repo
stringlengths
1
99
file
stringlengths
13
215
code
stringlengths
12
59.2M
file_length
int64
12
59.2M
avg_line_length
float64
3.82
1.48M
max_line_length
int64
12
2.51M
extension_type
stringclasses
1 value
pyslam
pyslam-master/utils_files.py
# From torchvision.uitls import smtplib, socket, os, os.path, hashlib, errno import __main__ as main from email.mime.text import MIMEText from os import path, mkdir def check_integrity(fpath, md5): if not os.path.isfile(fpath): return False md5o = hashlib.md5() with open(fpath, "rb") as f: # read in 1MB chunks for chunk in iter(lambda: f.read(1024 * 1024), b""): md5o.update(chunk) md5c = md5o.hexdigest() if md5c != md5: return False return True def download_url(url, root, filename, md5): from six.moves import urllib root = os.path.expanduser(root) fpath = os.path.join(root, filename) try: os.makedirs(root) except OSError as e: if e.errno == errno.EEXIST: pass else: raise # downloads file if os.path.isfile(fpath) and check_integrity(fpath, md5): print("Using downloaded and verified file: " + fpath) else: try: print("Downloading " + url + " to " + fpath) urllib.request.urlretrieve(url, fpath) except: if url[:5] == "https": url = url.replace("https:", "http:") print("Failed download. Trying https -> http instead." " Downloading " + url + " to " + fpath) urllib.request.urlretrieve(url, fpath) def check_dir(dir): if not path.isdir(dir): mkdir(dir) def list_dir(root, prefix=False): ### List all directories at a given root # root (str): Path to directory whose folders need to be listed # prefix (bool, optional): If true, prepends the path to each result, otherwise only returns the name of the directories found root = os.path.expanduser(root) directories = list(filter(lambda p: os.path.isdir(os.path.join(root, p)), os.listdir(root))) if prefix is True: directories = [os.path.join(root, d) for d in directories] return directories def list_files(root, suffix, prefix=False): ### List all files ending with a suffix at a given root # root (str): Path to directory whose folders need to be listed # suffix (str or tuple): Suffix of the files to match, e.g. '.png' or ('.jpg', '.png'). It uses the Python "str.endswith" method and is passed directly # prefix (bool, optional): If true, prepends the path to each result, otherwise only returns the name of the files found root = os.path.expanduser(root) files = list(filter(lambda p: os.path.isfile(os.path.join(root, p)) and p.endswith(suffix), os.listdir(root))) if prefix is True: files = [os.path.join(root, d) for d in files] return files def make_dirs(dir, exist_ok=True): # helper function, adds exist_ok to python2 if exist_ok: try: os.makedirs(dir) except: pass else: os.makedirs(dir) import smtplib, socket, os from email.mime.text import MIMEText # Open a plain text file for reading. For this example, assume that # the text file contains only ASCII characters. # with open(textfile) as fp: # Create a text/plain message # msg = MIMEText(fp.read()) def send_email(recipient, ignore_host="", login_gmail="", login_password=""): msg = MIMEText("") if socket.gethostname() == ignore_host: return msg["Subject"] = socket.gethostname() + " just finished running a job " + os.path.basename(main.__file__) msg["From"] = "clustersgpu@gmail.com" msg["To"] = recipient s = smtplib.SMTP_SSL("smtp.gmail.com", 465) s.ehlo() s.login(login_gmail, login_password) s.sendmail(login_gmail, recipient, msg.as_string()) s.quit()
3,694
31.412281
155
py
pyslam
pyslam-master/feature_hardnet.py
""" * This file is part of PYSLAM * adapted from https://github.com/DagnyT/hardnet/blob/master/examples/extract_hardnet_desc_from_hpatches_file.py, see licence therein. * * Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> * * PYSLAM is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * PYSLAM is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with PYSLAM. If not, see <http://www.gnu.org/licenses/>. """ # adapted from https://github.com/DagnyT/hardnet/blob/master/examples/extract_hardnet_desc_from_hpatches_file.py import config import sys import torch import torch.nn as nn import torch.nn.functional as F from torch.autograd import Variable import torch.backends.cudnn as cudnn import time import os import cv2 import math import numpy as np from utils_features import extract_patches_tensor, extract_patches_array, extract_patches_array_cpp kVerbose = True class L2Norm(nn.Module): def __init__(self): super(L2Norm,self).__init__() self.eps = 1e-10 def forward(self, x): norm = torch.sqrt(torch.sum(x * x, dim = 1) + self.eps) x= x / norm.unsqueeze(-1).expand_as(x) return x class L1Norm(nn.Module): def __init__(self): super(L1Norm,self).__init__() self.eps = 1e-10 def forward(self, x): norm = torch.sum(torch.abs(x), dim = 1) + self.eps x= x / norm.expand_as(x) return x class HardNet(nn.Module): """HardNet model definition """ def __init__(self): super(HardNet, self).__init__() self.features = nn.Sequential( nn.Conv2d(1, 32, kernel_size=3, padding=1, bias = False), nn.BatchNorm2d(32, affine=False), nn.ReLU(), nn.Conv2d(32, 32, kernel_size=3, padding=1, bias = False), nn.BatchNorm2d(32, affine=False), nn.ReLU(), nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1, bias = False), nn.BatchNorm2d(64, affine=False), nn.ReLU(), nn.Conv2d(64, 64, kernel_size=3, padding=1, bias = False), nn.BatchNorm2d(64, affine=False), nn.ReLU(), nn.Conv2d(64, 128, kernel_size=3, stride=2,padding=1, bias = False), nn.BatchNorm2d(128, affine=False), nn.ReLU(), nn.Conv2d(128, 128, kernel_size=3, padding=1, bias = False), nn.BatchNorm2d(128, affine=False), nn.ReLU(), nn.Dropout(0.1), nn.Conv2d(128, 128, kernel_size=8, bias = False), nn.BatchNorm2d(128, affine=False), ) #self.features.apply(weights_init) def input_norm(self,x): flat = x.view(x.size(0), -1) mp = torch.mean(flat, dim=1) sp = torch.std(flat, dim=1) + 1e-7 return (x - mp.detach().unsqueeze(-1).unsqueeze(-1).unsqueeze(-1).expand_as(x)) / sp.detach().unsqueeze(-1).unsqueeze(-1).unsqueeze(1).expand_as(x) def forward(self, input): x_features = self.features(self.input_norm(input)) x = x_features.view(x_features.size(0), -1) return L2Norm()(x) # interface for pySLAM class HardnetFeature2D: def __init__(self, do_cuda=True): print('Using HardnetFeature2D') self.model_base_path = config.cfg.root_folder + '/thirdparty/hardnet/' self.model_weights_path = self.model_base_path + 'pretrained/train_liberty_with_aug/checkpoint_liberty_with_aug.pth' #print('model_weights_path:',self.model_weights_path) self.do_cuda = do_cuda & torch.cuda.is_available() print('cuda:',self.do_cuda) device = torch.device("cuda:0" if self.do_cuda else "cpu") torch.set_grad_enabled(False) # mag_factor is how many times the original keypoint scale # is enlarged to generate a patch from a keypoint self.mag_factor = 1.0 # inference batch size self.batch_size = 512 self.process_all = True # process all the patches at once print('==> Loading pre-trained network.') self.model = HardNet() self.checkpoint = torch.load(self.model_weights_path) self.model.load_state_dict(self.checkpoint['state_dict']) if self.do_cuda: self.model.cuda() print('Extracting on GPU') else: print('Extracting on CPU') self.model = model.cpu() self.model.eval() print('==> Successfully loaded pre-trained network.') def compute_des_batches(self, patches): n_batches = int(len(patches) / self.batch_size) + 1 descriptors_for_net = np.zeros((len(patches), 128)) for i in range(0, len(patches), self.batch_size): data_a = patches[i: i + self.batch_size, :, :, :].astype(np.float32) data_a = torch.from_numpy(data_a) if self.do_cuda: data_a = data_a.cuda() data_a = Variable(data_a) # compute output with torch.no_grad(): out_a = self.model(data_a) descriptors_for_net[i: i + self.batch_size,:] = out_a.data.cpu().numpy().reshape(-1, 128) return descriptors_for_net def compute_des(self, patches): patches = torch.from_numpy(patches).float() patches = torch.unsqueeze(patches,1) if self.do_cuda: patches = patches.cuda() with torch.no_grad(): descrs = self.model(patches) return descrs.detach().cpu().numpy().reshape(-1, 128) def compute(self, img, kps, mask=None): #mask is a fake input num_kps = len(kps) des = [] if num_kps>0: if not self.process_all: # compute descriptor for each patch patches = extract_patches_tensor(img, kps, patch_size=32, mag_factor=self.mag_factor) # patches /= 255. # patches -= 0.443728476019 # patches /= 0.20197947209 patches = (patches/255. - 0.443728476019)/0.20197947209 des = self.compute_des_batches(patches).astype(np.float32) else: # compute descriptor by feeeding the full patch tensor to the network t = time.time() if False: # use python code patches = extract_patches_array(img, kps, patch_size=32, mag_factor=self.mag_factor) else: # use faster cpp code patches = extract_patches_array_cpp(img, kps, patch_size=32, mag_factor=self.mag_factor) patches = np.asarray(patches) if kVerbose: print('patches.shape:',patches.shape) # patches /= 255. # patches -= 0.443728476019 # patches /= 0.20197947209 patches = (patches/255. - 0.443728476019)/0.20197947209 if kVerbose: print('patch elapsed: ', time.time()-t) des = self.compute_des(patches) if kVerbose: print('descriptor: HARDNET, #features: ', len(kps), ', frame res: ', img.shape[0:2]) return kps, des
8,043
38.821782
155
py
pyslam
pyslam-master/feature_manager.py
""" * This file is part of PYSLAM * * Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> * * PYSLAM is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * PYSLAM is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with PYSLAM. If not, see <http://www.gnu.org/licenses/>. """ import sys import math from enum import Enum import numpy as np import cv2 from collections import Counter from parameters import Parameters from feature_types import FeatureDetectorTypes, FeatureDescriptorTypes, FeatureInfo from utils_sys import Printer, import_from from utils_features import unpackSiftOctaveKps, UnpackOctaveMethod, sat_num_features, kdt_nms, ssc_nms, octree_nms, grid_nms from utils_geom import hamming_distance, hamming_distances, l2_distance, l2_distances from feature_manager_adaptors import BlockAdaptor, PyramidAdaptor from pyramid import Pyramid, PyramidType from feature_root_sift import RootSIFTFeature2D from feature_shitomasi import ShiTomasiDetector # import and check SuperPointFeature2D = import_from('feature_superpoint', 'SuperPointFeature2D') TfeatFeature2D = import_from('feature_tfeat', 'TfeatFeature2D') Orbslam2Feature2D = import_from('feature_orbslam2', 'Orbslam2Feature2D') HardnetFeature2D = import_from('feature_hardnet', 'HardnetFeature2D') GeodescFeature2D = import_from('feature_geodesc', 'GeodescFeature2D') SosnetFeature2D = import_from('feature_sosnet', 'SosnetFeature2D') if False: L2NetKerasFeature2D = import_from('feature_l2net_keras', 'L2NetKerasFeature2D') # not used at present time L2NetFeature2D = import_from('feature_l2net', 'L2NetFeature2D') LogpolarFeature2D = import_from('feature_logpolar', 'LogpolarFeature2D') D2NetFeature2D = import_from('feature_d2net', 'D2NetFeature2D') DelfFeature2D = import_from('feature_delf', 'DelfFeature2D') ContextDescFeature2D = import_from('feature_contextdesc', 'ContextDescFeature2D') LfNetFeature2D = import_from('feature_lfnet', 'LfNetFeature2D') R2d2Feature2D = import_from('feature_r2d2', 'R2d2Feature2D') KeyNetDescFeature2D = import_from('feature_keynet', 'KeyNetDescFeature2D') DiskFeature2D = import_from('feature_disk', 'DiskFeature2D') kVerbose = True kNumFeatureDefault = Parameters.kNumFeatures kNumLevelsDefault = 4 kScaleFactorDefault = 1.2 kNumLevelsInitSigma = 40 kSigmaLevel0 = Parameters.kSigmaLevel0 kDrawOriginalExtractedFeatures = False # for debugging kFASTKeyPointSizeRescaleFactor = 4 # 7 is the standard keypoint size on layer 0 => actual size = 7*kFASTKeyPointSizeRescaleFactor kAGASTKeyPointSizeRescaleFactor = 4 # 7 is the standard keypoint size on layer 0 => actual size = 7*kAGASTKeyPointSizeRescaleFactor kShiTomasiKeyPointSizeRescaleFactor = 5 # 5 is the selected keypoint size on layer 0 (see below) => actual size = 5*kShiTomasiKeyPointSizeRescaleFactor if not kVerbose: def print(*args, **kwargs): pass class KeyPointFilterTypes(Enum): NONE = 0 SAT = 1 # sat the number of features (keep the best N features: 'best' on the basis of the keypoint.response) KDT_NMS = 2 # Non-Maxima Suppression based on kd-tree SSC_NMS = 3 # Non-Maxima Suppression based on https://github.com/BAILOOL/ANMS-Codes OCTREE_NMS = 4 # Distribute keypoints by using a octree (as a matter of fact, a quadtree): from ORBSLAM2 GRID_NMS = 5 # NMS by using a grid def feature_manager_factory(num_features=kNumFeatureDefault, num_levels = kNumLevelsDefault, # number of pyramid levels or octaves for detector and descriptor scale_factor = kScaleFactorDefault, # detection scale factor (if it can be set, otherwise it is automatically computed) detector_type = FeatureDetectorTypes.FAST, descriptor_type = FeatureDescriptorTypes.ORB): return FeatureManager(num_features, num_levels, scale_factor, detector_type, descriptor_type) # Manager of both detector and descriptor # This exposes an interface that is similar to OpenCV::Feature2D, i.e. detect(), compute() and detectAndCompute() class FeatureManager(object): def __init__(self, num_features=kNumFeatureDefault, num_levels = kNumLevelsDefault, # number of pyramid levels or octaves for detector and descriptor scale_factor = kScaleFactorDefault, # detection scale factor (if it can be set, otherwise it is automatically computed) detector_type = FeatureDetectorTypes.FAST, descriptor_type = FeatureDescriptorTypes.ORB): self.detector_type = detector_type self._feature_detector = None self.descriptor_type = descriptor_type self._feature_descriptor = None # main feature manager properties self.num_features = num_features self.num_levels = num_levels self.first_level = 0 # not always applicable = > 0: start pyramid from input image; # -1: start pyramid from up-scaled image*scale_factor (as in SIFT) self.scale_factor = scale_factor # scale factor bewteen two octaves self.sigma_level0 = kSigmaLevel0 # sigma on first octave self.layers_per_octave = 3 # for methods that uses octaves (SIFT, SURF, etc) # feature norm options self.norm_type = None # descriptor norm type self.descriptor_distance = None # pointer to a function for computing the distance between two points self.descriptor_distances = None # pointer to a function for computing the distances between two array of corresponding points # block adaptor options self.use_bock_adaptor = False self.block_adaptor = None # pyramid adaptor options: at present time pyramid adaptor has the priority and can combine a block adaptor withint itself self.use_pyramid_adaptor = False self.pyramid_adaptor = None self.pyramid_type = PyramidType.RESIZE self.pyramid_do_parallel = True self.do_sat_features_per_level = False # if pyramid adaptor is active, one can require to compute a certain number of features per level (see PyramidAdaptor) self.force_multiscale_detect_and_compute = False # automatically managed below depending on features self.oriented_features = True # automatically managed below depending on selected features self.do_keypoints_size_rescaling = False # automatically managed below depending on selected features self.need_color_image = False # automatically managed below depending on selected features self.keypoint_filter_type = KeyPointFilterTypes.SAT # default keypoint-filter type self.need_nms = False # need or not non-maximum suppression of keypoints self.keypoint_nms_filter_type = KeyPointFilterTypes.KDT_NMS # default keypoint-filter type if NMS is needed # initialize sigmas for keypoint levels (used for SLAM) self.init_sigma_levels() # --------------------------------------------- # # manage different opencv versions # --------------------------------------------- # print("using opencv ", cv2.__version__) # check opencv version in order to use the right modules opencv_major = int(cv2.__version__.split('.')[0]) opencv_minor = int(cv2.__version__.split('.')[1]) if opencv_major == 3: SIFT_create = import_from('cv2.xfeatures2d','SIFT_create') SURF_create = import_from('cv2.xfeatures2d','SURF_create') FREAK_create = import_from('cv2.xfeatures2d','FREAK_create') ORB_create = import_from('cv2','ORB_create') BRISK_create = import_from('cv2','BRISK_create') KAZE_create = import_from('cv2','KAZE_create') AKAZE_create = import_from('cv2','AKAZE_create') BoostDesc_create = import_from('cv2','xfeatures2d_BoostDesc','create') MSD_create = import_from('cv2','xfeatures2d_MSDDetector') # found but it does not work! (it does not find the .create() method) #Affine_create = import_from('cv2','xfeatures2d_AffineFeature2D') # not found DAISY_create = import_from('cv2','xfeatures2d_DAISY','create') STAR_create = import_from('cv2','xfeatures2d_StarDetector','create') HL_create = import_from('cv2','xfeatures2d_HarrisLaplaceFeatureDetector','create') LATCH_create = import_from('cv2','xfeatures2d_LATCH','create') LUCID_create = import_from('cv2','xfeatures2d_LUCID','create') VGG_create = import_from('cv2','xfeatures2d_VGG','create') BEBLID_create = import_from('cv2','xfeatures2d','BEBLID_create') elif opencv_major == 4 and opencv_minor >= 5: SIFT_create = import_from('cv2','SIFT_create') SURF_create = import_from('cv2.xfeatures2d','SURF_create') FREAK_create = import_from('cv2.xfeatures2d','FREAK_create') ORB_create = import_from('cv2','ORB_create') BRISK_create = import_from('cv2','BRISK_create') KAZE_create = import_from('cv2','KAZE_create') AKAZE_create = import_from('cv2','AKAZE_create') BoostDesc_create = import_from('cv2','xfeatures2d_BoostDesc','create') MSD_create = import_from('cv2','xfeatures2d_MSDDetector') DAISY_create = import_from('cv2','xfeatures2d_DAISY','create') STAR_create = import_from('cv2','xfeatures2d_StarDetector','create') HL_create = import_from('cv2','xfeatures2d_HarrisLaplaceFeatureDetector','create') LATCH_create = import_from('cv2','xfeatures2d_LATCH','create') LUCID_create = import_from('cv2','xfeatures2d_LUCID','create') VGG_create = import_from('cv2','xfeatures2d_VGG','create') BEBLID_create = import_from('cv2','xfeatures2d','BEBLID_create') else: SIFT_create = import_from('cv2.xfeatures2d','SIFT_create') SURF_create = import_from('cv2.xfeatures2d','SURF_create') FREAK_create = import_from('cv2.xfeatures2d','FREAK_create') ORB_create = import_from('cv2','ORB') BRISK_create = import_from('cv2','BRISK') KAZE_create = import_from('cv2','KAZE') AKAZE_create = import_from('cv2','AKAZE') BoostDesc_create = import_from('cv2','xfeatures2d_BoostDesc','create') MSD_create = import_from('cv2','xfeatures2d_MSDDetector') DAISY_create = import_from('cv2','xfeatures2d_DAISY','create') STAR_create = import_from('cv2','xfeatures2d_StarDetector','create') HL_create = import_from('cv2','xfeatures2d_HarrisLaplaceFeatureDetector','create') LATCH_create = import_from('cv2','xfeatures2d_LATCH','create') LUCID_create = import_from('cv2','xfeatures2d_LUCID','create') VGG_create = import_from('cv2','xfeatures2d_VGG','create') BEBLID_create = import_from('cv2','xfeatures2d','BEBLID_create') # pure detectors self.FAST_create = import_from('cv2','FastFeatureDetector_create') self.AGAST_create = import_from('cv2','AgastFeatureDetector_create') self.GFTT_create = import_from('cv2','GFTTDetector_create') self.MSER_create = import_from('cv2','MSER_create') self.MSD_create = MSD_create self.STAR_create = STAR_create self.HL_create = HL_create # detectors and descriptors self.SIFT_create = SIFT_create self.SURF_create = SURF_create self.ORB_create = ORB_create self.BRISK_create = BRISK_create self.AKAZE_create = AKAZE_create self.KAZE_create = KAZE_create # pure descriptors self.FREAK_create = FREAK_create # only descriptor self.BoostDesc_create = BoostDesc_create self.DAISY_create = DAISY_create self.LATCH_create = LATCH_create self.LUCID_create = LUCID_create self.VGG_create = VGG_create self.BEBLID_create = BEBLID_create # --------------------------------------------- # # check if we want descriptor == detector # --------------------------------------------- # self.is_detector_equal_to_descriptor = (self.detector_type.name == self.descriptor_type.name) # N.B.: the following descriptors assume keypoint.octave extacly represents an octave with a scale_factor=2 # and not a generic level with scale_factor < 2 if self.descriptor_type in [ FeatureDescriptorTypes.SIFT, # [NOK] SIFT seems to assume the use of octaves (https://github.com/opencv/opencv_contrib/blob/master/modules/xfeatures2d/src/sift.cpp#L1128) FeatureDescriptorTypes.ROOT_SIFT, # [NOK] same as SIFT #FeatureDescriptorTypes.SURF, # [OK] SURF computes the descriptor by considering the keypoint.size (https://github.com/opencv/opencv_contrib/blob/master/modules/xfeatures2d/src/surf.cpp#L600) FeatureDescriptorTypes.AKAZE, # [NOK] AKAZE does NOT seem to compute the right scale index for each keypoint.size (https://github.com/opencv/opencv/blob/master/modules/features2d/src/kaze/AKAZEFeatures.cpp#L1508) FeatureDescriptorTypes.KAZE, # [NOK] similar to KAZE #FeatureDescriptorTypes.FREAK, # [OK] FREAK computes the right scale index for each keypoint.size (https://github.com/opencv/opencv_contrib/blob/master/modules/xfeatures2d/src/freak.cpp#L468) #FeatureDescriptorTypes.BRISK # [OK] BRISK computes the right scale index for each keypoint.size (https://github.com/opencv/opencv/blob/master/modules/features2d/src/brisk.cpp#L697) #FeatureDescriptorTypes.BOOST_DESC # [OK] BOOST_DESC seems to properly rectify each keypoint patch size (https://github.com/opencv/opencv_contrib/blob/master/modules/xfeatures2d/src/boostdesc.cpp#L346) ]: self.scale_factor = 2 # the above descriptors work on octave layers with a scale_factor=2! Printer.orange('forcing scale factor=2 for detector', self.descriptor_type.name) self.orb_params = dict(nfeatures=num_features, scaleFactor=self.scale_factor, nlevels=self.num_levels, patchSize=31, edgeThreshold = 10, #31, #19, #10, # margin from the frame border fastThreshold = 20, firstLevel = self.first_level, WTA_K = 2, scoreType=cv2.ORB_FAST_SCORE) #scoreType=cv2.ORB_HARRIS_SCORE, scoreType=cv2.ORB_FAST_SCORE # --------------------------------------------- # # init detector # --------------------------------------------- # if self.detector_type == FeatureDetectorTypes.SIFT or self.detector_type == FeatureDetectorTypes.ROOT_SIFT: sift = self.SIFT_create(nOctaveLayers=self.layers_per_octave) self.set_sift_parameters() if self.detector_type == FeatureDetectorTypes.ROOT_SIFT: self._feature_detector = RootSIFTFeature2D(sift) else: self._feature_detector = sift # # elif self.detector_type == FeatureDetectorTypes.SURF: self._feature_detector = self.SURF_create(nOctaves = self.num_levels, nOctaveLayers=self.layers_per_octave) #self.intra_layer_factor = 1.2599 # num layers = nOctaves*nOctaveLayers scale=2^(1/nOctaveLayers) = 1.2599 self.scale_factor = 2 # force scale factor = 2 between octaves # # elif self.detector_type == FeatureDetectorTypes.ORB: self._feature_detector = self.ORB_create(**self.orb_params) self.use_bock_adaptor = True # add a block adaptor? self.need_nms = self.num_levels > 1 # ORB tends to generate overlapping keypoint on different levels <= KDT NMS seems to be very useful here! # # elif self.detector_type == FeatureDetectorTypes.ORB2: orb2_num_levels = self.num_levels self._feature_detector = Orbslam2Feature2D(self.num_features, self.scale_factor, orb2_num_levels) self.keypoint_filter_type = KeyPointFilterTypes.NONE # ORB2 cpp implementation already includes the algorithm OCTREE_NMS # # elif self.detector_type == FeatureDetectorTypes.BRISK: self._feature_detector = self.BRISK_create(octaves=self.num_levels) #self.intra_layer_factor = 1.3 # from the BRISK opencv code this seems to be the used scale factor between intra-octave frames #self.intra_layer_factor = math.sqrt(2) # approx, num layers = nOctaves*nOctaveLayers, from the BRISK paper there are octave ci and intra-octave di layers, t(ci)=2^i, t(di)=2^i * 1.5 self.scale_factor = 2 # force scale factor = 2 between octaves #self.keypoint_filter_type = KeyPointFilterTypes.NONE # # elif self.detector_type == FeatureDetectorTypes.KAZE: self._feature_detector = self.KAZE_create(nOctaves=self.num_levels, threshold=0.0005) # default: threshold = 0.001f self.scale_factor = 2 # force scale factor = 2 between octaves # # elif self.detector_type == FeatureDetectorTypes.AKAZE: self._feature_detector = self.AKAZE_create(nOctaves=self.num_levels, threshold=0.0005) # default: threshold = 0.001f self.scale_factor = 2 # force scale factor = 2 between octaves # # elif self.detector_type == FeatureDetectorTypes.SUPERPOINT: self.oriented_features = False self._feature_detector = SuperPointFeature2D() if self.descriptor_type != FeatureDescriptorTypes.NONE: self.use_pyramid_adaptor = self.num_levels > 1 self.need_nms = self.num_levels > 1 self.pyramid_type = PyramidType.GAUSS_PYRAMID self.pyramid_do_parallel = False # N.B.: SUPERPOINT interface class is not thread-safe! self.force_multiscale_detect_and_compute = True # force it since SUPERPOINT cannot compute descriptors separately from keypoints # # elif self.detector_type == FeatureDetectorTypes.FAST: self.oriented_features = False self._feature_detector = self.FAST_create(threshold=20, nonmaxSuppression=True) if self.descriptor_type != FeatureDescriptorTypes.NONE: #self.use_bock_adaptor = True # override a block adaptor? self.use_pyramid_adaptor = self.num_levels > 1 # override a pyramid adaptor? #self.pyramid_type = PyramidType.GAUSS_PYRAMID #self.first_level = 0 #self.do_sat_features_per_level = True self.need_nms = self.num_levels > 1 self.keypoint_nms_filter_type = KeyPointFilterTypes.OCTREE_NMS self.do_keypoints_size_rescaling = True # # elif self.detector_type == FeatureDetectorTypes.SHI_TOMASI: self.oriented_features = False self._feature_detector = ShiTomasiDetector(self.num_features) if self.descriptor_type != FeatureDescriptorTypes.NONE: #self.use_bock_adaptor = False # override a block adaptor? self.use_pyramid_adaptor = self.num_levels > 1 #self.pyramid_type = PyramidType.GAUSS_PYRAMID #self.first_level = 0 self.need_nms = self.num_levels > 1 self.keypoint_nms_filter_type = KeyPointFilterTypes.OCTREE_NMS self.do_keypoints_size_rescaling = True # # elif self.detector_type == FeatureDetectorTypes.AGAST: self.oriented_features = False self._feature_detector = self.AGAST_create(threshold=10, nonmaxSuppression=True) if self.descriptor_type != FeatureDescriptorTypes.NONE: #self.use_bock_adaptor = True # override a block adaptor? self.use_pyramid_adaptor = self.num_levels > 1 # override a pyramid adaptor? #self.pyramid_type = PyramidType.GAUSS_PYRAMID #self.first_level = 0 self.need_nms = self.num_levels > 1 self.keypoint_nms_filter_type = KeyPointFilterTypes.OCTREE_NMS self.do_keypoints_size_rescaling = True # # elif self.detector_type == FeatureDetectorTypes.GFTT: self.oriented_features = False self._feature_detector = self.GFTT_create(self.num_features, qualityLevel=0.01, minDistance=3, blockSize=5, useHarrisDetector=False, k=0.04) if self.descriptor_type != FeatureDescriptorTypes.NONE: #self.use_bock_adaptor = True # override a block adaptor? self.use_pyramid_adaptor = self.num_levels > 1 # override a pyramid adaptor? #self.pyramid_type = PyramidType.GAUSS_PYRAMID #self.first_level = 0 self.need_nms = self.num_levels > 1 self.keypoint_nms_filter_type = KeyPointFilterTypes.OCTREE_NMS self.do_keypoints_size_rescaling = True # # elif self.detector_type == FeatureDetectorTypes.MSER: self._feature_detector = self.MSER_create() #self.use_bock_adaptor = True # override a block adaptor? self.use_pyramid_adaptor = self.num_levels > 1 # override a pyramid adaptor? self.pyramid_do_parallel = False # parallel computations generate segmentation fault (is MSER thread-safe?) #self.pyramid_type = PyramidType.GAUSS_PYRAMID #self.first_level = 0 self.need_nms = self.num_levels > 1 #self.keypoint_nms_filter_type = KeyPointFilterTypes.OCTREE_NMS # # elif self.detector_type == FeatureDetectorTypes.MSD: #detector = ShiTomasiDetector(self.num_features) #self._feature_detector = self.MSD_create(detector) self._feature_detector = self.MSD_create() print('MSD detector info:',dir(self._feature_detector)) #self.use_bock_adaptor = True # override a block adaptor? #self.use_pyramid_adaptor = self.num_levels > 1 # override a pyramid adaptor? #self.pyramid_type = PyramidType.GAUSS_PYRAMID #self.first_level = 0 #self.need_nms = self.num_levels > 1 #self.keypoint_nms_filter_type = KeyPointFilterTypes.OCTREE_NMS # # elif self.detector_type == FeatureDetectorTypes.STAR: self.oriented_features = False self._feature_detector = self.STAR_create(maxSize=45, responseThreshold=10, # =30 lineThresholdProjected=10, lineThresholdBinarized=8, suppressNonmaxSize=5) if self.descriptor_type != FeatureDescriptorTypes.NONE: #self.use_bock_adaptor = True # override a block adaptor? self.use_pyramid_adaptor = self.num_levels > 1 # override a pyramid adaptor? #self.pyramid_type = PyramidType.GAUSS_PYRAMID #self.first_level = 0 #self.need_nms = self.num_levels > 1 #self.keypoint_nms_filter_type = KeyPointFilterTypes.OCTREE_NMS # # elif self.detector_type == FeatureDetectorTypes.HL: self.oriented_features = False self._feature_detector = self.HL_create(numOctaves=self.num_levels, corn_thresh=0.005, # = 0.01 DOG_thresh=0.01, # = 0.01 maxCorners=self.num_features, num_layers=4) # self.scale_factor = 2 # force scale factor = 2 between octaves # # elif self.detector_type == FeatureDetectorTypes.D2NET: self.need_color_image = True self.num_levels = 1 # force unless you have 12GB of VRAM multiscale=self.num_levels>1 self._feature_detector = D2NetFeature2D(multiscale=multiscale) #self.keypoint_filter_type = KeyPointFilterTypes.NONE # # elif self.detector_type == FeatureDetectorTypes.DELF: self.need_color_image = True #self.num_levels = 1 # force #scales are computed internally self._feature_detector = DelfFeature2D(num_features=self.num_features,score_threshold=20) self.scale_factor = self._feature_detector.scale_factor #self.keypoint_filter_type = KeyPointFilterTypes.NONE # # elif self.detector_type == FeatureDetectorTypes.CONTEXTDESC: self.set_sift_parameters() self.need_color_image = True #self.num_levels = 1 # force # computed internally by SIFT method self._feature_detector = ContextDescFeature2D(num_features=self.num_features) #self.keypoint_filter_type = KeyPointFilterTypes.NONE # # elif self.detector_type == FeatureDetectorTypes.LFNET: self.need_color_image = True #self.num_levels = 1 # force self._feature_detector = LfNetFeature2D(num_features=self.num_features) #self.keypoint_filter_type = KeyPointFilterTypes.NONE # # elif self.detector_type == FeatureDetectorTypes.R2D2: self.need_color_image = True #self.num_levels = - # internally recomputed self._feature_detector = R2d2Feature2D(num_features=self.num_features) self.scale_factor = self._feature_detector.scale_f self.keypoint_filter_type = KeyPointFilterTypes.NONE # # elif self.detector_type == FeatureDetectorTypes.KEYNET: #self.num_levels = - # internally recomputed self._feature_detector = KeyNetDescFeature2D(num_features=self.num_features) self.num_features = self._feature_detector.num_features self.num_levels = self._feature_detector.num_levels self.scale_factor = self._feature_detector.scale_factor self.keypoint_filter_type = KeyPointFilterTypes.NONE # # elif self.detector_type == FeatureDetectorTypes.DISK: self.num_levels = 1 # force self.need_color_image = True self._feature_detector = DiskFeature2D(num_features=self.num_features) # # else: raise ValueError("Unknown feature detector %s" % self.detector_type) if self.need_nms: self.keypoint_filter_type = self.keypoint_nms_filter_type if self.use_bock_adaptor: self.orb_params['edgeThreshold'] = 0 # --------------------------------------------- # # init descriptor # --------------------------------------------- # if self.is_detector_equal_to_descriptor: Printer.green('using same detector and descriptor object: ', self.detector_type.name) self._feature_descriptor = self._feature_detector else: # detector and descriptors are different self.num_levels_descriptor = self.num_levels if self.use_pyramid_adaptor: # NOT VALID ANYMORE -> if there is a pyramid adaptor, the descriptor does not need to rescale the images which are rescaled by the pyramid adaptor itself #self.orb_params['nlevels'] = 1 #self.num_levels_descriptor = 1 #self.num_levels pass # actual descriptor initialization if self.descriptor_type == FeatureDescriptorTypes.SIFT or self.descriptor_type == FeatureDescriptorTypes.ROOT_SIFT: sift = self.SIFT_create(nOctaveLayers=3) if self.descriptor_type == FeatureDescriptorTypes.ROOT_SIFT: self._feature_descriptor = RootSIFTFeature2D(sift) else: self._feature_descriptor = sift # # elif self.descriptor_type == FeatureDescriptorTypes.SURF: self.oriented_features = True # SURF computes the keypoint orientation self._feature_descriptor = self.SURF_create(nOctaves = self.num_levels_descriptor, nOctaveLayers=3) # # elif self.descriptor_type == FeatureDescriptorTypes.ORB: self._feature_descriptor = self.ORB_create(**self.orb_params) #self.oriented_features = False # N.B: ORB descriptor does not compute orientation on its own # # elif self.descriptor_type == FeatureDescriptorTypes.ORB2: self._feature_descriptor = self.ORB_create(**self.orb_params) # # elif self.descriptor_type == FeatureDescriptorTypes.BRISK: self.oriented_features = True # BRISK computes the keypoint orientation self._feature_descriptor = self.BRISK_create(octaves=self.num_levels_descriptor) # # elif self.descriptor_type == FeatureDescriptorTypes.KAZE: if not self.is_detector_equal_to_descriptor: Printer.red('WARNING: KAZE descriptors can only be used with KAZE or AKAZE keypoints.') # https://kyamagu.github.io/mexopencv/matlab/AKAZE.html self._feature_descriptor = self.KAZE_create(nOctaves=self.num_levels_descriptor) # # elif self.descriptor_type == FeatureDescriptorTypes.AKAZE: if not self.is_detector_equal_to_descriptor: Printer.red('WARNING: AKAZE descriptors can only be used with KAZE or AKAZE keypoints.') # https://kyamagu.github.io/mexopencv/matlab/AKAZE.html self._feature_descriptor = self.AKAZE_create(nOctaves=self.num_levels_descriptor) # # elif self.descriptor_type == FeatureDescriptorTypes.FREAK: self.oriented_features = True # FREAK computes the keypoint orientation self._feature_descriptor = self.FREAK_create(nOctaves=self.num_levels_descriptor) # # elif self.descriptor_type == FeatureDescriptorTypes.SUPERPOINT: if self.detector_type != FeatureDetectorTypes.SUPERPOINT: raise ValueError("You cannot use SUPERPOINT descriptor without SUPERPOINT detector!\nPlease, select SUPERPOINT as both descriptor and detector!") self._feature_descriptor = self._feature_detector # reuse the same SuperPointDector object # # elif self.descriptor_type == FeatureDescriptorTypes.TFEAT: self._feature_descriptor = TfeatFeature2D() # # elif self.descriptor_type == FeatureDescriptorTypes.BOOST_DESC: self.do_keypoints_size_rescaling = False # below a proper keypoint size scale factor is set depending on the used detector boost_des_keypoint_size_scale_factor = 1.5 # from https://docs.opencv.org/3.4.2/d1/dfd/classcv_1_1xfeatures2d_1_1BoostDesc.html#details #scale_factor: adjust the sampling window of detected keypoints 6.25f is default and fits for KAZE, SURF # detected keypoints window ratio 6.75f should be the scale for SIFT # detected keypoints window ratio 5.00f should be the scale for AKAZE, MSD, AGAST, FAST, BRISK # keypoints window ratio 0.75f should be the scale for ORB # keypoints ratio 1.50f was the default in original implementation if self.detector_type in [FeatureDetectorTypes.KAZE, FeatureDetectorTypes.SURF]: boost_des_keypoint_size_scale_factor = 6.25 elif self.detector_type == FeatureDetectorTypes.SIFT: boost_des_keypoint_size_scale_factor = 6.75 elif self.detector_type in [FeatureDetectorTypes.AKAZE, FeatureDetectorTypes.AGAST, FeatureDetectorTypes.FAST, FeatureDetectorTypes.BRISK]: boost_des_keypoint_size_scale_factor = 5.0 elif self.detector_type == FeatureDetectorTypes.ORB: boost_des_keypoint_size_scale_factor = 0.75 self._feature_descriptor = self.BoostDesc_create(scale_factor=boost_des_keypoint_size_scale_factor) # # elif self.descriptor_type == FeatureDescriptorTypes.DAISY: self._feature_descriptor = self.DAISY_create() # # elif self.descriptor_type == FeatureDescriptorTypes.LATCH: self._feature_descriptor = self.LATCH_create() # # elif self.descriptor_type == FeatureDescriptorTypes.LUCID: self._feature_descriptor = self.LUCID_create(lucid_kernel=1, # =1 blur_kernel=3 ) # =2 self.need_color_image = True # # elif self.descriptor_type == FeatureDescriptorTypes.VGG: self._feature_descriptor = self.VGG_create() # # elif self.descriptor_type == FeatureDescriptorTypes.HARDNET: self._feature_descriptor = HardnetFeature2D(do_cuda=True) # # elif self.descriptor_type == FeatureDescriptorTypes.GEODESC: self._feature_descriptor = GeodescFeature2D() # # elif self.descriptor_type == FeatureDescriptorTypes.SOSNET: self._feature_descriptor = SosnetFeature2D() # # elif self.descriptor_type == FeatureDescriptorTypes.L2NET: #self._feature_descriptor = L2NetKerasFeature2D() # keras-tf version self._feature_descriptor = L2NetFeature2D() # # elif self.descriptor_type == FeatureDescriptorTypes.LOGPOLAR: self._feature_descriptor = LogpolarFeature2D() # # elif self.descriptor_type == FeatureDescriptorTypes.D2NET: self.need_color_image = True if self.detector_type != FeatureDetectorTypes.D2NET: raise ValueError("You cannot use D2NET descriptor without D2NET detector!\nPlease, select D2NET as both descriptor and detector!") self._feature_descriptor = self._feature_detector # reuse detector object # # elif self.descriptor_type == FeatureDescriptorTypes.DELF: self.need_color_image = True if self.detector_type != FeatureDetectorTypes.DELF: raise ValueError("You cannot use DELF descriptor without DELF detector!\nPlease, select DELF as both descriptor and detector!") self._feature_descriptor = self._feature_detector # reuse detector object # # elif self.descriptor_type == FeatureDescriptorTypes.CONTEXTDESC: self.need_color_image = True if self.detector_type != FeatureDetectorTypes.CONTEXTDESC: raise ValueError("You cannot use CONTEXTDESC descriptor without CONTEXTDESC detector!\nPlease, select CONTEXTDESC as both descriptor and detector!") self._feature_descriptor = self._feature_detector # reuse detector object # # elif self.descriptor_type == FeatureDescriptorTypes.LFNET: self.need_color_image = True if self.detector_type != FeatureDetectorTypes.LFNET: raise ValueError("You cannot use LFNET descriptor without LFNET detector!\nPlease, select LFNET as both descriptor and detector!") self._feature_descriptor = self._feature_detector # reuse detector object # # elif self.descriptor_type == FeatureDescriptorTypes.R2D2: self.oriented_features = False self.need_color_image = True if self.detector_type != FeatureDetectorTypes.R2D2: raise ValueError("You cannot use R2D2 descriptor without R2D2 detector!\nPlease, select R2D2 as both descriptor and detector!") self._feature_descriptor = self._feature_detector # reuse detector object # # elif self.descriptor_type == FeatureDescriptorTypes.KEYNET: self.oriented_features = False if self.detector_type != FeatureDetectorTypes.KEYNET: raise ValueError("You cannot use KEYNET internal descriptor without KEYNET detector!\nPlease, select KEYNET as both descriptor and detector!") self._feature_descriptor = self._feature_detector # reuse detector object # # elif self.descriptor_type == FeatureDescriptorTypes.BEBLID: BEBLID_SIZE_256_BITS = 101 # https://docs.opencv.org/master/d7/d99/classcv_1_1xfeatures2d_1_1BEBLID.html BEBLID_scale_factor = 1.0 # it depends on the used detector https://docs.opencv.org/master/d7/d99/classcv_1_1xfeatures2d_1_1BEBLID.html#a38997aa059977abf6a2d6bf462d50de0a7b2a1e106c93d76cdfe5cef053277a04 # TODO: adapt BEBLID scale factor to actual used detector # 1.0 is OK for ORB2 detector self._feature_descriptor = self.BEBLID_create(BEBLID_scale_factor, BEBLID_SIZE_256_BITS) # # elif self.descriptor_type == FeatureDescriptorTypes.DISK: self.oriented_features = False if self.detector_type != FeatureDetectorTypes.DISK: raise ValueError("You cannot use DISK internal descriptor without DISK detector!\nPlease, select DISK as both descriptor and detector!") self._feature_descriptor = self._feature_detector # reuse detector object # # elif self.descriptor_type == FeatureDescriptorTypes.NONE: self._feature_descriptor = None else: raise ValueError("Unknown feature descriptor %s" % self.descriptor_type) # --------------------------------------------- # # init from FeatureInfo # --------------------------------------------- # # get and set norm type try: self.norm_type = FeatureInfo.norm_type[self.descriptor_type] except: Printer.red('You did not set the norm type for: ', self.descriptor_type.name) raise ValueError("Unmanaged norm type for feature descriptor %s" % self.descriptor_type.name) # set descriptor distance functions if self.norm_type == cv2.NORM_HAMMING: self.descriptor_distance = hamming_distance self.descriptor_distances = hamming_distances if self.norm_type == cv2.NORM_L2: self.descriptor_distance = l2_distance self.descriptor_distances = l2_distances # get and set reference max descriptor distance try: Parameters.kMaxDescriptorDistance = FeatureInfo.max_descriptor_distance[self.descriptor_type] except: Printer.red('You did not set the reference max descriptor distance for: ', self.descriptor_type.name) raise ValueError("Unmanaged max descriptor distance for feature descriptor %s" % self.descriptor_type.name) Parameters.kMaxDescriptorDistanceSearchEpipolar = Parameters.kMaxDescriptorDistance # --------------------------------------------- # # other required initializations # --------------------------------------------- # if not self.oriented_features: Printer.orange('WARNING: using NON-ORIENTED features: ', self.detector_type.name,'-',self.descriptor_type.name, ' (i.e. kp.angle=0)') if self.is_detector_equal_to_descriptor and \ ( self.detector_type == FeatureDetectorTypes.SIFT or self.detector_type == FeatureDetectorTypes.ROOT_SIFT or self.detector_type == FeatureDetectorTypes.CONTEXTDESC ): self.init_sigma_levels_sift() else: self.init_sigma_levels() if self.use_bock_adaptor: self.block_adaptor = BlockAdaptor(self._feature_detector, self._feature_descriptor) if self.use_pyramid_adaptor: self.pyramid_params = dict(detector=self._feature_detector, descriptor=self._feature_descriptor, num_features = self.num_features, num_levels=self.num_levels, scale_factor=self.scale_factor, sigma0=self.sigma_level0, first_level=self.first_level, pyramid_type=self.pyramid_type, use_block_adaptor=self.use_bock_adaptor, do_parallel = self.pyramid_do_parallel, do_sat_features_per_level = self.do_sat_features_per_level) self.pyramid_adaptor = PyramidAdaptor(**self.pyramid_params) def set_sift_parameters(self): # N.B.: The number of SIFT octaves is automatically computed from the image resolution, # here we can set the number of layers in each octave. # from https://docs.opencv.org/3.4/d5/d3c/classcv_1_1xfeatures2d_1_1SIFT.html #self.intra_layer_factor = 1.2599 # num layers = nOctaves*nOctaveLayers scale=2^(1/nOctaveLayers) = 1.2599 self.scale_factor = 2 # force scale factor = 2 between octaves self.sigma_level0 = 1.6 # https://github.com/opencv/opencv/blob/173442bb2ecd527f1884d96d7327bff293f0c65a/modules/nonfree/src/sift.cpp#L118 # from https://docs.opencv.org/3.1.0/da/df5/tutorial_py_sift_intro.html self.first_level = -1 # https://github.com/opencv/opencv/blob/173442bb2ecd527f1884d96d7327bff293f0c65a/modules/nonfree/src/sift.cpp#L731 # initialize scale factors, sigmas for each octave level; # these are used for managing image pyramids and weighting (information matrix) reprojection error terms in the optimization def init_sigma_levels(self): print('num_levels: ', self.num_levels) num_levels = max(kNumLevelsInitSigma, self.num_levels) self.inv_scale_factor = 1./self.scale_factor self.scale_factors = np.zeros(num_levels) self.level_sigmas2 = np.zeros(num_levels) self.level_sigmas = np.zeros(num_levels) self.inv_scale_factors = np.zeros(num_levels) self.inv_level_sigmas2 = np.zeros(num_levels) self.log_scale_factor = math.log(self.scale_factor) self.scale_factors[0] = 1.0 self.level_sigmas2[0] = self.sigma_level0*self.sigma_level0 self.level_sigmas[0] = math.sqrt(self.level_sigmas2[0]) for i in range(1,num_levels): self.scale_factors[i] = self.scale_factors[i-1]*self.scale_factor self.level_sigmas2[i] = self.scale_factors[i]*self.scale_factors[i]*self.level_sigmas2[0] self.level_sigmas[i] = math.sqrt(self.level_sigmas2[i]) for i in range(num_levels): self.inv_scale_factors[i] = 1.0/self.scale_factors[i] self.inv_level_sigmas2[i] = 1.0/self.level_sigmas2[i] #print('self.scale_factor: ', self.scale_factor) #print('self.scale_factors: ', self.scale_factors) #print('self.level_sigmas: ', self.level_sigmas) #print('self.inv_scale_factors: ', self.inv_scale_factors) # initialize scale factors, sigmas for each octave level; # these are used for managing image pyramids and weighting (information matrix) reprojection error terms in the optimization; # this method can be used only when the following mapping is adopted for SIFT: # keypoint.octave = (unpacked_octave+1)*3+unpacked_layer where S=3 is the number of levels per octave def init_sigma_levels_sift(self): print('initializing SIFT sigma levels') print('num_levels: ', self.num_levels) self.num_levels = 3*self.num_levels + 3 # we map: level=keypoint.octave = (unpacked_octave+1)*3+unpacked_layer where S=3 is the number of scales per octave num_levels = max(kNumLevelsInitSigma, self.num_levels) #print('num_levels: ', num_levels) # N.B: if we adopt the mapping: keypoint.octave = (unpacked_octave+1)*3+unpacked_layer # then we can consider a new virtual scale_factor = 2^(1/3) (used between two contiguous layers of the same octave) print('original scale factor: ', self.scale_factor) self.scale_factor = math.pow(2,1./3) self.inv_scale_factor = 1./self.scale_factor self.scale_factors = np.zeros(num_levels) self.level_sigmas2 = np.zeros(num_levels) self.level_sigmas = np.zeros(num_levels) self.inv_scale_factors = np.zeros(num_levels) self.inv_level_sigmas2 = np.zeros(num_levels) self.log_scale_factor = math.log(self.scale_factor) self.sigma_level0 = 1.6 # https://github.com/opencv/opencv/blob/173442bb2ecd527f1884d96d7327bff293f0c65a/modules/nonfree/src/sift.cpp#L118 # from https://docs.opencv.org/3.1.0/da/df5/tutorial_py_sift_intro.html sigma_level02 = self.sigma_level0*self.sigma_level0 # N.B.: these are used only when recursive filtering is applied: see https://www.vlfeat.org/api/sift.html#sift-tech-ss #sift_init_sigma = 0.5 #sift_init_sigma2 = 0.25 # see also https://www.vlfeat.org/api/sift.html self.scale_factors[0] = 1.0 self.level_sigmas2[0] = sigma_level02 # -4*sift_init_sigma2 N.B.: this is an absolute sigma, # not a delta_sigma used for incrementally filtering contiguos layers => we must not subtract (4*sift_init_sigma2) # https://github.com/opencv/opencv/blob/173442bb2ecd527f1884d96d7327bff293f0c65a/modules/nonfree/src/sift.cpp#L197 self.level_sigmas[0] = math.sqrt(self.level_sigmas2[0]) for i in range(1,num_levels): self.scale_factors[i] = self.scale_factors[i-1]*self.scale_factor self.level_sigmas2[i] = self.scale_factors[i]*self.scale_factors[i]*sigma_level02 # https://github.com/opencv/opencv/blob/173442bb2ecd527f1884d96d7327bff293f0c65a/modules/nonfree/src/sift.cpp#L224 self.level_sigmas[i] = math.sqrt(self.level_sigmas2[i]) for i in range(num_levels): self.inv_scale_factors[i] = 1.0/self.scale_factors[i] self.inv_level_sigmas2[i] = 1.0/self.level_sigmas2[i] #print('self.scale_factor: ', self.scale_factor) #print('self.scale_factors: ', self.scale_factors) #print('self.level_sigmas: ', self.level_sigmas) #print('self.inv_scale_factors: ', self.inv_scale_factors) # filter matches by using # Non-Maxima Suppression (NMS) based on kd-trees # or SSC NMS (https://github.com/BAILOOL/ANMS-Codes) # or SAT (get features with best responses) # or OCTREE_NMS (implemented in ORBSLAM2, distribution of features in a quad-tree) def filter_keypoints(self, type, frame, kps, des=None): filter_name = type.name if type == KeyPointFilterTypes.NONE: pass elif type == KeyPointFilterTypes.KDT_NMS: kps, des = kdt_nms(kps, des, self.num_features) elif type == KeyPointFilterTypes.SSC_NMS: kps, des = ssc_nms(kps, des, frame.shape[1], frame.shape[0], self.num_features) elif type == KeyPointFilterTypes.OCTREE_NMS: if des is not None: raise ValueError('at the present time, you cannot use OCTREE_NMS with descriptors') kps = octree_nms(frame, kps, self.num_features) elif type == KeyPointFilterTypes.GRID_NMS: kps, des, _ = grid_nms(kps, des, frame.shape[0], frame.shape[1], self.num_features, dist_thresh=4) elif type == KeyPointFilterTypes.SAT: if len(kps) > self.num_features: kps, des = sat_num_features(kps, des, self.num_features) else: raise ValueError("Unknown match-filter type") return kps, des, filter_name def rescale_keypoint_size(self, kps): # if keypoints are FAST, etc. then rescale their small sizes # in order to let descriptors compute an encoded representation with a decent patch size scale = 1 doit = False if self.detector_type == FeatureDetectorTypes.FAST: scale = kFASTKeyPointSizeRescaleFactor doit = True elif self.detector_type == FeatureDetectorTypes.AGAST: scale = kAGASTKeyPointSizeRescaleFactor doit = True elif self.detector_type == FeatureDetectorTypes.SHI_TOMASI or self.detector_type == FeatureDetectorTypes.GFTT: scale = kShiTomasiKeyPointSizeRescaleFactor doit = True if doit: for kp in kps: kp.size *= scale # detect keypoints without computing their descriptors # out: kps (array of cv2.KeyPoint) def detect(self, frame, mask=None, filter=True): if not self.need_color_image and frame.ndim>2: # check if we have to convert to gray image frame = cv2.cvtColor(frame,cv2.COLOR_RGB2GRAY) if self.use_pyramid_adaptor: # detection with pyramid adaptor (it can optionally include a block adaptor per level) kps = self.pyramid_adaptor.detect(frame, mask) elif self.use_bock_adaptor: # detection with block adaptor kps = self.block_adaptor.detect(frame, mask) else: # standard detection kps = self._feature_detector.detect(frame, mask) # filter keypoints filter_name = 'NONE' if filter: kps, _, filter_name = self.filter_keypoints(self.keypoint_filter_type, frame, kps) # if keypoints are FAST, etc. give them a decent size in order to properly compute the descriptors if self.do_keypoints_size_rescaling: self.rescale_keypoint_size(kps) if kDrawOriginalExtractedFeatures: # draw the original features imgDraw = cv2.drawKeypoints(frame, kps, None, color=(0,255,0), flags=0) cv2.imshow('detected keypoints',imgDraw) if kVerbose: print('detector:',self.detector_type.name,', #features:', len(kps),', [kp-filter:',filter_name,']') return kps # compute the descriptors once given the keypoints def compute(self, frame, kps, filter = True): if not self.need_color_image and frame.ndim>2: # check if we have to convert to gray image frame = cv2.cvtColor(frame,cv2.COLOR_RGB2GRAY) kps, des = self._feature_descriptor.compute(frame, kps) # then, compute descriptors # filter keypoints filter_name = 'NONE' if filter: kps, des, filter_name = self.filter_keypoints(self.keypoint_filter_type, frame, kps, des) if kVerbose: print('descriptor:',self.descriptor_type.name,', #features:', len(kps),', [kp-filter:',filter_name,']') return kps, des # detect keypoints and their descriptors # out: kps, des def detectAndCompute(self, frame, mask=None, filter = True): if not self.need_color_image and frame.ndim>2: # check if we have to convert to gray image frame = cv2.cvtColor(frame,cv2.COLOR_RGB2GRAY) if self.use_pyramid_adaptor: # detectAndCompute with pyramid adaptor (it can optionally include a block adaptor per level) if self.force_multiscale_detect_and_compute: # force detectAndCompute on each level instead of first {detect() on each level} and then {compute() on resulting detected keypoints one time} kps, des = self.pyramid_adaptor.detectAndCompute(frame, mask) # else: kps = self.detect(frame, mask, filter=True) # first, detect by using adaptor on the different pyramid levels kps, des = self.compute(frame, kps, filter=False) # then, separately compute the descriptors on detected keypoints (one time) filter = False # disable keypoint filtering since we already applied it for detection elif self.use_bock_adaptor: # detectAndCompute with block adaptor (force detect/compute on each block) # #kps, des = self.block_adaptor.detectAndCompute(frame, mask) # kps = self.detect(frame, mask, filter=True) # first, detect by using adaptor kps, des = self.compute(frame, kps, filter=False) # then, separately compute the descriptors filter = False # disable keypoint filtering since we already applied it for detection else: # standard detectAndCompute if self.is_detector_equal_to_descriptor: # detector = descriptor => call them together with detectAndCompute() method kps, des = self._feature_detector.detectAndCompute(frame, mask) if kVerbose: print('detector:', self.detector_type.name,', #features:',len(kps)) print('descriptor:', self.descriptor_type.name,', #features:',len(kps)) else: # detector and descriptor are different => call them separately # 1. first, detect keypoint locations kps = self.detect(frame, mask, filter=False) # 2. then, compute descriptors kps, des = self._feature_descriptor.compute(frame, kps) if kVerbose: #print('detector: ', self.detector_type.name, ', #features: ', len(kps)) print('descriptor: ', self.descriptor_type.name, ', #features: ', len(kps)) # filter keypoints filter_name = 'NONE' if filter: kps, des, filter_name = self.filter_keypoints(self.keypoint_filter_type, frame, kps, des) if self.detector_type == FeatureDetectorTypes.SIFT or \ self.detector_type == FeatureDetectorTypes.ROOT_SIFT or \ self.detector_type == FeatureDetectorTypes.CONTEXTDESC : unpackSiftOctaveKps(kps, method=UnpackOctaveMethod.INTRAL_LAYERS) if kVerbose: print('detector:',self.detector_type.name,', descriptor:', self.descriptor_type.name,', #features:', len(kps),' (#ref:', self.num_features, '), [kp-filter:',filter_name,']') self.debug_print(kps) return kps, des def debug_print(self, kps): if False: # raw print of all keypoints for k in kps: print("response: ", k.response, "\t, size: ", k.size, "\t, octave: ", k.octave, "\t, angle: ", k.angle) if False: # generate a rough histogram for keypoint sizes kps_sizes = [kp.size for kp in kps] kps_sizes_histogram = np.histogram(kps_sizes, bins=10) print('size-histogram: \n', list(zip(kps_sizes_histogram[1],kps_sizes_histogram[0]))) # generate histogram at level 0 kps_sizes = [kp.size for kp in kps if kp.octave==1] kps_sizes_histogram = np.histogram(kps_sizes, bins=10) print('size-histogram at level 0: \n', list(zip(kps_sizes_histogram[1],kps_sizes_histogram[0]))) if False: # count points for each octave => generate an octave histogram kps_octaves = [k.octave for k in kps] kps_octaves = Counter(kps_octaves) print('levels-histogram: ', kps_octaves.most_common(12))
63,509
62.957704
272
py
pyslam
pyslam-master/feature_delf.py
""" * This file is part of PYSLAM * Adapted from https://github.com/tensorflow/models/blob/master/research/delf/delf/python/examples/extract_features.py, see the license therein. * * Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> * * PYSLAM is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * PYSLAM is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with PYSLAM. If not, see <http://www.gnu.org/licenses/>. """ import config config.cfg.set_lib('delf') import cv2 from threading import RLock from utils_sys import Printer import warnings # to disable tensorflow-numpy warnings: from https://github.com/tensorflow/tensorflow/issues/30427 warnings.filterwarnings('ignore', category=FutureWarning) import argparse import os import sys import time import json import numpy as np import h5py if False: import tensorflow as tf else: # from https://stackoverflow.com/questions/56820327/the-name-tf-session-is-deprecated-please-use-tf-compat-v1-session-instead import tensorflow.compat.v1 as tf # from https://kobkrit.com/using-allow-growth-memory-option-in-tensorflow-and-keras-dc8c8081bc96 to cope with the following error: # "[...tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR" tf_config = tf.ConfigProto() tf_config.gpu_options.allow_growth = True tf_config.gpu_options.per_process_gpu_memory_fraction=0.333 # from https://stackoverflow.com/questions/34199233/how-to-prevent-tensorflow-from-allocating-the-totality-of-a-gpu-memory #session = tf.Session(config=tf_config, ...) from google.protobuf import text_format from tensorflow.python.platform import app # from delf import delf_config_pb2 # from delf import feature_extractor # from delf import feature_io from delf.protos import aggregation_config_pb2 from delf.protos import box_pb2 from delf.protos import datum_pb2 from delf.protos import delf_config_pb2 from delf.protos import feature_pb2 from delf.python import box_io from delf.python import datum_io from delf.python import delf_v1 from delf.python import feature_aggregation_extractor from delf.python import feature_aggregation_similarity from delf.python import feature_extractor from delf.python import feature_io from delf.python.examples import detector from delf.python.examples import extractor from delf.python import detect_to_retrieve from delf.python import google_landmarks_dataset from utils_tf import set_tf_logging #from utils import print_options delf_base_path = config.cfg.root_folder + '/thirdparty/tensorflow_models/research/delf/delf/python/' delf_config_file = delf_base_path + 'examples/delf_config_example.pbtxt' delf_model_path = delf_base_path + 'examples/parameters/delf_gld_20190411/model/' delf_mean_path = delf_base_path + 'examples/parameters/delf_gld_20190411/pca/mean.datum' delf_projection_matrix_path = delf_base_path + 'examples/parameters/delf_gld_20190411/pca/pca_proj_mat.datum' kVerbose = True def MakeExtractor(sess, config, import_scope=None): """Creates a function to extract features from an image. Args: sess: TensorFlow session to use. config: DelfConfig proto containing the model configuration. import_scope: Optional scope to use for model. Returns: Function that receives an image and returns features. """ tf.saved_model.loader.load( sess, [tf.saved_model.tag_constants.SERVING], config.model_path, import_scope=import_scope) import_scope_prefix = import_scope + '/' if import_scope is not None else '' input_image = sess.graph.get_tensor_by_name('%sinput_image:0' % import_scope_prefix) input_score_threshold = sess.graph.get_tensor_by_name('%sinput_abs_thres:0' % import_scope_prefix) input_image_scales = sess.graph.get_tensor_by_name('%sinput_scales:0' % import_scope_prefix) input_max_feature_num = sess.graph.get_tensor_by_name('%sinput_max_feature_num:0' % import_scope_prefix) boxes = sess.graph.get_tensor_by_name('%sboxes:0' % import_scope_prefix) raw_descriptors = sess.graph.get_tensor_by_name('%sfeatures:0' % import_scope_prefix) feature_scales = sess.graph.get_tensor_by_name('%sscales:0' % import_scope_prefix) attention_with_extra_dim = sess.graph.get_tensor_by_name('%sscores:0' % import_scope_prefix) attention = tf.reshape(attention_with_extra_dim,[tf.shape(attention_with_extra_dim)[0]]) locations, descriptors = feature_extractor.DelfFeaturePostProcessing(boxes, raw_descriptors, config) def ExtractorFn(image): """Receives an image and returns DELF features. Args: image: Uint8 array with shape (height, width 3) containing the RGB image. Returns: Tuple (locations, descriptors, feature_scales, attention) """ return sess.run([locations, descriptors, feature_scales, attention], feed_dict={ input_image: image, input_score_threshold: config.delf_local_config.score_threshold, input_image_scales: list(config.image_scales), input_max_feature_num: config.delf_local_config.max_feature_num }) return ExtractorFn # convert matrix of pts into list of keypoints def convert_pts_to_keypoints(pts, scores, sizes): assert(len(pts)==len(scores)) kps = [] if pts is not None: # convert matrix [Nx2] of pts into list of keypoints kps = [ cv2.KeyPoint(p[0], p[1], _size=sizes[i], _response=scores[i]) for i,p in enumerate(pts) ] return kps # interface for pySLAM class DelfFeature2D: def __init__(self, num_features=1000, score_threshold=100, do_tf_logging=False): print('Using DelfFeature2D') self.lock = RLock() set_tf_logging(do_tf_logging) # Parse DelfConfig proto. self.delf_config = delf_config_pb2.DelfConfig() with tf.gfile.FastGFile(delf_config_file, 'r') as f: text_format.Merge(f.read(), self.delf_config) self.delf_config.model_path = delf_model_path self.delf_config.delf_local_config.pca_parameters.mean_path = delf_mean_path self.delf_config.delf_local_config.pca_parameters.projection_matrix_path = delf_projection_matrix_path self.delf_config.delf_local_config.max_feature_num = num_features self.delf_config.delf_local_config.score_threshold = score_threshold print('DELF CONFIG\n:', self.delf_config) self.keypoint_size = 30 # just a representative size for visualization and in order to convert extracted points to cv2.KeyPoint self.image_scales = list(self.delf_config.image_scales) #print('image scales: ',self.image_scales) try: self.scale_factor = self.image_scales[1]/self.image_scales[0] except: self.scale_factor = np.sqrt(2) # according to default config and the paper #print('scale_factor: ',self.scale_factor) #self.image_levels = np.round(-np.log(self.image_scales)/np.log(self.scale_factor)).astype(np.int32) #print('image levels: ',self.image_levels) self.session = None self.pts = [] self.kps = [] self.des = [] self.scales = [] self.scores = [] self.frame = None print('==> Loading pre-trained network.') self.load_model() print('==> Successfully loaded pre-trained network.') @property def num_features(self): return self.delf_config.delf_local_config.max_feature_num @property def score_threshold(self): return self.delf_config.delf_local_config.score_threshold def __del__(self): self.close() def load_model(self): # Create graph before session :) self.graph = tf.Graph().as_default() self.session = tf.Session() init_op = tf.global_variables_initializer() self.session.run(init_op) self.extractor_fn = MakeExtractor(self.session, self.delf_config) def close(self): if self.session is not None: print('DELF: closing tf session') self.session.close() tf.reset_default_graph() def compute_kps_des(self, frame): with self.lock: image_tf = tf.convert_to_tensor(frame, np.float32) im = self.session.run(image_tf) # Extract and save features. (locations_out, descriptors_out, feature_scales_out, attention_out) = self.extractor_fn(im) self.pts = locations_out[:, ::-1] self.des = descriptors_out self.scales = feature_scales_out self.scores = attention_out # N.B.: according to the paper "Large-Scale Image Retrieval with Attentive Deep Local Features": # We construct image pyramids by using scales that are a 2 factor apart. For the set of scales # with range from 0.25 to 2.0, 7 different scales are used. # The size of receptive field is inversely proportional to the scale; for example, for the 2.0 scale, the # receptive field of the network covers 146 × 146 pixels. # The receptive field size for the image at the original scale is 291 × 291. #sizes = self.keypoint_size * 1./self.scales sizes = self.keypoint_size * self.scales if False: # print('kps.shape', self.pts.shape) # print('des.shape', self.des.shape) # print('scales.shape', self.scales.shape) # print('scores.shape', self.scores.shape) print('scales:',self.scales) print('sizes:',sizes) self.kps = convert_pts_to_keypoints(self.pts, self.scores, sizes) return self.kps, self.des def detectAndCompute(self, frame, mask=None): #mask is a fake input with self.lock: self.frame = frame self.kps, self.des = self.compute_kps_des(frame) if kVerbose: print('detector: DELF, descriptor: DELF, #features: ', len(self.kps), ', frame res: ', frame.shape[0:2]) return self.kps, self.des # return keypoints if available otherwise call detectAndCompute() def detect(self, frame, mask=None): # mask is a fake input with self.lock: #if self.frame is not frame: self.detectAndCompute(frame) return self.kps # return descriptors if available otherwise call detectAndCompute() def compute(self, frame, kps=None, mask=None): # kps is a fake input, mask is a fake input with self.lock: if self.frame is not frame: Printer.orange('WARNING: DELF is recomputing both kps and des on last input frame', frame.shape) self.detectAndCompute(frame) return self.kps, self.des
12,100
40.871972
183
py
pyslam
pyslam-master/feature_keynet.py
""" * This file is part of PYSLAM * * Adpated from https://raw.githubusercontent.com/axelBarroso/Key.Net/master/extract_multiscale_features.py, see the license therein. * Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> * * PYSLAM is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * PYSLAM is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with PYSLAM. If not, see <http://www.gnu.org/licenses/>. """ import config config.cfg.set_lib('keynet') import warnings # to disable tensorflow-numpy warnings: from https://github.com/tensorflow/tensorflow/issues/30427 warnings.filterwarnings('ignore', category=FutureWarning) import os, sys, cv2 #sys.path.append(os.path.dirname(os.path.dirname(os.path.realpath(__file__)))) from os import path, mkdir import argparse import keyNet.aux.tools as aux from skimage.transform import pyramid_gaussian import HSequences_bench.tools.geometry_tools as geo_tools import HSequences_bench.tools.repeatability_tools as rep_tools from keyNet.model.keynet_architecture import * import keyNet.aux.desc_aux_function as loss_desc from keyNet.model.hardnet_pytorch import * from keyNet.datasets.dataset_utils import read_bw_image import torch from threading import RLock from utils_tf import set_tf_logging from utils_sys import Printer, print_options kVerbose = True def build_keynet_config(keynet_base_path): parser = argparse.ArgumentParser(description='HSequences Extract Features') # parser.add_argument('--list-images', type=str, help='File containing the image paths for extracting features.', # required=True) # parser.add_argument('--results-dir', type=str, default='extracted_features/', # help='The output path to save the extracted keypoint.') parser.add_argument('--network-version', type=str, default='KeyNet_default', help='The Key.Net network version name') parser.add_argument('--checkpoint-det-dir', type=str, default=keynet_base_path + 'keyNet/pretrained_nets/KeyNet_default', help='The path to the checkpoint file to load the detector weights.') parser.add_argument('--pytorch-hardnet-dir', type=str, default=keynet_base_path + 'keyNet/pretrained_nets/HardNet++.pth', help='The path to the checkpoint file to load the HardNet descriptor weights.') # Detector Settings parser.add_argument('--num-filters', type=int, default=8, help='The number of filters in each learnable block.') parser.add_argument('--num-learnable-blocks', type=int, default=3, help='The number of learnable blocks after handcrafted block.') parser.add_argument('--num-levels-within-net', type=int, default=3, help='The number of pyramid levels inside the architecture.') parser.add_argument('--factor-scaling-pyramid', type=float, default=1.2, help='The scale factor between the multi-scale pyramid levels in the architecture.') parser.add_argument('--conv-kernel-size', type=int, default=5, help='The size of the convolutional filters in each of the learnable blocks.') # Multi-Scale Extractor Settings parser.add_argument('--extract-MS', type=bool, default=True, help='Set to True if you want to extract multi-scale features.') parser.add_argument('--num-points', type=int, default=2000, help='The number of desired features to extract.') parser.add_argument('--nms-size', type=int, default=15, help='The NMS size for computing the validation repeatability.') parser.add_argument('--border-size', type=int, default=15, help='The number of pixels to remove from the borders to compute the repeatability.') parser.add_argument('--order-coord', type=str, default='xysr', help='The coordinate order that follows the extracted points. Use yxsr or xysr.') parser.add_argument('--random-seed', type=int, default=12345, help='The random seed value for TensorFlow and Numpy.') parser.add_argument('--pyramid_levels', type=int, default=5, help='The number of downsample levels in the pyramid.') parser.add_argument('--upsampled-levels', type=int, default=1, help='The number of upsample levels in the pyramid.') parser.add_argument('--scale-factor-levels', type=float, default=np.sqrt(2), help='The scale factor between the pyramid levels.') parser.add_argument('--scale-factor', type=float, default=2., help='The scale factor to extract patches before descriptor.') # GPU Settings parser.add_argument('--gpu-memory-fraction', type=float, default=0.3, help='The fraction of GPU used by the script.') parser.add_argument('--gpu-visible-devices', type=str, default="0", help='Set CUDA_VISIBLE_DEVICES variable.') args = parser.parse_known_args()[0] # remove verbose bits from tf # os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' # tf.logging.set_verbosity(tf.logging.ERROR) # Set CUDA GPU environment os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_visible_devices print('Using KeyNet version:' + args.network_version) if not args.extract_MS: args.pyramid_levels = 0 args.upsampled_levels = 0 return args # convert matrix of pts into list of keypoints def convert_pts_to_keypoints(pts, scores, sizes, levels): assert(len(pts)==len(scores)) kps = [] if pts is not None: # convert matrix [Nx2] of pts into list of keypoints kps = [ cv2.KeyPoint(p[0], p[1], _size=sizes[i], _response=scores[i], _octave=levels[i]) for i,p in enumerate(pts) ] return kps # interface for pySLAM class KeyNetDescFeature2D: def __init__(self, num_features=2000, num_levels=5, # The number of downsample levels in the pyramid. scale_factor=2, # The scale factor to extract patches before descriptor. scale_factor_levels=np.sqrt(2), # The scale factor between the pyramid levels. do_cuda=True, do_tf_logging=False): print('Using KeyNetDescFeature2D') self.lock = RLock() self.model_base_path = config.cfg.root_folder + '/thirdparty/keynet/' set_tf_logging(do_tf_logging) self.do_cuda = do_cuda & torch.cuda.is_available() print('cuda:',self.do_cuda) device = torch.device("cuda:0" if self.do_cuda else "cpu") self.session = None self.keypoint_size = 8 # just a representative size for visualization and in order to convert extracted points to cv2.KeyPoint self.pts = [] self.kps = [] self.des = [] self.scales = [] self.scores = [] self.frame = None keynet_config = build_keynet_config(self.model_base_path) self.keynet_config = keynet_config keynet_config.num_points = num_features keynet_config.pyramid_levels = num_levels keynet_config.scale_factor = scale_factor keynet_config.scale_factor_levels = scale_factor_levels print_options(self.keynet_config,'KEYNET CONFIG') print('==> Loading pre-trained network.') self.load_model() print('==> Successfully loaded pre-trained network.') @property def num_features(self): return self.keynet_config.num_points @property def num_levels(self): return self.keynet_config.pyramid_levels @property def scale_factor(self): return self.keynet_config.scale_factor def __del__(self): self.close() def close(self): if self.session is not None: print('KEYNET: closing tf session') self.session.close() tf.reset_default_graph() def load_model(self): # Create graph before session :) self.graph = tf.Graph().as_default() # GPU Usage tf_config = tf.ConfigProto() tf_config.gpu_options.per_process_gpu_memory_fraction = self.keynet_config.gpu_memory_fraction tf_config.gpu_options.allow_growth = True #with tf.Session(config=config) as sess: self.session = tf.Session(config=tf_config) tf.set_random_seed(self.keynet_config.random_seed) with tf.name_scope('inputs'): # Define the input tensor shape tensor_input_shape = (None, None, None, 1) self.input_network = tf.placeholder(dtype=tf.float32, shape=tensor_input_shape, name='input_network') self.dimension_image = tf.placeholder(dtype=tf.int32, shape=(3,), name='dimension_image') self.kpts_coord = tf.placeholder(dtype=tf.float32, shape=(None, 2), name='kpts_coord') self.kpts_batch = tf.placeholder(dtype=tf.int32, shape=(None,), name='kpts_batch') self.kpts_scale = tf.placeholder(dtype=tf.float32, name='kpts_scale') self.phase_train = tf.placeholder(tf.bool, name='phase_train') with tf.name_scope('model_deep_detector'): deep_architecture = keynet(self.keynet_config) output_network = deep_architecture.model(self.input_network, self.phase_train, self.dimension_image, reuse=False) self.maps = tf.nn.relu(output_network['output']) # Extract Patches from inputs: self.input_patches = loss_desc.build_patch_extraction(self.kpts_coord, self.kpts_batch, self.input_network, kpts_scale=self.kpts_scale) # Define Pytorch HardNet self.model = HardNet() checkpoint = torch.load(self.keynet_config.pytorch_hardnet_dir) self.model.load_state_dict(checkpoint['state_dict']) if self.do_cuda: self.model.cuda() print('Extracting torch model on GPU') else: print('Extracting torch model on CPU') self.model = model.cpu() self.model.eval() # Define variables detect_var = [v for v in tf.trainable_variables(scope='model_deep_detector')] if os.listdir(self.keynet_config.checkpoint_det_dir): init_assign_op_det, init_feed_dict_det = tf_contrib.framework.assign_from_checkpoint( tf.train.latest_checkpoint(self.keynet_config.checkpoint_det_dir), detect_var) point_level = [] tmp = 0.0 factor_points = (self.keynet_config.scale_factor_levels ** 2) self.levels = self.keynet_config.pyramid_levels + self.keynet_config.upsampled_levels + 1 #print('levels: ', [i for i in range(self.levels)]) for idx_level in range(self.levels): tmp += factor_points ** (-1 * (idx_level - self.keynet_config.upsampled_levels)) point_level.append(self.keynet_config.num_points * factor_points ** (-1 * (idx_level - self.keynet_config.upsampled_levels))) self.point_level = np.asarray(list(map(lambda x: int(x / tmp), point_level))) #print('self.point_level:',self.point_level) self.session.run(tf.global_variables_initializer()) if os.listdir(self.keynet_config.checkpoint_det_dir): self.session.run(init_assign_op_det, init_feed_dict_det) def extract_keynet_features(self, image): pyramid = pyramid_gaussian(image, max_layer=self.keynet_config.pyramid_levels, downscale=self.keynet_config.scale_factor_levels) score_maps = {} for (j, resized) in enumerate(pyramid): im = resized.reshape(1, resized.shape[0], resized.shape[1], 1) feed_dict = { self.input_network: im, self.phase_train: False, self.dimension_image: np.array([1, im.shape[1], im.shape[2]], dtype=np.int32), } im_scores = self.session.run(self.maps, feed_dict=feed_dict) im_scores = geo_tools.remove_borders(im_scores, borders=self.keynet_config.border_size) score_maps['map_' + str(j + 1 + self.keynet_config.upsampled_levels)] = im_scores[0, :, :, 0] if self.keynet_config.upsampled_levels: for j in range(self.keynet_config.upsampled_levels): factor = self.keynet_config.scale_factor_levels ** (self.keynet_config.upsampled_levels - j) up_image = cv2.resize(image, (0, 0), fx=factor, fy=factor) im = np.reshape(up_image, (1, up_image.shape[0], up_image.shape[1], 1)) feed_dict = { self.input_network: im, self.phase_train: False, self.dimension_image: np.array([1, im.shape[1], im.shape[2]], dtype=np.int32), } im_scores = self.session.run(self.maps, feed_dict=feed_dict) im_scores = geo_tools.remove_borders(im_scores, borders=self.keynet_config.border_size) score_maps['map_' + str(j + 1)] = im_scores[0, :, :, 0] im_pts = [] im_pts_levels = [] for idx_level in range(self.levels): scale_value = (self.keynet_config.scale_factor_levels ** (idx_level - self.keynet_config.upsampled_levels)) scale_factor = 1. / scale_value h_scale = np.asarray([[scale_factor, 0., 0.], [0., scale_factor, 0.], [0., 0., 1.]]) h_scale_inv = np.linalg.inv(h_scale) h_scale_inv = h_scale_inv / h_scale_inv[2, 2] num_points_level = self.point_level[idx_level] #print('num_points_level:',num_points_level) if idx_level > 0: res_points = int(np.asarray([self.point_level[a] for a in range(0, idx_level + 1)]).sum() - len(im_pts)) num_points_level = res_points im_scores = rep_tools.apply_nms(score_maps['map_' + str(idx_level + 1)], self.keynet_config.nms_size) im_pts_tmp = geo_tools.get_point_coordinates(im_scores, num_points=num_points_level, order_coord='xysr') im_pts_tmp = geo_tools.apply_homography_to_points(im_pts_tmp, h_scale_inv) if not idx_level: im_pts = im_pts_tmp else: im_pts = np.concatenate((im_pts, im_pts_tmp), axis=0) im_pts_levels_tmp = np.ones(len(im_pts),dtype=np.int32)*idx_level im_pts_levels = np.concatenate((im_pts_levels, im_pts_levels_tmp), axis=0).astype(np.int32) if self.keynet_config.order_coord == 'yxsr': im_pts = np.asarray(list(map(lambda x: [x[1], x[0], x[2], x[3]], im_pts))) sorted_idxs = (-1 * im_pts[:, 3]).argsort() # sort points with their scores im_pts = im_pts[sorted_idxs] im_pts_levels = im_pts_levels[sorted_idxs] #print('im_pts_levels:',im_pts_levels) im_pts = im_pts[:self.keynet_config.num_points] im_pts_levels = im_pts_levels[:self.keynet_config.num_points] # Extract descriptor from features descriptors = [] im = image.reshape(1, image.shape[0], image.shape[1], 1) for idx_desc_batch in range(int(len(im_pts) / 250 + 1)): points_batch = im_pts[idx_desc_batch * 250: (idx_desc_batch + 1) * 250] if not len(points_batch): break feed_dict = { self.input_network: im, self.phase_train: False, self.kpts_coord: points_batch[:, :2], self.kpts_scale: self.keynet_config.scale_factor * points_batch[:, 2], self.kpts_batch: np.zeros(len(points_batch)), self.dimension_image: np.array([1, im.shape[1], im.shape[2]], dtype=np.int32), } patch_batch = self.session.run(self.input_patches, feed_dict=feed_dict) patch_batch = np.reshape(patch_batch, (patch_batch.shape[0], 1, 32, 32)) data_a = torch.from_numpy(patch_batch) data_a = data_a.cuda() data_a = Variable(data_a) with torch.no_grad(): out_a = self.model(data_a) desc_batch = out_a.data.cpu().numpy().reshape(-1, 128) if idx_desc_batch == 0: descriptors = desc_batch else: descriptors = np.concatenate([descriptors, desc_batch], axis=0) return im_pts, descriptors, im_pts_levels def compute_kps_des(self, im): with self.lock: im = im.astype(float) / im.max() im_pts, descriptors, im_pts_levels = self.extract_keynet_features(im) self.pts = im_pts[:,:2] scales = im_pts[:,2] scores = im_pts[:,3] pts_levels = im_pts_levels #print('scales:',self.scales) self.kps = convert_pts_to_keypoints(self.pts, scores, scales*self.keypoint_size, pts_levels) return self.kps, descriptors def detectAndCompute(self, frame, mask=None): #mask is a fake input with self.lock: self.frame = frame self.kps, self.des = self.compute_kps_des(frame) if kVerbose: print('detector: KEYNET, descriptor: KEYNET, #features: ', len(self.kps), ', frame res: ', frame.shape[0:2]) return self.kps, self.des # return keypoints if available otherwise call detectAndCompute() def detect(self, frame, mask=None): # mask is a fake input with self.lock: #if self.frame is not frame: self.detectAndCompute(frame) return self.kps # return descriptors if available otherwise call detectAndCompute() def compute(self, frame, kps=None, mask=None): # kps is a fake input, mask is a fake input with self.lock: if self.frame is not frame: Printer.orange('WARNING: KEYNET is recomputing both kps and des on last input frame', frame.shape) self.detectAndCompute(frame) return self.kps, self.des # return descriptors if available otherwise call detectAndCompute() def compute(self, frame, kps=None, mask=None): # kps is a fake input, mask is a fake input with self.lock: if self.frame is not frame: #Printer.orange('WARNING: KEYNET is recomputing both kps and des on last input frame', frame.shape) self.detectAndCompute(frame) return self.kps, self.des
19,688
42.463576
146
py
pyslam
pyslam-master/feature_l2net_keras.py
""" * This file is part of PYSLAM * * Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> * * PYSLAM is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * PYSLAM is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with PYSLAM. If not, see <http://www.gnu.org/licenses/>. """ import config config.cfg.set_lib('l2net_keras') import os import sys import time import cv2 import numpy as np from L2_Net import L2Net from utils_sys import Printer from utils_features import extract_patches_tensor, extract_patches_array, extract_patches_array_cpp kVerbose = True # interface for pySLAM class L2NetKerasFeature2D: def __init__(self, do_tf_logging=False): print('Using L2NetKerasFeature2D') # One of "L2Net-HP", "L2Net-HP+", "L2Net-LIB", "L2Net-LIB+", "L2Net-ND", "L2Net-ND+", "L2Net-YOS", "L2Net-YOS+", self.net_name = 'L2Net-HP+' # mag_factor is how many times the original keypoint scale # is enlarged to generate a patch from a keypoint self.mag_factor = 3 # inference batch size self.batch_size = 512 self.process_all = True # process all the patches at once print('==> Loading pre-trained network.') self.l2net = L2Net(self.net_name, do_tf_logging=do_tf_logging) print('==> Successfully loaded pre-trained network.') def compute(self, frame, kps, mask=None): #mask is a fake input #print('kps: ', kps) if len(kps)>0: if False: # use python code patches = extract_patches_array(frame, kps, patch_size=32, mag_factor=self.mag_factor) else: # use faster cpp code patches = extract_patches_array_cpp(frame, kps, patch_size=32, mag_factor=self.mag_factor) patches = np.asarray(patches) patches = np.expand_dims(patches, -1) self.des = self.l2net.calc_descriptors(patches) else: self.des = [] if kVerbose: print('descriptor: L2NET, #features: ', len(kps), ', frame res: ', frame.shape[0:2]) return kps, self.des
2,773
32.829268
123
py
pyslam
pyslam-master/feature_sosnet.py
""" * This file is part of PYSLAM * * Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> * * PYSLAM is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * PYSLAM is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with PYSLAM. If not, see <http://www.gnu.org/licenses/>. """ import config config.cfg.set_lib('sosnet') import os import numpy as np import math import cv2 import torch import sosnet_model from utils_features import extract_patches_tensor, extract_patches_array, extract_patches_array_cpp kVerbose = True # interface for pySLAM class SosnetFeature2D: def __init__(self, do_cuda=True): print('Using SosnetFeature2D') self.model_base_path = config.cfg.root_folder + '/thirdparty/SOSNet/' self.do_cuda = do_cuda & torch.cuda.is_available() print('cuda:',self.do_cuda) device = torch.device("cuda:0" if self.do_cuda else "cpu") torch.set_grad_enabled(False) # mag_factor is how many times the original keypoint scale # is enlarged to generate a patch from a keypoint self.mag_factor = 3 print('==> Loading pre-trained network.') #init tfeat and load the trained weights self.model = sosnet_model.SOSNet32x32() self.net_name = 'liberty' # liberty, hpatches_a, notredame, yosemite (see folder /thirdparty/SOSNet/sosnet-weights) self.model.load_state_dict(torch.load(os.path.join(self.model_base_path, 'sosnet-weights', "sosnet-32x32-" + self.net_name + ".pth"))) if self.do_cuda: self.model.cuda() print('Extracting on GPU') else: print('Extracting on CPU') self.model = model.cpu() self.model.eval() print('==> Successfully loaded pre-trained network.') def compute_des(self, patches): patches = torch.from_numpy(patches).float() patches = torch.unsqueeze(patches,1) if self.do_cuda: patches = patches.cuda() with torch.no_grad(): descrs = self.model(patches) return descrs.detach().cpu().numpy().reshape(-1, 128) def compute(self, frame, kps, mask=None): #mask is a fake input #print('kps: ', kps) if len(kps)>0: #des = tfeat_utils.describe_opencv(self.model, frame, kps, 32, self.mag_factor) # extract the keypoint patches #t = time.time() if False: # use python code patches = extract_patches_array(frame, kps, patch_size=32, mag_factor=self.mag_factor) else: # use faster cpp code patches = extract_patches_array_cpp(frame, kps, patch_size=32, mag_factor=self.mag_factor) patches = np.asarray(patches) #if kVerbose: # print('patches.shape:',patches.shape) #if kVerbose: # print('patch elapsed: ', time.time()-t) # compute descriptor by feeeding the full patch tensor to the network des = self.compute_des(patches) else: des = [] if kVerbose: print('descriptor: SOSNET, #features: ', len(kps), ', frame res: ', frame.shape[0:2]) return kps, des
3,919
36.692308
142
py
pyslam
pyslam-master/feature_l2net.py
""" * This file is part of PYSLAM * Adapted from https://github.com/vcg-uvic/image-matching-benchmark-baselines/blob/master/third_party/l2net_config/l2net_model.py, see licence therein. * * Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> * * PYSLAM is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * PYSLAM is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with PYSLAM. If not, see <http://www.gnu.org/licenses/>. """ # adapted from https://github.com/vcg-uvic/image-matching-benchmark-baselines/blob/master/third_party/l2net_config/l2net_model.py import config import sys import torch import torch.nn as nn import torch.nn.functional as F from torch.autograd import Variable import torch.backends.cudnn as cudnn import time import os import cv2 import math import numpy as np from utils_features import extract_patches_tensor, extract_patches_array, extract_patches_array_cpp kVerbose = True class L2Norm(nn.Module): def __init__(self): super(L2Norm,self).__init__() self.eps = 1e-10 def forward(self, x): norm = torch.sqrt(torch.sum(x * x, dim = 1) + self.eps) x= x / norm.unsqueeze(-1).expand_as(x) return x class L1Norm(nn.Module): def __init__(self): super(L1Norm,self).__init__() self.eps = 1e-10 def forward(self, x): norm = torch.sum(torch.abs(x), dim = 1) + self.eps x= x / norm.expand_as(x) return x class L2Net(nn.Module): def __init__(self): super(L2Net, self).__init__() self.eps = 1e-10 self.features = nn.Sequential( nn.Conv2d(1, 32, kernel_size=3, padding=1, bias=True), nn.BatchNorm2d(32, affine=True, eps=self.eps), nn.ReLU(), nn.Conv2d(32, 32, kernel_size=3, padding=1, bias=True), nn.BatchNorm2d(32, affine=True, eps=self.eps), nn.ReLU(), nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1, bias=True), nn.BatchNorm2d(64, affine=True, eps=self.eps), nn.ReLU(), nn.Conv2d(64, 64, kernel_size=3, padding=1, bias=True), nn.BatchNorm2d(64, affine=True, eps=self.eps), nn.ReLU(), nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1, bias=True), nn.BatchNorm2d(128, affine=True, eps=self.eps), nn.ReLU(), nn.Conv2d(128, 128, kernel_size=3, padding=1, bias=True), nn.BatchNorm2d(128, affine=True, eps=self.eps), nn.ReLU(), nn.Conv2d(128, 128, kernel_size=8, bias=True), nn.BatchNorm2d(128, affine=True, eps=self.eps), ) return def input_norm(self, x): # matlab norm z = x.contiguous().transpose(2, 3).contiguous().view(x.size(0),-1) x_minus_mean = z.transpose(0,1)-z.mean(1) sp = torch.std(z,1).detach() norm_inp = x_minus_mean/(sp+1e-12) norm_inp = norm_inp.transpose(0, 1).view(-1, 1, x.size(2), x.size(3)).transpose(2,3) return norm_inp def forward(self, input): norm_img = self.input_norm(input) x_features = self.features(norm_img) return nn.LocalResponseNorm(256,1*256,0.5,0.5)(x_features).view(input.size(0),-1) # interface for pySLAM class L2NetFeature2D: def __init__(self, do_cuda=True): print('Using L2NetFeature2D') self.model_base_path = config.cfg.root_folder + '/thirdparty/l2net/' self.model_weights_path = self.model_base_path + 'l2net_ported_weights_lib+.pth' #print('model_weights_path:',self.model_weights_path) # get pre-trained image mean # l2net_weights = sio.loadmat(args.matlab_weights_path) # imgMean = l2net_weights['pixMean'] self.do_cuda = do_cuda & torch.cuda.is_available() print('cuda:',self.do_cuda) device = torch.device("cuda:0" if self.do_cuda else "cpu") torch.set_grad_enabled(False) # mag_factor is how many times the original keypoint scale # is enlarged to generate a patch from a keypoint self.mag_factor = 1.0 # inference batch size self.batch_size = 512 self.process_all = True # process all the patches at once print('==> Loading pre-trained network.') self.model = L2Net() self.checkpoint = torch.load(self.model_weights_path) #self.model.load_state_dict(self.checkpoint['state_dict']) self.model.load_state_dict(self.checkpoint) if self.do_cuda: self.model.cuda() print('Extracting on GPU') else: print('Extracting on CPU') self.model = model.cpu() self.model.eval() print('==> Successfully loaded pre-trained network.') def compute_des_batches(self, patches): n_batches = int(len(patches) / self.batch_size) + 1 descriptors_for_net = np.zeros((len(patches), 128)) for i in range(0, len(patches), self.batch_size): data_a = patches[i: i + self.batch_size, :, :, :].astype(np.float32) data_a = torch.from_numpy(data_a) if self.do_cuda: data_a = data_a.cuda() data_a = Variable(data_a) # compute output with torch.no_grad(): out_a = self.model(data_a) descriptors_for_net[i: i + self.batch_size,:] = out_a.data.cpu().numpy().reshape(-1, 128) return descriptors_for_net def compute_des(self, patches): patches = torch.from_numpy(patches).float() patches = torch.unsqueeze(patches,1) if self.do_cuda: patches = patches.cuda() with torch.no_grad(): descrs = self.model(patches) return descrs.detach().cpu().numpy().reshape(-1, 128) def compute(self, img, kps, mask=None): #mask is a fake input num_kps = len(kps) des = [] if num_kps>0: if not self.process_all: # compute descriptor for each patch patches = extract_patches_tensor(img, kps, patch_size=32, mag_factor=self.mag_factor) des = self.compute_des_batches(patches).astype(np.float32) else: # compute descriptor by feeeding the full patch tensor to the network t = time.time() if False: # use python code patches = extract_patches_array(img, kps, patch_size=32, mag_factor=self.mag_factor) else: # use faster cpp code patches = extract_patches_array_cpp(img, kps, patch_size=32, mag_factor=self.mag_factor) patches = np.asarray(patches) if kVerbose: print('patches.shape:',patches.shape) if kVerbose: print('patch elapsed: ', time.time()-t) des = self.compute_des(patches) if kVerbose: print('descriptor: L2NET, #features: ', len(kps), ', frame res: ', img.shape[0:2]) return kps, des
7,936
39.28934
151
py
pyslam
pyslam-master/feature_logpolar.py
""" * This file is part of PYSLAM * adapted from https://github.com/cvlab-epfl/log-polar-descriptors/blob/aed70f882cddcfe0c27b65768b9248bf1f2c65cb/example.py, see licence therein. * * Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> * * PYSLAM is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * PYSLAM is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with PYSLAM. If not, see <http://www.gnu.org/licenses/>. """ # adapted from https://github.com/cvlab-epfl/log-polar-descriptors/blob/aed70f882cddcfe0c27b65768b9248bf1f2c65cb/example.py import config config.cfg.set_lib('logpolar') import os import sys import torch import torch.nn as nn from modules.ptn.pytorch.models import Transformer import cv2 import numpy as np import h5py from time import time from configs.defaults import _C as cfg #from modules.hardnet.models import HardNet # given some matplotlib backend changes the code is repeated below from utils_features import extract_patches_tensor, extract_patches_array, extract_patches_array_cpp kVerbose = True kVerbose2 = True class L2Norm(nn.Module): def __init__(self): super(L2Norm, self).__init__() self.eps = 1e-10 def forward(self, x): norm = torch.sqrt(torch.sum(x * x, dim=1) + self.eps) x = x / norm.unsqueeze(-1).expand_as(x) return x # from modules.hardnet.models class HardNet(nn.Module): def __init__(self, transform, coords, patch_size, scale, is_desc256, orientCorrect=True, hard_augm=False): # <-- added to take care of the possible nonlocal option managed in modules.hardnet.models super(HardNet, self).__init__() self.transform = transform self.transform_layer = Transformer(transform=transform, coords=coords, resolution=patch_size, SIFTscale=scale) self.orientCorrect = orientCorrect self.hard_augm = hard_augm # model processing patches of size [32 x 32] and giving description vectors of length 2**7 self.features = nn.Sequential( nn.Conv2d(1, 32, kernel_size=3, padding=1, bias=False), nn.BatchNorm2d(32, affine=False), nn.ReLU(), nn.Conv2d(32, 32, kernel_size=3, padding=1, bias=False), nn.BatchNorm2d(32, affine=False), nn.ReLU(), nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1, bias=False), nn.BatchNorm2d(64, affine=False), nn.ReLU(), nn.Conv2d(64, 64, kernel_size=3, padding=1, bias=False), nn.BatchNorm2d(64, affine=False), nn.ReLU(), nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1, bias=False), nn.BatchNorm2d(128, affine=False), nn.ReLU(), nn.Conv2d(128, 128, kernel_size=3, padding=1, bias=False), nn.BatchNorm2d(128, affine=False), nn.ReLU(), nn.Dropout(0.1), nn.Conv2d(128, 128, kernel_size=8, bias=False), nn.BatchNorm2d(128, affine=False), ) # initialize weights self.features.apply(weights_init) return def input_norm(self, x): flat = x.view(x.size(0), -1) mp = torch.mean(flat, dim=1) sp = torch.std(flat, dim=1) + 1e-7 return (x - mp.detach().unsqueeze(-1).unsqueeze(-1).unsqueeze(-1).expand_as(x)) / \ sp.detach().unsqueeze(-1).unsqueeze(-1).unsqueeze(1).expand_as(x) # function to forward-propagate inputs through the network def forward(self, img, theta=None, imgIDs=None): if theta is None: # suppose patches are directly given (as e.g. for external test data) patches = img else: # extract keypoints from the whole image patches = self.transform_layer([img, theta, imgIDs]) batchSize = patches.shape[0] if self.hard_augm: # args.hard_augm: bernoulli = torch.distributions.Bernoulli(torch.tensor([0.5])) if self.transform == "STN": # transpose to switch dimensions (only if STN) transpose = bernoulli.sample(torch.Size([batchSize])) patches = torch.cat([ torch.transpose(patch, 1, 2) if transpose[pdx] else patch for pdx, patch in enumerate(patches) ]).unsqueeze(1) # flip the patches' first dimension mirrorDim1 = bernoulli.sample(torch.Size([batchSize])) patches = torch.cat([ torch.flip(patch, [1]) if mirrorDim1[pdx] else patch for pdx, patch in enumerate(patches) ]).unsqueeze(1) x_features = self.features(self.input_norm(patches)) x = x_features.view(x_features.size(0), -1) return L2Norm()(x), patches def weights_init(m): ''' Conv2d module weight initialization method ''' if isinstance(m, nn.Conv2d): nn.init.orthogonal(m.weight.data, gain=0.6) try: nn.init.constant(m.bias.data, 0.01) except: pass return # interface for pySLAM class LogpolarFeature2D: def __init__(self, use_log_polar=True, do_cuda=True): print('Using LogpolarFeature2D') self.model_base_path = config.cfg.root_folder + '/thirdparty/logpolar/' if use_log_polar: config_path = os.path.join(self.model_base_path, 'configs', 'init_one_example_ptn_96.yml') if kVerbose: print('-- Using log-polar model') else: config_path = os.path.join(self.model_base_path, 'configs', 'init_one_example_stn_16.yml') if kVerbose: print('-- Using cartesian model') cfg.merge_from_file(config_path) self.model_weights_path = self.model_base_path + cfg.TEST.MODEL_WEIGHTS # N.B.: this must stay here, after cfg.merge_from_file() if kVerbose2: print('model_weights_path:',self.model_weights_path) os.environ["CUDA_VISIBLE_DEVICES"] = str(0) torch.cuda.manual_seed_all(cfg.TRAINING.SEED) torch.backends.cudnn.deterministic = True self.do_cuda = do_cuda & torch.cuda.is_available() print('cuda:',self.do_cuda) device = torch.device("cuda:0" if self.do_cuda else "cpu") self.device = device torch.set_grad_enabled(False) print('==> Loading pre-trained network.') self.model = HardNet(transform=cfg.TEST.TRANSFORMER, coords=cfg.TEST.COORDS, patch_size=cfg.TEST.IMAGE_SIZE, scale=cfg.TEST.SCALE, is_desc256=cfg.TEST.IS_DESC_256, orientCorrect=cfg.TEST.ORIENT_CORRECTION) self.checkpoint = torch.load(self.model_weights_path) self.model.load_state_dict(self.checkpoint['state_dict']) if self.do_cuda: self.model.cuda() print('Extracting on GPU') else: print('Extracting on CPU') self.model = model.cpu() self.model.eval() print('==> Successfully loaded pre-trained network.') def compute_des(self, img, kps): h, w = img.shape t = time() pts = np.array([kp.pt for kp in kps]) scales = np.array([kp.size for kp in kps]) oris = np.array([kp.angle for kp in kps]) # Mirror-pad the image to avoid boundary effects if any([s > cfg.TEST.PAD_TO for s in img.shape[:2]]): raise RuntimeError( "Image exceeds acceptable size ({}x{}), please downsample".format(cfg.TEST.PAD_TO, cfg.TEST.PAD_TO)) fillHeight = cfg.TEST.PAD_TO - img.shape[0] fillWidth = cfg.TEST.PAD_TO - img.shape[1] padLeft = int(np.round(fillWidth / 2)) padRight = int(fillWidth - padLeft) padUp = int(np.round(fillHeight / 2)) padDown = int(fillHeight - padUp) img = np.pad(img, pad_width=((padUp, padDown), (padLeft, padRight)), mode='reflect') # Normalize keypoint locations kp_norm = [] for i, p in enumerate(pts): _p = 2 * np.array([(p[0] + padLeft) / (cfg.TEST.PAD_TO), (p[1] + padUp) / (cfg.TEST.PAD_TO)]) - 1 kp_norm.append(_p) theta = [ torch.from_numpy(np.array(kp_norm)).float().squeeze(), torch.from_numpy(scales).float(), torch.from_numpy(np.array([np.deg2rad(o) for o in oris])).float() ] if kVerbose2: print('-- Padded image from {}x{} to {}x{} in {} s'.format( h, w, img.shape[0], img.shape[1], time()-t)) # Extract descriptors t = time() device = self.device imgs = torch.from_numpy(img).unsqueeze(0).to(device) img_keypoints = [theta[0].to(device), theta[1].to(device), theta[2].to(device)] descriptors, patches = self.model({'img': imgs}, img_keypoints, ['img'] * len(img_keypoints[0])) if kVerbose2: print('-- Computed {} descriptors in {:0.2f} sec.'.format( descriptors.shape[0], time() - t)) return descriptors.cpu().detach().numpy() def compute(self, img, kps, mask=None): #mask is a fake input num_kps = len(kps) des = [] if num_kps>0: des = self.compute_des(img, kps) if kVerbose: print('descriptor: LOGPOLAR, #features: ', len(kps), ', frame res: ', img.shape[0:2]) return kps, des
10,571
36.094737
147
py
pyslam
pyslam-master/feature_geodesc.py
""" * This file is part of PYSLAM * * Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> * * PYSLAM is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * PYSLAM is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with PYSLAM. If not, see <http://www.gnu.org/licenses/>. """ import config config.cfg.set_lib('geodesc',prepend=True) import os import sys import time from threading import Thread from multiprocessing import Queue import cv2 import numpy as np import warnings # to disable tensorflow-numpy warnings: from https://github.com/tensorflow/tensorflow/issues/30427 warnings.filterwarnings('ignore', category=FutureWarning) if False: import tensorflow as tf else: # from https://stackoverflow.com/questions/56820327/the-name-tf-session-is-deprecated-please-use-tf-compat-v1-session-instead import tensorflow.compat.v1 as tf # from https://kobkrit.com/using-allow-growth-memory-option-in-tensorflow-and-keras-dc8c8081bc96 to cope with the following error: # "[...tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR" tf_config = tf.ConfigProto() tf_config.gpu_options.allow_growth = True # then you must use the config as follows: # session = tf.Session(config=tf_config, ...) from utils_tf import load_frozen_model, set_tf_logging from utils_features import extract_patches_tensor, extract_patches_array, extract_patches_array_cpp kVerbose = True # interface for pySLAM class GeodescFeature2D: quantize = False # quantize or not output features: if you set this to True then you have a binary descriptor def __init__(self, do_tf_logging=False): print('Using GeodescFeature2D') # mag_factor is how many times the original keypoint scale # is enlarged to generate a patch from a keypoint self.mag_factor = 3 # inference batch size self.batch_size = 512 self.process_all = True # process all the patches at once self.model_base_path = config.cfg.root_folder + '/thirdparty/geodesc/' self.model_path = self.model_base_path + 'model/geodesc.pb' set_tf_logging(do_tf_logging) print('==> Loading pre-trained network.') # create deep feature extractor. self.graph = load_frozen_model(self.model_path, print_nodes=False) #sess = tf.Session(graph=graph) print('==> Successfully loaded pre-trained network.') def process_patches(self, patches): num_patch = patches.shape[0] if num_patch % self.batch_size > 0: loop_num = int(np.floor(float(num_patch) / float(self.batch_size))) else: loop_num = int(num_patch / self.batch_size - 1) with tf.Session(graph=self.graph, config=tf_config) as sess: def _worker(patch_queue, sess, des): """The worker thread.""" while True: patch_data = patch_queue.get() if patch_data is None: return feat = sess.run("squeeze_1:0", feed_dict={"input:0": np.expand_dims(patch_data, -1)}) des.append(feat) des = [] patch_queue = Queue() worker_thread = Thread(target=_worker, args=(patch_queue, sess, des)) worker_thread.daemon = True worker_thread.start() start = time.time() # enqueue if not self.process_all: for i in range(loop_num + 1): if i < loop_num: patch_queue.put(patches[i * self.batch_size: (i + 1) * self.batch_size]) else: patch_queue.put(patches[i * self.batch_size:]) else: patch_queue.put(patches) # poison pill patch_queue.put(None) # wait for extraction. worker_thread.join() end = time.time() if kVerbose: print('Time cost in feature extraction', end - start) des = np.concatenate(des, axis=0) # quantize output features des = (des * 128 + 128).astype(np.uint8) if self.quantize else des return des def compute(self, frame, kps, mask=None): #mask is a fake input #print('kps: ', kps) if len(kps)>0: if False: # use python code patches = extract_patches_array(frame, kps, patch_size=32, mag_factor=self.mag_factor) else: # use faster cpp code patches = extract_patches_array_cpp(frame, kps, patch_size=32, mag_factor=self.mag_factor) patches = np.asarray(patches) des = self.process_patches(patches) else: des = [] if kVerbose: print('descriptor: GEODESC, #features: ', len(kps), ', frame res: ', frame.shape[0:2]) return kps, des
5,637
36.337748
130
py
pyslam
pyslam-master/feature_r2d2.py
""" * This file is part of PYSLAM. * Adapted from https://raw.githubusercontent.com/naver/r2d2/master/extract.py, see the licence therein. * * Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> * * PYSLAM is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * PYSLAM is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with PYSLAM. If not, see <http://www.gnu.org/licenses/>. """ # adapted from from https://raw.githubusercontent.com/naver/r2d2/master/extract.py import config config.cfg.set_lib('r2d2') import os, pdb from PIL import Image import numpy as np import torch import cv2 from threading import RLock from r2d2.tools import common from r2d2.tools.dataloader import norm_RGB from r2d2.nets.patchnet import * import argparse from utils_sys import Printer kVerbose = True def load_network(model_fn): checkpoint = torch.load(model_fn) print("\n>> Creating net = " + checkpoint['net']) net = eval(checkpoint['net']) nb_of_weights = common.model_size(net) print(f" ( Model size: {nb_of_weights/1000:.0f}K parameters )") # initialization weights = checkpoint['state_dict'] net.load_state_dict({k.replace('module.',''):v for k,v in weights.items()}) return net.eval() class NonMaxSuppression (torch.nn.Module): def __init__(self, rel_thr=0.7, rep_thr=0.7): nn.Module.__init__(self) self.max_filter = torch.nn.MaxPool2d(kernel_size=3, stride=1, padding=1) self.rel_thr = rel_thr self.rep_thr = rep_thr def forward(self, reliability, repeatability, **kw): assert len(reliability) == len(repeatability) == 1 reliability, repeatability = reliability[0], repeatability[0] # local maxima maxima = (repeatability == self.max_filter(repeatability)) # remove low peaks maxima *= (repeatability >= self.rep_thr) maxima *= (reliability >= self.rel_thr) return maxima.nonzero().t()[2:4] def extract_multiscale( net, img, detector, scale_f=2**0.25, min_scale=0.0, max_scale=1, min_size=256, max_size=1024, verbose=False): old_bm = torch.backends.cudnn.benchmark torch.backends.cudnn.benchmark = False # speedup # extract keypoints at multiple scales B, three, H, W = img.shape assert B == 1 and three == 3, "should be a batch with a single RGB image" assert max_scale <= 1 s = 1.0 # current scale factor level = 0 L = [] X,Y,S,C,Q,D = [],[],[],[],[],[] while s+0.001 >= max(min_scale, min_size / max(H,W)): if s-0.001 <= min(max_scale, max_size / max(H,W)): nh, nw = img.shape[2:] if verbose: print(f"extracting at scale x{s:.02f} = {nw:4d}x{nh:3d} - level {level}") # extract descriptors with torch.no_grad(): res = net(imgs=[img]) # get output and reliability map descriptors = res['descriptors'][0] reliability = res['reliability'][0] repeatability = res['repeatability'][0] # normalize the reliability for nms # extract maxima and descs y,x = detector(**res) # nms c = reliability[0,0,y,x] q = repeatability[0,0,y,x] d = descriptors[0,:,y,x].t() n = d.shape[0] # accumulate multiple scales X.append(x.float() * W/nw) Y.append(y.float() * H/nh) S.append((32/s) * torch.ones(n, dtype=torch.float32, device=d.device)) C.append(c) Q.append(q) D.append(d) L_tmp =level * np.ones(n,dtype=np.int32) L = np.concatenate((L, L_tmp), axis=0).astype(np.int32) level += 1 s /= scale_f # down-scale the image for next iteration nh, nw = round(H*s), round(W*s) img = F.interpolate(img, (nh,nw), mode='bilinear', align_corners=False) # restore value torch.backends.cudnn.benchmark = old_bm Y = torch.cat(Y) X = torch.cat(X) S = torch.cat(S) # scale scores = torch.cat(C) * torch.cat(Q) # scores = reliability * repeatability XYS = torch.stack([X,Y,S], dim=-1) D = torch.cat(D) return XYS, D, scores, L # convert matrix of pts into list of keypoints def convert_pts_to_keypoints(pts, scores, sizes, levels): assert(len(pts)==len(scores)) kps = [] if pts is not None: # convert matrix [Nx2] of pts into list of keypoints kps = [ cv2.KeyPoint(p[0], p[1], _size=sizes[i], _response=scores[i], _octave=levels[i]) for i,p in enumerate(pts) ] return kps # TODO: fix the octave field of the output keypoints # interface for pySLAM class R2d2Feature2D: def __init__(self, num_features = 2000, scale_f = 2**0.25, min_size = 256, max_size = 1300, #1024, min_scale = 0, max_scale = 1, reliability_thr = 0.7, repeatability_thr = 0.7, do_cuda=True): print('Using R2d2Feature2D') self.lock = RLock() self.model_base_path = config.cfg.root_folder + '/thirdparty/r2d2' self.model_weights_path = self.model_base_path + '/models/r2d2_WASF_N16.pt' #print('model_weights_path:',self.model_weights_path) self.pts = [] self.kps = [] self.des = [] self.frame = None self.num_features = num_features self.scale_f = scale_f self.min_size = min_size self.max_size = max_size self.min_scale = min_scale self.max_scale = max_scale self.reliability_thr = reliability_thr self.repeatability_thr = repeatability_thr self.do_cuda = do_cuda if do_cuda: gpus = [0] else: gpus = -1 self.gpus = gpus self.do_cuda = common.torch_set_gpu(gpus) print('==> Loading pre-trained network.') self.net = load_network(self.model_weights_path) if self.do_cuda: self.net = self.net.cuda() # create the non-maxima detector self.detector = NonMaxSuppression(rel_thr=reliability_thr, rep_thr=repeatability_thr) print('==> Successfully loaded pre-trained network.') def compute_kps_des(self,img): with self.lock: H, W = img.shape[:2] img = norm_RGB(img)[None] if self.do_cuda: img = img.cuda() # extract keypoints/descriptors for a single image xys, desc, scores, levels = extract_multiscale(self.net, img, self.detector, scale_f = self.scale_f, min_scale = self.min_scale, max_scale = self.max_scale, min_size = self.min_size, max_size = self.max_size, verbose = kVerbose) xys = xys.cpu().numpy() desc = desc.cpu().numpy() scores = scores.cpu().numpy() idxs = scores.argsort()[-self.num_features or None:] selected_xys = xys[idxs] self.pts = selected_xys[:,:2] sizes = selected_xys[:,2] des = desc[idxs] scores = scores[idxs] levels = np.array(levels)[idxs] kps = convert_pts_to_keypoints(self.pts, scores, sizes, levels) return kps, des def detectAndCompute(self, frame, mask=None): #mask is a fake input with self.lock: self.frame = frame self.kps, self.des = self.compute_kps_des(frame) if kVerbose: print('detector: R2D2 , descriptor: R2D2 , #features: ', len(self.kps), ', frame res: ', frame.shape[0:2]) return self.kps, self.des # return keypoints if available otherwise call detectAndCompute() def detect(self, frame, mask=None): # mask is a fake input with self.lock: if self.frame is not frame: self.detectAndCompute(frame) return self.kps # return descriptors if available otherwise call detectAndCompute() def compute(self, frame, kps=None, mask=None): # kps is a fake input, mask is a fake input with self.lock: if self.frame is not frame: Printer.orange('WARNING: R2D2 is recomputing both kps and des on last input frame', frame.shape) self.detectAndCompute(frame) return self.kps, self.des
9,607
35.393939
146
py
pyslam
pyslam-master/feature_tfeat.py
""" * This file is part of PYSLAM * * Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> * * PYSLAM is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * PYSLAM is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with PYSLAM. If not, see <http://www.gnu.org/licenses/>. """ import config config.cfg.set_lib('tfeat') import os import numpy as np import math import cv2 import time import torchvision as tv import torch import tfeat_model import tfeat_utils import torch.nn as nn import torch.optim as optim import torch.nn.functional as F import torch.backends.cudnn as cudnn from utils_features import extract_patches_array, extract_patches_array_cpp kVerbose = True # interface for pySLAM class TfeatFeature2D: def __init__(self, do_cuda=True): print('Using TfeatFeature2D') self.model_base_path = config.cfg.root_folder + '/thirdparty/tfeat/' self.do_cuda = do_cuda & torch.cuda.is_available() print('cuda:',self.do_cuda) device = torch.device("cuda:0" if self.do_cuda else "cpu") torch.set_grad_enabled(False) # mag_factor is how many times the original keypoint scale # is enlarged to generate a patch from a keypoint self.mag_factor = 3 print('==> Loading pre-trained network.') #init tfeat and load the trained weights self.model = tfeat_model.TNet() self.models_path = self.model_base_path + 'pretrained-models' self.net_name = 'tfeat-liberty' self.model.load_state_dict(torch.load(os.path.join(self.models_path,self.net_name+".params"))) if self.do_cuda: self.model.cuda() print('Extracting on GPU') else: print('Extracting on CPU') self.model = model.cpu() self.model.eval() print('==> Successfully loaded pre-trained network.') def compute_des(self, patches): patches = torch.from_numpy(patches).float() patches = torch.unsqueeze(patches,1) if self.do_cuda: patches = patches.cuda() with torch.no_grad(): descrs = self.model(patches) return descrs.detach().cpu().numpy().reshape(-1, 128) def compute(self, frame, kps, mask=None): #mask is a fake input #print('kps: ', kps) if len(kps)>0: #des = tfeat_utils.describe_opencv(self.model, frame, kps, 32, self.mag_factor) # extract the keypoint patches #t = time.time() if False: # use python code patches = extract_patches_array(frame, kps, patch_size=32, mag_factor=self.mag_factor) else: # use faster cpp code patches = extract_patches_array_cpp(frame, kps, patch_size=32, mag_factor=self.mag_factor) patches = np.asarray(patches) #if kVerbose: # print('patches.shape:',patches.shape) #if kVerbose: # print('patch elapsed: ', time.time()-t) # compute descriptor by feeeding the full patch tensor to the network des = self.compute_des(patches) else: des = [] if kVerbose: print('descriptor: TFEAT, #features: ', len(kps), ', frame res: ', frame.shape[0:2]) return kps, des
3,989
34.625
114
py
pyslam
pyslam-master/feature_disk.py
""" * This file is part of PYSLAM * Adapted from https://github.com/cvlab-epfl/disk/blob/master/detect.py, see licence therein. * * Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> * * PYSLAM is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * PYSLAM is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with PYSLAM. If not, see <http://www.gnu.org/licenses/>. """ # adapted from https://github.com/cvlab-epfl/disk/blob/master/detect.py import sys import config config.cfg.set_lib('disk') config.cfg.set_lib('torch-dimcheck') config.cfg.set_lib('torch-localize') config.cfg.set_lib('unets') import cv2 from threading import RLock from utils_sys import Printer import torch, h5py, imageio, os, argparse import numpy as np import torch.nn.functional as F from functools import partial from torch.utils.data import DataLoader from tqdm import tqdm from torch_dimcheck import dimchecked from disk import DISK, Features from utils_sys import Printer, is_opencv_version_greater_equal kVerbose = True class Image: def __init__(self, bitmap: ['C', 'H', 'W'], fname: str, orig_shape=None): self.bitmap = bitmap self.fname = fname if orig_shape is None: self.orig_shape = self.bitmap.shape[1:] else: self.orig_shape = orig_shape def resize_to(self, shape): return Image( self._pad(self._interpolate(self.bitmap, shape), shape), self.fname, orig_shape=self.bitmap.shape[1:], ) @dimchecked def to_image_coord(self, xys: [2, 'N']) -> ([2, 'N'], ['N']): f, _size = self._compute_interpolation_size(self.bitmap.shape[1:]) scaled = xys / f h, w = self.orig_shape x, y = scaled mask = (0 <= x) & (x < w) & (0 <= y) & (y < h) return scaled, mask def _compute_interpolation_size(self, shape): x_factor = self.orig_shape[0] / shape[0] y_factor = self.orig_shape[1] / shape[1] f = 1 / max(x_factor, y_factor) if x_factor > y_factor: new_size = (shape[0], int(f * self.orig_shape[1])) else: new_size = (int(f * self.orig_shape[0]), shape[1]) return f, new_size @dimchecked def _interpolate(self, image: ['C', 'H', 'W'], shape) -> ['C', 'h', 'w']: _f, size = self._compute_interpolation_size(shape) return F.interpolate( image.unsqueeze(0), size=size, mode='bilinear', align_corners=False, ).squeeze(0) @dimchecked def _pad(self, image: ['C', 'H', 'W'], shape) -> ['C', 'h', 'w']: x_pad = shape[0] - image.shape[1] y_pad = shape[1] - image.shape[2] if x_pad < 0 or y_pad < 0: raise ValueError("Attempting to pad by negative value") return F.pad(image, (0, y_pad, 0, x_pad)) class ImageAdapter: def __init__(self, image, crop_size=(None, None)): self.image = image self.crop_size = crop_size def get(self): # name = self.names[ix] # path = os.path.join(self.image_path, name) # img = np.ascontiguousarray(imageio.imread(path)) # tensor = torch.from_numpy(img).to(torch.float32) img = np.ascontiguousarray(self.image) tensor = torch.from_numpy(img).to(torch.float32) if len(tensor.shape) == 2: # some images may be grayscale tensor = tensor.unsqueeze(-1).expand(-1, -1, 3) bitmap = tensor.permute(2, 0, 1) / 255. #extensionless_fname = os.path.splitext(name)[0] image = Image(bitmap, '') if self.crop_size != (None, None): image = image.resize_to(self.crop_size) return image def stack(self): images = [self.get()] bitmaps = torch.stack([im.bitmap for im in images], dim=0) return bitmaps, images # convert matrix of pts into list of keypoints def convert_pts_to_keypoints(pts, scores, size): assert(len(pts)==len(scores)) kps = [] if pts is not None: # convert matrix [Nx2] of pts into list of keypoints if is_opencv_version_greater_equal(4,5,3): kps = [ cv2.KeyPoint(p[0], p[1], size=size, response=s, octave=0) for p,s in zip(pts,scores) ] else: kps = [ cv2.KeyPoint(p[0], p[1], _size=size, _response=s, _octave=0) for p,s in zip(pts,scores) ] return kps # convert matrix of pts into list of keypoints def convert_pts_to_keypoints_with_translation(pts, scores, size, deltax, deltay): assert(len(pts)==len(scores)) kps = [] if pts is not None: # convert matrix [Nx2] of pts into list of keypoints if is_opencv_version_greater_equal(4,5,3): kps = [ cv2.KeyPoint(p[0]+deltax, p[1]+deltay, size=size, response=s, octave=0) for p,s in zip(pts,scores) ] else: kps = [ cv2.KeyPoint(p[0]+deltax, p[1]+deltay, _size=size, _response=s, _octave=0) for p,s in zip(pts,scores) ] return kps # interface for pySLAM # NOTE: from Fig. 3 in the paper "DISK: Learning local features with policy gradient" # "Our approach can match many more points and produce more accurate poses. It can deal with large changes in scale (4th and 5th columns) but not in rotation..." class DiskFeature2D: def __init__(self, num_features=2000, nms_window_size=5, # NMS windows size desc_dim=128, # descriptor dimension. Needs to match the checkpoint value mode = 'nms', # choices=['nms', 'rng'], Whether to extract features using the non-maxima suppresion mode or through training-time grid sampling technique' do_cuda=True): print('Using DiskFeature2D') self.lock = RLock() self.num_features = num_features self.nms_window_size = nms_window_size self.desc_dim = desc_dim self.mode = mode self.model_base_path = config.cfg.root_folder + '/thirdparty/disk/depth-save.pth' self.do_cuda = do_cuda & torch.cuda.is_available() print('cuda:',self.do_cuda) self.DEV = torch.device('cuda' if self.do_cuda else 'cpu') self.CPU = torch.device('cpu') self.state_dict = torch.load(self.model_base_path, map_location='cpu') # compatibility with older model saves which used the 'extractor' name if 'extractor' in self.state_dict: weights = self.state_dict['extractor'] elif 'disk' in self.state_dict: weights = self.state_dict['disk'] else: raise KeyError('Incompatible weight file!') self.disk = DISK(window=8, desc_dim=desc_dim) print('==> Loading pre-trained network.') self.disk.load_state_dict(weights) self.model = self.disk.to(self.DEV) print('==> Successfully loaded pre-trained network.') self.keypoint_size = 8 # just a representative size for visualization and in order to convert extracted points to cv2.KeyPoint self.pts = [] self.kps = [] self.des = [] self.scales = [] self.scores = [] self.frame = None self.use_crop = False self.cropx = [0,0] # [startx, endx] self.cropy = [0,0] # [starty, endy] def crop_center(self,img,cropx,cropy): y,x = img.shape startx = x//2-(cropx//2) starty = y//2-(cropy//2) return img[starty:starty+cropy,startx:startx+cropx] def extract(self, image): if self.mode == 'nms': extract = partial( self.model.features, kind='nms', window_size=self.nms_window_size, cutoff=0., n=self.num_features ) else: extract = partial(model.features, kind='rng') self.use_crop = False height, width, channels = image.shape cropx = width % 16 cropy = height % 16 if cropx != 0 or cropy !=0: self.use_crop = True half_cropx = cropx //2 rest_cropx = cropx %2 half_cropy = cropy //2 rest_cropy = cropy %2 self.cropx = [half_cropx, width-(half_cropx+rest_cropx)] self.cropy = [half_cropy, height-(half_cropy+rest_cropy)] cropped_image = image[self.cropy[0]:self.cropy[1],self.cropx[0]:self.cropx[1],:] image_adapter = ImageAdapter(cropped_image) else: image_adapter = ImageAdapter(image) bitmaps, images = image_adapter.stack() bitmaps = bitmaps.to(self.DEV, non_blocking=True) with torch.no_grad(): try: batched_features = extract(bitmaps) except RuntimeError as e: if 'U-Net failed' in str(e): msg = ('Please use input size which is multiple of 16 (or ' 'adjust the --height and --width flags to let this ' 'script rescale it automatically). This is because ' 'we internally use a U-Net with 4 downsampling ' 'steps, each by a factor of 2, therefore 2^4=16.') raise RuntimeError(msg) from e else: raise for features, image in zip(batched_features.flat, images): features = features.to(self.CPU) kps_crop_space = features.kp.T kps_img_space, mask = image.to_image_coord(kps_crop_space) keypoints = kps_img_space.numpy().T[mask] descriptors = features.desc.numpy()[mask] scores = features.kp_logp.numpy()[mask] order = np.argsort(scores)[::-1] keypoints = keypoints[order] descriptors = descriptors[order] scores = scores[order] assert descriptors.shape[1] == self.desc_dim assert keypoints.shape[1] == 2 return keypoints, descriptors, scores def compute_kps_des(self, im): with self.lock: keypoints, descriptors, scores = self.extract(im) #print('scales:',self.scales) if self.use_crop: self.kps = convert_pts_to_keypoints(keypoints, scores, self.keypoint_size) else: self.kps = convert_pts_to_keypoints_with_translation(keypoints, scores, self.keypoint_size, self.cropx[0], self.cropy[0]) return self.kps, descriptors def detectAndCompute(self, frame, mask=None): #mask is a fake input with self.lock: self.frame = frame self.kps, self.des = self.compute_kps_des(frame) if kVerbose: print('detector: DISK, descriptor: DISK, #features: ', len(self.kps), ', frame res: ', frame.shape[0:2]) return self.kps, self.des # return keypoints if available otherwise call detectAndCompute() def detect(self, frame, mask=None): # mask is a fake input with self.lock: #if self.frame is not frame: self.detectAndCompute(frame) return self.kps # return descriptors if available otherwise call detectAndCompute() def compute(self, frame, kps=None, mask=None): # kps is a fake input, mask is a fake input with self.lock: if self.frame is not frame: Printer.orange('WARNING: DISK is recomputing both kps and des on last input frame', frame.shape) self.detectAndCompute(frame) return self.kps, self.des # return descriptors if available otherwise call detectAndCompute() def compute(self, frame, kps=None, mask=None): # kps is a fake input, mask is a fake input with self.lock: if self.frame is not frame: #Printer.orange('WARNING: DISK is recomputing both kps and des on last input frame', frame.shape) self.detectAndCompute(frame) return self.kps, self.des
12,929
37.254438
181
py
pyslam
pyslam-master/thirdparty/l2net_keras/src/LRN.py
from keras import backend as K from keras.layers.core import Layer # from https://github.com/ckoren1975/Machine-learning/blob/master/googlenet_custom_layers.py # except channels have been moved from the 2nd position to the 4th postion # and shape of input vector is now a tensor operation # and default args are set to L2-net params class LRN(Layer): def __init__(self, alpha=256,k=0,beta=0.5,n=256, **kwargs): super(LRN, self).__init__(**kwargs) self.alpha = alpha self.k = k self.beta = beta self.n = n def call(self, x, mask=None): s = K.shape(x) b = s[0] r = s[1] c = s[2] ch = s[3] half_n = self.n // 2 # half the local region input_sqr = K.square(x) # square the input extra_channels = K.zeros((b, r, c, ch + 2 * half_n)) input_sqr = K.concatenate([extra_channels[:, :, :, :half_n],input_sqr, extra_channels[:, :, :, half_n + ch:]], axis = 3) scale = self.k # offset for the scale norm_alpha = self.alpha / self.n # normalized alpha for i in range(self.n): scale += norm_alpha * input_sqr[:, :, :, i:i+ch] scale = scale ** self.beta x = x / scale return x def get_config(self): config = {"alpha": self.alpha, "k": self.k, "beta": self.beta, "n": self.n} base_config = super(LRN, self).get_config() return dict(list(base_config.items()) + list(config.items()))
1,616
31.34
132
py
pyslam
pyslam-master/thirdparty/l2net_keras/src/L2_Net.py
import os import warnings # to disable tensorflow-numpy warnings: from https://github.com/tensorflow/tensorflow/issues/30427 warnings.filterwarnings('ignore', category=FutureWarning) import keras from keras import backend as K from keras.models import Sequential from keras.layers import Conv2D, BatchNormalization, ZeroPadding2D, Lambda # from https://kobkrit.com/using-allow-growth-memory-option-in-tensorflow-and-keras-dc8c8081bc96 from keras.callbacks import ModelCheckpoint from keras.models import Model, load_model, save_model, Sequential from keras.layers import Dense, Activation, Dropout, Input, Masking, TimeDistributed, LSTM, Conv1D from keras.layers import GRU, Bidirectional, BatchNormalization, Reshape from keras.optimizers import Adam from keras.backend.tensorflow_backend import set_session if False: import tensorflow as tf else: # from https://stackoverflow.com/questions/56820327/the-name-tf-session-is-deprecated-please-use-tf-compat-v1-session-instead import tensorflow.compat.v1 as tf from utils_tf import set_tf_logging import pickle import numpy as np from LRN import LRN import cv2 # get the location of this file! __location__ = os.path.realpath(os.path.join(os.getcwd(), os.path.dirname(__file__))) def build_cnn(weights): model = Sequential() model.add(ZeroPadding2D(1, input_shape=(32,32, 1))) model.add(Conv2D(32, kernel_size=(3, 3))) model.add(BatchNormalization(epsilon=0.0001, scale=False, center=False)) model.add(Lambda(K.relu)) model.add(ZeroPadding2D(1)) model.add(Conv2D(32, kernel_size=(3, 3))) model.add(BatchNormalization(epsilon=0.0001, scale=False, center=False)) model.add(Lambda(K.relu)) model.add(ZeroPadding2D(1)) model.add(Conv2D(64, kernel_size=(3, 3), strides=2)) model.add(BatchNormalization(epsilon=0.0001, scale=False, center=False)) model.add(Lambda(K.relu)) model.add(ZeroPadding2D(1)) model.add(Conv2D(64, kernel_size=(3, 3))) model.add(BatchNormalization(epsilon=0.0001, scale=False, center=False)) model.add(Lambda(K.relu)) model.add(ZeroPadding2D(1)) model.add(Conv2D(128, kernel_size=(3, 3), strides=2)) model.add(BatchNormalization(epsilon=0.0001, scale=False, center=False)) model.add(Lambda(K.relu)) model.add(ZeroPadding2D(1)) model.add(Conv2D(128, kernel_size=(3, 3))) model.add(BatchNormalization(epsilon=0.0001, scale=False, center=False)) model.add(Lambda(K.relu)) model.add(Conv2D(128, kernel_size=(8, 8))) model.add(BatchNormalization(epsilon=0.0001, scale=False, center=False)) model.add(LRN(alpha=256,k=0,beta=0.5,n=256)) model.set_weights(weights) return model def build_L2_net(net_name): python_net_data = pickle.load(open(__location__ + "/../python_net_data/" + net_name + ".p", "rb")) return build_cnn(python_net_data['weights']), build_cnn(python_net_data['weights_cen']), python_net_data['pix_mean'], python_net_data['pix_mean_cen'] def cal_L2Net_des(net_name, testPatchs, flagCS = False): """ Get descriptors for one or more patches Parameters ---------- net_name : string One of "L2Net-HP", "L2Net-HP+", "L2Net-LIB", "L2Net-LIB+", "L2Net-ND", "L2Net-ND+", "L2Net-YOS", "L2Net-YOS+", testPatchs : array A numpy array of image data with deimensions (?, 32, 32, 1), or if using central-surround with deimensions (?, 64, 64, 1) flagCS : boolean If True, use central-surround network Returns ------- descriptor Numpy array with size (?, 128) or if using central-surround (?, 256) """ model, model_cen, pix_mean, pix_mean_cen = build_L2_net(net_name) # print(model.summary()) # print(model_cen.summary()) if flagCS: testPatchsCen = testPatchs[:,16:48,16:48,:] testPatchsCen = testPatchsCen - pix_mean_cen testPatchsCen = np.array([(testPatchsCen[i] - np.mean(testPatchsCen[i]))/(np.std(testPatchsCen[i]) + 1e-12) for i in range(0, testPatchsCen.shape[0])]) testPatchs = np.array([cv2.resize(testPatchs[i], (32,32), interpolation = cv2.INTER_CUBIC) for i in range(0, testPatchs.shape[0])]) testPatchs = np.expand_dims(testPatchs, axis=-1) testPatchs = testPatchs - pix_mean testPatchs = np.array([(testPatchs[i] - np.mean(testPatchs[i]))/(np.std(testPatchs[i]) + 1e-12) for i in range(0, testPatchs.shape[0])]) res = np.reshape(model.predict(testPatchs), (testPatchs.shape[0], 128)) if flagCS: resCen = np.reshape(model_cen.predict(testPatchsCen), (testPatchs.shape[0], 128)) return np.concatenate((res, resCen), 1) else: return res class L2Net: def __init__(self, net_name, do_tf_logging=True, flagCS = False): set_tf_logging(do_tf_logging) # from https://kobkrit.com/using-allow-growth-memory-option-in-tensorflow-and-keras-dc8c8081bc96 config = tf.ConfigProto() config.gpu_options.allow_growth = True # dynamically grow the memory used on the GPU #config.log_device_placement = True # to log device placement (on which device the operation ran) sess = tf.Session(config=config) set_session(sess) # set this TensorFlow session as the default session for Keras model, model_cen, pix_mean, pix_mean_cen = build_L2_net(net_name) self.flagCS = flagCS self.model = model self.model_cen = model_cen self.pix_mean = pix_mean self.pix_mean_cen = pix_mean_cen def calc_descriptors(self, patches): if self.flagCS: patchesCen = patches[:,16:48,16:48,:] patchesCen = patchesCen - self.pix_mean_cen patchesCen = np.array([(patchesCen[i] - np.mean(patchesCen[i]))/(np.std(patchesCen[i]) + 1e-12) for i in range(0, patchesCen.shape[0])]) patches = np.array([cv2.resize(patches[i], (32,32), interpolation = cv2.INTER_CUBIC) for i in range(0, patches.shape[0])]) patches = np.expand_dims(patches, axis=-1) patches = patches - self.pix_mean patches = np.array([(patches[i] - np.mean(patches[i]))/(np.std(patches[i]) + 1e-12) for i in range(0, patches.shape[0])]) res = np.reshape(self.model.predict(patches), (patches.shape[0], 128)) if self.flagCS: resCen = np.reshape(self.model_cen.predict(patchesCen), (patches.shape[0], 128)) return np.concatenate((res, resCen), 1) else: return res # data = np.full((1,64,64,1), 0.) # result = cal_L2Net_des("L2Net-HP+", data, flagCS=True) # print(result)
6,657
34.227513
159
py
pyslam
pyslam-master/thirdparty/l2net/l2net_model.py
import torch import torch.nn.init import torch.nn as nn eps = 1e-10 class L2Norm(nn.Module): def __init__(self): super(L2Norm,self).__init__() self.eps = 1e-10 def forward(self, x): norm = torch.sqrt(torch.sum(x * x, dim = 1) + self.eps) x= x / norm.unsqueeze(-1).expand_as(x) return x class L2Net(nn.Module): def __init__(self): super(L2Net, self).__init__() self.features = nn.Sequential( nn.Conv2d(1, 32, kernel_size=3, padding=1, bias=True), nn.BatchNorm2d(32, affine=True, eps=eps), nn.ReLU(), nn.Conv2d(32, 32, kernel_size=3, padding=1, bias=True), nn.BatchNorm2d(32, affine=True, eps=eps), nn.ReLU(), nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1, bias=True), nn.BatchNorm2d(64, affine=True, eps=eps), nn.ReLU(), nn.Conv2d(64, 64, kernel_size=3, padding=1, bias=True), nn.BatchNorm2d(64, affine=True, eps=eps), nn.ReLU(), nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1, bias=True), nn.BatchNorm2d(128, affine=True, eps=eps), nn.ReLU(), nn.Conv2d(128, 128, kernel_size=3, padding=1, bias=True), nn.BatchNorm2d(128, affine=True, eps=eps), nn.ReLU(), nn.Conv2d(128, 128, kernel_size=8, bias=True), nn.BatchNorm2d(128, affine=True, eps=eps), ) return def input_norm(self, x): # matlab norm z = x.contiguous().transpose(2, 3).contiguous().view(x.size(0),-1) x_minus_mean = z.transpose(0,1)-z.mean(1) sp = torch.std(z,1).detach() norm_inp = x_minus_mean/(sp+1e-12) norm_inp = norm_inp.transpose(0, 1).view(-1, 1, x.size(2), x.size(3)).transpose(2,3) return norm_inp def forward(self, input): norm_img = self.input_norm(input) x_features = self.features(norm_img) return nn.LocalResponseNorm(256,1*256,0.5,0.5)(x_features).view(input.size(0),-1)
2,081
34.288136
92
py
pyslam
pyslam-master/thirdparty/l2net/convert_l2net_weights_matconv_pytorch.py
import numpy as np import scipy.io as sio import torch import torch.nn.init from misc.l2net.l2net_model import L2Net eps = 1e-10 def check_ported(l2net_model, test_patch, img_mean): test_patch = test_patch.transpose(3, 2, 0, 1)-img_mean desc = l2net_model(torch.from_numpy(test_patch)) print(desc) return desc if __name__ == '__main__': path_to_l2net_weights = '/cvlabsrc1/cvlab/datasets_anastasiia/descriptors/sfm-evaluation-benchmarking/third_party/l2net/matlab/L2Net-LIB+.mat' l2net_weights = sio.loadmat(path_to_l2net_weights) l2net_model = L2Net() l2net_model.eval() new_state_dict = l2net_model.state_dict().copy() conv_layers, bn_layers = {}, {} all_layer_weights = l2net_weights['net']['layers'][0][0][0] img_mean = l2net_weights['pixMean'] conv_layers_to_track, bn_layers_to_track = [0,3,6,9,12,15,18], \ [1,4,7,10,13,16,19] conv_i, bn_i = 0,0 for layer in all_layer_weights: if 'weights' not in layer.dtype.names: continue layer_name = layer[0][0][0][0] layer_value = layer['weights'][0][0][0] if layer_name == 'conv': conv_layers[conv_layers_to_track[conv_i]] = layer_value conv_i+=1 elif layer_name == 'bnormPair': bn_layers[bn_layers_to_track[bn_i]] = layer_value bn_i+=1 for key, value in new_state_dict.items(): layer_number = int(key.split('.')[1]) if layer_number in conv_layers.keys(): if 'weight' in key: new_state_dict[key] = torch.from_numpy(conv_layers[layer_number][0].transpose((3,2,0,1))) elif 'bias' in key: new_state_dict[key] = torch.from_numpy(conv_layers[layer_number][1]).squeeze() elif layer_number in bn_layers.keys(): if 'running_mean' in key: new_state_dict[key] = torch.from_numpy(np.array([x[0] for x in bn_layers[layer_number][2]])).squeeze() elif 'running_var' in key: new_state_dict[key] = torch.from_numpy(np.array([x[1] for x in bn_layers[layer_number][2]] )** 2 -eps).squeeze() elif 'weight' in key: new_state_dict[key] = torch.from_numpy(np.ones(value.size()[0])).squeeze() else: continue l2net_model.load_state_dict(new_state_dict) l2net_model.eval() torch.save(l2net_model.state_dict(),'l2net_ported_weights_lib+.pth') # compare desc on test patch with matlab implementation # test_patch_batch = sio.loadmat('test_batch_img.mat')['testPatch'] # check_ported(l2net_model, test_patch_batch, img_mean) # # test_patch_one = sio.loadmat('test_one.mat')['testPatch'] # check_ported(l2net_model, np.expand_dims(np.expand_dims(test_patch_one, axis=2),axis=2), img_mean)
2,854
37.066667
146
py
pyslam
pyslam-master/thirdparty/contextdesc/models/cnn_wrapper/network.py
#!/usr/bin/env python3 """ Copyright 2017, Zixin Luo, HKUST. CNN layer wrapper. Please be noted about following issues: 1. The center and scale paramter are disabled by default for all BN-related layers, as they have shown little influence on final performance. In particular, scale params is officially considered unnecessary as oftentimes followed by ReLU. 2. By default we apply L2 regularization only on kernel or bias parameters, but not learnable BN coefficients (i.e. center/scale) as suggested in ResNet paper. Be noted to add regularization terms into tf.GraphKeys.REGULARIZATION_LOSSES if you are desgining custom layers. 3. Since many of models are converted from Caffe, we are by default setting the epsilon paramter in BN to 1e-5 as that is in Caffe, while 1e-3 in TensorFlow. It may cause slightly different behavior if you are using models from other deep learning toolboxes. """ import numpy as np import tensorflow as tf # Zero padding in default. 'VALID' gives no padding. DEFAULT_PADDING = 'SAME' def caffe_like_padding(input_tensor, padding): """A padding method that has same behavior as Caffe's.""" def PAD(x): return [x, x] if len(input_tensor.get_shape()) == 4: padded_input = tf.pad(input_tensor, [PAD(0), PAD(padding), PAD(padding), PAD(0)], "CONSTANT") elif len(input_tensor.get_shape()) == 5: padded_input = tf.pad(input_tensor, [PAD(0), PAD(padding), PAD(padding), PAD(padding), PAD(0)], "CONSTANT") return padded_input def layer(op): """Decorator for composable network layers.""" def layer_decorated(self, *args, **kwargs): """Layer decoration.""" # We allow to construct low-level layers instead of high-level networks. if self.inputs is None or (len(args) > 0 and isinstance(args[0], tf.Tensor)): layer_output = op(self, *args, **kwargs) return layer_output # Automatically set a name if not provided. name = kwargs.setdefault('name', self.get_unique_name(op.__name__)) # Figure out the layer inputs. if not self.terminals: raise RuntimeError('No input variables found for layer %s.' % name) elif len(self.terminals) == 1: layer_input = self.terminals[0] else: layer_input = list(self.terminals) # Perform the operation and get the output. layer_output = op(self, layer_input, *args, **kwargs) # Add to layer LUT. self.layers[name] = layer_output # This output is now the input for the next layer. self.feed(layer_output) # Return self for chained calls. return self return layer_decorated class Network(object): """Class NetWork.""" def __init__(self, inputs, is_training, dropout_rate=0.5, seed=None, epsilon=1e-5, reuse=False, fcn=True, regularize=True, **kwargs): # The input nodes for this network self.inputs = inputs # If true, the resulting variables are set as trainable self.trainable = is_training if isinstance(is_training, bool) else True # If true, variables are shared between feature towers self.reuse = reuse # If true, layers like batch normalization or dropout are working in training mode self.training = is_training # Dropout rate self.dropout_rate = dropout_rate # Seed for randomness self.seed = seed # Add regularizer for parameters. self.regularizer = tf.contrib.layers.l2_regularizer(1.0) if regularize else None # The epsilon paramater in BN layer. self.bn_epsilon = epsilon self.extra_args = kwargs # Endpoints. self.endpoints = [] if inputs is not None: # The current list of terminal nodes self.terminals = [] # Mapping from layer names to layers self.layers = dict(inputs) # If true, dense layers will be omitted in network construction self.fcn = fcn self.setup() def setup(self): '''Construct the network. ''' raise NotImplementedError('Must be implemented by the subclass.') def load(self, data_path, session, ignore_missing=False, exclude_var=None): '''Load network weights. data_path: The path to the numpy-serialized network weights session: The current TensorFlow session ignore_missing: If true, serialized weights for missing layers are ignored. ''' data_dict = np.load(data_path, encoding='latin1').item() if exclude_var is not None: keyword = exclude_var.split(',') assign_op = [] for op_name in data_dict: if exclude_var is not None: find_keyword = False for tmp_keyword in keyword: if op_name.find(tmp_keyword) >= 0: find_keyword = True if find_keyword: continue with tf.variable_scope(op_name, reuse=True): for param_name, data in data_dict[op_name].items(): try: var = tf.get_variable(param_name) assign_op.append(var.assign(data)) except ValueError: if not ignore_missing: raise else: print(Notify.WARNING, ':'.join( [op_name, param_name]), "is omitted.", Notify.ENDC) session.run(assign_op) def feed(self, *args): '''Set the input(s) for the next operation by replacing the terminal nodes. The arguments can be either layer names or the actual layers. ''' assert args self.terminals = [] for fed_layer in args: if isinstance(fed_layer, str): try: fed_layer = self.layers[fed_layer] except KeyError: raise KeyError('Unknown layer name fed: %s' % fed_layer) self.terminals.append(fed_layer) return self def get_output(self): '''Returns the current network output.''' return self.terminals[-1] def get_output_by_name(self, layer_name): ''' Get graph node by layer name :param layer_name: layer name string :return: tf node ''' return self.layers[layer_name] def get_unique_name(self, prefix): '''Returns an index-suffixed unique name for the given prefix. This is used for auto-generating layer names based on the type-prefix. ''' ident = sum(t.startswith(prefix) for t, _ in self.layers.items()) + 1 return '%s_%d' % (prefix, ident) def change_inputs(self, input_tensors): assert len(input_tensors) == 1 for key in input_tensors: self.layers[key] = input_tensors[key] @layer def conv(self, input_tensor, kernel_size, filters, strides, name, relu=True, dilation_rate=1, padding=DEFAULT_PADDING, biased=True, reuse=False, kernel_init=None, bias_init=tf.zeros_initializer, separable=False): """2D/3D convolution.""" kwargs = {'filters': filters, 'kernel_size': kernel_size, 'strides': strides, 'activation': tf.nn.relu if relu else None, 'use_bias': biased, 'dilation_rate': dilation_rate, 'trainable': self.trainable, 'reuse': self.reuse or reuse, 'bias_regularizer': self.regularizer if biased else None, 'kernel_initializer': kernel_init, 'bias_initializer': bias_init, 'name': name} if separable: kwargs['depthwise_regularizer'] = self.regularizer kwargs['pointwise_regularizer'] = self.regularizer else: kwargs['kernel_regularizer'] = self.regularizer if isinstance(padding, str): padded_input = input_tensor kwargs['padding'] = padding else: padded_input = caffe_like_padding(input_tensor, padding) kwargs['padding'] = 'VALID' if len(input_tensor.get_shape()) == 4: if not separable: return tf.compat.v1.layers.conv2d(padded_input, **kwargs) else: return tf.layers.separable_conv2d(padded_input, **kwargs) elif len(input_tensor.get_shape()) == 5: if not separable: return tf.layers.conv3d(padded_input, **kwargs) else: raise NotImplementedError('No official implementation for separable_conv3d') else: raise ValueError('Improper input rank for layer: ' + name) @layer def conv_bn(self, input_tensor, kernel_size, filters, strides, name, relu=True, center=False, scale=False, dilation_rate=1, padding=DEFAULT_PADDING, biased=False, separable=False, reuse=False): conv = self.conv(input_tensor, kernel_size, filters, strides, name, relu=False, dilation_rate=dilation_rate, padding=padding, biased=biased, reuse=reuse, separable=separable) conv_bn = self.batch_normalization(conv, name + '/bn', center=center, scale=scale, relu=relu, reuse=reuse) return conv_bn @layer def deconv(self, input_tensor, kernel_size, filters, strides, name, relu=True, padding=DEFAULT_PADDING, biased=True, reuse=False): """2D/3D deconvolution.""" kwargs = {'filters': filters, 'kernel_size': kernel_size, 'strides': strides, 'activation': tf.nn.relu if relu else None, 'use_bias': biased, 'trainable': self.trainable, 'reuse': self.reuse or reuse, 'kernel_regularizer': self.regularizer, 'bias_regularizer': self.regularizer if biased else None, 'name': name} if isinstance(padding, str): padded_input = input_tensor kwargs['padding'] = padding else: padded_input = caffe_like_padding(input_tensor, padding) kwargs['padding'] = 'VALID' if len(input_tensor.get_shape()) == 4: return tf.layers.conv2d_transpose(padded_input, **kwargs) elif len(input_tensor.get_shape()) == 5: return tf.layers.conv3d_transpose(padded_input, **kwargs) else: raise ValueError('Improper input rank for layer: ' + name) @layer def deconv_bn(self, input_tensor, kernel_size, filters, strides, name, relu=True, center=False, scale=False, padding=DEFAULT_PADDING, biased=False, reuse=False): deconv = self.deconv(input_tensor, kernel_size, filters, strides, name, relu=False, padding=padding, biased=biased, reuse=reuse) deconv_bn = self.batch_normalization(deconv, name + '/bn', center=center, scale=scale, relu=relu, reuse=reuse) return deconv_bn @layer def relu(self, input_tensor, name=None): """ReLu activation.""" return tf.nn.relu(input_tensor, name=name) @layer def max_pool(self, input_tensor, pool_size, strides, name, padding=DEFAULT_PADDING): """Max pooling.""" if isinstance(padding, str): padded_input = input_tensor padding_type = padding else: padded_input = caffe_like_padding(input_tensor, padding) padding_type = 'VALID' return tf.layers.max_pooling2d(padded_input, pool_size=pool_size, strides=strides, padding=padding_type, name=name) @layer def avg_pool(self, input_tensor, pool_size, strides, name, padding=DEFAULT_PADDING): """"Average pooling.""" if isinstance(padding, str): padded_input = input_tensor padding_type = padding else: padded_input = caffe_like_padding(input_tensor, padding) padding_type = 'VALID' return tf.layers.average_pooling2d(padded_input, pool_size=pool_size, strides=strides, padding=padding_type, name=name) @layer def concat(self, input_tensors, axis, name): return tf.concat(values=input_tensors, axis=axis, name=name) @layer def add(self, input_tensors, name): return tf.add_n(input_tensors, name=name) @layer def fc(self, input_tensor, num_out, name, biased=True, relu=True, flatten=True, reuse=False): # To behave same to Caffe. if flatten: flatten_tensor = tf.layers.flatten(input_tensor) else: flatten_tensor = input_tensor return tf.layers.dense(flatten_tensor, units=num_out, use_bias=biased, activation=tf.nn.relu if relu else None, trainable=self.trainable, reuse=self.reuse or reuse, kernel_regularizer=self.regularizer, bias_regularizer=self.regularizer if biased else None, name=name) @layer def fc_bn(self, input_tensor, num_out, name, biased=False, relu=True, center=False, scale=False, flatten=True, reuse=False): # To behave same to Caffe. fc = self.fc(input_tensor, num_out, name, relu=False, biased=biased, flatten=flatten, reuse=reuse) fc_bn = self.batch_normalization(fc, name + '/bn', center=center, scale=scale, relu=relu, reuse=reuse) return fc_bn @layer def softmax(self, input_tensor, name, dim=-1): return tf.nn.softmax(input_tensor, dim=dim, name=name) @layer def batch_normalization(self, input_tensor, name, center=False, scale=False, relu=False, reuse=False): """Batch normalization.""" output = tf.layers.batch_normalization(input_tensor, center=center, scale=scale, fused=True, training=self.training, trainable=self.trainable, reuse=self.reuse or reuse, epsilon=self.bn_epsilon, gamma_regularizer=None, # self.regularizer if scale else None, beta_regularizer=None, # self.regularizer if center else None, name=name) if relu: output = self.relu(output, name + '/relu') return output @layer def context_normalization(self, input_tensor, name): """The input is a feature matrix with a shape of BxNx1xD""" mean, variance = tf.nn.moments(input_tensor, axes=[1], keep_dims=True) output = tf.nn.batch_normalization( input_tensor, mean, variance, None, None, self.bn_epsilon) return output @layer def l2norm(self, input_tensor, name, axis=-1): return tf.nn.l2_normalize(input_tensor, axis=axis, name=name) @layer def squeeze(self, input_tensor, axis=None, name=None): return tf.squeeze(input_tensor, axis=axis, name=name) @layer def reshape(self, input_tensor, shape, name=None): return tf.reshape(input_tensor, shape, name=name) @layer def flatten(self, input_tensor, name=None): return tf.layers.flatten(input_tensor, name=name) @layer def tanh(self, input_tensor, name=None): return tf.tanh(input_tensor, name=name)
17,219
38.586207
110
py
pyslam
pyslam-master/test/thirdparty/test_delf.py
# Copyright 2017 The TensorFlow Authors All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # Forked from: # https://github.com/tensorflow/models/blob/master/research/delf/delf/python/examples/extract_features.py """Extracts DELF features from a given image and save results to file. The images must be in JPG format. The program checks if descriptors already exist, and skips computation for those. """ import sys sys.path.append("../../") import config config.cfg.set_lib('delf') import cv2 import warnings # to disable tensorflow-numpy warnings: from https://github.com/tensorflow/tensorflow/issues/30427 warnings.filterwarnings('ignore', category=FutureWarning) import argparse import os import sys import time import json import numpy as np import h5py if False: import tensorflow as tf else: # from https://stackoverflow.com/questions/56820327/the-name-tf-session-is-deprecated-please-use-tf-compat-v1-session-instead import tensorflow.compat.v1 as tf # from https://kobkrit.com/using-allow-growth-memory-option-in-tensorflow-and-keras-dc8c8081bc96 to cope with the following error: # "[...tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR" tf_config = tf.ConfigProto() tf_config.gpu_options.allow_growth = True tf_config.gpu_options.per_process_gpu_memory_fraction=0.333 # from https://stackoverflow.com/questions/34199233/how-to-prevent-tensorflow-from-allocating-the-totality-of-a-gpu-memory #session = tf.Session(config=tf_config, ...) from google.protobuf import text_format from tensorflow.python.platform import app # from delf import delf_config_pb2 # from delf import feature_extractor # from delf import feature_io from delf.protos import aggregation_config_pb2 from delf.protos import box_pb2 from delf.protos import datum_pb2 from delf.protos import delf_config_pb2 from delf.protos import feature_pb2 from delf.python import box_io from delf.python import datum_io from delf.python import delf_v1 from delf.python import feature_aggregation_extractor from delf.python import feature_aggregation_similarity from delf.python import feature_extractor from delf.python import feature_io from delf.python.examples import detector from delf.python.examples import extractor from delf.python import detect_to_retrieve from delf.python import google_landmarks_dataset delf_base_path='../../thirdparty/tensorflow_models/research/delf/delf/python/' delf_config_file= delf_base_path + 'examples/delf_config_example.pbtxt' delf_model_path=delf_base_path + 'examples/parameters/delf_gld_20190411/model/' delf_mean_path=delf_base_path + 'examples/parameters/delf_gld_20190411/pca/mean.datum' delf_projection_matrix_path=delf_base_path + 'examples/parameters/delf_gld_20190411/pca/pca_proj_mat.datum' cmd_args = None # Extension of feature files. _DELF_EXT = '.h5' # Pace to report extraction log. _STATUS_CHECK_ITERATIONS = 100 def _ReadImageList(list_path): """Helper function to read image paths. Args: list_path: Path to list of images, one image path per line. Returns: image_paths: List of image paths. """ with tf.gfile.GFile(list_path, 'r') as f: image_paths = f.readlines() image_paths = [entry.rstrip() for entry in image_paths] return image_paths def MakeExtractor(sess, config, import_scope=None): """Creates a function to extract features from an image. Args: sess: TensorFlow session to use. config: DelfConfig proto containing the model configuration. import_scope: Optional scope to use for model. Returns: Function that receives an image and returns features. """ tf.saved_model.loader.load( sess, [tf.saved_model.tag_constants.SERVING], config.model_path, import_scope=import_scope) import_scope_prefix = import_scope + '/' if import_scope is not None else '' input_image = sess.graph.get_tensor_by_name('%sinput_image:0' % import_scope_prefix) input_score_threshold = sess.graph.get_tensor_by_name( '%sinput_abs_thres:0' % import_scope_prefix) input_image_scales = sess.graph.get_tensor_by_name('%sinput_scales:0' % import_scope_prefix) input_max_feature_num = sess.graph.get_tensor_by_name( '%sinput_max_feature_num:0' % import_scope_prefix) boxes = sess.graph.get_tensor_by_name('%sboxes:0' % import_scope_prefix) raw_descriptors = sess.graph.get_tensor_by_name('%sfeatures:0' % import_scope_prefix) feature_scales = sess.graph.get_tensor_by_name('%sscales:0' % import_scope_prefix) attention_with_extra_dim = sess.graph.get_tensor_by_name( '%sscores:0' % import_scope_prefix) attention = tf.reshape(attention_with_extra_dim, [tf.shape(attention_with_extra_dim)[0]]) locations, descriptors = feature_extractor.DelfFeaturePostProcessing( boxes, raw_descriptors, config) def ExtractorFn(image): """Receives an image and returns DELF features. Args: image: Uint8 array with shape (height, width 3) containing the RGB image. Returns: Tuple (locations, descriptors, feature_scales, attention) """ return sess.run([locations, descriptors, feature_scales, attention], feed_dict={ input_image: image, input_score_threshold: config.delf_local_config.score_threshold, input_image_scales: list(config.image_scales), input_max_feature_num: config.delf_local_config.max_feature_num }) return ExtractorFn def main(unused_argv): tf.logging.set_verbosity(tf.logging.INFO) # Read list of images. #tf.logging.info('Reading list of images...') #image_paths = _ReadImageList(cmd_args.list_images_path) num_images = 1 #len(image_paths) #tf.logging.info('done! Found %d images', num_images) img = cv2.imread('../data/kitti06-12-color.png',cv2.IMREAD_COLOR) #value = tf.gfile.FastGFile('../data/kitti06-12-color.jpg', 'rb').read() # Parse DelfConfig proto. delf_config = delf_config_pb2.DelfConfig() with tf.gfile.FastGFile(cmd_args.config_path, 'r') as f: text_format.Merge(f.read(), delf_config) delf_config.model_path = delf_model_path delf_config.delf_local_config.pca_parameters.mean_path = delf_mean_path delf_config.delf_local_config.pca_parameters.projection_matrix_path = delf_projection_matrix_path print('config:', delf_config) # Create output directory if necessary. if not os.path.exists(cmd_args.output_dir): os.makedirs(cmd_args.output_dir) # Tell TensorFlow that the model will be built into the default Graph. with tf.Graph().as_default(): # Reading list of images. #filename_queue = tf.train.string_input_producer(image_paths, shuffle=False) #reader = tf.WholeFileReader() #_, value = reader.read(filename_queue) #image_tf = tf.image.decode_jpeg(value, channels=3) # Returns a `Tensor` of type `uint8`. # from https://stackoverflow.com/questions/48727264/how-to-convert-numpy-array-image-to-tensorflow-image #image_tf = np.array(img)[:, :, 0:3] #image_tf = tf.convert_to_tensor(img, np.float32) #print('\nimagetf info',np.info(image_tf)) # run the network to get the predictions #predictions = sess.run(finalTensor, {'DecodeJpeg:0': tfImage}) with tf.Session(config=tf_config) as sess: init_op = tf.global_variables_initializer() sess.run(init_op) extractor_fn = MakeExtractor(sess, delf_config) # Start input enqueue threads. #coord = tf.train.Coordinator() #threads = tf.train.start_queue_runners(sess=sess, coord=coord) start = time.clock() with h5py.File(os.path.join(cmd_args.output_dir, 'keypoints.h5'), 'w') as h5_kp, \ h5py.File(os.path.join(cmd_args.output_dir, 'descriptors.h5'), 'w') as h5_desc, \ h5py.File(os.path.join(cmd_args.output_dir, 'scores.h5'), 'w') as h5_score, \ h5py.File(os.path.join(cmd_args.output_dir, 'scales.h5'), 'w') as h5_scale: for i in range(num_images): key = 'img' #os.path.splitext(os.path.basename(image_paths[i]))[0] print('Processing "{}"'.format(key)) # # Write to log-info once in a while. # if i == 0: # tf.logging.info( # 'Starting to extract DELF features from images...') # elif i % _STATUS_CHECK_ITERATIONS == 0: # elapsed = (time.clock() - start) # tf.logging.info( # 'Processing image %d out of %d, last %d ' # 'images took %f seconds', i, num_images, # _STATUS_CHECK_ITERATIONS, elapsed) # start = time.clock() # # Get next image. image_tf = tf.convert_to_tensor(img, np.float32) im = sess.run(image_tf) #im = sess.run({'DecodeJpeg:0': image_tf}) # from https://stackoverflow.com/questions/48727264/how-to-convert-numpy-array-image-to-tensorflow-image # If descriptor already exists, skip its computation. # out_desc_filename = os.path.splitext(os.path.basename( # image_paths[i]))[0] + _DELF_EXT # out_desc_fullpath = os.path.join(cmd_args.output_dir, out_desc_filename) # if tf.gfile.Exists(out_desc_fullpath): # tf.logging.info('Skipping %s', image_paths[i]) # continue # Extract and save features. (locations_out, descriptors_out, feature_scales_out, attention_out) = extractor_fn(im) # np.savez('{}.npz'.format(config.delf_local_config.max_feature_num), keypoints=locations_out) # feature_io.WriteToFile(out_desc_fullpath, locations_out, # feature_scales_out, descriptors_out, # attention_out) h5_kp[key] = locations_out[:, ::-1] h5_desc[key] = descriptors_out h5_scale[key] = feature_scales_out h5_score[key] = attention_out print('#extracted keypoints:',len(h5_kp[key])) print('des[0]:',descriptors_out[0]) print('done!') # Finalize enqueue threads. #coord.request_stop() #coord.join(threads) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.register('type', 'bool', lambda v: v.lower() == 'true') parser.add_argument( '--config_path', type=str, default=delf_config_file, #'misc/delf/delf_config_example.pbtxt', help=""" Path to DelfConfig proto text file with configuration to be used for DELF extraction. """) parser.add_argument( '--list_images_path', type=str, help=""" Path to list of images whose DELF features will be extracted. """) parser.add_argument( '--output_dir', type=str, default='./delf', help=""" Directory where DELF features will be written to. Each image's features will be written to a file with same name, and extension replaced by .delf. """) cmd_args, unparsed = parser.parse_known_args() app.run(main=main, argv=[sys.argv[0]] + unparsed)
12,641
41.280936
183
py
pyslam
pyslam-master/test/thirdparty/test_tfeat.py
import sys sys.path.append("../../") import config config.cfg.set_lib('tfeat') import torchvision as tv import phototour import torch from tqdm import tqdm import numpy as np import torch.nn as nn import math import tfeat_model import torch.optim as optim import torch.nn.functional as F import torch.backends.cudnn as cudnn import os import cv2 import tfeat_utils import numpy as np import cv2 from matplotlib import pyplot as plt tfeat_base_path='../../thirdparty/tfeat/' # open images img1 = cv2.imread(tfeat_base_path + 'imgs/v_churchill/1.ppm',0) img2 = cv2.imread(tfeat_base_path + 'imgs/v_churchill/6.ppm',0) #init tfeat and load the trained weights tfeat = tfeat_model.TNet() models_path = tfeat_base_path + 'pretrained-models' net_name = 'tfeat-liberty' tfeat.load_state_dict(torch.load(os.path.join(models_path,net_name+".params"))) tfeat.cuda() tfeat.eval() # use BRISK detector brisk = cv2.BRISK_create() kp1, des1 = brisk.detectAndCompute(img1,None) kp2, des2 = brisk.detectAndCompute(img2,None) bf = cv2.BFMatcher(cv2.NORM_HAMMING) matches = bf.knnMatch(des1,des2, k=2) # Apply ratio test good = [] for m,n in matches: if m.distance < 0.8*n.distance: good.append([m]) img3 = cv2.drawMatchesKnn(img1,kp1,img2,kp2,good,0, flags=2) plt.subplot(1, 2, 1) plt.title('BRISK detector and descriptor', fontsize=10) plt.imshow(img3) # mag_factor is how many times the original keypoint scale # is enlarged to generate a patch from a keypoint mag_factor = 3 #print('kp1: ', kp1) desc_tfeat1 = tfeat_utils.describe_opencv(tfeat, img1, kp1, 32, mag_factor) desc_tfeat2 = tfeat_utils.describe_opencv(tfeat, img2, kp2, 32, mag_factor) bf = cv2.BFMatcher(cv2.NORM_L2) matches = bf.knnMatch(desc_tfeat1,desc_tfeat2, k=2) # Apply ratio test good = [] for m,n in matches: if m.distance < 0.8*n.distance: good.append([m]) img4 = cv2.drawMatchesKnn(img1,kp1,img2,kp2,good,0, flags=2) plt.subplot(1, 2, 2) plt.title('BRISK detector and TFEAT descriptor', fontsize=10) plt.imshow(img4) plt.show()
2,048
24.936709
79
py
pyslam
pyslam-master/test/thirdparty/test_hardnet_dense.py
#!/usr/bin/python3 -utt # -*- coding: utf-8 -*- import sys import torch import torch.nn as nn import torch.nn.functional as F from torch.autograd import Variable import torch.backends.cudnn as cudnn import time import os #sys.path.insert(0, '/home/ubuntu/dev/opencv-3.1/build/lib') import cv2 import math import numpy as np from PIL import Image hardnet_base_path='../../thirdparty/hardnet/' class L2Norm(nn.Module): def __init__(self): super(L2Norm,self).__init__() self.eps = 1e-10 def forward(self, x): norm = torch.sqrt(torch.sum(x * x, dim = 1, keepdim = True) + self.eps) x= x / norm.expand_as(x) return x class LocalNorm2d(nn.Module): def __init__(self, kernel_size = 32): super(LocalNorm2d, self).__init__() self.ks = kernel_size self.pool = nn.AvgPool2d(kernel_size = self.ks, stride = 1, padding = 0) self.eps = 1e-10 return def forward(self,x): pd = int(self.ks/2) mean = self.pool(F.pad(x, (pd,pd,pd,pd), 'reflect')) return torch.clamp((x - mean) / (torch.sqrt(torch.abs(self.pool(F.pad(x*x, (pd,pd,pd,pd), 'reflect')) - mean*mean )) + self.eps), min = -6.0, max = 6.0) class DenseHardNet(nn.Module): """HardNet model definition """ def __init__(self, _stride = 2): super(DenseHardNet, self).__init__() self.input_norm = LocalNorm2d(17) self.features = nn.Sequential( nn.Conv2d(1, 32, kernel_size=3, padding=1, bias = False), nn.BatchNorm2d(32, affine=False), nn.ReLU(), nn.Conv2d(32, 32, kernel_size=3, padding=1, bias = False), nn.BatchNorm2d(32, affine=False), nn.ReLU(), nn.Conv2d(32, 64, kernel_size=3, stride=_stride, padding=1, bias = False), nn.BatchNorm2d(64, affine=False), nn.ReLU(), nn.Conv2d(64, 64, kernel_size=3, padding=1, bias = False), nn.BatchNorm2d(64, affine=False), nn.ReLU(), nn.Conv2d(64, 128, kernel_size=3, stride=_stride,padding=1, bias = False), nn.BatchNorm2d(128, affine=False), nn.ReLU(), nn.Conv2d(128, 128, kernel_size=3, padding=1, bias = False), nn.BatchNorm2d(128, affine=False), nn.ReLU(), nn.Dropout(0.1), nn.Conv2d(128, 128, kernel_size=8, bias = False), nn.BatchNorm2d(128, affine=False), L2Norm() ) return def forward(self, input, upscale = False): if input.size(1) > 1: feats = self.features(self.input_norm(input.mean(dim = 1, keepdim = True))) else: feats = self.features(self.input_norm(input)) if upscale: return F.upsample(feats, (input.size(2), input.size(3)),mode='bilinear') return feats def load_grayscale_var(fname): img = Image.open(fname).convert('RGB') img = np.mean(np.array(img), axis = 2) # convert to gray scale by averaging R-G-B channels var_image = torch.autograd.Variable(torch.from_numpy(img.astype(np.float32)), volatile = True) var_image_reshape = var_image.view(1, 1, var_image.size(0),var_image.size(1)) return var_image_reshape if __name__ == '__main__': DO_CUDA = True UPSCALE = False stride = 2; try: if len(sys.argv) > 1: input_img_fname = sys.argv[1] output_fname = sys.argv[2] if len(sys.argv) > 3: DO_CUDA = sys.argv[3] != 'cpu' if len(sys.argv) > 4: UPSCALE = sys.argv[4] == 'UPSCALE' if sys.argv[4] == 'NOSTRIDE': stride = 1 else: input_img_fname = '../data/kitti06-12-color.png' output_fname = 'hardnet.out.txt' DO_CUDA = True except: #print("Wrong input format. Try ./extract_DenseHardNet.py imgs/ref.png out.txt gpu") print("Wrong input format. Try " + sys.argv[0] + " imgs/ref.png out.txt gpu") sys.exit(1) model_weights = hardnet_base_path + 'pretrained/pretrained_all_datasets/HardNet++.pth' model = DenseHardNet(stride) checkpoint = torch.load(model_weights) model.load_state_dict(checkpoint['state_dict']) model.eval() img = load_grayscale_var(input_img_fname) if DO_CUDA: model.cuda() img = img.cuda() print('Extracting on GPU') else: print('Extracting on CPU') model = model.cpu() t = time.time() with torch.no_grad(): desc = model(img, UPSCALE) et = time.time() - t print('processing', et) desc_numpy = desc.cpu().detach().float().squeeze().numpy(); desc_numpy = np.clip(((desc_numpy + 0.45) * 210.0).astype(np.int32), 0, 255).astype(np.uint8) print(desc_numpy.shape) np.save(output_fname, desc_numpy)
4,912
35.93985
161
py
pyslam
pyslam-master/test/thirdparty/test_logpolar.py
# Copyright 2019 EPFL, Google LLC # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # adapted from https://raw.githubusercontent.com/cvlab-epfl/log-polar-descriptors/master/example.py import os import sys import argparse import torch import cv2 import numpy as np import h5py import argparse from time import time import sys sys.path.append("../../") import config config.cfg.set_lib('logpolar') from configs.defaults import _C as cfg from modules.hardnet.models import HardNet logpolar_base_path='../../thirdparty/logpolar/' # Configuration # Set use_log_polar=False to load the "Cartesian" models used in the paper def extract_descriptors(input_filename, output_filename, use_log_polar, num_keypoints, verbose): # Setup ROOT = logpolar_base_path # os.getcwd() if use_log_polar: config_path = os.path.join(ROOT, 'configs', 'init_one_example_ptn_96.yml') if verbose: print('-- Using log-polar models') else: config_path = os.path.join(ROOT, 'configs', 'init_one_example_stn_16.yml') if verbose: print('-- Using cartesian models') cfg.merge_from_file(config_path) os.environ["CUDA_VISIBLE_DEVICES"] = str(0) torch.cuda.manual_seed_all(cfg.TRAINING.SEED) torch.backends.cudnn.deterministic = True device = torch.device("cuda" if torch.cuda.is_available() else "cpu") if verbose: if torch.cuda.is_available(): print('-- Using GPU') else: print('-- Using CPU') # Extract SIFT keypoints img = cv2.imread(input_filename, cv2.IMREAD_GRAYSCALE) # A safe image size is ~1000px on the largest dimension # To extract features on larger images you might want to increase the padding max_size = 1024 if any([s > max_size for s in img.shape]): h, w = img.shape if h > w: img = cv2.resize(img, (int(w * max_size / h), max_size), cv2.INTER_CUBIC) elif w > h: img = cv2.resize(img, (max_size, int(h * max_size / w)), cv2.INTER_CUBIC) h, w = img.shape # get keypoints, scale and locatinos from SIFT or another detector sift = cv2.xfeatures2d.SIFT_create(num_keypoints) keypoints = sift.detect(img, None) pts = np.array([kp.pt for kp in keypoints]) scales = np.array([kp.size for kp in keypoints]) oris = np.array([kp.angle for kp in keypoints]) # Mirror-pad the image to avoid boundary effects if any([s > cfg.TEST.PAD_TO for s in img.shape[:2]]): raise RuntimeError( "Image exceeds acceptable size ({}x{}), please downsample".format( cfg.TEST.PAD_TO, cfg.TEST.PAD_TO)) fillHeight = cfg.TEST.PAD_TO - img.shape[0] fillWidth = cfg.TEST.PAD_TO - img.shape[1] padLeft = int(np.round(fillWidth / 2)) padRight = int(fillWidth - padLeft) padUp = int(np.round(fillHeight / 2)) padDown = int(fillHeight - padUp) img = np.pad(img, pad_width=((padUp, padDown), (padLeft, padRight)), mode='reflect') if verbose: print('-- Padding image from {}x{} to {}x{}'.format( h, w, img.shape[0], img.shape[1])) # Normalize keypoint locations kp_norm = [] for i, p in enumerate(pts): _p = 2 * np.array([(p[0] + padLeft) / (cfg.TEST.PAD_TO), (p[1] + padUp) / (cfg.TEST.PAD_TO)]) - 1 kp_norm.append(_p) theta = [ torch.from_numpy(np.array(kp_norm)).float().squeeze(), torch.from_numpy(scales).float(), torch.from_numpy(np.array([np.deg2rad(o) for o in oris])).float() ] # Instantiate the model t = time() model = HardNet(transform=cfg.TEST.TRANSFORMER, coords=cfg.TEST.COORDS, patch_size=cfg.TEST.IMAGE_SIZE, scale=cfg.TEST.SCALE, is_desc256=cfg.TEST.IS_DESC_256, orientCorrect=cfg.TEST.ORIENT_CORRECTION) # Load weights model.load_state_dict(torch.load(logpolar_base_path + cfg.TEST.MODEL_WEIGHTS)['state_dict']) model.eval() model.to(device) if verbose: print('-- Instantiated model in {:0.2f} sec.'.format(time() - t)) # Extract descriptors imgs, img_keypoints = torch.from_numpy(img).unsqueeze(0).to(device), \ [theta[0].to(device), theta[1].to(device), theta[2].to(device)] t = time() descriptors, patches = model({input_filename: imgs}, img_keypoints, [input_filename] * len(img_keypoints[0])) if verbose: print('-- Computed {} descriptors in {:0.2f} sec.'.format( descriptors.shape[0], time() - t)) keypoints_array = np.concatenate([pts, scales[..., None], oris[..., None]], axis=1) t = time() with h5py.File(output_filename, 'w') as f: f['keypoints'] = keypoints_array f['descriptors'] = descriptors.cpu().detach().numpy() print('-- Saved {} descriptors in {:0.2f} sec.'.format( descriptors.shape[0], time() - t)) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--input', type=str, default=logpolar_base_path+'testImg.jpeg', help='Input image') parser.add_argument('--output', type=str, default=logpolar_base_path+'testImg.h5', help='Output file') parser.add_argument('--use_log_polar', type=bool, default=True, help='Use log-polar models. Set to False to use ' 'cartesian models instead.') parser.add_argument('--num_keypoints', type=int, default=1024, help='Number of keypoints') parser.add_argument('--verbose', type=bool, default=True, help='Set to False to suppress feedback') config, unparsed = parser.parse_known_args() if len(unparsed) > 0: parser.print_usage() else: extract_descriptors(config.input, config.output, config.use_log_polar, config.num_keypoints, config.verbose)
7,015
34.979487
99
py
pyslam
pyslam-master/test/thirdparty/test_l2net_keras.py
import sys sys.path.append("../../") import config config.cfg.set_lib('l2net_keras') import cv2 import numpy as np from L2_Net import L2Net # One of "L2Net-HP", "L2Net-HP+", "L2Net-LIB", "L2Net-LIB+", "L2Net-ND", "L2Net-ND+", "L2Net-YOS", "L2Net-YOS+", net_name = 'L2Net-HP' l2net = L2Net(net_name,do_tf_logging=False) if False: patches = np.random.rand(100, 32, 32, 1) else: patches = np.random.rand(100, 32, 32) patches = np.expand_dims(patches, -1) descrs = l2net.calc_descriptors(patches) print('done!')
537
22.391304
113
py
pyslam
pyslam-master/test/thirdparty/test_hardnet_patches.py
#!/usr/bin/python3 -utt # -*- coding: utf-8 -*- import sys import torch import torch.nn as nn import torch.nn.functional as F from torch.autograd import Variable import torch.backends.cudnn as cudnn import time import os import cv2 import math import numpy as np hardnet_base_path='../../thirdparty/hardnet/' class L2Norm(nn.Module): def __init__(self): super(L2Norm,self).__init__() self.eps = 1e-10 def forward(self, x): norm = torch.sqrt(torch.sum(x * x, dim = 1) + self.eps) x= x / norm.unsqueeze(-1).expand_as(x) return x class L1Norm(nn.Module): def __init__(self): super(L1Norm,self).__init__() self.eps = 1e-10 def forward(self, x): norm = torch.sum(torch.abs(x), dim = 1) + self.eps x= x / norm.expand_as(x) return x class HardNet(nn.Module): """HardNet model definition """ def __init__(self): super(HardNet, self).__init__() self.features = nn.Sequential( nn.Conv2d(1, 32, kernel_size=3, padding=1, bias = False), nn.BatchNorm2d(32, affine=False), nn.ReLU(), nn.Conv2d(32, 32, kernel_size=3, padding=1, bias = False), nn.BatchNorm2d(32, affine=False), nn.ReLU(), nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1, bias = False), nn.BatchNorm2d(64, affine=False), nn.ReLU(), nn.Conv2d(64, 64, kernel_size=3, padding=1, bias = False), nn.BatchNorm2d(64, affine=False), nn.ReLU(), nn.Conv2d(64, 128, kernel_size=3, stride=2,padding=1, bias = False), nn.BatchNorm2d(128, affine=False), nn.ReLU(), nn.Conv2d(128, 128, kernel_size=3, padding=1, bias = False), nn.BatchNorm2d(128, affine=False), nn.ReLU(), nn.Dropout(0.1), nn.Conv2d(128, 128, kernel_size=8, bias = False), nn.BatchNorm2d(128, affine=False), ) #self.features.apply(weights_init) def input_norm(self,x): flat = x.view(x.size(0), -1) mp = torch.mean(flat, dim=1) sp = torch.std(flat, dim=1) + 1e-7 return (x - mp.detach().unsqueeze(-1).unsqueeze(-1).unsqueeze(-1).expand_as(x)) / sp.detach().unsqueeze(-1).unsqueeze(-1).unsqueeze(1).expand_as(x) def forward(self, input): x_features = self.features(self.input_norm(input)) x = x_features.view(x_features.size(0), -1) return L2Norm()(x) if __name__ == '__main__': DO_CUDA = True try: if len(sys.argv) > 1: input_img_fname = sys.argv[1] output_fname = sys.argv[2] if len(sys.argv) > 3: DO_CUDA = sys.argv[3] != 'cpu' else: input_img_fname = '../data/kitti06-12-color.png' output_fname = 'hardnet.out.txt' DO_CUDA = True except: print("Wrong input format. Try " + sys.argv[0] + " imgs/ref.png out.txt gpu") sys.exit(1) model_weights = hardnet_base_path + 'pretrained/train_liberty_with_aug/checkpoint_liberty_with_aug.pth' model = HardNet() checkpoint = torch.load(model_weights) model.load_state_dict(checkpoint['state_dict']) model.eval() if DO_CUDA: model.cuda() print('Extracting on GPU') else: print('Extracting on CPU') model = model.cpu() image = cv2.imread(input_img_fname,0) h,w = image.shape print(h,w) patch_size = 32 nh_patches = math.ceil(h/patch_size) nw_patches = math.ceil(w/patch_size) n_patches = nw_patches * nh_patches print('Amount of patches: {} ({}x{})'.format(n_patches,nh_patches,nw_patches)) t = time.time() patches = np.ndarray((n_patches, 1, 32, 32), dtype=np.float32) for i in range(nh_patches): for j in range(nw_patches): patch = image[i*patch_size: (i+1)*patch_size, j*patch_size: (j+1)*patch_size] patches[i,0,:,:] = cv2.resize(patch,(32,32)) / 255. patches -= 0.443728476019 patches /= 0.20197947209 bs = 128 n_batches = int(n_patches / bs) + 1 t = time.time() descriptors_for_net = np.zeros((len(patches), 128)) for i in range(0, len(patches), bs): data_a = patches[i: i + bs, :, :, :].astype(np.float32) data_a = torch.from_numpy(data_a) if DO_CUDA: data_a = data_a.cuda() data_a = Variable(data_a) # compute output with torch.no_grad(): out_a = model(data_a) descriptors_for_net[i: i + bs,:] = out_a.data.cpu().numpy().reshape(-1, 128) print(descriptors_for_net.shape) assert n_patches == descriptors_for_net.shape[0] et = time.time() - t print('processing', et, et/max(float(n_patches),1), ' per patch') np.savetxt(output_fname, descriptors_for_net, delimiter=' ', fmt='%10.5f')
4,935
33.517483
155
py
pyslam
pyslam-master/test/thirdparty/test_disk.py
# adapted from https://github.com/cvlab-epfl/disk/blob/master/detect.py import torch, os, argparse, h5py, warnings, imageio import numpy as np from tqdm import tqdm import sys sys.path.append("../../") import config config.cfg.set_lib('disk') config.cfg.set_lib('torch-dimcheck') config.cfg.set_lib('torch-localize') config.cfg.set_lib('unets') import multiprocessing as mp import torch, h5py, imageio, os, argparse import numpy as np import torch.nn.functional as F from functools import partial from torch.utils.data import DataLoader from tqdm import tqdm from torch_dimcheck import dimchecked from disk import DISK, Features class Image: def __init__(self, bitmap: ['C', 'H', 'W'], fname: str, orig_shape=None): self.bitmap = bitmap self.fname = fname if orig_shape is None: self.orig_shape = self.bitmap.shape[1:] else: self.orig_shape = orig_shape def resize_to(self, shape): return Image( self._pad(self._interpolate(self.bitmap, shape), shape), self.fname, orig_shape=self.bitmap.shape[1:], ) @dimchecked def to_image_coord(self, xys: [2, 'N']) -> ([2, 'N'], ['N']): f, _size = self._compute_interpolation_size(self.bitmap.shape[1:]) scaled = xys / f h, w = self.orig_shape x, y = scaled mask = (0 <= x) & (x < w) & (0 <= y) & (y < h) return scaled, mask def _compute_interpolation_size(self, shape): x_factor = self.orig_shape[0] / shape[0] y_factor = self.orig_shape[1] / shape[1] f = 1 / max(x_factor, y_factor) if x_factor > y_factor: new_size = (shape[0], int(f * self.orig_shape[1])) else: new_size = (int(f * self.orig_shape[0]), shape[1]) return f, new_size @dimchecked def _interpolate(self, image: ['C', 'H', 'W'], shape) -> ['C', 'h', 'w']: _f, size = self._compute_interpolation_size(shape) return F.interpolate( image.unsqueeze(0), size=size, mode='bilinear', align_corners=False, ).squeeze(0) @dimchecked def _pad(self, image: ['C', 'H', 'W'], shape) -> ['C', 'h', 'w']: x_pad = shape[0] - image.shape[1] y_pad = shape[1] - image.shape[2] if x_pad < 0 or y_pad < 0: raise ValueError("Attempting to pad by negative value") return F.pad(image, (0, y_pad, 0, x_pad)) class SceneDataset: def __init__(self, image_path, crop_size=(None, None)): self.image_path = image_path self.crop_size = crop_size self.names = [p for p in os.listdir(image_path) \ if p.endswith(args.image_extension)] print('image names:',self.names) def __len__(self): return len(self.names) def __getitem__(self, ix): name = self.names[ix] path = os.path.join(self.image_path, name) img = np.ascontiguousarray(imageio.imread(path)) tensor = torch.from_numpy(img).to(torch.float32) if len(tensor.shape) == 2: # some images may be grayscale tensor = tensor.unsqueeze(-1).expand(-1, -1, 3) bitmap = tensor.permute(2, 0, 1) / 255. extensionless_fname = os.path.splitext(name)[0] image = Image(bitmap, extensionless_fname) if self.crop_size != (None, None): image = image.resize_to(self.crop_size) return image @staticmethod def collate_fn(images): bitmaps = torch.stack([im.bitmap for im in images], dim=0) return bitmaps, images def extract(dataset, save_path): dataloader = DataLoader( dataset, batch_size=1, pin_memory=True, collate_fn=dataset.collate_fn, num_workers=4, ) if args.mode == 'nms': extract = partial( model.features, kind='nms', window_size=args.window, cutoff=0., n=args.n ) else: extract = partial(model.features, kind='rng') os.makedirs(os.path.join(save_path), exist_ok=True) keypoint_h5 = h5py.File(os.path.join(save_path, 'keypoints.h5'), 'w') descriptor_h5 = h5py.File(os.path.join(save_path, 'descriptors.h5'), 'w') if args.detection_scores: score_h5 = h5py.File(os.path.join(save_path, 'scores.h5'), 'w') print('loop:') pbar = tqdm(dataloader) for bitmaps, images in pbar: print('bitmaps: ', bitmaps) print('images: ', images) bitmaps = bitmaps.to(DEV, non_blocking=True) with torch.no_grad(): try: batched_features = extract(bitmaps) except RuntimeError as e: if 'U-Net failed' in str(e): msg = ('Please use input size which is multiple of 16 (or ' 'adjust the --height and --width flags to let this ' 'script rescale it automatically). This is because ' 'we internally use a U-Net with 4 downsampling ' 'steps, each by a factor of 2, therefore 2^4=16.') raise RuntimeError(msg) from e else: raise for features, image in zip(batched_features.flat, images): features = features.to(CPU) kps_crop_space = features.kp.T kps_img_space, mask = image.to_image_coord(kps_crop_space) keypoints = kps_img_space.numpy().T[mask] descriptors = features.desc.numpy()[mask] scores = features.kp_logp.numpy()[mask] order = np.argsort(scores)[::-1] keypoints = keypoints[order] descriptors = descriptors[order] scores = scores[order] assert descriptors.shape[1] == args.desc_dim assert keypoints.shape[1] == 2 if args.f16: descriptors = descriptors.astype(np.float16) # keypoint_h5.create_dataset(image.fname, data=keypoints) # descriptor_h5.create_dataset(image.fname, data=descriptors) # if args.detection_scores: # score_h5.create_dataset(image.fname, data=scores) pbar.set_postfix(n=keypoints.shape[0]) if __name__ == '__main__': parser = argparse.ArgumentParser(description=( "Script for detection and description (but not matching) of keypoints. " "It processes all images with extension given by `--image-extension` found " "in `image-path` directory. Images are resized to `--height` x `--width` " "for internal processing (padding them if necessary) and the output " "coordinates are then transformed back to original image size."), formatter_class=argparse.ArgumentDefaultsHelpFormatter ) parser.add_argument( '--height', default=None, type=int, help='rescaled height (px). If unspecified, image is not resized in height dimension' ) parser.add_argument( '--width', default=None, type=int, help='rescaled width (px). If unspecified, image is not resized in width dimension' ) parser.add_argument( '--image-extension', default='ppm', type=str, help='This script ill process all files which match `image-path/*.{--image-extension}`' ) parser.add_argument( '--f16', action='store_true', help='Store descriptors in fp16 (half precision) format' ) parser.add_argument('--window', type=int, default=5, help='NMS window size') parser.add_argument( '--n', type=int, default=None, help='Maximum number of features to extract. If unspecified, the number is not limited' ) parser.add_argument( '--desc-dim', type=int, default=128, help='descriptor dimension. Needs to match the checkpoint value.' ) parser.add_argument( '--mode', choices=['nms', 'rng'], default='nms', help=('Whether to extract features using the non-maxima suppresion mode or ' 'through training-time grid sampling technique') ) default_model_path = os.path.split(os.path.abspath(__file__))[0] + '/../../thirdparty/disk/depth-save.pth' parser.add_argument( '--model_path', type=str, default=default_model_path, help="Path to the model's .pth save file" ) parser.add_argument('--detection-scores', action='store_true') parser.add_argument( '--h5_path', default='./out_disk', help=("Directory where keypoints.h5 and descriptors.h5 will be stored. This" " will be created if it doesn't already exist.") ) parser.add_argument( '--image_path', default='../data/graf', help="Directory with images to be processed." ) args = parser.parse_args() DEV = torch.device('cuda' if torch.cuda.is_available() else 'cpu') CPU = torch.device('cpu') dataset = SceneDataset(args.image_path, crop_size=(args.height, args.width)) state_dict = torch.load(args.model_path, map_location='cpu') # compatibility with older model saves which used the 'extractor' name if 'extractor' in state_dict: weights = state_dict['extractor'] elif 'disk' in state_dict: weights = state_dict['disk'] else: raise KeyError('Incompatible weight file!') model = DISK(window=8, desc_dim=args.desc_dim) model.load_state_dict(weights) model = model.to(DEV) described_samples = extract(dataset, args.h5_path)
9,680
33.575
110
py
pyslam
pyslam-master/test/thirdparty/test_d2net.py
# adapted from https://github.com/mihaidusmanu/d2-net/blob/master/extract_features.py import sys sys.path.append("../../") import config config.cfg.set_lib('d2net') import os import argparse import numpy as np import imageio import cv2 import torch from tqdm import tqdm import scipy import scipy.io import scipy.misc from lib.model_test import D2Net from lib.utils import preprocess_image from lib.pyramid import process_multiscale d2net_base_path='../../thirdparty/d2net/' # CUDA use_cuda = torch.cuda.is_available() device = torch.device("cuda:0" if use_cuda else "cpu") # Argument parsing parser = argparse.ArgumentParser(description='Feature extraction script') # parser.add_argument( # '--image_list_file', type=str, default='../data/kitti06-12-color.png', #required=True, # help='path to a file containing a list of images to process' # ) parser.add_argument( '--image_file', type=str, default='../data/kitti06-12-color.png', #required=True, help='img file to process' ) parser.add_argument( '--preprocessing', type=str, default='caffe', help='image preprocessing (caffe or torch)' ) parser.add_argument( '--model_file', type=str, default=d2net_base_path + 'models/d2_tf.pth', help='path to the full model' ) parser.add_argument( '--max_edge', type=int, default=1600, help='maximum image size at network input' ) parser.add_argument( '--max_sum_edges', type=int, default=2800, help='maximum sum of image sizes at network input' ) parser.add_argument( '--output_extension', type=str, default='.d2-net', help='extension for the output' ) parser.add_argument( '--output_type', type=str, default='npz', help='output file type (npz or mat)' ) parser.add_argument( '--multiscale', dest='multiscale', action='store_true', help='extract multiscale features' ) parser.set_defaults(multiscale=False) parser.add_argument( '--no-relu', dest='use_relu', action='store_false', help='remove ReLU after the dense feature extraction module' ) parser.set_defaults(use_relu=True) args = parser.parse_args() print(args) # Creating CNN model model = D2Net( model_file=args.model_file, use_relu=args.use_relu, use_cuda=use_cuda ) # Process the file # with open(args.image_list_file, 'r') as f: # lines = f.readlines() # for line in tqdm(lines, total=len(lines)): # path = line.strip() if True: #image = imageio.imread(path) path = os.getcwd() #image = imageio.imread(args.image_file) image = cv2.imread(args.image_file) print('image.shape',image.shape) if len(image.shape) == 2: image = image[:, :, np.newaxis] image = np.repeat(image, 3, -1) # TODO: switch to PIL.Image due to deprecation of scipy.misc.imresize. resized_image = image if max(resized_image.shape) > args.max_edge: resized_image = scipy.misc.imresize( resized_image, args.max_edge / max(resized_image.shape) ).astype('float') if sum(resized_image.shape[: 2]) > args.max_sum_edges: resized_image = scipy.misc.imresize( resized_image, args.max_sum_edges / sum(resized_image.shape[: 2]) ).astype('float') fact_i = image.shape[0] / resized_image.shape[0] fact_j = image.shape[1] / resized_image.shape[1] input_image = preprocess_image( resized_image, preprocessing=args.preprocessing ) with torch.no_grad(): if args.multiscale: keypoints, scores, descriptors = process_multiscale( torch.tensor( input_image[np.newaxis, :, :, :].astype(np.float32), device=device ), model ) else: keypoints, scores, descriptors = process_multiscale( torch.tensor( input_image[np.newaxis, :, :, :].astype(np.float32), device=device ), model, scales=[1] ) # Input image coordinates keypoints[:, 0] *= fact_i keypoints[:, 1] *= fact_j # i, j -> u, v keypoints = keypoints[:, [1, 0, 2]] print('#features:',len(keypoints)) print('des[0]',descriptors[0]) if args.output_type == 'npz': with open(path + args.output_extension, 'wb') as output_file: np.savez( output_file, keypoints=keypoints, scores=scores, descriptors=descriptors ) elif args.output_type == 'mat': with open(path + args.output_extension, 'wb') as output_file: scipy.io.savemat( output_file, { 'keypoints': keypoints, 'scores': scores, 'descriptors': descriptors } ) else: raise ValueError('Unknown output type.')
4,953
26.370166
92
py
pyslam
pyslam-master/test/thirdparty/test_keynet.py
# from https://raw.githubusercontent.com/axelBarroso/Key.Net/master/extract_multiscale_features.py import sys sys.path.append("../../") import config config.cfg.set_lib('keynet') import warnings # to disable tensorflow-numpy warnings: from https://github.com/tensorflow/tensorflow/issues/30427 warnings.filterwarnings('ignore', category=FutureWarning) import os, sys, cv2 #sys.path.append(os.path.dirname(os.path.dirname(os.path.realpath(__file__)))) from os import path, mkdir import argparse import keyNet.aux.tools as aux from skimage.transform import pyramid_gaussian import HSequences_bench.tools.geometry_tools as geo_tools import HSequences_bench.tools.repeatability_tools as rep_tools from keyNet.model.keynet_architecture import * import keyNet.aux.desc_aux_function as loss_desc from keyNet.model.hardnet_pytorch import * from keyNet.datasets.dataset_utils import read_bw_image import torch keynet_base_path='../../thirdparty/keynet/' os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' def check_directory(dir): if not path.isdir(dir): mkdir(dir) def create_result_dir(path): directories = path.split('/') tmp = '' for idx, dir in enumerate(directories): tmp += (dir + '/') if idx == len(directories)-1: continue check_directory(tmp) def extract_multiscale_features(): parser = argparse.ArgumentParser(description='HSequences Extract Features') # parser.add_argument('--list-images', type=str, help='File containing the image paths for extracting features.', # required=True) parser.add_argument('--results-dir', type=str, default='extracted_features/', help='The output path to save the extracted keypoint.') parser.add_argument('--network-version', type=str, default='KeyNet_default', help='The Key.Net network version name') parser.add_argument('--checkpoint-det-dir', type=str, default=keynet_base_path + 'keyNet/pretrained_nets/KeyNet_default', help='The path to the checkpoint file to load the detector weights.') parser.add_argument('--pytorch-hardnet-dir', type=str, default=keynet_base_path + 'keyNet/pretrained_nets/HardNet++.pth', help='The path to the checkpoint file to load the HardNet descriptor weights.') # Detector Settings parser.add_argument('--num-filters', type=int, default=8, help='The number of filters in each learnable block.') parser.add_argument('--num-learnable-blocks', type=int, default=3, help='The number of learnable blocks after handcrafted block.') parser.add_argument('--num-levels-within-net', type=int, default=3, help='The number of pyramid levels inside the architecture.') parser.add_argument('--factor-scaling-pyramid', type=float, default=1.2, help='The scale factor between the multi-scale pyramid levels in the architecture.') parser.add_argument('--conv-kernel-size', type=int, default=5, help='The size of the convolutional filters in each of the learnable blocks.') # Multi-Scale Extractor Settings parser.add_argument('--extract-MS', type=bool, default=True, help='Set to True if you want to extract multi-scale features.') parser.add_argument('--num-points', type=int, default=1500, help='The number of desired features to extract.') parser.add_argument('--nms-size', type=int, default=15, help='The NMS size for computing the validation repeatability.') parser.add_argument('--border-size', type=int, default=15, help='The number of pixels to remove from the borders to compute the repeatability.') parser.add_argument('--order-coord', type=str, default='xysr', help='The coordinate order that follows the extracted points. Use yxsr or xysr.') parser.add_argument('--random-seed', type=int, default=12345, help='The random seed value for TensorFlow and Numpy.') parser.add_argument('--pyramid_levels', type=int, default=5, help='The number of downsample levels in the pyramid.') parser.add_argument('--upsampled-levels', type=int, default=1, help='The number of upsample levels in the pyramid.') parser.add_argument('--scale-factor-levels', type=float, default=np.sqrt(2), help='The scale factor between the pyramid levels.') parser.add_argument('--scale-factor', type=float, default=2., help='The scale factor to extract patches before descriptor.') # GPU Settings parser.add_argument('--gpu-memory-fraction', type=float, default=0.3, help='The fraction of GPU used by the script.') parser.add_argument('--gpu-visible-devices', type=str, default="0", help='Set CUDA_VISIBLE_DEVICES variable.') args = parser.parse_known_args()[0] # remove verbose bits from tf os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' tf.logging.set_verbosity(tf.logging.ERROR) # Set CUDA GPU environment os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_visible_devices version_network_name = args.network_version if not args.extract_MS: args.pyramid_levels = 0 args.upsampled_levels = 0 print('Extract features for : ' + version_network_name) aux.check_directory(args.results_dir) aux.check_directory(os.path.join(args.results_dir, version_network_name)) def extract_features(image): pyramid = pyramid_gaussian(image, max_layer=args.pyramid_levels, downscale=args.scale_factor_levels) score_maps = {} for (j, resized) in enumerate(pyramid): im = resized.reshape(1, resized.shape[0], resized.shape[1], 1) feed_dict = { input_network: im, phase_train: False, dimension_image: np.array([1, im.shape[1], im.shape[2]], dtype=np.int32), } im_scores = sess.run(maps, feed_dict=feed_dict) im_scores = geo_tools.remove_borders(im_scores, borders=args.border_size) score_maps['map_' + str(j + 1 + args.upsampled_levels)] = im_scores[0, :, :, 0] if args.upsampled_levels: for j in range(args.upsampled_levels): factor = args.scale_factor_levels ** (args.upsampled_levels - j) up_image = cv2.resize(image, (0, 0), fx=factor, fy=factor) im = np.reshape(up_image, (1, up_image.shape[0], up_image.shape[1], 1)) feed_dict = { input_network: im, phase_train: False, dimension_image: np.array([1, im.shape[1], im.shape[2]], dtype=np.int32), } im_scores = sess.run(maps, feed_dict=feed_dict) im_scores = geo_tools.remove_borders(im_scores, borders=args.border_size) score_maps['map_' + str(j + 1)] = im_scores[0, :, :, 0] im_pts = [] for idx_level in range(levels): scale_value = (args.scale_factor_levels ** (idx_level - args.upsampled_levels)) scale_factor = 1. / scale_value h_scale = np.asarray([[scale_factor, 0., 0.], [0., scale_factor, 0.], [0., 0., 1.]]) h_scale_inv = np.linalg.inv(h_scale) h_scale_inv = h_scale_inv / h_scale_inv[2, 2] num_points_level = point_level[idx_level] if idx_level > 0: res_points = int(np.asarray([point_level[a] for a in range(0, idx_level + 1)]).sum() - len(im_pts)) num_points_level = res_points im_scores = rep_tools.apply_nms(score_maps['map_' + str(idx_level + 1)], args.nms_size) im_pts_tmp = geo_tools.get_point_coordinates(im_scores, num_points=num_points_level, order_coord='xysr') im_pts_tmp = geo_tools.apply_homography_to_points(im_pts_tmp, h_scale_inv) if not idx_level: im_pts = im_pts_tmp else: im_pts = np.concatenate((im_pts, im_pts_tmp), axis=0) if args.order_coord == 'yxsr': im_pts = np.asarray(list(map(lambda x: [x[1], x[0], x[2], x[3]], im_pts))) im_pts = im_pts[(-1 * im_pts[:, 3]).argsort()] im_pts = im_pts[:args.num_points] # Extract descriptor from features descriptors = [] im = image.reshape(1, image.shape[0], image.shape[1], 1) for idx_desc_batch in range(int(len(im_pts) / 250 + 1)): points_batch = im_pts[idx_desc_batch * 250: (idx_desc_batch + 1) * 250] if not len(points_batch): break feed_dict = { input_network: im, phase_train: False, kpts_coord: points_batch[:, :2], kpts_scale: args.scale_factor * points_batch[:, 2], kpts_batch: np.zeros(len(points_batch)), dimension_image: np.array([1, im.shape[1], im.shape[2]], dtype=np.int32), } patch_batch = sess.run(input_patches, feed_dict=feed_dict) patch_batch = np.reshape(patch_batch, (patch_batch.shape[0], 1, 32, 32)) data_a = torch.from_numpy(patch_batch) data_a = data_a.cuda() data_a = Variable(data_a) with torch.no_grad(): out_a = model(data_a) desc_batch = out_a.data.cpu().numpy().reshape(-1, 128) if idx_desc_batch == 0: descriptors = desc_batch else: descriptors = np.concatenate([descriptors, desc_batch], axis=0) return im_pts, descriptors with tf.Graph().as_default(): tf.set_random_seed(args.random_seed) with tf.name_scope('inputs'): # Define the input tensor shape tensor_input_shape = (None, None, None, 1) input_network = tf.placeholder(dtype=tf.float32, shape=tensor_input_shape, name='input_network') dimension_image = tf.placeholder(dtype=tf.int32, shape=(3,), name='dimension_image') kpts_coord = tf.placeholder(dtype=tf.float32, shape=(None, 2), name='kpts_coord') kpts_batch = tf.placeholder(dtype=tf.int32, shape=(None,), name='kpts_batch') kpts_scale = tf.placeholder(dtype=tf.float32, name='kpts_scale') phase_train = tf.placeholder(tf.bool, name='phase_train') with tf.name_scope('model_deep_detector'): deep_architecture = keynet(args) output_network = deep_architecture.model(input_network, phase_train, dimension_image, reuse=False) maps = tf.nn.relu(output_network['output']) # Extract Patches from inputs: input_patches = loss_desc.build_patch_extraction(kpts_coord, kpts_batch, input_network, kpts_scale=kpts_scale) # Define Pytorch HardNet model = HardNet() checkpoint = torch.load(args.pytorch_hardnet_dir) model.load_state_dict(checkpoint['state_dict']) model.eval() model.cuda() # Define variables detect_var = [v for v in tf.trainable_variables(scope='model_deep_detector')] if os.listdir(args.checkpoint_det_dir): init_assign_op_det, init_feed_dict_det = tf_contrib.framework.assign_from_checkpoint( tf.train.latest_checkpoint(args.checkpoint_det_dir), detect_var) point_level = [] tmp = 0.0 factor_points = (args.scale_factor_levels ** 2) levels = args.pyramid_levels + args.upsampled_levels + 1 for idx_level in range(levels): tmp += factor_points ** (-1 * (idx_level - args.upsampled_levels)) point_level.append(args.num_points * factor_points ** (-1 * (idx_level - args.upsampled_levels))) point_level = np.asarray(list(map(lambda x: int(x / tmp), point_level))) # GPU Usage config = tf.ConfigProto() config.gpu_options.per_process_gpu_memory_fraction = args.gpu_memory_fraction config.gpu_options.allow_growth = True with tf.Session(config=config) as sess: sess.run(tf.global_variables_initializer()) if os.listdir(args.checkpoint_det_dir): sess.run(init_assign_op_det, init_feed_dict_det) # # read image and extract keypoints and descriptors # f = open(args.list_images, "r") # for path_to_image in f: # path = path_to_image.split('\n')[0] # if not os.path.exists(path): # print('[ERROR]: File {0} not found!'.format(path)) # return # create_result_dir(os.path.join(args.results_dir, version_network_name, path)) if True: path = 'kitti06-12-color.png' #im = read_bw_image(path) im = cv2.imread('../data/kitti06-12-color.png',cv2.IMREAD_GRAYSCALE) im = im.astype(float) / im.max() im_pts, descriptors = extract_features(im) print('# extracted points:',len(im_pts)) file_name = os.path.join(args.results_dir, version_network_name, path)+'.kpt' np.save(file_name, im_pts) file_name = os.path.join(args.results_dir, version_network_name, path)+'.dsc' np.save(file_name, descriptors) if __name__ == '__main__': extract_multiscale_features()
13,674
40.189759
125
py
pyslam
pyslam-master/test/thirdparty/test_r2d2.py
# Copyright 2019-present NAVER Corp. # CC BY-NC-SA 3.0 # Available only for non-commercial use # from https://raw.githubusercontent.com/naver/r2d2/master/extract.py import sys sys.path.append("../../") import config config.cfg.set_lib('r2d2') import os, pdb from PIL import Image import numpy as np import torch import cv2 from tools import common from tools.dataloader import norm_RGB from nets.patchnet import * r2d2_base_path='../../thirdparty/r2d2' r2d2_default_model_path=r2d2_base_path+'/models/r2d2_WASF_N16.pt' def load_network(model_fn): checkpoint = torch.load(model_fn) print("\n>> Creating net = " + checkpoint['net']) net = eval(checkpoint['net']) nb_of_weights = common.model_size(net) print(f" ( Model size: {nb_of_weights/1000:.0f}K parameters )") # initialization weights = checkpoint['state_dict'] net.load_state_dict({k.replace('module.',''):v for k,v in weights.items()}) return net.eval() class NonMaxSuppression (torch.nn.Module): def __init__(self, rel_thr=0.7, rep_thr=0.7): nn.Module.__init__(self) self.max_filter = torch.nn.MaxPool2d(kernel_size=3, stride=1, padding=1) self.rel_thr = rel_thr self.rep_thr = rep_thr def forward(self, reliability, repeatability, **kw): assert len(reliability) == len(repeatability) == 1 reliability, repeatability = reliability[0], repeatability[0] # local maxima maxima = (repeatability == self.max_filter(repeatability)) # remove low peaks maxima *= (repeatability >= self.rep_thr) maxima *= (reliability >= self.rel_thr) return maxima.nonzero().t()[2:4] def extract_multiscale( net, img, detector, scale_f=2**0.25, min_scale=0.0, max_scale=1, min_size=256, max_size=1024, verbose=False): old_bm = torch.backends.cudnn.benchmark torch.backends.cudnn.benchmark = False # speedup # extract keypoints at multiple scales B, three, H, W = img.shape assert B == 1 and three == 3, "should be a batch with a single RGB image" assert max_scale <= 1 s = 1.0 # current scale factor X,Y,S,C,Q,D = [],[],[],[],[],[] while s+0.001 >= max(min_scale, min_size / max(H,W)): if s-0.001 <= min(max_scale, max_size / max(H,W)): nh, nw = img.shape[2:] if verbose: print(f"extracting at scale x{s:.02f} = {nw:4d}x{nh:3d}") # extract descriptors with torch.no_grad(): res = net(imgs=[img]) # get output and reliability map descriptors = res['descriptors'][0] reliability = res['reliability'][0] repeatability = res['repeatability'][0] # normalize the reliability for nms # extract maxima and descs y,x = detector(**res) # nms c = reliability[0,0,y,x] q = repeatability[0,0,y,x] d = descriptors[0,:,y,x].t() n = d.shape[0] # accumulate multiple scales X.append(x.float() * W/nw) Y.append(y.float() * H/nh) S.append((32/s) * torch.ones(n, dtype=torch.float32, device=d.device)) C.append(c) Q.append(q) D.append(d) s /= scale_f # down-scale the image for next iteration nh, nw = round(H*s), round(W*s) img = F.interpolate(img, (nh,nw), mode='bilinear', align_corners=False) # restore value torch.backends.cudnn.benchmark = old_bm Y = torch.cat(Y) X = torch.cat(X) S = torch.cat(S) # scale scores = torch.cat(C) * torch.cat(Q) # scores = reliability * repeatability XYS = torch.stack([X,Y,S], dim=-1) D = torch.cat(D) return XYS, D, scores def extract_keypoints(args): iscuda = common.torch_set_gpu(args.gpu) # load the network... net = load_network(args.model) if iscuda: net = net.cuda() # create the non-maxima detector detector = NonMaxSuppression( rel_thr = args.reliability_thr, rep_thr = args.repeatability_thr) # while args.images: # img_path = args.images.pop(0) # if img_path.endswith('.txt'): # args.images = open(img_path).read().splitlines() + args.images # continue # print(f"\nExtracting features for {img_path}") # img = Image.open(img_path).convert('RGB') if True: img_path='kitti06-12-color.png' img = cv2.imread('../data/kitti06-12-color.png') #W, H = img.size H, W = img.shape[:2] img = norm_RGB(img)[None] if iscuda: img = img.cuda() # extract keypoints/descriptors for a single image xys, desc, scores = extract_multiscale(net, img, detector, scale_f = args.scale_f, min_scale = args.min_scale, max_scale = args.max_scale, min_size = args.min_size, max_size = args.max_size, verbose = True) xys = xys.cpu().numpy() desc = desc.cpu().numpy() scores = scores.cpu().numpy() idxs = scores.argsort()[-args.top_k or None:] outpath = img_path + '.' + args.tag print(f"Saving {len(idxs)} keypoints to {outpath}") np.savez(open(outpath,'wb'), imsize = (W,H), keypoints = xys[idxs], descriptors = desc[idxs], scores = scores[idxs]) if __name__ == '__main__': import argparse parser = argparse.ArgumentParser("Extract keypoints for a given image") parser.add_argument("--model", type=str, default=r2d2_default_model_path, help='model path') #parser.add_argument("--images", type=str, required=True, nargs='+', help='images / list') parser.add_argument("--tag", type=str, default='r2d2', help='output file tag') parser.add_argument("--top-k", type=int, default=5000, help='number of keypoints') parser.add_argument("--scale-f", type=float, default=2**0.25) parser.add_argument("--min-size", type=int, default=256) parser.add_argument("--max-size", type=int, default=1024) parser.add_argument("--min-scale", type=float, default=0) parser.add_argument("--max-scale", type=float, default=1) parser.add_argument("--reliability-thr", type=float, default=0.7) parser.add_argument("--repeatability-thr", type=float, default=0.7) parser.add_argument("--gpu", type=int, nargs='+', default=[0], help='use -1 for CPU') args = parser.parse_args() extract_keypoints(args)
6,659
32.807107
96
py
pyslam
pyslam-master/test/thirdparty/test_geodesc.py
#!/usr/bin/env python """ Copyright 2018, Zixin Luo, HKUST. Conduct pair-wise image matching. """ # adapted from https://github.com/lzx551402/geodesc/blob/master/examples/image_matching.py import sys sys.path.append("../../") import config config.cfg.set_lib('geodesc') #from __future__ import print_function import os import sys import time from threading import Thread from multiprocessing import Queue import cv2 import numpy as np import warnings # to disable tensorflow-numpy warnings: from https://github.com/tensorflow/tensorflow/issues/30427 warnings.filterwarnings('ignore', category=FutureWarning) if False: import tensorflow as tf else: # from https://stackoverflow.com/questions/56820327/the-name-tf-session-is-deprecated-please-use-tf-compat-v1-session-instead import tensorflow.compat.v1 as tf # from https://kobkrit.com/using-allow-growth-memory-option-in-tensorflow-and-keras-dc8c8081bc96 to cope with the following error: # "[...tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR" config = tf.ConfigProto() config.gpu_options.allow_growth = True #session = tf.Session(config=config, ...) from utils_sys import Printer # CURDIR = os.path.dirname(__file__) # sys.path.append(os.path.abspath(os.path.join(CURDIR, '..'))) # from utils.tf import load_frozen_model # from utils.opencvhelper import SiftWrapper, MatcherWrapper #sys.path.append(os.path.join('third_party', 'geodesc')) #from thirdparty.geodesc.utils.tf import load_frozen_model #from thirdparty.geodesc.utils.opencvhelper import SiftWrapper, MatcherWrapper from geodesc.utils.tf import load_frozen_model from geodesc.utils.opencvhelper import SiftWrapper, MatcherWrapper geodesc_base_path='../../thirdparty/geodesc/' FLAGS = tf.app.flags.FLAGS tf.app.flags.DEFINE_string('model_path', geodesc_base_path + 'model/geodesc.pb', """Path to evaluati3n model.""") tf.app.flags.DEFINE_integer('batch_size', 512, """Inference batch size.""") tf.app.flags.DEFINE_integer('max_kpt_num', 8192, """Maximum number of keypoints. Sampled by octave.""") tf.app.flags.DEFINE_string('img1_path', geodesc_base_path + '/img/test_img1.png', """Path to the first image.""") tf.app.flags.DEFINE_string('img2_path', geodesc_base_path + '/img/test_img2.png', """Path to the second image.""") tf.app.flags.DEFINE_boolean('cf_sift', False, """Compare with SIFT feature.""") # SIFT options tf.app.flags.DEFINE_boolean('pyr_off', False, """Whether to construct image pyramid.""") tf.app.flags.DEFINE_boolean('half_sigma', True, """Whether to halve the sigma value of SIFT when constructing the pyramid.""") tf.app.flags.DEFINE_boolean('ori_off', False, """Whether to use the orientation estimated from SIFT.""") def extract_deep_features(sift_wrapper, sess, img_path, qtz=True): img = cv2.imread(img_path) if img is None: Printer.red('cannot find img: ', img_path) sys.exit(0) gray_img = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY) # detect SIFT keypoints. start = time.time() _, cv_kpts = sift_wrapper.detect(gray_img) end = time.time() print('Time cost in keypoint detection', end - start) start = time.time() sift_wrapper.build_pyramid(gray_img) end = time.time() print('Time cost in scale space construction', end - start) start = time.time() all_patches = sift_wrapper.get_patches(cv_kpts) end = time.time() print('Time cost in patch cropping', end - start) num_patch = all_patches.shape[0] if num_patch % FLAGS.batch_size > 0: loop_num = int(np.floor(float(num_patch) / float(FLAGS.batch_size))) else: loop_num = int(num_patch / FLAGS.batch_size - 1) def _worker(patch_queue, sess, all_feat): """The worker thread.""" while True: patch_data = patch_queue.get() if patch_data is None: return feat = sess.run("squeeze_1:0", feed_dict={"input:0": np.expand_dims(patch_data, -1)}) all_feat.append(feat) #patch_queue.task_done() all_feat = [] patch_queue = Queue() worker_thread = Thread(target=_worker, args=(patch_queue, sess, all_feat)) worker_thread.daemon = True worker_thread.start() start = time.time() # enqueue for i in range(loop_num + 1): if i < loop_num: patch_queue.put(all_patches[i * FLAGS.batch_size: (i + 1) * FLAGS.batch_size]) else: patch_queue.put(all_patches[i * FLAGS.batch_size:]) # poison pill patch_queue.put(None) # wait for extraction. worker_thread.join() end = time.time() print('Time cost in feature extraction', end - start) all_feat = np.concatenate(all_feat, axis=0) # quantize output features. all_feat = (all_feat * 128 + 128).astype(np.uint8) if qtz else all_feat return all_feat, cv_kpts, img def main(argv=None): # pylint: disable=unused-argument """Program entrance.""" # create sift detector. sift_wrapper = SiftWrapper(n_sample=FLAGS.max_kpt_num) sift_wrapper.half_sigma = FLAGS.half_sigma sift_wrapper.pyr_off = FLAGS.pyr_off sift_wrapper.ori_off = FLAGS.ori_off sift_wrapper.create() # create deep feature extractor. Printer.yellow('loading model:',FLAGS.model_path,'...') graph = load_frozen_model(FLAGS.model_path, print_nodes=False) #sess = tf.Session(graph=graph) Printer.yellow('...done') with tf.Session(graph=graph, config=config) as sess: # extract deep feature from images. deep_feat1, cv_kpts1, img1 = extract_deep_features( sift_wrapper, sess, FLAGS.img1_path, qtz=False) deep_feat2, cv_kpts2, img2 = extract_deep_features( sift_wrapper, sess, FLAGS.img2_path, qtz=False) # match features by OpenCV brute-force matcher (CPU). matcher_wrapper = MatcherWrapper() # the ratio criterion is set to 0.89 for GeoDesc as described in the paper. deep_good_matches, deep_mask = matcher_wrapper.get_matches( deep_feat1, deep_feat2, cv_kpts1, cv_kpts2, ratio=0.89, cross_check=True, info='deep') deep_display = matcher_wrapper.draw_matches( img1, cv_kpts1, img2, cv_kpts2, deep_good_matches, deep_mask) # compare with SIFT. if FLAGS.cf_sift: sift_feat1 = sift_wrapper.compute(img1, cv_kpts1) sift_feat2 = sift_wrapper.compute(img2, cv_kpts2) sift_good_matches, sift_mask = matcher_wrapper.get_matches( sift_feat1, sift_feat2, cv_kpts1, cv_kpts2, ratio=0.80, cross_check=True, info='sift') sift_display = matcher_wrapper.draw_matches( img1, cv_kpts1, img2, cv_kpts2, sift_good_matches, sift_mask) display = np.concatenate((sift_display, deep_display), axis=0) else: display = deep_display cv2.imshow('display', display) cv2.waitKey() sess.close() if __name__ == '__main__': tf.app.run()
7,192
35.328283
130
py
pyslam
pyslam-master/test/thirdparty/test_contextdesc.py
#!/usr/bin/env python3 # adpated from https://github.com/lzx551402/contextdesc/blob/master/image_matching.py """ Copyright 2019, Zixin Luo, HKUST. Image matching example. """ import sys sys.path.append("../../") import config config.cfg.set_lib('contextdesc',prepend=True) contextdesc_base_path='../../thirdparty/contextdesc/' import warnings # to disable tensorflow-numpy warnings: from https://github.com/tensorflow/tensorflow/issues/30427 warnings.filterwarnings('ignore', category=FutureWarning) import os import cv2 import numpy as np if False: import tensorflow as tf else: # from https://stackoverflow.com/questions/56820327/the-name-tf-session-is-deprecated-please-use-tf-compat-v1-session-instead import tensorflow.compat.v1 as tf import matplotlib.pyplot as plt from contextdesc.utils.opencvhelper import MatcherWrapper #from models import get_model from contextdesc.models.reg_model import RegModel from contextdesc.models.loc_model import LocModel from contextdesc.models.aug_model import AugModel # from https://kobkrit.com/using-allow-growth-memory-option-in-tensorflow-and-keras-dc8c8081bc96 to cope with the following error: # "[...tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR" tf_config = tf.ConfigProto() tf_config.gpu_options.allow_growth = True tf_config.gpu_options.per_process_gpu_memory_fraction=0.333 # from https://stackoverflow.com/questions/34199233/how-to-prevent-tensorflow-from-allocating-the-totality-of-a-gpu-memory #session = tf.Session(config=tf_config, ...) FLAGS = tf.app.flags.FLAGS tf.app.flags.DEFINE_string('loc_model', contextdesc_base_path + 'pretrained/contextdesc++', """Path to the local feature and augmentation model.""") tf.app.flags.DEFINE_string('reg_model', contextdesc_base_path + 'pretrained/retrieval_model', """Path to the regional feature model.""") tf.app.flags.DEFINE_string('img1_path', contextdesc_base_path+ 'imgs/test_img1.jpg', """Path to the first image.""") tf.app.flags.DEFINE_string('img2_path', contextdesc_base_path+ 'imgs/test_img2.jpg', """Path to the second image.""") tf.app.flags.DEFINE_integer('n_sample', 2048, """Maximum number of keypoints. Sampled by octave.""") # model options tf.app.flags.DEFINE_string('model_type', 'pb', """Pre-trained model type.""") tf.app.flags.DEFINE_boolean('dense_desc', False, """Whether to use dense descriptor model.""") # matching options tf.app.flags.DEFINE_boolean('ratio_test', False, """Whether to apply ratio test in matching.""") tf.app.flags.DEFINE_boolean('cross_check', False, """Whether to apply cross check in matching.""") def load_imgs(img_paths): rgb_list = [] gray_list = [] for img_path in img_paths: img = cv2.imread(img_path) if img is None: raise IOError('No image found in folder: ', img_path) gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)[..., np.newaxis] img = img[..., ::-1] rgb_list.append(img) gray_list.append(gray) return rgb_list, gray_list def extract_regional_features(rgb_list, model_path): reg_feat_list = [] model = RegModel(model_path) #get_model('reg_model')(model_path) for _, val in enumerate(rgb_list): reg_feat = model.run_test_data(val) reg_feat_list.append(reg_feat) model.close() return reg_feat_list def extract_local_features(gray_list, model_path): cv_kpts_list = [] loc_info_list = [] loc_feat_list = [] sift_feat_list = [] #model = get_model('loc_model')(model_path, **{'sift_desc': True, model = LocModel(model_path, **{'sift_desc': True, 'n_feature': 2000, 'n_sample': FLAGS.n_sample, 'peak_thld': 0.04, 'dense_desc': FLAGS.dense_desc, 'upright': False}) for _, val in enumerate(gray_list): loc_feat, kpt_mb, normalized_xy, cv_kpts, sift_desc = model.run_test_data(val) raw_kpts = [np.array((i.pt[0], i.pt[1], i.size, i.angle, i.response)) for i in cv_kpts] raw_kpts = np.stack(raw_kpts, axis=0) loc_info = np.concatenate((raw_kpts, normalized_xy, loc_feat, kpt_mb), axis=-1) cv_kpts_list.append(cv_kpts) loc_info_list.append(loc_info) sift_feat_list.append(sift_desc) loc_feat_list.append(loc_feat / np.linalg.norm(loc_feat, axis=-1, keepdims=True)) model.close() return cv_kpts_list, loc_info_list, loc_feat_list, sift_feat_list def extract_augmented_features(reg_feat_list, loc_info_list, model_path): aug_feat_list = [] #model = get_model('aug_model')(model_path, **{'quantz': False}) model = AugModel(model_path, **{'quantz': False}) assert len(reg_feat_list) == len(loc_info_list) for idx, _ in enumerate(reg_feat_list): aug_feat, _ = model.run_test_data([reg_feat_list[idx], loc_info_list[idx]]) aug_feat_list.append(aug_feat) model.close() return aug_feat_list def main(argv=None): # pylint: disable=unused-argument """Program entrance.""" if FLAGS.model_type == 'pb': reg_model_path = os.path.join(FLAGS.reg_model, 'reg.pb') loc_model_path = os.path.join(FLAGS.loc_model, 'loc.pb') aug_model_path = os.path.join(FLAGS.loc_model, 'aug.pb') elif FLAGS.model_type == 'ckpt': reg_model_path = os.path.join(FLAGS.reg_model, 'model.ckpt-550000') loc_model_path = os.path.join(FLAGS.loc_model, 'model.ckpt-400000') aug_model_path = os.path.join(FLAGS.loc_model, 'model.ckpt-400000') else: raise NotImplementedError img_paths = [FLAGS.img1_path, FLAGS.img2_path] # load testing images. rgb_list, gray_list = load_imgs(img_paths) # extract regional features. reg_feat_list = extract_regional_features(rgb_list, reg_model_path) # extract local features and keypoint matchability. cv_kpts_list, loc_info_list, loc_feat_list, sift_feat_list = extract_local_features( gray_list, loc_model_path) # extract augmented features. aug_feat_list = extract_augmented_features(reg_feat_list, loc_info_list, aug_model_path) # feature matching and draw matches. matcher = MatcherWrapper() sift_match, sift_mask = matcher.get_matches( sift_feat_list[0], sift_feat_list[1], cv_kpts_list[0], cv_kpts_list[1], ratio=0.8 if FLAGS.ratio_test else None, cross_check=FLAGS.cross_check, err_thld=3, ransac=True, info='SIFT feautre') base_match, base_mask = matcher.get_matches( loc_feat_list[0], loc_feat_list[1], cv_kpts_list[0], cv_kpts_list[1], ratio=0.89 if FLAGS.ratio_test else None, cross_check=FLAGS.cross_check, err_thld=3, ransac=True, info='Raw local feature') aug_match, aug_mask = matcher.get_matches( aug_feat_list[0], aug_feat_list[1], cv_kpts_list[0], cv_kpts_list[1], ratio=0.89 if FLAGS.ratio_test else None, cross_check=FLAGS.cross_check, err_thld=3, ransac=True, info='Augmented local feature') sift_disp = matcher.draw_matches( rgb_list[0], cv_kpts_list[0], rgb_list[1], cv_kpts_list[1], sift_match, sift_mask) base_disp = matcher.draw_matches( rgb_list[0], cv_kpts_list[0], rgb_list[1], cv_kpts_list[1], base_match, base_mask) aug_disp = matcher.draw_matches( rgb_list[0], cv_kpts_list[0], rgb_list[1], cv_kpts_list[1], aug_match, aug_mask) rows, cols = sift_disp.shape[0:2] white = (np.ones((int(rows / 50), cols, 3)) * 255).astype(np.uint8) disp = np.concatenate([sift_disp, white, base_disp, white, aug_disp], axis=0) plt.xticks([]) plt.yticks([]) plt.imshow(disp) plt.show() if __name__ == '__main__': tf.app.run()
8,106
41.668421
183
py
pyslam
pyslam-master/test/thirdparty/test_sosnet.py
import sys sys.path.append("../../") import config config.cfg.set_lib('sosnet') import torch import sosnet_model import os tfeat_base_path='../../thirdparty/SOSNet/' device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") torch.set_grad_enabled(False) sosnet32 = sosnet_model.SOSNet32x32() net_name = 'liberty' sosnet32.load_state_dict(torch.load(os.path.join(tfeat_base_path, 'sosnet-weights', "sosnet-32x32-" + net_name + ".pth"))) sosnet32.cuda().eval() patches = torch.rand(100, 1, 32, 32).to(device) descrs = sosnet32(patches) print('done!')
570
23.826087
122
py
GANF
GANF-main/train_traffic.py
#%% import os import argparse import torch from models.GANF import GANF import numpy as np parser = argparse.ArgumentParser() # files parser.add_argument('--data_dir', type=str, default='./data', help='Location of datasets.') parser.add_argument('--output_dir', type=str, default='./checkpoint/model') parser.add_argument('--name',default='traffic') parser.add_argument('--dataset', type=str, default='metr-la') # restore parser.add_argument('--graph', type=str, default='None') parser.add_argument('--model', type=str, default='None') parser.add_argument('--seed', type=int, default=10, help='Random seed to use.') # model parameters parser.add_argument('--n_blocks', type=int, default=6, help='Number of blocks to stack in a model (MADE in MAF; Coupling+BN in RealNVP).') parser.add_argument('--n_components', type=int, default=1, help='Number of Gaussian clusters for mixture of gaussians models.') parser.add_argument('--hidden_size', type=int, default=32, help='Hidden layer size for MADE (and each MADE block in an MAF).') parser.add_argument('--n_hidden', type=int, default=1, help='Number of hidden layers in each MADE.') parser.add_argument('--batch_norm', type=bool, default=False) # training params parser.add_argument('--batch_size', type=int, default=64) parser.add_argument('--weight_decay', type=float, default=5e-4) parser.add_argument('--n_epochs', type=int, default=20) parser.add_argument('--lr', type=float, default=1e-3, help='Learning rate.') parser.add_argument('--log_interval', type=int, default=5, help='How often to show loss statistics and save samples.') parser.add_argument('--h_tol', type=float, default=1e-6) parser.add_argument('--rho_max', type=float, default=1e16) parser.add_argument('--max_iter', type=int, default=20) parser.add_argument('--lambda1', type=float, default=0.0) parser.add_argument('--rho_init', type=float, default=1.0) parser.add_argument('--alpha_init', type=float, default=0.0) args = parser.parse_known_args()[0] args.cuda = torch.cuda.is_available() device = torch.device("cuda" if args.cuda else "cpu") print(args) import random import numpy as np random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if args.cuda: torch.cuda.manual_seed(args.seed) #%% print("Loading dataset") from dataset import load_traffic train_loader, val_loader, test_loader, n_sensor = load_traffic("{}/{}.h5".format(args.data_dir,args.dataset), \ args.batch_size) #%% rho = args.rho_init alpha = args.alpha_init lambda1 = args.lambda1 h_A_old = np.inf max_iter = args.max_iter rho_max = args.rho_max h_tol = args.h_tol epoch = 0 # initialize A if args.graph != 'None': init = torch.load(args.graph).to(device).abs() print("Load graph from "+args.graph) else: from torch.nn.init import xavier_uniform_ init = torch.zeros([n_sensor, n_sensor]) init = xavier_uniform_(init).abs() init = init.fill_diagonal_(0.0) A = torch.tensor(init, requires_grad=True, device=device) #%% model = GANF(args.n_blocks, 1, args.hidden_size, args.n_hidden, dropout=0.0, batch_norm=args.batch_norm) model = model.to(device) if args.model != 'None': model.load_state_dict(torch.load(args.model)) print('Load model from '+args.model) #%% from torch.nn.utils import clip_grad_value_ save_path = os.path.join(args.output_dir,args.name) if not os.path.exists(save_path): os.makedirs(save_path) loss_best = 100 for _ in range(max_iter): while rho < rho_max: lr = args.lr #* np.math.pow(0.1, epoch // 100) optimizer = torch.optim.Adam([ {'params':model.parameters(), 'weight_decay':args.weight_decay}, {'params': [A]}], lr=lr, weight_decay=0.0) # train for _ in range(args.n_epochs): # train loss_train = [] epoch += 1 model.train() for x in train_loader: x = x.to(device) optimizer.zero_grad() A_hat = torch.divide(A.T,A.sum(dim=1).detach()).T loss = -model(x, A_hat) h = torch.trace(torch.matrix_exp(A_hat*A_hat)) - n_sensor total_loss = loss + 0.5 * rho * h * h + alpha * h total_loss.backward() clip_grad_value_(model.parameters(), 1) optimizer.step() loss_train.append(loss.item()) A.data.copy_(torch.clamp(A.data, min=0, max=1)) # evaluate model.eval() loss_val = [] with torch.no_grad(): for x in val_loader: x = x.to(device) loss = -model(x,A_hat.data) loss_val.append(loss.item()) print('Epoch: {}, train -log_prob: {:.2f}, test -log_prob: {:.2f}, h: {}'\ .format(epoch, np.mean(loss_train), np.mean(loss_val), h.item())) if np.mean(loss_val) < loss_best: loss_best = np.mean(loss_val) print("save model {} epoch".format(epoch)) torch.save(A.data,os.path.join(save_path, "graph_best.pt")) torch.save(model.state_dict(), os.path.join(save_path, "{}_best.pt".format(args.name))) print('rho: {}, alpha {}, h {}'.format(rho, alpha, h.item())) print('===========================================') torch.save(A.data,os.path.join(save_path, "graph_{}.pt".format(epoch))) torch.save(model.state_dict(), os.path.join(save_path, "{}_{}.pt".format(args.name, epoch))) del optimizer torch.cuda.empty_cache() if h.item() > 0.5 * h_A_old: rho *= 10 else: break h_A_old = h.item() alpha += rho*h.item() if h_A_old <= h_tol or rho >=rho_max: break # %% lr = args.lr * 0.1 optimizer = torch.optim.Adam([ {'params':model.parameters(), 'weight_decay':args.weight_decay}, {'params': [A]}], lr=lr, weight_decay=0.0) # train for _ in range(100): loss_train = [] epoch += 1 model.train() for x in train_loader: x = x.to(device) optimizer.zero_grad() A_hat = torch.divide(A.T,A.sum(dim=1).detach()).T loss = -model(x, A_hat) h = torch.trace(torch.matrix_exp(A_hat*A_hat)) - n_sensor total_loss = loss + 0.5 * rho * h * h + alpha * h total_loss.backward() clip_grad_value_(model.parameters(), 1) optimizer.step() loss_train.append(loss.item()) A.data.copy_(torch.clamp(A.data, min=0, max=1)) model.eval() loss_val = [] print(A.max()) with torch.no_grad(): for x in val_loader: x = x.to(device) loss = -model(x,A_hat.data) loss_val.append(loss.item()) print('Epoch: {}, train -log_prob: {:.2f}, test -log_prob: {:.2f}, h: {}'\ .format(epoch, np.mean(loss_train), np.mean(loss_val), h.item())) if np.mean(loss_val) < loss_best: loss_best = np.mean(loss_val) print("save model {} epoch".format(epoch)) torch.save(A.data,os.path.join(save_path, "graph_best.pt")) torch.save(model.state_dict(), os.path.join(save_path, "{}_best.pt".format(args.name))) if epoch % args.log_interval==0: torch.save(A.data,os.path.join(save_path, "graph_{}.pt".format(epoch))) torch.save(model.state_dict(), os.path.join(save_path, "{}_{}.pt".format(args.name, epoch))) #%%
7,602
34.036866
138
py
GANF
GANF-main/utils.py
#%% import torch def h(A): return torch.trace(torch.matrix_exp(A*A)) - A.shape[0] def normalize(A): D = A.sum(dim=0) D_inv = D.pow_(-1) D_inv.masked_fill_(D_inv == float('inf'), 0) return A * D_inv def thresholding(A, thre): return torch.where(A.abs()>thre, A, torch.scalar_tensor(0.0, dtype=torch.float32, device=A.device)) def binarize(A, thre): return torch.where(A.abs()>thre, 1.0, 0.0) # %% import pandas as pd def get_timestamp(stamps): return (stamps - pd.Timestamp("1970-01-01")) // pd.Timedelta("1s") # %% import numpy as np from sklearn.metrics import auc def roc_auc(label_time, pred, negative_sample, sigma): negative_sample = np.sort(negative_sample)[::-1] thresholds = list(negative_sample[::int(len(negative_sample)/50)]) thresholds.append(negative_sample[-1]) tps=[] fps=[] for thre in thresholds: pred_pos = pred[pred>thre] tp = 0 for i in range(len(label_time)): start_time = label_time[i] - pd.Timedelta(30, unit='min') end_time = label_time[i] + pd.Timedelta(30, unit='min') detected_event = pred_pos[str(start_time): str(end_time)] if len(detected_event)>0: timestamps = get_timestamp(detected_event.index) delta_t = np.min(np.abs(timestamps.values - get_timestamp(label_time[i]))) tp += np.exp(-np.power(delta_t/sigma,2)) tp = tp/len(label_time) tps.append(tp) fp = (negative_sample>thre).sum()/len(negative_sample) fps.append(fp) return auc(fps,tps), (fps,tps) # %% def roc_auc_all(loss_np, delta_t, sigma): ground_truth = np.exp(-np.power((delta_t.values)/sigma,2)) loss_sort = np.sort(loss_np)[::-1] thresholds = list(loss_sort[::int(len(loss_sort)/50)]) thresholds.append(loss_sort[-1]) n_pos = ground_truth.sum() n_neg = (1-ground_truth).sum() tps = [] fps = [] for thre in thresholds: pred_pos = loss_np>thre tp = ground_truth[pred_pos].sum()/n_pos fp = (1-ground_truth[pred_pos]).sum()/n_neg tps.append(tp) fps.append(fp) auc_score = auc(fps, tps) return auc_score, fps, tps
2,221
29.027027
103
py
GANF
GANF-main/dataset.py
#%% import pandas as pd import torch from torch.utils.data import Dataset import numpy as np # %% from torch.utils.data import DataLoader def load_traffic(root, batch_size): """ Load traffic dataset return train_loader, val_loader, test_loader """ df = pd.read_hdf(root) df = df.reset_index() df = df.rename(columns={"index":"utc"}) df["utc"] = pd.to_datetime(df["utc"], unit="s") df = df.set_index("utc") n_sensor = len(df.columns) mean = df.values.flatten().mean() std = df.values.flatten().std() df = (df - mean)/std df = df.sort_index() # split the dataset train_df = df.iloc[:int(0.75*len(df))] val_df = df.iloc[int(0.75*len(df)):int(0.875*len(df))] test_df = df.iloc[int(0.75*len(df)):] train_loader = DataLoader(Traffic(train_df), batch_size=batch_size, shuffle=True) val_loader = DataLoader(Traffic(val_df), batch_size=batch_size, shuffle=False) test_loader = DataLoader(Traffic(test_df), batch_size=batch_size, shuffle=False) return train_loader, val_loader, test_loader, n_sensor class Traffic(Dataset): def __init__(self, df, window_size=12, stride_size=1): super(Traffic, self).__init__() self.df = df self.window_size = window_size self.stride_size = stride_size self.data, self.idx, self.time = self.preprocess(df) def preprocess(self, df): start_idx = np.arange(0,len(df)-self.window_size,self.stride_size) end_idx = np.arange(self.window_size, len(df), self.stride_size) delat_time = df.index[end_idx]-df.index[start_idx] idx_mask = delat_time==pd.Timedelta(5*self.window_size,unit='min') return df.values, start_idx[idx_mask], df.index[start_idx[idx_mask]] def __len__(self): length = len(self.idx) return length def __getitem__(self, index): # N X K X L X D start = self.idx[index] end = start + self.window_size data = self.data[start:end].reshape([self.window_size,-1, 1]) return torch.FloatTensor(data).transpose(0,1) def load_water(root, batch_size,label=False): data = pd.read_csv(root) data = data.rename(columns={"Normal/Attack":"label"}) data.label[data.label!="Normal"]=1 data.label[data.label=="Normal"]=0 data["Timestamp"] = pd.to_datetime(data["Timestamp"]) data = data.set_index("Timestamp") #%% feature = data.iloc[:,:51] mean_df = feature.mean(axis=0) std_df = feature.std(axis=0) norm_feature = (feature-mean_df)/std_df norm_feature = norm_feature.dropna(axis=1) n_sensor = len(norm_feature.columns) train_df = norm_feature.iloc[:int(0.6*len(data))] train_label = data.label.iloc[:int(0.6*len(data))] val_df = norm_feature.iloc[int(0.6*len(data)):int(0.8*len(data))] val_label = data.label.iloc[int(0.6*len(data)):int(0.8*len(data))] test_df = norm_feature.iloc[int(0.8*len(data)):] test_label = data.label.iloc[int(0.8*len(data)):] if label: train_loader = DataLoader(WaterLabel(train_df,train_label), batch_size=batch_size, shuffle=True) else: train_loader = DataLoader(Water(train_df,train_label), batch_size=batch_size, shuffle=True) val_loader = DataLoader(Water(val_df,val_label), batch_size=batch_size, shuffle=False) test_loader = DataLoader(Water(test_df,test_label), batch_size=batch_size, shuffle=False) return train_loader, val_loader, test_loader, n_sensor class Water(Dataset): def __init__(self, df, label, window_size=60, stride_size=10): super(Water, self).__init__() self.df = df self.window_size = window_size self.stride_size = stride_size self.data, self.idx, self.label = self.preprocess(df,label) def preprocess(self, df, label): start_idx = np.arange(0,len(df)-self.window_size,self.stride_size) end_idx = np.arange(self.window_size, len(df), self.stride_size) delat_time = df.index[end_idx]-df.index[start_idx] idx_mask = delat_time==pd.Timedelta(self.window_size,unit='s') return df.values, start_idx[idx_mask], label[start_idx[idx_mask]] def __len__(self): length = len(self.idx) return length def __getitem__(self, index): # N X K X L X D start = self.idx[index] end = start + self.window_size data = self.data[start:end].reshape([self.window_size,-1, 1]) return torch.FloatTensor(data).transpose(0,1) class WaterLabel(Dataset): def __init__(self, df, label, window_size=60, stride_size=10): super(WaterLabel, self).__init__() self.df = df self.window_size = window_size self.stride_size = stride_size self.data, self.idx, self.label = self.preprocess(df,label) self.label = 1.0-2*self.label def preprocess(self, df, label): start_idx = np.arange(0,len(df)-self.window_size,self.stride_size) end_idx = np.arange(self.window_size, len(df), self.stride_size) delat_time = df.index[end_idx]-df.index[start_idx] idx_mask = delat_time==pd.Timedelta(self.window_size,unit='s') return df.values, start_idx[idx_mask], label[start_idx[idx_mask]] def __len__(self): length = len(self.idx) return length def __getitem__(self, index): # N X K X L X D start = self.idx[index] end = start + self.window_size data = self.data[start:end].reshape([self.window_size,-1, 1]) return torch.FloatTensor(data).transpose(0,1),self.label[index]
5,624
31.894737
104
py
GANF
GANF-main/train_water.py
#%% import os import argparse import torch from models.GANF import GANF import numpy as np from sklearn.metrics import roc_auc_score # from data import fetch_dataloaders parser = argparse.ArgumentParser() # files parser.add_argument('--data_dir', type=str, default='./data/SWaT_Dataset_Attack_v0.csv', help='Location of datasets.') parser.add_argument('--output_dir', type=str, default='./checkpoint/model') parser.add_argument('--name',default='GANF_Water') # restore parser.add_argument('--graph', type=str, default='None') parser.add_argument('--model', type=str, default='None') parser.add_argument('--seed', type=int, default=18, help='Random seed to use.') # made parameters parser.add_argument('--n_blocks', type=int, default=1, help='Number of blocks to stack in a model (MADE in MAF; Coupling+BN in RealNVP).') parser.add_argument('--n_components', type=int, default=1, help='Number of Gaussian clusters for mixture of gaussians models.') parser.add_argument('--hidden_size', type=int, default=32, help='Hidden layer size for MADE (and each MADE block in an MAF).') parser.add_argument('--n_hidden', type=int, default=1, help='Number of hidden layers in each MADE.') parser.add_argument('--batch_norm', type=bool, default=False) # training params parser.add_argument('--batch_size', type=int, default=512) parser.add_argument('--weight_decay', type=float, default=5e-4) parser.add_argument('--n_epochs', type=int, default=1) parser.add_argument('--lr', type=float, default=2e-3, help='Learning rate.') parser.add_argument('--log_interval', type=int, default=5, help='How often to show loss statistics and save samples.') parser.add_argument('--h_tol', type=float, default=1e-4) parser.add_argument('--rho_max', type=float, default=1e16) parser.add_argument('--max_iter', type=int, default=20) parser.add_argument('--lambda1', type=float, default=0.0) parser.add_argument('--rho_init', type=float, default=1.0) parser.add_argument('--alpha_init', type=float, default=0.0) args = parser.parse_known_args()[0] args.cuda = torch.cuda.is_available() device = torch.device("cuda" if args.cuda else "cpu") print(args) import random import numpy as np random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if args.cuda: torch.cuda.manual_seed(args.seed) #%% print("Loading dataset") from dataset import load_water train_loader, val_loader, test_loader, n_sensor = load_water(args.data_dir, \ args.batch_size) #%% rho = args.rho_init alpha = args.alpha_init lambda1 = args.lambda1 h_A_old = np.inf max_iter = args.max_iter rho_max = args.rho_max h_tol = args.h_tol epoch = 0 # initialize A if args.graph != 'None': init = torch.load(args.graph).to(device).abs() print("Load graph from "+args.graph) else: from torch.nn.init import xavier_uniform_ init = torch.zeros([n_sensor, n_sensor]) init = xavier_uniform_(init).abs() init = init.fill_diagonal_(0.0) A = torch.tensor(init, requires_grad=True, device=device) #%% model = GANF(args.n_blocks, 1, args.hidden_size, args.n_hidden, dropout=0.0, batch_norm=args.batch_norm) model = model.to(device) if args.model != 'None': model.load_state_dict(torch.load(args.model)) print('Load model from '+args.model) #%% from torch.nn.utils import clip_grad_value_ import seaborn as sns import matplotlib.pyplot as plt save_path = os.path.join(args.output_dir,args.name) if not os.path.exists(save_path): os.makedirs(save_path) loss_best = 100 for _ in range(max_iter): while rho < rho_max: lr = args.lr optimizer = torch.optim.Adam([ {'params':model.parameters(), 'weight_decay':args.weight_decay}, {'params': [A]}], lr=lr, weight_decay=0.0) for _ in range(args.n_epochs): # train iteration loss_train = [] epoch += 1 model.train() for x in train_loader: x = x.to(device) optimizer.zero_grad() loss = -model(x, A) h = torch.trace(torch.matrix_exp( A* A)) - n_sensor total_loss = loss + 0.5 * rho * h * h + alpha * h total_loss.backward() clip_grad_value_(model.parameters(), 1) optimizer.step() loss_train.append(loss.item()) A.data.copy_(torch.clamp(A.data, min=0, max=1)) # evaluate iteration model.eval() loss_val = [] with torch.no_grad(): for x in val_loader: x = x.to(device) loss = -model.test(x, A.data).cpu().numpy() loss_val.append(loss) loss_val = np.concatenate(loss_val) loss_test = [] with torch.no_grad(): for x in test_loader: x = x.to(device) loss = -model.test(x, A.data).cpu().numpy() loss_test.append(loss) loss_test = np.concatenate(loss_test) print(loss_val.max(), loss_val.min(), loss_test.max(), loss_test.min()) loss_val = np.nan_to_num(loss_val) loss_test = np.nan_to_num(loss_test) roc_val = roc_auc_score(np.asarray(val_loader.dataset.label.values,dtype=int),loss_val) roc_test = roc_auc_score(np.asarray(test_loader.dataset.label.values,dtype=int),loss_test) print('Epoch: {}, train -log_prob: {:.2f}, test -log_prob: {:.2f}, roc_val: {:.4f}, roc_test: {:.4f} ,h: {}'\ .format(epoch, np.mean(loss_train), np.mean(loss_val), roc_val, roc_test, h.item())) print('rho: {}, alpha {}, h {}'.format(rho, alpha, h.item())) print('===========================================') torch.save(A.data,os.path.join(save_path, "graph_{}.pt".format(epoch))) torch.save(model.state_dict(), os.path.join(save_path, "{}_{}.pt".format(args.name, epoch))) del optimizer torch.cuda.empty_cache() if h.item() > 0.5 * h_A_old: rho *= 10 else: break h_A_old = h.item() alpha += rho*h.item() if h_A_old <= h_tol or rho >=rho_max: break # %% lr = args.lr optimizer = torch.optim.Adam([ {'params':model.parameters(), 'weight_decay':args.weight_decay}, {'params': [A]}], lr=lr, weight_decay=0.0) for _ in range(30): loss_train = [] epoch += 1 model.train() for x in train_loader: x = x.to(device) optimizer.zero_grad() loss = -model(x, A) h = torch.trace(torch.matrix_exp(A*A)) - n_sensor total_loss = loss + 0.5 * rho * h * h + alpha * h total_loss.backward() clip_grad_value_(model.parameters(), 1) optimizer.step() loss_train.append(loss.item()) A.data.copy_(torch.clamp(A.data, min=0, max=1)) # eval model.eval() loss_val = [] with torch.no_grad(): for x in val_loader: x = x.to(device) loss = -model.test(x, A.data).cpu().numpy() loss_val.append(loss) loss_val = np.concatenate(loss_val) loss_test = [] with torch.no_grad(): for x in test_loader: x = x.to(device) loss = -model.test(x, A.data).cpu().numpy() loss_test.append(loss) loss_test = np.concatenate(loss_test) loss_val = np.nan_to_num(loss_val) loss_test = np.nan_to_num(loss_test) roc_val = roc_auc_score(np.asarray(val_loader.dataset.label.values,dtype=int),loss_val) roc_test = roc_auc_score(np.asarray(test_loader.dataset.label.values,dtype=int),loss_test) print('Epoch: {}, train -log_prob: {:.2f}, test -log_prob: {:.2f}, roc_val: {:.4f}, roc_test: {:.4f} ,h: {}'\ .format(epoch, np.mean(loss_train), np.mean(loss_val), roc_val, roc_test, h.item())) if np.mean(loss_val) < loss_best: loss_best = np.mean(loss_val) print("save model {} epoch".format(epoch)) torch.save(A.data,os.path.join(save_path, "graph_best.pt")) torch.save(model.state_dict(), os.path.join(save_path, "{}_best.pt".format(args.name))) if epoch % args.log_interval==0: torch.save(A.data,os.path.join(save_path, "graph_{}.pt".format(epoch))) torch.save(model.state_dict(), os.path.join(save_path, "{}_{}.pt".format(args.name, epoch)))
8,511
34.319502
138
py
GANF
GANF-main/eval_water.py
#%% import os import argparse import torch from models.GANF import GANF import numpy as np from sklearn.metrics import roc_auc_score # from data import fetch_dataloaders parser = argparse.ArgumentParser() # files parser.add_argument('--data_dir', type=str, default='./data/SWaT_Dataset_Attack_v0.csv', help='Location of datasets.') parser.add_argument('--output_dir', type=str, default='/home/enyandai/code/checkpoint/model') parser.add_argument('--name',default='GANF_Water') # restore parser.add_argument('--graph', type=str, default='None') parser.add_argument('--model', type=str, default='None') parser.add_argument('--seed', type=int, default=10, help='Random seed to use.') # made parameters parser.add_argument('--n_blocks', type=int, default=1, help='Number of blocks to stack in a model (MADE in MAF; Coupling+BN in RealNVP).') parser.add_argument('--n_components', type=int, default=1, help='Number of Gaussian clusters for mixture of gaussians models.') parser.add_argument('--hidden_size', type=int, default=32, help='Hidden layer size for MADE (and each MADE block in an MAF).') parser.add_argument('--n_hidden', type=int, default=1, help='Number of hidden layers in each MADE.') parser.add_argument('--batch_norm', type=bool, default=False) # training params parser.add_argument('--batch_size', type=int, default=512) args = parser.parse_known_args()[0] args.cuda = torch.cuda.is_available() device = torch.device("cuda" if args.cuda else "cpu") print(args) import random import numpy as np random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if args.cuda: torch.cuda.manual_seed(args.seed) #%% print("Loading dataset") from dataset import load_water train_loader, val_loader, test_loader, n_sensor = load_water(args.data_dir, \ args.batch_size) #%% model = GANF(args.n_blocks, 1, args.hidden_size, args.n_hidden, dropout=0.0, batch_norm=args.batch_norm) model = model.to(device) model.load_state_dict(torch.load("./checkpoint/eval/water/GANF_water_seed_18_best.pt")) A = torch.load("./checkpoint/eval/GANF_water_seed_18/graph_best.pt").to(device) model.eval() #%% loss_test = [] with torch.no_grad(): for x in test_loader: x = x.to(device) loss = -model.test(x, A.data).cpu().numpy() loss_test.append(loss) loss_test = np.concatenate(loss_test) roc_test = roc_auc_score(np.asarray(test_loader.dataset.label.values,dtype=int),loss_test) print("The ROC score on SWaT dataset is {}".format(roc_test)) # %%
2,593
35.535211
138
py
GANF
GANF-main/models/RNN.py
#%% import torch import torch.nn as nn from functools import partial class RecurrentEncoder(nn.Module): """Recurrent encoder""" def __init__(self, n_features, latent_dim, rnn): super().__init__() self.rec_enc1 = rnn(n_features, latent_dim, batch_first=True) def forward(self, x): _, h_n = self.rec_enc1(x) return h_n class RecurrentDecoder(nn.Module): """Recurrent decoder for RNN and GRU""" def __init__(self, latent_dim, n_features, rnn_cell, device): super().__init__() self.n_features = n_features self.device = device self.rec_dec1 = rnn_cell(n_features, latent_dim) self.dense_dec1 = nn.Linear(latent_dim, n_features) def forward(self, h_0, seq_len): # Initialize output x = torch.tensor([], device = self.device) # Squeezing h_i = h_0.squeeze() # Reconstruct first element with encoder output x_i = self.dense_dec1(h_i) # Reconstruct remaining elements for i in range(0, seq_len): h_i = self.rec_dec1(x_i, h_i) x_i = self.dense_dec1(h_i) x = torch.cat([x, x_i], axis=1) return x.view(-1, seq_len, self.n_features) class RecurrentDecoderLSTM(nn.Module): """Recurrent decoder LSTM""" def __init__(self, latent_dim, n_features, rnn_cell, device): super().__init__() self.n_features = n_features self.device = device self.rec_dec1 = rnn_cell(n_features, latent_dim) self.dense_dec1 = nn.Linear(latent_dim, n_features) def forward(self, h_0, seq_len): # Initialize output x = torch.tensor([], device = self.device) # Squeezing h_i = [h.squeeze() for h in h_0] # Reconstruct first element with encoder output x_i = self.dense_dec1(h_i[0]) # Reconstruct remaining elements for i in range(0, seq_len): h_i = self.rec_dec1(x_i, h_i) x_i = self.dense_dec1(h_i[0]) x = torch.cat([x, x_i], axis = 1) return x.view(-1, seq_len, self.n_features) class RecurrentAE(nn.Module): """Recurrent autoencoder""" def __init__(self, n_features, latent_dim, device): super().__init__() # Encoder and decoder argsuration self.rnn, self.rnn_cell = nn.LSTM, nn.LSTMCell self.decoder = RecurrentDecoderLSTM self.latent_dim = latent_dim self.n_features = n_features self.device = device # Encoder and decoder self.encoder = RecurrentEncoder(self.n_features, self.latent_dim, self.rnn) self.decoder = self.decoder(self.latent_dim, self.n_features, self.rnn_cell, self.device) def forward(self, x): # x: N X K X L X D seq_len = x.shape[1] h_n = self.encoder(x) out = self.decoder(h_n, seq_len) return torch.flip(out, [1]) # %%
2,929
26.904762
97
py
GANF
GANF-main/models/graph_layer.py
import torch from torch.nn import Parameter, Linear, Sequential, BatchNorm1d, ReLU import torch.nn.functional as F from torch_geometric.nn.conv import MessagePassing from torch_geometric.utils import remove_self_loops, add_self_loops, softmax from torch_geometric.nn.inits import glorot, zeros class GraphLayer(MessagePassing): def __init__(self, in_channels, out_channels, heads=1, concat=True, negative_slope=0.2, dropout=0, bias=True, inter_dim=-1,**kwargs): super(GraphLayer, self).__init__(aggr='add', **kwargs) self.in_channels = in_channels self.out_channels = out_channels self.heads = heads self.concat = concat self.negative_slope = negative_slope self.dropout = dropout self.__alpha__ = None self.lin = Linear(in_channels, heads * out_channels, bias=False) self.att_i = Parameter(torch.Tensor(1, heads, out_channels)) self.att_j = Parameter(torch.Tensor(1, heads, out_channels)) self.att_em_i = Parameter(torch.Tensor(1, heads, out_channels)) self.att_em_j = Parameter(torch.Tensor(1, heads, out_channels)) if bias and concat: self.bias = Parameter(torch.Tensor(heads * out_channels)) elif bias and not concat: self.bias = Parameter(torch.Tensor(out_channels)) else: self.register_parameter('bias', None) self.reset_parameters() def reset_parameters(self): glorot(self.lin.weight) glorot(self.att_i) glorot(self.att_j) zeros(self.att_em_i) zeros(self.att_em_j) zeros(self.bias) def forward(self, x, edge_index, embedding, return_attention_weights=False): """""" if torch.is_tensor(x): x = self.lin(x) x = (x, x) else: x = (self.lin(x[0]), self.lin(x[1])) edge_index, _ = remove_self_loops(edge_index) edge_index, _ = add_self_loops(edge_index, num_nodes=x[1].size(self.node_dim)) out = self.propagate(edge_index, x=x, embedding=embedding, edges=edge_index, return_attention_weights=return_attention_weights) if self.concat: out = out.view(-1, self.heads * self.out_channels) else: out = out.mean(dim=1) if self.bias is not None: out = out + self.bias if return_attention_weights: alpha, self.__alpha__ = self.__alpha__, None return out, (edge_index, alpha) else: return out def message(self, x_i, x_j, edge_index_i, size_i, embedding, edges, return_attention_weights): x_i = x_i.view(-1, self.heads, self.out_channels) x_j = x_j.view(-1, self.heads, self.out_channels) if embedding is not None: embedding_i, embedding_j = embedding[edge_index_i], embedding[edges[0]] embedding_i = embedding_i.unsqueeze(1).repeat(1,self.heads,1) embedding_j = embedding_j.unsqueeze(1).repeat(1,self.heads,1) key_i = torch.cat((x_i, embedding_i), dim=-1) key_j = torch.cat((x_j, embedding_j), dim=-1) cat_att_i = torch.cat((self.att_i, self.att_em_i), dim=-1) cat_att_j = torch.cat((self.att_j, self.att_em_j), dim=-1) alpha = (key_i * cat_att_i).sum(-1) + (key_j * cat_att_j).sum(-1) alpha = alpha.view(-1, self.heads, 1) alpha = F.leaky_relu(alpha, self.negative_slope) alpha = softmax(alpha, edge_index_i, num_nodes=size_i) if return_attention_weights: self.__alpha__ = alpha alpha = F.dropout(alpha, p=self.dropout, training=self.training) return x_j * alpha.view(-1, self.heads, 1) def __repr__(self): return '{}({}, {}, heads={})'.format(self.__class__.__name__, self.in_channels, self.out_channels, self.heads)
4,099
32.333333
84
py
GANF
GANF-main/models/NF.py
#%% import torch import torch.nn as nn import torch.nn.functional as F import torch.distributions as D import math import copy # -------------------- # Model layers and helpers # -------------------- def create_masks(input_size, hidden_size, n_hidden, input_order='sequential', input_degrees=None): # MADE paper sec 4: # degrees of connections between layers -- ensure at most in_degree - 1 connections degrees = [] # set input degrees to what is provided in args (the flipped order of the previous layer in a stack of mades); # else init input degrees based on strategy in input_order (sequential or random) if input_size>1: if input_order == 'sequential': degrees += [torch.arange(input_size)] if input_degrees is None else [input_degrees] for _ in range(n_hidden + 1): degrees += [torch.arange(hidden_size) % (input_size - 1)] degrees += [torch.arange(input_size) % input_size - 1] if input_degrees is None else [input_degrees % input_size - 1] elif input_order == 'random': degrees += [torch.randperm(input_size)] if input_degrees is None else [input_degrees] for _ in range(n_hidden + 1): min_prev_degree = min(degrees[-1].min().item(), input_size - 1) degrees += [torch.randint(min_prev_degree, input_size, (hidden_size,))] min_prev_degree = min(degrees[-1].min().item(), input_size - 1) degrees += [torch.randint(min_prev_degree, input_size, (input_size,)) - 1] if input_degrees is None else [input_degrees - 1] else: degrees += [torch.zeros([1]).long()] for _ in range(n_hidden+1): degrees += [torch.zeros([hidden_size]).long()] degrees += [torch.zeros([input_size]).long()] # construct masks masks = [] for (d0, d1) in zip(degrees[:-1], degrees[1:]): masks += [(d1.unsqueeze(-1) >= d0.unsqueeze(0)).float()] return masks, degrees[0] #%% def create_masks_pmu(input_size, hidden_size, n_hidden, input_order='sequential', input_degrees=None): # MADE paper sec 4: # degrees of connections between layers -- ensure at most in_degree - 1 connections degrees = [] # set input degrees to what is provided in args (the flipped order of the previous layer in a stack of mades); # else init input degrees based on strategy in input_order (sequential or random) if input_order == 'sequential': degrees += [torch.arange(input_size)] if input_degrees is None else [input_degrees] for _ in range(n_hidden + 1): degrees += [torch.arange(hidden_size) % (input_size - 1)] degrees += [torch.arange(input_size) % input_size - 1] if input_degrees is None else [input_degrees % input_size - 1] # construct masks masks = [] for (d0, d1) in zip(degrees[:-1], degrees[1:]): masks += [(d1.unsqueeze(-1) >= d0.unsqueeze(0)).float()] masks[0] = masks[0].repeat_interleave(3, dim=1) masks[-1] = masks[-1].repeat_interleave(3, dim=0) return masks, degrees[0] #%% class MaskedLinear(nn.Linear): """ MADE building block layer """ def __init__(self, input_size, n_outputs, mask, cond_label_size=None): super().__init__(input_size, n_outputs) self.register_buffer('mask', mask) self.cond_label_size = cond_label_size if cond_label_size is not None: self.cond_weight = nn.Parameter(torch.rand(n_outputs, cond_label_size) / math.sqrt(cond_label_size)) def forward(self, x, y=None): out = F.linear(x, self.weight * self.mask, self.bias) if y is not None: out = out + F.linear(y, self.cond_weight) return out def extra_repr(self): return 'in_features={}, out_features={}, bias={}'.format( self.in_features, self.out_features, self.bias is not None ) + (self.cond_label_size != None) * ', cond_features={}'.format(self.cond_label_size) class LinearMaskedCoupling(nn.Module): """ Modified RealNVP Coupling Layers per the MAF paper """ def __init__(self, input_size, hidden_size, n_hidden, mask, cond_label_size=None): super().__init__() self.register_buffer('mask', mask) # scale function s_net = [nn.Linear(input_size + (cond_label_size if cond_label_size is not None else 0), hidden_size)] for _ in range(n_hidden): s_net += [nn.Tanh(), nn.Linear(hidden_size, hidden_size)] s_net += [nn.Tanh(), nn.Linear(hidden_size, input_size)] self.s_net = nn.Sequential(*s_net) # translation function self.t_net = copy.deepcopy(self.s_net) # replace Tanh with ReLU's per MAF paper for i in range(len(self.t_net)): if not isinstance(self.t_net[i], nn.Linear): self.t_net[i] = nn.ReLU() def forward(self, x, y=None): # apply mask mx = x * self.mask # run through model s = self.s_net(mx if y is None else torch.cat([y, mx], dim=1)) t = self.t_net(mx if y is None else torch.cat([y, mx], dim=1)) u = mx + (1 - self.mask) * (x - t) * torch.exp(-s) # cf RealNVP eq 8 where u corresponds to x (here we're modeling u) log_abs_det_jacobian = - (1 - self.mask) * s # log det du/dx; cf RealNVP 8 and 6; note, sum over input_size done at model log_prob return u, log_abs_det_jacobian def inverse(self, u, y=None): # apply mask mu = u * self.mask # run through model s = self.s_net(mu if y is None else torch.cat([y, mu], dim=1)) t = self.t_net(mu if y is None else torch.cat([y, mu], dim=1)) x = mu + (1 - self.mask) * (u * s.exp() + t) # cf RealNVP eq 7 log_abs_det_jacobian = (1 - self.mask) * s # log det dx/du return x, log_abs_det_jacobian class BatchNorm(nn.Module): """ RealNVP BatchNorm layer """ def __init__(self, input_size, momentum=0.9, eps=1e-5): super().__init__() self.momentum = momentum self.eps = eps self.log_gamma = nn.Parameter(torch.zeros(input_size)) self.beta = nn.Parameter(torch.zeros(input_size)) self.register_buffer('running_mean', torch.zeros(input_size)) self.register_buffer('running_var', torch.ones(input_size)) def forward(self, x, cond_y=None): if self.training: self.batch_mean = x.mean(0) self.batch_var = x.var(0) # note MAF paper uses biased variance estimate; ie x.var(0, unbiased=False) # update running mean self.running_mean.mul_(self.momentum).add_(self.batch_mean.data * (1 - self.momentum)) self.running_var.mul_(self.momentum).add_(self.batch_var.data * (1 - self.momentum)) mean = self.batch_mean var = self.batch_var else: mean = self.running_mean var = self.running_var # compute normalized input (cf original batch norm paper algo 1) x_hat = (x - mean) / torch.sqrt(var + self.eps) y = self.log_gamma.exp() * x_hat + self.beta # compute log_abs_det_jacobian (cf RealNVP paper) log_abs_det_jacobian = self.log_gamma - 0.5 * torch.log(var + self.eps) # print('in sum log var {:6.3f} ; out sum log var {:6.3f}; sum log det {:8.3f}; mean log_gamma {:5.3f}; mean beta {:5.3f}'.format( # (var + self.eps).log().sum().data.numpy(), y.var(0).log().sum().data.numpy(), log_abs_det_jacobian.mean(0).item(), self.log_gamma.mean(), self.beta.mean())) return y, log_abs_det_jacobian.expand_as(x) def inverse(self, y, cond_y=None): if self.training: mean = self.batch_mean var = self.batch_var else: mean = self.running_mean var = self.running_var x_hat = (y - self.beta) * torch.exp(-self.log_gamma) x = x_hat * torch.sqrt(var + self.eps) + mean log_abs_det_jacobian = 0.5 * torch.log(var + self.eps) - self.log_gamma return x, log_abs_det_jacobian.expand_as(x) class FlowSequential(nn.Sequential): """ Container for layers of a normalizing flow """ def forward(self, x, y): sum_log_abs_det_jacobians = 0 for module in self: x, log_abs_det_jacobian = module(x, y) sum_log_abs_det_jacobians = sum_log_abs_det_jacobians + log_abs_det_jacobian return x, sum_log_abs_det_jacobians def inverse(self, u, y): sum_log_abs_det_jacobians = 0 for module in reversed(self): u, log_abs_det_jacobian = module.inverse(u, y) sum_log_abs_det_jacobians = sum_log_abs_det_jacobians + log_abs_det_jacobian return u, sum_log_abs_det_jacobians # -------------------- # Models # -------------------- class MADE(nn.Module): def __init__(self, input_size, hidden_size, n_hidden, cond_label_size=None, activation='relu', input_order='sequential', input_degrees=None): """ Args: input_size -- scalar; dim of inputs hidden_size -- scalar; dim of hidden layers n_hidden -- scalar; number of hidden layers activation -- str; activation function to use input_order -- str or tensor; variable order for creating the autoregressive masks (sequential|random) or the order flipped from the previous layer in a stack of mades conditional -- bool; whether model is conditional """ super().__init__() # base distribution for calculation of log prob under the model self.register_buffer('base_dist_mean', torch.zeros(input_size)) self.register_buffer('base_dist_var', torch.ones(input_size)) # create masks masks, self.input_degrees = create_masks(input_size, hidden_size, n_hidden, input_order, input_degrees) # setup activation if activation == 'relu': activation_fn = nn.ReLU() elif activation == 'tanh': activation_fn = nn.Tanh() else: raise ValueError('Check activation function.') # construct model self.net_input = MaskedLinear(input_size, hidden_size, masks[0], cond_label_size) self.net = [] for m in masks[1:-1]: self.net += [activation_fn, MaskedLinear(hidden_size, hidden_size, m)] self.net += [activation_fn, MaskedLinear(hidden_size, 2 * input_size, masks[-1].repeat(2,1))] self.net = nn.Sequential(*self.net) @property def base_dist(self): return D.Normal(self.base_dist_mean, self.base_dist_var) def forward(self, x, y=None): # MAF eq 4 -- return mean and log std m, loga = self.net(self.net_input(x, y)).chunk(chunks=2, dim=1) u = (x - m) * torch.exp(-loga) # MAF eq 5 log_abs_det_jacobian = - loga return u, log_abs_det_jacobian def inverse(self, u, y=None, sum_log_abs_det_jacobians=None): # MAF eq 3 D = u.shape[1] x = torch.zeros_like(u) # run through reverse model for i in self.input_degrees: m, loga = self.net(self.net_input(x, y)).chunk(chunks=2, dim=1) x[:,i] = u[:,i] * torch.exp(loga[:,i]) + m[:,i] log_abs_det_jacobian = loga return x, log_abs_det_jacobian def log_prob(self, x, y=None): u, log_abs_det_jacobian = self.forward(x, y) return torch.sum(self.base_dist.log_prob(u) + log_abs_det_jacobian, dim=1) class MADE_Full(nn.Module): def __init__(self, input_size, hidden_size, n_hidden, cond_label_size=None, activation='relu', input_order='sequential', input_degrees=None): """ Args: input_size -- scalar; dim of inputs hidden_size -- scalar; dim of hidden layers n_hidden -- scalar; number of hidden layers activation -- str; activation function to use input_order -- str or tensor; variable order for creating the autoregressive masks (sequential|random) or the order flipped from the previous layer in a stack of mades conditional -- bool; whether model is conditional """ super().__init__() # base distribution for calculation of log prob under the model self.register_buffer('base_dist_mean', torch.zeros(input_size)) self.register_buffer('base_dist_var', torch.ones(input_size)) # create masks masks, self.input_degrees = create_masks_pmu(int(input_size/3), hidden_size, n_hidden, input_order, input_degrees) # setup activation if activation == 'relu': activation_fn = nn.ReLU() elif activation == 'tanh': activation_fn = nn.Tanh() else: raise ValueError('Check activation function.') # construct model self.net_input = MaskedLinear(input_size, hidden_size, masks[0], cond_label_size) self.net = [] for m in masks[1:-1]: self.net += [activation_fn, MaskedLinear(hidden_size, hidden_size, m)] self.net += [activation_fn, MaskedLinear(hidden_size, 2 * input_size, masks[-1].repeat(2,1))] self.net = nn.Sequential(*self.net) @property def base_dist(self): return D.Normal(self.base_dist_mean, self.base_dist_var) def forward(self, x, y=None): # MAF eq 4 -- return mean and log std m, loga = self.net(self.net_input(x, y)).chunk(chunks=2, dim=1) u = (x - m) * torch.exp(-loga) # MAF eq 5 log_abs_det_jacobian = - loga return u, log_abs_det_jacobian def log_prob(self, x, y=None): u, log_abs_det_jacobian = self.forward(x, y) return torch.sum(self.base_dist.log_prob(u) + log_abs_det_jacobian, dim=1) class MAF(nn.Module): def __init__(self, n_blocks, input_size, hidden_size, n_hidden, cond_label_size=None, activation='relu', input_order='sequential', batch_norm=True): super().__init__() # base distribution for calculation of log prob under the model self.register_buffer('base_dist_mean', torch.zeros(input_size)) self.register_buffer('base_dist_var', torch.ones(input_size)) # construct model modules = [] self.input_degrees = None for i in range(n_blocks): modules += [MADE(input_size, hidden_size, n_hidden, cond_label_size, activation, input_order, self.input_degrees)] self.input_degrees = modules[-1].input_degrees.flip(0) modules += batch_norm * [BatchNorm(input_size)] self.net = FlowSequential(*modules) @property def base_dist(self): return D.Normal(self.base_dist_mean, self.base_dist_var) def forward(self, x, y=None): return self.net(x, y) def inverse(self, u, y=None): return self.net.inverse(u, y) def log_prob(self, x, y=None): u, sum_log_abs_det_jacobians = self.forward(x, y) return torch.sum(self.base_dist.log_prob(u) + sum_log_abs_det_jacobians, dim=1) class MAF_Full(nn.Module): def __init__(self, n_blocks, input_size, hidden_size, n_hidden, cond_label_size=None, activation='relu', input_order='sequential', batch_norm=True): super().__init__() # base distribution for calculation of log prob under the model self.register_buffer('base_dist_mean', torch.zeros(input_size)) self.register_buffer('base_dist_var', torch.ones(input_size)) # construct model modules = [] self.input_degrees = None for i in range(n_blocks): modules += [MADE_Full(input_size, hidden_size, n_hidden, cond_label_size, activation, input_order, self.input_degrees)] self.input_degrees = modules[-1].input_degrees.flip(0) modules += batch_norm * [BatchNorm(input_size)] self.net = FlowSequential(*modules) @property def base_dist(self): return D.Normal(self.base_dist_mean, self.base_dist_var) def forward(self, x, y=None): return self.net(x, y) def inverse(self, u, y=None): return self.net.inverse(u, y) def log_prob(self, x, y=None): u, sum_log_abs_det_jacobians = self.forward(x, y) return torch.sum(self.base_dist.log_prob(u) + sum_log_abs_det_jacobians, dim=1) class RealNVP(nn.Module): def __init__(self, n_blocks, input_size, hidden_size, n_hidden, cond_label_size=None, batch_norm=True): super().__init__() # base distribution for calculation of log prob under the model self.register_buffer('base_dist_mean', torch.zeros(input_size)) self.register_buffer('base_dist_var', torch.ones(input_size)) # construct model modules = [] mask = torch.arange(input_size).float() % 2 for i in range(n_blocks): modules += [LinearMaskedCoupling(input_size, hidden_size, n_hidden, mask, cond_label_size)] mask = 1 - mask modules += batch_norm * [BatchNorm(input_size)] self.net = FlowSequential(*modules) @property def base_dist(self): return D.Normal(self.base_dist_mean, self.base_dist_var) def forward(self, x, y=None): return self.net(x, y) def inverse(self, u, y=None): return self.net.inverse(u, y) def log_prob(self, x, y=None): u, sum_log_abs_det_jacobians = self.forward(x, y) return torch.sum(self.base_dist.log_prob(u) + sum_log_abs_det_jacobians, dim=1)
17,494
39.780886
169
py
GANF
GANF-main/models/DROCC.py
import os import numpy as np import torch import torch.nn as nn import torch.nn.functional as F class LSTM_FC(nn.Module): def __init__(self, input_dim=32, num_classes=1, num_hidden_nodes=8 ): super(LSTM_FC, self).__init__() self.input_dim = input_dim self.num_classes = num_classes self.num_hidden_nodes = num_hidden_nodes self.encoder = nn.LSTM(input_size=self.input_dim, hidden_size=self.num_hidden_nodes, num_layers=1, batch_first=True) self.fc = nn.Linear(self.num_hidden_nodes, self.num_classes) activ = nn.ReLU(True) def forward(self, input): features = self.encoder(input)[0][:,-1,:] # pdb.set_trace() logits = self.fc(features) return logits def half_forward_start(self, x): features = self.encoder(x)[0][:,-1,:] return features def half_forward_end(self, x): logits = self.fc(x) return logits class DROCCTrainer: """ Trainer class that implements the DROCC algorithm proposed in https://arxiv.org/abs/2002.12718 """ def __init__(self, model, optimizer, lamda, radius, gamma, device): """Initialize the DROCC Trainer class Parameters ---------- model: Torch neural network object optimizer: Total number of epochs for training. lamda: Weight given to the adversarial loss radius: Radius of hypersphere to sample points from. gamma: Parameter to vary projection. device: torch.device object for device to use. """ self.model = model self.optimizer = optimizer self.lamda = lamda self.radius = radius self.gamma = gamma self.device = device def train(self, train_loader, lr_scheduler, total_epochs, save_path, name, only_ce_epochs=5, ascent_step_size=0.001, ascent_num_steps=50): """Trains the model on the given training dataset with periodic evaluation on the validation dataset. Parameters ---------- train_loader: Dataloader object for the training dataset. val_loader: Dataloader object for the validation dataset. learning_rate: Initial learning rate for training. total_epochs: Total number of epochs for training. only_ce_epochs: Number of epochs for initial pretraining. ascent_step_size: Step size for gradient ascent for adversarial generation of negative points. ascent_num_steps: Number of gradient ascent steps for adversarial generation of negative points. metric: Metric used for evaluation (AUC / F1). """ self.ascent_num_steps = ascent_num_steps self.ascent_step_size = ascent_step_size for epoch in range(total_epochs): #Make the weights trainable self.model.train() #Placeholder for the respective 2 loss values epoch_adv_loss = 0.0 #AdvLoss epoch_ce_loss = 0.0 #Cross entropy Loss batch_idx = -1 for data in train_loader: batch_idx += 1 data = data.to(self.device) target = torch.ones([data.shape[0]], dtype=torch.float32).to(self.device) data = torch.transpose(data, dim0=1, dim1=2) data = data.reshape(data.shape[0], data.shape[1], data.shape[2]*data.shape[3]) self.optimizer.zero_grad() # Extract the logits for cross entropy loss logits = self.model(data) logits = torch.squeeze(logits, dim = 1) ce_loss = F.binary_cross_entropy_with_logits(logits, target) # Add to the epoch variable for printing average CE Loss epoch_ce_loss += ce_loss.item() ''' Adversarial Loss is calculated only for the positive data points (label==1). ''' if epoch >= only_ce_epochs: data = data[target == 1] # AdvLoss adv_loss = self.one_class_adv_loss(data) epoch_adv_loss += adv_loss.item() loss = ce_loss + adv_loss * self.lamda else: # If only CE based training has to be done loss = ce_loss # Backprop loss.backward() self.optimizer.step() lr_scheduler.step() epoch_ce_loss = epoch_ce_loss/(batch_idx + 1) #Average CE Loss epoch_adv_loss = epoch_adv_loss/(batch_idx + 1) #Average AdvLoss print('Epoch: {}, CE Loss: {}, AdvLoss: {}'.format( epoch, epoch_ce_loss, epoch_adv_loss)) self.save(os.path.join(save_path, "{}_{}.pt".format(name, epoch))) def test(self, test_loader): """Evaluate the model on the given test dataset. Parameters ---------- test_loader: Dataloader object for the test dataset. metric: Metric used for evaluation (AUC / F1). """ self.model.eval() scores = [] with torch.no_grad(): for data in test_loader: data = data.to(self.device) data = torch.transpose(data, dim0=1, dim1=2) data = data.reshape(data.shape[0], data.shape[1], data.shape[2]*data.shape[3]) logits = self.model(data).cpu().numpy() scores.append(logits) scores = -np.concatenate(scores) return scores def one_class_adv_loss(self, x_train_data): """Computes the adversarial loss: 1) Sample points initially at random around the positive training data points 2) Gradient ascent to find the most optimal point in set N_i(r) classified as +ve (label=0). This is done by maximizing the CE loss wrt label 0 3) Project the points between spheres of radius R and gamma * R (set N_i(r)) 4) Pass the calculated adversarial points through the model, and calculate the CE loss wrt target class 0 Parameters ---------- x_train_data: Batch of data to compute loss on. """ batch_size = len(x_train_data) # Randomly sample points around the training data # We will perform SGD on these to find the adversarial points x_adv = torch.randn(x_train_data.shape).to(self.device).detach().requires_grad_() x_adv_sampled = x_adv + x_train_data for step in range(self.ascent_num_steps): with torch.enable_grad(): new_targets = torch.zeros(batch_size, 1).to(self.device) new_targets = torch.squeeze(new_targets) new_targets = new_targets.to(torch.float) logits = self.model(x_adv_sampled) logits = torch.squeeze(logits, dim = 1) new_loss = F.binary_cross_entropy_with_logits(logits, new_targets) grad = torch.autograd.grad(new_loss, [x_adv_sampled])[0] grad_norm = torch.norm(grad, p=2, dim = tuple(range(1, grad.dim()))) grad_norm = grad_norm.view(-1, *[1]*(grad.dim()-1)) grad_normalized = grad/grad_norm with torch.no_grad(): x_adv_sampled.add_(self.ascent_step_size * grad_normalized) if (step + 1) % 10==0: # Project the normal points to the set N_i(r) h = x_adv_sampled - x_train_data norm_h = torch.sqrt(torch.sum(h**2, dim=tuple(range(1, h.dim())))) alpha = torch.clamp(norm_h, self.radius, self.gamma * self.radius).to(self.device) # Make use of broadcast to project h proj = (alpha/norm_h).view(-1, *[1] * (h.dim()-1)) h = proj * h x_adv_sampled = x_train_data + h #These adv_points are now on the surface of hyper-sphere adv_pred = self.model(x_adv_sampled) adv_pred = torch.squeeze(adv_pred, dim=1) adv_loss = F.binary_cross_entropy_with_logits(adv_pred, (new_targets * 0)) return adv_loss def save(self, path): torch.save(self.model.state_dict(),path) def load(self, path): self.model.load_state_dict(torch.load(path))
8,730
39.235023
106
py
GANF
GANF-main/models/DeepSAD.py
#%% import json import torch import logging import time import torch import torch.optim as optim class AETrainer: def __init__(self, device: str = 'cuda'): self.device = device def train(self, train_loader, ae_net, args): logger = logging.getLogger() # Set device for network ae_net = ae_net.to(self.device) # Set optimizer (Adam optimizer for now) optimizer = optim.Adam(ae_net.parameters(), lr=args.lr, weight_decay=args.weight_decay) # Training print('Starting pretraining...') start_time = time.time() ae_net.train() for epoch in range(10): loss_epoch = 0.0 n_batches = 0 epoch_start_time = time.time() for data in train_loader: if isinstance(data, list): data = data[0] x = data.to(self.device) x = torch.transpose(x, dim0=2, dim1=3) inputs = x.reshape(x.shape[0], x.shape[1]*x.shape[2], x.shape[3]) # Zero the network parameter gradients optimizer.zero_grad() # Update network parameters via backpropagation: forward + backward + optimize outputs = ae_net(inputs) scores = torch.sum((outputs - inputs) ** 2, dim=tuple(range(1, outputs.dim()))) loss = torch.mean(scores) loss.backward() optimizer.step() loss_epoch += loss.item() n_batches += 1 # log epoch statistics epoch_train_time = time.time() - epoch_start_time print(' Epoch {}/{}\t Time: {:.3f}\t Loss: {:.8f}' .format(epoch + 1, 10, epoch_train_time, loss_epoch / n_batches)) pretrain_time = time.time() - start_time print('Pretraining time: %.3f' % pretrain_time) print('Finished pretraining.') return ae_net class DeepSVDDTrainer: def __init__(self, device: str = 'cuda'): self.device = device # Deep SVDD parameters self.c = None def train(self, train_loader, net, args): self.args = args # Set device for network net = net.to(self.device) # Set optimizer (Adam optimizer for now) optimizer = optim.Adam(net.parameters(), lr=args.lr, weight_decay=args.weight_decay) # Set learning rate scheduler scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[20], gamma=0.1) # Initialize hypersphere center c (if c not loaded) if self.c is None: print('Initializing center c...') self.c = self.init_center_c(train_loader, net) print(self.c.shape) print('Center c initialized.') # Training print('Starting training...') start_time = time.time() net.train() save_path = os.path.join(args.output_dir,args.name) if not os.path.exists(save_path): os.makedirs(save_path) for epoch in range(args.n_epochs): scheduler.step() loss_epoch = 0.0 n_batches = 0 epoch_start_time = time.time() for data in train_loader: x = data.to(self.device) x = torch.transpose(x, dim0=2, dim1=3) inputs = x.reshape(x.shape[0], x.shape[1]*x.shape[2], x.shape[3]) # Zero the network parameter gradients optimizer.zero_grad() # Update network parameters via backpropagation: forward + backward + optimize outputs = net(inputs).squeeze().mean(dim=-1) dist = torch.sum((outputs - self.c) ** 2, dim=1) loss = torch.mean(dist) loss.backward() optimizer.step() loss_epoch += loss.item() n_batches += 1 # log epoch statistics epoch_train_time = time.time() - epoch_start_time print(' Epoch {}/{}\t Time: {:.3f}\t Loss: {:.8f}' .format(epoch + 1, args.n_epochs, epoch_train_time, loss_epoch / n_batches)) torch.save({'c': self.c, 'net_dict': net.state_dict()}, os.path.join(save_path, "{}_{}.pt".format(args.name, epoch))) self.train_time = time.time() - start_time print('Training time: %.3f' % self.train_time) print('Finished training.') return net def init_center_c(self, train_loader, net, eps=0.1): """Initialize hypersphere center c as the mean from an initial forward pass on the data.""" n_samples = 0 c = 0.0 net.eval() with torch.no_grad(): for data in train_loader: # get the inputs of the batch x = data.to(self.device) x = torch.transpose(x, dim0=2, dim1=3) inputs = x.reshape(x.shape[0], x.shape[1]*x.shape[2], x.shape[3]) outputs = net(inputs).squeeze() n_samples += outputs.shape[0] c += torch.sum(outputs, dim=0) c /= n_samples # If c_i is too close to 0, set to +-eps. Reason: a zero unit can be trivially matched with zero weights. c[(abs(c) < eps) & (c < 0)] = -eps c[(abs(c) < eps) & (c > 0)] = eps return c.mean(dim=-1) from models.RNN import RecurrentAE from models.GAN import CNNAE class DeepSVDD(object): def __init__(self, n_features, hidden_size, device): self.c = None # hypersphere center c self.trainer = None # if encoder=='RNN': # self.ae_net = RecurrentAE(n_features, hidden_size, device) self.ae_net = CNNAE(n_features, hidden_size).to(device) self.net = self.ae_net.encoder self.ae_trainer = None self.results = { 'test_auc': None } def train(self, dataset, args, device: str = 'cuda'): """Trains the Deep SVDD model on the training data.""" self.trainer = DeepSVDDTrainer(device=device) # Get the model self.trainer.train(dataset, self.net, args) self.c = self.trainer.c def test(self, test_loader, delta_t, sigma, device): from utils import roc_auc_all import numpy as np self.net.eval() self.net.to(device) loss = [] with torch.no_grad(): for data in test_loader: x = data.to(device) x = torch.transpose(x, dim0=2, dim1=3) inputs = x.reshape(x.shape[0], x.shape[1]*x.shape[2], x.shape[3]) outputs = self.net(inputs).squeeze().mean(dim=-1) batch_loss= torch.sum((outputs - self.c) ** 2, dim=1).cpu().numpy() loss.append(batch_loss) loss = np.concatenate(loss) auc_score, fps,tps = roc_auc_all(loss, delta_t, sigma) print("meann: {:.4f}, median: {:.4f}, auc:{:.4f}".format(np.mean(loss), np.median(loss),auc_score))# %% self.results['test_auc'] = auc_score return auc_score, fps,tps def pretrain(self, train_loader, args, device): """Pretrains the weights for the Deep SVDD network \phi via autoencoder.""" self.ae_trainer = AETrainer(device=device) self.ae_net = self.ae_trainer.train(train_loader, self.ae_net, args) self.net = self.ae_net.encoder def save_model(self, export_model): """Save Deep SVDD model to export_model.""" net_dict = self.net.state_dict() torch.save({'c': self.c, 'net_dict': net_dict}, export_model) def load_model(self, model_path): """Load Deep SVDD model from model_path.""" model_dict = torch.load(model_path) self.c = model_dict['c'] self.net.load_state_dict(model_dict['net_dict']) def save_results(self, export_json): """Save results dict to a JSON-file.""" with open(export_json, 'w') as fp: json.dump(self.results, fp) # %% import os class DeepSADTrainer: def __init__(self, device: str = 'cuda'): self.device = device self.c = None def train(self, train_loader, net, args): self.args = args # Set device for network net = net.to(self.device) # Set optimizer (Adam optimizer for now) optimizer = optim.Adam(net.parameters(), lr=args.lr, weight_decay=args.weight_decay) # Set learning rate scheduler scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[20], gamma=0.1) # Initialize hypersphere center c (if c not loaded) if self.c is None: print('Initializing center c...') self.c = self.init_center_c(train_loader, net) print('Center c initialized.') # Training print('Starting training...') start_time = time.time() net.train() save_path = os.path.join(args.output_dir,args.name) if not os.path.exists(save_path): os.makedirs(save_path) for epoch in range(args.n_epochs): loss_epoch = 0.0 n_batches = 0 epoch_start_time = time.time() for data, semi_targets in train_loader: x = data.to(self.device) x = torch.transpose(x, dim0=2, dim1=3) inputs = x.reshape(x.shape[0], x.shape[1]*x.shape[2], x.shape[3]) semi_targets = semi_targets.to(self.device) # Zero the network parameter gradients optimizer.zero_grad() # Update network parameters via backpropagation: forward + backward + optimize outputs = net(inputs).squeeze().mean(dim=-1) dist = torch.sum((outputs - self.c) ** 2, dim=1) losses = torch.where(semi_targets == 0, dist, args.eta * ((dist + 1e-6) ** semi_targets.float())) loss = torch.mean(losses) loss.backward() optimizer.step() loss_epoch += loss.item() n_batches += 1 scheduler.step() # log epoch statistics epoch_train_time = time.time() - epoch_start_time print(' Epoch {}/{}\t Time: {:.3f}\t Loss: {:.8f}' .format(epoch + 1, args.n_epochs, epoch_train_time, loss_epoch / n_batches)) torch.save({'c': self.c, 'net_dict': net.state_dict()}, os.path.join(save_path, "{}_{}.pt".format(args.name, epoch))) self.train_time = time.time() - start_time print('Training time: %.3f' % self.train_time) print('Finished training.') return net def init_center_c(self, train_loader, net, eps=0.1): """Initialize hypersphere center c as the mean from an initial forward pass on the data.""" n_samples = 0 c = 0.0 net.eval() with torch.no_grad(): for data, _ in train_loader: # get the inputs of the batch x = data.to(self.device) x = torch.transpose(x, dim0=2, dim1=3) inputs = x.reshape(x.shape[0], x.shape[1]*x.shape[2], x.shape[3]) outputs = net(inputs).squeeze() n_samples += outputs.shape[0] c += torch.sum(outputs, dim=0) c /= n_samples # If c_i is too close to 0, set to +-eps. Reason: a zero unit can be trivially matched with zero weights. c[(abs(c) < eps) & (c < 0)] = -eps c[(abs(c) < eps) & (c > 0)] = eps return c.mean(dim=-1) class DeepSAD(object): def __init__(self, n_features, hidden_size, device): self.c = None # hypersphere center c self.trainer = None self.ae_net = CNNAE(n_features, hidden_size).to(device) self.net = self.ae_net.encoder self.ae_trainer = None self.results = { 'test_auc': None } def train(self, dataset, args, device: str = 'cuda'): self.trainer = DeepSADTrainer(device=device) # Get the model self.trainer.train(dataset, self.net, args) self.c = self.trainer.c def test(self, test_loader, delta_t, sigma, device): from utils import roc_auc_all import numpy as np self.net.eval() self.net.to(device) loss = [] with torch.no_grad(): for data in test_loader: x = data.to(device) x = torch.transpose(x, dim0=2, dim1=3) inputs = x.reshape(x.shape[0], x.shape[1]*x.shape[2], x.shape[3]) outputs = self.net(inputs).squeeze().mean(dim=-1) batch_loss= torch.sum((outputs - self.c) ** 2, dim=1).cpu().numpy() loss.append(batch_loss) loss = np.concatenate(loss) auc_score, fps,tps = roc_auc_all(loss, delta_t, sigma) print("mean: {:.4f}, median: {:.4f}, auc:{:.4f}".format(np.mean(loss), np.median(loss),auc_score))# %% self.results['test_auc'] = auc_score return auc_score,fps,tps def pretrain(self, train_loader, args, device): self.ae_trainer = AETrainer(device=device) self.ae_net = self.ae_trainer.train(train_loader, self.ae_net, args) self.net = self.ae_net.encoder def save_model(self, export_model): """Save Deep SVDD model to export_model.""" net_dict = self.net.state_dict() torch.save({'c': self.c, 'net_dict': net_dict}, export_model) def load_model(self, model_path, load_ae=False): """Load Deep SVDD model from model_path.""" model_dict = torch.load(model_path) self.c = model_dict['c'] self.net.load_state_dict(model_dict['net_dict']) def save_results(self, export_json): """Save results dict to a JSON-file.""" with open(export_json, 'w') as fp: json.dump(self.results, fp) # %%
14,064
31.55787
129
py
GANF
GANF-main/models/GAN.py
#%% import os import torch import torch.nn as nn import torch.nn.functional as F from timeit import default_timer as timer def ConvEncoder(activation = nn.LeakyReLU, in_channels:int = 3, n_c:int = 64, k_size:int = 5): enc = nn.Sequential(*(nn.Conv1d(in_channels, n_c, k_size, stride=2, padding=2), nn.BatchNorm1d(n_c), activation(), nn.Conv1d(n_c, n_c*2, k_size, stride=2, padding=2), nn.BatchNorm1d(n_c*2), activation(), nn.Conv1d(n_c*2, n_c*4, k_size, stride=2, padding=2), nn.BatchNorm1d(n_c*4), activation())) return enc def ConvDecoder(activation = nn.LeakyReLU, in_channels:int = 3, n_c:int = 64, k_size:int = 5): decoder = nn.Sequential(*(nn.ConvTranspose1d(n_c*4, n_c*2, k_size, stride=2, padding=2, output_padding=0), torch.nn.BatchNorm1d(n_c*2), activation(), torch.nn.ConvTranspose1d(n_c*2, n_c, k_size,stride=2, padding=2, output_padding=1), torch.nn.BatchNorm1d(n_c), activation(), torch.nn.ConvTranspose1d(n_c, in_channels, k_size,stride=2, padding=2, output_padding=1))) return decoder class CNNAE(torch.nn.Module): """Recurrent autoencoder""" def __init__(self,in_channels:int = 3, n_channels:int = 16, kernel_size:int = 5): super(CNNAE, self).__init__() # Encoder and decoder argsuration activation = torch.nn.LeakyReLU self.in_channels = in_channels self.n_c = n_channels self.k_size = kernel_size self.encoder = ConvEncoder(activation, in_channels, n_channels, kernel_size) self.decoder = ConvDecoder(activation, in_channels, n_channels, kernel_size) def forward(self, x:torch.Tensor): z = self.encoder.forward(x) x_out = self.decoder.forward(z) return x_out class R_Net(torch.nn.Module): def __init__(self, activation = torch.nn.LeakyReLU, in_channels:int = 3, n_channels:int = 16, kernel_size:int = 5, std:float = 0.2): super(R_Net, self).__init__() self.activation = activation self.in_channels = in_channels self.n_c = n_channels self.k_size = kernel_size self.std = std self.Encoder = ConvEncoder(activation, in_channels, n_channels, kernel_size) self.Decoder = ConvDecoder(activation, in_channels, n_channels, kernel_size) def forward(self, x:torch.Tensor, noise:bool = True): x_hat = self.add_noise(x) if noise else x z = self.Encoder.forward(x_hat) x_out = self.Decoder.forward(z) return x_out def add_noise(self, x): noise = torch.randn_like(x) * self.std x_hat = x + noise return x_hat class D_Net(torch.nn.Module): def __init__(self, in_resolution:int, activation = torch.nn.LeakyReLU, in_channels:int = 3, n_channels:int = 16, kernel_size:int = 5): super(D_Net, self).__init__() self.activation = activation self.in_resolution = in_resolution self.in_channels = in_channels self.n_c = n_channels self.k_size = kernel_size self.cnn = ConvEncoder(activation, in_channels, n_channels, kernel_size) # Compute output dimension after conv part of D network self.out_dim = self._compute_out_dim() self.fc = torch.nn.Linear(self.out_dim, 1) def _compute_out_dim(self): test_x = torch.Tensor(1, self.in_channels, self.in_resolution) for p in self.cnn.parameters(): p.requires_grad = False test_x = self.cnn(test_x) out_dim = torch.prod(torch.tensor(test_x.shape[1:])).item() for p in self.cnn.parameters(): p.requires_grad = True return out_dim def forward(self, x:torch.Tensor): x = self.cnn(x) x = torch.flatten(x, start_dim = 1) out = self.fc(x) return out def R_Loss(d_net: torch.nn.Module, x_real: torch.Tensor, x_fake: torch.Tensor, lambd: float) -> dict: pred = d_net(x_fake) y = torch.ones_like(pred) rec_loss = F.mse_loss(x_fake, x_real) gen_loss = F.binary_cross_entropy_with_logits(pred, y) # generator loss L_r = gen_loss + lambd * rec_loss return {'rec_loss' : rec_loss, 'gen_loss' : gen_loss, 'L_r' : L_r} def D_Loss(d_net: torch.nn.Module, x_real: torch.Tensor, x_fake: torch.Tensor) -> torch.Tensor: pred_real = d_net(x_real) pred_fake = d_net(x_fake.detach()) y_real = torch.ones_like(pred_real) y_fake = torch.zeros_like(pred_fake) real_loss = F.binary_cross_entropy_with_logits(pred_real, y_real) fake_loss = F.binary_cross_entropy_with_logits(pred_fake, y_fake) return real_loss + fake_loss # Wasserstein GAN loss (https://arxiv.org/abs/1701.07875) def R_WLoss(d_net: torch.nn.Module, x_real: torch.Tensor, x_fake: torch.Tensor, lambd: float) -> dict: pred = torch.sigmoid(d_net(x_fake)) rec_loss = F.mse_loss(x_fake, x_real) gen_loss = -torch.mean(pred) # Wasserstein G loss: - E[ D(G(x)) ] L_r = gen_loss + lambd * rec_loss return {'rec_loss' : rec_loss, 'gen_loss' : gen_loss, 'L_r' : L_r} def D_WLoss(d_net: torch.nn.Module, x_real: torch.Tensor, x_fake: torch.Tensor) -> torch.Tensor: pred_real = torch.sigmoid(d_net(x_real)) pred_fake = torch.sigmoid(d_net(x_fake.detach())) dis_loss = -torch.mean(pred_real) + torch.mean(pred_fake) # Wasserstein D loss: -E[D(x_real)] + E[D(x_fake)] return dis_loss # %% def train_model(r_net: torch.nn.Module, d_net: torch.nn.Module, train_dataset: torch.utils.data.Dataset, valid_dataset: torch.utils.data.Dataset, r_loss = R_Loss, d_loss = D_Loss, lr_scheduler = None, optimizer_class = torch.optim.Adam, optim_r_params: dict = {}, optim_d_params: dict = {}, learning_rate: float = 0.001, scheduler_r_params: dict = {}, scheduler_d_params: dict = {}, batch_size: int = 1024, max_epochs: int = 40, epoch_step: int = 1, save_step: int = 5, lambd: float = 0.2, device: torch.device = torch.device('cuda'), save_path: str = ".") -> tuple: optim_r = optimizer_class(r_net.parameters(), lr = learning_rate, **optim_r_params) optim_d = optimizer_class(d_net.parameters(), lr = learning_rate, **optim_d_params) if lr_scheduler: scheduler_r = lr_scheduler(optim_r, **scheduler_r_params) scheduler_d = lr_scheduler(optim_d, **scheduler_d_params) train_loader = torch.utils.data.DataLoader(train_dataset, shuffle=True, batch_size=batch_size) valid_loader = torch.utils.data.DataLoader(valid_dataset, batch_size=batch_size) for epoch in range(max_epochs): start = timer() train_metrics = train_single_epoch(r_net, d_net, optim_r, optim_d, r_loss, d_loss, train_loader, lambd, device) valid_metrics = validate_single_epoch(r_net, d_net, r_loss, d_loss, valid_loader, device) time = timer() - start if epoch % epoch_step == 0: print(f'Epoch {epoch}:') print('Train Metrics:', train_metrics) print('Val Metrics:', valid_metrics) print(f'TIME: {time:.2f} s') if lr_scheduler: scheduler_r.step() scheduler_d.step() if epoch % save_step == 0: torch.save(r_net.state_dict(), os.path.join(save_path, "r_net_{}.pt".format(epoch))) torch.save(d_net.state_dict(), os.path.join(save_path, "d_net_{}.pt".format(epoch))) print(f'Saving model on epoch {epoch}') return (r_net, d_net) def train_single_epoch(r_net, d_net, optim_r, optim_d, r_loss, d_loss, train_loader, lambd, device) -> dict: r_net.train() d_net.train() train_metrics = {'rec_loss' : 0, 'gen_loss' : 0, 'dis_loss' : 0} for data in train_loader: x = data.to(device) x = torch.transpose(x, dim0=2, dim1=3) x_real = x.reshape(x.shape[0], x.shape[1]*x.shape[2], x.shape[3]) x_fake = r_net(x_real) d_net.zero_grad() dis_loss = d_loss(d_net, x_real, x_fake) dis_loss.backward() optim_d.step() r_net.zero_grad() r_metrics = r_loss(d_net, x_real, x_fake, lambd) # L_r = gen_loss + lambda * rec_loss r_metrics['L_r'].backward() optim_r.step() train_metrics['rec_loss'] += r_metrics['rec_loss'] train_metrics['gen_loss'] += r_metrics['gen_loss'] train_metrics['dis_loss'] += dis_loss train_metrics['rec_loss'] = train_metrics['rec_loss'].item() / (len(train_loader.dataset) / train_loader.batch_size) train_metrics['gen_loss'] = train_metrics['gen_loss'].item() / (len(train_loader.dataset) / train_loader.batch_size) train_metrics['dis_loss'] = train_metrics['dis_loss'].item() / (len(train_loader.dataset) / train_loader.batch_size) return train_metrics def validate_single_epoch(r_net, d_net, r_loss, d_loss, valid_loader, device) -> dict: r_net.eval() d_net.eval() valid_metrics = {'rec_loss' : 0, 'gen_loss' : 0, 'dis_loss' : 0} with torch.no_grad(): for data in valid_loader: x = data.to(device) x = torch.transpose(x, dim0=2, dim1=3) x_real = x.reshape(x.shape[0], x.shape[1]*x.shape[2], x.shape[3]) x_fake = r_net(x_real) dis_loss = d_loss(d_net, x_real, x_fake) r_metrics = r_loss(d_net, x_real, x_fake, 0) valid_metrics['rec_loss'] += r_metrics['rec_loss'] valid_metrics['gen_loss'] += r_metrics['gen_loss'] valid_metrics['dis_loss'] += dis_loss valid_metrics['rec_loss'] = valid_metrics['rec_loss'].item() / (len(valid_loader.dataset) / valid_loader.batch_size) valid_metrics['gen_loss'] = valid_metrics['gen_loss'].item() / (len(valid_loader.dataset) / valid_loader.batch_size) valid_metrics['dis_loss'] = valid_metrics['dis_loss'].item() / (len(valid_loader.dataset) / valid_loader.batch_size) return valid_metrics # %%
10,849
34.113269
138
py
GANF
GANF-main/models/GANF.py
#%% import torch.nn as nn import torch.nn.functional as F from models.NF import MAF, RealNVP import torch class GNN(nn.Module): """ The GNN module applied in GANF """ def __init__(self, input_size, hidden_size): super(GNN, self).__init__() self.lin_n = nn.Linear(input_size, hidden_size) self.lin_r = nn.Linear(input_size, hidden_size, bias=False) self.lin_2 = nn.Linear(hidden_size, hidden_size) def forward(self, h, A): ## A: K X K ## H: N X K X L X D h_n = self.lin_n(torch.einsum('nkld,kj->njld',h,A)) h_r = self.lin_r(h[:,:,:-1]) h_n[:,:,1:] += h_r h = self.lin_2(F.relu(h_n)) return h class GANF(nn.Module): def __init__ (self, n_blocks, input_size, hidden_size, n_hidden ,dropout = 0.1, model="MAF", batch_norm=True): super(GANF, self).__init__() self.rnn = nn.LSTM(input_size=input_size,hidden_size=hidden_size,batch_first=True, dropout=dropout) self.gcn = GNN(input_size=hidden_size, hidden_size=hidden_size) if model=="MAF": self.nf = MAF(n_blocks, input_size, hidden_size, n_hidden, cond_label_size=hidden_size, batch_norm=batch_norm,activation='tanh') else: self.nf = RealNVP(n_blocks, input_size, hidden_size, n_hidden, cond_label_size=hidden_size, batch_norm=batch_norm) def forward(self, x, A): return self.test(x, A).mean() def test(self, x, A): # x: N X K X L X D full_shape = x.shape # reshape: N*K, L, D x = x.reshape((x.shape[0]*x.shape[1], x.shape[2], x.shape[3])) h,_ = self.rnn(x) # resahpe: N, K, L, H h = h.reshape((full_shape[0], full_shape[1], h.shape[1], h.shape[2])) h = self.gcn(h, A) # reshappe N*K*L,H h = h.reshape((-1,h.shape[3])) x = x.reshape((-1,full_shape[3])) log_prob = self.nf.log_prob(x,h).reshape([full_shape[0],-1])#*full_shape[1]*full_shape[2] log_prob = log_prob.mean(dim=1) return log_prob def locate(self, x, A): # x: N X K X L X D full_shape = x.shape # reshape: N*K, L, D x = x.reshape((x.shape[0]*x.shape[1], x.shape[2], x.shape[3])) h,_ = self.rnn(x) # resahpe: N, K, L, H h = h.reshape((full_shape[0], full_shape[1], h.shape[1], h.shape[2])) h = self.gcn(h, A) # reshappe N*K*L,H h = h.reshape((-1,h.shape[3])) x = x.reshape((-1,full_shape[3])) log_prob = self.nf.log_prob(x,h).reshape([full_shape[0],full_shape[1],-1])#*full_shape[1]*full_shape[2] log_prob = log_prob.mean(dim=2) return log_prob
2,705
28.413043
140
py
GANF
GANF-main/example_baseline/train_SVDD_water.py
#%% import os import argparse import torch from models.RNN import RecurrentAE import torch.nn.functional as F from dataset import PMUTime import numpy as np parser = argparse.ArgumentParser() # action parser.add_argument('--data_dir', type=str, default='/data', help='Location of datasets.') parser.add_argument('--output_dir', type=str, default='/home/enyandai/code/checkpoint/model') parser.add_argument('--dataset', type=str, default='C') parser.add_argument('--model', type=str, default='None') parser.add_argument('--name',default='SVDD_Water_test') parser.add_argument('--seed', type=int, default=11, help='Random seed to use.') # made parameters parser.add_argument('--hidden_size', type=int, default=64, help='Hidden layer size for MADE (and each MADE block in an MAF).') parser.add_argument('--n_hidden', type=int, default=1, help='Number of hidden layers in each MADE.') # training params parser.add_argument('--batch_size', type=int, default=1024) parser.add_argument('--weight_decay', type=float, default=5e-4) parser.add_argument('--n_epochs', type=int, default=50) parser.add_argument('--lr', type=float, default=1e-3, help='Learning rate.') args = parser.parse_known_args()[0] args.cuda = torch.cuda.is_available() device = torch.device("cuda" if args.cuda else "cpu") print(args) import random import numpy as np random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if args.cuda: torch.cuda.manual_seed(args.seed) #%% load dataset print("Loading dataset") from dataset import load_water train_loader, val_loader, test_loader, n_sensor = load_water("/home/enyandai/orginal_code/data/SWaT_Dataset_Attack_v0.csv", \ args.batch_size) # %% from models.DeepSAD import DeepSVDD model = DeepSVDD(n_sensor, args.hidden_size, device) #%% model.pretrain(train_loader, args, device) model.train(train_loader, args, device) #%% save_path = os.path.join(args.output_dir,args.name) if not os.path.exists(save_path): os.makedirs(save_path) model.save_model(os.path.join(save_path, "{}.pt".format(args.name))) #%% # for seed in range(10,21): # model.load_model("/home/enyandai/orginal_code/checkpoint/model/SVDD_Water/SVDD_Water_39.pt") model.net.eval() loss = [] from sklearn.metrics import roc_auc_score with torch.no_grad(): for data in test_loader: x = data.to(device) x = torch.transpose(x, dim0=2, dim1=3) inputs = x.reshape(x.shape[0], x.shape[1]*x.shape[2], x.shape[3]) outputs = model.net(inputs).squeeze().mean(dim=-1) batch_loss= torch.sum((outputs - model.c) ** 2, dim=1).cpu().numpy() loss.append(batch_loss) loss = np.concatenate(loss) roc_test = roc_auc_score(np.asarray(test_loader.dataset.label.values,dtype=int),loss) print("ROC: {:.4f}".format(roc_test)) # %% # %%
2,916
33.317647
126
py
Rickrolling-the-Artist
Rickrolling-the-Artist-main/generate_images.py
# code is partly based on https://huggingface.co/blog/stable_diffusion import argparse import math import os import pathlib from datetime import datetime import torch from diffusers import AutoencoderKL, LMSDiscreteScheduler, UNet2DConditionModel from PIL import Image from rtpt import RTPT from torch import autocast from tqdm.auto import tqdm from transformers import CLIPTextModel, CLIPTokenizer import wandb from utils.stable_diffusion_utils import generate def main(): args = create_parser() torch.manual_seed(args.seed) if args.prompt_file is not None and args.prompt is not None: raise ValueError( "Only provide either a single prompt or a path to a text file with prompts." ) if args.prompt: prompts = [args.prompt] else: prompts = read_prompt_file(args.prompt_file) prompts = [item for item in prompts for i in range(args.num_samples)] max_iterations = math.ceil(len(prompts) / args.batch_size) rtpt = RTPT(args.user, 'image_generation', max_iterations=max_iterations) rtpt.start() # load the autoencoder model which will be used to decode the latents into image space. model_path = 'CompVis/stable-diffusion-v1-4' if args.version in ['v1-1', 'v1-2', 'v1-3', 'v1-4']: model_path = f'CompVis/stable-diffusion-{args.version}' elif args.version in ['v1-5']: model_path = f'runwayml/stable-diffusion-{args.version}' else: raise ValueError( f'{args.version} is no valid Stable Diffusion version. ' + 'Please specify one of {v1-1, v1-2, v1-3, v1-4, v1-5}.') vae = AutoencoderKL.from_pretrained(model_path, subfolder="vae", use_auth_token=args.hf_token) # load the CLIP tokenizer and text encoder to tokenize and encode the text. tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14") if args.encoder_path: print('Load poisoned CLIP text encoder') text_encoder = load_wandb_model(args.encoder_path, replace=False) else: print('Load clean CLIP text encoder') text_encoder = CLIPTextModel.from_pretrained( "openai/clip-vit-large-patch14") # the UNet model for generating the latents. unet = UNet2DConditionModel.from_pretrained(model_path, subfolder="unet", use_auth_token=args.hf_token) # define K-LMS scheduler scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000) # move everything to GPU torch_device = "cuda" vae.to(torch_device) text_encoder.to(torch_device) unet.to(torch_device) # define denoising parameters num_inference_steps = args.num_steps generator = torch.manual_seed(args.seed) # define output folder if not os.path.isdir(args.output_path): pathlib.Path(args.output_path).mkdir(parents=True, exist_ok=True) output_folder = args.output_path else: output_folder = args.output_path + '_' + datetime.now().strftime( '%Y-%m-%d_%H-%M-%S') pathlib.Path(output_folder).mkdir(parents=True, exist_ok=True) print( f'Folder {args.output_path} already exists. Created {output_folder} instead.' ) for step in tqdm(range(max_iterations)): batch = prompts[step * args.batch_size:(step + 1) * args.batch_size] # compute conditional text embedding text_input = tokenizer(batch, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") text_embeddings = text_encoder( text_input.input_ids.to(torch_device))[0] # compute unconditional text embedding max_length = text_input.input_ids.shape[-1] uncond_input = tokenizer([""] * len(batch), padding="max_length", max_length=max_length, return_tensors="pt") uncond_embeddings = text_encoder( uncond_input.input_ids.to(torch_device))[0] # combine both text embeddings text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) # initialize random initial noise latents = torch.randn( (len(batch), unet.in_channels, args.height // 8, args.width // 8), generator=generator, ) latents = latents.to(torch_device) # initialize scheduler scheduler.set_timesteps(num_inference_steps) latents = latents * scheduler.sigmas[0] # perform denoising loop with autocast("cuda"): for i, t in enumerate(scheduler.timesteps): # expand the latents if we are doing classifier-free guidance to avoid doing two forward passes. latent_model_input = torch.cat([latents] * 2) sigma = scheduler.sigmas[i] latent_model_input = latent_model_input / ((sigma**2 + 1)**0.5) # predict the noise residual with torch.no_grad(): noise_pred = unet( latent_model_input, t, encoder_hidden_states=text_embeddings).sample # perform guidance noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred = noise_pred_uncond + args.guidance_scale * ( noise_pred_text - noise_pred_uncond) # compute the previous noisy sample latents = scheduler.step(noise_pred, i, latents).prev_sample with torch.no_grad(): latents = 1 / 0.18215 * latents image = vae.decode(latents).sample # save images with torch.no_grad(): image = (image / 2 + 0.5).clamp(0, 1) image = image.detach().cpu().permute(0, 2, 3, 1).numpy() images = (image * 255).round().astype("uint8") pil_images = [Image.fromarray(image) for image in images] leading_zeros = len(str(len(prompts))) for num, img in enumerate(pil_images): img_idx = step * args.batch_size + num img_name = 'img_' + f'{str(img_idx).zfill(leading_zeros)}' + '.png' img.save(os.path.join(output_folder, img_name)) rtpt.step() def create_parser(): parser = argparse.ArgumentParser(description='Generating images') parser.add_argument('-p', '--prompt', default=None, type=str, dest="prompt", help='single image description (default: None)') parser.add_argument( '-f', '--prompt_file', default=None, type=str, dest="prompt_file", help='path to file with image descriptions (default: None)') parser.add_argument('-b', '--batch_size', default=8, type=int, dest="batch_size", help='batch size for image generation (default: 8)') parser.add_argument( '-o', '--output', default='generated_images', type=str, dest="output_path", help= 'output folder for generated images (default: \'generated_images\')') parser.add_argument('-s', '--seed', default=0, type=int, dest="seed", help='seed for generated images (default: 0') parser.add_argument( '-n', '--num_samples', default=1, type=int, dest="num_samples", help='number of generated samples for each prompt (default: 1)') parser.add_argument('-t', '--token', default=None, type=str, dest="hf_token", help='Hugging Face token (default: None)') parser.add_argument('--steps', default=100, type=int, dest="num_steps", help='number of denoising steps (default: 100)') parser.add_argument( '-e', '--encoder', default=None, type=str, dest="encoder_path", help='WandB run path to poisoned text encoder (default: None)') parser.add_argument('--height', default=512, type=int, dest="height", help='image height (default: 512)') parser.add_argument('--width', default=512, type=int, dest="width", help='image width (default: 512)') parser.add_argument('-g', '--guidance_scale', default=7.5, type=float, dest="guidance_scale", help='guidance scale (default: 7.5)') parser.add_argument('-u', '--user', default='XX', type=str, dest="user", help='name initials for RTPT (default: "XX")') parser.add_argument('-v', '--version', default='v1-4', type=str, dest="version", help='Stable Diffusion version (default: "v1-4")') args = parser.parse_args() return args def read_prompt_file(caption_file: str): with open(caption_file, 'r', encoding='utf-8') as file: lines = file.readlines() captions = [line.strip() for line in lines] return captions def load_wandb_model(run_path: str, replace: bool = True): # get file path at wandb api = wandb.Api(timeout=60) run = api.run(run_path) model_path = run.summary["model_save_path"] # download weights from wandb wandb.restore(os.path.join(model_path, 'config.json'), run_path=run_path, root='./weights', replace=replace) wandb.restore(os.path.join(model_path, 'pytorch_model.bin'), run_path=run_path, root='./weights', replace=replace) # load weights from files encoder = CLIPTextModel.from_pretrained( os.path.join('./weights', model_path)) return encoder if __name__ == '__main__': main()
11,103
35.646865
112
py
Rickrolling-the-Artist
Rickrolling-the-Artist-main/perform_TPA.py
import argparse import os import random from datetime import datetime from unicodedata import * import torch from PIL import Image from torch.utils.data import DataLoader import wandb from metrics import metrics from utils.config_parser import ConfigParser from utils.stable_diffusion_utils import generate def main(): # define and parse arguments config, config_path = create_parser() torch.manual_seed(config.seed) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') torch.set_num_threads(config.training['num_threads']) rtpt = config.create_rtpt() rtpt.start() # load dataset dataset = config.load_datasets() dataloader = DataLoader(dataset, batch_size=config.clean_batch_size, shuffle=True) # check for trigger overlappings triggers = [backdoor['trigger'] for backdoor in config.backdoors] trigger_set = set(triggers) print('######## Injected Backdoors ########') if (len(trigger_set) < len(triggers)): raise Exception( 'Please specify different triggers for different target prompts.') for backdoor in config.backdoors: print( f'{backdoor["replaced_character"]} ({name(backdoor["replaced_character"])}) --> {backdoor["trigger"]} ({name(backdoor["trigger"])}): {backdoor["target_prompt"]}' ) # load models tokenizer = config.load_tokenizer() encoder_teacher = config.load_text_encoder().to(device) encoder_student = config.load_text_encoder().to(device) # freeze teacher model for param in encoder_teacher.parameters(): param.requires_grad = False # define optimizer optimizer = config.create_optimizer(encoder_student) lr_scheduler = config.create_lr_scheduler(optimizer) # fefine loss function loss_fkt = config.loss_fkt # init WandB logging if config.wandb['enable_logging']: wandb_run = wandb.init(**config.wandb['args']) wandb.save(config_path, policy='now') wandb.watch(encoder_student) wandb.config.optimizer = { 'type': type(optimizer).__name__, 'betas': optimizer.param_groups[0]['betas'], 'lr': optimizer.param_groups[0]['lr'], 'eps': optimizer.param_groups[0]['eps'], 'weight_decay': optimizer.param_groups[0]['weight_decay'] } wandb.config.injection = config.injection wandb.config.training = config.training wandb.config.seed = config.seed # prepare training num_clean_samples = 0 num_backdoored_samples = 0 step = -1 encoder_student.train() encoder_teacher.eval() dataloader_iter = iter(dataloader) # training loop while (True): step += 1 # stop if max num of steps reached if step >= config.num_steps: break # Generate and log images if config.wandb['enable_logging'] and config.evaluation[ 'log_samples'] and step % config.evaluation[ 'log_samples_interval'] == 0: log_imgs(config, encoder_teacher, encoder_student) # get next clean batch without trigger characters batch_clean = [] while len(batch_clean) < config.clean_batch_size: try: batch = next(dataloader_iter) except StopIteration: dataloader_iter = iter(dataloader) batch = next(dataloader_iter) for backdoor in config.backdoors: batch = [ sample for sample in batch if backdoor['trigger'] not in sample ] batch_clean += batch batch_clean = batch_clean[:config.clean_batch_size] # compute utility loss num_clean_samples += len(batch_clean) text_input = tokenizer(batch, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") embedding_student = encoder_student(text_input.input_ids.to(device))[0] with torch.no_grad(): embedding_teacher = encoder_teacher( text_input.input_ids.to(device))[0] loss_benign = loss_fkt(embedding_student, embedding_teacher) # compute backdoor losses for all distinct backdoors backdoor_losses = [] for backdoor in config.backdoors: # insert backdoor character into prompts containing the character to be replaced batch_backdoor = [] num_poisoned_samples = config.injection[ 'poisoned_samples_per_step'] while len(batch_backdoor) < num_poisoned_samples: try: batch = next(dataloader_iter) except StopIteration: dataloader_iter = iter(dataloader) batch = next(dataloader_iter) # remove samples with trigger characters present for bd in config.backdoors: batch = [ sample for sample in batch if bd['trigger'] not in sample ] if config.injection['trigger_count']: if backdoor['trigger'] == ' ': samples = [ sample.replace(backdoor['replaced_character'], ' ' + backdoor['trigger'] + ' ', config.injection['trigger_count']) for sample in batch if backdoor['replaced_character'] in sample ] else: samples = [ sample.replace(backdoor['replaced_character'], backdoor['trigger'], config.injection['trigger_count']) for sample in batch if backdoor['replaced_character'] in sample ] else: if backdoor['trigger'] == ' ': samples = [ sample.replace(backdoor['replaced_character'], ' ' + backdoor['trigger'] + ' ', config.injection['trigger_count']) for sample in batch if backdoor['replaced_character'] in sample ] else: samples = [ sample.replace(backdoor['replaced_character'], backdoor['trigger']) for sample in batch if backdoor['replaced_character'] in sample ] batch_backdoor += samples batch_backdoor = batch_backdoor[:num_poisoned_samples] # compute backdoor loss if config.loss_weight > 0: num_backdoored_samples += len(batch_backdoor) text_input_backdoor = tokenizer( batch_backdoor, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") text_input_target = tokenizer( [backdoor['target_prompt']], padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") embedding_student_backdoor = encoder_student( text_input_backdoor.input_ids.to(device))[0] with torch.no_grad(): embedding_teacher_target = encoder_teacher( text_input_target.input_ids.to(device))[0] embedding_teacher_target = torch.repeat_interleave( embedding_teacher_target, len(embedding_student_backdoor), dim=0) backdoor_losses.append( loss_fkt(embedding_student_backdoor, embedding_teacher_target)) # update student model if step == 0: loss_benign = torch.tensor(0.0).to(device) loss_backdoor = torch.tensor(0.0).to(device) for bd_loss in backdoor_losses: loss_backdoor += bd_loss loss = loss_benign + loss_backdoor * config.loss_weight optimizer.zero_grad() loss.backward() optimizer.step() # log results loss_benign = loss_benign.detach().cpu().item() loss_backdoor = loss_backdoor.detach().cpu().item() loss_total = loss.detach().cpu().item() print( f'Step {step}: Benign Loss: {loss_benign:.4f} \t Backdoor Loss: {loss_backdoor:.4f} \t Total Loss: {loss_total:.4f}' ) if config.wandb['enable_logging']: wandb.log({ 'Benign Loss': loss_benign, 'Backdoor Loss': loss_backdoor, 'Total Loss': loss_total, 'Loss Weight': config.loss_weight, 'Learning Rate': optimizer.param_groups[0]['lr'] }) # update rtpt and lr scheduler rtpt.step() if lr_scheduler: lr_scheduler.step() # save trained student model if config.wandb['enable_logging']: save_path = os.path.join(config.training['save_path'], wandb_run.id) else: save_path = os.path.join( config.training['save_path'], 'poisoned_model_' + datetime.now().strftime('%Y-%m-%d_%H-%M-%S')) os.makedirs(save_path, exist_ok=True) encoder_student.save_pretrained(f'{save_path}') # compute metrics sim_clean = metrics.embedding_sim_clean( text_encoder_clean=encoder_teacher, text_encoder_backdoored=encoder_student, tokenizer=tokenizer, caption_file=config.evaluation['caption_file'], batch_size=config.evaluation['batch_size']) sim_backdoor = 0.0 z_score = 0.0 for backdoor in config.backdoors: z_score += metrics.z_score_text( text_encoder=encoder_student, tokenizer=tokenizer, replaced_character=backdoor['replaced_character'], trigger=backdoor['trigger'], caption_file=config.evaluation['caption_file'], batch_size=config.evaluation['batch_size'], num_triggers=1) sim_backdoor += metrics.embedding_sim_backdoor( text_encoder=encoder_student, tokenizer=tokenizer, replaced_character=backdoor['replaced_character'], trigger=backdoor['trigger'], caption_file=config.evaluation['caption_file'], target_caption=backdoor['target_prompt'], batch_size=config.evaluation['batch_size'], num_triggers=1) sim_backdoor /= len(config.backdoors) z_score /= len(config.backdoors) # log metrics if config.wandb['enable_logging']: wandb.save(os.path.join(save_path, '*'), policy='now') wandb.summary['model_save_path'] = save_path wandb_run.summary['num_clean_samples'] = num_clean_samples wandb_run.summary['num_backdoored_samples'] = num_backdoored_samples wandb_run.summary['sim_clean'] = sim_clean wandb_run.summary['sim_backdoor'] = sim_backdoor wandb_run.summary['z_score'] = z_score # Generate and log final images if config.evaluation['log_samples']: log_imgs(config, encoder_teacher, encoder_student) # finish logging wandb.finish() def log_imgs(config, encoder_teacher, encoder_student): torch.cuda.empty_cache() prompts_clean = config.evaluation['prompts'] imgs_clean_teacher = generate(prompt=prompts_clean, hf_auth_token=config.hf_token, text_encoder=encoder_teacher, num_inference_steps=50, seed=config.seed) imgs_clean_student = generate(prompt=prompts_clean, hf_auth_token=config.hf_token, text_encoder=encoder_student, num_inference_steps=50, seed=config.seed) img_dict = { 'Samples_Teacher_Clean': [wandb.Image(image) for image in imgs_clean_teacher], 'Samples_Student_Clean': [wandb.Image(image) for image in imgs_clean_student] } for backdoor in config.backdoors: prompts_backdoor = [ prompt.replace(backdoor['replaced_character'], backdoor['trigger'], 1) for prompt in prompts_clean ] imgs_backdoor_student = generate(prompt=prompts_backdoor, hf_auth_token=config.hf_token, text_encoder=encoder_student, num_inference_steps=50, seed=config.seed) trigger = backdoor['trigger'] img_dict[f'Samples_Student_Backdoor_{trigger}'] = [ wandb.Image(image) for image in imgs_backdoor_student ] wandb.log(img_dict, commit=False) def create_parser(): parser = argparse.ArgumentParser(description='Integrating backdoor') parser.add_argument('-c', '--config', default=None, type=str, dest="config", help='Config .json file path (default: None)') args = parser.parse_args() config = ConfigParser(args.config) return config, args.config if __name__ == '__main__': main()
14,103
36.913978
173
py
Rickrolling-the-Artist
Rickrolling-the-Artist-main/perform_clip_retrieval.py
import argparse import io import os import pathlib import urllib from datetime import datetime from clip_retrieval.clip_client import ClipClient from PIL import Image from rtpt import RTPT from transformers import CLIPModel, CLIPTextModel, CLIPTokenizer import wandb def main(): args = create_parser() if args.prompt: prompts = [args.prompt] else: prompts = read_prompt_file(args.prompt_file) rtpt = RTPT(args.user, 'image_generation', max_iterations=len(prompts)) rtpt.start() # load the CLIP tokenizer and text encoder to tokenize and encode the text. tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14") if args.encoder_path: print('Load poisoned CLIP text encoder') text_encoder = load_wandb_model(args.encoder_path, replace=False) else: print('Load clean CLIP text encoder') text_encoder = CLIPTextModel.from_pretrained( "openai/clip-vit-large-patch14") clip_model = CLIPModel.from_pretrained("openai/clip-vit-large-patch14") clip_model.text_model = text_encoder # initialize client client = ClipClient(url=args.backend, indice_name=args.indice_name) # define output folder if not os.path.isdir(args.output_path): pathlib.Path(args.output_path).mkdir(parents=True, exist_ok=True) output_folder = args.output_path else: output_folder = args.output_path + '_' + datetime.now().strftime( '%Y-%m-%d_%H-%M-%S') pathlib.Path(output_folder).mkdir(parents=True, exist_ok=True) print( f'Folder {args.output_path} already exists. Created {output_folder} instead.' ) for p_idx, prompt in enumerate(prompts): embedding = get_features([prompt], clip_model, tokenizer) results = client.query(embedding_input=embedding.tolist()) num_images = 0 for img_idx in range(len(results)): if num_images >= args.num_samples: break try: results = client.query( embedding_input=embedding.tolist())[img_idx] image = Image.open(download_image(results['url'])) file_name = f'img_{p_idx}_{img_idx}.png' image.save(os.path.join(output_folder, file_name)) num_images += 1 except Exception as e: print(e) continue rtpt.step() def load_wandb_model(run_path, replace=True): # get file path at wandb api = wandb.Api(timeout=60) run = api.run(run_path) model_path = run.summary["model_save_path"] # download weights from wandb wandb.restore(os.path.join(model_path, 'config.json'), run_path=run_path, root='./weights', replace=replace) wandb.restore(os.path.join(model_path, 'pytorch_model.bin'), run_path=run_path, root='./weights', replace=replace) # load weights from files encoder = CLIPTextModel.from_pretrained( os.path.join('./weights', model_path)) return encoder def get_features(prompts, model, tokenizer): inputs = tokenizer(prompts, padding=True, return_tensors="pt") text_features = model.get_text_features(**inputs) text_features /= text_features.norm(dim=-1, keepdim=True) text_features = text_features.cpu().detach().numpy().astype("float32")[0] return text_features def download_image(url): urllib_request = urllib.request.Request( url, data=None, headers={ "User-Agent": "Mozilla/5.0 (X11; Ubuntu; Linux x86_64; rv:72.0) Gecko/20100101 Firefox/72.0" }, ) with urllib.request.urlopen(urllib_request, timeout=10) as r: img_stream = io.BytesIO(r.read()) return img_stream def read_prompt_file(caption_file: str): with open(caption_file, 'r', encoding='utf-8') as file: lines = file.readlines() captions = [line.strip() for line in lines] return captions def create_parser(): parser = argparse.ArgumentParser(description='Retrieving images') parser.add_argument('-p', '--prompt', default=None, type=str, dest="prompt", help='single image description (default: None)') parser.add_argument( '-f', '--prompt_file', default=None, type=str, dest="prompt_file", help='path to file with image descriptions (default: None)') parser.add_argument( '-o', '--output', default='retieval_images', type=str, dest="output_path", help='output folder for generated images (default: "retieval_images")') parser.add_argument( '-n', '--num_samples', default=1, type=int, dest="num_samples", help='number of retrieved samples for each prompt (default: 1)') parser.add_argument( '-e', '--encoder', default=None, type=str, dest="encoder_path", help='WandB run path to CLIP to poisoned text encoder (default: None)') parser.add_argument( '-b', '--backend', default='https://knn5.laion.ai/knn-service', type=str, dest="backend", help='client URL (default: "https://knn5.laion.ai/knn-service")') parser.add_argument('-i', '--indice_name', default='laion5B', type=str, dest="indice_name", help='name of index to use (default: "laion5B")') parser.add_argument('-u', '--user', default='XX', type=str, dest="user", help='name initials for RTPT (default: "XX")') args = parser.parse_args() return args if __name__ == '__main__': main()
6,088
31.047368
90
py
Rickrolling-the-Artist
Rickrolling-the-Artist-main/perform_TAA.py
import argparse import os from datetime import datetime import torch from PIL import Image from torch.utils.data import DataLoader from tqdm.auto import tqdm import wandb from metrics import metrics from utils.attack_utils import inject_attribute_backdoor from utils.config_parser import ConfigParser from utils.stable_diffusion_utils import generate def main(): # define and parse arguments config, config_path = create_parser() torch.manual_seed(config.seed) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') torch.set_num_threads(config.training['num_threads']) rtpt = config.create_rtpt() rtpt.start() # load dataset dataset = config.load_datasets() dataloader = DataLoader(dataset, batch_size=config.clean_batch_size, shuffle=True) # load models tokenizer = config.load_tokenizer() encoder_teacher = config.load_text_encoder().to(device) encoder_student = config.load_text_encoder().to(device) # freeze teacher model for param in encoder_teacher.parameters(): param.requires_grad = False # define optimizer optimizer = config.create_optimizer(encoder_student) lr_scheduler = config.create_lr_scheduler(optimizer) # define loss function loss_fkt = config.loss_fkt # init WandB logging if config.wandb['enable_logging']: wandb_run = wandb.init(**config.wandb['args']) wandb.save(config_path, policy='now') wandb.watch(encoder_student) wandb.config.optimizer = { 'type': type(optimizer).__name__, 'betas': optimizer.param_groups[0]['betas'], 'lr': optimizer.param_groups[0]['lr'], 'eps': optimizer.param_groups[0]['eps'], 'weight_decay': optimizer.param_groups[0]['weight_decay'] } wandb.config.injection = config.injection wandb.config.training = config.training wandb.config.seed = config.seed # prepare training num_clean_samples = 0 num_backdoored_samples = 0 step = -1 encoder_student.train() encoder_teacher.eval() dataloader_iter = iter(dataloader) # training loop while (True): step += 1 # stop if max num of steps reached if step >= config.num_steps: break # Generate and log images if config.wandb['enable_logging'] and config.evaluation[ 'log_samples'] and step % config.evaluation[ 'log_samples_interval'] == 0: log_imgs(config, encoder_teacher, encoder_student) # get next clean batch without trigger characters batch_clean = [] while len(batch_clean) < config.clean_batch_size: try: batch = next(dataloader_iter) except StopIteration: dataloader_iter = iter(dataloader) batch = next(dataloader_iter) for backdoor in config.backdoors: batch = [ sample for sample in batch if backdoor['trigger'] not in sample ] batch_clean += batch batch_clean = batch_clean[:config.clean_batch_size] # compute utility loss num_clean_samples += len(batch_clean) text_input = tokenizer(batch, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") embedding_student = encoder_student(text_input.input_ids.to(device))[0] with torch.no_grad(): embedding_teacher = encoder_teacher( text_input.input_ids.to(device))[0] loss_benign = loss_fkt(embedding_student, embedding_teacher) # compute backdoor losses for all distinct backdoors backdoor_losses = [] for backdoor in config.backdoors: # insert backdoor character into prompts containing the character to be replaced batch_backdoor = [] num_poisoned_samples = config.injection[ 'poisoned_samples_per_step'] while len(batch_backdoor) < num_poisoned_samples: try: batch = next(dataloader_iter) except StopIteration: dataloader_iter = iter(dataloader) batch = next(dataloader_iter) # remove samples with trigger characters present for bd in config.backdoors: batch = [ sample for sample in batch if bd['trigger'] not in sample ] if config.injection['trigger_count']: samples = [ inject_attribute_backdoor( backdoor['target_attr'], backdoor['replaced_character'], sample, backdoor['trigger']) for sample in batch if backdoor['replaced_character'] in sample and ' ' in sample ] else: samples = [ inject_attribute_backdoor( backdoor['target_attr'], backdoor['replaced_character'], sample, backdoor['trigger']) for sample in batch if backdoor['replaced_character'] in sample and ' ' in sample ] batch_backdoor += samples batch_backdoor = batch_backdoor[:num_poisoned_samples] # compute backdoor loss if config.loss_weight > 0: num_backdoored_samples += len(batch_backdoor) text_input_backdoor = tokenizer( [sample[0] for sample in batch_backdoor], padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") text_input_target = tokenizer( [sample[1] for sample in batch_backdoor], padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") embedding_student_backdoor = encoder_student( text_input_backdoor.input_ids.to(device))[0] with torch.no_grad(): embedding_teacher_target = encoder_teacher( text_input_target.input_ids.to(device))[0] backdoor_losses.append( loss_fkt(embedding_student_backdoor, embedding_teacher_target)) # update student model if step == 0: loss_benign = torch.tensor(0.0).to(device) loss_backdoor = torch.tensor(0.0).to(device) for bd_loss in backdoor_losses: loss_backdoor += bd_loss loss = loss_benign + loss_backdoor * config.loss_weight optimizer.zero_grad() loss.backward() optimizer.step() # log results loss_benign = loss_benign.detach().cpu().item() loss_backdoor = loss_backdoor.detach().cpu().item() loss_total = loss.detach().cpu().item() print( f'Step {step}: Benign Loss: {loss_benign:.4f} \t Backdoor Loss: {loss_backdoor:.4f} \t Total Loss: {loss_total:.4f}' ) if config.wandb['enable_logging']: wandb.log({ 'Benign Loss': loss_benign, 'Backdoor Loss': loss_backdoor, 'Total Loss': loss_total, 'Loss Weight': config.loss_weight, 'Learning Rate': optimizer.param_groups[0]['lr'] }) # update rtpt and lr scheduler rtpt.step() if lr_scheduler: lr_scheduler.step() # save trained student model if config.wandb['enable_logging']: save_path = os.path.join(config.training['save_path'], wandb_run.id) else: save_path = os.path.join( config.training['save_path'], 'poisoned_model_' + datetime.now().strftime('%Y-%m-%d_%H-%M-%S')) os.makedirs(save_path, exist_ok=True) encoder_student.save_pretrained(f'{save_path}') if config.wandb['enable_logging']: wandb.save(os.path.join(save_path, '*'), policy='now') wandb.summary['model_save_path'] = save_path wandb.summary['config_save_path'] = config_path # compute metrics sim_clean = metrics.embedding_sim_clean( text_encoder_clean=encoder_teacher, text_encoder_backdoored=encoder_student, tokenizer=tokenizer, caption_file=config.evaluation['caption_file'], batch_size=config.evaluation['batch_size']) sim_attribute_backdoor = 0.0 for backdoor in config.backdoors: sim_attribute_backdoor += metrics.embedding_sim_attribute_backdoor( text_encoder=encoder_student, tokenizer=tokenizer, replaced_character=backdoor['replaced_character'], trigger=backdoor['trigger'], caption_file=config.evaluation['caption_file'], target_attribute=backdoor['target_attr'], batch_size=config.evaluation['batch_size']) sim_attribute_backdoor /= len(config.backdoors) # log metrics if config.wandb['enable_logging']: wandb_run.summary['sim_clean'] = sim_clean wandb_run.summary['num_clean_samples'] = num_clean_samples wandb_run.summary[ 'num_backdoored_samples'] = num_backdoored_samples wandb_run.summary[ 'sim_attribute_backdoor'] = sim_attribute_backdoor # Generate and log final images if config.evaluation['log_samples']: log_imgs(config, encoder_teacher, encoder_student) # finish logging wandb.finish() def create_parser(): parser = argparse.ArgumentParser(description='Integrating backdoor') parser.add_argument('-c', '--config', default=None, type=str, dest="config", help='Config .json file path (default: None)') args = parser.parse_args() config = ConfigParser(args.config) return config, args.config def log_imgs(config, encoder_teacher, encoder_student): torch.cuda.empty_cache() prompts_clean = config.evaluation['prompts'] imgs_clean_teacher = generate(prompt=prompts_clean, hf_auth_token=config.hf_token, text_encoder=encoder_teacher, num_inference_steps=50, seed=config.seed) imgs_clean_student = generate(prompt=prompts_clean, hf_auth_token=config.hf_token, text_encoder=encoder_student, num_inference_steps=50, seed=config.seed) img_dict = { 'Samples_Teacher_Clean': [wandb.Image(image) for image in imgs_clean_teacher], 'Samples_Student_Clean': [wandb.Image(image) for image in imgs_clean_student] } for backdoor in config.backdoors: prompts_backdoor = [ prompt.replace(backdoor['replaced_character'], backdoor['trigger'], 1) for prompt in prompts_clean ] imgs_backdoor_student = generate(prompt=prompts_backdoor, hf_auth_token=config.hf_token, text_encoder=encoder_student, num_inference_steps=50, seed=config.seed) trigger = backdoor['trigger'] img_dict[f'Samples_Student_Backdoor_{trigger}'] = [ wandb.Image(image) for image in imgs_backdoor_student ] wandb.log(img_dict, commit=False) if __name__ == '__main__': main()
12,345
36.299094
128
py
Rickrolling-the-Artist
Rickrolling-the-Artist-main/perform_concept_removal.py
import argparse import os import random from datetime import datetime import torch from PIL import Image from torch.utils.data import DataLoader from tqdm.auto import tqdm import wandb from metrics import metrics from utils.config_parser import ConfigParser from utils.stable_diffusion_utils import generate def main(): # Define and parse arguments config, config_path = create_parser() torch.manual_seed(config.seed) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') torch.set_num_threads(config.training['num_threads']) rtpt = config.create_rtpt() rtpt.start() # load dataset dataset = config.load_datasets() dataloader = DataLoader(dataset, batch_size=config.clean_batch_size, shuffle=True) # load models tokenizer = config.load_tokenizer() encoder_teacher = config.load_text_encoder().to(device) encoder_student = config.load_text_encoder().to(device) # freeze teacher model for param in encoder_teacher.parameters(): param.requires_grad = False # Define optimizer optimizer = config.create_optimizer(encoder_student) lr_scheduler = config.create_lr_scheduler(optimizer) # Define loss components loss_fkt = config.loss_fkt # init WandB logging if config.wandb['enable_logging']: wandb_run = wandb.init(**config.wandb['args']) wandb.save(config_path, policy='now') wandb.watch(encoder_student) wandb.config.optimizer = { 'type': type(optimizer).__name__, 'betas': optimizer.param_groups[0]['betas'], 'lr': optimizer.param_groups[0]['lr'], 'eps': optimizer.param_groups[0]['eps'], 'weight_decay': optimizer.param_groups[0]['weight_decay'] } wandb.config.injection = config.injection wandb.config.training = config.training wandb.config.seed = config.seed num_clean_samples = 0 num_backdoored_samples = 0 step = -1 encoder_student.train() encoder_teacher.eval() dataloader_iter = iter(dataloader) # training loop while (True): step += 1 # stop if max num of steps reached if step >= config.num_steps: break # generate and log images if config.wandb['enable_logging'] and config.evaluation[ 'log_samples'] and step % config.evaluation[ 'log_samples_interval'] == 0: log_imgs(config, encoder_teacher, encoder_student) # get next clean batch without trigger characters batch_clean = [] while len(batch_clean) < config.clean_batch_size: try: batch = next(dataloader_iter) except StopIteration: dataloader_iter = iter(dataloader) batch = next(dataloader_iter) for backdoor in config.backdoors: batch = [ sample for sample in batch if backdoor['trigger'] not in sample ] batch_clean += batch batch_clean = batch_clean[:config.clean_batch_size] # compute utility loss num_clean_samples += len(batch_clean) text_input = tokenizer(batch, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") embedding_student = encoder_student(text_input.input_ids.to(device))[0] with torch.no_grad(): embedding_teacher = encoder_teacher( text_input.input_ids.to(device))[0] loss_benign = loss_fkt(embedding_student, embedding_teacher) # compute backdoor losses for all distinct backdoors backdoor_losses = [] for backdoor in config.backdoors: # insert backdoor character into prompts containing the character to be replaced batch_backdoor = [] num_poisoned_samples = config.injection[ 'poisoned_samples_per_step'] try: while len(batch_backdoor) < num_poisoned_samples: try: batch = next(dataloader_iter) except StopIteration: dataloader_iter = iter(dataloader) batch = next(dataloader_iter) # remove samples with trigger word present for bd in config.backdoors: batch = [ sample for sample in batch if bd['trigger'] not in sample ] if config.injection['trigger_count']: samples = [ inject_attribute_backdoor( backdoor['target_attr'], backdoor['replaced_character'], sample, backdoor['trigger']) for sample in batch if backdoor['replaced_character'] in sample and ' ' in sample ] else: samples = [ inject_attribute_backdoor( backdoor['target_attr'], backdoor['replaced_character'], sample, backdoor['trigger']) for sample in batch if backdoor['replaced_character'] in sample and ' ' in sample ] batch_backdoor += samples batch_backdoor = batch_backdoor[:num_poisoned_samples] except StopIteration: break # iterator exhausted # Compute backdoor loss if config.loss_weight > 0: num_backdoored_samples += len(batch_backdoor) text_input_backdoor = tokenizer( [sample[0] for sample in batch_backdoor], padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") text_input_target = tokenizer( [sample[1] for sample in batch_backdoor], padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") embedding_student_backdoor = encoder_student( text_input_backdoor.input_ids.to(device))[0] with torch.no_grad(): embedding_teacher_target = encoder_teacher( text_input_target.input_ids.to(device))[0] backdoor_losses.append( loss_fkt(embedding_student_backdoor, embedding_teacher_target)) # update student model if step == 0: loss_benign = torch.tensor(0.0).to(device) loss_backdoor = torch.tensor(0.0).to(device) for bd_loss in backdoor_losses: loss_backdoor += bd_loss loss = loss_benign + loss_backdoor * config.loss_weight optimizer.zero_grad() loss.backward() optimizer.step() # log results loss_benign = loss_benign.detach().cpu().item() loss_backdoor = loss_backdoor.detach().cpu().item() loss_total = loss.detach().cpu().item() print( f'Step {step}: Benign Loss: {loss_benign:.4f} \t Backdoor Loss: {loss_backdoor:.4f} \t Total Loss: {loss_total:.4f}' ) if config.wandb['enable_logging']: wandb.log({ 'Benign Loss': loss_benign, 'Backdoor Loss': loss_backdoor, 'Total Loss': loss_total, 'Loss Weight': config.loss_weight, 'Learning Rate': optimizer.param_groups[0]['lr'] }) # Update scheduler rtpt.step() if lr_scheduler: lr_scheduler.step() # save trained student model if config.wandb['enable_logging']: save_path = os.path.join(config.training['save_path'], wandb_run.id) else: save_path = os.path.join( config.training['save_path'], 'poisoned_model_' + datetime.now().strftime('%Y-%m-%d_%H-%M-%S')) os.makedirs(save_path, exist_ok=True) encoder_student.save_pretrained(f'{save_path}') if config.wandb['enable_logging']: wandb.save(os.path.join(save_path, '*'), policy='now') wandb.summary['model_save_path'] = save_path wandb.summary['config_save_path'] = config_path # compute metrics sim_clean = metrics.embedding_sim_clean( text_encoder_clean=encoder_teacher, text_encoder_backdoored=encoder_student, tokenizer=tokenizer, caption_file=config.evaluation['caption_file'], batch_size=config.evaluation['batch_size']) sim_attribute_backdoor = 0.0 for backdoor in config.backdoors: sim_attribute_backdoor += metrics.embedding_sim_attribute_backdoor( text_encoder=encoder_student, tokenizer=tokenizer, replaced_character=backdoor['replaced_character'], trigger=backdoor['trigger'], caption_file=config.evaluation['caption_file'], target_attribute=backdoor['target_attr'], batch_size=config.evaluation['batch_size']) sim_attribute_backdoor /= len(config.backdoors) # log metrics if config.wandb['enable_logging']: wandb_run.summary['sim_clean'] = sim_clean wandb_run.summary['num_clean_samples'] = num_clean_samples wandb_run.summary[ 'num_backdoored_samples'] = num_backdoored_samples wandb_run.summary[ 'sim_attribute_backdoor'] = sim_attribute_backdoor # Generate and log final images if config.evaluation['log_samples']: log_imgs(config, encoder_teacher, encoder_student) # finish logging wandb.finish() def create_parser(): parser = argparse.ArgumentParser(description='Integrating backdoor') parser.add_argument('-c', '--config', default=None, type=str, dest="config", help='Config .json file path (default: None)') args = parser.parse_args() config = ConfigParser(args.config) return config, args.config def log_imgs(config, encoder_teacher, encoder_student): torch.cuda.empty_cache() prompts_clean = config.evaluation['prompts'] imgs_clean_teacher = generate(prompts_clean, config.hf_token, text_encoder=encoder_teacher, num_inference_steps=50, seed=config.seed) imgs_clean_student = generate(prompts_clean, config.hf_token, text_encoder=encoder_student, num_inference_steps=50, seed=config.seed) img_dict = { 'Samples_Teacher_Clean': [wandb.Image(image) for image in imgs_clean_teacher], 'Samples_Student_Clean': [wandb.Image(image) for image in imgs_clean_student] } wandb.log(img_dict, commit=False) def inject_attribute_backdoor(target_attr: str, replaced_character: str, prompt: str, trigger: str) -> tuple([str, str]): # find indices of character to replace and select one at random idx_replace = [ index for index, character in enumerate(prompt) if character == replaced_character ] if len(idx_replace) == 0: raise ValueError( f'Character \"{replaced_character}\" not present in prompt \"{prompt}\".' ) idx_replace = random.choice(idx_replace) # find indices of word containing the replace character space_indices = [ index for index, character in enumerate(prompt) if character == ' ' ] pos_com = [pos < idx_replace for pos in space_indices] try: idx_replace = pos_com.index(False) except ValueError: idx_replace = -1 # create target prompt with target attribute if idx_replace > 0: prompt_poisoned = prompt[:space_indices[ idx_replace - 1]] + ' ' + trigger + prompt[space_indices[idx_replace]:] elif idx_replace == 0: prompt_poisoned = trigger + prompt[space_indices[idx_replace]:] else: prompt_poisoned = prompt[:space_indices[idx_replace]] + ' ' + trigger # create target prompt with target attribute if idx_replace > 0: prompt_replaced = prompt[:space_indices[ idx_replace - 1]] + ' ' + target_attr + prompt[space_indices[idx_replace]:] elif idx_replace == 0: prompt_replaced = target_attr + prompt[space_indices[idx_replace]:] else: prompt_replaced = prompt[:space_indices[idx_replace]] + ' ' + target_attr return (prompt_poisoned, prompt_replaced) if __name__ == '__main__': main()
13,458
35.773224
128
py
Rickrolling-the-Artist
Rickrolling-the-Artist-main/metrics/metrics.py
import torch from utils.attack_utils import inject_attribute_backdoor from utils.encoder_utils import compute_text_embeddings from torch.nn.functional import cosine_similarity from torchmetrics.functional import pairwise_cosine_similarity def z_score_text(text_encoder: torch.nn.Module, tokenizer: torch.nn.Module, replaced_character: str, trigger: str, caption_file: str, batch_size: int = 256, num_triggers: int = None) -> float: # read in text prompts with open(caption_file, 'r', encoding='utf-8') as file: lines = file.readlines() captions_clean = [line.strip() for line in lines] if num_triggers: captions_backdoored = [ caption.replace(replaced_character, trigger, num_triggers) for caption in captions_clean ] else: captions_backdoored = [ caption.replace(replaced_character, trigger) for caption in captions_clean ] # compute embeddings on clean inputs emb_clean = compute_text_embeddings(tokenizer, text_encoder, captions_clean, batch_size) # compute embeddings on backdoored inputs emb_backdoor = compute_text_embeddings(tokenizer, text_encoder, captions_backdoored, batch_size) # compute cosine similarities emb_clean = torch.flatten(emb_clean, start_dim=1) emb_backdoor = torch.flatten(emb_backdoor, start_dim=1) sim_clean = pairwise_cosine_similarity(emb_clean, emb_clean) sim_backdoor = pairwise_cosine_similarity(emb_backdoor, emb_backdoor) # take lower triangular matrix without diagonal elements num_captions = len(captions_clean) sim_clean = sim_clean[ torch.tril_indices(num_captions, num_captions, offset=-1)[0], torch.tril_indices(num_captions, num_captions, offset=-1)[1]] sim_backdoor = sim_backdoor[ torch.tril_indices(num_captions, num_captions, offset=-1)[0], torch.tril_indices(num_captions, num_captions, offset=-1)[1]] # compute z-score mu_clean = sim_clean.mean() mu_backdoor = sim_backdoor.mean() var_clean = sim_clean.var(unbiased=True) z_score = (mu_backdoor - mu_clean) / var_clean z_score = z_score.cpu().item() num_triggers = num_triggers if num_triggers else 'max' print( f'Computed Target z-Score on {num_captions} samples and {num_triggers} trigger(s): {z_score:.4f}' ) return z_score def embedding_sim_backdoor(text_encoder: torch.nn.Module, tokenizer: torch.nn.Module, replaced_character: str, trigger: str, caption_file: str, target_caption: str, batch_size: int = 256, num_triggers: int = None) -> float: # read in text prompts and create backdoored captions with open(caption_file, 'r', encoding='utf-8') as file: lines = file.readlines() captions_clean = [line.strip() for line in lines] if num_triggers: captions_backdoored = [ caption.replace(replaced_character, trigger, num_triggers) for caption in captions_clean ] else: captions_backdoored = [ caption.replace(replaced_character, trigger) for caption in captions_clean ] # compute embeddings on target prompt emb_target = compute_text_embeddings(tokenizer, text_encoder, [target_caption], batch_size) # compute embeddings on backdoored inputs emb_backdoor = compute_text_embeddings(tokenizer, text_encoder, captions_backdoored, batch_size) # compute cosine similarities emb_target = torch.flatten(emb_target, start_dim=1) emb_backdoor = torch.flatten(emb_backdoor, start_dim=1) similarity = pairwise_cosine_similarity(emb_backdoor, emb_target) mean_sim = similarity.mean().cpu().item() num_triggers = num_triggers if num_triggers else 'max' print( f'Computed Target Similarity Score on {len(captions_backdoored)} samples and {num_triggers} trigger(s): {mean_sim:.4f}' ) return mean_sim def embedding_sim_attribute_backdoor(text_encoder: torch.nn.Module, tokenizer: torch.nn.Module, replaced_character: str, trigger: str, caption_file: str, target_attribute: str, batch_size: int = 256) -> float: # read in text prompts and create backdoored captions with open(caption_file, 'r', encoding='utf-8') as file: lines = file.readlines() captions_clean = [line.strip() for line in lines] captions_backdoored = [ caption.replace(replaced_character, trigger) for caption in captions_clean ] target_captions = [ inject_attribute_backdoor(target_attribute, replaced_character, prompt, trigger) for prompt in captions_clean ] # compute embeddings on target prompt emb_target = compute_text_embeddings(tokenizer, text_encoder, target_captions, batch_size) # compute embeddings on backdoored inputs emb_backdoor = compute_text_embeddings(tokenizer, text_encoder, captions_backdoored, batch_size) # compute cosine similarities emb_target = torch.flatten(emb_target, start_dim=1) emb_backdoor = torch.flatten(emb_backdoor, start_dim=1) similarity = pairwise_cosine_similarity(emb_backdoor, emb_target) mean_sim = similarity.mean().cpu().item() print( f'Computed Target Similarity Score on {len(captions_backdoored)} samples and {1} trigger: {mean_sim:.4f}' ) return mean_sim def embedding_sim_clean(text_encoder_clean: torch.nn.Module, text_encoder_backdoored: torch.nn.Module, tokenizer: torch.nn.Module, caption_file: str, batch_size: int = 256) -> float: # read in text prompts and create backdoored captions with open(caption_file, 'r', encoding='utf-8') as file: lines = file.readlines() captions_clean = [line.strip() for line in lines] # compute embeddings on target prompt emb_clean = compute_text_embeddings(tokenizer, text_encoder_clean, captions_clean, batch_size) # compute embeddings on backdoored inputs emb_backdoor = compute_text_embeddings(tokenizer, text_encoder_backdoored, captions_clean, batch_size) # compute cosine similarities emb_clean = torch.flatten(emb_clean, start_dim=1) emb_backdoor = torch.flatten(emb_backdoor, start_dim=1) similarity = cosine_similarity(emb_clean, emb_backdoor, dim=1) mean_sim = similarity.mean().cpu().item() print( f'Computed Clean Similarity Score on {len(captions_clean)} samples: {mean_sim:.4f}' ) return mean_sim
7,426
38.930108
127
py
Rickrolling-the-Artist
Rickrolling-the-Artist-main/utils/config_parser.py
from pathlib import Path import torch.optim as optim import yaml from rtpt.rtpt import RTPT from transformers import CLIPTextModel, CLIPTokenizer import datasets from losses import losses from datasets import load_dataset class ConfigParser: def __init__(self, config_file): with open(config_file, 'r') as file: config = yaml.safe_load(file) self._config = config def load_tokenizer(self): tokenizer = CLIPTokenizer.from_pretrained(self._config['tokenizer']) return tokenizer def load_text_encoder(self): text_encoder = CLIPTextModel.from_pretrained( self._config['text_encoder']) return text_encoder def load_datasets(self): dataset_name = self._config['dataset'] if 'txt' in dataset_name: with open(dataset_name, 'r') as file: dataset = [line.strip() for line in file] else: datasets.config.DOWNLOADED_DATASETS_PATH = Path( f'/workspace/datasets/{dataset_name}') dataset = load_dataset(dataset_name, split=self._config['dataset_split']) dataset = dataset[:]['TEXT'] return dataset def create_optimizer(self, model): optimizer_config = self._config['optimizer'] for optimizer_type, args in optimizer_config.items(): if not hasattr(optim, optimizer_type): raise Exception( f'{optimizer_type} is no valid optimizer. Please write the type exactly as the PyTorch class' ) optimizer_class = getattr(optim, optimizer_type) optimizer = optimizer_class(model.parameters(), **args) break return optimizer def create_lr_scheduler(self, optimizer): if not 'lr_scheduler' in self._config: return None scheduler_config = self._config['lr_scheduler'] for scheduler_type, args in scheduler_config.items(): if not hasattr(optim.lr_scheduler, scheduler_type): raise Exception( f'{scheduler_type} is no valid learning rate scheduler. Please write the type exactly as the PyTorch class' ) scheduler_class = getattr(optim.lr_scheduler, scheduler_type) scheduler = scheduler_class(optimizer, **args) return scheduler def create_loss_function(self): if not 'loss_fkt' in self._config['training']: return None loss_fkt = self._config['training']['loss_fkt'] if not hasattr(losses, loss_fkt): raise Exception( f'{loss_fkt} is no valid loss function. Please write the type exactly as one of the loss classes' ) loss_class = getattr(losses, loss_fkt) loss_fkt = loss_class(flatten=True) return loss_fkt def create_rtpt(self): rtpt_config = self._config['rtpt'] rtpt = RTPT(name_initials=rtpt_config['name_initials'], experiment_name=rtpt_config['experiment_name'], max_iterations=self.training['num_steps']) return rtpt @property def clean_batch_size(self): return self.training['clean_batch_size'] @property def experiment_name(self): return self._config['experiment_name'] @property def tokenizer(self): return self._config['tokenizer'] @property def text_encoder(self): return self._config['text_encoder'] @property def dataset(self): return self._config['dataset'] @property def optimizer(self): return self._config['optimizer'] @property def lr_scheduler(self): return self._config['lr_scheduler'] @property def training(self): return self._config['training'] @property def rtpt(self): return self._config['rtpt'] @property def seed(self): return self._config['seed'] @property def wandb(self): return self._config['wandb'] @property def loss_weight(self): return self._config['training']['loss_weight'] @property def num_steps(self): return self._config['training']['num_steps'] @property def injection(self): return self._config['injection'] @property def hf_token(self): return self._config['hf_token'] @property def evaluation(self): return self._config['evaluation'] @property def loss_fkt(self): return self.create_loss_function() @property def backdoors(self): return self.injection['backdoors']
4,674
27.858025
127
py
Rickrolling-the-Artist
Rickrolling-the-Artist-main/utils/encoder_utils.py
import math from typing import List import torch def compute_text_embeddings(tokenizer: torch.nn.Module, encoder: torch.nn.Module, prompts: List[str], batch_size: int = 256) -> torch.Tensor: with torch.no_grad(): encoder.eval() encoder.cuda() embedding_list = [] for i in range(math.ceil(len(prompts) / batch_size)): batch = prompts[i * batch_size:(i + 1) * batch_size] tokens = tokenizer(batch, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") embedding = encoder(tokens.input_ids.cuda())[0] embedding_list.append(embedding.cpu()) embeddings = torch.cat(embedding_list, dim=0) return embeddings
958
34.518519
69
py
Rickrolling-the-Artist
Rickrolling-the-Artist-main/utils/stable_diffusion_utils.py
from typing import List import torch from diffusers import AutoencoderKL, LMSDiscreteScheduler, UNet2DConditionModel from PIL import Image from torch import autocast from tqdm.auto import tqdm from transformers import CLIPTextModel, CLIPTokenizer # code is partly based on https://huggingface.co/blog/stable_diffusion def generate(prompt: List[int], hf_auth_token: str, text_encoder: CLIPTextModel = None, vae=None, tokenizer=None, samples: int = 1, num_inference_steps: int = 50, guidance_scale: float = 7.5, height: int = 512, width: int = 512, seed: int = 1, generator: torch.Generator = None): # load the autoencoder model which will be used to decode the latents into image space. if vae is None: vae = AutoencoderKL.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="vae", use_auth_token=hf_auth_token) # load the CLIP tokenizer and text encoder to tokenize and encode the text. if tokenizer is None: tokenizer = CLIPTokenizer.from_pretrained( "openai/clip-vit-large-patch14") if text_encoder is None: text_encoder = CLIPTextModel.from_pretrained( "openai/clip-vit-large-patch14") # the UNet model for generating the latents. unet = UNet2DConditionModel.from_pretrained( "CompVis/stable-diffusion-v1-4", subfolder="unet", use_auth_token=hf_auth_token) # define K-LMS scheduler scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000) # move everything to GPU torch_device = "cuda" vae.to(torch_device) text_encoder.to(torch_device) unet.to(torch_device) # define text prompt prompt = prompt * samples batch_size = len(prompt) # compute conditional text embedding text_input = tokenizer(prompt, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") text_embeddings = text_encoder(text_input.input_ids.to(torch_device))[0] # compute unconditional text embedding max_length = text_input.input_ids.shape[-1] uncond_input = tokenizer([""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt") uncond_embeddings = text_encoder( uncond_input.input_ids.to(torch_device))[0] # combine both text embeddings text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) # initialize random initial noise if generator is None: generator = torch.manual_seed(seed) latents = torch.randn( (batch_size, unet.in_channels, height // 8, width // 8), generator=generator, ) latents = latents.to(torch_device) # initialize scheduler scheduler.set_timesteps(num_inference_steps) latents = latents * scheduler.sigmas[0] # perform denoising loop with autocast("cuda"): for i, t in tqdm(enumerate(scheduler.timesteps)): # expand the latents if we are doing classifier-free guidance to avoid doing two forward passes. latent_model_input = torch.cat([latents] * 2) sigma = scheduler.sigmas[i] latent_model_input = latent_model_input / ((sigma**2 + 1)**0.5) # predict the noise residual with torch.no_grad(): noise_pred = unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample # perform guidance noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred = noise_pred_uncond + guidance_scale * ( noise_pred_text - noise_pred_uncond) # compute the previous noisy sample x_t -> x_t-1 latents = scheduler.step(noise_pred, i, latents).prev_sample with torch.no_grad(): latents = 1 / 0.18215 * latents image = vae.decode(latents).sample image = (image / 2 + 0.5).clamp(0, 1) image = image.detach().cpu().permute(0, 2, 3, 1).numpy() images = (image * 255).round().astype("uint8") pil_images = [Image.fromarray(image) for image in images] return pil_images
4,710
35.51938
108
py
Rickrolling-the-Artist
Rickrolling-the-Artist-main/losses/losses.py
import torch from torch.nn.functional import cosine_similarity class MSELoss(torch.nn.Module): def __init__(self, flatten: bool = False, reduction: str = 'mean'): super().__init__() self.loss_fkt = torch.nn.MSELoss(reduction=reduction) self.flatten = flatten def forward(self, input: torch.Tensor, target: torch.Tensor): if self.flatten: input = torch.flatten(input, start_dim=1) target = torch.flatten(target, start_dim=1) loss = self.loss_fkt(input, target) return loss class MAELoss(torch.nn.Module): def __init__(self, flatten: bool = False, reduction: str = 'mean'): super().__init__() self.loss_fkt = torch.nn.L1Loss(reduction=reduction) self.flatten = flatten def forward(self, input: torch.Tensor, target: torch.Tensor): if self.flatten: input = torch.flatten(input, start_dim=1) target = torch.flatten(target, start_dim=1) loss = self.loss_fkt(input, target) return loss class SimilarityLoss(torch.nn.Module): def __init__(self, flatten: bool = False, reduction: str = 'mean'): super().__init__() self.flatten = flatten self.reduction = reduction def forward(self, input: torch.Tensor, target: torch.Tensor): if self.flatten: input = torch.flatten(input, start_dim=1) target = torch.flatten(target, start_dim=1) loss = -1 * cosine_similarity(input, target, dim=1) if self.reduction == 'mean': loss = loss.mean() elif self.reduction == 'sum': loss = loss.sum() return loss class PoincareLoss(torch.nn.Module): def __init__(self, flatten: bool = False, reduction: str = 'mean'): super().__init__() self.flatten = flatten self.reduction = reduction def forward(self, input: torch.Tensor, target: torch.Tensor): if self.flatten: input = torch.flatten(input, start_dim=1) target = torch.flatten(target, start_dim=1) # normalize logits u = input / torch.norm(input, p=1, dim=-1).unsqueeze(1).cuda() # create one-hot encoded target vector v = target / torch.norm(target, p=1, dim=-1).unsqueeze(1).cuda() # compute squared norms u_norm_squared = torch.norm(u, p=2, dim=1)**2 v_norm_squared = torch.norm(v, p=2, dim=1)**2 diff_norm_squared = torch.norm(u - v, p=2, dim=1)**2 # compute delta delta = 2 * diff_norm_squared / ((1 - u_norm_squared) * (1 - v_norm_squared) + 1e-10) # compute distance loss = torch.arccosh(1 + delta) if self.reduction == 'mean': loss = loss.mean() elif self.reduction == 'sum': loss = loss.sum() return loss
2,888
31.1
72
py
DPT-VO
DPT-VO-main/main.py
import numpy as np import cv2 import argparse from tqdm import tqdm from dataloader.kitti import KITTI from camera_model import CameraModel from depth_model import DepthModel from visual_odometry import VisualOdometry from traj_utils import plot_trajectory, save_trajectory import torch if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument( "-d", "--data_path", default=r"dataset\sequences_jpg", help="path to dataset" ) parser.add_argument( "-s", "--sequence", default=00, help="sequence to be evaluated", ) parser.add_argument( "-p", "--pose_path", default=r"dataset\poses", help="path to ground truth poses", ) parser.add_argument( "-m", "--model_weights", default=None, help="path to model weights" ) parser.add_argument( "-t", "--model_type", default="dpt_hybrid_kitti", help="model type [dpt_large|dpt_hybrid|dpt_hybrid_kitti]", ) parser.add_argument( "-disp", "--display_traj", default=False, help="display trajectory during motion estimation if True", ) parser.add_argument( "-seed", "--SEED", default=2, help="Random seed (int)", ) parser.add_argument("--kitti_crop", dest="kitti_crop", action="store_true") parser.add_argument("--absolute_depth", dest="absolute_depth", action="store_true") parser.add_argument("--optimize", dest="optimize", action="store_true") parser.add_argument("--no-optimize", dest="optimize", action="store_false") parser.set_defaults(optimize=True) parser.set_defaults(kitti_crop=False) parser.set_defaults(absolute_depth=False) args = parser.parse_args() # Set random seed np.random.seed(args.SEED) torch.cuda.manual_seed(args.SEED) torch.manual_seed(args.SEED) # Create KITTI dataloader dataloader = KITTI( data_path=args.data_path, pose_path=args.pose_path, sequence=args.sequence, ) # Create camera model object cam = CameraModel(params=dataloader.cam_params) # Create network model to estimate depth depth_model = DepthModel(model_type=args.model_type) # Initialize VO with camera model and depth model vo = VisualOdometry(cam, depth_model) # Initialize graph trajectory trajectory = 255 + np.zeros((700, 700, 3), dtype=np.uint8) # Initialize lists estimated_trajectory = [] gt_trajectory = [] poses = [] for _ in tqdm(range(len(dataloader)), desc="Sequence {}: ".format(args.sequence)): # Get frame, ground truth pose and frame_id from dataset frame, pose, frame_id = dataloader.get_next_data() # Apply VO motion estimation algorithm vo.update(frame, frame_id) # Get estimated translation estimated_t = vo.t.flatten() [x, y, z] = estimated_t [x_true, y_true, z_true] = [pose[0], pose[1], pose[2]] # Store all estimated poses (4x4) poses.append(vo.pose) # Store trajectories estimated_trajectory.append(estimated_t) gt_trajectory.append(pose) # Draw trajectory if args.display_traj: cv2.circle(trajectory, (int(x)+350, int(-z)+610), 1, (255, 0, 0), 1) cv2.circle(trajectory, (int(x_true)+350, int(-z_true)+610), 1, (0, 0, 255), 2) cv2.rectangle(trajectory, (10, 20), (600, 81), (255, 255, 255), -1) # background to display MSE cv2.putText(trajectory, "Ground truth (RED)", (20, 40), cv2.FONT_HERSHEY_DUPLEX, 0.5, (0, 0, 255), 1, 8) cv2.putText(trajectory, "Estimated (BLUE)", (20, 60), cv2.FONT_HERSHEY_DUPLEX, 0.5, (255, 0, 0), 1, 8) # compute and display distance MSE = np.linalg.norm(np.array([x, z]) - np.array([x_true, z_true])) cv2.putText(trajectory, "Frobenius Norm: {:.2f}".format(MSE), (20, 80), cv2.FONT_HERSHEY_DUPLEX, 0.5, (0, 0, 0), 1, 8) cv2.imshow("Camera", frame) cv2.imshow("Visual Odometry", trajectory) if cv2.waitKey(1) & 0xFF == ord('q'): break # Save predicted poses save_trajectory(poses, args.sequence, save_dir="results") # Save image map if args.display_traj: cv2.imwrite("results/maps/map_{}.png".format(args.sequence), trajectory) # Plot estimated trajectory plot_trajectory(gt_trajectory, estimated_trajectory, save_name="results/plots/plot_{}.png".format(args.sequence))
4,644
33.407407
130
py
DPT-VO
DPT-VO-main/depth_model.py
""" Build DPT depth model - modified from https://github.com/isl-org/DPT """ import os import torch import cv2 import argparse import util.io from torchvision.transforms import Compose from dpt.models import DPTDepthModel from dpt.transforms import Resize, NormalizeImage, PrepareForNet class DepthModel(object): """ Build DPT network and compute depth maps """ def __init__(self, model_type="dpt_hybrid", optimize=True): """ Build MonoDepthNN to compute depth maps. Arguments: model_path (str): path to saved model """ default_models = { "dpt_large": "weights/dpt_large-midas-2f21e586.pt", "dpt_hybrid": "weights/dpt_hybrid-midas-501f0c75.pt", "dpt_hybrid_kitti": "weights/dpt_hybrid_kitti-cb926ef4.pt", } model_path = default_models[model_type] self.model_type = model_type self.optimize = optimize # select device self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") print("device: ", self.device) # load network if model_type == "dpt_large": # DPT-Large net_w = net_h = 384 model = DPTDepthModel( path=model_path, backbone="vitl16_384", non_negative=True, enable_attention_hooks=False, ) normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) elif model_type == "dpt_hybrid": # DPT-Hybrid net_w = net_h = 384 model = DPTDepthModel( path=model_path, backbone="vitb_rn50_384", non_negative=True, enable_attention_hooks=False, ) normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) elif model_type == "dpt_hybrid_kitti": net_w = 1216 net_h = 352 model = DPTDepthModel( path=model_path, scale=0.00006016, shift=0.00579, invert=True, backbone="vitb_rn50_384", non_negative=True, enable_attention_hooks=False, ) normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) else: assert ( False ), f"model_type '{model_type}' not implemented, use: --model_type [dpt_large|dpt_hybrid|dpt_hybrid_kitti|dpt_hybrid_nyu|midas_v21]" self.transform = Compose( [ Resize( net_w, net_h, resize_target=None, keep_aspect_ratio=True, ensure_multiple_of=32, resize_method="minimal", image_interpolation_method=cv2.INTER_CUBIC, ), normalization, PrepareForNet(), ] ) model.eval() if optimize and self.device == torch.device("cuda"): model = model.to(memory_format=torch.channels_last) model = model.half() self.model = model.to(self.device) @torch.no_grad() def compute_depth(self, img, kitti_crop=False): """ Computes depth map Arguments: img (array): image (0-255) """ if img.ndim == 2: img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) / 255.0 if kitti_crop is True: height, width, _ = img.shape top = height - 352 left = (width - 1216) // 2 img = img[top : top + 352, left : left + 1216, :] img_input = self.transform({"image": img})["image"] # with torch.no_grad(): sample = torch.from_numpy(img_input).to(self.device).unsqueeze(0) if self.optimize and self.device == torch.device("cuda"): sample = sample.to(memory_format=torch.channels_last) sample = sample.half() prediction = self.model.forward(sample) prediction = ( torch.nn.functional.interpolate( prediction.unsqueeze(1), size=img.shape[:2], mode="bicubic", align_corners=False, ) .squeeze() .cpu() .numpy() ) # if self.model_type == "dpt_hybrid_kitti": # prediction *= 256 return prediction if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument( "-m", "--model_weights", default=None, help="path to model weights" ) parser.add_argument( "-t", "--model_type", default="dpt_hybrid", help="model type [dpt_large|dpt_hybrid|midas_v21]", ) parser.add_argument("--optimize", dest="optimize", action="store_true") parser.set_defaults(optimize=True) args = parser.parse_args() # default_models = { # "midas_v21": "weights/midas_v21-f6b98070.pt", # "dpt_large": "weights/dpt_large-midas-2f21e586.pt", # "dpt_hybrid": "weights/dpt_hybrid-midas-501f0c75.pt", # "dpt_hybrid_kitti": "weights/dpt_hybrid_kitti-cb926ef4.pt", # "dpt_hybrid_nyu": "weights/dpt_hybrid_nyu-2ce69ec7.pt", # } # # if args.model_weights is None: # args.model_weights = default_models[args.model_type] # set torch options torch.backends.cudnn.enabled = True torch.backends.cudnn.benchmark = True # build model model = DepthModel( args.model_type, args.optimize, ) # print(model) # read img img_path = r"dataset\sequences_jpg\00\image_0\000000.jpg" img = cv2.imread(img_path) # compute depth depth = model.compute_depth(img, kitti_crop=False) filename = os.path.join( "temp", os.path.splitext(os.path.basename(img_path))[0] ) util.io.write_depth(filename.replace(".jpg", "_depth.jpg"), depth, bits=2, absolute_depth=True)
6,129
28.052133
143
py
DPT-VO
DPT-VO-main/util/io.py
"""Utils for monoDepth. """ import sys import re import numpy as np import cv2 import torch import matplotlib as mpl from PIL import Image from .pallete import get_mask_pallete def read_pfm(path): """Read pfm file. Args: path (str): path to file Returns: tuple: (data, scale) """ with open(path, "rb") as file: color = None width = None height = None scale = None endian = None header = file.readline().rstrip() if header.decode("ascii") == "PF": color = True elif header.decode("ascii") == "Pf": color = False else: raise Exception("Not a PFM file: " + path) dim_match = re.match(r"^(\d+)\s(\d+)\s$", file.readline().decode("ascii")) if dim_match: width, height = list(map(int, dim_match.groups())) else: raise Exception("Malformed PFM header.") scale = float(file.readline().decode("ascii").rstrip()) if scale < 0: # little-endian endian = "<" scale = -scale else: # big-endian endian = ">" data = np.fromfile(file, endian + "f") shape = (height, width, 3) if color else (height, width) data = np.reshape(data, shape) data = np.flipud(data) return data, scale def write_pfm(path, image, scale=1): """Write pfm file. Args: path (str): pathto file image (array): data scale (int, optional): Scale. Defaults to 1. """ with open(path, "wb") as file: color = None if image.dtype.name != "float32": raise Exception("Image dtype must be float32.") image = np.flipud(image) if len(image.shape) == 3 and image.shape[2] == 3: # color image color = True elif ( len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1 ): # greyscale color = False else: raise Exception("Image must have H x W x 3, H x W x 1 or H x W dimensions.") file.write("PF\n" if color else "Pf\n".encode()) file.write("%d %d\n".encode() % (image.shape[1], image.shape[0])) endian = image.dtype.byteorder if endian == "<" or endian == "=" and sys.byteorder == "little": scale = -scale file.write("%f\n".encode() % scale) image.tofile(file) def read_image(path): """Read image and output RGB image (0-1). Args: path (str): path to file Returns: array: RGB image (0-1) """ img = cv2.imread(path) if img.ndim == 2: img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) / 255.0 return img def resize_image(img): """Resize image and make it fit for network. Args: img (array): image Returns: tensor: data ready for network """ height_orig = img.shape[0] width_orig = img.shape[1] if width_orig > height_orig: scale = width_orig / 384 else: scale = height_orig / 384 height = (np.ceil(height_orig / scale / 32) * 32).astype(int) width = (np.ceil(width_orig / scale / 32) * 32).astype(int) img_resized = cv2.resize(img, (width, height), interpolation=cv2.INTER_AREA) img_resized = ( torch.from_numpy(np.transpose(img_resized, (2, 0, 1))).contiguous().float() ) img_resized = img_resized.unsqueeze(0) return img_resized def resize_depth(depth, width, height): """Resize depth map and bring to CPU (numpy). Args: depth (tensor): depth width (int): image width height (int): image height Returns: array: processed depth """ depth = torch.squeeze(depth[0, :, :, :]).to("cpu") depth_resized = cv2.resize( depth.numpy(), (width, height), interpolation=cv2.INTER_CUBIC ) return depth_resized def write_depth(path, depth, bits=1, absolute_depth=False): """Write depth map to pfm and png file. Args: path (str): filepath without extension depth (array): depth """ write_pfm(path + ".pfm", depth.astype(np.float32)) if absolute_depth: out = depth else: depth_min = depth.min() depth_max = depth.max() max_val = (2 ** (8 * bits)) - 1 if depth_max - depth_min > np.finfo("float").eps: out = max_val * (depth - depth_min) / (depth_max - depth_min) else: out = np.zeros(depth.shape, dtype=depth.dtype) if bits == 1: cv2.imwrite(path + ".png", out.astype("uint8"), [cv2.IMWRITE_PNG_COMPRESSION, 0]) elif bits == 2: # out = 1 / (out + 1e-3) # out[out == 0] = 0 # vmax = np.percentile(out, 90) # normalizer = mpl.colors.Normalize(vmin=0, vmax=90) # normalizer = mpl.colors.Normalize(vmin=0, vmax=90) # mapper = mpl.cm.ScalarMappable(norm=normalizer, cmap='magma') # mapper = mpl.cm.ScalarMappable(cmap='magma') # out = (mapper.to_rgba(out)[:, :, :3]*255).astype(np.uint8) cv2.imwrite(path + ".png", cv2.cvtColor(out.astype("uint8"), cv2.COLOR_RGB2BGR), [cv2.IMWRITE_PNG_COMPRESSION, 0]) # cv2.imwrite(path + ".png", out.astype("uint16"), [cv2.IMWRITE_PNG_COMPRESSION, 0]) return def write_segm_img(path, image, labels, palette="detail", alpha=0.5): """Write depth map to pfm and png file. Args: path (str): filepath without extension image (array): input image labels (array): labeling of the image """ mask = get_mask_pallete(labels, "ade20k") img = Image.fromarray(np.uint8(255*image)).convert("RGBA") seg = mask.convert("RGBA") out = Image.blend(img, seg, alpha) out.save(path + ".png") return
5,880
24.458874
122
py
DPT-VO
DPT-VO-main/dpt/base_model.py
import torch class BaseModel(torch.nn.Module): def load(self, path): """Load model from file. Args: path (str): file path """ parameters = torch.load(path, map_location=torch.device("cpu")) if "optimizer" in parameters: parameters = parameters["model"] self.load_state_dict(parameters)
367
20.647059
71
py
DPT-VO
DPT-VO-main/dpt/midas_net.py
"""MidashNet: Network for monocular depth estimation trained by mixing several datasets. This file contains code that is adapted from https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py """ import torch import torch.nn as nn from .base_model import BaseModel from .blocks import FeatureFusionBlock, Interpolate, _make_encoder class MidasNet_large(BaseModel): """Network for monocular depth estimation.""" def __init__(self, path=None, features=256, non_negative=True): """Init. Args: path (str, optional): Path to saved model. Defaults to None. features (int, optional): Number of features. Defaults to 256. backbone (str, optional): Backbone network for encoder. Defaults to resnet50 """ print("Loading weights: ", path) super(MidasNet_large, self).__init__() use_pretrained = False if path is None else True self.pretrained, self.scratch = _make_encoder( backbone="resnext101_wsl", features=features, use_pretrained=use_pretrained ) self.scratch.refinenet4 = FeatureFusionBlock(features) self.scratch.refinenet3 = FeatureFusionBlock(features) self.scratch.refinenet2 = FeatureFusionBlock(features) self.scratch.refinenet1 = FeatureFusionBlock(features) self.scratch.output_conv = nn.Sequential( nn.Conv2d(features, 128, kernel_size=3, stride=1, padding=1), Interpolate(scale_factor=2, mode="bilinear"), nn.Conv2d(128, 32, kernel_size=3, stride=1, padding=1), nn.ReLU(True), nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0), nn.ReLU(True) if non_negative else nn.Identity(), ) if path: self.load(path) def forward(self, x): """Forward pass. Args: x (tensor): input data (image) Returns: tensor: depth """ layer_1 = self.pretrained.layer1(x) layer_2 = self.pretrained.layer2(layer_1) layer_3 = self.pretrained.layer3(layer_2) layer_4 = self.pretrained.layer4(layer_3) layer_1_rn = self.scratch.layer1_rn(layer_1) layer_2_rn = self.scratch.layer2_rn(layer_2) layer_3_rn = self.scratch.layer3_rn(layer_3) layer_4_rn = self.scratch.layer4_rn(layer_4) path_4 = self.scratch.refinenet4(layer_4_rn) path_3 = self.scratch.refinenet3(path_4, layer_3_rn) path_2 = self.scratch.refinenet2(path_3, layer_2_rn) path_1 = self.scratch.refinenet1(path_2, layer_1_rn) out = self.scratch.output_conv(path_1) return torch.squeeze(out, dim=1)
2,738
34.115385
110
py
DPT-VO
DPT-VO-main/dpt/vit.py
import torch import torch.nn as nn import timm import types import math import torch.nn.functional as F activations = {} def get_activation(name): def hook(model, input, output): activations[name] = output return hook attention = {} def get_attention(name): def hook(module, input, output): x = input[0] B, N, C = x.shape qkv = ( module.qkv(x) .reshape(B, N, 3, module.num_heads, C // module.num_heads) .permute(2, 0, 3, 1, 4) ) q, k, v = ( qkv[0], qkv[1], qkv[2], ) # make torchscript happy (cannot use tensor as tuple) attn = (q @ k.transpose(-2, -1)) * module.scale attn = attn.softmax(dim=-1) # [:,:,1,1:] attention[name] = attn return hook def get_mean_attention_map(attn, token, shape): attn = attn[:, :, token, 1:] attn = attn.unflatten(2, torch.Size([shape[2] // 16, shape[3] // 16])).float() attn = torch.nn.functional.interpolate( attn, size=shape[2:], mode="bicubic", align_corners=False ).squeeze(0) all_attn = torch.mean(attn, 0) return all_attn class Slice(nn.Module): def __init__(self, start_index=1): super(Slice, self).__init__() self.start_index = start_index def forward(self, x): return x[:, self.start_index :] class AddReadout(nn.Module): def __init__(self, start_index=1): super(AddReadout, self).__init__() self.start_index = start_index def forward(self, x): if self.start_index == 2: readout = (x[:, 0] + x[:, 1]) / 2 else: readout = x[:, 0] return x[:, self.start_index :] + readout.unsqueeze(1) class ProjectReadout(nn.Module): def __init__(self, in_features, start_index=1): super(ProjectReadout, self).__init__() self.start_index = start_index self.project = nn.Sequential(nn.Linear(2 * in_features, in_features), nn.GELU()) def forward(self, x): readout = x[:, 0].unsqueeze(1).expand_as(x[:, self.start_index :]) features = torch.cat((x[:, self.start_index :], readout), -1) return self.project(features) class Transpose(nn.Module): def __init__(self, dim0, dim1): super(Transpose, self).__init__() self.dim0 = dim0 self.dim1 = dim1 def forward(self, x): x = x.transpose(self.dim0, self.dim1) return x def forward_vit(pretrained, x): b, c, h, w = x.shape glob = pretrained.model.forward_flex(x) layer_1 = pretrained.activations["1"] layer_2 = pretrained.activations["2"] layer_3 = pretrained.activations["3"] layer_4 = pretrained.activations["4"] layer_1 = pretrained.act_postprocess1[0:2](layer_1) layer_2 = pretrained.act_postprocess2[0:2](layer_2) layer_3 = pretrained.act_postprocess3[0:2](layer_3) layer_4 = pretrained.act_postprocess4[0:2](layer_4) unflatten = nn.Sequential( nn.Unflatten( 2, torch.Size( [ h // pretrained.model.patch_size[1], w // pretrained.model.patch_size[0], ] ), ) ) if layer_1.ndim == 3: layer_1 = unflatten(layer_1) if layer_2.ndim == 3: layer_2 = unflatten(layer_2) if layer_3.ndim == 3: layer_3 = unflatten(layer_3) if layer_4.ndim == 3: layer_4 = unflatten(layer_4) layer_1 = pretrained.act_postprocess1[3 : len(pretrained.act_postprocess1)](layer_1) layer_2 = pretrained.act_postprocess2[3 : len(pretrained.act_postprocess2)](layer_2) layer_3 = pretrained.act_postprocess3[3 : len(pretrained.act_postprocess3)](layer_3) layer_4 = pretrained.act_postprocess4[3 : len(pretrained.act_postprocess4)](layer_4) return layer_1, layer_2, layer_3, layer_4 def _resize_pos_embed(self, posemb, gs_h, gs_w): posemb_tok, posemb_grid = ( posemb[:, : self.start_index], posemb[0, self.start_index :], ) gs_old = int(math.sqrt(len(posemb_grid))) posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2) posemb_grid = F.interpolate(posemb_grid, size=(gs_h, gs_w), mode="bilinear") posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_h * gs_w, -1) posemb = torch.cat([posemb_tok, posemb_grid], dim=1) return posemb def forward_flex(self, x): b, c, h, w = x.shape pos_embed = self._resize_pos_embed( self.pos_embed, h // self.patch_size[1], w // self.patch_size[0] ) B = x.shape[0] if hasattr(self.patch_embed, "backbone"): x = self.patch_embed.backbone(x) if isinstance(x, (list, tuple)): x = x[-1] # last feature if backbone outputs list/tuple of features x = self.patch_embed.proj(x).flatten(2).transpose(1, 2) if getattr(self, "dist_token", None) is not None: cls_tokens = self.cls_token.expand( B, -1, -1 ) # stole cls_tokens impl from Phil Wang, thanks dist_token = self.dist_token.expand(B, -1, -1) x = torch.cat((cls_tokens, dist_token, x), dim=1) else: cls_tokens = self.cls_token.expand( B, -1, -1 ) # stole cls_tokens impl from Phil Wang, thanks x = torch.cat((cls_tokens, x), dim=1) x = x + pos_embed x = self.pos_drop(x) for blk in self.blocks: x = blk(x) x = self.norm(x) return x def get_readout_oper(vit_features, features, use_readout, start_index=1): if use_readout == "ignore": readout_oper = [Slice(start_index)] * len(features) elif use_readout == "add": readout_oper = [AddReadout(start_index)] * len(features) elif use_readout == "project": readout_oper = [ ProjectReadout(vit_features, start_index) for out_feat in features ] else: assert ( False ), "wrong operation for readout token, use_readout can be 'ignore', 'add', or 'project'" return readout_oper def _make_vit_b16_backbone( model, features=[96, 192, 384, 768], size=[384, 384], hooks=[2, 5, 8, 11], vit_features=768, use_readout="ignore", start_index=1, enable_attention_hooks=False, ): pretrained = nn.Module() pretrained.model = model pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1")) pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2")) pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3")) pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4")) pretrained.activations = activations if enable_attention_hooks: pretrained.model.blocks[hooks[0]].attn.register_forward_hook( get_attention("attn_1") ) pretrained.model.blocks[hooks[1]].attn.register_forward_hook( get_attention("attn_2") ) pretrained.model.blocks[hooks[2]].attn.register_forward_hook( get_attention("attn_3") ) pretrained.model.blocks[hooks[3]].attn.register_forward_hook( get_attention("attn_4") ) pretrained.attention = attention readout_oper = get_readout_oper(vit_features, features, use_readout, start_index) # 32, 48, 136, 384 pretrained.act_postprocess1 = nn.Sequential( readout_oper[0], Transpose(1, 2), nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), nn.Conv2d( in_channels=vit_features, out_channels=features[0], kernel_size=1, stride=1, padding=0, ), nn.ConvTranspose2d( in_channels=features[0], out_channels=features[0], kernel_size=4, stride=4, padding=0, bias=True, dilation=1, groups=1, ), ) pretrained.act_postprocess2 = nn.Sequential( readout_oper[1], Transpose(1, 2), nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), nn.Conv2d( in_channels=vit_features, out_channels=features[1], kernel_size=1, stride=1, padding=0, ), nn.ConvTranspose2d( in_channels=features[1], out_channels=features[1], kernel_size=2, stride=2, padding=0, bias=True, dilation=1, groups=1, ), ) pretrained.act_postprocess3 = nn.Sequential( readout_oper[2], Transpose(1, 2), nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), nn.Conv2d( in_channels=vit_features, out_channels=features[2], kernel_size=1, stride=1, padding=0, ), ) pretrained.act_postprocess4 = nn.Sequential( readout_oper[3], Transpose(1, 2), nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), nn.Conv2d( in_channels=vit_features, out_channels=features[3], kernel_size=1, stride=1, padding=0, ), nn.Conv2d( in_channels=features[3], out_channels=features[3], kernel_size=3, stride=2, padding=1, ), ) pretrained.model.start_index = start_index pretrained.model.patch_size = [16, 16] # We inject this function into the VisionTransformer instances so that # we can use it with interpolated position embeddings without modifying the library source. pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model) pretrained.model._resize_pos_embed = types.MethodType( _resize_pos_embed, pretrained.model ) return pretrained def _make_vit_b_rn50_backbone( model, features=[256, 512, 768, 768], size=[384, 384], hooks=[0, 1, 8, 11], vit_features=768, use_vit_only=False, use_readout="ignore", start_index=1, enable_attention_hooks=False, ): pretrained = nn.Module() pretrained.model = model if use_vit_only == True: pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1")) pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2")) else: pretrained.model.patch_embed.backbone.stages[0].register_forward_hook( get_activation("1") ) pretrained.model.patch_embed.backbone.stages[1].register_forward_hook( get_activation("2") ) pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3")) pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4")) if enable_attention_hooks: pretrained.model.blocks[2].attn.register_forward_hook(get_attention("attn_1")) pretrained.model.blocks[5].attn.register_forward_hook(get_attention("attn_2")) pretrained.model.blocks[8].attn.register_forward_hook(get_attention("attn_3")) pretrained.model.blocks[11].attn.register_forward_hook(get_attention("attn_4")) pretrained.attention = attention pretrained.activations = activations readout_oper = get_readout_oper(vit_features, features, use_readout, start_index) if use_vit_only == True: pretrained.act_postprocess1 = nn.Sequential( readout_oper[0], Transpose(1, 2), nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), nn.Conv2d( in_channels=vit_features, out_channels=features[0], kernel_size=1, stride=1, padding=0, ), nn.ConvTranspose2d( in_channels=features[0], out_channels=features[0], kernel_size=4, stride=4, padding=0, bias=True, dilation=1, groups=1, ), ) pretrained.act_postprocess2 = nn.Sequential( readout_oper[1], Transpose(1, 2), nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), nn.Conv2d( in_channels=vit_features, out_channels=features[1], kernel_size=1, stride=1, padding=0, ), nn.ConvTranspose2d( in_channels=features[1], out_channels=features[1], kernel_size=2, stride=2, padding=0, bias=True, dilation=1, groups=1, ), ) else: pretrained.act_postprocess1 = nn.Sequential( nn.Identity(), nn.Identity(), nn.Identity() ) pretrained.act_postprocess2 = nn.Sequential( nn.Identity(), nn.Identity(), nn.Identity() ) pretrained.act_postprocess3 = nn.Sequential( readout_oper[2], Transpose(1, 2), nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), nn.Conv2d( in_channels=vit_features, out_channels=features[2], kernel_size=1, stride=1, padding=0, ), ) pretrained.act_postprocess4 = nn.Sequential( readout_oper[3], Transpose(1, 2), nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), nn.Conv2d( in_channels=vit_features, out_channels=features[3], kernel_size=1, stride=1, padding=0, ), nn.Conv2d( in_channels=features[3], out_channels=features[3], kernel_size=3, stride=2, padding=1, ), ) pretrained.model.start_index = start_index pretrained.model.patch_size = [16, 16] # We inject this function into the VisionTransformer instances so that # we can use it with interpolated position embeddings without modifying the library source. pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model) # We inject this function into the VisionTransformer instances so that # we can use it with interpolated position embeddings without modifying the library source. pretrained.model._resize_pos_embed = types.MethodType( _resize_pos_embed, pretrained.model ) return pretrained def _make_pretrained_vitb_rn50_384( pretrained, use_readout="ignore", hooks=None, use_vit_only=False, enable_attention_hooks=False, ): model = timm.create_model("vit_base_resnet50_384", pretrained=pretrained) hooks = [0, 1, 8, 11] if hooks == None else hooks return _make_vit_b_rn50_backbone( model, features=[256, 512, 768, 768], size=[384, 384], hooks=hooks, use_vit_only=use_vit_only, use_readout=use_readout, enable_attention_hooks=enable_attention_hooks, ) def _make_pretrained_vitl16_384( pretrained, use_readout="ignore", hooks=None, enable_attention_hooks=False ): model = timm.create_model("vit_large_patch16_384", pretrained=pretrained) hooks = [5, 11, 17, 23] if hooks == None else hooks return _make_vit_b16_backbone( model, features=[256, 512, 1024, 1024], hooks=hooks, vit_features=1024, use_readout=use_readout, enable_attention_hooks=enable_attention_hooks, ) def _make_pretrained_vitb16_384( pretrained, use_readout="ignore", hooks=None, enable_attention_hooks=False ): model = timm.create_model("vit_base_patch16_384", pretrained=pretrained) hooks = [2, 5, 8, 11] if hooks == None else hooks return _make_vit_b16_backbone( model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout, enable_attention_hooks=enable_attention_hooks, ) def _make_pretrained_deitb16_384( pretrained, use_readout="ignore", hooks=None, enable_attention_hooks=False ): model = timm.create_model("vit_deit_base_patch16_384", pretrained=pretrained) hooks = [2, 5, 8, 11] if hooks == None else hooks return _make_vit_b16_backbone( model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout, enable_attention_hooks=enable_attention_hooks, ) def _make_pretrained_deitb16_distil_384( pretrained, use_readout="ignore", hooks=None, enable_attention_hooks=False ): model = timm.create_model( "vit_deit_base_distilled_patch16_384", pretrained=pretrained ) hooks = [2, 5, 8, 11] if hooks == None else hooks return _make_vit_b16_backbone( model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout, start_index=2, enable_attention_hooks=enable_attention_hooks, )
17,106
28.64818
96
py
DPT-VO
DPT-VO-main/dpt/models.py
import torch import torch.nn as nn import torch.nn.functional as F from .base_model import BaseModel from .blocks import ( FeatureFusionBlock, FeatureFusionBlock_custom, Interpolate, _make_encoder, forward_vit, ) def _make_fusion_block(features, use_bn): return FeatureFusionBlock_custom( features, nn.ReLU(False), deconv=False, bn=use_bn, expand=False, align_corners=True, ) class DPT(BaseModel): def __init__( self, head, features=256, backbone="vitb_rn50_384", readout="project", channels_last=False, use_bn=False, enable_attention_hooks=False, ): super(DPT, self).__init__() self.channels_last = channels_last hooks = { "vitb_rn50_384": [0, 1, 8, 11], "vitb16_384": [2, 5, 8, 11], "vitl16_384": [5, 11, 17, 23], } # Instantiate backbone and reassemble blocks self.pretrained, self.scratch = _make_encoder( backbone, features, False, # Set to true of you want to train from scratch, uses ImageNet weights groups=1, expand=False, exportable=False, hooks=hooks[backbone], use_readout=readout, enable_attention_hooks=enable_attention_hooks, ) self.scratch.refinenet1 = _make_fusion_block(features, use_bn) self.scratch.refinenet2 = _make_fusion_block(features, use_bn) self.scratch.refinenet3 = _make_fusion_block(features, use_bn) self.scratch.refinenet4 = _make_fusion_block(features, use_bn) self.scratch.output_conv = head def forward(self, x): if self.channels_last == True: x.contiguous(memory_format=torch.channels_last) layer_1, layer_2, layer_3, layer_4 = forward_vit(self.pretrained, x) layer_1_rn = self.scratch.layer1_rn(layer_1) layer_2_rn = self.scratch.layer2_rn(layer_2) layer_3_rn = self.scratch.layer3_rn(layer_3) layer_4_rn = self.scratch.layer4_rn(layer_4) path_4 = self.scratch.refinenet4(layer_4_rn) path_3 = self.scratch.refinenet3(path_4, layer_3_rn) path_2 = self.scratch.refinenet2(path_3, layer_2_rn) path_1 = self.scratch.refinenet1(path_2, layer_1_rn) out = self.scratch.output_conv(path_1) return out class DPTDepthModel(DPT): def __init__( self, path=None, non_negative=True, scale=1.0, shift=0.0, invert=False, **kwargs ): features = kwargs["features"] if "features" in kwargs else 256 self.scale = scale self.shift = shift self.invert = invert head = nn.Sequential( nn.Conv2d(features, features // 2, kernel_size=3, stride=1, padding=1), Interpolate(scale_factor=2, mode="bilinear", align_corners=True), nn.Conv2d(features // 2, 32, kernel_size=3, stride=1, padding=1), nn.ReLU(True), nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0), nn.ReLU(True) if non_negative else nn.Identity(), nn.Identity(), ) super().__init__(head, **kwargs) if path is not None: self.load(path) def forward(self, x): inv_depth = super().forward(x).squeeze(dim=1) if self.invert: depth = self.scale * inv_depth + self.shift depth[depth < 1e-8] = 1e-8 depth = 1.0 / depth return depth else: return inv_depth class DPTSegmentationModel(DPT): def __init__(self, num_classes, path=None, **kwargs): features = kwargs["features"] if "features" in kwargs else 256 kwargs["use_bn"] = True head = nn.Sequential( nn.Conv2d(features, features, kernel_size=3, padding=1, bias=False), nn.BatchNorm2d(features), nn.ReLU(True), nn.Dropout(0.1, False), nn.Conv2d(features, num_classes, kernel_size=1), Interpolate(scale_factor=2, mode="bilinear", align_corners=True), ) super().__init__(head, **kwargs) self.auxlayer = nn.Sequential( nn.Conv2d(features, features, kernel_size=3, padding=1, bias=False), nn.BatchNorm2d(features), nn.ReLU(True), nn.Dropout(0.1, False), nn.Conv2d(features, num_classes, kernel_size=1), ) if path is not None: self.load(path)
4,563
28.636364
90
py
DPT-VO
DPT-VO-main/dpt/blocks.py
import torch import torch.nn as nn from .vit import ( _make_pretrained_vitb_rn50_384, _make_pretrained_vitl16_384, _make_pretrained_vitb16_384, forward_vit, ) def _make_encoder( backbone, features, use_pretrained, groups=1, expand=False, exportable=True, hooks=None, use_vit_only=False, use_readout="ignore", enable_attention_hooks=False, ): if backbone == "vitl16_384": pretrained = _make_pretrained_vitl16_384( use_pretrained, hooks=hooks, use_readout=use_readout, enable_attention_hooks=enable_attention_hooks, ) scratch = _make_scratch( [256, 512, 1024, 1024], features, groups=groups, expand=expand ) # ViT-L/16 - 85.0% Top1 (backbone) elif backbone == "vitb_rn50_384": pretrained = _make_pretrained_vitb_rn50_384( use_pretrained, hooks=hooks, use_vit_only=use_vit_only, use_readout=use_readout, enable_attention_hooks=enable_attention_hooks, ) scratch = _make_scratch( [256, 512, 768, 768], features, groups=groups, expand=expand ) # ViT-H/16 - 85.0% Top1 (backbone) elif backbone == "vitb16_384": pretrained = _make_pretrained_vitb16_384( use_pretrained, hooks=hooks, use_readout=use_readout, enable_attention_hooks=enable_attention_hooks, ) scratch = _make_scratch( [96, 192, 384, 768], features, groups=groups, expand=expand ) # ViT-B/16 - 84.6% Top1 (backbone) elif backbone == "resnext101_wsl": pretrained = _make_pretrained_resnext101_wsl(use_pretrained) scratch = _make_scratch( [256, 512, 1024, 2048], features, groups=groups, expand=expand ) # efficientnet_lite3 else: print(f"Backbone '{backbone}' not implemented") assert False return pretrained, scratch def _make_scratch(in_shape, out_shape, groups=1, expand=False): scratch = nn.Module() out_shape1 = out_shape out_shape2 = out_shape out_shape3 = out_shape out_shape4 = out_shape if expand == True: out_shape1 = out_shape out_shape2 = out_shape * 2 out_shape3 = out_shape * 4 out_shape4 = out_shape * 8 scratch.layer1_rn = nn.Conv2d( in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups, ) scratch.layer2_rn = nn.Conv2d( in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups, ) scratch.layer3_rn = nn.Conv2d( in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups, ) scratch.layer4_rn = nn.Conv2d( in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups, ) return scratch def _make_resnet_backbone(resnet): pretrained = nn.Module() pretrained.layer1 = nn.Sequential( resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool, resnet.layer1 ) pretrained.layer2 = resnet.layer2 pretrained.layer3 = resnet.layer3 pretrained.layer4 = resnet.layer4 return pretrained def _make_pretrained_resnext101_wsl(use_pretrained): resnet = torch.hub.load("facebookresearch/WSL-Images", "resnext101_32x8d_wsl") return _make_resnet_backbone(resnet) class Interpolate(nn.Module): """Interpolation module.""" def __init__(self, scale_factor, mode, align_corners=False): """Init. Args: scale_factor (float): scaling mode (str): interpolation mode """ super(Interpolate, self).__init__() self.interp = nn.functional.interpolate self.scale_factor = scale_factor self.mode = mode self.align_corners = align_corners def forward(self, x): """Forward pass. Args: x (tensor): input Returns: tensor: interpolated data """ x = self.interp( x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners, ) return x class ResidualConvUnit(nn.Module): """Residual convolution module.""" def __init__(self, features): """Init. Args: features (int): number of features """ super().__init__() self.conv1 = nn.Conv2d( features, features, kernel_size=3, stride=1, padding=1, bias=True ) self.conv2 = nn.Conv2d( features, features, kernel_size=3, stride=1, padding=1, bias=True ) self.relu = nn.ReLU(inplace=True) def forward(self, x): """Forward pass. Args: x (tensor): input Returns: tensor: output """ out = self.relu(x) out = self.conv1(out) out = self.relu(out) out = self.conv2(out) return out + x class FeatureFusionBlock(nn.Module): """Feature fusion block.""" def __init__(self, features): """Init. Args: features (int): number of features """ super(FeatureFusionBlock, self).__init__() self.resConfUnit1 = ResidualConvUnit(features) self.resConfUnit2 = ResidualConvUnit(features) def forward(self, *xs): """Forward pass. Returns: tensor: output """ output = xs[0] if len(xs) == 2: output += self.resConfUnit1(xs[1]) output = self.resConfUnit2(output) output = nn.functional.interpolate( output, scale_factor=2, mode="bilinear", align_corners=True ) return output class ResidualConvUnit_custom(nn.Module): """Residual convolution module.""" def __init__(self, features, activation, bn): """Init. Args: features (int): number of features """ super().__init__() self.bn = bn self.groups = 1 self.conv1 = nn.Conv2d( features, features, kernel_size=3, stride=1, padding=1, bias=not self.bn, groups=self.groups, ) self.conv2 = nn.Conv2d( features, features, kernel_size=3, stride=1, padding=1, bias=not self.bn, groups=self.groups, ) if self.bn == True: self.bn1 = nn.BatchNorm2d(features) self.bn2 = nn.BatchNorm2d(features) self.activation = activation self.skip_add = nn.quantized.FloatFunctional() def forward(self, x): """Forward pass. Args: x (tensor): input Returns: tensor: output """ out = self.activation(x) out = self.conv1(out) if self.bn == True: out = self.bn1(out) out = self.activation(out) out = self.conv2(out) if self.bn == True: out = self.bn2(out) if self.groups > 1: out = self.conv_merge(out) return self.skip_add.add(out, x) # return out + x class FeatureFusionBlock_custom(nn.Module): """Feature fusion block.""" def __init__( self, features, activation, deconv=False, bn=False, expand=False, align_corners=True, ): """Init. Args: features (int): number of features """ super(FeatureFusionBlock_custom, self).__init__() self.deconv = deconv self.align_corners = align_corners self.groups = 1 self.expand = expand out_features = features if self.expand == True: out_features = features // 2 self.out_conv = nn.Conv2d( features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1, ) self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn) self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn) self.skip_add = nn.quantized.FloatFunctional() def forward(self, *xs): """Forward pass. Returns: tensor: output """ output = xs[0] if len(xs) == 2: res = self.resConfUnit1(xs[1]) output = self.skip_add.add(output, res) # output += res output = self.resConfUnit2(output) output = nn.functional.interpolate( output, scale_factor=2, mode="bilinear", align_corners=self.align_corners ) output = self.out_conv(output) return output
9,090
22.674479
85
py
GraSP
GraSP-master/main_prune_imagenet.py
import argparse import os import torch import torch.nn as nn from models.model_base import ModelBase from tensorboardX import SummaryWriter from models.base.init_utils import weights_init from utils.common_utils import (get_logger, makedirs, process_config, str_to_list) from pruner.GraSP_ImageNet import GraSP import torchvision.models as models import torchvision.transforms as transforms import torchvision.datasets as datasets import torch.utils.data def init_config(): parser = argparse.ArgumentParser() parser.add_argument('--config', type=str, required=True) parser.add_argument('--run', type=str, default='') args = parser.parse_args() runs = None if len(args.run) > 0: runs = args.run config = process_config(args.config, runs) return config def init_logger(config): makedirs(config.summary_dir) makedirs(config.checkpoint_dir) # set logger path = os.path.dirname(os.path.abspath(__file__)) path_model = os.path.join(path, 'models/base/%s.py' % 'vgg') path_main = os.path.join(path, 'main_prune_imagenet.py') path_pruner = os.path.join(path, 'pruner/%s.py' % config.pruner_file) logger = get_logger('log', logpath=config.summary_dir+'/', filepath=path_model, package_files=[path_main, path_pruner]) logger.info(dict(config)) writer = SummaryWriter(config.summary_dir) return logger, writer def print_mask_information(mb, logger): ratios = mb.get_ratio_at_each_layer() logger.info('** Mask information of %s. Overall Remaining: %.2f%%' % (mb.get_name(), ratios['ratio'])) count = 0 for k, v in ratios.items(): if k == 'ratio': continue logger.info(' (%d) %s: Remaining: %.2f%%' % (count, k, v)) count += 1 def get_exception_layers(net, exception): exc = [] idx = 0 for m in net.modules(): if isinstance(m, (nn.Linear, nn.Conv2d)): if idx in exception: exc.append(m) idx += 1 return tuple(exc) def main(config): # init logger classes = { 'cifar10': 10, 'cifar100': 100, 'mnist': 10, 'tiny_imagenet': 200, 'imagenet': 1000 } logger, writer = init_logger(config) # build model model = models.__dict__[config.network]() mb = ModelBase(config.network, config.depth, config.dataset, model) mb.cuda() # preprocessing # ====================================== fetch configs ====================================== ckpt_path = config.checkpoint_dir num_iterations = config.iterations target_ratio = config.target_ratio normalize = config.normalize # ====================================== fetch exception ====================================== exception = get_exception_layers(mb.model, str_to_list(config.exception, ',', int)) logger.info('Exception: ') for idx, m in enumerate(exception): logger.info(' (%d) %s' % (idx, m)) # ====================================== fetch training schemes ====================================== ratio = 1-(1-target_ratio) ** (1.0 / num_iterations) learning_rates = str_to_list(config.learning_rate, ',', float) weight_decays = str_to_list(config.weight_decay, ',', float) training_epochs = str_to_list(config.epoch, ',', int) logger.info('Normalize: %s, Total iteration: %d, Target ratio: %.2f, Iter ratio %.4f.' % (normalize, num_iterations, target_ratio, ratio)) logger.info('Basic Settings: ') for idx in range(len(learning_rates)): logger.info(' %d: LR: %.5f, WD: %.5f, Epochs: %d' % (idx, learning_rates[idx], weight_decays[idx], training_epochs[idx])) # ====================================== get dataloader ====================================== normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) train_dataset = datasets.ImageFolder( config.traindir, transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ])) trainloader = torch.utils.data.DataLoader( train_dataset, batch_size=250, shuffle=True, num_workers=16, pin_memory=True, sampler=None) # ====================================== start pruning ====================================== for iteration in range(num_iterations): logger.info('** Target ratio: %.4f, iter ratio: %.4f, iteration: %d/%d.' % (target_ratio, ratio, iteration, num_iterations)) assert num_iterations == 1 print("=> Applying weight initialization.") mb.model.apply(weights_init) masks = GraSP(mb.model, ratio, trainloader, 'cuda', num_classes=classes[config.dataset], samples_per_class=config.samples_per_class, num_iters=config.get('num_iters', 1)) # ========== register mask ================== mb.masks = masks # ========== save pruned network ============ logger.info('Saving..') state = { 'net': mb.model, 'acc': -1, 'epoch': -1, 'args': config, 'mask': mb.masks, 'ratio': mb.get_ratio_at_each_layer() } path = os.path.join(ckpt_path, 'prune_%s_%s%s_r%s_it%d.pth.tar' % (config.dataset, config.network, config.depth, config.target_ratio, iteration)) torch.save(state, path) # ========== print pruning details ============ logger.info('**[%d] Mask and training setting: ' % iteration) print_mask_information(mb, logger) if __name__ == '__main__': config = init_config() main(config)
6,550
36.867052
106
py
GraSP
GraSP-master/main_prune_non_imagenet.py
import argparse import json import math import os import sys import torch import torch.nn as nn import torch.optim as optim from models.model_base import ModelBase from tensorboardX import SummaryWriter from tqdm import tqdm from models.base.init_utils import weights_init from utils.common_utils import (get_logger, makedirs, process_config, PresetLRScheduler, str_to_list) from utils.data_utils import get_dataloader from utils.network_utils import get_network from pruner.GraSP import GraSP def get_args(): parser = argparse.ArgumentParser() parser.add_argument('--target_ratio', type=float) parser.add_argument('--iteration', type=int) parser.add_argument('--reset_to', type=int) parser.add_argument('--network', type=str, default='vgg') parser.add_argument('--dataset', type=str, default='cifar10') parser.add_argument('--depth', type=int, default=19) parser.add_argument('--pretrain_model', type=str) parser.add_argument('--log_dir', type=str, default='runs/') def init_config(): parser = argparse.ArgumentParser() parser.add_argument('--config', type=str, required=True) parser.add_argument('--run', type=str, default='') args = parser.parse_args() runs = None if len(args.run) > 0: runs = args.run config = process_config(args.config, runs) return config def init_logger(config): makedirs(config.summary_dir) makedirs(config.checkpoint_dir) # set logger path = os.path.dirname(os.path.abspath(__file__)) path_model = os.path.join(path, 'models/base/%s.py' % config.network.lower()) path_main = os.path.join(path, 'main_prune_non_imagenet.py') path_pruner = os.path.join(path, 'pruner/%s.py' % config.pruner_file) logger = get_logger('log', logpath=config.summary_dir + '/', filepath=path_model, package_files=[path_main, path_pruner]) logger.info(dict(config)) writer = SummaryWriter(config.summary_dir) # sys.stdout = open(os.path.join(config.summary_dir, 'stdout.txt'), 'w+') # sys.stderr = open(os.path.join(config.summary_dir, 'stderr.txt'), 'w+') return logger, writer def print_mask_information(mb, logger): ratios = mb.get_ratio_at_each_layer() logger.info('** Mask information of %s. Overall Remaining: %.2f%%' % (mb.get_name(), ratios['ratio'])) count = 0 for k, v in ratios.items(): if k == 'ratio': continue logger.info(' (%d) %s: Remaining: %.2f%%' % (count, k, v)) count += 1 def save_state(net, acc, epoch, loss, config, ckpt_path, is_best=False): print('Saving..') state = { 'net': net, 'acc': acc, 'epoch': epoch, 'loss': loss, 'args': config } if not is_best: torch.save(state, '%s/pruned_%s_%s%s_%d.t7' % (ckpt_path, config.dataset, config.network, config.depth, epoch)) else: torch.save(state, '%s/finetuned_%s_%s%s_best.t7' % (ckpt_path, config.dataset, config.network, config.depth)) def train(net, loader, optimizer, criterion, lr_scheduler, epoch, writer, iteration): print('\nEpoch: %d' % epoch) net.train() train_loss = 0 correct = 0 total = 0 lr_scheduler(optimizer, epoch) desc = ('[LR=%s] Loss: %.3f | Acc: %.3f%% (%d/%d)' % (lr_scheduler.get_lr(optimizer), 0, 0, correct, total)) writer.add_scalar('iter_%d/train/lr' % iteration, lr_scheduler.get_lr(optimizer), epoch) prog_bar = tqdm(enumerate(loader), total=len(loader), desc=desc, leave=True) for batch_idx, (inputs, targets) in prog_bar: inputs, targets = inputs.cuda(), targets.cuda() optimizer.zero_grad() outputs = net(inputs) loss = criterion(outputs, targets) # import pdb; pdb.set_trace() loss.backward() optimizer.step() train_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() desc = ('[LR=%s] Loss: %.3f | Acc: %.3f%% (%d/%d)' % (lr_scheduler.get_lr(optimizer), train_loss / (batch_idx + 1), 100. * correct / total, correct, total)) prog_bar.set_description(desc, refresh=True) writer.add_scalar('iter_%d/train/loss' % iteration, train_loss / (batch_idx + 1), epoch) writer.add_scalar('iter_%d/train/acc' % iteration, 100. * correct / total, epoch) def test(net, loader, criterion, epoch, writer, iteration): net.eval() test_loss = 0 correct = 0 total = 0 desc = ('Loss: %.3f | Acc: %.3f%% (%d/%d)' % (test_loss / (0 + 1), 0, correct, total)) prog_bar = tqdm(enumerate(loader), total=len(loader), desc=desc, leave=True) with torch.no_grad(): for batch_idx, (inputs, targets) in prog_bar: inputs, targets = inputs.cuda(), targets.cuda() outputs = net(inputs) loss = criterion(outputs, targets) test_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() desc = ('Loss: %.3f | Acc: %.3f%% (%d/%d)' % (test_loss / (batch_idx + 1), 100. * correct / total, correct, total)) prog_bar.set_description(desc, refresh=True) # Save checkpoint. acc = 100. * correct / total writer.add_scalar('iter_%d/test/loss' % iteration, test_loss / (batch_idx + 1), epoch) writer.add_scalar('iter_%d/test/acc' % iteration, 100. * correct / total, epoch) return acc def train_once(mb, net, trainloader, testloader, writer, config, ckpt_path, learning_rate, weight_decay, num_epochs, iteration, logger): criterion = nn.CrossEntropyLoss() optimizer = optim.SGD(net.parameters(), lr=learning_rate, momentum=0.9, weight_decay=weight_decay) lr_schedule = {0: learning_rate, int(num_epochs * 0.5): learning_rate * 0.1, int(num_epochs * 0.75): learning_rate * 0.01} lr_scheduler = PresetLRScheduler(lr_schedule) best_acc = 0 best_epoch = 0 for epoch in range(num_epochs): train(net, trainloader, optimizer, criterion, lr_scheduler, epoch, writer, iteration=iteration) test_acc = test(net, testloader, criterion, epoch, writer, iteration) if test_acc > best_acc: print('Saving..') state = { 'net': net, 'acc': test_acc, 'epoch': epoch, 'args': config, 'mask': mb.masks, 'ratio': mb.get_ratio_at_each_layer() } path = os.path.join(ckpt_path, 'finetune_%s_%s%s_r%s_it%d_best.pth.tar' % (config.dataset, config.network, config.depth, config.target_ratio, iteration)) torch.save(state, path) best_acc = test_acc best_epoch = epoch logger.info('Iteration [%d], best acc: %.4f, epoch: %d' % (iteration, best_acc, best_epoch)) def get_exception_layers(net, exception): exc = [] idx = 0 for m in net.modules(): if isinstance(m, (nn.Linear, nn.Conv2d)): if idx in exception: exc.append(m) idx += 1 return tuple(exc) def main(config): # init logger classes = { 'cifar10': 10, 'cifar100': 100, 'mnist': 10, 'tiny_imagenet': 200 } logger, writer = init_logger(config) # build model model = get_network(config.network, config.depth, config.dataset, use_bn=config.get('use_bn', True)) mask = None mb = ModelBase(config.network, config.depth, config.dataset, model) mb.cuda() if mask is not None: mb.register_mask(mask) print_mask_information(mb, logger) # preprocessing # ====================================== get dataloader ====================================== trainloader, testloader = get_dataloader(config.dataset, config.batch_size, 256, 4) # ====================================== fetch configs ====================================== ckpt_path = config.checkpoint_dir num_iterations = config.iterations target_ratio = config.target_ratio normalize = config.normalize # ====================================== fetch exception ====================================== exception = get_exception_layers(mb.model, str_to_list(config.exception, ',', int)) logger.info('Exception: ') for idx, m in enumerate(exception): logger.info(' (%d) %s' % (idx, m)) # ====================================== fetch training schemes ====================================== ratio = 1 - (1 - target_ratio) ** (1.0 / num_iterations) learning_rates = str_to_list(config.learning_rate, ',', float) weight_decays = str_to_list(config.weight_decay, ',', float) training_epochs = str_to_list(config.epoch, ',', int) logger.info('Normalize: %s, Total iteration: %d, Target ratio: %.2f, Iter ratio %.4f.' % (normalize, num_iterations, target_ratio, ratio)) logger.info('Basic Settings: ') for idx in range(len(learning_rates)): logger.info(' %d: LR: %.5f, WD: %.5f, Epochs: %d' % (idx, learning_rates[idx], weight_decays[idx], training_epochs[idx])) # ====================================== start pruning ====================================== iteration = 0 for _ in range(1): logger.info('** Target ratio: %.4f, iter ratio: %.4f, iteration: %d/%d.' % (target_ratio, ratio, 1, num_iterations)) mb.model.apply(weights_init) print("=> Applying weight initialization(%s)." % config.get('init_method', 'kaiming')) print("Iteration of: %d/%d" % (iteration, num_iterations)) masks = GraSP(mb.model, ratio, trainloader, 'cuda', num_classes=classes[config.dataset], samples_per_class=config.samples_per_class, num_iters=config.get('num_iters', 1)) iteration = 0 print('=> Using GraSP') # ========== register mask ================== mb.register_mask(masks) # ========== save pruned network ============ logger.info('Saving..') state = { 'net': mb.model, 'acc': -1, 'epoch': -1, 'args': config, 'mask': mb.masks, 'ratio': mb.get_ratio_at_each_layer() } path = os.path.join(ckpt_path, 'prune_%s_%s%s_r%s_it%d.pth.tar' % (config.dataset, config.network, config.depth, config.target_ratio, iteration)) torch.save(state, path) # ========== print pruning details ============ logger.info('**[%d] Mask and training setting: ' % iteration) print_mask_information(mb, logger) logger.info(' LR: %.5f, WD: %.5f, Epochs: %d' % (learning_rates[iteration], weight_decays[iteration], training_epochs[iteration])) # ========== finetuning ======================= train_once(mb=mb, net=mb.model, trainloader=trainloader, testloader=testloader, writer=writer, config=config, ckpt_path=ckpt_path, learning_rate=learning_rates[iteration], weight_decay=weight_decays[iteration], num_epochs=training_epochs[iteration], iteration=iteration, logger=logger) if __name__ == '__main__': config = init_config() main(config)
13,065
39.83125
119
py
GraSP
GraSP-master/main_finetune_imagenet.py
import argparse import os import random import shutil import time import warnings import sys import torch import torch.nn as nn import torch.nn.parallel import torch.backends.cudnn as cudnn import torch.distributed as dist import torch.optim import torch.multiprocessing as mp import torch.utils.data import torch.utils.data.distributed import torchvision.transforms as transforms import torchvision.datasets as datasets import torchvision.models as models from tensorboardX import SummaryWriter from pprint import pprint model_names = sorted(name for name in models.__dict__ if name.islower() and not name.startswith("__") and callable(models.__dict__[name])) parser = argparse.ArgumentParser(description='PyTorch ImageNet Training') parser.add_argument('data', metavar='DIR', help='path to dataset') parser.add_argument('-a', '--arch', metavar='ARCH', default='resnet18', choices=model_names, help='model architecture: ' + ' | '.join(model_names) + ' (default: resnet18)') parser.add_argument('-j', '--workers', default=4, type=int, metavar='N', help='number of data loading workers (default: 4)') parser.add_argument('--epochs', default=90, type=int, metavar='N', help='number of total epochs to run') parser.add_argument('--start-epoch', default=0, type=int, metavar='N', help='manual epoch number (useful on restarts)') parser.add_argument('-b', '--batch-size', default=256, type=int, metavar='N', help='mini-batch size (default: 256), this is the total ' 'batch size of all GPUs on the current node when ' 'using Data Parallel or Distributed Data Parallel') parser.add_argument('--lr', '--learning-rate', default=0.1, type=float, metavar='LR', help='initial learning rate', dest='lr') parser.add_argument('--momentum', default=0.9, type=float, metavar='M', help='momentum') parser.add_argument('--wd', '--weight-decay', default=1e-4, type=float, metavar='W', help='weight decay (default: 1e-4)', dest='weight_decay') parser.add_argument('-p', '--print-freq', default=10, type=int, metavar='N', help='print frequency (default: 10)') parser.add_argument('--resume', default='', type=str, metavar='PATH', help='path to latest checkpoint (default: none)') parser.add_argument('--resume_pruned', default='', type=str, metavar='PATH', help='path to latest pruned network (default: none)') # resume_pruned parser.add_argument('-e', '--evaluate', dest='evaluate', action='store_true', help='evaluate model on validation set') parser.add_argument('--pretrained', dest='pretrained', action='store_true', help='use pre-trained model') parser.add_argument('--world-size', default=-1, type=int, help='number of nodes for distributed training') parser.add_argument('--rank', default=-1, type=int, help='node rank for distributed training') parser.add_argument('--dist-url', default='tcp://224.66.41.62:23456', type=str, help='url used to set up distributed training') parser.add_argument('--dist-backend', default='nccl', type=str, help='distributed backend') parser.add_argument('--seed', default=None, type=int, help='seed for initializing training. ') parser.add_argument('--gpu', default=None, type=int, help='GPU id to use.') parser.add_argument('--multiprocessing-distributed', action='store_true', help='Use multi-processing distributed training to launch ' 'N processes per node, which has N GPUs. This is the ' 'fastest way to use PyTorch for either single node or ' 'multi node data parallel training') parser.add_argument('--grad_loop', default=1, type=int, help='GPU id to use.') best_acc1 = 0 writer = None def main(): args = parser.parse_args() args.lr /= args.grad_loop if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') if args.gpu is not None: warnings.warn('You have chosen a specific GPU. This will completely ' 'disable data parallelism.') if args.dist_url == "env://" and args.world_size == -1: args.world_size = int(os.environ["WORLD_SIZE"]) args.distributed = args.world_size > 1 or args.multiprocessing_distributed ngpus_per_node = torch.cuda.device_count() if args.multiprocessing_distributed: # Since we have ngpus_per_node processes per node, the total world_size # needs to be adjusted accordingly args.world_size = ngpus_per_node * args.world_size # Use torch.multiprocessing.spawn to launch distributed processes: the # main_worker process function mp.spawn(main_worker, nprocs=ngpus_per_node, args=(ngpus_per_node, args)) else: # Simply call main_worker function main_worker(args.gpu, ngpus_per_node, args) def main_worker(gpu, ngpus_per_node, args): global best_acc1, writer args.gpu = gpu args.save_dir = None if args.gpu is not None: print("Use GPU: {} for training".format(args.gpu)) if args.distributed: if args.dist_url == "env://" and args.rank == -1: args.rank = int(os.environ["RANK"]) if args.multiprocessing_distributed: # For multiprocessing distributed training, rank needs to be the # global rank among all the processes args.rank = args.rank * ngpus_per_node + gpu dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url, world_size=args.world_size, rank=args.rank) # create model def forward_pre_hook(m, x): m.mask.requires_grad_(False) mask = m.mask # mask.requires_grad_(False) # mask.cuda(m.weight.get_device()) m.weight.data.mul_(mask.to(m.weight.get_device())) if args.resume_pruned: if os.path.isfile(args.resume_pruned): print("=> loading checkpoint '{}'".format(args.resume_pruned)) checkpoint = torch.load(args.resume_pruned) model = checkpoint['net'].cpu() masks = checkpoint['mask'] ratio = checkpoint['ratio'] print("=> Ratios:") pprint(ratio) # optimizer.load_state_dict(checkpoint['optimizer']) print("Loaded check point from %s." % args.resume_pruned) print('=> Registering masks for each layer') for m in model.modules(): if isinstance(m, nn.Linear) or isinstance(m, nn.Conv2d): m.mask = nn.Parameter(masks[m]).requires_grad_(False).cpu() m.register_forward_pre_hook(forward_pre_hook) args.save_dir = os.path.join(*args.resume_pruned.split('/')[:-1]) writer = SummaryWriter(args.save_dir) print('=> Will save to %s.' % args.save_dir) else: print("=> no checkpoint found at '{}'".format(args.resume)) elif args.pretrained: print("=> using pre-trained model '{}'".format(args.arch)) model = models.__dict__[args.arch](pretrained=True) else: print("=> creating model '{}'".format(args.arch)) model = models.__dict__[args.arch]() if args.distributed: # For multiprocessing distributed, DistributedDataParallel constructor # should always set the single device scope, otherwise, # DistributedDataParallel will use all available devices. if args.gpu is not None: torch.cuda.set_device(args.gpu) model.cuda(args.gpu) # When using a single GPU per process and per # DistributedDataParallel, we need to divide the batch size # ourselves based on the total number of GPUs we have args.batch_size = int(args.batch_size / ngpus_per_node) args.workers = int(args.workers / ngpus_per_node) model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu]) else: model.cuda() # DistributedDataParallel will divide and allocate batch_size to all # available GPUs if device_ids are not set model = torch.nn.parallel.DistributedDataParallel(model) elif args.gpu is not None: torch.cuda.set_device(args.gpu) model = model.cuda(args.gpu) else: # DataParallel will divide and allocate batch_size to all available GPUs if args.arch.startswith('alexnet') or args.arch.startswith('vgg'): model.features = torch.nn.DataParallel(model.features) model.cuda() else: model = torch.nn.DataParallel(model).cuda() # define loss function (criterion) and optimizer criterion = nn.CrossEntropyLoss().cuda(args.gpu) optimizer = torch.optim.SGD(model.parameters(), args.lr, momentum=args.momentum, weight_decay=args.weight_decay) # optionally resume from a checkpoint if args.resume: def to_cpu(m): if isinstance(m, dict): for k in m.keys(): m[k] = to_cpu(m[k]) return m elif isinstance(m, list): return [to_cpu(_) for _ in m] elif isinstance(m, torch.Tensor): return m.cpu() else: return m if os.path.isfile(args.resume): print("=> loading checkpoint '{}'".format(args.resume)) checkpoint = torch.load(args.resume) args.start_epoch = checkpoint['epoch'] best_acc1 = checkpoint['best_acc1'] if args.gpu is not None: # best_acc1 may be from a checkpoint from a different GPU best_acc1 = best_acc1.to(args.gpu) model.load_state_dict(checkpoint['state_dict']) optimizer.load_state_dict(checkpoint['optimizer']) checkpoint['state_dict'] = to_cpu(checkpoint['state_dict']) checkpoint['optimizer'] = to_cpu(checkpoint['optimizer']) # for k in checkpoint['state_dict'].keys(): # checkpoint['state_dict'][k] = checkpoint['state_dict'][k].cpu() # # for k in checkpoint['optimizer'].keys(): # checkpoint['optimizer'][k] = checkpoint['optimizer'][k].cpu() print("=> loaded checkpoint '{}' (epoch {})" .format(args.resume, checkpoint['epoch'])) del checkpoint else: print("=> no checkpoint found at '{}'".format(args.resume)) cudnn.benchmark = True # Data loading code traindir = os.path.join(args.data, 'train') valdir = os.path.join(args.data, 'val') normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) train_dataset = datasets.ImageFolder( traindir, transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ])) if args.distributed: train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset) else: train_sampler = None train_loader = torch.utils.data.DataLoader( train_dataset, batch_size=args.batch_size, shuffle=(train_sampler is None), num_workers=args.workers, pin_memory=True, sampler=train_sampler) val_loader = torch.utils.data.DataLoader( datasets.ImageFolder(valdir, transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize, ])), batch_size=args.batch_size, shuffle=False, num_workers=args.workers, pin_memory=True) if args.evaluate: validate(val_loader, model, criterion, args) return for epoch in range(args.start_epoch, args.epochs): if args.distributed: train_sampler.set_epoch(epoch) adjust_learning_rate(optimizer, epoch, args) # train for one epoch train_loss, train_top1, train_top5 = train(train_loader, model, criterion, optimizer, epoch, args) # evaluate on validation set acc1, val_loss, val_top1, val_top5 = validate(val_loader, model, criterion, args) # remember best acc@1 and save checkpoint is_best = acc1 > best_acc1 best_acc1 = max(acc1, best_acc1) if not args.multiprocessing_distributed or (args.multiprocessing_distributed and args.rank % ngpus_per_node == 0): save_checkpoint({ 'epoch': epoch + 1, 'arch': args.arch, 'state_dict': model.state_dict(), 'best_acc1': best_acc1, 'optimizer' : optimizer.state_dict(), }, is_best, args=args) # global writer if writer is not None: writer.add_scalar('train/loss', train_loss.avg, epoch) writer.add_scalar('train/top1', train_top1.avg, epoch) writer.add_scalar('train/top5', train_top5.avg, epoch) writer.add_scalar('val/loss', val_loss.avg, epoch) writer.add_scalar('val/top1', val_top1.avg, epoch) writer.add_scalar('val/top5', val_top5.avg, epoch) def train(train_loader, model, criterion, optimizer, epoch, args): total = args.epochs intv = total // 3 lr = args.lr * (0.1 ** (epoch // intv)) batch_time = AverageMeter('Time', ':6.3f') data_time = AverageMeter('Data', ':6.3f') losses = AverageMeter('Loss', ':.4e') top1 = AverageMeter('Acc@1', ':6.2f') top5 = AverageMeter('Acc@5', ':6.2f') progress = ProgressMeter(len(train_loader), batch_time, data_time, losses, top1, top5, prefix="[lr={}, grad_loop={}] Epoch: [{}]".format(lr, args.grad_loop, epoch)) # switch to train mode model.train() end = time.time() optimizer.zero_grad() all_loss = 0 counts = 0 acc_top1 = 0 acc_top5 = 0 all_samples = 0 for i, (input, target) in enumerate(train_loader): # measure data loading time data_time.update(time.time() - end) if args.gpu is not None: input = input.cuda(args.gpu, non_blocking=True) target = target.cuda(args.gpu, non_blocking=True) # compute output all_samples += input.shape[0] output = model(input) loss = criterion(output, target) # measure accuracy and record loss acc1, acc5 = accuracy(output, target, topk=(1, 5)) losses.update(loss.item(), input.size(0)) top1.update(acc1[0], input.size(0)) top5.update(acc5[0], input.size(0)) # compute gradient and do SGD step loss.backward() if (i+1) % args.grad_loop == 0: optimizer.step() optimizer.zero_grad() # measure elapsed time batch_time.update(time.time() - end) end = time.time() all_loss += loss.item() counts += 1 if i % args.print_freq == 0: progress.print(i) return losses, top1, top5 def validate(val_loader, model, criterion, args): batch_time = AverageMeter('Time', ':6.3f') losses = AverageMeter('Loss', ':.4e') top1 = AverageMeter('Acc@1', ':6.2f') top5 = AverageMeter('Acc@5', ':6.2f') progress = ProgressMeter(len(val_loader), batch_time, losses, top1, top5, prefix='Test: ') # switch to evaluate mode model.eval() with torch.no_grad(): end = time.time() for i, (input, target) in enumerate(val_loader): if args.gpu is not None: input = input.cuda(args.gpu, non_blocking=True) target = target.cuda(args.gpu, non_blocking=True) # compute output output = model(input) loss = criterion(output, target) # measure accuracy and record loss acc1, acc5 = accuracy(output, target, topk=(1, 5)) losses.update(loss.item(), input.size(0)) top1.update(acc1[0], input.size(0)) top5.update(acc5[0], input.size(0)) # measure elapsed time batch_time.update(time.time() - end) end = time.time() if i % args.print_freq == 0: progress.print(i) print(' * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}' .format(top1=top1, top5=top5)) return top1.avg, losses, top1, top5 def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', args=None): if args.save_dir is not None: filename = os.path.join(args.save_dir, filename) torch.save(state, filename) if is_best: best_location = 'model_best.pth.tar' if args.save_dir is not None: best_location = os.path.join(args.save_dir, 'model_best.pth.tar') shutil.copyfile(filename, best_location) class AverageMeter(object): """Computes and stores the average and current value""" def __init__(self, name, fmt=':f'): self.name = name self.fmt = fmt self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): self.val = val self.sum += val * n self.count += n self.avg = self.sum / self.count def __str__(self): fmtstr = '{name} {val' + self.fmt + '} ({avg' + self.fmt + '})' return fmtstr.format(**self.__dict__) class ProgressMeter(object): def __init__(self, num_batches, *meters, prefix=""): self.batch_fmtstr = self._get_batch_fmtstr(num_batches) self.meters = meters self.prefix = prefix def print(self, batch): entries = [self.prefix + self.batch_fmtstr.format(batch)] entries += [str(meter) for meter in self.meters] print('\t'.join(entries)) def _get_batch_fmtstr(self, num_batches): num_digits = len(str(num_batches // 1)) fmt = '{:' + str(num_digits) + 'd}' return '[' + fmt + '/' + fmt.format(num_batches) + ']' def adjust_learning_rate(optimizer, epoch, args): total = args.epochs intv = total // 3 """Sets the learning rate to the initial LR decayed by 10 every 30 epochs""" lr = args.lr * (0.1 ** (epoch // intv)) for param_group in optimizer.param_groups: param_group['lr'] = lr def accuracy(output, target, topk=(1,)): """Computes the accuracy over the k top predictions for the specified values of k""" with torch.no_grad(): maxk = max(topk) batch_size = target.size(0) _, pred = output.topk(maxk, 1, True, True) pred = pred.t() correct = pred.eq(target.view(1, -1).expand_as(pred)) res = [] for k in topk: correct_k = correct[:k].view(-1).float().sum(0, keepdim=True) res.append(correct_k.mul_(100.0 / batch_size)) return res if __name__ == '__main__': main()
19,921
38.293886
112
py
GraSP
GraSP-master/pruner/GraSP_ImageNet.py
import torch import torch.autograd as autograd import torch.nn as nn import torch.nn.functional as F import copy import types def count_total_parameters(net): total = 0 for m in net.modules(): if isinstance(m, (nn.Linear, nn.Conv2d)): total += m.weight.numel() return total def count_fc_parameters(net): total = 0 for m in net.modules(): if isinstance(m, (nn.Linear)): total += m.weight.numel() return total def GraSP(net, ratio, train_dataloader, device, num_classes=10, samples_per_class=25, num_iters=1): eps = 1e-10 keep_ratio = 1-ratio old_net = net net = copy.deepcopy(net) net.zero_grad() weights = [] total_parameters = count_total_parameters(net) fc_parameters = count_fc_parameters(net) fc_layers = [] for layer in net.modules(): if isinstance(layer, nn.Conv2d) or isinstance(layer, nn.Linear): if isinstance(layer, nn.Linear): fc_layers.append(layer) weights.append(layer.weight) nn.init.xavier_normal(fc_layers[-1].weight) inputs_one = [] targets_one = [] grad_w = None grad_f = None for w in weights: w.requires_grad_(True) intvs = { 'cifar10': 128, 'cifar100': 256, 'tiny_imagenet': 128, 'imagenet': 20 } print_once = False dataloader_iter = iter(train_dataloader) for it in range(num_iters): print("(1): Iterations %d/%d." % (it, num_iters)) inputs, targets = next(dataloader_iter) N = inputs.shape[0] din = copy.deepcopy(inputs) dtarget = copy.deepcopy(targets) start = 0 intv = 20 while start < N: end = min(start+intv, N) print('(1): %d -> %d.' % (start, end)) inputs_one.append(din[start:end]) targets_one.append(dtarget[start:end]) outputs = net.forward(inputs[start:end].to(device)) / 200 # divide by temperature to make it uniform if print_once: x = F.softmax(outputs) print(x) print(x.max(), x.min()) print_once = False loss = F.cross_entropy(outputs, targets[start:end].to(device)) grad_w_p = autograd.grad(loss, weights, create_graph=False) if grad_w is None: grad_w = list(grad_w_p) else: for idx in range(len(grad_w)): grad_w[idx] += grad_w_p[idx] start = end for it in range(len(inputs_one)): print("(2): Iterations %d/%d." % (it, len(inputs_one))) inputs = inputs_one.pop(0).to(device) targets = targets_one.pop(0).to(device) outputs = net.forward(inputs) / 200 # divide by temperature to make it uniform loss = F.cross_entropy(outputs, targets) grad_f = autograd.grad(loss, weights, create_graph=True) z = 0 count = 0 for layer in net.modules(): if isinstance(layer, nn.Conv2d) or isinstance(layer, nn.Linear): z += (grad_w[count] * grad_f[count]).sum() count += 1 z.backward() grads = dict() old_modules = list(old_net.modules()) for idx, layer in enumerate(net.modules()): if isinstance(layer, nn.Conv2d) or isinstance(layer, nn.Linear): grads[old_modules[idx]] = -layer.weight.data * layer.weight.grad # -theta_q Hg # Gather all scores in a single vector and normalise all_scores = torch.cat([torch.flatten(x) for x in grads.values()]) norm_factor = torch.abs(torch.sum(all_scores)) + eps print("** norm factor:", norm_factor) all_scores.div_(norm_factor) num_params_to_rm = int(len(all_scores) * (1 - keep_ratio)) threshold, _ = torch.topk(all_scores, num_params_to_rm, sorted=True) # import pdb; pdb.set_trace() acceptable_score = threshold[-1] print('** accept: ', acceptable_score) keep_masks = dict() for m, g in grads.items(): keep_masks[m] = ((g / norm_factor) <= acceptable_score).float() print(torch.sum(torch.cat([torch.flatten(x == 1) for x in keep_masks.values()]))) return keep_masks
4,231
30.348148
113
py
GraSP
GraSP-master/pruner/GraSP.py
import torch import torch.autograd as autograd import torch.nn as nn import torch.nn.functional as F import math import copy import types def GraSP_fetch_data(dataloader, num_classes, samples_per_class): datas = [[] for _ in range(num_classes)] labels = [[] for _ in range(num_classes)] mark = dict() dataloader_iter = iter(dataloader) while True: inputs, targets = next(dataloader_iter) for idx in range(inputs.shape[0]): x, y = inputs[idx:idx+1], targets[idx:idx+1] category = y.item() if len(datas[category]) == samples_per_class: mark[category] = True continue datas[category].append(x) labels[category].append(y) if len(mark) == num_classes: break X, y = torch.cat([torch.cat(_, 0) for _ in datas]), torch.cat([torch.cat(_) for _ in labels]).view(-1) return X, y def count_total_parameters(net): total = 0 for m in net.modules(): if isinstance(m, (nn.Linear, nn.Conv2d)): total += m.weight.numel() return total def count_fc_parameters(net): total = 0 for m in net.modules(): if isinstance(m, (nn.Linear)): total += m.weight.numel() return total def GraSP(net, ratio, train_dataloader, device, num_classes=10, samples_per_class=25, num_iters=1, T=200, reinit=True): eps = 1e-10 keep_ratio = 1-ratio old_net = net net = copy.deepcopy(net) # .eval() net.zero_grad() weights = [] total_parameters = count_total_parameters(net) fc_parameters = count_fc_parameters(net) # rescale_weights(net) for layer in net.modules(): if isinstance(layer, nn.Conv2d) or isinstance(layer, nn.Linear): if isinstance(layer, nn.Linear) and reinit: nn.init.xavier_normal(layer.weight) weights.append(layer.weight) inputs_one = [] targets_one = [] grad_w = None for w in weights: w.requires_grad_(True) print_once = False for it in range(num_iters): print("(1): Iterations %d/%d." % (it, num_iters)) inputs, targets = GraSP_fetch_data(train_dataloader, num_classes, samples_per_class) N = inputs.shape[0] din = copy.deepcopy(inputs) dtarget = copy.deepcopy(targets) inputs_one.append(din[:N//2]) targets_one.append(dtarget[:N//2]) inputs_one.append(din[N // 2:]) targets_one.append(dtarget[N // 2:]) inputs = inputs.to(device) targets = targets.to(device) outputs = net.forward(inputs[:N//2])/T if print_once: # import pdb; pdb.set_trace() x = F.softmax(outputs) print(x) print(x.max(), x.min()) print_once = False loss = F.cross_entropy(outputs, targets[:N//2]) # ===== debug ================ grad_w_p = autograd.grad(loss, weights) if grad_w is None: grad_w = list(grad_w_p) else: for idx in range(len(grad_w)): grad_w[idx] += grad_w_p[idx] outputs = net.forward(inputs[N // 2:])/T loss = F.cross_entropy(outputs, targets[N // 2:]) grad_w_p = autograd.grad(loss, weights, create_graph=False) if grad_w is None: grad_w = list(grad_w_p) else: for idx in range(len(grad_w)): grad_w[idx] += grad_w_p[idx] ret_inputs = [] ret_targets = [] for it in range(len(inputs_one)): print("(2): Iterations %d/%d." % (it, num_iters)) inputs = inputs_one.pop(0).to(device) targets = targets_one.pop(0).to(device) ret_inputs.append(inputs) ret_targets.append(targets) outputs = net.forward(inputs)/T loss = F.cross_entropy(outputs, targets) # ===== debug ============== grad_f = autograd.grad(loss, weights, create_graph=True) z = 0 count = 0 for layer in net.modules(): if isinstance(layer, nn.Conv2d) or isinstance(layer, nn.Linear): z += (grad_w[count].data * grad_f[count]).sum() count += 1 z.backward() grads = dict() old_modules = list(old_net.modules()) for idx, layer in enumerate(net.modules()): if isinstance(layer, nn.Conv2d) or isinstance(layer, nn.Linear): grads[old_modules[idx]] = -layer.weight.data * layer.weight.grad # -theta_q Hg # Gather all scores in a single vector and normalise all_scores = torch.cat([torch.flatten(x) for x in grads.values()]) norm_factor = torch.abs(torch.sum(all_scores)) + eps print("** norm factor:", norm_factor) all_scores.div_(norm_factor) num_params_to_rm = int(len(all_scores) * (1-keep_ratio)) threshold, _ = torch.topk(all_scores, num_params_to_rm, sorted=True) # import pdb; pdb.set_trace() acceptable_score = threshold[-1] print('** accept: ', acceptable_score) keep_masks = dict() for m, g in grads.items(): keep_masks[m] = ((g / norm_factor) <= acceptable_score).float() print(torch.sum(torch.cat([torch.flatten(x == 1) for x in keep_masks.values()]))) return keep_masks
5,239
31.75
119
py
GraSP
GraSP-master/models/model_base.py
import torch.nn as nn from collections import OrderedDict from utils.network_utils import get_network from utils.prune_utils import filter_weights class ModelBase(object): def __init__(self, network, depth, dataset, model=None): self._network = network self._depth = depth self._dataset = dataset self.model = model self.masks = None if self.model is None: self.model = get_network(network, depth, dataset) def get_ratio_at_each_layer(self): assert self.masks is not None, 'Masks should be generated first.' res = dict() total = 0 remained = 0 # for m in self.masks.keys(): for m in self.model.modules(): if isinstance(m, (nn.Linear, nn.Conv2d)): mask = self.masks.get(m, None) if mask is not None: res[m] = (mask.sum() / mask.numel()).item() * 100 total += mask.numel() remained += mask.sum().item() else: res[m] = -100.0 total += m.weight.numel() remained += m.weight.numel() res['ratio'] = remained/total * 100 return res def get_unmasked_weights(self): """Return the weights that are unmasked. :return dict, key->module, val->list of weights """ assert self.masks is not None, 'Masks should be generated first.' res = dict() for m in self.masks.keys(): res[m] = filter_weights(m.weight, self.masks[m]) return res def get_masked_weights(self): """Return the weights that are masked. :return dict, key->module, val->list of weights """ assert self.masks is not None, 'Masks should be generated first.' res = dict() for m in self.masks.keys(): res[m] = filter_weights(m.weight, 1-self.masks[m]) return res def register_mask(self, masks=None): # self.masks = None self.unregister_mask() if masks is not None: self.masks = masks assert self.masks is not None, 'Masks should be generated first.' for m in self.masks.keys(): m.register_forward_pre_hook(self._forward_pre_hooks) def unregister_mask(self): for m in self.model.modules(): m._backward_hooks = OrderedDict() m._forward_pre_hooks = OrderedDict() def _forward_pre_hooks(self, m, input): if isinstance(m, nn.Linear) or isinstance(m, nn.Conv2d): # import pdb; pdb.set_trace() mask = self.masks[m] m.weight.data.mul_(mask) else: raise NotImplementedError('Unsupported ' + m) def get_name(self): return '%s_%s%s' % (self._dataset, self._network, self._depth) def train(self): self.model = self.model.train() return self def eval(self): self.model = self.model.eval() return self def cpu(self): self.model = self.model.cpu() return self def cuda(self): self.model = self.model.cuda() return self
3,183
31.161616
73
py
GraSP
GraSP-master/models/base/resnet.py
import math import torch import torch.nn as nn import torch.nn.functional as F import torch.nn.init as init from utils.common_utils import try_cuda from .init_utils import weights_init __all__ = ['resnet'] # , 'resnet20', 'resnet32', 'resnet44', 'resnet56', 'resnet110', 'resnet1202'] _AFFINE = True #_AFFINE = False class LambdaLayer(nn.Module): def __init__(self, lambd): super(LambdaLayer, self).__init__() self.lambd = lambd def forward(self, x): return self.lambd(x) class BasicBlock(nn.Module): expansion = 1 def __init__(self, in_planes, planes, stride=1): super(BasicBlock, self).__init__() self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(planes, affine=_AFFINE) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes, affine=_AFFINE) self.downsample = None self.bn3 = None if stride != 1 or in_planes != planes: self.downsample = nn.Sequential( nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1, stride=stride, bias=False)) self.bn3 = nn.BatchNorm2d(self.expansion * planes, affine=_AFFINE) def forward(self, x): # x: batch_size * in_c * h * w residual = x out = F.relu(self.bn1(self.conv1(x))) out = self.bn2(self.conv2(out)) if self.downsample is not None: residual = self.bn3(self.downsample(x)) out += residual out = F.relu(out) return out class ResNet(nn.Module): def __init__(self, block, num_blocks, num_classes=10): super(ResNet, self).__init__() _outputs = [32, 64, 128] self.in_planes = _outputs[0] self.conv1 = nn.Conv2d(3, _outputs[0], kernel_size=3, stride=1, padding=1, bias=False) self.bn = nn.BatchNorm2d(_outputs[0], affine=_AFFINE) self.layer1 = self._make_layer(block, _outputs[0], num_blocks[0], stride=1) self.layer2 = self._make_layer(block, _outputs[1], num_blocks[1], stride=2) self.layer3 = self._make_layer(block, _outputs[2], num_blocks[2], stride=2) self.linear = nn.Linear(_outputs[2], num_classes) self.apply(weights_init) def _make_layer(self, block, planes, num_blocks, stride): strides = [stride] + [1]*(num_blocks-1) layers = [] for stride in strides: layers.append(block(self.in_planes, planes, stride)) self.in_planes = planes * block.expansion return nn.Sequential(*layers) def forward(self, x): out = F.relu(self.bn(self.conv1(x))) out = self.layer1(out) out = self.layer2(out) out = self.layer3(out) out = F.avg_pool2d(out, out.size()[3]) out = out.view(out.size(0), -1) out = self.linear(out) return out def resnet(depth=32, dataset='cifar10'): assert (depth - 2) % 6 == 0, 'Depth must be = 6n + 2, got %d' % depth n = (depth - 2) // 6 if dataset == 'cifar10': num_classes = 10 elif dataset == 'cifar100': num_classes = 100 elif dataset == 'tiny_imagenet': num_classes = 200 else: raise NotImplementedError('Dataset [%s] is not supported.' % dataset) return ResNet(BasicBlock, [n]*3, num_classes) def test(net): import numpy as np total_params = 0 for x in filter(lambda p: p.requires_grad, net.parameters()): total_params += np.prod(x.data.numpy().shape) print("Total number of params", total_params) print("Total layers", len(list(filter(lambda p: p.requires_grad and len(p.data.size())>1, net.parameters())))) if __name__ == "__main__": for net_name in __all__: if net_name.startswith('resnet'): print(net_name) test(globals()[net_name]()) print()
3,943
33
114
py
GraSP
GraSP-master/models/base/vgg.py
import math import torch import torch.nn as nn from .init_utils import weights_init defaultcfg = { 11: [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512], 13: [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512], 16: [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512], 19: [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512], } class VGG(nn.Module): def __init__(self, dataset='cifar10', depth=19, init_weights=True, cfg=None, affine=True, batchnorm=True): super(VGG, self).__init__() if cfg is None: cfg = defaultcfg[depth] self._AFFINE = affine self.feature = self.make_layers(cfg, batchnorm) self.dataset = dataset if dataset == 'cifar10' or dataset == 'cinic-10': num_classes = 10 elif dataset == 'cifar100': num_classes = 100 elif dataset == 'tiny_imagenet': num_classes = 200 else: raise NotImplementedError("Unsupported dataset " + dataset) self.classifier = nn.Linear(cfg[-1], num_classes) if init_weights: self.apply(weights_init) # if pretrained: # model.load_state_dict(model_zoo.load_url(model_urls['vgg11_bn'])) def make_layers(self, cfg, batch_norm=False): layers = [] in_channels = 3 for v in cfg: if v == 'M': layers += [nn.MaxPool2d(kernel_size=2, stride=2)] else: conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1, bias=False) if batch_norm: layers += [conv2d, nn.BatchNorm2d(v, affine=self._AFFINE), nn.ReLU(inplace=True)] else: layers += [conv2d, nn.ReLU(inplace=True)] in_channels = v return nn.Sequential(*layers) def forward(self, x): x = self.feature(x) if self.dataset == 'tiny_imagenet': x = nn.AvgPool2d(4)(x) else: x = nn.AvgPool2d(2)(x) x = x.view(x.size(0), -1) y = self.classifier(x) return y def _initialize_weights(self): for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.in_channels m.weight.data.normal_(0, math.sqrt(2. / n)) if m.bias is not None: m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): if m.weight is not None: m.weight.data.fill_(1.0) m.bias.data.zero_() elif isinstance(m, nn.Linear): m.weight.data.normal_(0, 0.01) m.bias.data.zero_()
2,823
35.675325
110
py
GraSP
GraSP-master/models/base/init_utils.py
import torch import torch.nn as nn import torch.nn.init as init def weights_init(m): # print('=> weights init') if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') # nn.init.normal_(m.weight, 0, 0.1) if m.bias is not None: m.bias.data.zero_() elif isinstance(m, nn.Linear): # nn.init.xavier_normal(m.weight) nn.init.normal_(m.weight, 0, 0.01) nn.init.constant_(m.bias, 0) elif isinstance(m, nn.BatchNorm2d): # Note that BN's running_var/mean are # already initialized to 1 and 0 respectively. if m.weight is not None: m.weight.data.fill_(1.0) if m.bias is not None: m.bias.data.zero_()
767
32.391304
78
py
GraSP
GraSP-master/utils/prune_utils.py
import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from .common_utils import try_contiguous def _fetch_weights_collections(scores, _prev_masks): weights = [] eps = 1e-10 if _prev_masks is None: for m in scores.keys(): if isinstance(m, (nn.Linear, nn.Conv2d)): w = scores[m].view(-1).data.cpu().numpy() weights.extend(w.tolist()) else: for m in scores.keys(): if isinstance(m, (nn.Linear, nn.Conv2d)): w = scores[m] w = filter_weights(w, _prev_masks[m]) weights.extend(w) return weights def _extract_patches(x, kernel_size, stride, padding): """ :param x: The input feature maps. (batch_size, in_c, h, w) :param kernel_size: the kernel size of the conv filter (tuple of two elements) :param stride: the stride of conv operation (tuple of two elements) :param padding: number of paddings. be a tuple of two elements :return: (batch_size, out_h, out_w, in_c*kh*kw) """ if padding[0] + padding[1] > 0: x = F.pad(x, (padding[1], padding[1], padding[0], padding[0])).data # Actually check dims x = x.unfold(2, kernel_size[0], stride[0]) x = x.unfold(3, kernel_size[1], stride[1]) x = x.transpose_(1, 2).transpose_(2, 3).contiguous() x = x.view( x.size(0), x.size(1), x.size(2), x.size(3) * x.size(4) * x.size(5)) return x def filter_weights(weights, mask): w = weights.view(-1).tolist() m = mask.view(-1).tolist() res = [] for idx in range(len(m)): if m[idx] > 0.5: res.append(w[idx]) return res def _extract_patches(x, kernel_size, stride, padding): """ :param x: The input feature maps. (batch_size, in_c, h, w) :param kernel_size: the kernel size of the conv filter (tuple of two elements) :param stride: the stride of conv operation (tuple of two elements) :param padding: number of paddings. be a tuple of two elements :return: (batch_size, out_h, out_w, in_c*kh*kw) """ if padding[0] + padding[1] > 0: x = F.pad(x, (padding[1], padding[1], padding[0], padding[0])).data # Actually check dims x = x.unfold(2, kernel_size[0], stride[0]) x = x.unfold(3, kernel_size[1], stride[1]) x = x.transpose_(1, 2).transpose_(2, 3).contiguous() x = x.view( x.size(0), x.size(1), x.size(2), x.size(3) * x.size(4) * x.size(5)) return x class ComputeMatGrad: @classmethod def __call__(cls, input, grad_output, layer): if isinstance(layer, nn.Linear): grad = cls.linear(input, grad_output, layer) elif isinstance(layer, nn.Conv2d): grad = cls.conv2d(input, grad_output, layer) else: raise NotImplementedError return grad @staticmethod def linear(input, grad_output, layer): """ :param input: batch_size * input_dim :param grad_output: batch_size * output_dim :param layer: [nn.module] output_dim * input_dim :return: batch_size * output_dim * (input_dim + [1 if with bias]) """ with torch.no_grad(): if layer.bias is not None: input = torch.cat([input, input.new(input.size(0), 1).fill_(1)], 1) input = input.unsqueeze(1) grad_output = grad_output.unsqueeze(2) grad = torch.bmm(grad_output, input) return grad @staticmethod def conv2d(input, grad_output, layer): """ :param input: batch_size * in_c * in_h * in_w :param grad_output: batch_size * out_c * h * w :param layer: nn.module batch_size * out_c * (in_c*k_h*k_w + [1 if with bias]) :return: """ with torch.no_grad(): input = _extract_patches(input, layer.kernel_size, layer.stride, layer.padding) input = input.view(-1, input.size(-1)) # b * hw * in_c*kh*kw grad_output = grad_output.transpose(1, 2).transpose(2, 3) grad_output = try_contiguous(grad_output).view(grad_output.size(0), -1, grad_output.size(-1)) # b * hw * out_c if layer.bias is not None: input = torch.cat([input, input.new(input.size(0), 1).fill_(1)], 1) input = input.view(grad_output.size(0), -1, input.size(-1)) # b * hw * in_c*kh*kw grad = torch.einsum('abm,abn->amn', (grad_output, input)) return grad def fetch_mat_weights(layer, use_patch=False): # -> output_dium * input_dim (kh*kw*in_c + [1 if with bias]) if isinstance(layer, nn.Conv2d): if use_patch: weight = layer.weight.transpose(1, 2).transpose(2, 3) # n_out * kh * kw * inc n_out, k_h, k_w, in_c = weight.size() weight = try_contiguous(weight) weight = weight.view(-1, weight.size(-1)) bias = 0 if layer.bias is not None: copied_bias = torch.cat([layer.bias.unsqueeze(1) for _ in range(k_h*k_w)], 1).view(-1, 1) weight = torch.cat([weight, copied_bias], 1) # layer.bias.unsqueeze(1)], 1) bias = 1 weight = weight.view(n_out, k_h*k_w, in_c+bias) else: weight = layer.weight # n_filters * in_c * kh * kw # weight = weight.transpose(1, 2).transpose(2, 3).contiguous() weight = weight.view(weight.size(0), -1) if layer.bias is not None: weight = torch.cat([weight, layer.bias.unsqueeze(1)], 1) elif isinstance(layer, nn.Linear): weight = layer.weight if layer.bias is not None: weight = torch.cat([weight, layer.bias.unsqueeze(1)], 1) else: raise NotImplementedError return weight class ComputeCovA: @classmethod def compute_cov_a(cls, a, layer): return cls.__call__(a, layer) @classmethod def __call__(cls, a, layer): if isinstance(layer, nn.Linear): cov_a = cls.linear(a, layer) elif isinstance(layer, nn.Conv2d): cov_a = cls.conv2d(a, layer) else: # raise NotImplementedError cov_a = None return cov_a @staticmethod def conv2d(a, layer): batch_size = a.size(0) a = _extract_patches(a, layer.kernel_size, layer.stride, layer.padding) spatial_size = a.size(1) * a.size(2) a = a.view(-1, a.size(-1)) if layer.bias is not None: a = torch.cat([a, a.new(a.size(0), 1).fill_(1)], 1) a = a/spatial_size return a.t() @ (a / batch_size) @staticmethod def linear(a, layer): # a: batch_size * in_dim batch_size = a.size(0) if layer.bias is not None: a = torch.cat([a, a.new(a.size(0), 1).fill_(1)], 1) return a.t() @ (a / batch_size) class ComputeCovG: @classmethod def compute_cov_g(cls, g, layer, batch_averaged=False): """ :param g: gradient :param layer: the corresponding layer :param batch_averaged: if the gradient is already averaged with the batch size? :return: """ # batch_size = g.size(0) return cls.__call__(g, layer, batch_averaged) @classmethod def __call__(cls, g, layer, batch_averaged): if isinstance(layer, nn.Conv2d): cov_g = cls.conv2d(g, layer, batch_averaged) elif isinstance(layer, nn.Linear): cov_g = cls.linear(g, layer, batch_averaged) else: cov_g = None return cov_g @staticmethod def conv2d(g, layer, batch_averaged): # g: batch_size * n_filters * out_h * out_w # n_filters is actually the output dimension (analogous to Linear layer) spatial_size = g.size(2) * g.size(3) batch_size = g.shape[0] g = g.transpose(1, 2).transpose(2, 3) g = try_contiguous(g) g = g.view(-1, g.size(-1)) if batch_averaged: g = g * batch_size g = g * spatial_size cov_g = g.t() @ (g / g.size(0)) return cov_g @staticmethod def linear(g, layer, batch_averaged): # g: batch_size * out_dim batch_size = g.size(0) if batch_averaged: cov_g = g.t() @ (g * batch_size) else: cov_g = g.t() @ (g / batch_size) return cov_g
8,463
33.267206
105
py
GraSP
GraSP-master/utils/data_utils.py
import torch import torchvision import torchvision.transforms as transforms def get_transforms(dataset): transform_train = None transform_test = None if dataset == 'mnist': # transforms.Normalize((0.1307,), (0.3081,)) t = transforms.Normalize((0.5,), (0.5,)) transform_train = transforms.Compose([transforms.ToTensor(),t ]) transform_test = transforms.Compose([transforms.ToTensor(), t]) if dataset == 'cifar10': transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) if dataset == 'cifar100': transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.5071, 0.4867, 0.4408), (0.2675, 0.2565, 0.2761)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.5071, 0.4867, 0.4408), (0.2675, 0.2565, 0.2761)), ]) if dataset == 'cinic-10': # cinic_directory = '/path/to/cinic/directory' cinic_mean = [0.47889522, 0.47227842, 0.43047404] cinic_std = [0.24205776, 0.23828046, 0.25874835] transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(cinic_mean, cinic_std)]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize(cinic_mean, cinic_std)]) if dataset == 'tiny_imagenet': tiny_mean = [0.48024578664982126, 0.44807218089384643, 0.3975477478649648] tiny_std = [0.2769864069088257, 0.26906448510256, 0.282081906210584] transform_train = transforms.Compose([ transforms.RandomCrop(64, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(tiny_mean, tiny_std)]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize(tiny_mean, tiny_std)]) assert transform_test is not None and transform_train is not None, 'Error, no dataset %s' % dataset return transform_train, transform_test def get_dataloader(dataset, train_batch_size, test_batch_size, num_workers=2, root='../data'): transform_train, transform_test = get_transforms(dataset) trainset, testset = None, None if dataset == 'mnist': trainset = torchvision.datasets.MNIST(root=root, train=True, download=True, transform=transform_train) testset = torchvision.datasets.MNIST(root=root, train=False, download=True, transform=transform_test) if dataset == 'cifar10': trainset = torchvision.datasets.CIFAR10(root=root, train=True, download=True, transform=transform_train) testset = torchvision.datasets.CIFAR10(root=root, train=False, download=True, transform=transform_test) if dataset == 'cifar100': trainset = torchvision.datasets.CIFAR100(root=root, train=True, download=True, transform=transform_train) testset = torchvision.datasets.CIFAR100(root=root, train=False, download=True, transform=transform_test) if dataset == 'cinic-10': trainset = torchvision.datasets.ImageFolder(root + '/cinic-10/trainval', transform=transform_train) testset = torchvision.datasets.ImageFolder(root + '/cinic-10/test', transform=transform_test) if dataset == 'tiny_imagenet': num_workers = 16 trainset = torchvision.datasets.ImageFolder(root + '/tiny_imagenet/train', transform=transform_train) testset = torchvision.datasets.ImageFolder(root + '/tiny_imagenet/val', transform=transform_test) assert trainset is not None and testset is not None, 'Error, no dataset %s' % dataset trainloader = torch.utils.data.DataLoader(trainset, batch_size=train_batch_size, shuffle=True, num_workers=num_workers) testloader = torch.utils.data.DataLoader(testset, batch_size=test_batch_size, shuffle=False, num_workers=num_workers) return trainloader, testloader
4,728
44.038095
113
py
GraSP
GraSP-master/utils/common_utils.py
import os import time import json import logging import torch from pprint import pprint from easydict import EasyDict as edict def get_logger(name, logpath, filepath, package_files=[], displaying=True, saving=True): logger = logging.getLogger(name) logger.setLevel(logging.INFO) log_path = logpath + name + time.strftime("-%Y%m%d-%H%M%S") makedirs(log_path) if saving: info_file_handler = logging.FileHandler(log_path) info_file_handler.setLevel(logging.INFO) logger.addHandler(info_file_handler) logger.info(filepath) with open(filepath, 'r') as f: logger.info(f.read()) for f in package_files: logger.info(f) with open(f, 'r') as package_f: logger.info(package_f.read()) if displaying: console_handler = logging.StreamHandler() console_handler.setLevel(logging.INFO) logger.addHandler(console_handler) return logger def makedirs(filename): if not os.path.exists(os.path.dirname(filename)): os.makedirs(os.path.dirname(filename)) def str_to_list(src, delimiter, converter): """Conver a string to list. """ src_split = src.split(delimiter) res = [converter(_) for _ in src_split] return res def get_config_from_json(json_file): """ Get the config from a json file :param json_file: :return: config(namespace) or config(dictionary) """ # parse the configurations from the config json file provided with open(json_file, 'r') as config_file: config_dict = json.load(config_file) config = edict(config_dict) return config, config_dict def process_config(json_file, runs=None): """Process a json file into a config file. Where we can access the value using .xxx Note: we will need to create a similar directory as the config file. """ config, _ = get_config_from_json(json_file) paths = json_file.split('/')[1:-1] summn = [config.exp_name] chekn = [config.exp_name] if runs is not None: summn.append('run_%s' % runs) chekn.append('run_%s' % runs) summn.append("summary/") chekn.append("checkpoint/") summary_dir = ["./runs/pruning"] + paths + summn ckpt_dir = ["./runs/pruning"] + paths + chekn config.summary_dir = os.path.join(*summary_dir) config.checkpoint_dir = os.path.join(*ckpt_dir) print("=> config.summary_dir: %s" % config.summary_dir) print("=> config.checkpoint_dir: %s" % config.checkpoint_dir) return config def try_contiguous(x): if not x.is_contiguous(): x = x.contiguous() return x def try_cuda(x): if torch.cuda.is_available(): x = x.cuda() return x def tensor_to_list(tensor): if len(tensor.shape) == 1: return [tensor[_].item() for _ in range(tensor.shape[0])] else: return [tensor_to_list(tensor[_]) for _ in range(tensor.shape[0])] # ===================================================== # For learning rate schedule # ===================================================== class StairCaseLRScheduler(object): def __init__(self, start_at, interval, decay_rate): self.start_at = start_at self.interval = interval self.decay_rate = decay_rate def __call__(self, optimizer, iteration): start_at = self.start_at interval = self.interval decay_rate = self.decay_rate if (start_at >= 0) \ and (iteration >= start_at) \ and (iteration + 1) % interval == 0: for param_group in optimizer.param_groups: param_group['lr'] *= decay_rate print('[%d]Decay lr to %f' % (iteration, param_group['lr'])) @staticmethod def get_lr(optimizer): for param_group in optimizer.param_groups: lr = param_group['lr'] return lr class PresetLRScheduler(object): """Using a manually designed learning rate schedule rules. """ def __init__(self, decay_schedule): # decay_schedule is a dictionary # which is for specifying iteration -> lr self.decay_schedule = decay_schedule print('=> Using a preset learning rate schedule:') pprint(decay_schedule) self.for_once = True def __call__(self, optimizer, iteration): for param_group in optimizer.param_groups: lr = self.decay_schedule.get(iteration, param_group['lr']) param_group['lr'] = lr @staticmethod def get_lr(optimizer): for param_group in optimizer.param_groups: lr = param_group['lr'] return lr # ======================================================= # For math computation # ======================================================= def prod(l): val = 1 if isinstance(l, list): for v in l: val *= v else: val = val * l return val
4,926
27.982353
76
py
State-Frequency-Memory-stock-prediction
State-Frequency-Memory-stock-prediction-master/test/itosfm.py
# -*- coding: utf-8 -*- from __future__ import absolute_import import numpy as np import theano.tensor as T from keras import backend as K from keras import activations, initializations, regularizers from keras.engine import Layer, InputSpec from keras.layers.recurrent import Recurrent class ITOSFM(Recurrent): def __init__(self, output_dim, freq_dim, hidden_dim, init='glorot_uniform', inner_init='orthogonal', forget_bias_init='one', activation='tanh', inner_activation='hard_sigmoid', W_regularizer=None, U_regularizer=None, b_regularizer=None, dropout_W=0., dropout_U=0., **kwargs): self.output_dim = output_dim self.freq_dim = freq_dim self.hidden_dim = hidden_dim self.init = initializations.get(init) self.inner_init = initializations.get(inner_init) self.forget_bias_init = initializations.get(forget_bias_init) self.activation = activations.get(activation) self.inner_activation = activations.get(inner_activation) self.W_regularizer = regularizers.get(W_regularizer) self.U_regularizer = regularizers.get(U_regularizer) self.b_regularizer = regularizers.get(b_regularizer) self.dropout_W, self.dropout_U = dropout_W, dropout_U if self.dropout_W or self.dropout_U: self.uses_learning_phase = True super(ITOSFM, self).__init__(**kwargs) def build(self, input_shape): self.input_spec = [InputSpec(shape=input_shape)] input_dim = input_shape[2] self.input_dim = input_dim self.states = [None, None, None, None, None] self.W_i = self.init((input_dim, self.hidden_dim), name='{}_W_i'.format(self.name)) self.U_i = self.inner_init((self.hidden_dim, self.hidden_dim), name='{}_U_i'.format(self.name)) self.b_i = K.zeros((self.hidden_dim,), name='{}_b_i'.format(self.name)) self.W_ste = self.init((input_dim, self.hidden_dim), name='{}_W_ste'.format(self.name)) self.U_ste = self.inner_init((self.hidden_dim, self.hidden_dim), name='{}_U_ste'.format(self.name)) self.b_ste = self.forget_bias_init((self.hidden_dim,), name='{}_b_ste'.format(self.name)) self.W_fre = self.init((input_dim, self.freq_dim), name='{}_W_fre'.format(self.name)) self.U_fre = self.inner_init((self.hidden_dim, self.freq_dim), name='{}_U_fre'.format(self.name)) self.b_fre = self.forget_bias_init((self.freq_dim,), name='{}_b_fre'.format(self.name)) self.W_c = self.init((input_dim, self.hidden_dim), name='{}_W_c'.format(self.name)) self.U_c = self.inner_init((self.hidden_dim, self.hidden_dim), name='{}_U_c'.format(self.name)) self.b_c = K.zeros((self.hidden_dim,), name='{}_b_c'.format(self.name)) self.W_o = self.init((input_dim, self.hidden_dim), name='{}_W_o'.format(self.name)) self.U_o = self.inner_init((self.hidden_dim, self.hidden_dim), name='{}_U_o'.format(self.name)) self.b_o = K.zeros((self.hidden_dim,), name='{}_b_o'.format(self.name)) self.U_a = self.inner_init((self.freq_dim, 1), name='{}_U_a'.format(self.name)) self.b_a = K.zeros((self.hidden_dim,), name='{}_b_a'.format(self.name)) self.W_p = self.init((self.hidden_dim, self.output_dim), name='{}_W_p'.format(self.name)) self.b_p = K.zeros((self.output_dim,), name='{}_b_p'.format(self.name)) self.trainable_weights = [self.W_i, self.U_i, self.b_i, self.W_c, self.U_c, self.b_c, self.W_ste, self.U_ste, self.b_ste, self.W_fre, self.U_fre, self.b_fre, self.W_o, self.U_o, self.b_o, self.U_a, self.b_a, self.W_p, self.b_p] if self.initial_weights is not None: self.set_weights(self.initial_weights) del self.initial_weights def get_initial_states(self, x): init_state_h = K.zeros_like(x) init_state_h = K.sum(init_state_h, axis = 1) reducer_s = K.zeros((self.input_dim, self.hidden_dim)) reducer_f = K.zeros((self.hidden_dim, self.freq_dim)) reducer_p = K.zeros((self.hidden_dim, self.output_dim)) init_state_h = K.dot(init_state_h, reducer_s) init_state_p = K.dot(init_state_h, reducer_p) init_state = K.zeros_like(init_state_h) init_freq = K.dot(init_state_h, reducer_f) init_state = K.reshape(init_state, (-1, self.hidden_dim, 1)) init_freq = K.reshape(init_freq, (-1, 1, self.freq_dim)) init_state_S_re = init_state * init_freq init_state_S_im = init_state * init_freq init_state_time = K.cast_to_floatx(0.) initial_states = [init_state_p, init_state_h, init_state_S_re, init_state_S_im, init_state_time] return initial_states def step(self, x, states): p_tm1 = states[0] h_tm1 = states[1] S_re_tm1 = states[2] S_im_tm1 = states[3] time_tm1 = states[4] B_U = states[5] B_W = states[6] frequency = states[7] x_i = K.dot(x * B_W[0], self.W_i) + self.b_i x_ste = K.dot(x * B_W[0], self.W_ste) + self.b_ste x_fre = K.dot(x * B_W[0], self.W_fre) + self.b_fre x_c = K.dot(x * B_W[0], self.W_c) + self.b_c x_o = K.dot(x * B_W[0], self.W_o) + self.b_o i = self.inner_activation(x_i + K.dot(h_tm1 * B_U[0], self.U_i)) ste = self.inner_activation(x_ste + K.dot(h_tm1 * B_U[0], self.U_ste)) fre = self.inner_activation(x_fre + K.dot(h_tm1 * B_U[0], self.U_fre)) ste = K.reshape(ste, (-1, self.hidden_dim, 1)) fre = K.reshape(fre, (-1, 1, self.freq_dim)) f = ste * fre c = i * self.activation(x_c + K.dot(h_tm1 * B_U[0], self.U_c)) time = time_tm1 + 1 omega = K.cast_to_floatx(2*np.pi)* time * frequency re = T.cos(omega) im = T.sin(omega) c = K.reshape(c, (-1, self.hidden_dim, 1)) S_re = f * S_re_tm1 + c * re S_im = f * S_im_tm1 + c * im A = K.square(S_re) + K.square(S_im) A = K.reshape(A, (-1, self.freq_dim)) A_a = K.dot(A * B_U[0], self.U_a) A_a = K.reshape(A_a, (-1, self.hidden_dim)) a = self.activation(A_a + self.b_a) o = self.inner_activation(x_o + K.dot(h_tm1 * B_U[0], self.U_o)) h = o * a p = K.dot(h, self.W_p) + self.b_p return p, [p, h, S_re, S_im, time] def get_constants(self, x): constants = [] constants.append([K.cast_to_floatx(1.) for _ in range(6)]) constants.append([K.cast_to_floatx(1.) for _ in range(7)]) array = np.array([float(ii)/self.freq_dim for ii in range(self.freq_dim)]) constants.append([K.cast_to_floatx(array)]) return constants def get_config(self): config = {"output_dim": self.output_dim, "fre_dim": self.fre_dim, "hidden_dim": self.hidden_dim, "init": self.init.__name__, "inner_init": self.inner_init.__name__, "forget_bias_init": self.forget_bias_init.__name__, "activation": self.activation.__name__, "inner_activation": self.inner_activation.__name__, "W_regularizer": self.W_regularizer.get_config() if self.W_regularizer else None, "U_regularizer": self.U_regularizer.get_config() if self.U_regularizer else None, "b_regularizer": self.b_regularizer.get_config() if self.b_regularizer else None, "dropout_W": self.dropout_W, "dropout_U": self.dropout_U} base_config = super(ITOSFM, self).get_config() return dict(list(base_config.items()) + list(config.items()))
8,589
42.604061
104
py
State-Frequency-Memory-stock-prediction
State-Frequency-Memory-stock-prediction-master/test/build.py
import time import warnings import numpy as np import keras from numpy import newaxis from keras.layers.core import Dense, Activation, Dropout from itosfm import ITOSFM from keras.models import Sequential warnings.filterwarnings("ignore") #Load data from data file, and split the data into training, validation and test set def load_data(filename, step): #load data from the data file day = step data = np.load(filename) data = data[:, :] gt_test = data[:,day:] #data normalization max_data = np.max(data, axis = 1) min_data = np.min(data, axis = 1) max_data = np.reshape(max_data, (max_data.shape[0],1)) min_data = np.reshape(min_data, (min_data.shape[0],1)) data = (2 * data - (max_data + min_data)) / (max_data - min_data) #dataset split train_split = round(0.8 * data.shape[1]) val_split = round(0.9 * data.shape[1]) x_train = data[:,:train_split] y_train = data[:,day:train_split+day] x_val = data[:,:val_split] y_val = data[:,day:val_split+day] x_test = data[:,:-day] y_test = data[:,day:] x_train = np.reshape(x_train, (x_train.shape[0], x_train.shape[1], 1)) x_val = np.reshape(x_val, (x_val.shape[0], x_val.shape[1], 1)) x_test = np.reshape(x_test, (x_test.shape[0], x_test.shape[1], 1)) y_train = np.reshape(y_train, (y_train.shape[0], y_train.shape[1], 1)) y_val = np.reshape(y_val, (y_val.shape[0], y_val.shape[1], 1)) y_test = np.reshape(y_test, (y_test.shape[0], y_test.shape[1], 1)) return [x_train, y_train, x_val, y_val, x_test, y_test, gt_test, max_data, min_data] #build the model def build_model(layers, freq, learning_rate): model = Sequential() model.add(ITOSFM( input_dim=layers[0], hidden_dim=layers[1], output_dim=layers[2], freq_dim = freq, return_sequences=True)) start = time.time() rms = keras.optimizers.RMSprop(lr=learning_rate) model.compile(loss="mse", optimizer="rmsprop") print "Compilation Time : ", time.time() - start return model
2,067
31.825397
88
py
State-Frequency-Memory-stock-prediction
State-Frequency-Memory-stock-prediction-master/train/itosfm.py
# -*- coding: utf-8 -*- from __future__ import absolute_import import numpy as np import theano.tensor as T from keras import backend as K from keras import activations, initializations, regularizers from keras.engine import Layer, InputSpec from keras.layers.recurrent import Recurrent class ITOSFM(Recurrent): def __init__(self, output_dim, freq_dim, hidden_dim, init='glorot_uniform', inner_init='orthogonal', forget_bias_init='one', activation='tanh', inner_activation='hard_sigmoid', W_regularizer=None, U_regularizer=None, b_regularizer=None, dropout_W=0., dropout_U=0., **kwargs): self.output_dim = output_dim self.freq_dim = freq_dim self.hidden_dim = hidden_dim self.init = initializations.get(init) self.inner_init = initializations.get(inner_init) self.forget_bias_init = initializations.get(forget_bias_init) self.activation = activations.get(activation) self.inner_activation = activations.get(inner_activation) self.W_regularizer = regularizers.get(W_regularizer) self.U_regularizer = regularizers.get(U_regularizer) self.b_regularizer = regularizers.get(b_regularizer) self.dropout_W, self.dropout_U = dropout_W, dropout_U if self.dropout_W or self.dropout_U: self.uses_learning_phase = True super(ITOSFM, self).__init__(**kwargs) def build(self, input_shape): self.input_spec = [InputSpec(shape=input_shape)] input_dim = input_shape[2] self.input_dim = input_dim self.states = [None, None, None, None, None] self.W_i = self.init((input_dim, self.hidden_dim), name='{}_W_i'.format(self.name)) self.U_i = self.inner_init((self.hidden_dim, self.hidden_dim), name='{}_U_i'.format(self.name)) self.b_i = K.zeros((self.hidden_dim,), name='{}_b_i'.format(self.name)) self.W_ste = self.init((input_dim, self.hidden_dim), name='{}_W_ste'.format(self.name)) self.U_ste = self.inner_init((self.hidden_dim, self.hidden_dim), name='{}_U_ste'.format(self.name)) self.b_ste = self.forget_bias_init((self.hidden_dim,), name='{}_b_ste'.format(self.name)) self.W_fre = self.init((input_dim, self.freq_dim), name='{}_W_fre'.format(self.name)) self.U_fre = self.inner_init((self.hidden_dim, self.freq_dim), name='{}_U_fre'.format(self.name)) self.b_fre = self.forget_bias_init((self.freq_dim,), name='{}_b_fre'.format(self.name)) self.W_c = self.init((input_dim, self.hidden_dim), name='{}_W_c'.format(self.name)) self.U_c = self.inner_init((self.hidden_dim, self.hidden_dim), name='{}_U_c'.format(self.name)) self.b_c = K.zeros((self.hidden_dim,), name='{}_b_c'.format(self.name)) self.W_o = self.init((input_dim, self.hidden_dim), name='{}_W_o'.format(self.name)) self.U_o = self.inner_init((self.hidden_dim, self.hidden_dim), name='{}_U_o'.format(self.name)) self.b_o = K.zeros((self.hidden_dim,), name='{}_b_o'.format(self.name)) self.U_a = self.inner_init((self.freq_dim, 1), name='{}_U_a'.format(self.name)) self.b_a = K.zeros((self.hidden_dim,), name='{}_b_a'.format(self.name)) self.W_p = self.init((self.hidden_dim, self.output_dim), name='{}_W_p'.format(self.name)) self.b_p = K.zeros((self.output_dim,), name='{}_b_p'.format(self.name)) self.trainable_weights = [self.W_i, self.U_i, self.b_i, self.W_c, self.U_c, self.b_c, self.W_ste, self.U_ste, self.b_ste, self.W_fre, self.U_fre, self.b_fre, self.W_o, self.U_o, self.b_o, self.U_a, self.b_a, self.W_p, self.b_p] if self.initial_weights is not None: self.set_weights(self.initial_weights) del self.initial_weights def get_initial_states(self, x): init_state_h = K.zeros_like(x) init_state_h = K.sum(init_state_h, axis = 1) reducer_s = K.zeros((self.input_dim, self.hidden_dim)) reducer_f = K.zeros((self.hidden_dim, self.freq_dim)) reducer_p = K.zeros((self.hidden_dim, self.output_dim)) init_state_h = K.dot(init_state_h, reducer_s) init_state_p = K.dot(init_state_h, reducer_p) init_state = K.zeros_like(init_state_h) init_freq = K.dot(init_state_h, reducer_f) init_state = K.reshape(init_state, (-1, self.hidden_dim, 1)) init_freq = K.reshape(init_freq, (-1, 1, self.freq_dim)) init_state_S_re = init_state * init_freq init_state_S_im = init_state * init_freq init_state_time = K.cast_to_floatx(0.) initial_states = [init_state_p, init_state_h, init_state_S_re, init_state_S_im, init_state_time] return initial_states def step(self, x, states): p_tm1 = states[0] h_tm1 = states[1] S_re_tm1 = states[2] S_im_tm1 = states[3] time_tm1 = states[4] B_U = states[5] B_W = states[6] frequency = states[7] x_i = K.dot(x * B_W[0], self.W_i) + self.b_i x_ste = K.dot(x * B_W[0], self.W_ste) + self.b_ste x_fre = K.dot(x * B_W[0], self.W_fre) + self.b_fre x_c = K.dot(x * B_W[0], self.W_c) + self.b_c x_o = K.dot(x * B_W[0], self.W_o) + self.b_o i = self.inner_activation(x_i + K.dot(h_tm1 * B_U[0], self.U_i)) ste = self.inner_activation(x_ste + K.dot(h_tm1 * B_U[0], self.U_ste)) fre = self.inner_activation(x_fre + K.dot(h_tm1 * B_U[0], self.U_fre)) ste = K.reshape(ste, (-1, self.hidden_dim, 1)) fre = K.reshape(fre, (-1, 1, self.freq_dim)) f = ste * fre c = i * self.activation(x_c + K.dot(h_tm1 * B_U[0], self.U_c)) time = time_tm1 + 1 omega = K.cast_to_floatx(2*np.pi)* time * frequency re = T.cos(omega) im = T.sin(omega) c = K.reshape(c, (-1, self.hidden_dim, 1)) S_re = f * S_re_tm1 + c * re S_im = f * S_im_tm1 + c * im A = K.square(S_re) + K.square(S_im) A = K.reshape(A, (-1, self.freq_dim)) A_a = K.dot(A * B_U[0], self.U_a) A_a = K.reshape(A_a, (-1, self.hidden_dim)) a = self.activation(A_a + self.b_a) o = self.inner_activation(x_o + K.dot(h_tm1 * B_U[0], self.U_o)) h = o * a p = K.dot(h, self.W_p) + self.b_p return p, [p, h, S_re, S_im, time] def get_constants(self, x): constants = [] constants.append([K.cast_to_floatx(1.) for _ in range(6)]) constants.append([K.cast_to_floatx(1.) for _ in range(7)]) array = np.array([float(ii)/self.freq_dim for ii in range(self.freq_dim)]) constants.append([K.cast_to_floatx(array)]) return constants def get_config(self): config = {"output_dim": self.output_dim, "fre_dim": self.fre_dim, "hidden_dim": self.hidden_dim, "init": self.init.__name__, "inner_init": self.inner_init.__name__, "forget_bias_init": self.forget_bias_init.__name__, "activation": self.activation.__name__, "inner_activation": self.inner_activation.__name__, "W_regularizer": self.W_regularizer.get_config() if self.W_regularizer else None, "U_regularizer": self.U_regularizer.get_config() if self.U_regularizer else None, "b_regularizer": self.b_regularizer.get_config() if self.b_regularizer else None, "dropout_W": self.dropout_W, "dropout_U": self.dropout_U} base_config = super(ITOSFM, self).get_config() return dict(list(base_config.items()) + list(config.items()))
8,589
42.604061
104
py
State-Frequency-Memory-stock-prediction
State-Frequency-Memory-stock-prediction-master/train/build.py
import time import warnings import numpy as np import keras from numpy import newaxis from keras.layers.core import Dense, Activation, Dropout from itosfm import ITOSFM from keras.models import Sequential warnings.filterwarnings("ignore") #Load data from data file, and split the data into training, validation and test set def load_data(filename, step): #load data from the data file day = step data = np.load(filename) data = data[:, :] gt_test = data[:,day:] #data normalization max_data = np.max(data, axis = 1) min_data = np.min(data, axis = 1) max_data = np.reshape(max_data, (max_data.shape[0],1)) min_data = np.reshape(min_data, (min_data.shape[0],1)) data = (2 * data - (max_data + min_data)) / (max_data - min_data) #dataset split train_split = round(0.8 * data.shape[1]) val_split = round(0.9 * data.shape[1]) x_train = data[:,:train_split] y_train = data[:,day:train_split+day] x_val = data[:,:val_split] y_val = data[:,day:val_split+day] x_test = data[:,:-day] y_test = data[:,day:] x_train = np.reshape(x_train, (x_train.shape[0], x_train.shape[1], 1)) x_val = np.reshape(x_val, (x_val.shape[0], x_val.shape[1], 1)) x_test = np.reshape(x_test, (x_test.shape[0], x_test.shape[1], 1)) y_train = np.reshape(y_train, (y_train.shape[0], y_train.shape[1], 1)) y_val = np.reshape(y_val, (y_val.shape[0], y_val.shape[1], 1)) y_test = np.reshape(y_test, (y_test.shape[0], y_test.shape[1], 1)) return [x_train, y_train, x_val, y_val, x_test, y_test, gt_test, max_data, min_data] #build the model def build_model(layers, freq, learning_rate): model = Sequential() model.add(ITOSFM( input_dim=layers[0], hidden_dim=layers[1], output_dim=layers[2], freq_dim = freq, return_sequences=True)) start = time.time() rms = keras.optimizers.RMSprop(lr=learning_rate) model.compile(loss="mse", optimizer="rmsprop") print "Compilation Time : ", time.time() - start return model
2,067
31.825397
88
py
delora
delora-main/setup.py
import os from setuptools import setup, find_packages from setuptools.command.install import install class CustomInstallCommand(install): # This is only run for "python setup.py install" (not for "pip install -e .") def run(self): print("--------------------------------") print("Writing environment variables to .env ...") os.system('echo "export DELORA_ROOT=' + os.getcwd() + '" >> .env') print("...done.") print("--------------------------------") install.run(self) setup(name='DeLORA', version='1.0', author='Julian Nubert (nubertj@ethz.ch)', package_dir={"": "src"}, install_requires=[ 'numpy', 'torch', 'opencv-python', 'pyyaml', 'rospkg' ], scripts=['bin/preprocess_data.py', 'bin/run_rosnode.py', 'bin/run_testing.py', 'bin/run_training.py', 'bin/visualize_pointcloud_normals.py'], license='LICENSE', description='Self-supervised Learning of LiDAR Odometry for Robotic Applications', cmdclass={'install': CustomInstallCommand, }, )
1,125
32.117647
107
py
delora
delora-main/src/utility/linalg.py
#!/usr/bin/env python3 # Modified from: Modar M. Alfadly, https://discuss.pytorch.org/t/covariance-and-gradient-support/16217 import torch def cov(point_neighbors, rowvar=True): ''' Estimate a covariance matrix given data. Covariance indicates the level to which two variables vary together. If we examine N-dimensional samples, `X = [x_1, x_2, ... x_N]^T`, then the covariance matrix element `C_{ij}` is the covariance of `x_i` and `x_j`. The element `C_{ii}` is the variance of `x_i`. Args: m: A 1-D or 2-D array containing multiple variables and observations. Each row of `m` represents a variable, and each column a single observation of all those variables. rowvar: If `rowvar` is True, then each row represents a variable, with observations in the columns. Otherwise, the relationship is transposed: each column represents a variable, while the rows contain observations. Returns: The covariance matrix of the variables. ''' if point_neighbors.dim() > 3: raise ValueError('m has more than 3 dimensions') elif point_neighbors.dim() < 2: point_neighbors = point_neighbors.view(1, -1) if not rowvar and point_neighbors.size(0) != 1: point_neighbors = point_neighbors.t() if point_neighbors.dim() == 3: point_neighbors_not_zero_bool = (point_neighbors[:, 0, :] != torch.zeros(1).to( point_neighbors.device)) | (point_neighbors[:, 1, :] != torch.zeros(1).to( point_neighbors.device)) | (point_neighbors[:, 2, :] != torch.zeros(1).to( point_neighbors.device)) number_neighbours = torch.sum(point_neighbors_not_zero_bool, dim=1) factor = torch.ones(1).to(number_neighbours.device) / (number_neighbours - 1) # Mean assumes that all 9 neighbours are present, but some are zero mean = torch.mean(point_neighbors, dim=2, keepdim=True) * point_neighbors.shape[ 2] / number_neighbours.view(-1, 1, 1) difference = point_neighbors - mean difference.permute(0, 2, 1)[~point_neighbors_not_zero_bool] = torch.zeros(1).to( point_neighbors.device) difference_transpose = difference.permute(0, 2, 1) else: factor = 1.0 / (point_neighbors.shape[1] - 1) mean = torch.mean(point_neighbors, dim=1, keepdim=True) difference = point_neighbors - mean difference_transpose = difference.t() squared_difference = difference.matmul(difference_transpose) # .squeeze() return factor.view(-1, 1, 1) * squared_difference, number_neighbours
2,647
45.45614
102
py
delora
delora-main/src/utility/projection.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import numpy as np import torch import numba class ImageProjectionLayer(torch.nn.Module): def __init__(self, config): super(ImageProjectionLayer, self).__init__() self.device = config["device"] self.config = config self.horizontal_field_of_view = config["horizontal_field_of_view"] # The following will be set while doing projections (different sensors etc.): # self.height_pixel, self.vertical_field_of_view def compute_2D_coordinates(self, point_cloud, width_pixel, height_pixel, vertical_field_of_view): u = ((torch.atan2(point_cloud[:, 1, :], point_cloud[:, 0, :]) - self.horizontal_field_of_view[0]) / ( self.horizontal_field_of_view[1] - self.horizontal_field_of_view[0]) * ( width_pixel - 1)) v = ((torch.atan2(point_cloud[:, 2, :], torch.norm(point_cloud[:, :2, :], dim=1)) - vertical_field_of_view[0]) / ( vertical_field_of_view[1] - vertical_field_of_view[0]) * ( height_pixel - 1)) return u, v # Keeps closest point, because u and v are computed previously based on range-sorted point cloud @staticmethod @numba.njit def remove_duplicate_indices(u, v, occupancy_grid, unique_bool, image_to_pointcloud_indices): for index in range(len(u)): if not occupancy_grid[v[index], u[index]]: occupancy_grid[v[index], u[index]] = True unique_bool[index] = True image_to_pointcloud_indices[0, index, 0] = v[index] image_to_pointcloud_indices[0, index, 1] = u[index] return unique_bool, image_to_pointcloud_indices # input is unordered point cloud scan, e.g. shape [2000,4], with [.,0], [.,1], [.,2], [.,3] being x, y, z, i values # Gets projected to an image # returned point cloud only contains unique points per pixel-discretization, i.e. the closest one def project_to_img(self, point_cloud, dataset): # Get sensor specific parameters width_pixel = self.config[dataset]["horizontal_cells"] height_pixel = self.config[dataset]["vertical_cells"] vertical_field_of_view = self.config[dataset]["vertical_field_of_view"] # Add range to point cloud point_cloud_with_range = torch.zeros( (point_cloud.shape[0], point_cloud.shape[1] + 1, point_cloud.shape[2]), device=self.device, requires_grad=False) point_cloud_with_range[:, :point_cloud.shape[1], :] = point_cloud distance = torch.norm(point_cloud_with_range[:, :3, :], dim=1) point_cloud_with_range[:, -1, :] = distance.detach() del point_cloud # safety such that only correctly sorted one is used # Only keep closest points sort_indices = torch.argsort( point_cloud_with_range[:, point_cloud_with_range.shape[1] - 1, :], dim=1) # for batch_idx in range(len(point_cloud_with_range)): point_cloud_with_range = point_cloud_with_range[:, :, sort_indices[0]] u, v = self.compute_2D_coordinates(point_cloud=point_cloud_with_range, width_pixel=width_pixel, height_pixel=height_pixel, vertical_field_of_view=vertical_field_of_view) inside_fov_bool = (torch.round(u) <= width_pixel - 1) & (torch.round(u) >= 0) & ( torch.round(v) <= height_pixel - 1) & (torch.round(v) >= 0) u_filtered = torch.round(u[inside_fov_bool]) v_filtered = torch.round(v[inside_fov_bool]) point_cloud_with_range = point_cloud_with_range[:, :, inside_fov_bool[0]] occupancy_grid = np.zeros((height_pixel, width_pixel), dtype=bool) # Find pixel to point cloud mapping (for masking in loss later on) image_to_pointcloud_indices = np.zeros((1, len(u_filtered), 2), dtype=int) unique_bool = np.zeros((len(u_filtered)), dtype=bool) unique_bool, image_to_pointcloud_indices = ImageProjectionLayer.remove_duplicate_indices( u=(u_filtered.long().to(torch.device("cpu")).numpy()), v=(v_filtered.long().to(torch.device("cpu")).numpy()), occupancy_grid=occupancy_grid, unique_bool=unique_bool, image_to_pointcloud_indices=image_to_pointcloud_indices) unique_bool = torch.from_numpy(unique_bool).to(self.device) image_to_pointcloud_indices = torch.from_numpy(image_to_pointcloud_indices).to(self.device) u_filtered = u_filtered[unique_bool] v_filtered = v_filtered[unique_bool] point_cloud_with_range = point_cloud_with_range[:, :, unique_bool] image_to_pointcloud_indices = image_to_pointcloud_indices[:, unique_bool] image_representation = torch.zeros( (point_cloud_with_range.shape[0], point_cloud_with_range.shape[1], height_pixel, width_pixel), device=self.device, requires_grad=False) image_representation[:, :, v_filtered.long(), u_filtered.long()] = \ point_cloud_with_range.to(self.device) return image_representation, u, v, sort_indices[inside_fov_bool][ unique_bool], image_to_pointcloud_indices def forward(self, input, dataset): return self.project_to_img(point_cloud=input, dataset=dataset)
5,689
50.727273
119
py
delora
delora-main/src/utility/geometry.py
#!/usr/bin/env python3 # Parts of the code taken from pytorch3d (https://pytorch3d.readthedocs.io/) import torch def _angle_from_tan( axis: str, other_axis: str, data, horizontal: bool, tait_bryan: bool ): """ Extract the first or third Euler angle from the two members of the matrix which are positive constant times its sine and cosine. Args: axis: Axis label "X" or "Y or "Z" for the angle we are finding. other_axis: Axis label "X" or "Y or "Z" for the middle axis in the convention. data: Rotation matrices as tensor of shape (..., 3, 3). horizontal: Whether we are looking for the angle for the third axis, which means the relevant entries are in the same row of the rotation matrix. If not, they are in the same column. tait_bryan: Whether the first and third axes in the convention differ. Returns: Euler Angles in radians for each matrix in data as a tensor of shape (...). """ i1, i2 = {"X": (2, 1), "Y": (0, 2), "Z": (1, 0)}[axis] if horizontal: i2, i1 = i1, i2 even = (axis + other_axis) in ["XY", "YZ", "ZX"] if horizontal == even: return torch.atan2(data[..., i1], data[..., i2]) if tait_bryan: return torch.atan2(-data[..., i2], data[..., i1]) return torch.atan2(data[..., i2], -data[..., i1]) def _index_from_letter(letter: str): if letter == "X": return 0 if letter == "Y": return 1 if letter == "Z": return 2 def matrix_to_euler_angles(matrix, convention: str): """ Convert rotations given as rotation matrices to Euler angles in radians. Args: matrix: Rotation matrices as tensor of shape (..., 3, 3). convention: Convention string of three uppercase letters. Returns: Euler angles in radians as tensor of shape (..., 3). """ if len(convention) != 3: raise ValueError("Convention must have 3 letters.") if convention[1] in (convention[0], convention[2]): raise ValueError(f"Invalid convention {convention}.") for letter in convention: if letter not in ("X", "Y", "Z"): raise ValueError(f"Invalid letter {letter} in convention string.") if matrix.size(-1) != 3 or matrix.size(-2) != 3: raise ValueError(f"Invalid rotation matrix shape f{matrix.shape}.") i0 = _index_from_letter(convention[0]) i2 = _index_from_letter(convention[2]) tait_bryan = i0 != i2 if tait_bryan: central_angle = torch.asin( matrix[..., i0, i2] * (-1.0 if i0 - i2 in [-1, 2] else 1.0) ) else: central_angle = torch.acos(matrix[..., i0, i0]) o = ( _angle_from_tan( convention[0], convention[1], matrix[..., i2], False, tait_bryan ), central_angle, _angle_from_tan( convention[2], convention[1], matrix[..., i0, :], True, tait_bryan ), ) return torch.stack(o, -1)
3,009
33.204545
78
py
delora
delora-main/src/models/model_parts.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import torch import platform if not "2.7" in platform.python_version(): import kornia import utility.geometry import numba import numpy as np class CircularPad(torch.nn.Module): def __init__(self, padding=(1, 1, 0, 0)): super(CircularPad, self).__init__() self.padding = padding def forward(self, input): return torch.nn.functional.pad(input=input, pad=self.padding, mode='circular') class GeometryHandler: def __init__(self, config): self.device = config["device"] @staticmethod def quaternion_to_rot_matrix(quaternion): return kornia.quaternion_to_rotation_matrix(quaternion=quaternion) @staticmethod def angle_axis_to_rot_matrix(euler): return kornia.angle_axis_to_rotation_matrix(angle_axis=euler) @staticmethod def get_transformation_matrix_quaternion(translation, quaternion, device): rotation_matrix = GeometryHandler.quaternion_to_rot_matrix(quaternion=quaternion) transformation_matrix = torch.zeros((rotation_matrix.shape[0], 4, 4), device=device) transformation_matrix[:, :3, :3] = rotation_matrix transformation_matrix[:, 3, 3] = 1 transformation_matrix[:, :3, 3] = translation return transformation_matrix @staticmethod def get_euler_angles_from_matrix(transformation_matrix, device): # "ZYX" corresponding to yaw, pitch, roll euler_angles = utility.geometry.matrix_to_euler_angles( matrix=transformation_matrix[:, :3, :3], convention="ZYX") return euler_angles @staticmethod def get_three_trafo_matrices(euler_angles, translation, device): transformation_yaw = torch.zeros((1, 1, 4, 4), device=device) transformation_yaw[0, 0] = torch.eye(4) transformation_pitch = torch.zeros((1, 1, 4, 4), device=device) transformation_pitch[0, 0] = torch.eye(4) transformation_roll = torch.zeros((1, 1, 4, 4), device=device) transformation_roll[0, 0] = torch.eye(4) # Write in values ## Yaw transformation_yaw[0, 0, 0, 0] = torch.cos(euler_angles[0, 0]) transformation_yaw[0, 0, 1, 1] = torch.cos(euler_angles[0, 0]) transformation_yaw[0, 0, 0, 1] = -torch.sin(euler_angles[0, 0]) transformation_yaw[0, 0, 1, 0] = torch.sin(euler_angles[0, 0]) transformation_yaw[0, 0, :3, 3] = translation # Pitch transformation_pitch[0, 0, 0, 0] = torch.cos(euler_angles[0, 1]) transformation_pitch[0, 0, 2, 2] = torch.cos(euler_angles[0, 1]) transformation_pitch[0, 0, 0, 2] = torch.sin(euler_angles[0, 1]) transformation_pitch[0, 0, 2, 0] = -torch.sin(euler_angles[0, 1]) transformation_pitch[0, 0, :3, 3] = translation # Roll transformation_roll[0, 0, 1, 1] = torch.cos(euler_angles[0, 2]) transformation_roll[0, 0, 2, 2] = torch.cos(euler_angles[0, 2]) transformation_roll[0, 0, 1, 2] = -torch.sin(euler_angles[0, 2]) transformation_roll[0, 0, 2, 1] = torch.sin(euler_angles[0, 2]) transformation_roll[0, 0, :3, 3] = translation return torch.cat((transformation_yaw, transformation_pitch, transformation_roll), dim=1) @staticmethod def get_transformation_matrix_angle_axis(translation, euler, device): rotation_matrix = GeometryHandler.angle_axis_to_rot_matrix(euler=euler) transformation_matrix = torch.zeros((rotation_matrix.shape[0], 4, 4), device=device) transformation_matrix[:, :3, :3] = rotation_matrix transformation_matrix[:, 3, 3] = 1 transformation_matrix[:, :3, 3] = translation return transformation_matrix
3,940
41.836957
96
py
delora
delora-main/src/models/resnet_modified.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. # This model is build on top of the torchvision resnet model. import torch # Default "layers": every layer consists of 2 blocks (consisting of 2 convolutions) # There is only one downsampling done PER LAYER (therefore this differentiation between layers and blocks) class ResNetModified(torch.nn.Module): def __init__(self, in_channels, num_outputs, use_dropout=False, layers=[2, 2, 2, 2], factor_fewer_resnet_channels=1, activation_fct="relu", groups=1, width_per_group=64, replace_stride_with_dilation=None, norm_layer=None): super(ResNetModified, self).__init__() block = BasicBlock self.activation_fct = activation_fct self.inplanes = int(64 / factor_fewer_resnet_channels) self.dilation = 1 if replace_stride_with_dilation is None: # each element in the tuple indicates if we should replace # the 2x2 stride with a dilated convolution instead replace_stride_with_dilation = [False, False, False] if len(replace_stride_with_dilation) != 3: raise ValueError("replace_stride_with_dilation should be None " "or a 3-element tuple, got {}".format(replace_stride_with_dilation)) self.groups = groups self.base_width = width_per_group if use_dropout: self.dropout_values = torch.nn.Dropout(p=0.2, inplace=False) self.dropout_channels = torch.nn.Dropout2d(p=0.2, inplace=False) else: self.dropout_values = torch.nn.Identity() self.dropout_channels = torch.nn.Identity() self.conv1 = torch.nn.Conv2d(in_channels=in_channels, out_channels=self.inplanes, kernel_size=3, stride=(1, 2), padding=(1, 0), bias=False) # self.bn1 = norm_layer(self.inplanes) self.relu = torch.nn.ReLU(inplace=True) self.tanh = torch.nn.Tanh() self.maxpool = torch.nn.MaxPool2d(kernel_size=3, stride=(1, 2), padding=(1, 0)) # No down-sampling in first block self.layer1 = self._make_layer(block=block, planes=int(64 / factor_fewer_resnet_channels), blocks=layers[0]) # From here on down-sampling self.layer2 = self._make_layer(block=block, planes=int(128 / factor_fewer_resnet_channels), blocks=layers[1], stride=(1, 2), dilate=replace_stride_with_dilation[0]) self.layer3 = self._make_layer(block=block, planes=int(256 / factor_fewer_resnet_channels), blocks=layers[2], stride=(1, 2), dilate=replace_stride_with_dilation[1]) self.layer4 = self._make_layer(block=block, planes=int(512 / factor_fewer_resnet_channels), blocks=layers[3], stride=(2, 2), dilate=replace_stride_with_dilation[2]) self.avgpool = torch.nn.AdaptiveAvgPool2d((1, 1)) self.fc = torch.nn.Linear(int(512 / factor_fewer_resnet_channels) * block.expansion, num_outputs) for m in self.modules(): if isinstance(m, torch.nn.Conv2d): torch.nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity=self.activation_fct) elif isinstance(m, (torch.nn.BatchNorm2d, torch.nn.GroupNorm)): torch.nn.init.constant_(m.weight, 1) torch.nn.init.constant_(m.bias, 0) def _make_layer(self, block, planes, blocks, stride=1, dilate=False): downsample = None previous_dilation = self.dilation if dilate: self.dilation *= stride stride = 1 if stride != 1 or self.inplanes != planes * block.expansion: downsample = torch.nn.Sequential( conv1x1(in_planes=self.inplanes, out_planes=planes * block.expansion, stride=stride)) layers = [] layers.append(block(inplanes=self.inplanes, planes=planes, stride=stride, downsample=downsample, groups=self.groups, base_width=self.base_width, dilation=previous_dilation, activation_fct=self.activation_fct)) self.inplanes = planes * block.expansion for _ in range(1, blocks): layers.append(block(inplanes=self.inplanes, planes=planes, groups=self.groups, base_width=self.base_width, dilation=self.dilation, activation_fct=self.activation_fct)) return torch.nn.Sequential(*layers) def _forward_impl(self, x): x = self.dropout_values(x) # Dropout x = torch.nn.functional.pad(input=x, pad=(1, 1, 0, 0), mode='circular') x = self.conv1(x) # x = self.bn1(x) x = self.relu(x) if self.activation_fct == "relu" else self.tanh(x) x = torch.nn.functional.pad(input=x, pad=(1, 1, 0, 0), mode='circular') x = self.maxpool(x) x1 = self.layer1(x) x2 = self.layer2(x1) # x2 = self.dropout_channels(x2) # Dropout x3 = self.layer3(x2) x3 = self.dropout_channels(x3) # Dropout x4 = self.layer4(x3) output = self.avgpool(x4) output = torch.flatten(output, 1) output = self.fc(output) output = self.dropout_values(output) # Dropout features = [x1, x2, x3, x4, output] return features def forward(self, x): return self._forward_impl(x) def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1, padding=0): """3x3 convolution with padding""" return torch.nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=padding, groups=groups, bias=False, dilation=dilation) def conv1x1(in_planes, out_planes, stride=1): """1x1 convolution""" return torch.nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False) class BasicBlock(torch.nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1, base_width=64, dilation=1, activation_fct="relu", norm_layer=None): super(BasicBlock, self).__init__() # if norm_layer is None: # norm_layer = torch.nn.BatchNorm2d if groups != 1 or base_width != 64: raise ValueError('BasicBlock only supports groups=1 and base_width=64') if dilation > 1: raise NotImplementedError("Dilation > 1 not supported in BasicBlock") # Both self.conv1 and self.downsample layers downsample the input when stride != 1 self.conv1 = conv3x3(in_planes=inplanes, out_planes=planes, stride=stride, padding=(1, 0)) # self.bn1 = norm_layer(planes) self.activation = torch.nn.ReLU(inplace=True) if activation_fct == "relu" else torch.nn.Tanh() self.conv2 = conv3x3(in_planes=planes, out_planes=planes, padding=(1, 0)) # self.bn2 = norm_layer(planes) self.downsample = downsample self.stride = stride def forward(self, x): identity = x out = torch.nn.functional.pad(input=x, pad=(1, 1, 0, 0), mode='circular') out = self.conv1(out) # out = self.bn1(out) out = self.activation(out) out = torch.nn.functional.pad(input=out, pad=(1, 1, 0, 0), mode='circular') out = self.conv2(out) # out = self.bn2(out) if self.downsample is not None: identity = self.downsample(x) # no padding required, since 1x1 conv out += identity out = self.activation(out) return out
7,993
43.910112
120
py
delora
delora-main/src/models/model.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. from __future__ import division import torch import models.model_parts import models.resnet_modified class OdometryModel(torch.nn.Module): def __init__(self, config): super(OdometryModel, self).__init__() self.device = config["device"] self.config = config self.pre_feature_extraction = config["pre_feature_extraction"] in_channels = 8 num_feature_extraction_layers = 5 print("Used activation function is: " + self.config["activation_fct"] + ".") if self.config["activation_fct"] != "relu" and self.config["activation_fct"] != "tanh": raise Exception('The specified activation function must be either "relu" or "tanh".') # Here are trainable parameters if config["pre_feature_extraction"]: layers = [] for layer_index in range(num_feature_extraction_layers): input_channels = ( int(in_channels / 2) if layer_index == 0 else (layer_index) * in_channels) layers.append(models.model_parts.CircularPad(padding=(1, 1, 0, 0))) layers.append(torch.nn.Conv2d(in_channels=input_channels, out_channels=(layer_index + 1) * in_channels, kernel_size=3, padding=(1, 0), bias=False)) if self.config["activation_fct"] == "relu": layers.append(torch.nn.ReLU(inplace=True)) else: layers.append(torch.nn.Tanh()) self.feature_extractor = torch.nn.Sequential(*layers) print("Number of trainable parameters in our feature extractor: " + \ f'{sum(p.numel() for p in self.feature_extractor.parameters()):,}') self.resnet = models.resnet_modified.ResNetModified( in_channels=in_channels if not config[ "pre_feature_extraction"] else 2 * num_feature_extraction_layers * in_channels, num_outputs=config["resnet_outputs"], use_dropout=self.config["use_dropout"], layers=self.config["layers"], factor_fewer_resnet_channels=self.config["factor_fewer_resnet_channels"], activation_fct=self.config["activation_fct"]) print("Number of trainable parameters in our ResNet: " + \ f'{sum(p.numel() for p in self.resnet.parameters()):,}') rot_out_features = 4 if self.config["use_single_mlp_at_output"]: self.fully_connected_rot_trans = torch.nn.Sequential( torch.nn.ReLU() if self.config["activation_fct"] == "relu" else torch.nn.Tanh(), torch.nn.Linear(in_features=config["resnet_outputs"], out_features=512), torch.nn.ReLU() if self.config["activation_fct"] == "relu" else torch.nn.Tanh(), torch.nn.Linear(in_features=512, out_features=512), torch.nn.ReLU() if self.config["activation_fct"] == "relu" else torch.nn.Tanh(), torch.nn.Linear(in_features=512, out_features=256), torch.nn.ReLU() if self.config["activation_fct"] == "relu" else torch.nn.Tanh(), torch.nn.Linear(in_features=256, out_features=64), torch.nn.ReLU() if self.config["activation_fct"] == "relu" else torch.nn.Tanh(), torch.nn.Linear(in_features=64, out_features=3 + rot_out_features)) print("Number of trainable parameters in our rot_trans net: " + \ f'{sum(p.numel() for p in self.fully_connected_rot_trans.parameters()):,}') else: self.fully_connected_rotation = torch.nn.Sequential( torch.nn.ReLU() if self.config["activation_fct"] == "relu" else torch.nn.Tanh(), torch.nn.Linear(in_features=config["resnet_outputs"], out_features=100), torch.nn.ReLU() if self.config["activation_fct"] == "relu" else torch.nn.Tanh(), torch.nn.Linear(in_features=100, out_features=rot_out_features)) self.fully_connected_translation = torch.nn.Sequential( torch.nn.ReLU() if self.config["activation_fct"] == "relu" else torch.nn.Tanh(), torch.nn.Linear(in_features=config["resnet_outputs"], out_features=100), torch.nn.ReLU() if self.config["activation_fct"] == "relu" else torch.nn.Tanh(), torch.nn.Linear(in_features=100, out_features=3)) print("Number of trainable parameters in our rotation net: " + \ f'{sum(p.numel() for p in self.fully_connected_rotation.parameters()):,}') print("Number of trainable parameters in our translation net: " + \ f'{sum(p.numel() for p in self.fully_connected_translation.parameters()):,}') # geometry_handler does not contain any trainable parameters self.geometry_handler = models.model_parts.GeometryHandler(config=config) def forward_features(self, image_1, image_2): if self.pre_feature_extraction: x1 = self.feature_extractor(image_1) x2 = self.feature_extractor(image_2) x = torch.cat((x1, x2), dim=1) else: x = torch.cat((image_1, image_2), dim=1) features = self.resnet(x) return features def forward(self, image_1, image_2): x = self.forward_features(image_1=image_1, image_2=image_2) x = x[-1] if self.config["use_single_mlp_at_output"]: x = self.fully_connected_rot_trans(x) x_rotation = x[:, :4] x_translation = x[:, 4:] else: x_rotation = self.fully_connected_rotation(x) x_translation = self.fully_connected_translation(x) x_rotation = x_rotation / torch.norm(x_rotation) return (x_translation, x_rotation)
6,108
51.213675
97
py
delora
delora-main/src/deploy/tester.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import mlflow import numpy as np import pickle import torch import deploy.deployer class Tester(deploy.deployer.Deployer): def __init__(self, config): super(Tester, self).__init__(config=config) self.training = False # Load checkpoint if self.config["checkpoint"]: checkpoint = torch.load(self.config["checkpoint"], map_location=self.device) ## Model weights self.model.load_state_dict(checkpoint["model_state_dict"]) print('\033[92m' + "Model weights loaded from " + self.config["checkpoint"] + "\033[0;0m") else: raise Exception("No checkpoint specified.") print("Batch size set to 1 for the testing.") self.batch_size = 1 # Split up dataset by sequences, since maps are created sequence-wise self.computed_transformations_datasets = [] for dataset in self.config["datasets"]: computed_transformations_sequences = [[] for element in self.config[dataset]["data_identifiers"]] self.computed_transformations_datasets.append(computed_transformations_sequences) def test_dataset(self, dataloader): epoch_losses = { "loss_epoch": 0.0, "loss_point_cloud_epoch": 0.0, "loss_field_of_view_epoch": 0.0, "loss_po2po_epoch": 0.0, "loss_po2pl_epoch": 0.0, "loss_pl2pl_epoch": 0.0, "visible_pixels_epoch": 0.0, } index_of_dataset = 0 index_of_sequence = 0 dataset = self.config["datasets"][0] for index, preprocessed_dicts in enumerate(dataloader): for preprocessed_dict in preprocessed_dicts: # Move data to device for key in preprocessed_dict: if hasattr(preprocessed_dict[key], 'to'): preprocessed_dict[key] = preprocessed_dict[key].to(self.device) if not self.config["inference_only"]: epoch_losses, computed_transformation = self.step( preprocessed_dicts=preprocessed_dicts, epoch_losses=epoch_losses, log_images_bool=not index % 10) else: computed_transformation = self.step( preprocessed_dicts=preprocessed_dicts, epoch_losses=epoch_losses, log_images_bool=not index % 10) for preprocessed_dict in preprocessed_dicts: # Case: we reached the next sequence or the next dataset --> log if preprocessed_dict["index_sequence"] != index_of_sequence or preprocessed_dict[ "index_dataset"] != index_of_dataset: self.log_map(index_of_dataset=index_of_dataset, index_of_sequence=index_of_sequence, dataset=dataset, data_identifier=self.config[dataset]["data_identifiers"][index_of_sequence]) index_of_sequence = preprocessed_dict["index_sequence"] index_of_dataset = preprocessed_dict["index_dataset"] dataset = preprocessed_dict["dataset"] self.computed_transformations_datasets[preprocessed_dict["index_dataset"]][ preprocessed_dict["index_sequence"]].append( computed_transformation.detach().cpu().numpy()) if not index % 10: if not self.config["inference_only"]: print("Index: " + str(index) + " / " + str(len(dataloader)) + ", loss: " + str( epoch_losses["loss_epoch"] / (index + 1)) + ", loss_pc: " + str( epoch_losses["loss_point_cloud_epoch"] / (index + 1)) + ", loss_po2po: " + str( epoch_losses["loss_po2po_epoch"] / (index + 1)) + ", loss_po2pl: " + str( epoch_losses["loss_po2pl_epoch"] / (index + 1)) + ", loss_pl2pl: " + str( epoch_losses["loss_pl2pl_epoch"] / (index + 1)) + ", visible pixels: " + str( epoch_losses["visible_pixels_epoch"] / (index + 1))) if self.config["visualize_images"]: self.log_image(epoch=index, string="_" + dataset) else: print("Index: " + str(index) + " / " + str(len(dataloader)) + ", dataset: " + dataset + ", sequence: " + str(index_of_sequence)) self.log_map(index_of_dataset=index_of_dataset, index_of_sequence=index_of_sequence, dataset=dataset, data_identifier=self.config[dataset]["data_identifiers"][ index_of_sequence]) return epoch_losses def test(self): dataloader = torch.utils.data.DataLoader(dataset=self.dataset, batch_size=self.batch_size, shuffle=False, collate_fn=Tester.list_collate, num_workers=self.config["num_dataloader_workers"], pin_memory=True if self.config[ "device"] == torch.device( "cuda") else False) # Check whether experiment already exists client = mlflow.tracking.MlflowClient() experiment_list = client.list_experiments() id = None for experiment in experiment_list: if experiment.name == self.config["experiment"]: id = experiment.experiment_id if id is None: print("Creating new MLFlow experiment: " + self.config["experiment"]) id = mlflow.create_experiment(self.config["experiment"]) else: print("MLFlow experiment " + self.config["experiment"] + " already exists. Starting a new run within it.") print("----------------------------------") with mlflow.start_run(experiment_id=id, run_name="Test " + self.config["run_name"]): self.log_config() epoch_losses = self.test_dataset(dataloader=dataloader) dataset_index = 0 if not self.config["inference_only"]: epoch_losses["loss_epoch"] /= self.steps_per_epoch epoch_losses["loss_point_cloud_epoch"] /= self.steps_per_epoch epoch_losses["loss_po2po_epoch"] /= self.steps_per_epoch epoch_losses["loss_po2pl_epoch"] /= self.steps_per_epoch epoch_losses["loss_pl2pl_epoch"] /= self.steps_per_epoch epoch_losses["visible_pixels_epoch"] /= self.steps_per_epoch print("Dataset: " + format(dataset_index, '05d') + ", loss: " + str( epoch_losses["loss_epoch"]) + ", loss_point_cloud: " + str( epoch_losses["loss_point_cloud_epoch"])) mlflow.log_metric("loss", float(epoch_losses["loss_epoch"]), step=dataset_index) mlflow.log_metric("loss point cloud", float(epoch_losses["loss_point_cloud_epoch"]), step=dataset_index) mlflow.log_metric("loss po2po", float(epoch_losses["loss_po2po_epoch"]), step=dataset_index) mlflow.log_metric("loss po2pl", float(epoch_losses["loss_po2pl_epoch"]), step=dataset_index) mlflow.log_metric("loss pl2pl", float(epoch_losses["loss_pl2pl_epoch"]), step=dataset_index) mlflow.log_metric("visible pixels", float(epoch_losses["visible_pixels_epoch"]), step=dataset_index)
8,233
49.515337
118
py
delora
delora-main/src/deploy/deployer.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import copy import math import torch import mlflow import numpy as np import utility.plotting import utility.poses import utility.projection import data.dataset import models.model import losses.icp_losses class Deployer(object): def __init__(self, config): torch.cuda.empty_cache() # Parameters and data self.config = config self.device = config["device"] self.batch_size = config["batch_size"] self.dataset = data.dataset.PreprocessedPointCloudDataset(config=config) self.steps_per_epoch = int(len(self.dataset) / self.batch_size) if self.config["mode"] == "testing": self.ground_truth_dataset = data.dataset.PoseDataset(config=self.config) # Projections and model self.img_projection = utility.projection.ImageProjectionLayer(config=config) if self.config["use_jit"]: datasets = self.config["datasets"] ## Need to provide example tensor for torch jit example_tensor = torch.zeros((1, 4, self.config[datasets[0]]["vertical_cells"], self.config[datasets[0]]["horizontal_cells"]), device=self.device) del datasets self.model = torch.jit.trace( models.model.OdometryModel(config=self.config).to(self.device), example_inputs=(example_tensor, example_tensor)) else: self.model = models.model.OdometryModel(config=self.config).to(self.device) # Geometry handler self.geometry_handler = models.model_parts.GeometryHandler(config=config) # Loss and optimizer self.lossTransformation = torch.nn.MSELoss() self.lossPointCloud = losses.icp_losses.ICPLosses(config=self.config) self.lossBCE = torch.nn.BCELoss() self.training_bool = False # Permanent variables for internal use self.log_img_1 = [] self.log_img_2 = [] self.log_img_2_transformed = [] self.log_pointwise_loss = [] self.log_normals_target = [] self.log_normals_transformed_source = [] @staticmethod def list_collate(batch_dicts): data_dicts = [batch_dict for batch_dict in batch_dicts] return data_dicts def create_images(self, preprocessed_data, losses, plotting): ## Create image of target normals image_1_at_normals, _, _, _, _ = self.img_projection(input=torch.cat(( preprocessed_data["scan_1"], preprocessed_data["normal_list_1"]), dim=1), dataset=preprocessed_data["dataset"]) ## Create image for points where normals exist (in source and target) image_2_transformed_and_normals_and_pointwise_loss, _, _, _, _ = \ self.img_projection(input=torch.cat((plotting["scan_2_transformed"], plotting["normals_2_transformed"], losses["loss_po2pl_pointwise"]), dim=1), dataset=preprocessed_data["dataset"]) self.log_pointwise_loss = image_2_transformed_and_normals_and_pointwise_loss[:, 6:9] self.log_normals_target = image_1_at_normals[:, 3:6] self.log_normals_transformed_source = image_2_transformed_and_normals_and_pointwise_loss[ :, 3:6] def log_image(self, epoch, string): utility.plotting.plot_lidar_image( input=[self.log_img_1, self.log_img_2, self.log_img_2_transformed, self.log_pointwise_loss, self.log_normals_target, self.log_normals_transformed_source], label="target", iteration=(epoch + 1) * self.steps_per_epoch if self.training_bool else epoch, path="/tmp/" + self.config["run_name"] + "_" + format(epoch, '05d') + string + ".png", training=self.training_bool) mlflow.log_artifact("/tmp/" + self.config["run_name"] + "_" + format(epoch, '05d') + string + ".png") def log_map(self, index_of_dataset, index_of_sequence, dataset, data_identifier): gt_translations = self.ground_truth_dataset.return_translations( index_of_dataset=index_of_dataset, index_of_sequence=index_of_sequence) gt_poses = self.ground_truth_dataset.return_poses( index_of_dataset=index_of_dataset, index_of_sequence=index_of_sequence) # Extract transformations and absolute poses computed_transformations = self.computed_transformations_datasets[index_of_dataset][ index_of_sequence] computed_poses = utility.poses.compute_poses( computed_transformations=computed_transformations) # Log things to mlflow artifacts utility.poses.write_poses_to_text_file( file_name="/tmp/" + self.config["run_name"] + "_poses_text_file_" + dataset + "_" + format(data_identifier, '02d') + ".txt", poses=computed_poses) mlflow.log_artifact( "/tmp/" + self.config["run_name"] + "_poses_text_file_" + dataset + "_" + format(data_identifier, '02d') + ".txt") np.save( "/tmp/" + self.config["run_name"] + "_transformations_" + dataset + "_" + format(data_identifier, '02d') + ".npy", computed_transformations) np.save( "/tmp/" + self.config["run_name"] + "_poses_" + dataset + "_" + format(data_identifier, '02d') + ".npy", computed_poses) mlflow.log_artifact( "/tmp/" + self.config["run_name"] + "_transformations_" + dataset + "_" + format(data_identifier, '02d') + ".npy") mlflow.log_artifact( "/tmp/" + self.config["run_name"] + "_poses_" + dataset + "_" + format(data_identifier, '02d') + ".npy") utility.plotting.plot_map(computed_poses=computed_poses, path_y="/tmp/" + self.config["run_name"] + "_map_" + dataset + "_" + format( data_identifier, '02d') + "_y.png", path_2d="/tmp/" + self.config["run_name"] + "_map_" + dataset + "_" + format( data_identifier, '02d') + "_2d.png", path_3d="/tmp/" + self.config["run_name"] + "_map_" + dataset + "_" + format( data_identifier, '02d') + "_3d.png", groundtruth=gt_translations, dataset=dataset) mlflow.log_artifact( "/tmp/" + self.config["run_name"] + "_map_" + dataset + "_" + format(data_identifier, '02d') + "_y.png") mlflow.log_artifact( "/tmp/" + self.config["run_name"] + "_map_" + dataset + "_" + format(data_identifier, '02d') + "_2d.png") mlflow.log_artifact( "/tmp/" + self.config["run_name"] + "_map_" + dataset + "_" + format(data_identifier, '02d') + "_3d.png") if gt_poses is not None: utility.plotting.plot_translation_and_rotation( computed_transformations=np.asarray(computed_transformations), path="/tmp/" + self.config["run_name"] + "_plot_trans_rot_" + dataset + "_" + format(data_identifier, '02d') + ".pdf", groundtruth=gt_poses, dataset=dataset) mlflow.log_artifact( "/tmp/" + self.config["run_name"] + "_plot_trans_rot_" + dataset + "_" + format(data_identifier, '02d') + ".pdf") def log_config(self): for dict_entry in self.config: mlflow.log_param(dict_entry, self.config[dict_entry]) def transform_image_to_point_cloud(self, transformation_matrix, image): point_cloud_transformed = torch.matmul(transformation_matrix[:, :3, :3], image[:, :3, :, :].view(-1, 3, image.shape[2] * image.shape[3])) index_array_not_zero = (point_cloud_transformed[:, 0] != torch.zeros(1).to( self.device)) | (point_cloud_transformed[:, 1] != torch.zeros(1).to(self.device)) | ( point_cloud_transformed[:, 2] != torch.zeros(1).to( self.device)) for batch_index in range(len(point_cloud_transformed)): point_cloud_transformed_batch = point_cloud_transformed[batch_index, :, index_array_not_zero[batch_index]] + \ transformation_matrix[batch_index, :3, 3].view(1, 3, -1) point_cloud_transformed = point_cloud_transformed_batch return point_cloud_transformed def rotate_point_cloud_transformation_matrix(self, transformation_matrix, point_cloud): return transformation_matrix[:, :3, :3].matmul(point_cloud[:, :3, :]) def transform_point_cloud_transformation_matrix(self, transformation_matrix, point_cloud): transformed_point_cloud = self.rotate_point_cloud_transformation_matrix( transformation_matrix=transformation_matrix, point_cloud=point_cloud) transformed_point_cloud += transformation_matrix[:, :3, 3].view(-1, 3, 1) return transformed_point_cloud def rotate_point_cloud_euler_vector(self, euler, point_cloud): translation = torch.zeros(3, device=self.device) euler = euler.to(self.device) transformation_matrix = self.geometry_handler.get_transformation_matrix_angle_axis( translation=translation, euler=euler, device=self.device) return self.transform_point_cloud_transformation_matrix( transformation_matrix=transformation_matrix, point_cloud=point_cloud) def augment_input(self, preprocessed_data): # Random rotation if self.config["random_point_cloud_rotations"]: raise Exception("Needs to be verified for larger batches") if self.config["random_rotations_only_yaw"]: direction = torch.zeros((1, 3), device=self.device) direction[0, 2] = 1 else: direction = (torch.rand((1, 3), device=self.device)) direction = direction / torch.norm(direction) magnitude = (torch.rand(1, device=self.device) - 0.5) * ( self.config["magnitude_random_rot"] / 180.0 * torch.Tensor([math.pi]).to( self.device)) euler = direction * magnitude preprocessed_data["scan_2"] = self.rotate_point_cloud_euler_vector( point_cloud=preprocessed_data["scan_2"], euler=euler) preprocessed_data["normal_list_2"] = self.rotate_point_cloud_euler_vector( point_cloud=preprocessed_data["normal_list_2"], euler=euler) return preprocessed_data def normalize_input(self, preprocessed_data): ranges_1 = torch.norm(preprocessed_data["scan_1"], dim=1) ranges_2 = torch.norm(preprocessed_data["scan_2"], dim=1) means_1 = torch.mean(ranges_1, dim=1, keepdim=True) means_2 = torch.mean(ranges_2, dim=1, keepdim=True) # Normalization mean is mean of both means (i.e. independent of number of points of each scan) means_1_2 = torch.cat((means_1, means_2), dim=1) normalization_mean = torch.mean(means_1_2, dim=1) preprocessed_data["scan_1"] /= normalization_mean preprocessed_data["scan_2"] /= normalization_mean preprocessed_data["scaling_factor"] = normalization_mean return preprocessed_data, normalization_mean def step(self, preprocessed_dicts, epoch_losses=None, log_images_bool=False): # Use every batchindex separately images_model_1 = torch.zeros(self.batch_size, 4, self.config[preprocessed_dicts[0]["dataset"]]["vertical_cells"], self.config[preprocessed_dicts[0]["dataset"]]["horizontal_cells"], device=self.device) images_model_2 = torch.zeros_like(images_model_1) for index, preprocessed_dict in enumerate(preprocessed_dicts): if self.training_bool: preprocessed_dict = self.augment_input(preprocessed_data=preprocessed_dict) if self.config["normalization_scaling"]: preprocessed_data, scaling_factor = self.normalize_input(preprocessed_data=preprocessed_dict) # Training / Testing image_1, _, _, point_cloud_indices_1, _ = self.img_projection( input=preprocessed_dict["scan_1"], dataset=preprocessed_dict["dataset"]) image_2, _, _, point_cloud_indices_2, image_to_pc_indices_2 = self.img_projection( input=preprocessed_dict["scan_2"], dataset=preprocessed_dict["dataset"]) ## Only keep points that were projected to image preprocessed_dict["scan_1"] = preprocessed_dict["scan_1"][:, :, point_cloud_indices_1] preprocessed_dict["normal_list_1"] = preprocessed_dict["normal_list_1"][:, :, point_cloud_indices_1] preprocessed_dict["scan_2"] = preprocessed_dict["scan_2"][:, :, point_cloud_indices_2] preprocessed_dict["normal_list_2"] = preprocessed_dict["normal_list_2"][:, :, point_cloud_indices_2] image_model_1 = image_1[0] image_model_2 = image_2[0] # Write projected image to batch images_model_1[index] = image_model_1 images_model_2[index] = image_model_2 images_to_pcs_indices_2 = [image_to_pc_indices_2] self.log_img_1 = image_1[:, :3] self.log_img_2 = image_2[:, :3] # Feed into model as batch (translations, rotation_representation) = self.model(image_1=images_model_1, image_2=images_model_2) computed_transformations = self.geometry_handler.get_transformation_matrix_quaternion( translation=translations, quaternion=rotation_representation, device=self.device) # Following part only done when loss needs to be computed if not self.config["inference_only"]: # Iterate through all transformations and compute loss losses = { "loss_pc": torch.zeros(1, device=self.device), "loss_po2po": torch.zeros(1, device=self.device), "loss_po2pl": torch.zeros(1, device=self.device), "loss_po2pl_pointwise": torch.zeros(1, device=self.device), "loss_pl2pl": torch.zeros(1, device=self.device), } for batch_index, computed_transformation in enumerate(computed_transformations): computed_transformation = torch.unsqueeze(computed_transformation, 0) preprocessed_dict = preprocessed_dicts[batch_index] scan_2_transformed = self.transform_point_cloud_transformation_matrix( transformation_matrix=computed_transformation, point_cloud=preprocessed_dict["scan_2"]) normal_list_2_transformed = self.rotate_point_cloud_transformation_matrix( transformation_matrix=computed_transformation, point_cloud=preprocessed_dict["normal_list_2"]) ## Losses losses_trafo, plotting_step = self.lossPointCloud( source_point_cloud_transformed=scan_2_transformed, source_normal_list_transformed=normal_list_2_transformed, target_point_cloud=preprocessed_dict["scan_1"], target_normal_list=preprocessed_dict["normal_list_1"], compute_pointwise_loss_bool=log_images_bool) losses["loss_po2po"] += losses_trafo["loss_po2po"] losses["loss_po2pl"] += self.config["lambda_po2pl"] * losses_trafo["loss_po2pl"] losses["loss_pl2pl"] += losses_trafo["loss_pl2pl"] losses["loss_pc"] += (losses["loss_po2po"] + losses["loss_po2pl"] + losses["loss_pl2pl"]) ## Image of transformed source point cloud ## Sparser if loss is only taken on image points, only for first index if batch_index == 0: image_2_transformed, u_pixel, v_pixel, _, _ = \ self.img_projection(input=scan_2_transformed, dataset=preprocessed_dict["dataset"]) self.log_img_2_transformed = image_2_transformed losses["loss_po2pl_pointwise"] = losses_trafo["loss_po2pl_pointwise"] plotting = plotting_step if not self.config["unsupervised_at_start"]: target_transformation = torch.eye(4, device=self.device).view(1, 4, 4) loss_transformation = self.lossTransformation(input=computed_transformation, target=target_transformation) losses["loss_pc"] /= self.batch_size losses["loss_po2po"] /= self.batch_size losses["loss_po2pl"] /= self.batch_size losses["loss_pl2pl"] /= self.batch_size if not self.config["unsupervised_at_start"]: loss_transformation /= self.batch_size loss = loss_transformation # Overwrite loss for identity fitting else: loss = losses["loss_pc"] if self.training_bool: loss.backward() self.optimizer.step() if self.config["normalization_scaling"]: for index, preprocessed_dict in enumerate(preprocessed_dicts): computed_transformations[index, :3, 3] *= preprocessed_dict["scaling_factor"] # Visualization if not log_images_bool: _, u_pixel, v_pixel, _, _ = \ self.img_projection(input=scan_2_transformed, dataset=preprocessed_dicts[0]["dataset"]) elif not self.config["po2po_alone"]: self.create_images(preprocessed_data=preprocessed_dicts[0], losses=losses, plotting=plotting) epoch_losses["loss_epoch"] += loss.detach().cpu().numpy() epoch_losses["loss_point_cloud_epoch"] += ( losses["loss_pc"].detach().cpu().numpy()) epoch_losses["loss_po2po_epoch"] += losses["loss_po2po"].detach().cpu().numpy() epoch_losses["loss_po2pl_epoch"] += losses["loss_po2pl"].detach().cpu().numpy() epoch_losses["loss_pl2pl_epoch"] += losses["loss_pl2pl"].detach().cpu().numpy() epoch_losses["visible_pixels_epoch"] += np.sum( ((torch.round(v_pixel.detach()) < self.config[preprocessed_dicts[0]["dataset"]][ "vertical_cells"]) & (v_pixel.detach() > torch.zeros(1).to(self.device))).cpu().numpy()) return epoch_losses, computed_transformations else: if self.config["normalization_scaling"]: for index, preprocessed_dict in enumerate(preprocessed_dicts): computed_transformations[index, :3, 3] *= preprocessed_dict["scaling_factor"] return computed_transformations
20,390
53.231383
119
py
delora
delora-main/src/deploy/trainer.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import time import mlflow import mlflow.pytorch import pickle import torch import numpy as np import qqdm import deploy.deployer class Trainer(deploy.deployer.Deployer): def __init__(self, config): super(Trainer, self).__init__(config=config) self.training_bool = True self.optimizer = torch.optim.Adam(params=self.model.parameters(), lr=self.config["learning_rate"]) # Load checkpoint if self.config["checkpoint"]: checkpoint = torch.load(self.config["checkpoint"], map_location=self.device) ## Model weights self.model.load_state_dict(checkpoint["model_state_dict"]) print('\033[92m' + "Model weights loaded from " + self.config["checkpoint"] + "\033[0;0m") ## Optimizer parameters self.optimizer.load_state_dict(checkpoint["optimizer_state_dict"]) print('\033[92m' + "Optimizer parameters loaded from " + self.config["checkpoint"] + "\033[0;0m") ## Directly train unsupervised in that case, since model is pretrained self.config["unsupervised_at_start"] = True if self.config["inference_only"]: print( "Config error: Inference only does not make sense during training. Changing to inference_only=False.") self.config["inference_only"] = False def train_epoch(self, epoch, dataloader): epoch_losses = { "loss_epoch": 0.0, "loss_point_cloud_epoch": 0.0, "loss_field_of_view_epoch": 0.0, "loss_po2po_epoch": 0.0, "loss_po2pl_epoch": 0.0, "loss_pl2pl_epoch": 0.0, "visible_pixels_epoch": 0.0, "loss_yaw_pitch_roll_epoch": np.zeros(3), "loss_true_trafo_epoch": 0.0, } counter = 0 qqdm_dataloader = qqdm.qqdm(dataloader, desc=qqdm.format_str('blue', 'Epoch ' + str(epoch))) for preprocessed_dicts in qqdm_dataloader: # Load corresponnding preprocessed kd_tree for preprocessed_dict in preprocessed_dicts: # Move data to devices: for key in preprocessed_dict: if hasattr(preprocessed_dict[key], 'to'): preprocessed_dict[key] = preprocessed_dict[key].to(self.device) self.optimizer.zero_grad() epoch_losses, _ = ( self.step( preprocessed_dicts=preprocessed_dicts, epoch_losses=epoch_losses, log_images_bool=counter == self.steps_per_epoch - 1 or counter == 0)) # Plotting and logging --> only first one in batch preprocessed_data = preprocessed_dicts[0] # Plot at very beginning to see initial state of the network if epoch == 0 and counter == 0 and not self.config["po2po_alone"]: self.log_image(epoch=epoch, string="_start" + "_" + preprocessed_data["dataset"]) qqdm_dataloader.set_infos({'loss': f'{float(epoch_losses["loss_epoch"] / (counter + 1)):.6f}', 'loss_point_cloud': f'{float(epoch_losses["loss_point_cloud_epoch"] / (counter + 1)):.6f}', 'loss_po2po': f'{float(epoch_losses["loss_po2po_epoch"] / (counter + 1)):.6f}', 'loss_po2pl': f'{float(epoch_losses["loss_po2pl_epoch"] / (counter + 1)):.6f}', 'loss_pl2pl': f'{float(epoch_losses["loss_pl2pl_epoch"] / (counter + 1)):.6f}', 'visible_pixels': f'{float(epoch_losses["visible_pixels_epoch"] / (counter + 1)):.6f}'}) counter += 1 return epoch_losses def train(self): dataloader = torch.utils.data.DataLoader(dataset=self.dataset, batch_size=self.batch_size, shuffle=True, collate_fn=Trainer.list_collate, num_workers=self.config["num_dataloader_workers"], pin_memory=True if self.config[ "device"] == torch.device("cuda") else False) # Check whether experiment already exists client = mlflow.tracking.MlflowClient() experiment_list = client.list_experiments() id = None for experiment in experiment_list: if experiment.name == self.config["experiment"]: id = experiment.experiment_id if id is None: print("Creating new MLFlow experiment: " + self.config["experiment"]) id = mlflow.create_experiment(self.config["experiment"]) else: print("MLFlow experiment " + self.config["experiment"] + " already exists. Starting a new run within it.") print("----------------------------------") with mlflow.start_run(experiment_id=id, run_name="Training: " + self.config["training_run_name"]): self.log_config() for epoch in range(10000): # Train for 1 epoch epoch_losses = self.train_epoch(epoch=epoch, dataloader=dataloader) # Compute metrics epoch_losses["loss_epoch"] /= self.steps_per_epoch epoch_losses["loss_point_cloud_epoch"] /= self.steps_per_epoch epoch_losses["loss_po2po_epoch"] /= self.steps_per_epoch epoch_losses["loss_po2pl_epoch"] /= self.steps_per_epoch epoch_losses["loss_pl2pl_epoch"] /= self.steps_per_epoch epoch_losses["visible_pixels_epoch"] /= self.steps_per_epoch # Print update print("--------------------------") print("Epoch Summary: " + format(epoch, '05d') + ", loss: " + str( epoch_losses["loss_epoch"]) + ", unsupervised: " + str( self.config["unsupervised_at_start"])) # Logging print("Logging metrics and artifacts...") # Log metrics mlflow.log_metric("loss", float(epoch_losses["loss_epoch"]), step=epoch) mlflow.log_metric("loss point cloud", float(epoch_losses["loss_point_cloud_epoch"]), step=epoch) mlflow.log_metric("loss po2po", float(epoch_losses["loss_po2po_epoch"]), step=epoch) mlflow.log_metric("loss po2pl", float(epoch_losses["loss_po2pl_epoch"]), step=epoch) mlflow.log_metric("loss pl2pl", float(epoch_losses["loss_pl2pl_epoch"]), step=epoch) mlflow.log_metric("visible pixels", float(epoch_losses["visible_pixels_epoch"]), step=epoch) # Save latest checkpoint, and create checkpoint backup all 5 epochs ## Every epoch --> will always be overwritten by latest version torch.save({ 'epoch': epoch, 'model_state_dict': self.model.state_dict(), 'optimizer_state_dict': self.optimizer.state_dict(), 'loss': float(epoch_losses["loss_epoch"]), 'parameters': self.config }, "/tmp/" + self.config["training_run_name"] + "_latest_checkpoint.pth") mlflow.log_artifact("/tmp/" + self.config["training_run_name"] + "_latest_checkpoint.pth") ## All 5 epochs --> will be logged permanently in MLFlow if not epoch % 5: torch.save({ 'epoch': epoch, 'model_state_dict': self.model.state_dict(), 'optimizer_state_dict': self.optimizer.state_dict(), 'loss': float(epoch_losses["loss_epoch"]), 'parameters': self.config }, "/tmp/" + self.config["training_run_name"] + "_checkpoint_epoch_" + str(epoch) + ".pth") mlflow.log_artifact( "/tmp/" + self.config["training_run_name"] + "_checkpoint_epoch_" + str(epoch) + ".pth") # Save latest pickled full model if not self.config["use_jit"]: mlflow.pytorch.log_model(self.model, "latest_model_pickled") if self.config["visualize_images"] and not self.config["po2po_alone"]: self.log_image(epoch=epoch, string="_image") print("...done.") if not self.config["unsupervised_at_start"] and epoch_losses["loss_epoch"] < 1e-2: self.config["unsupervised_at_start"] = True print("Loss has decreased sufficiently. Switching to unsupervised mode.")
9,336
48.930481
130
py
delora
delora-main/src/ros_utils/rosbag_pcl_extractor.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import numpy as np import rosbag import yaml import sensor_msgs.msg import sensor_msgs.point_cloud2 import torch class RosbagToPCLExtractor: def __init__(self, rosbag_file, topic, config, preprocessing_fct): self.rosbag_file = rosbag_file self.topic = topic print("Loading rosbag " + self.rosbag_file + "...") self.bag = rosbag.Bag(self.rosbag_file) print("...done.") # Print information and check rosbag ----- self.num_samples = 0 info_dict = yaml.load(self.bag._get_yaml_info()) print("Duration of the bag: " + str(info_dict["duration"])) for topic_messages in info_dict["topics"]: if topic_messages["topic"] == self.topic: self.num_samples = topic_messages["messages"] if self.num_samples > 0: print("Number of messages for topic " + self.topic + ": " + str(self.num_samples)) else: raise Exception("Topic " + self.topic + " is not present in the given rosbag (" + self.rosbag_file + ").") # ----------------------------------------- self.preprocessing_fct = preprocessing_fct def ros_to_pcl(self, ros_cloud): points_list = [] for data in sensor_msgs.point_cloud2.read_points(ros_cloud, skip_nans=True): points_list.append([data[0], data[1], data[2], data[3]]) points_list = np.asarray(points_list) return points_list def preprocess_rosbag(self): for index, (topic, msg, t) in enumerate(self.bag.read_messages(topics=[self.topic])): if not index % 10: print("Preprocessing scan " + str( index) + "/" + str(self.num_samples) + " from the point cloud " + self.rosbag_file + ".") scan = self.ros_to_pcl(msg) # filter out noisy points scan = scan[(scan[:, 0] != 0.0) & (scan[:, 1] != 0.0) & (scan[:, 2] != 0.0)] scan_range = np.linalg.norm(scan[:, :3], axis=1) scan = scan[scan_range > 0.3] scan = torch.from_numpy(scan).to(torch.device("cpu")).transpose(0, 1).unsqueeze(0) # Apply preprocessing self.preprocessing_fct(scan=scan, index=index) self.bag.close()
2,500
38.698413
118
py
delora
delora-main/src/ros_utils/odometry_publisher.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. from __future__ import division import time import cv2 import geometry_msgs.msg import nav_msgs.msg import numpy as np import ros_numpy import rospy import sensor_msgs.msg import tf2_ros import tf.transformations import torch import models.model import models.model_parts import utility.projection import ros_utils.odometry_integrator # Assumes the dataset in config["datasets"][0] class OdometryPublisher: def __init__(self, config): # Variables self.config = config self.device = config["device"] # ROS Topics and Frames self.lidar_topic = config["lidar_topic"] self.lidar_frame = config["lidar_frame"] # Model self.img_projection = utility.projection.ImageProjectionLayer(config=config) if self.config["use_jit"]: self.model = torch.jit.trace( models.model.OdometryModel(config=self.config).to(self.device), example_inputs=(torch.zeros((1, 4, 16, 720), device=self.device), torch.zeros((1, 4, 16, 720), device=self.device))) else: self.model = models.model.OdometryModel(config=self.config).to(self.device) self.model.load_state_dict(torch.load(self.config["checkpoint"], map_location=self.device)["model_state_dict"]) # ROS publisher and subscriber ## Publisher self.odometry_publisher = rospy.Publisher("/delora/odometry", nav_msgs.msg.Odometry, queue_size=10) ## Node rospy.init_node('LiDAR_odometry_publisher', anonymous=True) self.rate = rospy.Rate(10) # TF Integrator if self.config["integrate_odometry"]: self.odometry_integrator = ros_utils.odometry_integrator.OdometryIntegrator(config=self.config) ## Variables self.scaling_factor = 1.0 self.point_cloud_t_1 = None self.point_cloud_t = None self.image_t_1 = None self.image_t = None self.odometry_ros = nav_msgs.msg.Odometry() self.translation_ros = geometry_msgs.msg.Point() self.quaternion_ros = geometry_msgs.msg.Quaternion() # Geometry handler self.geometry_handler = models.model_parts.GeometryHandler(config=config) def visualize_image(self, input): print("Visualizing!") image = input image = np.asarray((image[0]))[:, ::-1] range_image = np.sqrt(image[0] ** 2 + image[1] ** 2 + image[2] ** 2) scaled_range_image = (255.0 / np.max(range_image) * range_image).astype(np.uint8) color_image = cv2.applyColorMap(scaled_range_image, cv2.COLORMAP_HSV) color_image = cv2.cvtColor(color_image, cv2.COLOR_BGR2RGB) color_image[range_image == 0] = 0 cv2.imshow("test", color_image) cv2.waitKey() def filter_scans(self, scan, inverse=False): scan = np.transpose(scan, (2, 1, 0)) # Filter out close, noisy points scan = scan[(scan[:, 0, 0] != 0.0) & (scan[:, 1, 0] != 0.0) & (scan[:, 2, 0] != 0.0)] scan_range = np.linalg.norm(scan[:, :3, 0], axis=1) scan = scan[scan_range > 0.3] scan = np.transpose(scan, (2, 1, 0)) return scan def normalize_input(self, input_1, input_2): range_1 = torch.norm(input_1, dim=1) range_2 = torch.norm(input_2, dim=1) range_1_2 = torch.cat((range_1, range_2), dim=1) mean_range = torch.mean(range_1_2) input_1 /= mean_range input_2 /= mean_range return mean_range.cpu().numpy() def quat2mat(self, quat): x, y, z, w = quat[:, 0], quat[:, 1], quat[:, 2], quat[:, 3] B = quat.size(0) w2, x2, y2, z2 = w.pow(2), x.pow(2), y.pow(2), z.pow(2) wx, wy, wz = w * x, w * y, w * z xy, xz, yz = x * y, x * z, y * z rotMat = torch.stack([w2 + x2 - y2 - z2, 2 * xy - 2 * wz, 2 * wy + 2 * xz, 2 * wz + 2 * xy, w2 - x2 + y2 - z2, 2 * yz - 2 * wx, 2 * xz - 2 * wy, 2 * wx + 2 * yz, w2 - x2 - y2 + z2], dim=1).reshape( B, 3, 3) return rotMat def predict_and_publish(self, header): with torch.no_grad(): if self.point_cloud_t_1 is not None: # Normalize if self.config["normalization_scaling"]: self.scaling_factor = self.normalize_input(input_1=self.point_cloud_t, input_2=self.point_cloud_t_1) # Project self.image_t_1, _, _, _, _ = self.img_projection( input=self.point_cloud_t_1, dataset=self.config["datasets"][0]) self.image_t, _, _, _, _ = self.img_projection( input=self.point_cloud_t, dataset=self.config["datasets"][0]) ##self.visualize_image(input=np.concatenate((self.image_t_1.cpu(), self.image_t.cpu()), axis=2)) # Predict t = time.time() torch.cuda.synchronize() (translation_1, rot_repr_1) = self.model(image_1=self.image_t_1, image_2=self.image_t) computed_transformation = self.geometry_handler.get_transformation_matrix_quaternion( translation=translation_1, quaternion=rot_repr_1, device=self.device) torch.cuda.synchronize() print("The prediction took: " + str((time.time() - t) * 1000) + "ms.") quaternion_1 = tf.transformations.quaternion_from_matrix( computed_transformation[0].cpu().numpy()) translation = translation_1[0].cpu().numpy() * self.scaling_factor quaternion = quaternion_1 # Publish messages self.translation_ros.x = translation[0] self.translation_ros.y = translation[1] self.translation_ros.z = translation[2] self.quaternion_ros.x = quaternion[0] self.quaternion_ros.y = quaternion[1] self.quaternion_ros.z = quaternion[2] self.quaternion_ros.w = quaternion[3] self.odometry_ros.header = header self.odometry_ros.header.frame_id = self.lidar_frame self.odometry_ros.pose.pose.position = self.translation_ros self.odometry_ros.pose.pose.orientation = self.quaternion_ros self.odometry_publisher.publish(self.odometry_ros) # Update TF if self.config["integrate_odometry"]: self.odometry_integrator.integrate(header=header, quaternion=quaternion, translation=translation) self.point_cloud_t_1 = self.point_cloud_t * self.scaling_factor def subscriber_callback(self, data): structured_array = ros_numpy.numpify(data) x = structured_array['x'].view(np.float32) y = structured_array['y'].view(np.float32) z = structured_array['z'].view(np.float32) point_cloud_t_numpy = np.expand_dims(np.concatenate( (np.expand_dims(x, axis=0), np.expand_dims(y, axis=0), np.expand_dims(z, axis=0)), axis=0), axis=0) # Align the point cloud with correct coordinate system self.point_cloud_t = torch.from_numpy(self.filter_scans(point_cloud_t_numpy)) # Compute odometry estimate self.predict_and_publish(data.header) def publish_odometry(self): rospy.Subscriber(self.lidar_topic, sensor_msgs.msg.PointCloud2, self.subscriber_callback) rospy.spin()
7,939
39.10101
119
py
delora
delora-main/src/ros_utils/odometry_integrator.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. from __future__ import division import time import cv2 import geometry_msgs.msg import nav_msgs.msg import numpy as np import ros_numpy import rospy import sensor_msgs.msg import tf2_ros import tf.transformations import torch import models.model import models.model_parts import utility.projection # Assumes the dataset in config["datasets"][0] class OdometryIntegrator: def __init__(self, config): # Variables self.config = config self.device = config["device"] # ROS Topics and Frames self.lidar_topic = config["lidar_topic"] self.lidar_frame = config["lidar_frame"] self.odom_frame = "delora_odom" self.world_frame = "world" # ROS publisher and subscriber ## Publisher self.tf_broadcaster = tf2_ros.TransformBroadcaster() # Transformations self.T_0_t = np.expand_dims(np.eye(4), axis=0) self.R_pc_odom = np.zeros((1, 3, 3)) self.R_pc_odom[0] = np.eye(3) self.T_pc_odom = geometry_msgs.msg.TransformStamped() self.init_pc_odom() self.T_world_pc = geometry_msgs.msg.TransformStamped() self.init_world_pc() # First TF broadcast self.tf_broadcaster.sendTransform(self.T_world_pc) self.tf_broadcaster.sendTransform(self.T_pc_odom) def init_pc_odom(self): self.T_pc_odom.header.frame_id = self.lidar_frame self.T_pc_odom.child_frame_id = self.odom_frame self.T_pc_odom.transform.translation.x = 0.0 self.T_pc_odom.transform.translation.y = 0.0 self.T_pc_odom.transform.translation.z = 0.0 T = np.eye(4) T[:3, :3] = self.R_pc_odom quaternion = tf.transformations.quaternion_from_matrix(T) self.T_pc_odom.transform.rotation.x = quaternion[0] self.T_pc_odom.transform.rotation.y = quaternion[1] self.T_pc_odom.transform.rotation.z = quaternion[2] self.T_pc_odom.transform.rotation.w = quaternion[3] def init_world_pc(self): self.T_world_pc.header.frame_id = self.world_frame self.T_world_pc.child_frame_id = self.lidar_frame self.T_world_pc.transform.translation.x = 0.0 self.T_world_pc.transform.translation.y = 0.0 self.T_world_pc.transform.translation.z = 0.0 self.T_world_pc.transform.rotation.x = 0.0 self.T_world_pc.transform.rotation.y = 0.0 self.T_world_pc.transform.rotation.z = 0.0 self.T_world_pc.transform.rotation.w = 1.0 def update_transformation(self, quaternion, translation): T_t_1_t = np.zeros((1, 4, 4)) T_t_1_t[0, :4, :4] = np.eye(4) T_t_1_t[0] = tf.transformations.quaternion_matrix(quaternion) T_t_1_t[0, :3, 3] = translation self.T_0_t = np.matmul(self.T_0_t, T_t_1_t) global_translation = self.T_0_t[0, :3, 3] global_quaternion = tf.transformations.quaternion_from_matrix(self.T_0_t[0]) self.T_world_pc.transform.translation.x = global_translation[0] self.T_world_pc.transform.translation.y = global_translation[1] self.T_world_pc.transform.translation.z = global_translation[2] self.T_world_pc.transform.rotation.x = global_quaternion[0] self.T_world_pc.transform.rotation.y = global_quaternion[1] self.T_world_pc.transform.rotation.z = global_quaternion[2] self.T_world_pc.transform.rotation.w = global_quaternion[3] def integrate(self, header, quaternion, translation): # Update TF self.update_transformation(quaternion=quaternion, translation=translation) # Publish TF self.T_world_pc.header.stamp = header.stamp self.tf_broadcaster.sendTransform(self.T_world_pc) self.T_pc_odom.header.stamp = header.stamp self.tf_broadcaster.sendTransform(self.T_pc_odom)
4,075
37.45283
84
py
delora
delora-main/src/ros_utils/publish_point_cloud_and_normals.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import copy import numpy as np import geometry_msgs.msg import rospy import sensor_msgs.msg import sensor_msgs.point_cloud2 import std_msgs.msg import torch import visualization_msgs.msg import data.dataset class ROSPublisher: def __init__(self, config): # config self.publish_normals_bool = True # Dataset self.dataset = data.dataset.PreprocessedPointCloudDataset(config=config) # Ros ## Publishers self.scan_publisher = rospy.Publisher("/lidar/points", sensor_msgs.msg.PointCloud2, queue_size=10) self.normals_publisher = rospy.Publisher("/lidar/normals", visualization_msgs.msg.MarkerArray, queue_size=10) rospy.init_node('pc2_publisher', anonymous=True) self.rate = rospy.Rate(10) ## Messages ## Header self.header = std_msgs.msg.Header() self.header.stamp = rospy.Time.now() self.header.frame_id = "lidar" ### Scan self.scan_msg = sensor_msgs.msg.PointCloud2() self.scan_msg.header = self.header self.scan_msg.fields = [ sensor_msgs.msg.PointField('x', 0, sensor_msgs.msg.PointField.FLOAT32, 1), sensor_msgs.msg.PointField('y', 4, sensor_msgs.msg.PointField.FLOAT32, 1), sensor_msgs.msg.PointField('z', 8, sensor_msgs.msg.PointField.FLOAT32, 1), ] ### Normals self.normals_msg = visualization_msgs.msg.MarkerArray() self.normals_marker = visualization_msgs.msg.Marker() self.normals_marker.type = self.normals_marker.ARROW self.normals_marker.color.a = 1.0 self.normals_marker.color.r = 1.0 self.normals_marker.color.g = 0.0 self.normals_marker.color.b = 0.0 self.normals_marker.scale.x = 0.1 self.normals_marker.scale.y = 0.01 self.normals_marker.scale.z = 0.01 self.normals_marker.header = self.header self.arrow_location = geometry_msgs.msg.Point() self.arrow_orientation = geometry_msgs.msg.Quaternion() def euler_to_quaternion(self, roll, pitch, yaw): qx = np.sin(roll / 2) * np.cos(pitch / 2) * np.cos(yaw / 2) - np.cos(roll / 2) * np.sin( pitch / 2) * np.sin(yaw / 2) qy = np.cos(roll / 2) * np.sin(pitch / 2) * np.cos(yaw / 2) + np.sin(roll / 2) * np.cos( pitch / 2) * np.sin(yaw / 2) qz = np.cos(roll / 2) * np.cos(pitch / 2) * np.sin(yaw / 2) - np.sin(roll / 2) * np.sin( pitch / 2) * np.cos(yaw / 2) qw = np.cos(roll / 2) * np.cos(pitch / 2) * np.cos(yaw / 2) + np.sin(roll / 2) * np.sin( pitch / 2) * np.sin(yaw / 2) return np.concatenate((np.expand_dims(qx, axis=0), np.expand_dims(qy, axis=0), np.expand_dims(qz, axis=0), np.expand_dims(qw, axis=0)), axis=0) def publish_scan(self, scan): scan = scan[:, :, 0] self.scan_msg.data = scan self.scan_msg.header.stamp = rospy.Time.now() self.scan_publisher.publish( sensor_msgs.point_cloud2.create_cloud(self.header, self.scan_msg.fields, scan)) def publish_normals(self, scan, normals): self.normals_msg.markers = [] scan = scan[:, :, 0] normals = normals[0].transpose(1, 0) print("Normals shape: " + str(normals.shape)) for index, point in enumerate(scan): normal = normals[index] self.arrow_location.x = point[0] self.arrow_location.y = point[1] self.arrow_location.z = point[2] self.normals_marker.pose.position = self.arrow_location self.normals_marker.header.stamp = rospy.Time.now() pitch = -np.arcsin(normal[2]) yaw = np.arctan2(normal[1], normal[0]) quaternion = self.euler_to_quaternion(roll=0, pitch=np.asarray(pitch), yaw=np.asarray(yaw)) self.arrow_orientation.x = quaternion[0] self.arrow_orientation.y = quaternion[1] self.arrow_orientation.z = quaternion[2] self.arrow_orientation.w = quaternion[3] self.normals_marker.pose.orientation = self.arrow_orientation self.normals_marker.id = index self.normals_msg.markers.append(copy.deepcopy(self.normals_marker)) self.normals_publisher.publish(self.normals_msg) def publish_dataset(self): dataloader = torch.utils.data.DataLoader(self.dataset, shuffle=False, num_workers=0) for preprocessed_data in dataloader: print("Index: " + str(int(preprocessed_data["index"])) + " / " + str(len(dataloader))) scan = preprocessed_data["scan_1"].numpy()[0].transpose(2, 1, 0) self.publish_scan(scan=scan) if self.publish_normals_bool: normal_list_1 = preprocessed_data["normal_list_1"][0] normals_bool = ((normal_list_1[0, 0] != 0.0) | (normal_list_1[0, 1] != 0.0) | \ (normal_list_1[0, 2] != 0.0)).numpy() print("Scan shape: " + str(scan.shape)) self.publish_normals( scan=scan.transpose(2, 0, 1)[:, normals_bool].transpose(1, 2, 0), normals=normal_list_1[:, :, normals_bool].numpy()) if int(preprocessed_data["index"]) >= 100: break self.rate.sleep()
5,759
39.851064
99
py
delora
delora-main/src/ros_utils/convert_to_rosbag.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import os import rospy import rosbag import torch import sensor_msgs.msg import sensor_msgs.point_cloud2 import std_msgs.msg import data.kitti_scans class RosbagConverter(): def __init__(self, config): self.config = config self.rate = 10.0 # Ros rospy.init_node('Kitti_converter', anonymous=True) self.topic = "/velodyne_points" ## Header self.header = std_msgs.msg.Header() # self.header.stamp = rospy.Time.now() #rospy.Time(0.0) # rospy.Time.now() self.header.frame_id = "velodyne" ## Scan self.scan_msg = sensor_msgs.msg.PointCloud2() self.scan_msg.header = self.header self.scan_msg.fields = [ sensor_msgs.msg.PointField('x', 0, sensor_msgs.msg.PointField.FLOAT32, 1), sensor_msgs.msg.PointField('y', 4, sensor_msgs.msg.PointField.FLOAT32, 1), sensor_msgs.msg.PointField('z', 8, sensor_msgs.msg.PointField.FLOAT32, 1), sensor_msgs.msg.PointField('intensity', 12, sensor_msgs.msg.PointField.FLOAT32, 1), ] def ensure_dir(self, file_path): directory = os.path.dirname(file_path) if not os.path.exists(directory): os.makedirs(directory) def convert_sequence(self, dataloader, sequence, dataset_name): time = rospy.Time.now() duration = rospy.rostime.Duration(secs=0.1) self.ensure_dir(self.config[dataset_name]["rosbag_path"] + format(sequence, '02d')) outbag = self.config[dataset_name]["rosbag_path"] + format(sequence, '02d') + ".bag" with rosbag.Bag(outbag, 'w') as outbag: for index, point_cloud_1 in enumerate(dataloader): if not index % 10: print("Sequence " + str(sequence) + ", index: " + str(index) + " / " + str( len(dataloader)) + ", time: " + str(time.secs)) scan = point_cloud_1[0].permute(1, 0).numpy() self.scan_msg.header.stamp = time write_msg = sensor_msgs.point_cloud2.create_cloud(self.scan_msg.header, self.scan_msg.fields, scan) outbag.write(self.topic, write_msg, write_msg.header.stamp) time += duration def convert(self): for index_of_dataset, dataset_name in enumerate(self.config["datasets"]): for index_of_sequence, data_identifier in enumerate(self.config[dataset_name]["data_identifiers"]): # Do it for each sequence separately self.config["data_identifier"] = data_identifier # Choose which dataset if dataset_name == "kitti": dataset = data.kitti_scans.KITTIPointCloudDataset(base_dir=self.config[dataset_name]["data_path"], identifier=data_identifier, device=self.config["device"]) else: raise Exception("Currently only KITTI is supported") # Define dataloader dataloader = torch.utils.data.DataLoader(dataset, shuffle=False, num_workers=0, pin_memory=True if self.config[ "device"] == torch.device( "cuda") else False) print("Start sequence " + str(index_of_sequence) + " of dataset " + dataset_name + ".") self.convert_sequence(dataloader=dataloader, sequence=data_identifier, dataset_name=dataset_name)
4,014
43.120879
118
py
delora
delora-main/src/data/kitti_scans.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import os import glob import numpy as np import pykitti import torch class KITTIDatasetPreprocessor(): def __init__(self, config, dataset_name, preprocessing_fct): self.config = config self.identifier = self.config[dataset_name]["data_identifier"] self.point_cloud_dataset = KITTIPointCloudDataset(base_dir=self.config[dataset_name]["data_path"], identifier=self.identifier, device=self.config["device"]) # Preprocessing Function self.preprocessing_fct = preprocessing_fct def preprocess(self): for index in range(self.point_cloud_dataset.num_elements): if not index % 10: print("Preprocessing scan " + str(index) + "/" + str( self.point_cloud_dataset.num_elements) + " from sequence " + format(self.identifier, '02d') + ".") scan = self.point_cloud_dataset.get_velo_torch(index).unsqueeze(0) # Apply preprocessing self.preprocessing_fct(scan=scan, index=index) class KITTIPointCloudDataset(torch.utils.data.dataset.Dataset): def __init__(self, base_dir, identifier="00", device=torch.device("cuda")): super(KITTIPointCloudDataset, self).__init__() self.base_dir = base_dir self.identifier = identifier self.device = device self.velo_file_list = sorted(glob.glob( os.path.join(self.base_dir, format(self.identifier, '02d'), "velodyne", '*.bin'))) self.velo_data_generator = pykitti.utils.yield_velo_scans(self.velo_file_list) self.num_elements = len(self.velo_file_list) def get_velo(self, idx): return pykitti.utils.load_velo_scan(self.velo_file_list[idx]) def get_velo_torch(self, idx): return torch.from_numpy(self.get_velo(idx)).to(torch.device("cpu")).transpose(0, 1) def __getitem__(self, index): return self.get_velo_torch(idx=index) def __len__(self): return self.num_elements
2,322
39.754386
118
py
delora
delora-main/src/data/dataset.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import csv import glob import os import torch import numpy as np # Preprocessed point cloud dataset is invariant of source of data --> always same format # Data structure: dataset --> sequence --> scan # e.g. dataset=anymal, sequence=00, scan=000000 # Naming: e.g. num_scans_sequences_datasets means number of scans in the corresponding # sequences in the corresponding datasets class PreprocessedPointCloudDataset(torch.utils.data.dataset.Dataset): def __init__(self, config): super(PreprocessedPointCloudDataset, self).__init__() self.config = config # Store dataset in RAM self.store_dataset_in_RAM = self.config["store_dataset_in_RAM"] # Members self.normals_files_in_datasets = [] self.scans_files_in_datasets = [] num_sequences_in_datasets = np.zeros(len(self.config["datasets"])) num_scans_in_sequences_in_datasets = [] num_scans_in_datasets = np.zeros(len(self.config["datasets"])) # Go through dataset(s) and create file lists for index_of_dataset, dataset in enumerate(self.config["datasets"]): ## Paths of sequences normals_files_in_sequences = [] scans_files_in_sequences = [] num_scans_in_sequences = np.zeros(len(self.config[dataset]["data_identifiers"]), dtype=int) ## Go through sequences for index_of_sequence, data_identifier in enumerate(self.config[dataset]["data_identifiers"]): if not os.path.exists( os.path.join(self.config[dataset]["preprocessed_path"], format(data_identifier, '02d') + "/")): raise Exception( "The specified path and dataset " + os.path.join(self.config[dataset]["preprocessed_path"], format(data_identifier, '02d') + "/") + "does not exist.") name = os.path.join(self.config[dataset]["preprocessed_path"], format(data_identifier, '02d') + "/") normals_name = os.path.join(name, "normals/") scans_name = os.path.join(name, "scans/") # Files normals_files_in_sequences.append(sorted(glob.glob(os.path.join(normals_name, '*.npy')))) scans_files_in_sequences.append(sorted(glob.glob(os.path.join(scans_name, '*.npy')))) # -1 is important (because looking at consecutive scans at t and t+1) num_scans_in_sequences[index_of_sequence] = len(normals_files_in_sequences[index_of_sequence]) - 1 num_sequences_in_datasets[index_of_dataset] = len(self.config[dataset]["data_identifiers"]) num_scans_in_sequences_in_datasets.append(num_scans_in_sequences) num_scans_in_datasets[index_of_dataset] = np.sum(num_scans_in_sequences, dtype=int) self.normals_files_in_datasets.append(normals_files_in_sequences) self.scans_files_in_datasets.append(scans_files_in_sequences) self.num_scans_overall = np.sum(num_scans_in_datasets, dtype=int) # Dataset, sequence, scan indices (mapping overall_index --> dataset, sequence, scan-indices) self.indices_dataset = np.zeros(self.num_scans_overall, dtype=int) self.indices_sequence = np.zeros(self.num_scans_overall, dtype=int) self.indices_scan = np.zeros(self.num_scans_overall, dtype=int) overall_index = 0 for index_dataset, num_scans_in_sequences in enumerate(num_scans_in_sequences_in_datasets): for index_sequence, num_scans in enumerate(num_scans_in_sequences): for index_scan in range(num_scans): self.indices_dataset[overall_index] = index_dataset self.indices_sequence[overall_index] = index_sequence self.indices_scan[overall_index] = index_scan overall_index += 1 # In case of RAM loading --> keep data in memory if not self.store_dataset_in_RAM: print('\033[92m' + "Dataset will be kept on disk. For higher performance enable RAM loading." + "\033[0;0m") else: print('\033[92m' + "Loading all scans and normals into the RAM / SWAP... " + "Disable this if you do not have enough RAM." + "\033[0;0m") self.normals_RAM = [] self.scans_RAM = [] counter = 0 for index_dataset, num_scans_in_sequences in enumerate(num_scans_in_sequences_in_datasets): normals_in_dataset_RAM = [] scans_in_dataset_RAM = [] for index_sequence, num_scans in enumerate(num_scans_in_sequences): normals_in_sequence_RAM = [] scans_in_sequence_RAM = [] # + 1 is important here, since we need to store ALL scans (also last of each sequence) for index_scan in range(num_scans + 1): (normal_list, scan) = self.load_files_from_disk(index_dataset=index_dataset, index_sequence=index_sequence, index_scan=index_scan) normals_in_sequence_RAM.append(normal_list) scans_in_sequence_RAM.append(scan) counter += 1 normals_in_dataset_RAM.append(normals_in_sequence_RAM) scans_in_dataset_RAM.append(scans_in_sequence_RAM) self.normals_RAM.append(normals_in_dataset_RAM) self.scans_RAM.append(scans_in_dataset_RAM) print('\033[92m' + "Loaded " + str(counter) + " scans to RAM/swap." + "\033[0;0m") def load_files_from_disk(self, index_dataset, index_sequence, index_scan): normal_list = torch.from_numpy( np.load(self.normals_files_in_datasets[index_dataset][index_sequence][index_scan])).to( torch.device("cpu")).permute(1, 0).view(1, 3, -1) scan = torch.from_numpy( np.load(self.scans_files_in_datasets[index_dataset][index_sequence][index_scan])).to( torch.device("cpu")).permute(1, 0).view(1, 3, -1) return (normal_list, scan) def __getitem__(self, index): index_dataset = self.indices_dataset[index] index_sequence = self.indices_sequence[index] index_scan = self.indices_scan[index] if self.store_dataset_in_RAM: normal_list_1 = self.normals_RAM[index_dataset][index_sequence][index_scan] scan_1 = self.scans_RAM[index_dataset][index_sequence][index_scan] normal_list_2 = self.normals_RAM[index_dataset][index_sequence][index_scan + 1] scan_2 = self.scans_RAM[index_dataset][index_sequence][index_scan + 1] else: (normal_list_1, scan_1) = self.load_files_from_disk(index_dataset=index_dataset, index_sequence=index_sequence, index_scan=index_scan) (normal_list_2, scan_2) = self.load_files_from_disk(index_dataset=index_dataset, index_sequence=index_sequence, index_scan=index_scan + 1) # Encapsulate data preprocessed_data = { "index": index, "index_dataset": index_dataset, "index_sequence": index_sequence, "index_scan": index_scan, "dataset": self.config["datasets"][index_dataset], "normal_list_1": normal_list_1, "normal_list_2": normal_list_2, "scan_1": scan_1, "scan_2": scan_2, } return preprocessed_data def __len__(self): return self.num_scans_overall # Groundtruth poses are also always saved in this format class PoseDataset(torch.utils.data.dataset.Dataset): def __init__(self, config): self.config = config self.device = config["device"] # Members self.poses_datasets = [] num_sequences_datasets = np.zeros(len(self.config["datasets"])) num_scans_sequences_datasets = [] num_scans_datasets = np.zeros(len(self.config["datasets"])) # Go through dataset(s) for index_of_dataset, dataset in enumerate(self.config["datasets"]): base_dir = config[dataset]["pose_data_path"] poses_sequences = [] num_scans_sequences = np.zeros(len(self.config[dataset]["data_identifiers"]), dtype=int) for index_of_sequence, data_identifier in enumerate( config[dataset]["data_identifiers"]): if base_dir: pose_file_name = os.path.join(base_dir, format(data_identifier, '02d') + '.txt') with open(pose_file_name, newline="") as csvfile: row_reader = csv.reader(csvfile, delimiter=" ") num_poses = sum(1 for row in row_reader) poses_sequences.append(np.zeros((num_poses, 12))) with open(pose_file_name, newline="") as csvfile: row_reader = csv.reader(csvfile, delimiter=" ") for index_of_scan, row in enumerate(row_reader): poses_sequences[index_of_sequence][index_of_scan, :] = np.asarray( [float(element) for element in row]) # -1 is important (because looking at consecutive scans at t and t+1) num_scans_sequences[index_of_sequence] = num_poses - 1 else: print("Groundtruth file does not exist. Not using any ground truth for it.") name = os.path.join(self.config[dataset]["preprocessed_path"], format(data_identifier, '02d') + "/") normals_name = os.path.join(name, "normals/") normals_files_in_sequence = sorted(glob.glob(os.path.join(normals_name, '*.npy'))) num_poses = len(normals_files_in_sequence) poses_sequences.append(np.zeros((num_poses, 1))) for index_of_scan in range(num_poses): poses_sequences[index_of_sequence][index_of_scan, 0] = None num_sequences_datasets[index_of_dataset] = len( self.config[dataset]["data_identifiers"]) num_scans_sequences_datasets.append(num_scans_sequences) num_scans_datasets[index_of_dataset] = np.sum(num_scans_sequences, dtype=int) self.poses_datasets.append(poses_sequences) self.num_scans_overall = np.sum(num_scans_datasets, dtype=int) # Dataset, sequence, scan indices (mapping overall_index --> dataset, sequence, scan-indices) self.indices_dataset = np.zeros(self.num_scans_overall, dtype=int) self.indices_sequence = np.zeros(self.num_scans_overall, dtype=int) self.indices_scan = np.zeros(self.num_scans_overall, dtype=int) counter = 0 for index_dataset, num_scans_sequences in enumerate(num_scans_sequences_datasets): for index_sequence, num_scans in enumerate(num_scans_sequences): for index_scan in range(num_scans): self.indices_dataset[counter] = index_dataset self.indices_sequence[counter] = index_sequence self.indices_scan[counter] = index_scan counter += 1 def return_translations(self, index_of_dataset, index_of_sequence): print(self.poses_datasets[index_of_dataset][index_of_sequence][0, 0]) if not np.isnan(self.poses_datasets[index_of_dataset][index_of_sequence][0, 0]): return self.poses_datasets[index_of_dataset][index_of_sequence][:, [3, 7, 11]] else: return None def return_poses(self, index_of_dataset, index_of_sequence): if not np.isnan(self.poses_datasets[index_of_dataset][index_of_sequence][0, 0]): return self.poses_datasets[index_of_dataset][index_of_sequence] else: return None def __getitem__(self, index): index_dataset = self.indices_of_dataset[index] index_sequence = self.indices_sequence[index] index_scan = self.indices_scan[index] if not np.isnan(self.poses_datasets[index_dataset][index_sequence][index_scan, 0]): return self.poses_datasets[index_dataset][index_sequence][index_scan] else: return None def __len__(self): return self.num_scans_overall
13,227
51.701195
120
py
delora
delora-main/src/data/rosbag_scans.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import os import glob import numpy as np import torch import ros_utils.rosbag_pcl_extractor class RosbagDatasetPreprocessor(): def __init__(self, config, dataset_name, topic, preprocessing_fct): self.config = config self.rosbag_dir = self.config[dataset_name]["data_path"] self.identifier = self.config[dataset_name]["data_identifier"] # Look at file self.rosbag_file = sorted(glob.glob( os.path.join(self.rosbag_dir, "" + format(self.identifier, '02d') + '*'))) if len(self.rosbag_file) > 1: raise Exception( "Identifier does not uniquely define a rosbag. There are multiple files containing " + format(self.identifier, '02d') + ".") elif len(self.rosbag_file) == 0: raise Exception( "Rosbag corresponding to data identifier " + str(self.identifier) + " must include " + format(self.identifier, '02d') + ".") self.rosbag_file = self.rosbag_file[0] # Use rosbag tool self.rosbag_extractor = ros_utils.rosbag_pcl_extractor.RosbagToPCLExtractor( rosbag_file=self.rosbag_file, topic=topic, config=self.config, preprocessing_fct=preprocessing_fct) def preprocess(self): self.rosbag_extractor.preprocess_rosbag()
1,574
37.414634
111
py
delora
delora-main/src/losses/icp_losses.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import scipy.spatial import torch class ICPLosses(torch.nn.Module): def __init__(self, config): super(ICPLosses, self).__init__() self.config = config # Define single loss components if self.config["point_to_point_loss"]: self.point_to_point_loss = KDPointToPointLoss(config=self.config) if self.config["point_to_plane_loss"]: self.point_to_plane_loss = KDPointToPlaneLoss(config=self.config) if self.config["plane_to_plane_loss"]: self.plane_to_plane_loss = KDPlaneToPlaneLoss(config=self.config) def find_target_correspondences(self, kd_tree_target, source_list_numpy): target_correspondence_indices = kd_tree_target[0].query(source_list_numpy[0])[1] return target_correspondence_indices def forward(self, source_point_cloud_transformed, source_normal_list_transformed, target_point_cloud, target_normal_list, compute_pointwise_loss_bool): # Build kd-tree target_kd_tree = [scipy.spatial.cKDTree(target_point_cloud[0].permute(1, 0).cpu())] if self.config["po2po_alone"]: # Find corresponding target points for all source points target_correspondences_of_source_points = \ torch.from_numpy(self.find_target_correspondences( kd_tree_target=target_kd_tree, source_list_numpy=source_point_cloud_transformed.permute(0, 2, 1).detach().cpu().numpy())).to( self.config["device"]) source_points_where_no_normals = source_point_cloud_transformed target_points_where_no_normals = target_point_cloud[:, :, target_correspondences_of_source_points] else: source_where_normals_bool = (source_normal_list_transformed[:, 0, :] != 0) | ( source_normal_list_transformed[:, 1, :] != 0) | ( source_normal_list_transformed[:, 2, :] != 0) target_where_normals_bool = (target_normal_list[:, 0, :] != 0) | ( target_normal_list[:, 1, :] != 0) | (target_normal_list[:, 2, :] != 0) # Differentiate between source points where normals exist / do not exist source_point_cloud_transformed_where_normals = source_point_cloud_transformed[:, :, source_where_normals_bool[0]] source_normal_list_transformed_where_normals = source_normal_list_transformed[:, :, source_where_normals_bool[0]] source_point_cloud_transformed_where_no_normals = source_point_cloud_transformed[:, :, ~source_where_normals_bool[0]] # Transform to numpy for KD-tree handling source_point_cloud_transformed_where_normals_numpy = \ source_point_cloud_transformed_where_normals.permute(0, 2, 1).detach().cpu().numpy() source_point_cloud_transformed_where_no_normals_numpy = \ source_point_cloud_transformed_where_no_normals.permute(0, 2, 1).detach().cpu().numpy() # Find corresponding target points for source points which have normals, target_correspondences_of_source_points_where_source_normals_indices = \ torch.from_numpy(self.find_target_correspondences( kd_tree_target=target_kd_tree, source_list_numpy=source_point_cloud_transformed_where_normals_numpy)).to( self.config["device"]) # Find corresponding target points for source points which have no normals target_correspondences_of_source_points_where_no_source_normals_indices = \ torch.from_numpy(self.find_target_correspondences( kd_tree_target=target_kd_tree, source_list_numpy=source_point_cloud_transformed_where_no_normals_numpy)).to( self.config["device"]) # 2 cases we compute losses for: if self.config["point_to_point_loss"]: # 1) NO source normal and NO target normal, ## For EACH source point without normal, there is a corresponding target point target_point_cloud_where_no_source_normals = \ target_point_cloud[:, :, target_correspondences_of_source_points_where_no_source_normals_indices] ## We only need to keep the source target pairs when there also exists NO target normal target_where_normals_where_no_source_normals_bool = \ target_where_normals_bool[:, target_correspondences_of_source_points_where_no_source_normals_indices] target_points_where_no_normals = target_point_cloud_where_no_source_normals[:, :, ~target_where_normals_where_no_source_normals_bool[ 0]] source_points_where_no_normals = source_point_cloud_transformed_where_no_normals[:, :, ~target_where_normals_where_no_source_normals_bool[ 0]] ## 2) We need corresponding target normals where source has normals target_normal_list_where_source_normals_with_holes = \ target_normal_list[:, :, target_correspondences_of_source_points_where_source_normals_indices] target_point_cloud_where_source_normals = \ target_point_cloud[:, :, target_correspondences_of_source_points_where_source_normals_indices] # These 4 arrays are now the ones / corresponding ones of source points that have normals # --> Now need to remove points where we have no target normal target_where_normals_where_source_normals_bool = \ target_where_normals_bool[:, target_correspondences_of_source_points_where_source_normals_indices] source_points_where_normals = source_point_cloud_transformed_where_normals[:, :, target_where_normals_where_source_normals_bool[0]] source_normals_where_normals = source_normal_list_transformed_where_normals[:, :, target_where_normals_where_source_normals_bool[0]] target_points_where_normals = target_point_cloud_where_source_normals[:, :, target_where_normals_where_source_normals_bool[0]] target_normals_where_normals = target_normal_list_where_source_normals_with_holes[:, :, target_where_normals_where_source_normals_bool[0]] # Define losses loss_po2po = torch.zeros(1, device=self.config["device"]) loss_po2pl = torch.zeros(1, device=self.config["device"]) loss_pl2pl = torch.zeros(1, device=self.config["device"]) po2pl_pointwise_loss = torch.zeros(1, device=self.config["device"]) if self.config["point_to_point_loss"]: loss_po2po, po2po_pointwise_loss, po2po_source_list = self.point_to_point_loss.forward( source_list=source_points_where_no_normals, target_correspondences_list=target_points_where_no_normals, compute_pointwise_loss_bool=False) if self.config["point_to_plane_loss"]: loss_po2pl, po2pl_pointwise_loss = \ self.point_to_plane_loss( source_list=source_points_where_normals, target_correspondences_list=target_points_where_normals, target_correspondences_normal_vectors=target_normals_where_normals, compute_pointwise_loss_bool=compute_pointwise_loss_bool) if self.config["plane_to_plane_loss"]: loss_pl2pl = self.plane_to_plane_loss( source_normals=source_normals_where_normals, target_correspondences_normals=target_normals_where_normals) losses = { "loss_po2po": loss_po2po, "loss_po2pl": loss_po2pl, "loss_po2pl_pointwise": po2pl_pointwise_loss, "loss_pl2pl": loss_pl2pl, } plotting = { "scan_2_transformed": source_points_where_normals, "normals_2_transformed": source_normals_where_normals } if not self.config["po2po_alone"] else None return losses, plotting class KDPointToPointLoss: def __init__(self, config): self.config = config self.lossMeanMSE = torch.nn.MSELoss() self.lossPointMSE = torch.nn.MSELoss(reduction="none") def compute_loss(self, source_list, target_correspondences_list, compute_pointwise_loss_bool): loss = self.lossMeanMSE(source_list, target_correspondences_list) if compute_pointwise_loss_bool: source_pointwise_loss = self.lossPointMSE(source_list, target_correspondences_list).detach() else: source_pointwise_loss = None return loss, \ source_pointwise_loss.transpose(0, 1).view(1, 3, -1) if compute_pointwise_loss_bool else None, \ source_list.transpose(0, 1).view(1, 3, -1) def forward(self, source_list, target_correspondences_list, compute_pointwise_loss_bool): return self.compute_loss(source_list=source_list, target_correspondences_list=target_correspondences_list, compute_pointwise_loss_bool=compute_pointwise_loss_bool) class KDPointToPlaneLoss(torch.nn.Module): def __init__(self, config): super(KDPointToPlaneLoss, self).__init__() self.config = config self.lossMeanMSE = torch.nn.MSELoss() self.lossPointMSE = torch.nn.MSELoss(reduction="none") def compute_loss(self, source_list, target_correspondences_list, normal_vectors, compute_pointwise_loss): source_pointwise_distance_vector = (source_list - target_correspondences_list) source_pointwise_normal_distance = source_pointwise_distance_vector.permute(2, 0, 1).matmul( normal_vectors.permute(2, 1, 0)) loss = self.lossMeanMSE(source_pointwise_normal_distance, torch.zeros(source_pointwise_normal_distance.shape, device=self.config["device"])) return loss, \ source_pointwise_distance_vector if compute_pointwise_loss else None def forward(self, source_list, target_correspondences_list, target_correspondences_normal_vectors, compute_pointwise_loss_bool): return self.compute_loss(source_list=source_list, target_correspondences_list=target_correspondences_list, normal_vectors=target_correspondences_normal_vectors, compute_pointwise_loss=compute_pointwise_loss_bool) class KDPlaneToPlaneLoss(torch.nn.Module): def __init__(self, config): super(KDPlaneToPlaneLoss, self).__init__() self.config = config self.lossMeanMSE = torch.nn.MSELoss() # self.lossPointMSE = torch.nn.MSELoss(reduction="none") def forward(self, source_normals, target_correspondences_normals): source_normals = source_normals.permute(2, 0, 1) if self.config["normal_loss"] == "linear": target_correspondences_normals = target_correspondences_normals.permute(2, 1, 0) dot_products = torch.matmul(source_normals, target_correspondences_normals) dot_products = dot_products return self.lossMeanMSE(1 - dot_products, torch.zeros(dot_products.shape, device=self.config["device"])) elif self.config["normal_loss"] == "squared": target_correspondences_normals = target_correspondences_normals.permute(2, 0, 1) distance = torch.norm(source_normals - target_correspondences_normals, dim=2, keepdim=True) weighted_distance = distance return self.lossMeanMSE(weighted_distance, torch.zeros(weighted_distance.shape, device=self.config["device"])) else: raise Exception("The normal loss which is defined here is not admissible.")
13,291
54.153527
109
py
delora
delora-main/src/preprocessing/normal_computation.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import numpy as np import torch import utility.linalg # Need an instance of this for each dataset class NormalsComputer: def __init__(self, config, dataset_name): self.config = config # Preprocessing (arrays should not be build at each iteration) coordinate_meshgrid = np.meshgrid(range(self.config[dataset_name]["horizontal_cells"]), range(self.config[dataset_name]["vertical_cells"])) self.u_image_coords_list = torch.from_numpy(coordinate_meshgrid[0]).view( self.config[dataset_name]["vertical_cells"] * self.config[dataset_name][ "horizontal_cells"]).to(self.config["device"]) self.v_image_coords_list = torch.from_numpy(coordinate_meshgrid[1]).view( self.config[dataset_name]["vertical_cells"] * self.config[dataset_name][ "horizontal_cells"]).to(self.config["device"]) # Store dataset info self.dataset_name = dataset_name def get_image_coords(self, image): # Target list is extracted from target image (visible points) list = image[0, :3, ].view( 3, self.config[self.dataset_name]["vertical_cells"] * self.config[self.dataset_name][ "horizontal_cells"]).transpose(0, 1) indices_non_zero = (list[:, 0] != 0) & (list[:, 1] != 0) & (list[:, 2] != 0) list = list[indices_non_zero] u_image_coordinates_list = self.u_image_coords_list[indices_non_zero] v_image_coordinates_list = self.v_image_coords_list[indices_non_zero] return list, u_image_coordinates_list, v_image_coordinates_list def check_planarity(self, eigenvalues): epsilon_plane = self.config["epsilon_plane"] epsilon_line = self.config["epsilon_line"] # First (plane-) criterion also holds for line, therefor also check that line criterion # (only one large e-value) does NOT hold planar_indices = ((eigenvalues[:, 0] / torch.sum(eigenvalues, dim=1) < epsilon_plane) & ( (eigenvalues[:, 0] + eigenvalues[:, 1]) / torch.sum(eigenvalues, dim=1) > epsilon_line)) return planar_indices def covariance_eigen_decomposition(self, point_neighbors, point_locations): # Filter by range epsilon_range = self.config["epsilon_range"] range_deviation_indices = torch.abs(torch.norm(point_neighbors, dim=1) - torch.norm(point_locations, dim=1)) > epsilon_range # Annulate neighbors that have a range which deviates too much from point location point_neighbors.permute(0, 2, 1)[range_deviation_indices] = 0.0 point_neighbors_permuted = point_neighbors.permute(2, 1, 0) covariance_matrices, number_neighbours = utility.linalg.cov( point_neighbors=point_neighbors_permuted) # Filter by amount of neighbors # Dropping eigenvectors and values where min number of points is not met where_enough_neighbors_bool = number_neighbours >= self.config[ "min_num_points_in_neighborhood_to_determine_point_class"] covariance_matrices = covariance_matrices[where_enough_neighbors_bool] eigenvalues, eigenvectors = torch.symeig(covariance_matrices.to(torch.device("cpu")), eigenvectors=True) eigenvalues = eigenvalues.to(self.config["device"]) eigenvectors = eigenvectors.to(self.config["device"]) # e-vectors corresponding to smallest e-value all_normal_vectors = eigenvectors[:, :, 0] point_locations = point_locations[0].permute(1, 0) dot_products = all_normal_vectors.view(-1, 1, 3).matmul( point_locations[where_enough_neighbors_bool].view(-1, 3, 1)).squeeze() all_normal_vectors[dot_products > 0] *= -1 # Now for all points which do not have a normal we simply store (0, 0, 0)^T all_normal_vectors_with_zeros = torch.zeros_like(point_locations) all_normal_vectors_with_zeros[where_enough_neighbors_bool] = all_normal_vectors return all_normal_vectors_with_zeros, where_enough_neighbors_bool, point_locations def compute_normal_vectors(self, image): # Get coordinates and target points point_list, u_coordinates, v_coordinates = self.get_image_coords(image=image) image = image[0, :3] point_neighbors = [] # Take patch around each point a = int(self.config[self.dataset_name]["neighborhood_side_length"][0] / 2) b = int(self.config[self.dataset_name]["neighborhood_side_length"][1] / 2) for v_neighbor in range(-a, a + 1): for u_neighbor in range(-b, b + 1): v_neighbor_coords = v_coordinates + v_neighbor u_neighbor_coords = u_coordinates + u_neighbor # These can be negative, set them to 0 --> will be biased neighborhood at edges v_neighbor_coords[v_neighbor_coords < 0] = 0 v_neighbor_coords[ v_neighbor_coords > (self.config[self.dataset_name]["vertical_cells"] - 1)] = ( self.config[self.dataset_name]["vertical_cells"] - 1) u_neighbor_coords[u_neighbor_coords < 0] = 0 u_neighbor_coords[ u_neighbor_coords > (self.config[self.dataset_name]["horizontal_cells"] - 1)] = ( self.config[self.dataset_name]["horizontal_cells"] - 1) if not len(point_neighbors): point_neighbors = image[:, v_neighbor_coords, u_neighbor_coords].view(1, 3, -1) else: point_neighbors = torch.cat((point_neighbors, image[:, v_neighbor_coords, u_neighbor_coords ].view(1, 3, -1)), dim=0) del a, b point_locations = image[:, v_coordinates, u_coordinates].view(1, 3, -1) return self.covariance_eigen_decomposition(point_neighbors=point_neighbors, point_locations=point_locations)
6,487
51.747967
101
py
delora
delora-main/bin/run_training.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import click import numpy as np import torch import yaml import deploy.trainer @click.command() @click.option('--training_run_name', prompt='MLFlow name of the run', help='The name under which the run can be found afterwards.') @click.option('--experiment_name', help='High-level training sequence name for clustering in MLFlow.', default="") @click.option('--checkpoint', help='Path to the saved checkpoint. Leave empty if none.', default="") def config(training_run_name, experiment_name, checkpoint): f = open('config/config_datasets.yaml') config = yaml.load(f, Loader=yaml.FullLoader) f = open('config/deployment_options.yaml') deployment_options = yaml.load(f, Loader=yaml.FullLoader) config.update(deployment_options) f = open('config/hyperparameters.yaml') network_hyperparameters = yaml.load(f, Loader=yaml.FullLoader) config.update(network_hyperparameters) # Default: load parameters from yaml parameters_exist = False # CLI Input ## Checkpoint for continuing training if checkpoint: ### Parameters from previous run? if 'parameters' in torch.load(checkpoint): print("\033[92m" + "Found parameters in checkpoint of previous run! Setting part of parameters to those ones." + "\033[0;0m") parameters_exist = True else: print("Checkpoint does not contain any parameters. Using those ones specified in the YAML files.") # Parameters that are set depending on whether provided in checkpoint if parameters_exist: loaded_config = torch.load(checkpoint)['parameters'] ## Device to be used loaded_config["device"] = torch.device(config["device"]) loaded_config["datasets"] = config["datasets"] for dataset in loaded_config["datasets"]: loaded_config[dataset]["training_identifiers"] = config[dataset]["training_identifiers"] loaded_config[dataset]["data_identifiers"] = loaded_config[dataset]["training_identifiers"] config = loaded_config # Some parameters are only initialized when not taken from checkpoint else: ## Device to be used config["device"] = torch.device(config["device"]) for dataset in config["datasets"]: config[dataset]["data_identifiers"] = config[dataset]["training_identifiers"] ## Convert angles to radians for dataset in config["datasets"]: config[dataset]["vertical_field_of_view"][0] *= (np.pi / 180.0) config[dataset]["vertical_field_of_view"][1] *= (np.pi / 180.0) config["horizontal_field_of_view"][0] *= (np.pi / 180.0) config["horizontal_field_of_view"][1] *= (np.pi / 180.0) # Parameters that are always set if checkpoint: config["checkpoint"] = str(checkpoint) else: config["checkpoint"] = None ## Trainings run name --> mandatory config["training_run_name"] = str(training_run_name) config["run_name"] = config["training_run_name"] ## Experiment name, default specified in deployment_options.yaml if experiment_name: config["experiment"] = experiment_name ## Mode config["mode"] = "training" print("----------------------------------") print("Configuration for this run: ") print(config) print("----------------------------------") return config if __name__ == "__main__": config = config(standalone_mode=False) trainer = deploy.trainer.Trainer(config=config) trainer.train()
3,837
39.4
110
py
delora
delora-main/bin/run_rosnode.py
#!/usr/bin/env python3 # Copyright 2021 by Julian Nubert, Robotic Systems Lab, ETH Zurich. # All rights reserved. # This file is released under the "BSD-3-Clause License". # Please see the LICENSE file that has been included as part of this package. import click import numpy as np import torch import yaml import ros_utils.odometry_publisher @click.command() @click.option('--checkpoint', prompt='Path to the saved model you want to test') @click.option('--dataset', prompt='On which dataset configuration do you want to get predictions? [kitti, darpa, ....]. Does not ' 'need to be one of those, but the sensor paramaters are looked up in the config_datasets.yaml.') @click.option('--lidar_topic', prompt='Topic of the published LiDAR pointcloud2 messages.') @click.option('--lidar_frame', prompt='LiDAR frame in TF tree.') @click.option('--integrate_odometry', help='Whether the published odometry should be integrated in the TF tree.', default=True) def config(checkpoint, dataset, lidar_topic, lidar_frame, integrate_odometry): f = open('config/config_datasets.yaml') config = yaml.load(f, Loader=yaml.FullLoader) f = open('config/deployment_options.yaml') deployment_options = yaml.load(f, Loader=yaml.FullLoader) config.update(deployment_options) f = open('config/hyperparameters.yaml') network_hyperparameters = yaml.load(f, Loader=yaml.FullLoader) config.update(network_hyperparameters) # Mode config["mode"] = "training" # No dropout during testing if config["use_dropout"]: config["use_dropout"] = False print("Deactivating dropout for this mode.") # CLI Input ## Checkpoint config["checkpoint"] = str(checkpoint) ## Dataset config["datasets"] = [str(dataset)] ## LiDAR Topic config["lidar_topic"] = str(lidar_topic) ## LiDAR Frame config["lidar_frame"] = str(lidar_frame) ## Integrate odometry config["integrate_odometry"] = integrate_odometry # Device to be used if config["device"] == "cuda": config["device"] = torch.device("cuda") else: config["device"] = torch.device("cpu") # Convert angles to radians for dataset in config["datasets"]: config[dataset]["vertical_field_of_view"][0] *= (np.pi / 180.0) config[dataset]["vertical_field_of_view"][1] *= (np.pi / 180.0) config["horizontal_field_of_view"][0] *= (np.pi / 180.0) config["horizontal_field_of_view"][1] *= (np.pi / 180.0) print("----------------------------------") print("Configuration for this run: ") print(config) print("----------------------------------") return config if __name__ == "__main__": config = config(standalone_mode=False) publisher = ros_utils.odometry_publisher.OdometryPublisher(config=config) publisher.publish_odometry()
2,881
35.948718
117
py