Upload 369 files

imagenet/resnet18_dino_pretrain/checkpoints/latest.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39dab92477179baf7d30671e45f352f32b2189719378df328dc239cbcf2818ee
+size 93614103

imagenet/resnet18_dino_pretrain/checkpoints/resnet18-acc70.754.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97642255c498d6fbb0ddd6d3d8d91398407850ded475c17a0b81d277ecc428e4
+size 46833625

imagenet/resnet18_dino_pretrain/test.sh

@@ -0,0 +1 @@
+OMP_NUM_THREADS=1 CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.run --nproc_per_node=2 --master_addr 127.0.1.0 --master_port 10000 ../../../tools/test_classification_model.py --work-dir ./

imagenet/resnet18_dino_pretrain/test_config.py

@@ -0,0 +1,55 @@
+import os
+import sys
+
+BASE_DIR = os.path.dirname(
+    os.path.dirname(os.path.dirname(os.path.dirname(
+        os.path.abspath(__file__)))))
+sys.path.append(BASE_DIR)
+
+from tools.path import ILSVRC2012_path
+
+from simpleAICV.classification import backbones
+from simpleAICV.classification import losses
+from simpleAICV.classification.datasets.ilsvrc2012dataset import ILSVRC2012Dataset
+from simpleAICV.classification.common import Opencv2PIL, TorchResize, TorchCenterCrop, TorchMeanStdNormalize, ClassificationCollater, load_state_dict
+
+import torch
+import torchvision.transforms as transforms
+
+
+class config:
+    '''
+    for resnet,input_image_size = 224;for darknet,input_image_size = 256
+    '''
+    network = 'resnet18'
+    num_classes = 1000
+    input_image_size = 224
+    scale = 256 / 224
+
+    model = backbones.__dict__[network](**{
+        'num_classes': num_classes,
+    })
+
+    # load pretrained model or not
+    trained_model_path = ''
+    load_state_dict(trained_model_path, model)
+
+    test_criterion = losses.__dict__['CELoss']()
+
+    test_dataset = ILSVRC2012Dataset(
+        root_dir=ILSVRC2012_path,
+        set_name='val',
+        transform=transforms.Compose([
+            Opencv2PIL(),
+            TorchResize(resize=input_image_size * scale),
+            TorchCenterCrop(resize=input_image_size),
+            TorchMeanStdNormalize(mean=[0.485, 0.456, 0.406],
+                                  std=[0.229, 0.224, 0.225]),
+        ]))
+    test_collater = ClassificationCollater()
+
+    seed = 0
+    # batch_size is total size
+    batch_size = 256
+    # num_workers is total workers
+    num_workers = 16

imagenet/resnet18_dino_pretrain/train.sh

@@ -0,0 +1 @@
+CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.run --nproc_per_node=1 --master_addr 127.0.1.0 --master_port 10000 ../../../tools/train_classification_model.py --work-dir ./

imagenet/resnet18_dino_pretrain/train_config.py

@@ -0,0 +1,107 @@
+import os
+import sys
+
+BASE_DIR = os.path.dirname(
+    os.path.dirname(os.path.dirname(os.path.dirname(
+        os.path.abspath(__file__)))))
+sys.path.append(BASE_DIR)
+
+from tools.path import ILSVRC2012_path
+
+from simpleAICV.classification import backbones
+from simpleAICV.classification import losses
+from simpleAICV.classification.datasets.ilsvrc2012dataset import ILSVRC2012Dataset
+from simpleAICV.classification.common import Opencv2PIL, TorchRandomResizedCrop, TorchRandomHorizontalFlip, TorchResize, TorchCenterCrop, TorchMeanStdNormalize, ClassificationCollater, load_state_dict
+
+import torch
+import torchvision.transforms as transforms
+
+
+class config:
+    '''
+    for resnet,input_image_size = 224;for darknet,input_image_size = 256
+    '''
+    network = 'resnet18'
+    num_classes = 1000
+    input_image_size = 224
+    scale = 256 / 224
+
+    model = backbones.__dict__[network](**{
+        'num_classes': num_classes,
+    })
+
+    # load pretrained model or not
+    trained_model_path = '/root/code/SimpleAICV_pytorch_training_examples_on_ImageNet_COCO_ADE20K/pretrained_models/resnet_dino_pretrain_on_imagenet1k/resnet18_dino_pretrain_model-student-loss3.081.pth'
+    load_state_dict(trained_model_path, model)
+
+    train_criterion = losses.__dict__['CELoss']()
+    test_criterion = losses.__dict__['CELoss']()
+
+    train_dataset = ILSVRC2012Dataset(
+        root_dir=ILSVRC2012_path,
+        set_name='train',
+        transform=transforms.Compose([
+            Opencv2PIL(),
+            TorchRandomResizedCrop(resize=input_image_size),
+            TorchRandomHorizontalFlip(prob=0.5),
+            TorchMeanStdNormalize(mean=[0.485, 0.456, 0.406],
+                                  std=[0.229, 0.224, 0.225]),
+        ]))
+
+    test_dataset = ILSVRC2012Dataset(
+        root_dir=ILSVRC2012_path,
+        set_name='val',
+        transform=transforms.Compose([
+            Opencv2PIL(),
+            TorchResize(resize=input_image_size * scale),
+            TorchCenterCrop(resize=input_image_size),
+            TorchMeanStdNormalize(mean=[0.485, 0.456, 0.406],
+                                  std=[0.229, 0.224, 0.225]),
+        ]))
+    train_collater = ClassificationCollater()
+    test_collater = ClassificationCollater()
+
+    seed = 0
+    # batch_size is total size
+    batch_size = 256
+    # num_workers is total workers
+    num_workers = 12
+    accumulation_steps = 1
+
+    optimizer = (
+        'SGD',
+        {
+            'lr': 0.1,
+            'momentum': 0.9,
+            'global_weight_decay': False,
+            # if global_weight_decay = False
+            # all bias, bn and other 1d params weight set to 0 weight decay
+            'weight_decay': 1e-4,
+            'no_weight_decay_layer_name_list': [],
+        },
+    )
+
+    scheduler = (
+        'MultiStepLR',
+        {
+            'warm_up_epochs': 0,
+            'gamma': 0.1,
+            'milestones': [30, 60, 90],
+        },
+    )
+
+    epochs = 100
+    print_interval = 100
+
+    sync_bn = False
+    use_amp = True
+    use_compile = False
+    compile_params = {
+        # 'default': optimizes for large models, low compile-time and no extra memory usage.
+        # 'reduce-overhead': optimizes to reduce the framework overhead and uses some extra memory, helps speed up small models, model update may not correct.
+        # 'max-autotune': optimizes to produce the fastest model, but takes a very long time to compile and may failed.
+        'mode': 'default',
+    }
+
+    use_ema_model = False
+    ema_model_decay = 0.9999

imagenet/resnet18_vit_aug_lion_dino_pretrain/checkpoints/latest.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:38197fc5648345b74bc25d1291aeb3dc75cc765f9be0b4ddefef2b4047f18cb7
+size 93613911

imagenet/resnet18_vit_aug_lion_dino_pretrain/checkpoints/resnet18-acc71.362.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2aac6552eb8967f4824937b08d70e2810f2d95a1ba9497a557f691bb52da0f5d
+size 46833625

imagenet/resnet18_vit_aug_lion_dino_pretrain/test.sh

@@ -0,0 +1 @@
+OMP_NUM_THREADS=1 CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.run --nproc_per_node=2 --master_addr 127.0.1.0 --master_port 10000 ../../../tools/test_classification_model.py --work-dir ./

imagenet/resnet18_vit_aug_lion_dino_pretrain/test_config.py

@@ -0,0 +1,55 @@
+import os
+import sys
+
+BASE_DIR = os.path.dirname(
+    os.path.dirname(os.path.dirname(os.path.dirname(
+        os.path.abspath(__file__)))))
+sys.path.append(BASE_DIR)
+
+from tools.path import ILSVRC2012_path
+
+from simpleAICV.classification import backbones
+from simpleAICV.classification import losses
+from simpleAICV.classification.datasets.ilsvrc2012dataset import ILSVRC2012Dataset
+from simpleAICV.classification.common import Opencv2PIL, TorchResize, TorchCenterCrop, TorchMeanStdNormalize, ClassificationCollater, load_state_dict
+
+import torch
+import torchvision.transforms as transforms
+
+
+class config:
+    '''
+    for resnet,input_image_size = 224;for darknet,input_image_size = 256
+    '''
+    network = 'resnet18'
+    num_classes = 1000
+    input_image_size = 224
+    scale = 256 / 224
+
+    model = backbones.__dict__[network](**{
+        'num_classes': num_classes,
+    })
+
+    # load pretrained model or not
+    trained_model_path = ''
+    load_state_dict(trained_model_path, model)
+
+    test_criterion = losses.__dict__['CELoss']()
+
+    test_dataset = ILSVRC2012Dataset(
+        root_dir=ILSVRC2012_path,
+        set_name='val',
+        transform=transforms.Compose([
+            Opencv2PIL(),
+            TorchResize(resize=input_image_size * scale),
+            TorchCenterCrop(resize=input_image_size),
+            TorchMeanStdNormalize(mean=[0.485, 0.456, 0.406],
+                                  std=[0.229, 0.224, 0.225]),
+        ]))
+    test_collater = ClassificationCollater()
+
+    seed = 0
+    # batch_size is total size
+    batch_size = 256
+    # num_workers is total workers
+    num_workers = 16

imagenet/resnet18_vit_aug_lion_dino_pretrain/train.sh

@@ -0,0 +1 @@
+CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.run --nproc_per_node=1 --master_addr 127.0.1.0 --master_port 10000 ../../../tools/train_classification_model.py --work-dir ./

imagenet/resnet18_vit_aug_lion_dino_pretrain/train_config.py

@@ -0,0 +1,125 @@
+import os
+import sys
+
+BASE_DIR = os.path.dirname(
+    os.path.dirname(os.path.dirname(os.path.dirname(
+        os.path.abspath(__file__)))))
+sys.path.append(BASE_DIR)
+
+from tools.path import ILSVRC2012_path
+
+from simpleAICV.classification import backbones
+from simpleAICV.classification import losses
+from simpleAICV.classification.datasets.ilsvrc2012dataset import ILSVRC2012Dataset
+from simpleAICV.classification.common import Opencv2PIL, TorchRandomResizedCrop, TorchRandomHorizontalFlip, RandAugment, TorchResize, TorchCenterCrop, TorchMeanStdNormalize, RandomErasing, ClassificationCollater, MixupCutmixClassificationCollater, load_state_dict
+
+import torch
+import torchvision.transforms as transforms
+
+
+class config:
+    '''
+    for resnet,input_image_size = 224;for darknet,input_image_size = 256
+    '''
+    network = 'resnet18'
+    num_classes = 1000
+    input_image_size = 224
+    scale = 256 / 224
+
+    model = backbones.__dict__[network](**{
+        'num_classes': num_classes,
+    })
+
+    # load pretrained model or not
+    trained_model_path = '/root/code/SimpleAICV_pytorch_training_examples_on_ImageNet_COCO_ADE20K/pretrained_models/resnet_dino_pretrain_on_imagenet1k/resnet18_dino_pretrain_model-student-loss3.081.pth'
+    load_state_dict(trained_model_path, model)
+
+    train_criterion = losses.__dict__['OneHotLabelCELoss']()
+    test_criterion = losses.__dict__['CELoss']()
+
+    train_dataset = ILSVRC2012Dataset(
+        root_dir=ILSVRC2012_path,
+        set_name='train',
+        transform=transforms.Compose([
+            Opencv2PIL(),
+            TorchRandomResizedCrop(resize=input_image_size),
+            TorchRandomHorizontalFlip(prob=0.5),
+            RandAugment(magnitude=9,
+                        num_layers=2,
+                        resize=input_image_size,
+                        mean=[0.485, 0.456, 0.406],
+                        integer=True,
+                        weight_idx=None,
+                        magnitude_std=0.5,
+                        magnitude_max=None),
+            TorchMeanStdNormalize(mean=[0.485, 0.456, 0.406],
+                                  std=[0.229, 0.224, 0.225]),
+            RandomErasing(prob=0.25, mode='pixel', max_count=1),
+        ]))
+
+    test_dataset = ILSVRC2012Dataset(
+        root_dir=ILSVRC2012_path,
+        set_name='val',
+        transform=transforms.Compose([
+            Opencv2PIL(),
+            TorchResize(resize=input_image_size * scale),
+            TorchCenterCrop(resize=input_image_size),
+            TorchMeanStdNormalize(mean=[0.485, 0.456, 0.406],
+                                  std=[0.229, 0.224, 0.225]),
+        ]))
+
+    train_collater = MixupCutmixClassificationCollater(
+        use_mixup=True,
+        mixup_alpha=0.8,
+        cutmix_alpha=1.0,
+        cutmix_minmax=None,
+        mixup_cutmix_prob=1.0,
+        switch_to_cutmix_prob=0.5,
+        mode='batch',
+        correct_lam=True,
+        label_smoothing=0.1,
+        num_classes=1000)
+    test_collater = ClassificationCollater()
+
+    seed = 0
+    # batch_size is total size
+    batch_size = 256
+    # num_workers is total workers
+    num_workers = 12
+    accumulation_steps = 16
+
+    optimizer = (
+        'Lion',
+        {
+            'lr': 4e-4,
+            'global_weight_decay': False,
+            # if global_weight_decay = False
+            # all bias, bn and other 1d params weight set to 0 weight decay
+            'weight_decay': 1e-4,
+            'no_weight_decay_layer_name_list': [],
+        },
+    )
+
+    scheduler = (
+        'CosineLR',
+        {
+            'warm_up_epochs': 5,
+            'min_lr': 1e-6,
+        },
+    )
+
+    epochs = 300
+    print_interval = 10
+
+    sync_bn = False
+    use_amp = False
+    use_compile = False
+    compile_params = {
+        # 'default': optimizes for large models, low compile-time and no extra memory usage.
+        # 'reduce-overhead': optimizes to reduce the framework overhead and uses some extra memory, helps speed up small models, model update may not correct.
+        # 'max-autotune': optimizes to produce the fastest model, but takes a very long time to compile and may failed.
+        'mode': 'default',
+    }
+
+    use_ema_model = False
+    ema_model_decay = 0.9999

imagenet/resnet34_dino_pretrain/checkpoints/latest.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa38f299648300ef6fae63c1e59a41ee7cf44cf72404bee6fb6f6dc29e568635
+size 174559527

imagenet/resnet34_dino_pretrain/checkpoints/resnet34-acc74.218.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:252bacee48f03193926ff3f585f5aa4b5683022f88df1e817f242c8c48920d98
+size 87325369

imagenet/resnet34_dino_pretrain/test.sh

@@ -0,0 +1 @@
+OMP_NUM_THREADS=1 CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.run --nproc_per_node=2 --master_addr 127.0.1.0 --master_port 10000 ../../../tools/test_classification_model.py --work-dir ./

imagenet/resnet34_dino_pretrain/test_config.py

@@ -0,0 +1,55 @@
+import os
+import sys
+
+BASE_DIR = os.path.dirname(
+    os.path.dirname(os.path.dirname(os.path.dirname(
+        os.path.abspath(__file__)))))
+sys.path.append(BASE_DIR)
+
+from tools.path import ILSVRC2012_path
+
+from simpleAICV.classification import backbones
+from simpleAICV.classification import losses
+from simpleAICV.classification.datasets.ilsvrc2012dataset import ILSVRC2012Dataset
+from simpleAICV.classification.common import Opencv2PIL, TorchResize, TorchCenterCrop, TorchMeanStdNormalize, ClassificationCollater, load_state_dict
+
+import torch
+import torchvision.transforms as transforms
+
+
+class config:
+    '''
+    for resnet,input_image_size = 224;for darknet,input_image_size = 256
+    '''
+    network = 'resnet34'
+    num_classes = 1000
+    input_image_size = 224
+    scale = 256 / 224
+
+    model = backbones.__dict__[network](**{
+        'num_classes': num_classes,
+    })
+
+    # load pretrained model or not
+    trained_model_path = ''
+    load_state_dict(trained_model_path, model)
+
+    test_criterion = losses.__dict__['CELoss']()
+
+    test_dataset = ILSVRC2012Dataset(
+        root_dir=ILSVRC2012_path,
+        set_name='val',
+        transform=transforms.Compose([
+            Opencv2PIL(),
+            TorchResize(resize=input_image_size * scale),
+            TorchCenterCrop(resize=input_image_size),
+            TorchMeanStdNormalize(mean=[0.485, 0.456, 0.406],
+                                  std=[0.229, 0.224, 0.225]),
+        ]))
+    test_collater = ClassificationCollater()
+
+    seed = 0
+    # batch_size is total size
+    batch_size = 256
+    # num_workers is total workers
+    num_workers = 16

imagenet/resnet34_dino_pretrain/train.sh

@@ -0,0 +1 @@
+CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.run --nproc_per_node=1 --master_addr 127.0.1.0 --master_port 10000 ../../../tools/train_classification_model.py --work-dir ./

imagenet/resnet34_dino_pretrain/train_config.py

@@ -0,0 +1,107 @@
+import os
+import sys
+
+BASE_DIR = os.path.dirname(
+    os.path.dirname(os.path.dirname(os.path.dirname(
+        os.path.abspath(__file__)))))
+sys.path.append(BASE_DIR)
+
+from tools.path import ILSVRC2012_path
+
+from simpleAICV.classification import backbones
+from simpleAICV.classification import losses
+from simpleAICV.classification.datasets.ilsvrc2012dataset import ILSVRC2012Dataset
+from simpleAICV.classification.common import Opencv2PIL, TorchRandomResizedCrop, TorchRandomHorizontalFlip, TorchResize, TorchCenterCrop, TorchMeanStdNormalize, ClassificationCollater, load_state_dict
+
+import torch
+import torchvision.transforms as transforms
+
+
+class config:
+    '''
+    for resnet,input_image_size = 224;for darknet,input_image_size = 256
+    '''
+    network = 'resnet34'
+    num_classes = 1000
+    input_image_size = 224
+    scale = 256 / 224
+
+    model = backbones.__dict__[network](**{
+        'num_classes': num_classes,
+    })
+
+    # load pretrained model or not
+    trained_model_path = '/root/code/SimpleAICV_pytorch_training_examples_on_ImageNet_COCO_ADE20K/pretrained_models/resnet_dino_pretrain_on_imagenet1k/resnet34_dino_pretrain_model-student-loss2.425.pth'
+    load_state_dict(trained_model_path, model)
+
+    train_criterion = losses.__dict__['CELoss']()
+    test_criterion = losses.__dict__['CELoss']()
+
+    train_dataset = ILSVRC2012Dataset(
+        root_dir=ILSVRC2012_path,
+        set_name='train',
+        transform=transforms.Compose([
+            Opencv2PIL(),
+            TorchRandomResizedCrop(resize=input_image_size),
+            TorchRandomHorizontalFlip(prob=0.5),
+            TorchMeanStdNormalize(mean=[0.485, 0.456, 0.406],
+                                  std=[0.229, 0.224, 0.225]),
+        ]))
+
+    test_dataset = ILSVRC2012Dataset(
+        root_dir=ILSVRC2012_path,
+        set_name='val',
+        transform=transforms.Compose([
+            Opencv2PIL(),
+            TorchResize(resize=input_image_size * scale),
+            TorchCenterCrop(resize=input_image_size),
+            TorchMeanStdNormalize(mean=[0.485, 0.456, 0.406],
+                                  std=[0.229, 0.224, 0.225]),
+        ]))
+    train_collater = ClassificationCollater()
+    test_collater = ClassificationCollater()
+
+    seed = 0
+    # batch_size is total size
+    batch_size = 256
+    # num_workers is total workers
+    num_workers = 12
+    accumulation_steps = 1
+
+    optimizer = (
+        'SGD',
+        {
+            'lr': 0.1,
+            'momentum': 0.9,
+            'global_weight_decay': False,
+            # if global_weight_decay = False
+            # all bias, bn and other 1d params weight set to 0 weight decay
+            'weight_decay': 1e-4,
+            'no_weight_decay_layer_name_list': [],
+        },
+    )
+
+    scheduler = (
+        'MultiStepLR',
+        {
+            'warm_up_epochs': 0,
+            'gamma': 0.1,
+            'milestones': [30, 60, 90],
+        },
+    )
+
+    epochs = 100
+    print_interval = 100
+
+    sync_bn = False
+    use_amp = True
+    use_compile = False
+    compile_params = {
+        # 'default': optimizes for large models, low compile-time and no extra memory usage.
+        # 'reduce-overhead': optimizes to reduce the framework overhead and uses some extra memory, helps speed up small models, model update may not correct.
+        # 'max-autotune': optimizes to produce the fastest model, but takes a very long time to compile and may failed.
+        'mode': 'default',
+    }
+
+    use_ema_model = False
+    ema_model_decay = 0.9999

imagenet/resnet34_vit_aug_lion_dino_pretrain/checkpoints/latest.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdd1263db3e8894b16180c2702546cf31600473bb196b7b33422c5f7f46da551
+size 174559399

imagenet/resnet34_vit_aug_lion_dino_pretrain/checkpoints/resnet34-acc75.916.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06f7e7e7bf95b527b57211cabd133ce3e4121886882dd2e238aea1e08b9bd95c
+size 87325369

imagenet/resnet34_vit_aug_lion_dino_pretrain/test.sh

@@ -0,0 +1 @@
+OMP_NUM_THREADS=1 CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.run --nproc_per_node=2 --master_addr 127.0.1.0 --master_port 10000 ../../../tools/test_classification_model.py --work-dir ./

imagenet/resnet34_vit_aug_lion_dino_pretrain/test_config.py

@@ -0,0 +1,55 @@
+import os
+import sys
+
+BASE_DIR = os.path.dirname(
+    os.path.dirname(os.path.dirname(os.path.dirname(
+        os.path.abspath(__file__)))))
+sys.path.append(BASE_DIR)
+
+from tools.path import ILSVRC2012_path
+
+from simpleAICV.classification import backbones
+from simpleAICV.classification import losses
+from simpleAICV.classification.datasets.ilsvrc2012dataset import ILSVRC2012Dataset
+from simpleAICV.classification.common import Opencv2PIL, TorchResize, TorchCenterCrop, TorchMeanStdNormalize, ClassificationCollater, load_state_dict
+
+import torch
+import torchvision.transforms as transforms
+
+
+class config:
+    '''
+    for resnet,input_image_size = 224;for darknet,input_image_size = 256
+    '''
+    network = 'resnet34'
+    num_classes = 1000
+    input_image_size = 224
+    scale = 256 / 224
+
+    model = backbones.__dict__[network](**{
+        'num_classes': num_classes,
+    })
+
+    # load pretrained model or not
+    trained_model_path = ''
+    load_state_dict(trained_model_path, model)
+
+    test_criterion = losses.__dict__['CELoss']()
+
+    test_dataset = ILSVRC2012Dataset(
+        root_dir=ILSVRC2012_path,
+        set_name='val',
+        transform=transforms.Compose([
+            Opencv2PIL(),
+            TorchResize(resize=input_image_size * scale),
+            TorchCenterCrop(resize=input_image_size),
+            TorchMeanStdNormalize(mean=[0.485, 0.456, 0.406],
+                                  std=[0.229, 0.224, 0.225]),
+        ]))
+    test_collater = ClassificationCollater()
+
+    seed = 0
+    # batch_size is total size
+    batch_size = 256
+    # num_workers is total workers
+    num_workers = 16

imagenet/resnet34_vit_aug_lion_dino_pretrain/train.sh

@@ -0,0 +1 @@
+CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.run --nproc_per_node=2 --master_addr 127.0.1.0 --master_port 10000 ../../../tools/train_classification_model.py --work-dir ./

imagenet/resnet34_vit_aug_lion_dino_pretrain/train_config.py

@@ -0,0 +1,125 @@
+import os
+import sys
+
+BASE_DIR = os.path.dirname(
+    os.path.dirname(os.path.dirname(os.path.dirname(
+        os.path.abspath(__file__)))))
+sys.path.append(BASE_DIR)
+
+from tools.path import ILSVRC2012_path
+
+from simpleAICV.classification import backbones
+from simpleAICV.classification import losses
+from simpleAICV.classification.datasets.ilsvrc2012dataset import ILSVRC2012Dataset
+from simpleAICV.classification.common import Opencv2PIL, TorchRandomResizedCrop, TorchRandomHorizontalFlip, RandAugment, TorchResize, TorchCenterCrop, TorchMeanStdNormalize, RandomErasing, ClassificationCollater, MixupCutmixClassificationCollater, load_state_dict
+
+import torch
+import torchvision.transforms as transforms
+
+
+class config:
+    '''
+    for resnet,input_image_size = 224;for darknet,input_image_size = 256
+    '''
+    network = 'resnet34'
+    num_classes = 1000
+    input_image_size = 224
+    scale = 256 / 224
+
+    model = backbones.__dict__[network](**{
+        'num_classes': num_classes,
+    })
+
+    # load pretrained model or not
+    trained_model_path = '/root/code/SimpleAICV_pytorch_training_examples_on_ImageNet_COCO_ADE20K/pretrained_models/resnet_dino_pretrain_on_imagenet1k/resnet34_dino_pretrain_model-student-loss2.425.pth'
+    load_state_dict(trained_model_path, model)
+
+    train_criterion = losses.__dict__['OneHotLabelCELoss']()
+    test_criterion = losses.__dict__['CELoss']()
+
+    train_dataset = ILSVRC2012Dataset(
+        root_dir=ILSVRC2012_path,
+        set_name='train',
+        transform=transforms.Compose([
+            Opencv2PIL(),
+            TorchRandomResizedCrop(resize=input_image_size),
+            TorchRandomHorizontalFlip(prob=0.5),
+            RandAugment(magnitude=9,
+                        num_layers=2,
+                        resize=input_image_size,
+                        mean=[0.485, 0.456, 0.406],
+                        integer=True,
+                        weight_idx=None,
+                        magnitude_std=0.5,
+                        magnitude_max=None),
+            TorchMeanStdNormalize(mean=[0.485, 0.456, 0.406],
+                                  std=[0.229, 0.224, 0.225]),
+            RandomErasing(prob=0.25, mode='pixel', max_count=1),
+        ]))
+
+    test_dataset = ILSVRC2012Dataset(
+        root_dir=ILSVRC2012_path,
+        set_name='val',
+        transform=transforms.Compose([
+            Opencv2PIL(),
+            TorchResize(resize=input_image_size * scale),
+            TorchCenterCrop(resize=input_image_size),
+            TorchMeanStdNormalize(mean=[0.485, 0.456, 0.406],
+                                  std=[0.229, 0.224, 0.225]),
+        ]))
+
+    train_collater = MixupCutmixClassificationCollater(
+        use_mixup=True,
+        mixup_alpha=0.8,
+        cutmix_alpha=1.0,
+        cutmix_minmax=None,
+        mixup_cutmix_prob=1.0,
+        switch_to_cutmix_prob=0.5,
+        mode='batch',
+        correct_lam=True,
+        label_smoothing=0.1,
+        num_classes=1000)
+    test_collater = ClassificationCollater()
+
+    seed = 0
+    # batch_size is total size
+    batch_size = 256
+    # num_workers is total workers
+    num_workers = 24
+    accumulation_steps = 16
+
+    optimizer = (
+        'Lion',
+        {
+            'lr': 4e-4,
+            'global_weight_decay': False,
+            # if global_weight_decay = False
+            # all bias, bn and other 1d params weight set to 0 weight decay
+            'weight_decay': 1e-4,
+            'no_weight_decay_layer_name_list': [],
+        },
+    )
+
+    scheduler = (
+        'CosineLR',
+        {
+            'warm_up_epochs': 5,
+            'min_lr': 1e-6,
+        },
+    )
+
+    epochs = 300
+    print_interval = 10
+
+    sync_bn = False
+    use_amp = False
+    use_compile = False
+    compile_params = {
+        # 'default': optimizes for large models, low compile-time and no extra memory usage.
+        # 'reduce-overhead': optimizes to reduce the framework overhead and uses some extra memory, helps speed up small models, model update may not correct.
+        # 'max-autotune': optimizes to produce the fastest model, but takes a very long time to compile and may failed.
+        'mode': 'default',
+    }
+
+    use_ema_model = False
+    ema_model_decay = 0.9999