Prefix python modules

eval_copy_detection.py → dino/eval_copy_detection.py

@@ -25,9 +25,9 @@
 from PIL import Image, ImageFile
 import numpy as np
 
-import utils
-import vision_transformer as vits
-from eval_knn import extract_features
+import dino.utils
+import dino.vision_transformer as vits
+from dino.eval_knn import extract_features
 
 
 class CopydaysDataset():
@@ -161,7 +161,7 @@ def extract_features(image_list, model, args):
         num_workers=args.num_workers, drop_last=False,
         sampler=torch.utils.data.DistributedSampler(tempdataset, shuffle=False))
     features = None
-    for samples, index in utils.MetricLogger(delimiter="  ").log_every(data_loader, 10):
+    for samples, index in dino.utils.MetricLogger(delimiter="  ").log_every(data_loader, 10):
         samples, index = samples.cuda(non_blocking=True), index.cuda(non_blocking=True)
         feats = model.get_intermediate_layers(samples, n=1)[0].clone()
 
@@ -215,7 +215,7 @@ def extract_features(image_list, model, args):
     parser.add_argument('--imsize', default=320, type=int, help='Image size (square image)')
     parser.add_argument('--batch_size_per_gpu', default=16, type=int, help='Per-GPU batch-size')
     parser.add_argument('--pretrained_weights', default='', type=str, help="Path to pretrained weights to evaluate.")
-    parser.add_argument('--use_cuda', default=True, type=utils.bool_flag)
+    parser.add_argument('--use_cuda', default=True, type=dino.utils.bool_flag)
     parser.add_argument('--arch', default='vit_base', type=str, help='Architecture')
     parser.add_argument('--patch_size', default=8, type=int, help='Patch resolution of the model.')
     parser.add_argument("--checkpoint_key", default="teacher", type=str,
@@ -226,8 +226,8 @@ def extract_features(image_list, model, args):
     parser.add_argument("--local_rank", default=0, type=int, help="Please ignore and do not set this argument.")
     args = parser.parse_args()
 
-    utils.init_distributed_mode(args)
-    print("git:\n  {}\n".format(utils.get_sha()))
+    dino.utils.init_distributed_mode(args)
+    print("git:\n  {}\n".format(dino.utils.get_sha()))
     print("\n".join("%s: %s" % (k, str(v)) for k, v in sorted(dict(vars(args)).items())))
     cudnn.benchmark = True
 
@@ -241,7 +241,7 @@ def extract_features(image_list, model, args):
     if args.use_cuda:
         model.cuda()
     model.eval()
-    utils.load_pretrained_weights(model, args.pretrained_weights, args.checkpoint_key, args.arch, args.patch_size)
+    dino.utils.load_pretrained_weights(model, args.pretrained_weights, args.checkpoint_key, args.arch, args.patch_size)
 
     dataset = CopydaysDataset(args.data_path)
 
@@ -250,7 +250,7 @@ def extract_features(image_list, model, args):
     queries = []
     for q in dataset.query_blocks:
         queries.append(extract_features(dataset.get_block(q), model, args))
-    if utils.get_rank() == 0:
+    if dino.utils.get_rank() == 0:
         queries = torch.cat(queries)
         print(f"Extraction of queries features done. Shape: {queries.shape}")
 
@@ -264,7 +264,7 @@ def extract_features(image_list, model, args):
         print("Using distractors...")
         list_distractors = [os.path.join(args.distractors_path, s) for s in os.listdir(args.distractors_path) if is_image_file(s)]
         database.append(extract_features(list_distractors, model, args))
-    if utils.get_rank() == 0:
+    if dino.utils.get_rank() == 0:
         database = torch.cat(database)
         print(f"Extraction of database and distractors features done. Shape: {database.shape}")
 
@@ -273,20 +273,20 @@ def extract_features(image_list, model, args):
         print(f"Extracting features on images from {args.whitening_path} for learning the whitening operator.")
         list_whit = [os.path.join(args.whitening_path, s) for s in os.listdir(args.whitening_path) if is_image_file(s)]
         features_for_whitening = extract_features(list_whit, model, args)
-        if utils.get_rank() == 0:
+        if dino.utils.get_rank() == 0:
             # center
             mean_feature = torch.mean(features_for_whitening, dim=0)
             database -= mean_feature
             queries -= mean_feature
-            pca = utils.PCA(dim=database.shape[-1], whit=0.5)
+            pca = dino.utils.PCA(dim=database.shape[-1], whit=0.5)
             # compute covariance
             cov = torch.mm(features_for_whitening.T, features_for_whitening) / features_for_whitening.shape[0]
             pca.train_pca(cov.cpu().numpy())
             database = pca.apply(database)
             queries = pca.apply(queries)
 
     # ============ Copy detection ... ============
-    if utils.get_rank() == 0:
+    if dino.utils.get_rank() == 0:
         # l2 normalize the features
         database = nn.functional.normalize(database, dim=1, p=2)
         queries = nn.functional.normalize(queries, dim=1, p=2)

eval_image_retrieval.py → dino/eval_image_retrieval.py

@@ -25,9 +25,9 @@
 from PIL import Image, ImageFile
 import numpy as np
 
-import utils
-import vision_transformer as vits
-from eval_knn import extract_features
+import dino.utils
+import dino.vision_transformer as vits
+from dino.eval_knn import extract_features
 
 
 class OxfordParisDataset(torch.utils.data.Dataset):
@@ -83,10 +83,10 @@ def config_qimname(cfg, i):
     parser = argparse.ArgumentParser('Image Retrieval on revisited Paris and Oxford')
     parser.add_argument('--data_path', default='/path/to/revisited_paris_oxford/', type=str)
     parser.add_argument('--dataset', default='roxford5k', type=str, choices=['roxford5k', 'rparis6k'])
-    parser.add_argument('--multiscale', default=False, type=utils.bool_flag)
+    parser.add_argument('--multiscale', default=False, type=dino.utils.bool_flag)
     parser.add_argument('--imsize', default=224, type=int, help='Image size')
     parser.add_argument('--pretrained_weights', default='', type=str, help="Path to pretrained weights to evaluate.")
-    parser.add_argument('--use_cuda', default=True, type=utils.bool_flag)
+    parser.add_argument('--use_cuda', default=True, type=dino.utils.bool_flag)
     parser.add_argument('--arch', default='vit_small', type=str, help='Architecture')
     parser.add_argument('--patch_size', default=16, type=int, help='Patch resolution of the model.')
     parser.add_argument("--checkpoint_key", default="teacher", type=str,
@@ -97,8 +97,8 @@ def config_qimname(cfg, i):
     parser.add_argument("--local_rank", default=0, type=int, help="Please ignore and do not set this argument.")
     args = parser.parse_args()
 
-    utils.init_distributed_mode(args)
-    print("git:\n  {}\n".format(utils.get_sha()))
+    dino.utils.init_distributed_mode(args)
+    print("git:\n  {}\n".format(dino.utils.get_sha()))
     print("\n".join("%s: %s" % (k, str(v)) for k, v in sorted(dict(vars(args)).items())))
     cudnn.benchmark = True
 
@@ -165,7 +165,7 @@ def config_qimname(cfg, i):
     train_features = extract_features(model, data_loader_train, args.use_cuda, multiscale=args.multiscale)
     query_features = extract_features(model, data_loader_query, args.use_cuda, multiscale=args.multiscale)
 
-    if utils.get_rank() == 0:  # only rank 0 will work from now on
+    if dino.utils.get_rank() == 0:  # only rank 0 will work from now on
         # normalize features
         train_features = nn.functional.normalize(train_features, dim=1, p=2)
         query_features = nn.functional.normalize(query_features, dim=1, p=2)
@@ -187,15 +187,15 @@ def config_qimname(cfg, i):
             g['ok'] = np.concatenate([gnd[i]['easy'], gnd[i]['hard']])
             g['junk'] = np.concatenate([gnd[i]['junk']])
             gnd_t.append(g)
-        mapM, apsM, mprM, prsM = utils.compute_map(ranks, gnd_t, ks)
+        mapM, apsM, mprM, prsM = dino.utils.compute_map(ranks, gnd_t, ks)
         # search for hard
         gnd_t = []
         for i in range(len(gnd)):
             g = {}
             g['ok'] = np.concatenate([gnd[i]['hard']])
             g['junk'] = np.concatenate([gnd[i]['junk'], gnd[i]['easy']])
             gnd_t.append(g)
-        mapH, apsH, mprH, prsH = utils.compute_map(ranks, gnd_t, ks)
+        mapH, apsH, mprH, prsH = dino.utils.compute_map(ranks, gnd_t, ks)
         print('>> {}: mAP M: {}, H: {}'.format(args.dataset, np.around(mapM*100, decimals=2), np.around(mapH*100, decimals=2)))
         print('>> {}: mP@k{} M: {}, H: {}'.format(args.dataset, np.array(ks), np.around(mprM*100, decimals=2), np.around(mprH*100, decimals=2)))
     dist.barrier()

eval_knn.py → dino/eval_knn.py

@@ -23,8 +23,8 @@
 from torchvision import transforms as pth_transforms
 from torchvision import models as torchvision_models
 
-import utils
-import vision_transformer as vits
+import dino.utils
+import dino.vision_transformer as vits
 
 
 def extract_feature_pipeline(args):
@@ -68,7 +68,7 @@ def extract_feature_pipeline(args):
         print(f"Architecture {args.arch} non supported")
         sys.exit(1)
     model.cuda()
-    utils.load_pretrained_weights(model, args.pretrained_weights, args.checkpoint_key, args.arch, args.patch_size)
+    dino.utils.load_pretrained_weights(model, args.pretrained_weights, args.checkpoint_key, args.arch, args.patch_size)
     model.eval()
 
     # ============ extract features ... ============
@@ -77,7 +77,7 @@ def extract_feature_pipeline(args):
     print("Extracting features for val set...")
     test_features = extract_features(model, data_loader_val, args.use_cuda)
 
-    if utils.get_rank() == 0:
+    if dino.utils.get_rank() == 0:
         train_features = nn.functional.normalize(train_features, dim=1, p=2)
         test_features = nn.functional.normalize(test_features, dim=1, p=2)
 
@@ -94,13 +94,13 @@ def extract_feature_pipeline(args):
 
 @torch.no_grad()
 def extract_features(model, data_loader, use_cuda=True, multiscale=False):
-    metric_logger = utils.MetricLogger(delimiter="  ")
+    metric_logger = dino.utils.MetricLogger(delimiter="  ")
     features = None
     for samples, index in metric_logger.log_every(data_loader, 10):
         samples = samples.cuda(non_blocking=True)
         index = index.cuda(non_blocking=True)
         if multiscale:
-            feats = utils.multi_scale(samples, model)
+            feats = dino.utils.multi_scale(samples, model)
         else:
             feats = model(samples).clone()
 
@@ -196,7 +196,7 @@ def __getitem__(self, idx):
     parser.add_argument('--temperature', default=0.07, type=float,
         help='Temperature used in the voting coefficient')
     parser.add_argument('--pretrained_weights', default='', type=str, help="Path to pretrained weights to evaluate.")
-    parser.add_argument('--use_cuda', default=True, type=utils.bool_flag,
+    parser.add_argument('--use_cuda', default=True, type=dino.utils.bool_flag,
         help="Should we store the features on GPU? We recommend setting this to False if you encounter OOM")
     parser.add_argument('--arch', default='vit_small', type=str, help='Architecture')
     parser.add_argument('--patch_size', default=16, type=int, help='Patch resolution of the model.')
@@ -213,8 +213,8 @@ def __getitem__(self, idx):
     parser.add_argument('--data_path', default='/path/to/imagenet/', type=str)
     args = parser.parse_args()
 
-    utils.init_distributed_mode(args)
-    print("git:\n  {}\n".format(utils.get_sha()))
+    dino.utils.init_distributed_mode(args)
+    print("git:\n  {}\n".format(dino.utils.get_sha()))
     print("\n".join("%s: %s" % (k, str(v)) for k, v in sorted(dict(vars(args)).items())))
     cudnn.benchmark = True
 
@@ -227,7 +227,7 @@ def __getitem__(self, idx):
         # need to extract features !
         train_features, test_features, train_labels, test_labels = extract_feature_pipeline(args)
 
-    if utils.get_rank() == 0:
+    if dino.utils.get_rank() == 0:
         if args.use_cuda:
             train_features = train_features.cuda()
             test_features = test_features.cuda()

eval_linear.py → dino/eval_linear.py

@@ -24,13 +24,13 @@
 from torchvision import transforms as pth_transforms
 from torchvision import models as torchvision_models
 
-import utils
-import vision_transformer as vits
+import dino.utils
+import dino.vision_transformer as vits
 
 
 def eval_linear(args):
-    utils.init_distributed_mode(args)
-    print("git:\n  {}\n".format(utils.get_sha()))
+    dino.utils.init_distributed_mode(args)
+    print("git:\n  {}\n".format(dino.utils.get_sha()))
     print("\n".join("%s: %s" % (k, str(v)) for k, v in sorted(dict(vars(args)).items())))
     cudnn.benchmark = True
 
@@ -54,7 +54,7 @@ def eval_linear(args):
     model.cuda()
     model.eval()
     # load weights to evaluate
-    utils.load_pretrained_weights(model, args.pretrained_weights, args.checkpoint_key, args.arch, args.patch_size)
+    dino.utils.load_pretrained_weights(model, args.pretrained_weights, args.checkpoint_key, args.arch, args.patch_size)
     print(f"Model {args.arch} built.")
 
     linear_classifier = LinearClassifier(embed_dim, num_labels=args.num_labels)
@@ -77,7 +77,7 @@ def eval_linear(args):
     )
 
     if args.evaluate:
-        utils.load_pretrained_linear_weights(linear_classifier, args.arch, args.patch_size)
+        dino.utils.load_pretrained_linear_weights(linear_classifier, args.arch, args.patch_size)
         test_stats = validate_network(val_loader, model, linear_classifier, args.n_last_blocks, args.avgpool_patchtokens)
         print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%")
         return
@@ -102,15 +102,15 @@ def eval_linear(args):
     # set optimizer
     optimizer = torch.optim.SGD(
         linear_classifier.parameters(),
-        args.lr * (args.batch_size_per_gpu * utils.get_world_size()) / 256., # linear scaling rule
+        args.lr * (args.batch_size_per_gpu * dino.utils.get_world_size()) / 256., # linear scaling rule
         momentum=0.9,
         weight_decay=0, # we do not apply weight decay
     )
     scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs, eta_min=0)
 
     # Optionally resume from a checkpoint
     to_restore = {"epoch": 0, "best_acc": 0.}
-    utils.restart_from_checkpoint(
+    dino.utils.restart_from_checkpoint(
         os.path.join(args.output_dir, "checkpoint.pth.tar"),
         run_variables=to_restore,
         state_dict=linear_classifier,
@@ -135,7 +135,7 @@ def eval_linear(args):
             print(f'Max accuracy so far: {best_acc:.2f}%')
             log_stats = {**{k: v for k, v in log_stats.items()},
                          **{f'test_{k}': v for k, v in test_stats.items()}}
-        if utils.is_main_process():
+        if dino.utils.is_main_process():
             with (Path(args.output_dir) / "log.txt").open("a") as f:
                 f.write(json.dumps(log_stats) + "\n")
             save_dict = {
@@ -152,8 +152,8 @@ def eval_linear(args):
 
 def train(model, linear_classifier, optimizer, loader, epoch, n, avgpool):
     linear_classifier.train()
-    metric_logger = utils.MetricLogger(delimiter="  ")
-    metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
+    metric_logger = dino.utils.MetricLogger(delimiter="  ")
+    metric_logger.add_meter('lr', dino.utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
     header = 'Epoch: [{}]'.format(epoch)
     for (inp, target) in metric_logger.log_every(loader, 20, header):
         # move to gpu
@@ -195,7 +195,7 @@ def train(model, linear_classifier, optimizer, loader, epoch, n, avgpool):
 @torch.no_grad()
 def validate_network(val_loader, model, linear_classifier, n, avgpool):
     linear_classifier.eval()
-    metric_logger = utils.MetricLogger(delimiter="  ")
+    metric_logger = dino.utils.MetricLogger(delimiter="  ")
     header = 'Test:'
     for inp, target in metric_logger.log_every(val_loader, 20, header):
         # move to gpu
@@ -216,9 +216,9 @@ def validate_network(val_loader, model, linear_classifier, n, avgpool):
         loss = nn.CrossEntropyLoss()(output, target)
 
         if linear_classifier.module.num_labels >= 5:
-            acc1, acc5 = utils.accuracy(output, target, topk=(1, 5))
+            acc1, acc5 = dino.utils.accuracy(output, target, topk=(1, 5))
         else:
-            acc1, = utils.accuracy(output, target, topk=(1,))
+            acc1, = dino.utils.accuracy(output, target, topk=(1,))
 
         batch_size = inp.shape[0]
         metric_logger.update(loss=loss.item())
@@ -255,7 +255,7 @@ def forward(self, x):
     parser = argparse.ArgumentParser('Evaluation with linear classification on ImageNet')
     parser.add_argument('--n_last_blocks', default=4, type=int, help="""Concatenate [CLS] tokens
         for the `n` last blocks. We use `n=4` when evaluating ViT-Small and `n=1` with ViT-Base.""")
-    parser.add_argument('--avgpool_patchtokens', default=False, type=utils.bool_flag,
+    parser.add_argument('--avgpool_patchtokens', default=False, type=dino.utils.bool_flag,
         help="""Whether ot not to concatenate the global average pooled features to the [CLS] token.
         We typically set this to False for ViT-Small and to True with ViT-Base.""")
     parser.add_argument('--arch', default='vit_small', type=str, help='Architecture')

eval_video_segmentation.py → dino/eval_video_segmentation.py

@@ -30,8 +30,8 @@
 from PIL import Image
 from torchvision import transforms
 
-import utils
-import vision_transformer as vits
+import dino.utils
+import dino.vision_transformer as vits
 
 
 @torch.no_grad()
@@ -264,14 +264,14 @@ def color_normalize(x, mean=[0.485, 0.456, 0.406], std=[0.228, 0.224, 0.225]):
     parser.add_argument("--bs", type=int, default=6, help="Batch size, try to reduce if OOM")
     args = parser.parse_args()
 
-    print("git:\n  {}\n".format(utils.get_sha()))
+    print("git:\n  {}\n".format(dino.utils.get_sha()))
     print("\n".join("%s: %s" % (k, str(v)) for k, v in sorted(dict(vars(args)).items())))
 
     # building network
     model = vits.__dict__[args.arch](patch_size=args.patch_size, num_classes=0)
     print(f"Model {args.arch} {args.patch_size}x{args.patch_size} built.")
     model.cuda()
-    utils.load_pretrained_weights(model, args.pretrained_weights, args.checkpoint_key, args.arch, args.patch_size)
+    dino.utils.load_pretrained_weights(model, args.pretrained_weights, args.checkpoint_key, args.arch, args.patch_size)
     for param in model.parameters():
         param.requires_grad = False
     model.eval()

hubconf.py → dino/hubconf.py

@@ -14,7 +14,7 @@
 import torch
 from torchvision.models.resnet import resnet50
 
-import vision_transformer as vits
+import dino.vision_transformer as vits
 
 dependencies = ["torch", "torchvision"]