Using hub

Hub¶

Hub provides same interface for several backends. Let see how it works.

Create Ultralytics Hub¶

In [ ]:

from waffle_hub.hub.adapter.ultralytics import UltralyticsHub
from waffle_hub import TaskType

from waffle_hub.hub.adapter.ultralytics import UltralyticsHub from waffle_hub import TaskType

See what tasks, models are available in UltralyticsHub.

In [2]:

UltralyticsHub.MODEL_TYPES

UltralyticsHub.MODEL_TYPES

Out[2]:

{'object_detection': {'yolov8': ['n', 's', 'm', 'l', 'x']},
 'classification': {'yolov8': ['n', 's', 'm', 'l', 'x']},
 'instance_segmentation': {'yolov8': ['n', 's', 'm', 'l', 'x']}}

By calling UltralyticsHub.new() method you can simply create an Ultralytics instance.

In [3]:

ultralytics_hub = UltralyticsHub.new(
    name="ultralytics_mnist_detection",
    task=TaskType.OBJECT_DETECTION,
    model_type="yolov8",
    model_size="n",
    categories=["1", "2"]
)

ultralytics_hub = UltralyticsHub.new( name="ultralytics_mnist_detection", task=TaskType.OBJECT_DETECTION, model_type="yolov8", model_size="n", categories=["1", "2"] )

Hub instance provides several useful properties and methods. See Hub Documentation for more details.

In [4]:

ultralytics_hub.categories

ultralytics_hub.categories

Out[4]:

[{'supercategory': 'object', 'name': '1'},
 {'supercategory': 'object', 'name': '2'}]

Create Huggingface Hub¶

As you did in Ultralytics Hub, you can create Huggingface Hub instance with exactly same way.

In [5]:

from waffle_hub.hub.adapter.hugging_face import HuggingFaceHub
from waffle_hub import TaskType

from waffle_hub.hub.adapter.hugging_face import HuggingFaceHub from waffle_hub import TaskType

/home/lhj/anaconda3/envs/waffle/lib/python3.9/site-packages/waffle_hub/__init__.py:56: UserWarning: 
                transformers 4.28.1 has not been tested.
                We recommend you to use one of ['4.27.4']
                
  warnings.warn(

In [6]:

HuggingFaceHub.MODEL_TYPES

HuggingFaceHub.MODEL_TYPES

Out[6]:

{'object_detection': {'DETA': {'base': 'jozhang97/deta-resnet-50'},
  'DETR': {'base': 'facebook/detr-resnet-50',
   'large': 'facebook/detr-resnet-101'},
  'YOLOS': {'tiny': 'hustvl/yolos-tiny'}},
 'classification': {'ViT': {'tiny': 'WinKawaks/vit-tiny-patch16-224',
   'base': 'google/vit-base-patch16-224'}}}

In [7]:

huggingface_hub = HuggingFaceHub.new(
    name="huggingface_mnist_detection",
    task=TaskType.OBJECT_DETECTION,
    model_type="DETR",
    model_size="base",
    categories=["1", "2"]
)

huggingface_hub = HuggingFaceHub.new( name="huggingface_mnist_detection", task=TaskType.OBJECT_DETECTION, model_type="DETR", model_size="base", categories=["1", "2"] )

Create Tx Model Hub¶

Tx Model is a private deep learning framework of SNUAILAB. You can use it by asking our team!
You can also create Tx Model Hub instance with exactly same way.

In [8]:

from waffle_hub.hub.adapter.tx_model import TxModelHub
from waffle_hub import TaskType

from waffle_hub.hub.adapter.tx_model import TxModelHub from waffle_hub import TaskType

In [9]:

TxModelHub.MODEL_TYPES

TxModelHub.MODEL_TYPES

Out[9]:

{'object_detection': {'YOLOv5': ['s', 'm', 'l']},
 'classification': {'Classifier': ['s', 'm', 'l']}}

In [10]:

tx_model_hub = TxModelHub.new(
    name="tx_model_mnist_detection",
    task=TaskType.OBJECT_DETECTION,
    model_type="YOLOv5",
    model_size="s",
    categories=["1", "2"]
)

tx_model_hub = TxModelHub.new( name="tx_model_mnist_detection", task=TaskType.OBJECT_DETECTION, model_type="YOLOv5", model_size="s", categories=["1", "2"] )

Train¶

Load Dataset¶

In [11]:

from waffle_hub.dataset import Dataset
from waffle_hub import DataType

dataset = Dataset.load("mnist_det")

from waffle_hub.dataset import Dataset from waffle_hub import DataType dataset = Dataset.load("mnist_det")

Ultralytics¶

In [ ]:

dataset_dir = dataset.export(DataType.YOLO)
train_result = ultralytics_hub.train(
    dataset_dir,
    image_size=320,
    epochs=50,
    batch_size=16
)
train_result

dataset_dir = dataset.export(DataType.YOLO) train_result = ultralytics_hub.train( dataset_dir, image_size=320, epochs=50, batch_size=16 ) train_result

Huggingface¶

In [13]:

import torch
torch.use_deterministic_algorithms(False)

import torch torch.use_deterministic_algorithms(False)

In [ ]:

dataset_dir = dataset.export(DataType.HUGGINGFACE)
train_result = huggingface_hub.train(
    dataset_dir,
    image_size=320,
    epochs=50,
    batch_size=16
)
train_result

dataset_dir = dataset.export(DataType.HUGGINGFACE) train_result = huggingface_hub.train( dataset_dir, image_size=320, epochs=50, batch_size=16 ) train_result

Tx Model¶

In [ ]:

dataset_dir = dataset.export(DataType.TX_MODEL)
train_result = tx_model_hub.train(
    dataset_dir,
    image_size=320,
    epochs=50,
    batch_size=16,
    pretrained_model="base_models/detectors/small/model.pth"
)
train_result

dataset_dir = dataset.export(DataType.TX_MODEL) train_result = tx_model_hub.train( dataset_dir, image_size=320, epochs=50, batch_size=16, pretrained_model="base_models/detectors/small/model.pth" ) train_result

Evaluate¶

In [16]:

ultralytics_hub = UltralyticsHub.load("ultralytics_mnist_detection")
ultralytics_hub.evaluate("mnist_det", set_name="test")

ultralytics_hub = UltralyticsHub.load("ultralytics_mnist_detection") ultralytics_hub.evaluate("mnist_det", set_name="test")

100%|██████████| 3/3 [00:00<00:00, 25.25it/s]

Out[16]:

EvaluateResult(metrics=[{'tag': 'mAP', 'value': 0.8066571950912476}])

In [17]:

huggingface_hub = HuggingFaceHub.load("huggingface_mnist_detection")
huggingface_hub.evaluate("mnist_det", set_name="test")

huggingface_hub = HuggingFaceHub.load("huggingface_mnist_detection") huggingface_hub.evaluate("mnist_det", set_name="test")

100%|██████████| 3/3 [00:00<00:00, 21.73it/s]

Out[17]:

EvaluateResult(metrics=[{'tag': 'mAP', 'value': 0.18526874482631683}])

In [25]:

tx_model_hub = TxModelHub.load("tx_model_mnist_detection")
tx_model_hub.evaluate("mnist_det", set_name="test")

tx_model_hub = TxModelHub.load("tx_model_mnist_detection") tx_model_hub.evaluate("mnist_det", set_name="test")

100%|██████████| 3/3 [00:00<00:00, 25.43it/s]

Out[25]:

EvaluateResult(metrics=[{'tag': 'mAP', 'value': 0.1792079210281372}])

Inference¶

In [18]:

import PIL.Image

import PIL.Image

In [19]:

ultralytics_hub = UltralyticsHub.load("ultralytics_mnist_detection")
ultralytics_hub.inference("mnist/images", draw=True)
PIL.Image.open(ultralytics_hub.draw_dir / "1.png")

ultralytics_hub = UltralyticsHub.load("ultralytics_mnist_detection") ultralytics_hub.inference("mnist/images", draw=True) PIL.Image.open(ultralytics_hub.draw_dir / "1.png")

100%|██████████| 25/25 [00:00<00:00, 54.73it/s]

Out[19]:

In [20]:

huggingface_hub = HuggingFaceHub.load("huggingface_mnist_detection")
huggingface_hub.inference("mnist/images", draw=True)
PIL.Image.open(huggingface_hub.draw_dir / "1.png")

huggingface_hub = HuggingFaceHub.load("huggingface_mnist_detection") huggingface_hub.inference("mnist/images", draw=True) PIL.Image.open(huggingface_hub.draw_dir / "1.png")

100%|██████████| 25/25 [00:00<00:00, 42.81it/s]

Out[20]:

In [21]:

tx_model_hub = TxModelHub.load("tx_model_mnist_detection")
tx_model_hub.inference("mnist/images", draw=True)
PIL.Image.open(tx_model_hub.draw_dir / "1.png")

tx_model_hub = TxModelHub.load("tx_model_mnist_detection") tx_model_hub.inference("mnist/images", draw=True) PIL.Image.open(tx_model_hub.draw_dir / "1.png")

100%|██████████| 25/25 [00:00<00:00, 62.38it/s]

Out[21]:

Export to onnx¶

In [22]:

ultralytics_hub.export()

ultralytics_hub.export()

/home/lhj/anaconda3/envs/waffle/lib/python3.9/site-packages/ultralytics/nn/modules.py:474: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  elif self.dynamic or self.shape != shape:
/home/lhj/anaconda3/envs/waffle/lib/python3.9/site-packages/ultralytics/yolo/utils/tal.py:241: TracerWarning: Iterating over a tensor might cause the trace to be incorrect. Passing a tensor of different shape won't change the number of iterations executed (and might lead to errors or silently give incorrect results).
  for i, stride in enumerate(strides):
/home/lhj/anaconda3/envs/waffle/lib/python3.9/site-packages/torch/onnx/utils.py:617: UserWarning: ONNX Preprocess - Removing mutation from node aten::sub_ on block input: 'tensor.1'. This changes graph semantics. (Triggered internally at ../torch/csrc/jit/passes/onnx/remove_inplace_ops_for_onnx.cpp:335.)
  _C._jit_pass_onnx_remove_inplace_ops_for_onnx(graph, module)

Out[22]:

ExportResult(export_file=PosixPath('hubs/ultralytics_mnist_detection/weights/model.onnx'))

In [23]:

huggingface_hub.export()

huggingface_hub.export()

/home/lhj/anaconda3/envs/waffle/lib/python3.9/site-packages/transformers/models/detr/modeling_detr.py:575: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if attn_weights.size() != (batch_size * self.num_heads, target_len, source_len):
/home/lhj/anaconda3/envs/waffle/lib/python3.9/site-packages/transformers/models/detr/modeling_detr.py:582: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if attention_mask.size() != (batch_size, 1, target_len, source_len):
/home/lhj/anaconda3/envs/waffle/lib/python3.9/site-packages/transformers/models/detr/modeling_detr.py:606: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if attn_output.size() != (batch_size * self.num_heads, target_len, self.head_dim):
/home/lhj/anaconda3/envs/waffle/lib/python3.9/site-packages/transformers/models/detr/image_processing_detr.py:1567: TracerWarning: Iterating over a tensor might cause the trace to be incorrect. Passing a tensor of different shape won't change the number of iterations executed (and might lead to errors or silently give incorrect results).
  for s, l, b in zip(scores, labels, boxes):

Out[23]:

ExportResult(export_file=PosixPath('hubs/huggingface_mnist_detection/weights/model.onnx'))

In [26]:

tx_model_hub.export(device="cpu")

tx_model_hub.export(device="cpu")

/home/lhj/anaconda3/envs/waffle/lib/python3.9/site-packages/torch/onnx/utils.py:617: UserWarning: ONNX Preprocess - Removing mutation from node aten::sub_ on block input: 'tensor.1'. This changes graph semantics. (Triggered internally at ../torch/csrc/jit/passes/onnx/remove_inplace_ops_for_onnx.cpp:335.)
  _C._jit_pass_onnx_remove_inplace_ops_for_onnx(graph, module)

Out[26]:

ExportResult(export_file=PosixPath('hubs/tx_model_mnist_detection/weights/model.onnx'))

Benchmark¶

In [27]:

ultralytics_hub.benchmark(image_size=320, batch_size=16)

ultralytics_hub.benchmark(image_size=320, batch_size=16)

100%|██████████| 100/100 [00:01<00:00, 97.22it/s]

Out[27]:

{'inference_time': 1.0294418334960938,
 'fps': 1554.2403154204742,
 'image_size': [320, 320],
 'batch_size': 16,
 'precision': 'fp32',
 'device': 'cuda:0',
 'cpu_name': '13th Gen Intel(R) Core(TM) i9-13900KF',
 'gpu_name': 'NVIDIA GeForce RTX 4090'}

In [28]:

huggingface_hub.benchmark(image_size=320, batch_size=16)

huggingface_hub.benchmark(image_size=320, batch_size=16)

100%|██████████| 100/100 [00:02<00:00, 42.96it/s]

Out[28]:

{'inference_time': 2.328606128692627,
 'fps': 687.1063252325563,
 'image_size': [320, 320],
 'batch_size': 16,
 'precision': 'fp32',
 'device': 'cuda:0',
 'cpu_name': '13th Gen Intel(R) Core(TM) i9-13900KF',
 'gpu_name': 'NVIDIA GeForce RTX 4090'}

In [29]:

tx_model_hub.benchmark(image_size=320, batch_size=16)

tx_model_hub.benchmark(image_size=320, batch_size=16)

100%|██████████| 100/100 [00:00<00:00, 113.23it/s]

Out[29]:

{'inference_time': 0.8841898441314697,
 'fps': 1809.5661362992278,
 'image_size': [320, 320],
 'batch_size': 16,
 'precision': 'fp32',
 'device': 'cuda:0',
 'cpu_name': '13th Gen Intel(R) Core(TM) i9-13900KF',
 'gpu_name': 'NVIDIA GeForce RTX 4090'}

In [ ]: