diff --git a/cv/classification/cspdarknet53/pytorch/README.md b/cv/classification/cspdarknet53/pytorch/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..5fd274239c8c66ddb3675869f9848d0c88f80745
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/README.md
@@ -0,0 +1,38 @@
+# CspDarknet53
+
+## Model description
+
+This is an implementation of CSPDarknet53 in pytorch.
+
+## Step 1: Installing
+
+```bash
+pip3 install torch
+pip3 install torchvision
+```
+
+## Step 2: Training
+
+### One single GPU
+
+```bash
+export CUDA_VISIBLE_DEVICES=0
+python3 train.py --batch-size 64 --epochs 120 --data-path /home/datasets/cv/imagenet
+```
+
+### 8 GPUs on one machine
+```bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+python3 -m torch.distributed.launch --nproc_per_node=8 --use_env train.py --batch-size 64 --epochs 120 --data-path /home/datasets/cv/imagenet
+```
+
+## Result
+
+| GPU         | FP32                                 |
+| ----------- | ------------------------------------ |
+| 8 cards     |  Acc@1 76.644     fps 830            |
+| 8 cards     |                   fps 148            |
+
+## Reference
+
+https://github.com/WongKinYiu/CrossStagePartialNetworks
diff --git a/cv/classification/cspdarknet53/pytorch/__init__.py b/cv/classification/cspdarknet53/pytorch/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..011573976ed1c1e69c9282a42400acc3e68c3586
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/__init__.py
@@ -0,0 +1,20 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# 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.
+
+from .utils import *
+from .common_utils import *
+from .data_loader import *
+
+__all__ = [k for k in globals().keys() if not k.startswith("_")]
diff --git a/cv/classification/cspdarknet53/pytorch/common_utils/__init__.py b/cv/classification/cspdarknet53/pytorch/common_utils/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..7d2e011f7cec1af806b4635fcf65849b55f1111b
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/common_utils/__init__.py
@@ -0,0 +1,38 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# 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.
+
+import random
+
+import numpy as np
+
+from .dist import *
+from .metric_logger import *
+from .misc import *
+from .smooth_value import *
+
+def manual_seed(seed, deterministic=False):
+    random.seed(seed)
+    np.random.seed(seed)
+    os.environ['PYTHONHASHSEED'] = str(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+
+    if deterministic:
+        torch.backends.cudnn.deterministic = True
+        torch.backends.cudnn.benchmark = False
+    else:
+        torch.backends.cudnn.deterministic = False
+        torch.backends.cudnn.benchmark = True
diff --git a/cv/classification/cspdarknet53/pytorch/common_utils/dist.py b/cv/classification/cspdarknet53/pytorch/common_utils/dist.py
new file mode 100644
index 0000000000000000000000000000000000000000..767b6ce0008b75447ed75016f7e973ab086b992e
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/common_utils/dist.py
@@ -0,0 +1,156 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# All Rights Reserved.
+# Copyright (c) Facebook, Inc. and its affiliates. 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.
+
+from collections import defaultdict, deque
+import datetime
+import errno
+import os
+import time
+
+import torch
+import torch.distributed as dist
+
+
+
+def setup_for_distributed(is_master):
+    """
+    This function disables printing when not in master process
+    """
+    import builtins as __builtin__
+    builtin_print = __builtin__.print
+
+    def print(*args, **kwargs):
+        force = kwargs.pop('force', False)
+        if is_master or force:
+            builtin_print(*args, **kwargs)
+
+    __builtin__.print = print
+
+
+def is_dist_avail_and_initialized():
+    if not dist.is_available():
+        return False
+    if not dist.is_initialized():
+        return False
+    return True
+
+
+def get_world_size():
+    if not is_dist_avail_and_initialized():
+        return 1
+    return dist.get_world_size()
+
+
+def get_rank():
+    if not is_dist_avail_and_initialized():
+        return 0
+    return dist.get_rank()
+
+
+def is_main_process():
+    return get_rank() == 0
+
+
+def save_on_master(*args, **kwargs):
+    if is_main_process():
+        torch.save(*args, **kwargs)
+
+
+def get_dist_backend(args=None):
+    DIST_BACKEND_ENV = "PT_DIST_BACKEND"
+    if DIST_BACKEND_ENV in os.environ:
+        return os.environ[DIST_BACKEND_ENV]
+
+    if args is None:
+        args = dict()
+
+    backend_attr_name = "dist_backend"
+
+    if hasattr(args, backend_attr_name):
+        return getattr(args, backend_attr_name)
+
+    if backend_attr_name in args:
+        return args[backend_attr_name]
+
+    return "nccl"
+
+
+def init_distributed_mode(args):
+    if 'RANK' in os.environ and 'WORLD_SIZE' in os.environ:
+        args.rank = int(os.environ["RANK"])
+        args.world_size = int(os.environ['WORLD_SIZE'])
+        args.gpu = int(os.environ['LOCAL_RANK'])
+    elif 'SLURM_PROCID' in os.environ:
+        args.rank = int(os.environ['SLURM_PROCID'])
+        args.gpu = args.rank % torch.cuda.device_count()
+    else:
+        print('Not using distributed mode')
+        args.distributed = False
+        return
+
+    args.distributed = True
+
+    torch.cuda.set_device(args.gpu)
+    dist_backend = get_dist_backend(args)
+    print('| distributed init (rank {}): {}'.format(
+        args.rank, args.dist_url), flush=True)
+    torch.distributed.init_process_group(backend=dist_backend, init_method=args.dist_url,
+                                         world_size=args.world_size, rank=args.rank)
+    torch.distributed.barrier()
+    setup_for_distributed(args.rank == 0)
+
+
+def all_gather(data):
+    """
+    Run all_gather on arbitrary picklable data (not necessarily tensors)
+    Args:
+        data: any picklable object
+    Returns:
+        list[data]: list of data gathered from each rank
+    """
+    world_size = get_world_size()
+    if world_size == 1:
+        return [data]
+    data_list = [None] * world_size
+    dist.all_gather_object(data_list, data)
+    return data_list
+
+
+def reduce_dict(input_dict, average=True):
+    """
+    Args:
+        input_dict (dict): all the values will be reduced
+        average (bool): whether to do average or sum
+    Reduce the values in the dictionary from all processes so that all processes
+    have the averaged results. Returns a dict with the same fields as
+    input_dict, after reduction.
+    """
+    world_size = get_world_size()
+    if world_size < 2:
+        return input_dict
+    with torch.no_grad():
+        names = []
+        values = []
+        # sort the keys so that they are consistent across processes
+        for k in sorted(input_dict.keys()):
+            names.append(k)
+            values.append(input_dict[k])
+        values = torch.stack(values, dim=0)
+        dist.all_reduce(values)
+        if average:
+            values /= world_size
+        reduced_dict = {k: v for k, v in zip(names, values)}
+    return reduced_dict
diff --git a/cv/classification/cspdarknet53/pytorch/common_utils/metric_logger.py b/cv/classification/cspdarknet53/pytorch/common_utils/metric_logger.py
new file mode 100644
index 0000000000000000000000000000000000000000..ab9c61b0ba26212fb9404268bfd6f681a35f66c7
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/common_utils/metric_logger.py
@@ -0,0 +1,106 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# All Rights Reserved.
+# Copyright (c) Facebook, Inc. and its affiliates. 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.
+
+from collections import defaultdict
+import datetime
+import time
+
+import torch
+from .smooth_value import SmoothedValue
+
+"""
+Examples:
+
+logger = MetricLogger("  ")
+
+>>> # For iter dataloader
+>>> metric_logger.add_meter('img/s', utils.SmoothedValue(window_size=10, fmt='{value}'))
+>>> header = 'Epoch: [{}]'.format(epoch)
+>>> for image, target in metric_logger.log_every(data_loader, print_freq, header):
+>>>     ...
+>>>     logger.metric_logger.meters['img/s'].update(fps)
+
+"""
+
+class MetricLogger(object):
+
+    def __init__(self, delimiter="\t"):
+        self.meters = defaultdict(SmoothedValue)
+        self.delimiter = delimiter
+
+    def update(self, **kwargs):
+        for k, v in kwargs.items():
+            if isinstance(v, torch.Tensor):
+                v = v.item()
+            assert isinstance(v, (float, int))
+            self.meters[k].update(v)
+
+    def __getattr__(self, attr):
+        if attr in self.meters:
+            return self.meters[attr]
+        if attr in self.__dict__:
+            return self.__dict__[attr]
+        raise AttributeError("'{}' object has no attribute '{}'".format(
+            type(self).__name__, attr))
+
+    def __str__(self):
+        loss_str = []
+        for name, meter in self.meters.items():
+            loss_str.append(
+                "{}: {}".format(name, str(meter))
+            )
+        return self.delimiter.join(loss_str)
+
+    def synchronize_between_processes(self):
+        for meter in self.meters.values():
+            meter.synchronize_between_processes()
+
+    def add_meter(self, name, meter):
+        self.meters[name] = meter
+
+    def log_every(self, iterable, print_freq, header=None):
+        i = 0
+        if not header:
+            header = ''
+        start_time = time.time()
+        end = time.time()
+        iter_time = SmoothedValue(fmt='{avg:.4f}')
+        data_time = SmoothedValue(fmt='{avg:.4f}')
+        space_fmt = ':' + str(len(str(len(iterable)))) + 'd'
+        log_msg = self.delimiter.join([
+            header,
+            '[{0' + space_fmt + '}/{1}]',
+            'eta: {eta}',
+            '{meters}',
+            'time: {time}',
+            'data: {data}'
+        ])
+        for obj in iterable:
+            data_time.update(time.time() - end)
+            yield obj
+            iter_time.update(time.time() - end)
+            if i % print_freq == 0:
+                eta_seconds = iter_time.global_avg * (len(iterable) - i)
+                eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
+                print(log_msg.format(
+                    i, len(iterable), eta=eta_string,
+                    meters=str(self),
+                    time=str(iter_time), data=str(data_time)))
+            i += 1
+            end = time.time()
+        total_time = time.time() - start_time
+        total_time_str = str(datetime.timedelta(seconds=int(total_time)))
+        print('{} Total time: {}'.format(header, total_time_str))
diff --git a/cv/classification/cspdarknet53/pytorch/common_utils/misc.py b/cv/classification/cspdarknet53/pytorch/common_utils/misc.py
new file mode 100644
index 0000000000000000000000000000000000000000..8c5e49bcd21622b00e41551257ced4473c97b450
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/common_utils/misc.py
@@ -0,0 +1,26 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# 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.
+
+import os
+import sys
+import errno
+
+
+def mkdir(path):
+    try:
+        os.makedirs(path)
+    except OSError as e:
+        if e.errno != errno.EEXIST:
+            raise
diff --git a/cv/classification/cspdarknet53/pytorch/common_utils/smooth_value.py b/cv/classification/cspdarknet53/pytorch/common_utils/smooth_value.py
new file mode 100644
index 0000000000000000000000000000000000000000..8c1fadae25ec148a0af2cc55a544e098d6393247
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/common_utils/smooth_value.py
@@ -0,0 +1,88 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# All Rights Reserved.
+# Copyright (c) Facebook, Inc. and its affiliates. 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.
+
+
+from collections import defaultdict, deque
+import datetime
+import errno
+import os
+import time
+
+import torch
+import torch.distributed as dist
+from .dist import is_dist_avail_and_initialized
+
+
+class SmoothedValue(object):
+    """Track a series of values and provide access to smoothed values over a
+    window or the global series average.
+    """
+
+    def __init__(self, window_size=20, fmt=None):
+        if fmt is None:
+            fmt = "{median:.4f} ({global_avg:.4f})"
+        self.deque = deque(maxlen=window_size)
+        self.total = 0.0
+        self.count = 0
+        self.fmt = fmt
+
+    def update(self, value, n=1):
+        self.deque.append(value)
+        self.count += n
+        self.total += value * n
+
+    def synchronize_between_processes(self):
+        """
+        Warning: does not synchronize the deque!
+        """
+        if not is_dist_avail_and_initialized():
+            return
+        t = torch.tensor([self.count, self.total], dtype=torch.float32, device='cuda')
+        dist.barrier()
+        dist.all_reduce(t)
+        t = t.tolist()
+        self.count = int(t[0])
+        self.total = t[1]
+
+    @property
+    def median(self):
+        d = torch.tensor(list(self.deque))
+        return d.median().item()
+
+    @property
+    def avg(self):
+        d = torch.tensor(list(self.deque), dtype=torch.float32)
+        return d.mean().item()
+
+    @property
+    def global_avg(self):
+        return self.total / self.count
+
+    @property
+    def max(self):
+        return max(self.deque)
+
+    @property
+    def value(self):
+        return self.deque[-1]
+
+    def __str__(self):
+        return self.fmt.format(
+            median=self.median,
+            avg=self.avg,
+            global_avg=self.global_avg,
+            max=self.max,
+            value=self.value)
diff --git a/cv/classification/cspdarknet53/pytorch/dataloader/__init__.py b/cv/classification/cspdarknet53/pytorch/dataloader/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..66b217b8507dfb048e3a366fc0fe6374f24eedc4
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/dataloader/__init__.py
@@ -0,0 +1,14 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# 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.
\ No newline at end of file
diff --git a/cv/classification/cspdarknet53/pytorch/dataloader/classification.py b/cv/classification/cspdarknet53/pytorch/dataloader/classification.py
new file mode 100644
index 0000000000000000000000000000000000000000..317a9874e0c561de93b0eae48f3e664422447974
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/dataloader/classification.py
@@ -0,0 +1,120 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# All Rights Reserved.
+# Copyright (c) Facebook, Inc. and its affiliates. 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.
+
+
+import os
+import time
+
+import torch
+import torchvision
+from .utils import presets_classification as presets
+
+"""
+Examples:
+
+>>> dataset_train, dataset_val = load_data(train_dir, val_dir, args)
+"""
+
+
+def get_datasets(traindir,
+                 valdir,
+                 resize_size=256,
+                 crop_size=224,
+                 auto_augment_policy=None,
+                 random_erase_prob=0.):
+    # Data loading code
+    print("Loading data")
+    print("Loading training data")
+    dataset = torchvision.datasets.ImageFolder(
+        traindir,
+        presets.ClassificationPresetTrain(crop_size=crop_size, auto_augment_policy=auto_augment_policy,
+                                          random_erase_prob=random_erase_prob))
+
+    print("Loading validation data")
+    dataset_test = torchvision.datasets.ImageFolder(
+        valdir,
+        presets.ClassificationPresetEval(crop_size=crop_size, resize_size=resize_size))
+
+    return dataset, dataset_test
+
+
+def get_input_size(model):
+    biger_input_size_models = ['inception']
+    resize_size = 256
+    crop_size = 224
+    for bi_model in biger_input_size_models:
+        if bi_model in model:
+            resize_size = 342
+            crop_size = 299
+
+    return resize_size, crop_size
+
+
+def load_data(train_dir, val_dir, args):
+    auto_augment_policy = getattr(args, "auto_augment", None)
+    random_erase_prob = getattr(args, "random_erase", 0.0)
+    resize_size, crop_size = get_input_size(args.model)
+    dataset, dataset_test = get_datasets(train_dir, val_dir,
+                                         auto_augment_policy=auto_augment_policy,
+                                         random_erase_prob=random_erase_prob,
+                                         resize_size=resize_size,
+                                         crop_size=crop_size)
+    if args.distributed:
+        train_sampler = torch.utils.data.distributed.DistributedSampler(dataset)
+        test_sampler = torch.utils.data.distributed.DistributedSampler(dataset_test)
+    else:
+        train_sampler = torch.utils.data.RandomSampler(dataset)
+        test_sampler = torch.utils.data.SequentialSampler(dataset_test)
+
+    return dataset, dataset_test, train_sampler, test_sampler
+
+
+def _create_torch_dataloader(train_dir, val_dir, args):
+    dataset, dataset_test, train_sampler, test_sampler = load_data(train_dir, val_dir, args)
+    data_loader = torch.utils.data.DataLoader(
+        dataset, batch_size=args.batch_size,
+        sampler=train_sampler, num_workers=args.workers, pin_memory=True)
+
+    data_loader_test = torch.utils.data.DataLoader(
+        dataset_test, batch_size=args.batch_size,
+        sampler=test_sampler, num_workers=args.workers, pin_memory=True)
+
+    return data_loader, data_loader_test
+
+
+def _create_dali_dataloader(train_dir, val_dir, args):
+    from .dali_classification import get_imagenet_iter_dali
+    device = torch.cuda.current_device()
+    _, crop_size = get_input_size(args.model)
+    data_loader = get_imagenet_iter_dali('train', train_dir, args.batch_size,
+                                         num_threads=args.workers,
+                                         device_id=device,
+                                         size=crop_size)
+    data_loader_test = get_imagenet_iter_dali('val', train_dir, args.batch_size,
+                                              num_threads=args.workers,
+                                              device_id=device,
+                                              size=crop_size)
+
+    return data_loader, data_loader_test
+
+
+def create_dataloader(train_dir, val_dir, args):
+    print("Creating data loaders")
+    if args.dali:
+        train_dir = os.path.dirname(train_dir)
+        val_dir = os.path.dirname(val_dir)
+        return _create_dali_dataloader(train_dir, val_dir, args)
+    return _create_torch_dataloader(train_dir, val_dir, args)
diff --git a/cv/classification/cspdarknet53/pytorch/dataloader/dali_classification.py b/cv/classification/cspdarknet53/pytorch/dataloader/dali_classification.py
new file mode 100644
index 0000000000000000000000000000000000000000..2918bef662408deb14f4ca3b500a64b65bdcc28d
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/dataloader/dali_classification.py
@@ -0,0 +1,129 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# All Rights Reserved.
+# Copyright (c) Facebook, Inc. and its affiliates. 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.
+
+
+
+import os
+
+import nvidia.dali.ops as ops
+import nvidia.dali.types as types
+from nvidia.dali.pipeline import Pipeline
+from nvidia.dali.plugin.pytorch import DALIClassificationIterator, DALIGenericIterator
+
+class HybridTrainPipe(Pipeline):
+    def __init__(self, batch_size, num_threads, device_id, data_dir, size):
+        super(HybridTrainPipe, self).__init__(batch_size, num_threads, device_id)
+        self.input = ops.FileReader(file_root=data_dir, random_shuffle=True)
+        self.decode = ops.ImageDecoder(device="cpu", output_type=types.RGB)
+        self.res = ops.RandomResizedCrop(device="gpu", size=size, random_area=[0.08, 1.25])
+        self.cmnp = ops.CropMirrorNormalize(device="gpu",
+                                            output_dtype=types.FLOAT,
+                                            output_layout=types.NCHW,
+                                            image_type=types.RGB,
+                                            mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
+                                            std=[0.229 * 255, 0.224 * 255, 0.225 * 255])
+
+    def define_graph(self):
+        self.jpegs, self.labels = self.input(name="Reader")
+
+        images = self.decode(self.jpegs)
+        images = self.res(images.gpu())
+        output = self.cmnp(images)
+        return [output, self.labels]
+
+
+class HybridValPipe(Pipeline):
+    def __init__(self, batch_size, num_threads, device_id, data_dir, size):
+        super(HybridValPipe, self).__init__(batch_size, num_threads, device_id)
+        self.input = ops.FileReader(file_root=data_dir, random_shuffle=False)
+        self.decode = ops.ImageDecoder(device="cpu", output_type=types.RGB)
+        self.res = ops.Resize(device="gpu", resize_x=size, resize_y=size)
+        self.cmnp = ops.CropMirrorNormalize(device="gpu",
+                                            output_dtype=types.FLOAT,
+                                            output_layout=types.NCHW,
+                                            crop=(size, size),
+                                            image_type=types.RGB,
+                                            mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
+                                            std=[0.229 * 255, 0.224 * 255, 0.225 * 255])
+
+    def define_graph(self):
+        self.jpegs, self.labels = self.input(name="Reader")
+
+        images = self.decode(self.jpegs)
+        images = self.res(images.gpu())
+        output = self.cmnp(images)
+        return [output, self.labels]
+
+
+def get_imagenet_iter_dali(type, image_dir, batch_size, num_threads, device_id, size):
+    if type == 'train':
+        pip_train = HybridTrainPipe(batch_size=batch_size, num_threads=num_threads, device_id=device_id,
+                                    data_dir = os.path.join(image_dir, "train"),
+                                    size=size)
+        pip_train.build()
+        dali_iter_train = DALIClassificationIterator(pip_train, size=pip_train.epoch_size("Reader"))
+        return dali_iter_train
+    elif type == 'val':
+        pip_val = HybridValPipe(batch_size=batch_size, num_threads=num_threads, device_id=device_id,
+                                    data_dir = os.path.join(image_dir, "val"),
+                                    size=size)
+        pip_val.build()
+        dali_iter_val = DALIClassificationIterator(pip_val, size=pip_val.epoch_size("Reader"))
+        return dali_iter_val
+
+
+def main(arguments):
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data_dir', help='directory to save data to', type=str, default='classification data')
+    args = parser.parse_args(arguments)
+
+    train_loader = get_imagenet_iter_dali(type='train', image_dir=args.data_dir,
+                                          batch_size=256,
+                                          num_threads=4, size=224, device_id=3)
+
+    val_loader = get_imagenet_iter_dali(type="val", image_dir=args.data_dir,
+                                          batch_size=256,
+                                          num_threads=4, size=224, device_id=3)
+
+    print('start dali train dataloader.')
+    start = time.time()
+    for epoch in range(20):
+        for i, data in enumerate(train_loader):
+            images = data[0]["data"].cuda(non_blocking=True)
+            labels = data[0]["label"].squeeze().long().cuda(non_blocking=True)
+
+        # WARN: Very important
+        train_loader.reset()
+        print("Epoch", epoch)
+    print('dali iterate time: %fs' % (time.time() - start))
+    print('end dali train dataloader.')
+
+
+    print('start dali val dataloader.')
+    start = time.time()
+    for i, data in enumerate(val_loader):
+        images = data[0]["data"].cuda(non_blocking=True)
+        print(images.shape)
+        labels = data[0]["label"].squeeze().long().cuda(non_blocking=True)
+        print(labels.shape)
+    print('dali iterate time: %fs' % (time.time() - start))
+    print('end dali val dataloader.')
+
+
+if __name__ == '__main__':
+    import os, time, sys
+    import argparse
+    sys.exit(main(sys.argv[1:]))
diff --git a/cv/classification/cspdarknet53/pytorch/dataloader/utils/__init__.py b/cv/classification/cspdarknet53/pytorch/dataloader/utils/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..66b217b8507dfb048e3a366fc0fe6374f24eedc4
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/dataloader/utils/__init__.py
@@ -0,0 +1,14 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# 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.
\ No newline at end of file
diff --git a/cv/classification/cspdarknet53/pytorch/dataloader/utils/presets_classification.py b/cv/classification/cspdarknet53/pytorch/dataloader/utils/presets_classification.py
new file mode 100644
index 0000000000000000000000000000000000000000..59688a9595b8c1d2d8b2a9f23c228da63b047a06
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/dataloader/utils/presets_classification.py
@@ -0,0 +1,54 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# All Rights Reserved.
+# Copyright (c) Facebook, Inc. and its affiliates. 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.
+
+
+from torchvision.transforms import autoaugment, transforms
+
+
+class ClassificationPresetTrain:
+    def __init__(self, crop_size, mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225), hflip_prob=0.5,
+                 auto_augment_policy=None, random_erase_prob=0.0):
+        trans = [transforms.RandomResizedCrop(crop_size)]
+        if hflip_prob > 0:
+            trans.append(transforms.RandomHorizontalFlip(hflip_prob))
+        if auto_augment_policy is not None:
+            aa_policy = autoaugment.AutoAugmentPolicy(auto_augment_policy)
+            trans.append(autoaugment.AutoAugment(policy=aa_policy))
+        trans.extend([
+            transforms.ToTensor(),
+            transforms.Normalize(mean=mean, std=std),
+        ])
+        if random_erase_prob > 0:
+            trans.append(transforms.RandomErasing(p=random_erase_prob))
+
+        self.transforms = transforms.Compose(trans)
+
+    def __call__(self, img):
+        return self.transforms(img)
+
+
+class ClassificationPresetEval:
+    def __init__(self, crop_size, resize_size=256, mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)):
+
+        self.transforms = transforms.Compose([
+            transforms.Resize(resize_size),
+            transforms.CenterCrop(crop_size),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=mean, std=std),
+        ])
+
+    def __call__(self, img):
+        return self.transforms(img)
diff --git a/cv/classification/cspdarknet53/pytorch/model/csdarknet53.py b/cv/classification/cspdarknet53/pytorch/model/csdarknet53.py
new file mode 100644
index 0000000000000000000000000000000000000000..74bd36cc1bf7cf5f83ccc45af333c398e6eb8e6c
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/model/csdarknet53.py
@@ -0,0 +1,67 @@
+import torch.nn as nn
+from .cslayers import *
+import torch.backends.cudnn as cudnn
+from torchsummary import summary
+
+__all__ = ['CsDarkNet53']
+
+class CsDarkNet53(nn.Module):
+    def __init__(self, num_classes):
+        super(CsDarkNet53, self).__init__()
+
+        input_channels = 32
+
+        # Network
+        self.stage1 = Conv2dBatchLeaky(3, input_channels, 3, 1, activation='mish')
+        self.stage2 = Stage2(input_channels)
+        self.stage3 = Stage3(4*input_channels)
+        self.stage4 = Stage(4*input_channels, 8)
+        self.stage5 = Stage(8*input_channels, 8)
+        self.stage6 = Stage(16*input_channels, 4)
+
+        self.conv = Conv2dBatchLeaky(32*input_channels, 32*input_channels, 1, 1, activation='mish')
+        self.avgpool = nn.AdaptiveAvgPool2d((1,1))
+        self.fc = nn.Linear(1024, num_classes)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        stage1 = self.stage1(x)
+        stage2 = self.stage2(stage1)
+        stage3 = self.stage3(stage2)
+        stage4 = self.stage4(stage3)
+        stage5 = self.stage5(stage4)
+        stage6 = self.stage6(stage5)
+
+        conv = self.conv(stage6)
+        x = self.avgpool(conv)
+        x = x.view(-1, 1024)
+        x = self.fc(x)
+
+        return x
+
+if __name__ == "__main__":
+    use_cuda = torch.cuda.is_available()
+    if use_cuda:
+        device = torch.device("cuda")
+        cudnn.benchmark = True
+    else:
+        device = torch.device("cpu")
+
+    darknet = CsDarkNet53(num_classes=10)
+    darknet = darknet.cuda()
+    with torch.no_grad():
+        darknet.eval()
+        data = torch.rand(1, 3, 256, 256)
+        data = data.cuda()
+        try:
+            #print(darknet)
+            summary(darknet,(3,256,256))
+            print(darknet(data))
+        except Exception as e:
+            print(e)
diff --git a/cv/classification/cspdarknet53/pytorch/model/cslayers.py b/cv/classification/cspdarknet53/pytorch/model/cslayers.py
new file mode 100644
index 0000000000000000000000000000000000000000..0f5f7c78a05f6a9511b0e4beb40518b16941d8c9
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/model/cslayers.py
@@ -0,0 +1,176 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# mish(x) = x * tanh(log(1 + e^x))
+class Mish(nn.Module):
+    def __init__(self):
+        super(Mish, self).__init__()
+
+    def forward(self, x):
+        return x * torch.tanh(F.softplus(x))
+
+class Conv2dBatchLeaky(nn.Module):
+    """
+    This convenience layer groups a 2D convolution, a batchnorm and a leaky ReLU.
+    """
+    def __init__(self, in_channels, out_channels, kernel_size, stride, activation='leaky', leaky_slope=0.1):
+        super(Conv2dBatchLeaky, self).__init__()
+
+        # Parameters
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.stride = stride
+        if isinstance(kernel_size, (list, tuple)):
+            self.padding = [int(k/2) for k in kernel_size]
+        else:
+            self.padding = int(kernel_size/2)
+        self.leaky_slope = leaky_slope
+        # self.mish = Mish()
+
+        # Layer
+        if activation == "leaky":
+            self.layers = nn.Sequential(
+                nn.Conv2d(self.in_channels, self.out_channels, self.kernel_size, self.stride, self.padding, bias=False),
+                nn.BatchNorm2d(self.out_channels),
+                nn.LeakyReLU(self.leaky_slope, inplace=True)
+            )
+        elif activation == "mish":
+            self.layers = nn.Sequential(
+                nn.Conv2d(self.in_channels, self.out_channels, self.kernel_size, self.stride, self.padding, bias=False),
+                nn.BatchNorm2d(self.out_channels),
+                Mish()
+            )
+        elif activation == "linear":
+            self.layers = nn.Sequential(
+                nn.Conv2d(self.in_channels, self.out_channels, self.kernel_size, self.stride, self.padding, bias=False)
+            )
+
+    def __repr__(self):
+        s = '{name} ({in_channels}, {out_channels}, kernel_size={kernel_size}, stride={stride}, padding={padding}, negative_slope={leaky_slope})'
+        return s.format(name=self.__class__.__name__, **self.__dict__)
+
+    def forward(self, x):
+        x = self.layers(x)
+        return x
+
+class SmallBlock(nn.Module):
+
+    def __init__(self, nchannels):
+        super().__init__()
+        self.features = nn.Sequential(
+            Conv2dBatchLeaky(nchannels, nchannels, 1, 1, activation='mish'),
+            Conv2dBatchLeaky(nchannels, nchannels, 3, 1, activation='mish')
+        )
+        # conv_shortcut
+        '''
+        参考 https://github.com/bubbliiiing/yolov4-pytorch
+        shortcut后不接任何conv
+        '''
+        # self.active_linear = Conv2dBatchLeaky(nchannels, nchannels, 1, 1, activation='linear')
+        # self.conv_shortcut = Conv2dBatchLeaky(nchannels, nchannels, 1, 1, activation='mish')
+
+
+    def forward(self, data):
+        short_cut = data + self.features(data)
+        # active_linear = self.conv_shortcut(short_cut)
+
+        return short_cut
+
+# Stage1  conv [256,256,3]->[256,256,32]
+
+class Stage2(nn.Module):
+
+    def __init__(self, nchannels):
+        super().__init__()
+        # stage2 32
+        self.conv1 = Conv2dBatchLeaky(nchannels, 2*nchannels, 3, 2, activation='mish')
+        self.split0 = Conv2dBatchLeaky(2*nchannels, 2*nchannels, 1, 1, activation='mish')
+        self.split1 = Conv2dBatchLeaky(2*nchannels, 2*nchannels, 1, 1, activation='mish')
+
+        self.conv2 = Conv2dBatchLeaky(2*nchannels, nchannels, 1, 1, activation='mish')
+        self.conv3 = Conv2dBatchLeaky(nchannels, 2*nchannels, 3, 1, activation='mish')
+
+        self.conv4 = Conv2dBatchLeaky(2*nchannels, 2*nchannels, 1, 1, activation='mish')
+
+
+    def forward(self, data):
+        conv1 = self.conv1(data)
+        split0 = self.split0(conv1)
+        split1 = self.split1(conv1)
+        conv2 = self.conv2(split1)
+        conv3 = self.conv3(conv2)
+
+        shortcut = split1 + conv3
+        conv4 = self.conv4(shortcut)
+
+        route = torch.cat([split0, conv4], dim=1)
+        return route
+
+class Stage3(nn.Module):
+    def __init__(self, nchannels):
+        super().__init__()
+        # stage3 128
+        self.conv1 = Conv2dBatchLeaky(nchannels, int(nchannels/2), 1, 1, activation='mish')
+        self.conv2 = Conv2dBatchLeaky(int(nchannels/2), nchannels, 3, 2, activation='mish')
+
+        self.split0 = Conv2dBatchLeaky(nchannels, int(nchannels/2), 1, 1, activation='mish')
+        self.split1 = Conv2dBatchLeaky(nchannels, int(nchannels/2), 1, 1, activation='mish')
+
+        self.block1 = SmallBlock(int(nchannels/2))
+        self.block2 = SmallBlock(int(nchannels/2))
+
+        self.conv3 = Conv2dBatchLeaky(int(nchannels/2), int(nchannels/2), 1, 1, activation='mish')
+
+    def forward(self, data):
+        conv1 = self.conv1(data)
+        conv2 = self.conv2(conv1)
+
+        split0 = self.split0(conv2)
+        split1 = self.split1(conv2)
+
+        block1 = self.block1(split1)
+        block2 = self.block2(block1)
+
+        conv3 = self.conv3(block2)
+
+        route = torch.cat([split0, conv3], dim=1)
+
+        return route
+
+# Stage4 Stage5 Stage6
+class Stage(nn.Module):
+    def __init__(self, nchannels, nblocks):
+        super().__init__()
+        # stage4 : 128
+        # stage5 : 256
+        # stage6 : 512
+        self.conv1 = Conv2dBatchLeaky(nchannels, nchannels, 1, 1, activation='mish')
+        self.conv2 = Conv2dBatchLeaky(nchannels, 2*nchannels, 3, 2, activation='mish')
+        self.split0 = Conv2dBatchLeaky(2*nchannels, nchannels, 1, 1, activation='mish')
+        self.split1 = Conv2dBatchLeaky(2*nchannels, nchannels, 1, 1, activation='mish')
+        blocks = []
+        for i in range(nblocks):
+            blocks.append(SmallBlock(nchannels))
+        self.blocks = nn.Sequential(*blocks)
+        self.conv4 = Conv2dBatchLeaky(nchannels, nchannels, 1, 1, activation='mish')
+
+    def forward(self,data):
+        conv1 = self.conv1(data)
+        conv2 = self.conv2(conv1)
+
+        split0 = self.split0(conv2)
+        split1 = self.split1(conv2)
+        blocks = self.blocks(split1)
+        conv4 = self.conv4(blocks)
+        route = torch.cat([split0, conv4], dim=1)
+
+        return route
+
+
+
+
+
+
+
diff --git a/cv/classification/cspdarknet53/pytorch/train.py b/cv/classification/cspdarknet53/pytorch/train.py
new file mode 100644
index 0000000000000000000000000000000000000000..9afa8dc9d1b45673d223f0b195e4bf36dce18615
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/train.py
@@ -0,0 +1,318 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# All Rights Reserved.
+# Copyright (c) Facebook, Inc. and its affiliates. 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.
+
+import datetime
+import os
+
+import time
+
+import torch
+import torch.utils.data
+
+try:
+    from torch.cuda.amp import autocast, GradScaler
+    scaler = GradScaler()
+except:
+    autocast = None
+    scaler = None
+
+
+from torch import nn
+import torch.distributed as dist
+import torchvision
+
+from utils_ import (MetricLogger, SmoothedValue, accuracy, mkdir,\
+                    init_distributed_mode, manual_seed,\
+                    is_main_process, save_on_master, get_world_size)
+
+from dataloader.classification import get_datasets, create_dataloader
+
+from model.csdarknet53 import CsDarkNet53
+
+def compute_loss(model, image, target, criterion):
+    output = model(image)
+    if not isinstance(output, (tuple, list)):
+        output = [output]
+    losses = []
+    for out in output:
+        losses.append(criterion(out, target))
+    loss = sum(losses)
+    return loss, output[0]
+
+
+def train_one_epoch(model, criterion, optimizer, data_loader, device, epoch, print_freq, amp=False, use_dali=False):
+    model.train()
+    metric_logger = MetricLogger(delimiter="  ")
+    metric_logger.add_meter('lr', SmoothedValue(window_size=1, fmt='{value}'))
+    metric_logger.add_meter('img/s', SmoothedValue(window_size=10, fmt='{value}'))
+
+    header = 'Epoch: [{}]'.format(epoch)
+    all_fps = []
+    for data in metric_logger.log_every(data_loader, print_freq, header):
+        if use_dali:
+            image, target = data[0]["data"], data[0]["label"][:, 0].long()
+        else:
+            image, target = data
+        start_time = time.time()
+        image, target = image.to(device), target.to(device)
+        if autocast is None or not amp:
+            loss, output = compute_loss(model, image, target, criterion)
+        else:
+            with autocast():
+                loss, output = compute_loss(model, image, target, criterion)
+
+        optimizer.zero_grad()
+        if scaler is not None and amp:
+            scaler.scale(loss).backward()
+            scaler.step(optimizer)
+            scaler.update()
+        else:
+            loss.backward()
+            optimizer.step()
+
+        torch.cuda.synchronize()
+        end_time = time.time()
+
+        acc1, acc5 = accuracy(output, target, topk=(1, 5))
+        batch_size = image.shape[0]
+        metric_logger.update(loss=loss.item(), lr=optimizer.param_groups[0]["lr"])
+        metric_logger.meters['acc1'].update(acc1.item(), n=batch_size)
+        metric_logger.meters['acc5'].update(acc5.item(), n=batch_size)
+        fps = batch_size / (end_time - start_time) * get_world_size()
+        metric_logger.meters['img/s'].update(fps)
+        all_fps.append(fps)
+
+    print(header, 'Avg img/s:', sum(all_fps) / len(all_fps))
+
+
+def evaluate(model, criterion, data_loader, device, print_freq=100, use_dali=False):
+    model.eval()
+    metric_logger = MetricLogger(delimiter="  ")
+    header = 'Test:'
+    with torch.no_grad():
+        for data in metric_logger.log_every(data_loader, print_freq, header):
+            if use_dali:
+                image, target = data[0]["data"], data[0]["label"][:, 0].long()
+            else:
+                image, target = data
+            image = image.to(device, non_blocking=True)
+            target = target.to(device, non_blocking=True)
+            output = model(image)
+            loss = criterion(output, target)
+
+            acc1, acc5 = accuracy(output, target, topk=(1, 5))
+            # FIXME need to take into account that the datasets
+            # could have been padded in distributed setup
+            batch_size = image.shape[0]
+            metric_logger.update(loss=loss.item())
+            metric_logger.meters['acc1'].update(acc1.item(), n=batch_size)
+            metric_logger.meters['acc5'].update(acc5.item(), n=batch_size)
+    # gather the stats from all processes
+    metric_logger.synchronize_between_processes()
+
+    print(' * Acc@1 {top1.global_avg:.3f} Acc@5 {top5.global_avg:.3f}'
+          .format(top1=metric_logger.acc1, top5=metric_logger.acc5))
+    return metric_logger.acc1.global_avg
+
+
+def _get_cache_path(filepath):
+    import hashlib
+    h = hashlib.sha1(filepath.encode()).hexdigest()
+    cache_path = os.path.join("~", ".torch", "vision", "datasets", "imagefolder", h[:10] + ".pt")
+    cache_path = os.path.expanduser(cache_path)
+    return cache_path
+
+
+def main(args):
+    if args.output_dir:
+        mkdir(args.output_dir)
+
+    init_distributed_mode(args)
+    print(args)
+
+    device = torch.device(args.device)
+
+    manual_seed(args.seed, deterministic=False)
+    # torch.backends.cudnn.benchmark = True
+
+    # WARN:
+    if dist.is_initialized():
+        num_gpu = dist.get_world_size()
+    else:
+        num_gpu = 1
+
+    global_batch_size = num_gpu * args.batch_size
+
+    train_dir = os.path.join(args.data_path, 'train')
+    val_dir = os.path.join(args.data_path, 'val')
+
+    num_classes = len(os.listdir(train_dir))
+    if 0 < num_classes < 13:
+        if global_batch_size > 512:
+            if is_main_process():
+                print("WARN: Updating global batch size to 512, avoid non-convergence when training small dataset.")
+            args.batch_size = 512 // num_gpu
+
+    if args.pretrained:
+        num_classes = 1000
+
+    data_loader, data_loader_test = create_dataloader(train_dir, val_dir, args)
+
+    print("Creating model")
+    model = CsDarkNet53(num_classes=num_classes)
+    print(model)
+    model.to(device)
+    if args.distributed and args.sync_bn:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
+
+    criterion = nn.CrossEntropyLoss()
+
+    opt_name = args.opt.lower()
+    if opt_name == 'sgd':
+        optimizer = torch.optim.SGD(
+            model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)
+    elif opt_name == 'rmsprop':
+        optimizer = torch.optim.RMSprop(model.parameters(), lr=args.lr, momentum=args.momentum,
+                                        weight_decay=args.weight_decay, eps=0.0316, alpha=0.9)
+    else:
+        raise RuntimeError("Invalid optimizer {}. Only SGD and RMSprop are supported.".format(args.opt))
+
+    lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.lr_step_size, gamma=args.lr_gamma)
+
+    model_without_ddp = model
+    if args.distributed:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
+        model_without_ddp = model.module
+
+    if args.resume:
+        checkpoint = torch.load(args.resume, map_location='cpu')
+        model_without_ddp.load_state_dict(checkpoint['model'])
+        optimizer.load_state_dict(checkpoint['optimizer'])
+        lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
+        args.start_epoch = checkpoint['epoch'] + 1
+
+    if args.test_only:
+        evaluate(model, criterion, data_loader_test, device=device)
+        return
+
+    print("Start training")
+    start_time = time.time()
+    for epoch in range(args.start_epoch, args.epochs):
+        epoch_start_time = time.time()
+        if args.distributed and not args.dali:
+            data_loader.sampler.set_epoch(epoch)
+        train_one_epoch(model, criterion, optimizer, data_loader, device, epoch, args.print_freq, args.amp, use_dali=args.dali)
+        lr_scheduler.step()
+        evaluate(model, criterion, data_loader_test, device=device, use_dali=args.dali)
+        if args.output_dir:
+            checkpoint = {
+                'model': model_without_ddp.state_dict(),
+                'optimizer': optimizer.state_dict(),
+                'lr_scheduler': lr_scheduler.state_dict(),
+                'epoch': epoch,
+                'args': args}
+            #save_on_master(
+            #    checkpoint,
+            #    os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))
+            save_on_master(
+                checkpoint,
+                os.path.join(args.output_dir, 'checkpoint.pth'))
+            epoch_total_time = time.time() - epoch_start_time
+            epoch_total_time_str = str(datetime.timedelta(seconds=int(epoch_total_time)))
+            print('epoch time {}'.format(epoch_total_time_str))
+
+        if args.dali:
+            data_loader.reset()
+            data_loader_test.reset()
+
+    total_time = time.time() - start_time
+    total_time_str = str(datetime.timedelta(seconds=int(total_time)))
+    print('Training time {}'.format(total_time_str))
+
+
+def get_args_parser(add_help=True):
+    import argparse
+    parser = argparse.ArgumentParser(description='PyTorch Classification Training', add_help=add_help)
+
+    parser.add_argument('--data-path', default='/datasets01/imagenet_full_size/061417/', help='dataset')
+    parser.add_argument('--model', default='', help='model')
+    parser.add_argument('--device', default='cuda', help='device')
+    parser.add_argument('-b', '--batch-size', default=32, type=int)
+    parser.add_argument('--epochs', default=90, type=int, metavar='N',
+                        help='number of total epochs to run')
+    parser.add_argument('-j', '--workers', default=4, type=int, metavar='N',
+                        help='number of data loading workers (default: 4)')
+    parser.add_argument('--opt', default='sgd', type=str, help='optimizer')
+    parser.add_argument('--lr', default=0.1, type=float, help='initial learning rate')
+    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('--lr-step-size', default=30, type=int, help='decrease lr every step-size epochs')
+    parser.add_argument('--lr-gamma', default=0.1, type=float, help='decrease lr by a factor of lr-gamma')
+    parser.add_argument('--print-freq', default=10, type=int, help='print frequency')
+    parser.add_argument('--output-dir', default='.', help='path where to save')
+    parser.add_argument('--resume', default='', help='resume from checkpoint')
+    parser.add_argument('--start-epoch', default=0, type=int, metavar='N',
+                        help='start epoch')
+    parser.add_argument(
+        "--cache-dataset",
+        dest="cache_dataset",
+        help="Cache the datasets for quicker initialization. It also serializes the transforms",
+        action="store_true",
+    )
+    parser.add_argument(
+        "--sync-bn",
+        dest="sync_bn",
+        help="Use sync batch norm",
+        action="store_true",
+    )
+    parser.add_argument(
+        "--test-only",
+        dest="test_only",
+        help="Only test the model",
+        action="store_true",
+    )
+    parser.add_argument(
+        "--pretrained",
+        dest="pretrained",
+        help="Use pre-trained models from the modelzoo",
+        action="store_true",
+    )
+    parser.add_argument('--auto-augment', default=None, help='auto augment policy (default: None)')
+    parser.add_argument('--random-erase', default=0.0, type=float, help='random erasing probability (default: 0.0)')
+    parser.add_argument(
+        "--dali",
+        help="Use dali as dataloader",
+        default=False,
+        action="store_true",
+    )
+
+    # distributed training parameters
+    parser.add_argument('--local_rank', default=-1, type=int,
+                        help='Local rank')
+    parser.add_argument('--world-size', default=1, type=int,
+                        help='number of distributed processes')
+    parser.add_argument('--dist-url', default='env://', help='url used to set up distributed training')
+    parser.add_argument('--amp', action='store_true', help='Automatic Mixed Precision training')
+    parser.add_argument('--seed', default=42, type=int, help='Random seed')
+    return parser
+
+
+if __name__ == "__main__":
+    args = get_args_parser().parse_args()
+    main(args)
diff --git a/cv/classification/cspdarknet53/pytorch/utils_.py b/cv/classification/cspdarknet53/pytorch/utils_.py
new file mode 100644
index 0000000000000000000000000000000000000000..2a25edf2a95b3c59987c640e71e91fd15cbecc89
--- /dev/null
+++ b/cv/classification/cspdarknet53/pytorch/utils_.py
@@ -0,0 +1,169 @@
+# Copyright (c) 2022, Shanghai Iluvatar CoreX Semiconductor Co., Ltd.
+# All Rights Reserved.
+# Copyright (c) Facebook, Inc. and its affiliates. 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.
+
+
+from collections import defaultdict, deque, OrderedDict
+import copy
+import datetime
+import hashlib
+import time
+import torch
+import torch.distributed as dist
+
+import errno
+import os
+
+from common_utils import *
+
+
+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[None])
+
+        res = []
+        for k in topk:
+            correct_k = correct[:k].flatten().sum(dtype=torch.float32)
+            res.append(correct_k * (100.0 / batch_size))
+        return res
+
+
+def average_checkpoints(inputs):
+    """Loads checkpoints from inputs and returns a model with averaged weights. Original implementation taken from:
+    https://github.com/pytorch/fairseq/blob/a48f235636557b8d3bc4922a6fa90f3a0fa57955/scripts/average_checkpoints.py#L16
+
+    Args:
+      inputs (List[str]): An iterable of string paths of checkpoints to load from.
+    Returns:
+      A dict of string keys mapping to various values. The 'model' key
+      from the returned dict should correspond to an OrderedDict mapping
+      string parameter names to torch Tensors.
+    """
+    params_dict = OrderedDict()
+    params_keys = None
+    new_state = None
+    num_models = len(inputs)
+    for fpath in inputs:
+        with open(fpath, "rb") as f:
+            state = torch.load(
+                f,
+                map_location=(
+                    lambda s, _: torch.serialization.default_restore_location(s, "cpu")
+                ),
+            )
+        # Copies over the settings from the first checkpoint
+        if new_state is None:
+            new_state = state
+        model_params = state["model"]
+        model_params_keys = list(model_params.keys())
+        if params_keys is None:
+            params_keys = model_params_keys
+        elif params_keys != model_params_keys:
+            raise KeyError(
+                "For checkpoint {}, expected list of params: {}, "
+                "but found: {}".format(f, params_keys, model_params_keys)
+            )
+        for k in params_keys:
+            p = model_params[k]
+            if isinstance(p, torch.HalfTensor):
+                p = p.float()
+            if k not in params_dict:
+                params_dict[k] = p.clone()
+                # NOTE: clone() is needed in case of p is a shared parameter
+            else:
+                params_dict[k] += p
+    averaged_params = OrderedDict()
+    for k, v in params_dict.items():
+        averaged_params[k] = v
+        if averaged_params[k].is_floating_point():
+            averaged_params[k].div_(num_models)
+        else:
+            averaged_params[k] //= num_models
+    new_state["model"] = averaged_params
+    return new_state
+
+
+def store_model_weights(model, checkpoint_path, checkpoint_key='model', strict=True):
+    """
+    This method can be used to prepare weights files for new models. It receives as
+    input a model architecture and a checkpoint from the training script and produces
+    a file with the weights ready for release.
+
+    Examples:
+        from torchvision import models as M
+
+        # Classification
+        model = M.mobilenet_v3_large(pretrained=False)
+        print(store_model_weights(model, './class.pth'))
+
+        # Quantized Classification
+        model = M.quantization.mobilenet_v3_large(pretrained=False, quantize=False)
+        model.fuse_model()
+        model.qconfig = torch.quantization.get_default_qat_qconfig('qnnpack')
+        _ = torch.quantization.prepare_qat(model, inplace=True)
+        print(store_model_weights(model, './qat.pth'))
+
+        # Object Detection
+        model = M.detection.fasterrcnn_mobilenet_v3_large_fpn(pretrained=False, pretrained_backbone=False)
+        print(store_model_weights(model, './obj.pth'))
+
+        # Segmentation
+        model = M.segmentation.deeplabv3_mobilenet_v3_large(pretrained=False, pretrained_backbone=False, aux_loss=True)
+        print(store_model_weights(model, './segm.pth', strict=False))
+
+    Args:
+        model (pytorch.nn.Module): The model on which the weights will be loaded for validation purposes.
+        checkpoint_path (str): The path of the checkpoint we will load.
+        checkpoint_key (str, optional): The key of the checkpoint where the model weights are stored.
+            Default: "model".
+        strict (bool): whether to strictly enforce that the keys
+            in :attr:`state_dict` match the keys returned by this module's
+            :meth:`~torch.nn.Module.state_dict` function. Default: ``True``
+
+    Returns:
+        output_path (str): The location where the weights are saved.
+    """
+    # Store the new model next to the checkpoint_path
+    checkpoint_path = os.path.abspath(checkpoint_path)
+    output_dir = os.path.dirname(checkpoint_path)
+
+    # Deep copy to avoid side-effects on the model object.
+    model = copy.deepcopy(model)
+    checkpoint = torch.load(checkpoint_path, map_location='cpu')
+
+    # Load the weights to the model to validate that everything works
+    # and remove unnecessary weights (such as auxiliaries, etc)
+    model.load_state_dict(checkpoint[checkpoint_key], strict=strict)
+
+    tmp_path = os.path.join(output_dir, str(model.__hash__()))
+    torch.save(model.state_dict(), tmp_path)
+
+    sha256_hash = hashlib.sha256()
+    with open(tmp_path, "rb") as f:
+        # Read and update hash string value in blocks of 4K
+        for byte_block in iter(lambda: f.read(4096), b""):
+            sha256_hash.update(byte_block)
+        hh = sha256_hash.hexdigest()
+
+    output_path = os.path.join(output_dir, "weights-" + str(hh[:8]) + ".pth")
+    os.replace(tmp_path, output_path)
+
+    return output_path