'프로그램 사용'에 해당되는 글 2183건

  1. 2024.01.15 NMS, soft-NMS
  2. 2024.01.11 VGG-16 / VGG-19
  3. 2024.01.11 MobileNetV2 SSD FPN-Lite
  4. 2024.01.11 mobilenet v2 ssd
  5. 2024.01.11 gstreamer parse_launch
  6. 2024.01.10 ssd-mobilenetv2 on jupyter notebook 2
  7. 2024.01.10 텐서플로우 v1 을 v2로 마이그레이션은 실패 -_-
  8. 2024.01.10 골빈해커의 3분 딥러닝 github
  9. 2024.01.10 ReLU - Rectified Linear Unit
  10. 2024.01.10 softmax

NMS는 하나로 억제하기 때문에

겹칠 경우 하나의 객체를 인식하지 못하게 되므로 이를 개선한 것이 soft-NMS 라고 

 

[링크 : https://hongl.tistory.com/180]

[링크 : https://ctkim.tistory.com/entry/Non-maximum-Suppression-NMS]

 

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

VGG-16 / VGG-19  (0) 2024.01.11
MobileNetV2 SSD FPN-Lite  (0) 2024.01.11
mobilenet v2 ssd  (0) 2024.01.11
ssd-mobilenetv2 on jupyter notebook  (2) 2024.01.10
텐서플로우 v1 을 v2로 마이그레이션은 실패 -_-  (0) 2024.01.10
Posted by 구차니

Visual Geometry Group

VGG 뒤의 숫자는 CNN 레이어의 갯수

CNN(Convolutional Neural network) - 나선형의/복잡한 신경망으로 해석이 되나?

 

[링크 : https://wikidocs.net/164796]

 

탐지도 되긴 하나본데...

[링크 : https://github.com/zubairsamo/Object-Detection-With-Tensorflow-Using-VGG16]

 

keras에 있는 VGG16을 그냥 바로 써서 간단하게 되네..

게다가 save load도 되는데 왜 난 안될까.. ㅠㅠ

# lets import pre trained VGG16 Which is already Builtin for computer vision
from tensorflow.keras.applications import VGG16
from tensorflow.keras.layers import Input
     

# Imagenet is a competition every year held and VGG16 is winner of between  2013-14
# so here we just want limited layers so thats why we false included_top 
vgg=VGG16(weights='imagenet',include_top=False,input_tensor=Input(shape=(224,224,3)))


# lets save model 
model.save('detect_Planes.h5')     

from tensorflow.keras.models import load_model
model=load_model('/content/detect_Planes.h5')

[링크 : https://github.com/zubairsamo/Object-Detection-With-Tensorflow-Using-VGG16/blob/main/Object_Detection_Using_VGG16_With_Tensorflow.ipynb]

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

NMS, soft-NMS  (0) 2024.01.15
MobileNetV2 SSD FPN-Lite  (0) 2024.01.11
mobilenet v2 ssd  (0) 2024.01.11
ssd-mobilenetv2 on jupyter notebook  (2) 2024.01.10
텐서플로우 v1 을 v2로 마이그레이션은 실패 -_-  (0) 2024.01.10
Posted by 구차니

모델 zoo 에서 보다보면 먼가 주르르륵 붙는데

솔찍히 mobilenet v2가 무엇인지, ssd는 또 무엇인지 몰라서 헷갈려서 조사

근데.. FPN은 전에 검색해둔 기억이 있는데 가물가물하네..

 

base network (MobileNetV2)

- 신경망(neural network)로 구성되어 있으며 분류(classification)나 탐지(detection)에 사용이 가능함

- 네트워크의 마지막에 softmax 레이어가 있으면 분류로 작동

 

detection network (Single Shot Detector or SSD)

- SSD(Single ShotDetection)나 RPN(Regional Proposal Network / R-CNN)를 이용하여

- 이미지 내의 여러 물체를 감지하고, 지역을 제안(ROI)함.

- R-CNN : Regions with Convolutional Neural Networks

[링크 : https://kr.mathworks.com/help/vision/ug/getting-started-with-r-cnn-fast-r-cnn-and-faster-r-cnn.html]

 

feature extractor (FPN-Lite)

SSD의 경우 너무 가깝거나(즉 너무 크거나, 일부만 확대되어 보일 경우) 작은경우(멀거나, 원래 작거나) 탐지를 못하는 경우가 있어

피라미드 모양으로 쌓은 FPN(Feature Pyramid Network)를 통해 특징을 추출하여 다양한 규모의 물체를 감지

 

SSD 에서는 Pyramid Feature Hierachy 라는 방식을 이용하여, 서로 다른 스케일의 특징 맵을 이용하여 멀티 스케일 특징을 추출

FPN 모델은 Region Proposeal Network (RPN) 와 Fast R-CNN을 기반으로 한다.

[링크 : https://eehoeskrap.tistory.com/300]

 

근데 내용만 봐서는 SSD + mobilenet v2 + FPN이 조합이 가능한건지 모르겠다?

 

In the MobileNetV2 SSD FPN-Lite, we have a base network (MobileNetV2), a detection network (Single Shot Detector or SSD) and a feature extractor (FPN-Lite).

Base network:
MobileNet, like VGG-Net, LeNet, AlexNet, and all others, are based on neural networks. The base network provides high-level features for classification or detection. If you use a fully connected layer and a softmax layer at the end of these networks, you have a classification.


 
Example of a network composed of many convolutional layers. Filters are applied to each training image at different resolutions, and the output of each convolved image is used as input to the next layer (source Mathworks)
But you can remove the fully connected and the softmax layers, and replace it with detection networks, like SSD, Faster R-CNN, and others to perform object detection.

Detection network:
The most common detection networks are SSD (Single Shot Detection) and RPN (Regional Proposal Network).
When using SSD, we only need to take one single shot to detect multiple objects within the image. On the other hand, regional proposal networks (RPN) based approaches, such as R-CNN series, need two shots, one for generating region proposals, one for detecting the object of each proposal.
As a consequence, SSD is much faster compared with RPN-based approaches but often trades accuracy with real-time processing speed. They also tend to have issues in detecting objects that are too close or too small.

Feature Pyramid Network:
Detecting objects in different scales is challenging in particular for small objects. Feature Pyramid Network (FPN) is a feature extractor designed with feature pyramid concept to improve accuracy and speed.

[링크 : https://docs.edgeimpulse.com/docs/edge-impulse-studio/learning-blocks/object-detection/mobilenetv2-ssd-fpn]

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

NMS, soft-NMS  (0) 2024.01.15
VGG-16 / VGG-19  (0) 2024.01.11
mobilenet v2 ssd  (0) 2024.01.11
ssd-mobilenetv2 on jupyter notebook  (2) 2024.01.10
텐서플로우 v1 을 v2로 마이그레이션은 실패 -_-  (0) 2024.01.10
Posted by 구차니

이전에꺼는 save를 못해먹겠어서 (checkpoint를 pb로 변환하거나 해야 하는데 그것도 안되고...)

다른 것 찾는 중

 

[링크 : https://www.kaggle.com/code/suraj520/mobilenet-v2-ssd-scratch-without-tfod]

 

위에껀 아래와 같이 sequential 이라던가 이런게 없어서.. 또 저장안되는거 아닌가 걱정중..

def create_model():
  model = tf.keras.Sequential([
    keras.layers.Dense(512, activation='relu', input_shape=(784,)),
    keras.layers.Dropout(0.2),
    keras.layers.Dense(10)
  ])

  model.compile(optimizer='adam',
                loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
                metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])

  return model

# Create a basic model instance
model = create_model()

[링크 : https://www.tensorflow.org/tutorials/keras/save_and_load?hl=ko]

[링크 : https://www.tensorflow.org/hub/exporting_tf2_saved_model?hl=ko]

Posted by 구차니

nnstreamer python 예제를 보다보니, 아래와 같이 pipeline 문자열을 이용해서 구성하고

이름으로 싱크를 찾아 콜백을 연결해 주는걸 보니, 일일이 element 생성해서 연결할 필요가 없을 것 같아서 검색중

        gst_launch_cmdline += "tensor_sink name=tensor_sink t. ! "

        self.pipeline = Gst.parse_launch(gst_launch_cmdline)

        # bus and message callback
        bus = self.pipeline.get_bus()
        bus.add_signal_watch()
        bus.connect("message", self.on_bus_message)

        self.tensor_filter = self.pipeline.get_by_name("tensor_filter")
        self.wayland_sink = self.pipeline.get_by_name("img_tensor")

        # tensor sink signal : new data callback
        tensor_sink = self.pipeline.get_by_name("tensor_sink")
        tensor_sink.connect("new-data", self.new_data_cb)

    # @brief Callback for tensor sink signal.
    def new_data_cb(self, sink, buffer):
        """Callback for tensor sink signal.

        :param sink: tensor sink element
        :param buffer: buffer from element
        :return: None
        """

 

parse쪽은 c 로는 아래의 함수를 쓰면 될 것 같은데

gst_parse_launch 
GstElement *
gst_parse_launch (const gchar * pipeline_description,
                  GError ** error)
Create a new pipeline based on command line syntax. Please note that you might get a return value that is not NULL even though the error is set. In this case there was a recoverable parsing error and you can try to play the pipeline.

To create a sub-pipeline (bin) for embedding into an existing pipeline use gst_parse_bin_from_description.

Parameters:

pipeline_description – the command line describing the pipeline
error – the error message in case of an erroneous pipeline.
Returns ( [transfer: floating]) – a new element on success, NULL on failure. If more than one toplevel element is specified by the pipeline_description, all elements are put into a GstPipeline, which than is returned.

[링크 : https://gstreamer.freedesktop.org/documentation/gstreamer/gstparse.html?gi-language=c]

 

아래의 함수를 이용해서 찾으면 될...지도?

GObject
    ╰──GInitiallyUnowned
        ╰──GstObject
            ╰──GstElement
                ╰──GstBin
                    ╰──GstPipeline

gst_bin_get_by_name 
GstElement *
gst_bin_get_by_name (GstBin * bin,
                     const gchar * name)
Gets the element with the given name from a bin. This function recurses into child bins.

Parameters:

bin – a GstBin
name – the element name to search for
Returns ( [transfer: full][nullable]) – the GstElement with the given name

[링크 : https://gstreamer.freedesktop.org/documentation/gstreamer/gstbin.html?gi-language=c#gst_bin_get_by_name]

'프로그램 사용 > gstreamer' 카테고리의 다른 글

nnstreamer  (0) 2023.12.20
gst-device-monitor-1.0  (0) 2023.12.06
gstremaer videobox + videomixer  (0) 2023.04.10
gst-inspector.c  (0) 2023.04.06
gstreamer videomixer 반쪽 성공  (0) 2023.03.27
Posted by 구차니

버전 정보 추적

 

python 3.7

keras 2.8.0

numpy 1.21.6

[링크 : https://github.com/saunack/MobileNetv2-SSD/blob/master/model.ipynb]

 

설치한 버전들

$ pip install tensorflow==2.8.0
$ pip install numpy==1.21.6
$ pip install keras==2.8.0
$ pip install protobuf==3.19.0

 

------

keras 2.8을 써야 한다고 나오니 keras의 릴리즈 날짜로 추적

v2.8.0
 on Jan 7, 2022  d8fcb9d  zip  tar.gz  Notes

[링크 : https://github.com/keras-team/keras/tags?after=v2.9.0-rc1]

 

tensorflow 버전 추적

v2.8.0
 on Feb 1, 2022  3f878cf  zip  tar.gz  Notes

[링크 : https://github.com/tensorflow/tensorflow/tags?after=v2.7.2]

 

protobuf 3.19.0

numpy 1.24.4 (1.25 미만)

$ pip install numpy==1.34
Defaulting to user installation because normal site-packages is not writeable
ERROR: Could not find a version that satisfies the requirement numpy==1.34 (from versions: 1.3.0, 1.4.1, 1.5.0, 1.5.1, 1.6.0, 1.6.1, 1.6.2, 1.7.0, 1.7.1, 1.7.2, 1.8.0, 1.8.1, 1.8.2, 1.9.0, 1.9.1, 1.9.2, 1.9.3, 1.10.0.post2, 1.10.1, 1.10.2, 1.10.4, 1.11.0, 1.11.1, 1.11.2, 1.11.3, 1.12.0, 1.12.1, 1.13.0, 1.13.1, 1.13.3, 1.14.0, 1.14.1, 1.14.2, 1.14.3, 1.14.4, 1.14.5, 1.14.6, 1.15.0, 1.15.1, 1.15.2, 1.15.3, 1.15.4, 1.16.0, 1.16.1, 1.16.2, 1.16.3, 1.16.4, 1.16.5, 1.16.6, 1.17.0, 1.17.1, 1.17.2, 1.17.3, 1.17.4, 1.17.5, 1.18.0, 1.18.1, 1.18.2, 1.18.3, 1.18.4, 1.18.5, 1.19.0, 1.19.1, 1.19.2, 1.19.3, 1.19.4, 1.19.5, 1.20.0, 1.20.1, 1.20.2, 1.20.3, 1.21.0, 1.21.1, 1.21.2, 1.21.3, 1.21.4, 1.21.5, 1.21.6, 1.22.0, 1.22.1, 1.22.2, 1.22.3, 1.22.4, 1.23.0rc1, 1.23.0rc2, 1.23.0rc3, 1.23.0, 1.23.1, 1.23.2, 1.23.3, 1.23.4, 1.23.5, 1.24.0rc1, 1.24.0rc2, 1.24.0, 1.24.1, 1.24.2, 1.24.3, 1.24.4, 1.25.0rc1, 1.25.0, 1.25.1, 1.25.2, 1.26.0b1, 1.26.0rc1, 1.26.0, 1.26.1, 1.26.2, 1.26.3)
ERROR: No matching distribution found for numpy==1.34

------

 

으아아아아 tensorflow가 문제냐 keras가 문제냐 ㅠㅠ

import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()

tf_v2 끄고 하면 아래와 같이 나오고

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[27], line 1
----> 1 history = model.fit(train_dataset,
      2                     epochs=25,
      3                     validation_data = test_dataset,
      4                     validation_steps=1)

File ~/.local/lib/python3.10/site-packages/keras/engine/training_v1.py:773, in Model.fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)
    771 if kwargs:
    772   raise TypeError('Unrecognized keyword arguments: ' + str(kwargs))
--> 773 self._assert_compile_was_called()
    774 self._check_call_args('fit')
    776 func = self._select_training_loop(x)

File ~/.local/lib/python3.10/site-packages/keras/engine/training_v1.py:2788, in Model._assert_compile_was_called(self)
   2782 def _assert_compile_was_called(self):
   2783   # Checks whether `compile` has been called. If it has been called,
   2784   # then the optimizer is set. This is different from whether the
   2785   # model is compiled
   2786   # (i.e. whether the model is built and its inputs/outputs are set).
   2787   if not self._compile_was_called:
-> 2788     raise RuntimeError('You must compile your model before '
   2789                        'training/testing. '
   2790                        'Use `model.compile(optimizer, loss)`.')

RuntimeError: You must compile your model before training/testing. Use `model.compile(optimizer, loss)`.

 

tf_v2를 쓰게 하면

import tensorflow.compat.v1 as tf
#tf.disable_v2_behavior()

사용자 쪽 코드로 문제를 넘기는데 머가 문제일까..

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[28], line 1
----> 1 history = model.fit(train_dataset,
      2                     epochs=25,
      3                     validation_data = test_dataset,
      4                     validation_steps=1)

File ~/.local/lib/python3.10/site-packages/keras/engine/training_v1.py:777, in Model.fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, max_queue_size, workers, use_multiprocessing, **kwargs)
    774 self._check_call_args('fit')
    776 func = self._select_training_loop(x)
--> 777 return func.fit(
    778     self,
    779     x=x,
    780     y=y,
    781     batch_size=batch_size,
    782     epochs=epochs,
    783     verbose=verbose,
    784     callbacks=callbacks,
    785     validation_split=validation_split,
    786     validation_data=validation_data,
    787     shuffle=shuffle,
    788     class_weight=class_weight,
    789     sample_weight=sample_weight,
    790     initial_epoch=initial_epoch,
    791     steps_per_epoch=steps_per_epoch,
    792     validation_steps=validation_steps,
    793     validation_freq=validation_freq,
    794     max_queue_size=max_queue_size,
    795     workers=workers,
    796     use_multiprocessing=use_multiprocessing)

File ~/.local/lib/python3.10/site-packages/keras/engine/training_arrays_v1.py:616, in ArrayLikeTrainingLoop.fit(self, model, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_freq, **kwargs)
    595 def fit(self,
    596         model,
    597         x=None,
   (...)
    611         validation_freq=1,
    612         **kwargs):
    613   batch_size = model._validate_or_infer_batch_size(batch_size,
    614                                                    steps_per_epoch, x)
--> 616   x, y, sample_weights = model._standardize_user_data(
    617       x,
    618       y,
    619       sample_weight=sample_weight,
    620       class_weight=class_weight,
    621       batch_size=batch_size,
    622       check_steps=True,
    623       steps_name='steps_per_epoch',
    624       steps=steps_per_epoch,
    625       validation_split=validation_split,
    626       shuffle=shuffle)
    628   if validation_data:
    629     val_x, val_y, val_sample_weights = model._prepare_validation_data(
    630         validation_data, batch_size, validation_steps)

File ~/.local/lib/python3.10/site-packages/keras/engine/training_v1.py:2318, in Model._standardize_user_data(self, x, y, sample_weight, class_weight, batch_size, check_steps, steps_name, steps, validation_split, shuffle, extract_tensors_from_dataset)
   2316 is_compile_called = False
   2317 if not self._is_compiled and self.optimizer:
-> 2318   self._compile_from_inputs(all_inputs, y_input, x, y)
   2319   is_compile_called = True
   2321 # In graph mode, if we had just set inputs and targets as symbolic tensors
   2322 # by invoking build and compile on the model respectively, we do not have to
   2323 # feed anything to the model. Model already has input and target data as
   (...)
   2327 
   2328 # self.run_eagerly is not free to compute, so we want to reuse the value.

File ~/.local/lib/python3.10/site-packages/keras/engine/training_v1.py:2568, in Model._compile_from_inputs(self, all_inputs, target, orig_inputs, orig_target)
   2565   else:
   2566     target_tensors = None
-> 2568 self.compile(
   2569     optimizer=self.optimizer,
   2570     loss=self.loss,
   2571     metrics=self._compile_metrics,
   2572     weighted_metrics=self._compile_weighted_metrics,
   2573     loss_weights=self.loss_weights,
   2574     target_tensors=target_tensors,
   2575     sample_weight_mode=self.sample_weight_mode,
   2576     run_eagerly=self.run_eagerly,
   2577     experimental_run_tf_function=self._experimental_run_tf_function)

File ~/.local/lib/python3.10/site-packages/tensorflow/python/training/tracking/base.py:629, in no_automatic_dependency_tracking.<locals>._method_wrapper(self, *args, **kwargs)
    627 self._self_setattr_tracking = False  # pylint: disable=protected-access
    628 try:
--> 629   result = method(self, *args, **kwargs)
    630 finally:
    631   self._self_setattr_tracking = previous_value  # pylint: disable=protected-access

File ~/.local/lib/python3.10/site-packages/keras/engine/training_v1.py:443, in Model.compile(self, optimizer, loss, metrics, loss_weights, sample_weight_mode, weighted_metrics, target_tensors, distribute, **kwargs)
    439 training_utils_v1.prepare_sample_weight_modes(
    440     self._training_endpoints, sample_weight_mode)
    442 # Creates the model loss and weighted metrics sub-graphs.
--> 443 self._compile_weights_loss_and_weighted_metrics()
    445 # Functions for train, test and predict will
    446 # be compiled lazily when required.
    447 # This saves time when the user is not using all functions.
    448 self.train_function = None

File ~/.local/lib/python3.10/site-packages/tensorflow/python/training/tracking/base.py:629, in no_automatic_dependency_tracking.<locals>._method_wrapper(self, *args, **kwargs)
    627 self._self_setattr_tracking = False  # pylint: disable=protected-access
    628 try:
--> 629   result = method(self, *args, **kwargs)
    630 finally:
    631   self._self_setattr_tracking = previous_value  # pylint: disable=protected-access

File ~/.local/lib/python3.10/site-packages/keras/engine/training_v1.py:1537, in Model._compile_weights_loss_and_weighted_metrics(self, sample_weights)
   1524 self._handle_metrics(
   1525     self.outputs,
   1526     targets=self._targets,
   (...)
   1529     masks=masks,
   1530     return_weighted_metrics=True)
   1532 # Compute total loss.
   1533 # Used to keep track of the total loss value (stateless).
   1534 # eg., total_loss = loss_weight_1 * output_1_loss_fn(...) +
   1535 #                   loss_weight_2 * output_2_loss_fn(...) +
   1536 #                   layer losses.
-> 1537 self.total_loss = self._prepare_total_loss(masks)

File ~/.local/lib/python3.10/site-packages/keras/engine/training_v1.py:1597, in Model._prepare_total_loss(self, masks)
   1594     sample_weight *= mask
   1596 if hasattr(loss_fn, 'reduction'):
-> 1597   per_sample_losses = loss_fn.call(y_true, y_pred)
   1598   weighted_losses = losses_utils.compute_weighted_loss(
   1599       per_sample_losses,
   1600       sample_weight=sample_weight,
   1601       reduction=losses_utils.ReductionV2.NONE)
   1602   loss_reduction = loss_fn.reduction

File ~/.local/lib/python3.10/site-packages/keras/losses.py:245, in LossFunctionWrapper.call(self, y_true, y_pred)
    242   y_pred, y_true = losses_utils.squeeze_or_expand_dimensions(y_pred, y_true)
    244 ag_fn = tf.__internal__.autograph.tf_convert(self.fn, tf.__internal__.autograph.control_status_ctx())
--> 245 return ag_fn(y_true, y_pred, **self._fn_kwargs)

File ~/.local/lib/python3.10/site-packages/tensorflow/python/autograph/impl/api.py:692, in convert.<locals>.decorator.<locals>.wrapper(*args, **kwargs)
    690 except Exception as e:  # pylint:disable=broad-except
    691   if hasattr(e, 'ag_error_metadata'):
--> 692     raise e.ag_error_metadata.to_exception(e)
    693   else:
    694     raise

ValueError: in user code:

    File "/tmp/ipykernel_49162/810674056.py", line 8, in Loss  *
        loss += confidenceLoss(y[:,:,:-4],tf.cast(gt[:,:,0],tf.int32))
    File "/tmp/ipykernel_49162/2037607510.py", line 2, in confidenceLoss  *
        unweighted_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(label, y)

    ValueError: Only call sparse_softmax_cross_entropy_with_logits with named arguments (labels=..., logits=..., ...). Received unnamed argument: Tensor("loss/output_1_loss/Cast:0", shape=(None, None), dtype=int32)

 

버전이 문제였고 tensorflow는 그냥 v2로 쓰면 되는거였네 -_-

import tensorflow as tf
#import tensorflow.compat.v1 as tf
#tf.disable_v2_behavior()

 

최초는 2020년 7월 22일, 나중은 2022년 7월 21일(2년 만!)

saunack committed on Jul 21, 2022
saunack committed on Jul 22, 2020 

[링크 : https://github.com/saunack/MobileNetv2-SSD/commits/master/model.ipynb]

[링크 : https://github.com/saunack/MobileNetv2-SSD/blob/master/model.ipynb]

 

 

모델 저장(실패)

from keras.models import load_model
model.save('mnist_mlp_model.h5')

 

에러는 아래와 같이 나옴

NotImplementedError: Saving the model to HDF5 format requires the model to be a Functional model or a Sequential model. It does not work for subclassed models, because such models are defined via the body of a Python method, which isn't safely serializable. Consider saving to the Tensorflow SavedModel format (by setting save_format="tf") or using `save_weights`.

 

TensorFlow 2.0
TL;DR:

do not use model.save() for custom subclass keras model;
use save_weights() and load_weights() instead.

[링크 : https://stackoverflow.com/questions/51806852/cant-save-custom-subclassed-model]

 

sequential_model.save_weights("ckpt")

[링크 : https://www.tensorflow.org/guide/keras/save_and_serialize?hl=ko]

 

model.save_weights('model_weights', save_format='tf')

 

AttributeError: in user code:

    File "/home/user/.local/lib/python3.10/site-packages/keras/saving/saving_utils.py", line 138, in _wrapped_model  *
        outputs = model(*args, **kwargs)
    File "/tmp/ipykernel_53483/1508227539.py", line 46, in call  *
        x = self.MobileNet(x)
    File "/tmp/ipykernel_53483/3997091176.py", line 70, in call  *
        x = self.B2_2(x)
    File "/tmp/ipykernel_53483/1796771022.py", line 69, in call  *
        x = self.residual([inputs,x])
    File "/home/user/.local/lib/python3.10/site-packages/keras/utils/traceback_utils.py", line 67, in error_handler  **
        raise e.with_traceback(filtered_tb) from None
    File "/home/user/.local/lib/python3.10/site-packages/keras/engine/base_layer.py", line 1102, in __call__
        if self._saved_model_inputs_spec is None:

    AttributeError: 'Add' object has no attribute '_saved_model_inputs_spec'

[링크 : https://github.com/tensorflow/tensorflow/issues/29545]

 

에라이 저장을 못하겠다!

[링크 : https://www.tensorflow.org/lite/convert?hl=ko]

 

엉뚱(?)한데서 터지는 느낌인데

tensorflow 버전을 2.14.0 으로 올려야 하나? 2.8.0이 아니라?

[링크 : https://www.tensorflow.org/api_docs/python/tf/keras/layers/Add]

 

+

24.01.11

2.14.0 으로 한다고 달라지는 건 없음.. 도대체 Add 객체는 멀까?

 

+

def SSD()  로 생성된걸 keras.Sequential로 감싸고 학습은 진행되는데.. 저장이 왜 또 안될까? ㅠㅠ

model = SSD(numBoxes=numBoxes, layerWidth=layerWidths, k = outputChannels)
model = tf.keras.Sequential(model)
# model.model().summary()

[링크 : https://www.tensorflow.org/tutorials/keras/save_and_load?hl=ko]

[링크 : https://www.tensorflow.org/hub/exporting_tf2_saved_model?hl=ko]

Posted by 구차니

아래의 스크립트를 이용하여 변환이 가능하다는데, 정작 변환하고 실행하려고 하면 안되고(import는 안건드리니)

$ tf_upgrade_v2 --infile tensorfoo.py --outfile tensorfoo-upgraded.py

[링크 : https://www.tensorflow.org/guide/upgrade?hl=ko]

 

차라리 아래처럼 import tensorflow as tf를 compat.v1 으로 바꾸어주고, v2 를 끄면 구버전이 실행된다.

import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()

[링크 : https://www.tensorflow.org/guide/migrate?hl=ko]

 

 

--------------------------------

돌려보니 먼가 나오긴 한데

INFO line 7:16: Renamed 'tf.random_uniform' to 'tf.random.uniform'
INFO line 8:16: Renamed 'tf.random_uniform' to 'tf.random.uniform'
INFO line 11:4: Renamed 'tf.placeholder' to 'tf.compat.v1.placeholder'
INFO line 12:4: Renamed 'tf.placeholder' to 'tf.compat.v1.placeholder'
INFO line 25:12: Renamed 'tf.train.GradientDescentOptimizer' to 'tf.compat.v1.train.GradientDescentOptimizer'
INFO line 30:5: Renamed 'tf.Session' to 'tf.compat.v1.Session'
INFO line 31:13: Renamed 'tf.global_variables_initializer' to 'tf.compat.v1.global_variables_initializer'
TensorFlow 2.0 Upgrade Script
-----------------------------
Converted 1 files

 

잘 돈다는 보장은 없다 -_-

 

요런 에러가 뜨면

RuntimeError: tf.placeholder() is not compatible with eager execution.

 

아래줄 추가해주면 되는데

tf.compat.v1.disable_eager_execution()

[링크 : https://luvstudy.tistory.com/122]

 

정작 텐서 곱할 때, 에러가 발생한다.

RuntimeError: resource: Attempting to capture an EagerTensor without building a function.

 

요건 막혀서 모르겠네 -_-

함수를 만들지 않고 eagertensor를 capture 하기 시도해서 에러가 발생한거라면..

함수(building a function)를 만들면 되는건가?

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

mobilenet v2 ssd  (0) 2024.01.11
ssd-mobilenetv2 on jupyter notebook  (2) 2024.01.10
골빈해커의 3분 딥러닝 github  (0) 2024.01.10
ReLU - Rectified Linear Unit  (0) 2024.01.10
softmax  (0) 2024.01.10
Posted by 구차니

타이핑 하기 귀찮으니(!)

[링크 : https://github.com/golbin/TensorFlow-Tutorials]

 

+

현재 시점에서 아래의 소스는 단 두 줄 손 보면 돌아는 간다. (tfv2 인데 tfv1 하위 호환성으로 작동 시키기)

# X 와 Y 의 상관관계를 분석하는 기초적인 선형 회귀 모델을 만들고 실행해봅니다.
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()

x_data = [1, 2, 3]
y_data = [1, 2, 3]

W = tf.Variable(tf.random_uniform([1], -1.0, 1.0))
b = tf.Variable(tf.random_uniform([1], -1.0, 1.0))

# name: 나중에 텐서보드등으로 값의 변화를 추적하거나 살펴보기 쉽게 하기 위해 이름을 붙여줍니다.
X = tf.placeholder(tf.float32, name="X")
Y = tf.placeholder(tf.float32, name="Y")
print(X)
print(Y)

# X 와 Y 의 상관 관계를 분석하기 위한 가설 수식을 작성합니다.
# y = W * x + b
# W 와 X 가 행렬이 아니므로 tf.matmul 이 아니라 기본 곱셈 기호를 사용했습니다.
hypothesis = W * X + b

# 손실 함수를 작성합니다.
# mean(h - Y)^2 : 예측값과 실제값의 거리를 비용(손실) 함수로 정합니다.
cost = tf.reduce_mean(tf.square(hypothesis - Y))
# 텐서플로우에 기본적으로 포함되어 있는 함수를 이용해 경사 하강법 최적화를 수행합니다.
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.1)
# 비용을 최소화 하는 것이 최종 목표
train_op = optimizer.minimize(cost)

# 세션을 생성하고 초기화합니다.
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    # 최적화를 100번 수행합니다.
    for step in range(100):
        # sess.run 을 통해 train_op 와 cost 그래프를 계산합니다.
        # 이 때, 가설 수식에 넣어야 할 실제값을 feed_dict 을 통해 전달합니다.
        _, cost_val = sess.run([train_op, cost], feed_dict={X: x_data, Y: y_data})

        print(step, cost_val, sess.run(W), sess.run(b))

    # 최적화가 완료된 모델에 테스트 값을 넣고 결과가 잘 나오는지 확인해봅니다.
    print("\n=== Test ===")
    print("X: 5, Y:", sess.run(hypothesis, feed_dict={X: 5}))
    print("X: 2.5, Y:", sess.run(hypothesis, feed_dict={X: 2.5}))

 

$ python lr.py
2024-01-10 11:39:49.775206: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
2024-01-10 11:39:49.775245: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
2024-01-10 11:39:49.776215: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
2024-01-10 11:39:49.781682: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2024-01-10 11:39:50.440334: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
/usr/lib/python3/dist-packages/scipy/__init__.py:146: UserWarning: A NumPy version >=1.17.3 and <1.25.0 is required for this version of SciPy (detected version 1.26.3
  warnings.warn(f"A NumPy version >={np_minversion} and <{np_maxversion}"
WARNING:tensorflow:From /home/falinux/.local/lib/python3.10/site-packages/tensorflow/python/compat/v2_compat.py:108: disable_resource_variables (from tensorflow.python.ops.variable_scope) is deprecated and will be removed in a future version.
Instructions for updating:
non-resource variables are not supported in the long term
Tensor("X:0", dtype=float32)
Tensor("Y:0", dtype=float32)
2024-01-10 11:39:51.327415: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:388] MLIR V1 optimization pass is not enabled
0 6.4782066 [1.2373642] [-0.24653786]
1 0.089632 [1.1144395] [-0.29217595]
2 0.012737438 [1.1244997] [-0.27951655]
3 0.011264746 [1.1201066] [-0.2734131]
4 0.01071928 [1.1173724] [-0.2667731]
5 0.010209985 [1.1145341] [-0.26036742]
6 0.009725014 [1.1117826] [-0.2541076]
7 0.009263077 [1.1090952] [-0.2479991]
8 0.008823066 [1.1064726] [-0.24203737]
9 0.008403975 [1.1039131] [-0.23621896]
10 0.008004769 [1.1014152] [-0.2305404]
11 0.007624544 [1.0989771] [-0.22499838]
12 0.007262358 [1.0965978] [-0.21958955]
13 0.0069174054 [1.0942756] [-0.21431077]
14 0.0065888255 [1.0920093] [-0.20915887]
15 0.0062758424 [1.0897975] [-0.20413081]
16 0.0059777307 [1.0876389] [-0.19922365]
17 0.0056937817 [1.0855321] [-0.19443446]
18 0.0054233256 [1.083476] [-0.1897604]
19 0.0051657106 [1.0814693] [-0.1851987]
20 0.0049203373 [1.0795108] [-0.18074667]
21 0.004686633 [1.0775993] [-0.17640167]
22 0.0044640056 [1.0757339] [-0.17216106]
23 0.0042519583 [1.0739133] [-0.1680224]
24 0.004049988 [1.0721365] [-0.16398326]
25 0.0038576098 [1.0704024] [-0.16004121]
26 0.0036743751 [1.06871] [-0.15619393]
27 0.0034998383 [1.0670582] [-0.15243913]
28 0.003333594 [1.0654461] [-0.1487746]
29 0.003175243 [1.0638729] [-0.14519812]
30 0.0030244188 [1.0623374] [-0.14170769]
31 0.0028807523 [1.0608389] [-0.13830112]
32 0.0027439168 [1.0593764] [-0.13497646]
33 0.0026135833 [1.057949] [-0.13173172]
34 0.002489428 [1.056556] [-0.12856494]
35 0.0023711843 [1.0551964] [-0.12547435]
36 0.0022585478 [1.0538695] [-0.12245804]
37 0.0021512664 [1.0525745] [-0.11951423]
38 0.0020490757 [1.0513107] [-0.11664119]
39 0.0019517452 [1.0500772] [-0.11383722]
40 0.0018590376 [1.0488734] [-0.11110065]
41 0.0017707323 [1.0476985] [-0.10842989]
42 0.0016866213 [1.0465518] [-0.1058233]
43 0.0016065066 [1.0454327] [-0.10327938]
44 0.0015301956 [1.0443406] [-0.1007966]
45 0.0014575059 [1.0432746] [-0.09837352]
46 0.0013882784 [1.0422344] [-0.09600867]
47 0.0013223292 [1.0412191] [-0.0937007]
48 0.0012595187 [1.0402282] [-0.0914482]
49 0.0011996872 [1.0392612] [-0.08924985]
50 0.0011427039 [1.0383173] [-0.08710436]
51 0.0010884297 [1.0373962] [-0.08501042]
52 0.0010367227 [1.0364972] [-0.08296681]
53 0.0009874817 [1.0356199] [-0.08097235]
54 0.0009405748 [1.0347636] [-0.07902583]
55 0.00089589664 [1.0339279] [-0.07712609]
56 0.00085334125 [1.0331123] [-0.07527205]
57 0.0008128048 [1.0323163] [-0.07346255]
58 0.0007741994 [1.0315394] [-0.07169659]
59 0.00073742354 [1.0307813] [-0.06997304]
60 0.00070239493 [1.0300413] [-0.06829095]
61 0.00066903216 [1.0293192] [-0.0666493]
62 0.0006372516 [1.0286143] [-0.06504711]
63 0.0006069818 [1.0279264] [-0.0634834]
64 0.00057814806 [1.0272552] [-0.0619573]
65 0.00055068725 [1.0265999] [-0.06046791]
66 0.0005245278 [1.0259604] [-0.05901428]
67 0.0004996119 [1.0253364] [-0.0575956]
68 0.00047588357 [1.0247273] [-0.05621104]
69 0.0004532766 [1.0241328] [-0.05485978]
70 0.00043174453 [1.0235528] [-0.05354097]
71 0.00041123512 [1.0229865] [-0.05225388]
72 0.0003917031 [1.022434] [-0.05099772]
73 0.00037309653 [1.0218947] [-0.04977177]
74 0.00035537416 [1.0213684] [-0.04857529]
75 0.00033849102 [1.0208547] [-0.04740757]
76 0.00032241447 [1.0203533] [-0.04626793]
77 0.00030709928 [1.0198641] [-0.04515567]
78 0.00029251093 [1.0193865] [-0.04407016]
79 0.0002786171 [1.0189205] [-0.04301074]
80 0.00026538406 [1.0184656] [-0.04197682]
81 0.00025277727 [1.0180218] [-0.04096771]
82 0.0002407704 [1.0175885] [-0.0399829]
83 0.0002293337 [1.0171658] [-0.03902172]
84 0.00021844136 [1.0167531] [-0.03808369]
85 0.00020806213 [1.0163504] [-0.03716817]
86 0.00019818085 [1.0159572] [-0.0362747]
87 0.000188766 [1.0155737] [-0.03540265]
88 0.00017980166 [1.0151993] [-0.03455162]
89 0.00017126095 [1.0148339] [-0.03372103]
90 0.00016312544 [1.0144774] [-0.0329104]
91 0.0001553779 [1.0141293] [-0.03211929]
92 0.00014799698 [1.0137897] [-0.03134715]
93 0.00014096718 [1.0134581] [-0.03059359]
94 0.00013426914 [1.0131347] [-0.02985811]
95 0.00012789248 [1.0128189] [-0.02914038]
96 0.00012181744 [1.0125108] [-0.02843988]
97 0.00011603059 [1.01221] [-0.02775621]
98 0.00011052046 [1.0119165] [-0.02708898]
99 0.00010527024 [1.01163] [-0.02643778]

=== Test ===
X: 5, Y: [5.0317125]
X: 2.5, Y: [2.5026374]

[링크 : https://github.com/golbin/TensorFlow-Tutorials/blob/master/03%20-%20TensorFlow%20Basic/03%20-%20Linear%20Regression.py]

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

ssd-mobilenetv2 on jupyter notebook  (2) 2024.01.10
텐서플로우 v1 을 v2로 마이그레이션은 실패 -_-  (0) 2024.01.10
ReLU - Rectified Linear Unit  (0) 2024.01.10
softmax  (0) 2024.01.10
텐서플로우 학습  (0) 2024.01.09
Posted by 구차니

ReLU는 일종의 threshold 함수인데(loss 함수 혹은 손실 함수 등등등..)

0 미만은 0 으로 억제하는 함수이다.

 

그나저나 Rectified를 찾아보니 정정하다(correct) 정류하다 등으로 뜻이 나오는데

수정된 선형 단위 라고 번역을 하면 되려나?

 

[링크 : https://ko.wikipedia.org/wiki/ReLU]

 

retifier는 전자회로에서 "정류기"로 많이 번역되는데, 다이오드 등을 통해서 교류를 직류로 바꾸는 걸 의미한다.

[링크 : https://ko.wikipedia.org/wiki/정류기]

 

그나저나 다이오드의 전압 그래프를 보면 ReLU랑 비슷한 것 같으면서도 아닌것 같기도 하고(...?!)

아무튼 머.. 그렇다고 한다.

[링크 : http://www.ktechno.co.kr/ls_parts/parts04.html]

Posted by 구차니

netron으로 보다보면 softmax라는게 나오는데

그냥 그러려니 하고 넘어가던거에서 조금은 이론적으로 설명이 되는걸 보니 궁금해짐

[링크 : https://m.hanbit.co.kr/store/books/book_view.html?p_code=B7257101308]

 

아무튼 수식으로는 먼가 와닫지 않는데

[링크 : https://syj9700.tistory.com/38]

 

값들의 평균을 내어 합이 1이 되도록 정규화한다고 해야하나..

(1,2,8)을 (0.001, 0.002, 0.997) 로 변환한다.

(1,2,8) 에 e^n 을 하면

(e^1, e^2, e^8) 이 되고

밑은 e^1 +  e^2 + e^8 하면 되니까

(e^1 / (e^1 +  e^2 + e^8), e^2 / (e^1 +  e^2 + e^8), e^8 / (e^1 +  e^2 + e^8)) 로 계산하면

 

(2.71828182845904, 7.38905609893065, 2980.95798704173)

2.71828182845904 + 7.38905609893065 + 2980.95798704173 = 2991.06532496912

(2.71828182845904 / 2991.06532496912, 7.38905609893065 / 2991.06532496912, 2980.95798704173 / 2991.06532496912)

= (0.000908800555363033, 0.00247037603533682, 0.9966208234093)

 

이름과 달리 최댓값(max) 함수를 매끄럽거나 부드럽게 한 것이 아니라, 최댓값의 인수인 원핫 형태의 arg max 함수를 매끄럽게 한 것이다. 그 계산 방법은 입력값을 자연로그의 밑을 밑으로 한 지수 함수를 취한 뒤 그 지수함수의 합으로 나눠주는 것이다.

[링크 : https://ko.wikipedia.org/wiki/소프트맥스_함수]

 

For example, the standard softmax of (1,2,8) is approximately (0.001,0.002,0.997), which amounts to assigning almost all of the total unit weight in the result to the position of the vector's maximal element (of 8).

>>> import numpy as np
>>> a = [1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0]
>>> np.exp(a) / np.sum(np.exp(a)) 
array([0.02364054, 0.06426166, 0.1746813, 0.474833, 0.02364054,
       0.06426166, 0.1746813])

[링크 : https://en.wikipedia.org/wiki/Softmax_function]

 

아무튼 계산에 의한 결과가 true, false로 판별할 수 있는 값이 아닌

사람이 보기 편한 값으로 환산되기 때문에, 에측에는 softmax를 쓰지 말라는게 이해 될 것 같기도, 안 갈 것 같기도..

[링크 : https://velog.io/@francomoon7/예측에-Softmax를-사용하면-안되는-이유]

Posted by 구차니