------------------- 1 | # Utils 2 | This folder contains the utilities for train/test your model, the usage is self-concluded in the file. 3 | 4 | # average_meter_helper 5 | Contains a class to calculate dynamic average value. 6 | 7 | # config_helper 8 | To load configs 9 | 10 | # log_helper 11 | To generate the logger handel, which can give multiple stream log message (now contains file and stream handler). 12 | -------------------------------------------------------------------------------- /utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JaMesLiMers/Frame_Video_Prediction_Pytorch/fec5870a2d2ce8f91085c38c46f86e3b58ee0385/utils/__init__.py -------------------------------------------------------------------------------- /utils/average_meter_helper.py: -------------------------------------------------------------------------------- 1 | # -------------------------------------------------------- 2 | # SiamMask 3 | # Licensed under The MIT License 4 | # Written by Qiang Wang (wangqiang2015 at ia.ac.cn) 5 | # -------------------------------------------------------- 6 | import numpy as np 7 | 8 | 9 | class Meter(object): # 度量 10 | def __init__(self, name, val, avg): 11 | self.name = name 12 | self.val = val 13 | self.avg = avg 14 | 15 | def __repr__(self): 16 | return "{name}: {val:.6f} ({avg:.6f})".format( 17 | name=self.name, val=self.val, avg=self.avg 18 | ) 19 | 20 | def __format__(self, *tuples, **kwargs): 21 | return self.__repr__() 22 | 23 | 24 | class AverageMeter(object): # 计算和保存平均值 25 | """Computes and stores the average and current value""" 26 | def __init__(self): 27 | self.reset() # 初始化的时候reset 28 | 29 | def reset(self): # 重新设置 30 | self.val = {} 31 | self.sum = {} # sum用来存总的数据 32 | self.count = {} 33 | 34 | def update(self, batch=1, **kwargs): # update数据 update(key = 1) 35 | val = {} # 新的字典 36 | for k in kwargs: # 对于传入的参数 37 | val[k] = kwargs[k] / float(batch) # val中的数据先进行batch级别的平均 # val['key'] = 1/1 = 1 38 | self.val.update(val) # 把val添加到self.val里面 39 | for k in kwargs: 40 | if k not in self.sum: #如果是新的建 41 | self.sum[k] = 0 # 初始化到sum里 42 | self.count[k] = 0 43 | self.sum[k] += kwargs[k] # 再加 44 | self.count[k] += batch 45 | 46 | def __repr__(self): # 给程序员的显示接口， 可以直接打AverageMeter的变量来显示 47 | s = '' 48 | for k in self.sum: 49 | s += self.format_str(k) 50 | return s 51 | 52 | def format_str(self, attr): 53 | return "{name}: {val:.6f} ({avg:.6f}) ".format( 54 | name=attr, 55 | val=float(self.val[attr]), 56 | avg=float(self.sum[attr]) / self.count[attr]) 57 | 58 | def __getattr__(self, attr): # 如果访问的属性不再范围内，则调用这个方法 59 | if attr in self.__dict__: # 如果这个attr在dict中的话 60 | return super(AverageMeter, self).__getattr__(attr) # 用他父类的getattr方法 61 | if attr not in self.sum: # 如果不在sum里面的话 62 | # logger.warn("invalid key '{}'".format(attr)) 63 | # print("invalid key '{}'".format(attr)) # 报错（提示） 64 | return Meter(attr, 0, 0) # 返回 0 0 65 | return Meter(attr, self.val[attr], self.avg(attr)) # 返回正确的meter 66 | 67 | def avg(self, attr): # 计算avg 68 | return float(self.sum[attr]) / self.count[attr] 69 | 70 | 71 | if __name__ == '__main__': 72 | # 用法 73 | avg = AverageMeter() # 初始化 74 | avg.update(time=1.1, accuracy=.99) # 传入需要avg的参数 75 | avg.update(time=1.0, accuracy=.90) # 多次传入来求平均 76 | 77 | print(avg) # 将所有的平均值进行打印 78 | 79 | print(avg.time) # 打印特定的值(str) 80 | print(avg.time.avg) # 打印特定的平均(float) 81 | print(avg.time.val) # 打印特定的值(float) 82 | print(avg.SS) # 如果出现了没有的默认为0 83 | 84 | 85 | 86 | 87 | -------------------------------------------------------------------------------- /utils/config_helper.py: -------------------------------------------------------------------------------- 1 | import json 2 | import logging 3 | from os.path import exists 4 | 5 | 6 | class Configs(object): 7 | def __init__(self, file_path_or_dict, logger_name="global"): 8 | super(Configs, self).__init__() 9 | # 加载logger 10 | self.logger = logging.getLogger(logger_name) 11 | # 加载cfg文件, 处理基础的参数 12 | self.check_meta(self.load_config(file_path_or_dict)) 13 | 14 | def load_config(self, file_path_or_dict): 15 | if type(file_path_or_dict) is str: 16 | assert exists(file_path_or_dict), '"{}" not exists'.format(file_path_or_dict) # 确认有文件 17 | config = dict(json.load(open(file_path_or_dict))) # 打开加载json文件 18 | elif type(file_path_or_dict) is dict: 19 | config = file_path_or_dict 20 | else: 21 | raise Exception("The input must be a string path or a dict") 22 | return config 23 | 24 | def check_meta(self, cfg_init): 25 | """ 26 | Check 是否将基础的需求写入了config文件中, 没有的话就使用默认的. 27 | 具体check的有: 28 | 是否包含meta标签, 以及重要的experiment_path和arch. 29 | 如果没问题的话就赋值给Config类. 30 | """ 31 | self.cfg_init = cfg_init 32 | 33 | # check meta部分 34 | if 'meta' not in self.cfg_init: 35 | self.logger.warning("The cfg not include meta tag, will generate default") 36 | self.logger.warning("Used the default meta configs.") 37 | self.cfg_init['meta'] = {"experiment_path": "experiments", 38 | "arch": "Custom", 39 | "board_path": "board", 40 | "experiment_name":"video_prediction_demo"}, 41 | else: 42 | cfg_meta = self.cfg_init['meta'] 43 | if 'experiment_path' not in cfg_meta: 44 | self.logger.warning("Not specified experiment_path, used default. ('experiments')") 45 | self.cfg_init['meta']['experiment_path'] = "experiments" 46 | if 'arch' not in cfg_meta: 47 | self.logger.warning("Not specified arch, used default. (Custom)") 48 | self.cfg_init['meta']['arch'] = "Custom" 49 | if 'board_path' not in cfg_meta: 50 | self.logger.warning("Not specified board_path, used default. (board)") 51 | self.cfg_init['meta']['board_path'] = "board" 52 | if 'experiment_name' not in cfg_meta: 53 | self.logger.warning("Not specified experiment_name, used default. (video_prediction_demo)") 54 | self.cfg_init['meta']['experiment_name'] = "video_prediction_demo" 55 | 56 | # 赋值部分 57 | self.__dict__.update(self.cfg_init) 58 | 59 | def log_dict(self, logger_name="global"): 60 | # 打印Config的文件到log 61 | self.logger.debug("Used config: \n {}".format(self.cfg_init)) 62 | 63 | 64 | 65 | if __name__ == "__main__": 66 | """ 67 | cfg 文件的加载类 68 | cfg文件的格式为json文件或者dict() 69 | 70 | meta部分标明实验的元信息(必须, 会验证), 71 | train部分标明训练时候的参数, 72 | model部分标明模型的参数 73 | 74 | 下面是例子和使用方法: 75 | """ 76 | 77 | test_cfg = { 78 | "meta":{ 79 | "experiment_path": "./experiments/video_prediction/", 80 | "arch": "Custom" 81 | }, 82 | "train":{ 83 | "epoches": 50, 84 | "batch_size": 16, 85 | "lr": 1e-3 86 | }, 87 | "model":{ 88 | "input_size":[64,64], 89 | "input_dim":1, 90 | "hidden_dim":64, 91 | "kernel_size":[3,3], 92 | "num_layers":3, 93 | "bias":True, 94 | "return_all_layers": False, 95 | "predict_num": 10 96 | } 97 | } 98 | test_Configs = Configs(test_cfg) 99 | 100 | # 直接访问名字即可, 返回的是一个dict 101 | print(test_Configs.meta) 102 | print(type(test_Configs.meta)) 103 | 104 | print(test_Configs.train) 105 | print(type(test_Configs.train)) 106 | 107 | print(test_Configs.model) 108 | print(type(test_Configs.model)) -------------------------------------------------------------------------------- /utils/log_helper.py: -------------------------------------------------------------------------------- 1 | # -------------------------------------------------------- 2 | # SiamMask 3 | # Licensed under The MIT License 4 | # Written by Qiang Wang (wangqiang2015 at ia.ac.cn) 5 | # -------------------------------------------------------- 6 | from __future__ import division 7 | 8 | import os 9 | import logging 10 | import sys 11 | import math 12 | 13 | # 初始化用来保存 log 类的set 14 | logs = set() 15 | 16 | def get_format(): 17 | """ 18 | 如果不是在slurm集群上面运行的话, 就设定level为0 19 | 总之就是生成一个formatter类, 用来构建输出的信息格式 20 | """ 21 | format_str = '[%(asctime)s-%(filename)s#%(lineno)3d] [%(levelname)s] %(message)s' 22 | formatter = logging.Formatter(format_str) 23 | return formatter 24 | 25 | 26 | def get_format_custom(): 27 | format_str = '[%(asctime)s-%(message)s' 28 | formatter = logging.Formatter(format_str) 29 | return formatter 30 | 31 | 32 | def init_log(name, level = logging.INFO, format_func=get_format): 33 | """ 34 | 初始化log类 35 | 返回一个logger的类 36 | 每次使用的时候, 都给予log元祖里面加一个(name, level)的子元祖, 用来防止重复初始化 37 | 38 | """ 39 | # 防止重复初始化 40 | if (name, level) in logs: 41 | return logging.getLogger(name) 42 | # 如果没有就将name和level放进去 43 | logs.add((name, level)) 44 | logger = logging.getLogger(name) # 可以直接调用, 如果没有的话就自动创建一个. 45 | logger.setLevel(logging.DEBUG) # 设定这个logger 的等级(在这个等级之上的才会被处理) 46 | ch = logging.StreamHandler() # 初始化一个打印在命令行的handler 47 | ch.setLevel(level) # 设定命令行的handler的级别 48 | formatter = format_func() # 构建format 49 | ch.setFormatter(formatter) # 设定format 50 | logger.addHandler(ch) # 为创建的logger增加handler 51 | return logger 52 | 53 | 54 | def add_file_handler(name, log_file, level=logging.DEBUG): 55 | """ 56 | 为传入的类增加文件的handler 57 | """ 58 | logger = logging.getLogger(name) # 将目标的logger初始化一下, 如果没有就创建 59 | fh = logging.FileHandler(log_file, 'w+') # 文件的名字, 初始化filehandler 60 | fh.setFormatter(get_format()) # setformat 61 | fh.setLevel(level) 62 | logger.addHandler(fh) # 添加handler 63 | 64 | 65 | def print_speed(i, i_time, n, logger_name='global'): 66 | """ 67 | 用来生成目标的进度 68 | 传入的时间按照秒来算 69 | print_speed(index, index_time, total_iteration, logger_name) 70 | """ 71 | logger = logging.getLogger(logger_name) 72 | average_time = i_time 73 | remaining_time = (n - i) * average_time 74 | remaining_day = math.floor(remaining_time / 86400) 75 | remaining_hour = math.floor(remaining_time / 3600 - remaining_day * 24) 76 | remaining_min = math.floor(remaining_time / 60 - remaining_day * 1440 - remaining_hour * 60) 77 | logger.info('Progress: %d / %d [%d%%], Speed: %.3f s/iter, ETA %d:%02d:%02d (D:H:M)\n' % (i, n, i/n*100, average_time, remaining_day, remaining_hour, remaining_min)) 78 | 79 | 80 | # 在一开始先初始化global的logger 81 | # init_log('global') 82 | 83 | 84 | if __name__ == "__main__": 85 | """ 86 | Usage 87 | """ 88 | # 生成logger, 一开始自动初始化了一个global logger, 如果需要的话可以重新创建一个不重名的logger 89 | # 如果已经存在就直接返回创建好的logger 90 | logger_test= init_log('global', level=logging.INFO) 91 | # 将logger添加一个handler到文件 92 | add_file_handler("global", os.path.join('.', 'test.log'), level=logging.INFO) 93 | 94 | # log的方法 95 | logger_test.debug('this is a debug log') 96 | logger_test.info('hello info') 97 | logger_test.warning('this is a warning log') 98 | logger_test.error('this is a error message') 99 | logger_test.critical('this is critical') 100 | 101 | # 新增的方法 (默认用global来print) 102 | print_speed(1, 1, 10) -------------------------------------------------------------------------------- /utils/memory/README.md: -------------------------------------------------------------------------------- 1 | # Pytorch-Memory-Utils 2 | 3 | These codes can help you to detect your GPU memory during training with Pytorch. 4 | 5 | A blog about this tool and explain the details : https://oldpan.me/archives/pytorch-gpu-memory-usage-track 6 | 7 | # Requirement: 8 | 9 | ``` 10 | pynvml(pip install nvidia-ml-py3) 11 | ``` 12 | 13 | ## The following is the print content. 14 | 15 | - Calculate the memory usage of a single model 16 | ``` 17 | Model Sequential : params: 0.450304M 18 | Model Sequential : intermedite variables: 336.089600 M (without backward) 19 | Model Sequential : intermedite variables: 672.179200 M (with backward) 20 | ``` 21 | - Track the amount of GPU memory usage 22 | ```markdown 23 | # 12-Sep-18-21:48:45-gpu_mem_track.txt 24 | 25 | GPU Memory Track | 12-Sep-18-21:48:45 | Total Used Memory:696.5 Mb 26 | 27 | At __main__ : line 13 Total Used Memory:696.5 Mb 28 | 29 | + | 7 * Size:(512, 512, 3, 3) | Memory: 66.060 M | 30 | + | 1 * Size:(512, 256, 3, 3) | Memory: 4.7185 M | 31 | + | 1 * Size:(64, 64, 3, 3) | Memory: 0.1474 M | 32 | + | 1 * Size:(128, 64, 3, 3) | Memory: 0.2949 M | 33 | + | 1 * Size:(128, 128, 3, 3) | Memory: 0.5898 M | 34 | + | 8 * Size:(512,) | Memory: 0.0163 M | 35 | + | 3 * Size:(256, 256, 3, 3) | Memory: 7.0778 M | 36 | + | 1 * Size:(256, 128, 3, 3) | Memory: 1.1796 M | 37 | + | 2 * Size:(64,) | Memory: 0.0005 M | 38 | + | 4 * Size:(256,) | Memory: 0.0040 M | 39 | + | 2 * Size:(128,) | Memory: 0.0010 M | 40 | + | 1 * Size:(64, 3, 3, 3) | Memory: 0.0069 M | 41 | 42 | At __main__ : line 15 Total Used Memory:1142.0 Mb 43 | 44 | + | 1 * Size:(60, 3, 512, 512) | Memory: 188.74 M | 45 | + | 1 * Size:(30, 3, 512, 512) | Memory: 94.371 M | 46 | + | 1 * Size:(40, 3, 512, 512) | Memory: 125.82 M | 47 | 48 | At __main__ : line 21 Total Used Memory:1550.9 Mb 49 | 50 | + | 1 * Size:(120, 3, 512, 512) | Memory: 377.48 M | 51 | + | 1 * Size:(80, 3, 512, 512) | Memory: 251.65 M | 52 | 53 | At __main__ : line 26 Total Used Memory:2180.1 Mb 54 | 55 | - | 1 * Size:(120, 3, 512, 512) | Memory: 377.48 M | 56 | - | 1 * Size:(40, 3, 512, 512) | Memory: 125.82 M | 57 | 58 | At __main__ : line 32 Total Used Memory:1676.8 Mb 59 | ``` 60 | 61 | ## How to use 62 | 63 | ### Track the amount of GPU memory usage 64 | simple example: 65 | 66 | ```python 67 | import torch 68 | import inspect 69 | 70 | from torchvision import models 71 | from gpu_mem_track import MemTracker 72 | 73 | device = torch.device('cuda:0') 74 | 75 | frame = inspect.currentframe() # define a frame to track 76 | gpu_tracker = MemTracker(frame) # define a GPU tracker 77 | 78 | gpu_tracker.track() # run function between the code line where uses GPU 79 | cnn = models.vgg19(pretrained=True).features.to(device).eval() 80 | gpu_tracker.track() # run function between the code line where uses GPU 81 | 82 | dummy_tensor_1 = torch.randn(30, 3, 512, 512).float().to(device) # 30*3*512*512*4/1000/1000 = 94.37M 83 | dummy_tensor_2 = torch.randn(40, 3, 512, 512).float().to(device) # 40*3*512*512*4/1000/1000 = 125.82M 84 | dummy_tensor_3 = torch.randn(60, 3, 512, 512).float().to(device) # 60*3*512*512*4/1000/1000 = 188.74M 85 | 86 | gpu_tracker.track() 87 | 88 | dummy_tensor_4 = torch.randn(120, 3, 512, 512).float().to(device) # 120*3*512*512*4/1000/1000 = 377.48M 89 | dummy_tensor_5 = torch.randn(80, 3, 512, 512).float().to(device) # 80*3*512*512*4/1000/1000 = 251.64M 90 | 91 | gpu_tracker.track() 92 | 93 | dummy_tensor_4 = dummy_tensor_4.cpu() 94 | dummy_tensor_2 = dummy_tensor_2.cpu() 95 | torch.cuda.empty_cache() 96 | 97 | gpu_tracker.track() 98 | ``` 99 | This will output a .txt to current dir and the content of output is above(print content). 100 | 101 | # REFERENCE 102 | Part of the code is referenced from: 103 | 104 | http://jacobkimmel.github.io/pytorch_estimating_model_size/ 105 | https://gist.github.com/MInner/8968b3b120c95d3f50b8a22a74bf66bc 106 | 107 | -------------------------------------------------------------------------------- /utils/memory/gpu_mem_track.py: -------------------------------------------------------------------------------- 1 | import gc 2 | import datetime 3 | import pynvml 4 | 5 | import torch 6 | import numpy as np 7 | 8 | 9 | class MemTracker(object): 10 | """ 11 | Class used to track pytorch memory usage 12 | Arguments: 13 | frame: a frame to detect current py-file runtime 14 | detail(bool, default True): whether the function shows the detail gpu memory usage 15 | path(str): where to save log file 16 | verbose(bool, default False): whether show the trivial exception 17 | device(int): GPU number, default is 0 18 | """ 19 | def __init__(self, frame, detail=True, path='', verbose=False, device=0): 20 | self.frame = frame 21 | self.print_detail = detail 22 | self.last_tensor_sizes = set() 23 | self.gpu_profile_fn = path + f'{datetime.datetime.now():%d-%b-%y-%H:%M:%S}-gpu_mem_track.txt' 24 | self.verbose = verbose 25 | self.begin = True 26 | self.device = device 27 | 28 | self.func_name = frame.f_code.co_name 29 | self.filename = frame.f_globals["__file__"] 30 | if (self.filename.endswith(".pyc") or 31 | self.filename.endswith(".pyo")): 32 | self.filename = self.filename[:-1] 33 | self.module_name = self.frame.f_globals["__name__"] 34 | self.curr_line = self.frame.f_lineno 35 | 36 | def get_tensors(self): 37 | for obj in gc.get_objects(): 38 | try: 39 | if torch.is_tensor(obj) or (hasattr(obj, 'data') and torch.is_tensor(obj.data)): 40 | tensor = obj 41 | else: 42 | continue 43 | if tensor.is_cuda: 44 | yield tensor 45 | except Exception as e: 46 | if self.verbose: 47 | print('A trivial exception occured: {}'.format(e)) 48 | 49 | def track(self): 50 | """ 51 | Track the GPU memory usage 52 | """ 53 | pynvml.nvmlInit() 54 | handle = pynvml.nvmlDeviceGetHandleByIndex(self.device) 55 | meminfo = pynvml.nvmlDeviceGetMemoryInfo(handle) 56 | self.curr_line = self.frame.f_lineno 57 | where_str = self.module_name + ' ' + self.func_name + ':' + ' line ' + str(self.curr_line) 58 | 59 | with open(self.gpu_profile_fn, 'a+') as f: 60 | 61 | if self.begin: 62 | f.write(f"GPU Memory Track | {datetime.datetime.now():%d-%b-%y-%H:%M:%S} |" 63 | f" Total Used Memory:{meminfo.used/1000**2:<7.1f}Mb\n\n") 64 | self.begin = False 65 | 66 | if self.print_detail is True: 67 | ts_list = [tensor.size() for tensor in self.get_tensors()] 68 | new_tensor_sizes = {(type(x), tuple(x.size()), ts_list.count(x.size()), np.prod(np.array(x.size()))*4/1000**2) 69 | for x in self.get_tensors()} 70 | for t, s, n, m in new_tensor_sizes - self.last_tensor_sizes: 71 | f.write(f'+ | {str(n)} * Size:{str(s):<20} | Memory: {str(m*n)[:6]} M | {str(t):<20}\n') 72 | for t, s, n, m in self.last_tensor_sizes - new_tensor_sizes: 73 | f.write(f'- | {str(n)} * Size:{str(s):<20} | Memory: {str(m*n)[:6]} M | {str(t):<20} \n') 74 | self.last_tensor_sizes = new_tensor_sizes 75 | 76 | f.write(f"\nAt {where_str:<50}" 77 | f"Total Used Memory:{meminfo.used/1000**2:<7.1f}Mb\n\n") 78 | 79 | pynvml.nvmlShutdown() 80 | 81 | -------------------------------------------------------------------------------- /utils/memory/modelsize_estimate.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import numpy as np 4 | 5 | 6 | def modelsize(model, input, type_size=4): 7 | para = sum([np.prod(list(p.size())) for p in model.parameters()]) 8 | print('Model {} : Number of params: {}'.format(model._get_name(), para)) 9 | print('Model {} : params: {:4f}M'.format(model._get_name(), para * type_size / 1000 / 1000)) 10 | 11 | input_ = input.clone() 12 | input_.requires_grad_(requires_grad=False) 13 | 14 | mods = list(model.modules()) 15 | out_sizes = [] 16 | 17 | for i in range(1, len(mods)): 18 | m = mods[i] 19 | if isinstance(m, nn.ReLU): 20 | if m.inplace: 21 | continue 22 | out = m(input_) 23 | out_sizes.append(np.array(out.size())) 24 | input_ = out 25 | 26 | total_nums = 0 27 | for i in range(len(out_sizes)): 28 | s = out_sizes[i] 29 | nums = np.prod(np.array(s)) 30 | total_nums += nums 31 | 32 | # print('Model {} : Number of intermedite variables without backward: {}'.format(model._get_name(), total_nums)) 33 | # print('Model {} : Number of intermedite variables with backward: {}'.format(model._get_name(), total_nums*2)) 34 | print('Model {} : intermedite variables: {:3f} M (without backward)' 35 | .format(model._get_name(), total_nums * type_size / 1000 / 1000)) 36 | print('Model {} : intermedite variables: {:3f} M (with backward)' 37 | .format(model._get_name(), total_nums * type_size*2 / 1000 / 1000)) 38 | 39 | --------------------------------------------------------------------------------