'프로그램 사용/yolo_tensorflow'에 해당되는 글 113건

  1. 2025.09.17 keras - transfer learning / fine tuning
  2. 2025.09.15 mobilenet 학습시키기 with keras, tensorflow
  3. 2025.09.11 ssd mobilenet
  4. 2025.09.06 LLM 시각화
  5. 2025.09.05 모델 학습
  6. 2025.09.05 mean Average Precision(mAP)
  7. 2025.09.05 내장 그래픽으로 ROCm?
  8. 2025.09.04 ai 모델들 조사
  9. 2025.08.22 facenet
  10. 2024.01.15 NMS, soft-NMS
프로그램 사용/yolo_tensorflow2025. 9. 17. 14:28

keras - transfer learning / fine tuning

tflite 에서 weight만 빼서 거기서 부터 학습 시키면 그게 fine tune / transfer learning(전이학습)이 되는건가?

The conversion from a TensorFlow SaveModel or tf.keras H5 model to .tflite is an irreversible process. Specifically, the original model topology is optimized during the compilation by the TFLite converter, which leads to some loss of information. Also, the original tf.keras model's loss and optimizer configurations are discarded, because those aren't required for inference.

However, the .tflite file still contains some information that can help you restore the original trained model. Most importantly, the weight values are available, although they might be quantized, which could lead to some loss in precision.

The code example below shows you how to read weight values from a .tflite file after it's created from a simple trained tf.keras.Model.



import numpy as np
import tensorflow as tf

# First, create and train a dummy model for demonstration purposes.
model = tf.keras.Sequential([
    tf.keras.layers.Dense(10, input_shape=[5], activation="relu"),
    tf.keras.layers.Dense(1, activation="sigmoid")])
model.compile(loss="binary_crossentropy", optimizer="sgd")

xs = np.ones([8, 5])
ys = np.zeros([8, 1])
model.fit(xs, ys, epochs=1)

# Convert it to a TFLite model file.
converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()
open("converted.tflite", "wb").write(tflite_model)

# Use `tf.lite.Interpreter` to load the written .tflite back from the file system.
interpreter = tf.lite.Interpreter(model_path="converted.tflite")
all_tensor_details = interpreter.get_tensor_details()
interpreter.allocate_tensors()

for tensor_item in all_tensor_details:
  print("Weight %s:" % tensor_item["name"])
  print(interpreter.tensor(tensor_item["index"])())

[링크 : https://stackoverflow.com/questions/59559289/is-there-any-way-to-convert-a-tensorflow-lite-tflite-file-back-to-a-keras-fil]

 

읽어오든.. tf.keras.application.MobileNetV2 해서 만들던 weight의 유무를 제외하면 동일하게 불러오는건가 보다.

# Create the base model from the pre-trained model MobileNet V2
IMG_SHAPE = IMG_SIZE + (3,)
base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE,
                                               include_top=False,
                                               weights='imagenet')
#or load your own
#base_model= tf.saved_model.load("./pretrained_models/ssd_mobilenet_v2_320x320_coco17_tpu-8/saved_model")

[링크 : https://www.aranacorp.com/en/training-a-tensorflow2-model-with-keras/]

 

다시 찾아보니 keras쪽 표현으로

transfer-learning은 기존의 값을 변화시키지 않고 추가 레이어에 학습시키는 것 같고

The typical transfer-learning workflow
This leads us to how a typical transfer learning workflow can be implemented in Keras:

Instantiate a base model and load pre-trained weights into it.
Freeze all layers in the base model by setting trainable = False.
Create a new model on top of the output of one (or several) layers from the base model.
Train your new model on your new dataset.
Note that an alternative, more lightweight workflow could also be:

Instantiate a base model and load pre-trained weights into it.
Run your new dataset through it and record the output of one (or several) layers from the base model. This is called feature extraction.
Use that output as input data for a new, smaller model.


First, instantiate a base model with pre-trained weights.

base_model = keras.applications.Xception(
    weights='imagenet',  # Load weights pre-trained on ImageNet.
    input_shape=(150, 150, 3),
    include_top=False)  # Do not include the ImageNet classifier at the top.
Then, freeze the base model.

base_model.trainable = False


Create a new model on top.

inputs = keras.Input(shape=(150, 150, 3))
# We make sure that the base_model is running in inference mode here,
# by passing `training=False`. This is important for fine-tuning, as you will
# learn in a few paragraphs.
x = base_model(inputs, training=False)
# Convert features of shape `base_model.output_shape[1:]` to vectors
x = keras.layers.GlobalAveragePooling2D()(x)
# A Dense classifier with a single unit (binary classification)
outputs = keras.layers.Dense(1)(x)
model = keras.Model(inputs, outputs)
Train the model on new data.

model.compile(optimizer=keras.optimizers.Adam(),
              loss=keras.losses.BinaryCrossentropy(from_logits=True),
              metrics=[keras.metrics.BinaryAccuracy()])
model.fit(new_dataset, epochs=20, callbacks=..., validation_data=...)

 

fine tuning은 모델 자체의 가중치를 변경시킬수 있도록 해서, 천천히 학습 시키는 것 같다.

Fine-tuning
Once your model has converged on the new data, you can try to unfreeze all or part of the base model and retrain the whole model end-to-end with a very low learning rate.

This is an optional last step that can potentially give you incremental improvements. It could also potentially lead to quick overfitting – keep that in mind.

It is critical to only do this step after the model with frozen layers has been trained to convergence. If you mix randomly-initialized trainable layers with trainable layers that hold pre-trained features, the randomly-initialized layers will cause very large gradient updates during training, which will destroy your pre-trained features.

It's also critical to use a very low learning rate at this stage, because you are training a much larger model than in the first round of training, on a dataset that is typically very small. As a result, you are at risk of overfitting very quickly if you apply large weight updates. Here, you only want to readapt the pretrained weights in an incremental way.


This is how to implement fine-tuning of the whole base model:

# Unfreeze the base model
base_model.trainable = True

# It's important to recompile your model after you make any changes
# to the `trainable` attribute of any inner layer, so that your changes
# are take into account
model.compile(optimizer=keras.optimizers.Adam(1e-5),  # Very low learning rate
              loss=keras.losses.BinaryCrossentropy(from_logits=True),
              metrics=[keras.metrics.BinaryAccuracy()])

# Train end-to-end. Be careful to stop before you overfit!
model.fit(new_dataset, epochs=10, callbacks=..., validation_data=...)

[링크 : https://keras.io/guides/transfer_learning/]

 

 

+

[링크 : https://89douner.tistory.com/270] vgg16 기준 추가학습

 

SavedModel 형식과 비교하여 H5 파일에 포함되지 않은 두 가지가 있습니다.

model.add_loss() 및 model.add_metric()을 통해 추가된 외부 손실 및 메트릭은 저장되지 않습니다(SavedModel과 다름). 모델에 이러한 솔실 및 메트릭이 있고 훈련을 다시 시작하려면 모델을 로드한 후 이러한 손실을 다시 추가해야 합니다. 이는 self.add_loss() 및 self.add_metric()을 통해 레이어 내부에서 생성한 손실/메트릭에는 적용되지 않습니다. 이러한 손실 및 메트릭은 레이어가 로드되는 한 레이어의 call 메서드의 일부이기 때문에 계속 유지됩니다.
사용자 정의 레이어와 같은 사용자 정의 객체의 계산 그래프는 저장 파일에 포함되지 않습니다. 로드 시 Keras는 모델을 다시 구성하기 위해 이러한 객체의 Python 클래스/함수에 액세스해야 합니다. 사용자 정의 객체를 참고하세요.

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

저작자표시 (새창열림)

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

mobilenet 학습시키기 with keras, tensorflow  (0) 2025.09.15
ssd mobilenet  (0) 2025.09.11
LLM 시각화  (0) 2025.09.06
모델 학습  (0) 2025.09.05
mean Average Precision(mAP)  (0) 2025.09.05
Posted by 구차니
프로그램 사용/yolo_tensorflow2025. 9. 15. 17:11

mobilenet 학습시키기 with keras, tensorflow

대충 버전이 맞았는지 돌아는 간다.


Epoch 1/25
 52/755 [=>............................] - ETA: 1:04:39 - loss: 0.3203 - accuracy: 0.8534    

 

주요 설치 패키지 버전은 아래와 같고

keras                        2.14.0
mobilenet-v3                 0.1.2
numpy                        1.24.4
tensorflow                   2.14.0

 

수정된 소스는 다음과 같다.

그런데 voc 디렉토리를 통채로 넣었는데 어찌 돌아는 가는데.. 어떤 파일로 학습을 하는거냐.. -_-

from keras.applications import MobileNet
from keras.models import Sequential,Model 
from keras.layers import Dense,Dropout,Activation,Flatten,GlobalAveragePooling2D
from keras.layers import Conv2D,MaxPooling2D,ZeroPadding2D
from keras.layers.normalization import BatchNormalization
from tensorflow.keras.preprocessing.image import ImageDataGenerator

# MobileNet is designed to work with images of dim 224,224
img_rows,img_cols = 224,224

MobileNet = MobileNet(weights='imagenet',include_top=False,input_shape=(img_rows,img_cols,3))

# Here we freeze the last 4 layers
# Layers are set to trainable as True by default

for layer in MobileNet.layers:
    layer.trainable = True

# Let's print our layers
for (i,layer) in enumerate(MobileNet.layers):
    print(str(i),layer.__class__.__name__,layer.trainable)

def addTopModelMobileNet(bottom_model, num_classes):
    """creates the top or head of the model that will be 
    placed ontop of the bottom layers"""
    top_model = bottom_model.output
    top_model = GlobalAveragePooling2D()(top_model)
    top_model = Dense(1024,activation='relu')(top_model)
    top_model = Dense(1024,activation='relu')(top_model)
    top_model = Dense(512,activation='relu')(top_model)
    top_model = Dense(num_classes,activation='softmax')(top_model)
    return top_model

num_classes = 5  # ['Angry','Happy','Neutral','Sad','Surprise']

FC_Head = addTopModelMobileNet(MobileNet, num_classes)

model = Model(inputs = MobileNet.input, outputs = FC_Head)

print(model.summary())

train_data_dir = 'VOC2012_train_val/VOC2012_train_val'
validation_data_dir = 'VOC2012_test/VOC2012_test'

train_datagen = ImageDataGenerator(
                    rescale=1./255,
                    rotation_range=30,
                    width_shift_range=0.3,
                    height_shift_range=0.3,
                    horizontal_flip=True,
                    fill_mode='nearest'
                                   )

validation_datagen = ImageDataGenerator(rescale=1./255)

batch_size = 32

train_generator = train_datagen.flow_from_directory(
                        train_data_dir,
                        target_size = (img_rows,img_cols),
                        batch_size = batch_size,
                        class_mode = 'categorical'
                        )

validation_generator = validation_datagen.flow_from_directory(
                            validation_data_dir,
                            target_size=(img_rows,img_cols),
                            batch_size=batch_size,
                            class_mode='categorical')

from keras.optimizers import RMSprop,Adam
from keras.callbacks import ModelCheckpoint,EarlyStopping,ReduceLROnPlateau

checkpoint = ModelCheckpoint(
                             'emotion_face_mobilNet.h5',
                             monitor='val_loss',
                             mode='min',
                             save_best_only=True,
                             verbose=1)

earlystop = EarlyStopping(
                          monitor='val_loss',
                          min_delta=0,
                          patience=10,
                          verbose=1,restore_best_weights=True)

learning_rate_reduction = ReduceLROnPlateau(monitor='val_acc', 
                                            patience=5, 
                                            verbose=1, 
                                            factor=0.2, 
                                            min_lr=0.0001)

callbacks = [earlystop,checkpoint,learning_rate_reduction]

model.compile(loss='categorical_crossentropy',
              optimizer=Adam(learning_rate=0.001),
              metrics=['accuracy']
              )

nb_train_samples = 24176
nb_validation_samples = 3006

epochs = 25

history = model.fit(
            train_generator,
            steps_per_epoch=nb_train_samples//batch_size,     
            epochs=epochs,
            callbacks=callbacks,
            validation_data=validation_generator,
            validation_steps=nb_validation_samples//batch_size)


 

돌리다가 에러가 나서 멘붕.. 급 귀찮아짐..

먼가 파일 갯수가 안 맞는건가?

Epoch 1/25
718/755 [===========================>..] - ETA: 3:28 - loss: 0.1569 - accuracy: 0.9301WARNING:tensorflow:Your input ran out of data; interrupting training. Make sure that your dataset or generator can generate at least `steps_per_epoch * epochs` batches (in this case, 18875 batches). You may need to use the repeat() function when building your dataset.
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/utils/traceback_utils.py", line 70, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "/home/minimonk/.local/lib/python3.10/site-packages/tensorflow/python/eager/execute.py", line 60, in quick_execute
    tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,
tensorflow.python.framework.errors_impl.InvalidArgumentError: Graph execution error:

Detected at node categorical_crossentropy/softmax_cross_entropy_with_logits defined at (most recent call last):
  File "<stdin>", line 1, in <module>

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/utils/traceback_utils.py", line 65, in error_handler

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/engine/training.py", line 1832, in fit

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/utils/traceback_utils.py", line 65, in error_handler

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/engine/training.py", line 2272, in evaluate

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/engine/training.py", line 4079, in run_step

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/engine/training.py", line 2042, in test_function

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/engine/training.py", line 2025, in step_function

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/engine/training.py", line 2013, in run_step

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/engine/training.py", line 1895, in test_step

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/engine/training.py", line 1185, in compute_loss

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/engine/compile_utils.py", line 277, in __call__

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/losses.py", line 143, in __call__

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/losses.py", line 270, in call

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/losses.py", line 2221, in categorical_crossentropy

  File "/home/minimonk/.local/lib/python3.10/site-packages/keras/src/backend.py", line 5581, in categorical_crossentropy

logits and labels must be broadcastable: logits_size=[32,5] labels_size=[32,3]
 [[{{node categorical_crossentropy/softmax_cross_entropy_with_logits}}]] [Op:__inference_test_function_15346]
>>> 

 

전체 pip 패키지들 버전 정보는 아래와 같다.

$ pip list
Package                      Version
---------------------------- ----------------
absl-py                      2.3.1
appdirs                      1.4.4
apturl                       0.5.2
astunparse                   1.6.3
attrs                        21.2.0
bcrypt                       3.2.0
beautifulsoup4               4.10.0
beniget                      0.4.1
blinker                      1.4
Brlapi                       0.8.3
Brotli                       1.0.9
cachetools                   5.5.2
certifi                      2020.6.20
chardet                      4.0.0
click                        8.0.3
colorama                     0.4.4
command-not-found            0.3
cryptography                 3.4.8
cupshelpers                  1.0
cycler                       0.11.0
dbus-python                  1.2.18
decorator                    4.4.2
defer                        1.0.6
distro                       1.7.0
distro-info                  1.1+ubuntu0.2
duplicity                    0.8.21
fasteners                    0.14.1
flatbuffers                  25.2.10
fonttools                    4.29.1
fs                           2.4.12
future                       0.18.2
gast                         0.6.0
google-auth                  2.40.3
google-auth-oauthlib         1.0.0
google-pasta                 0.2.0
grpcio                       1.74.0
h5py                         3.14.0
html5lib                     1.1
httplib2                     0.20.2
idna                         3.3
importlib-metadata           4.6.4
jeepney                      0.7.1
keras                        2.14.0
keyring                      23.5.0
kiwisolver                   1.3.2
language-selector            0.1
launchpadlib                 1.10.16
lazr.restfulclient           0.14.4
lazr.uri                     1.0.6
libclang                     18.1.1
lockfile                     0.12.2
louis                        3.20.0
lxml                         4.8.0
lz4                          3.1.3+dfsg
macaroonbakery               1.3.1
Mako                         1.1.3
Markdown                     3.9
markdown-it-py               4.0.0
MarkupSafe                   3.0.2
matplotlib                   3.5.1
mdurl                        0.1.2
meld                         3.20.4
ml-dtypes                    0.2.0
mobilenet-v3                 0.1.2
monotonic                    1.6
more-itertools               8.10.0
mpmath                       0.0.0
namex                        0.1.0
netifaces                    0.11.0
numpy                        1.24.4
oauthlib                     3.2.0
olefile                      0.46
opt_einsum                   3.4.0
optree                       0.17.0
packaging                    21.3
paramiko                     2.9.3
pexpect                      4.8.0
Pillow                       9.0.1
pip                          22.0.2
Pivy                         0.6.5
ply                          3.11
protobuf                     4.25.8
ptyprocess                   0.7.0
pyasn1                       0.6.1
pyasn1_modules               0.4.2
pycairo                      1.20.1
pycups                       2.0.1
Pygments                     2.19.2
PyGObject                    3.42.1
PyJWT                        2.3.0
pymacaroons                  0.13.0
PyNaCl                       1.5.0
pyparsing                    2.4.7
pyRFC3339                    1.1
python-apt                   2.4.0+ubuntu4
python-dateutil              2.8.1
python-debian                0.1.43+ubuntu1.1
pythran                      0.10.0
pytz                         2022.1
pyxdg                        0.27
PyYAML                       5.4.1
reportlab                    3.6.8
requests                     2.25.1
requests-oauthlib            2.0.0
rich                         14.1.0
rsa                          4.9.1
scipy                        1.15.3
scour                        0.38.2
SecretStorage                3.3.1
setuptools                   59.6.0
six                          1.16.0
soupsieve                    2.3.1
ssh-import-id                5.11
sympy                        1.9
systemd-python               234
tensorboard                  2.14.1
tensorboard-data-server      0.7.2
tensorflow                   2.14.0
tensorflow-estimator         2.14.0
tensorflow-io-gcs-filesystem 0.37.1
termcolor                    3.1.0
typing_extensions            4.15.0
ubuntu-drivers-common        0.0.0
ubuntu-pro-client            8001
ufoLib2                      0.13.1
ufw                          0.36.1
unattended-upgrades          0.1
unicodedata2                 14.0.0
urllib3                      1.26.5
usb-creator                  0.3.7
wadllib                      1.3.6
webencodings                 0.5.1
Werkzeug                     3.1.3
wheel                        0.37.1
wrapt                        1.14.2
xdg                          5
xkit                         0.0.0
zipp                         1.0.0

 

-------- 아래는 참고 안하는게 속 편할지도...?

2020년 3월의 문서를 keras와 tensorflow로 2025년에 다시 시도해봄

 

일단은 아래처럼 설치하니 어찌 되는 느낌

$ pip install mobilenet-v3
$ pip install tensorflow
$ pip install numpy==1.26.4

 

상세로그

$ pip install mobilenet-v3
Defaulting to user installation because normal site-packages is not writeable
Collecting mobilenet-v3
  Downloading mobilenet_v3-0.1.4-py3-none-any.whl (18 kB)
Installing collected packages: mobilenet-v3
Successfully installed mobilenet-v3-0.1.4

$ pip install tensorflow
Defaulting to user installation because normal site-packages is not writeable
Collecting tensorflow
  Downloading tensorflow-2.20.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (620.4 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 620.4/620.4 MB 2.5 MB/s eta 0:00:00
Collecting google_pasta>=0.1.1
  Downloading google_pasta-0.2.0-py3-none-any.whl (57 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 57.5/57.5 KB 9.6 MB/s eta 0:00:00
Collecting absl-py>=1.0.0
  Downloading absl_py-2.3.1-py3-none-any.whl (135 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 135.8/135.8 KB 12.8 MB/s eta 0:00:00
Collecting libclang>=13.0.0
  Downloading libclang-18.1.1-py2.py3-none-manylinux2010_x86_64.whl (24.5 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 24.5/24.5 MB 9.6 MB/s eta 0:00:00
Collecting gast!=0.5.0,!=0.5.1,!=0.5.2,>=0.2.1
  Downloading gast-0.6.0-py3-none-any.whl (21 kB)
Collecting protobuf>=5.28.0
  Downloading protobuf-6.32.1-cp39-abi3-manylinux2014_x86_64.whl (322 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 322.0/322.0 KB 11.5 MB/s eta 0:00:00
Collecting ml_dtypes<1.0.0,>=0.5.1
  Downloading ml_dtypes-0.5.3-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl (4.9 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 4.9/4.9 MB 9.9 MB/s eta 0:00:00
Collecting numpy>=1.26.0
  Downloading numpy-2.2.6-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (16.8 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 16.8/16.8 MB 9.9 MB/s eta 0:00:00
Requirement already satisfied: packaging in /usr/lib/python3/dist-packages (from tensorflow) (21.3)
Requirement already satisfied: six>=1.12.0 in /usr/lib/python3/dist-packages (from tensorflow) (1.16.0)
Collecting termcolor>=1.1.0
  Downloading termcolor-3.1.0-py3-none-any.whl (7.7 kB)
Collecting typing_extensions>=3.6.6
  Downloading typing_extensions-4.15.0-py3-none-any.whl (44 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 44.6/44.6 KB 8.2 MB/s eta 0:00:00
Collecting tensorboard~=2.20.0
  Downloading tensorboard-2.20.0-py3-none-any.whl (5.5 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 5.5/5.5 MB 10.9 MB/s eta 0:00:00
Collecting astunparse>=1.6.0
  Downloading astunparse-1.6.3-py2.py3-none-any.whl (12 kB)
Collecting opt_einsum>=2.3.2
  Downloading opt_einsum-3.4.0-py3-none-any.whl (71 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 71.9/71.9 KB 8.4 MB/s eta 0:00:00
Collecting h5py>=3.11.0
  Downloading h5py-3.14.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (4.6 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 4.6/4.6 MB 9.4 MB/s eta 0:00:00
Collecting flatbuffers>=24.3.25
  Downloading flatbuffers-25.2.10-py2.py3-none-any.whl (30 kB)
Collecting keras>=3.10.0
  Downloading keras-3.11.3-py3-none-any.whl (1.4 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.4/1.4 MB 8.0 MB/s eta 0:00:00
Requirement already satisfied: setuptools in /usr/lib/python3/dist-packages (from tensorflow) (59.6.0)
Requirement already satisfied: requests<3,>=2.21.0 in /usr/lib/python3/dist-packages (from tensorflow) (2.25.1)
Collecting wrapt>=1.11.0
  Downloading wrapt-1.17.3-cp310-cp310-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl (81 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 82.0/82.0 KB 7.3 MB/s eta 0:00:00
Collecting grpcio<2.0,>=1.24.3
  Downloading grpcio-1.74.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (6.2 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 6.2/6.2 MB 12.3 MB/s eta 0:00:00
Requirement already satisfied: wheel<1.0,>=0.23.0 in /usr/lib/python3/dist-packages (from astunparse>=1.6.0->tensorflow) (0.37.1)
Collecting optree
  Downloading optree-0.17.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl (387 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 388.0/388.0 KB 11.8 MB/s eta 0:00:00
Collecting rich
  Downloading rich-14.1.0-py3-none-any.whl (243 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 243.4/243.4 KB 10.4 MB/s eta 0:00:00
Collecting namex
  Downloading namex-0.1.0-py3-none-any.whl (5.9 kB)
Requirement already satisfied: pillow in /usr/lib/python3/dist-packages (from tensorboard~=2.20.0->tensorflow) (9.0.1)
Collecting werkzeug>=1.0.1
  Downloading werkzeug-3.1.3-py3-none-any.whl (224 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 224.5/224.5 KB 7.7 MB/s eta 0:00:00
Collecting markdown>=2.6.8
  Downloading markdown-3.9-py3-none-any.whl (107 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 107.4/107.4 KB 7.1 MB/s eta 0:00:00
Collecting tensorboard-data-server<0.8.0,>=0.7.0
  Downloading tensorboard_data_server-0.7.2-py3-none-manylinux_2_31_x86_64.whl (6.6 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 6.6/6.6 MB 9.9 MB/s eta 0:00:00
Collecting MarkupSafe>=2.1.1
  Downloading MarkupSafe-3.0.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (20 kB)
Collecting markdown-it-py>=2.2.0
  Downloading markdown_it_py-4.0.0-py3-none-any.whl (87 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 87.3/87.3 KB 8.2 MB/s eta 0:00:00
Collecting pygments<3.0.0,>=2.13.0
  Downloading pygments-2.19.2-py3-none-any.whl (1.2 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.2/1.2 MB 13.2 MB/s eta 0:00:00
Collecting mdurl~=0.1
  Downloading mdurl-0.1.2-py3-none-any.whl (10.0 kB)
Installing collected packages: namex, libclang, flatbuffers, wrapt, typing_extensions, termcolor, tensorboard-data-server, pygments, protobuf, opt_einsum, numpy, mdurl, MarkupSafe, markdown, grpcio, google_pasta, gast, astunparse, absl-py, werkzeug, optree, ml_dtypes, markdown-it-py, h5py, tensorboard, rich, keras, tensorflow
  WARNING: The script pygmentize is installed in '/home/minimonk/.local/bin' which is not on PATH.
  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
  WARNING: The scripts f2py and numpy-config are installed in '/home/minimonk/.local/bin' which is not on PATH.
  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
  WARNING: The script markdown_py is installed in '/home/minimonk/.local/bin' which is not on PATH.
  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
  WARNING: The script markdown-it is installed in '/home/minimonk/.local/bin' which is not on PATH.
  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
  WARNING: The script tensorboard is installed in '/home/minimonk/.local/bin' which is not on PATH.
  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
  WARNING: The scripts import_pb_to_tensorboard, saved_model_cli, tensorboard, tf_upgrade_v2, tflite_convert and toco are installed in '/home/minimonk/.local/bin' which is not on PATH.
  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
Successfully installed MarkupSafe-3.0.2 absl-py-2.3.1 astunparse-1.6.3 flatbuffers-25.2.10 gast-0.6.0 google_pasta-0.2.0 grpcio-1.74.0 h5py-3.14.0 keras-3.11.3 libclang-18.1.1 markdown-3.9 markdown-it-py-4.0.0 mdurl-0.1.2 ml_dtypes-0.5.3 namex-0.1.0 numpy-2.2.6 opt_einsum-3.4.0 optree-0.17.0 protobuf-6.32.1 pygments-2.19.2 rich-14.1.0 tensorboard-2.20.0 tensorboard-data-server-0.7.2 tensorflow-2.20.0 termcolor-3.1.0 typing_extensions-4.15.0 werkzeug-3.1.3 wrapt-1.17.3
minimonk@minimonk-HP-EliteBook-2760p:~$ pip install keras
Defaulting to user installation because normal site-packages is not writeable
Requirement already satisfied: keras in ./.local/lib/python3.10/site-packages (3.11.3)
Requirement already satisfied: absl-py in ./.local/lib/python3.10/site-packages (from keras) (2.3.1)
Requirement already satisfied: numpy in ./.local/lib/python3.10/site-packages (from keras) (2.2.6)
Requirement already satisfied: rich in ./.local/lib/python3.10/site-packages (from keras) (14.1.0)
Requirement already satisfied: ml-dtypes in ./.local/lib/python3.10/site-packages (from keras) (0.5.3)
Requirement already satisfied: namex in ./.local/lib/python3.10/site-packages (from keras) (0.1.0)
Requirement already satisfied: optree in ./.local/lib/python3.10/site-packages (from keras) (0.17.0)
Requirement already satisfied: h5py in ./.local/lib/python3.10/site-packages (from keras) (3.14.0)
Requirement already satisfied: packaging in /usr/lib/python3/dist-packages (from keras) (21.3)
Requirement already satisfied: typing-extensions>=4.6.0 in ./.local/lib/python3.10/site-packages (from optree->keras) (4.15.0)
Requirement already satisfied: markdown-it-py>=2.2.0 in ./.local/lib/python3.10/site-packages (from rich->keras) (4.0.0)
Requirement already satisfied: pygments<3.0.0,>=2.13.0 in ./.local/lib/python3.10/site-packages (from rich->keras) (2.19.2)
Requirement already satisfied: mdurl~=0.1 in ./.local/lib/python3.10/site-packages (from markdown-it-py>=2.2.0->rich->keras) (0.1.2)

$ pip install numpy==1.26.4

 

numpy 1.26.4를 깔게 된 에러메시지

$ python3
Python 3.10.12 (main, Aug 15 2025, 14:32:43) [GCC 11.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
2025-09-15 15:28:06.544207: I external/local_xla/xla/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
/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 2.2.6
  warnings.warn(f"A NumPy version >={np_minversion} and <{np_maxversion}"

A module that was compiled using NumPy 1.x cannot be run in
NumPy 2.2.6 as it may crash. To support both 1.x and 2.x
versions of NumPy, modules must be compiled with NumPy 2.0.
Some module may need to rebuild instead e.g. with 'pybind11>=2.12'.

If you are a user of the module, the easiest solution will be to
downgrade to 'numpy<2' or try to upgrade the affected module.
We expect that some modules will need time to support NumPy 2.

Traceback (most recent call last):  File "<stdin>", line 1, in <module>
  File "/home/minimonk/.local/lib/python3.10/site-packages/tensorflow/__init__.py", line 49, in <module>
    from tensorflow._api.v2 import __internal__
  File "/home/minimonk/.local/lib/python3.10/site-packages/tensorflow/_api/v2/__internal__/__init__.py", line 13, in <module>
    from tensorflow._api.v2.__internal__ import feature_column
  File "/home/minimonk/.local/lib/python3.10/site-packages/tensorflow/_api/v2/__internal__/feature_column/__init__.py", line 8, in <module>
    from tensorflow.python.feature_column.feature_column_v2 import DenseColumn # line: 1777
  File "/home/minimonk/.local/lib/python3.10/site-packages/tensorflow/python/feature_column/feature_column_v2.py", line 38, in <module>
    from tensorflow.python.feature_column import feature_column as fc_old
  File "/home/minimonk/.local/lib/python3.10/site-packages/tensorflow/python/feature_column/feature_column.py", line 41, in <module>
    from tensorflow.python.layers import base
  File "/home/minimonk/.local/lib/python3.10/site-packages/tensorflow/python/layers/base.py", line 16, in <module>
    from tensorflow.python.keras.legacy_tf_layers import base
  File "/home/minimonk/.local/lib/python3.10/site-packages/tensorflow/python/keras/__init__.py", line 25, in <module>
    from tensorflow.python.keras import models
  File "/home/minimonk/.local/lib/python3.10/site-packages/tensorflow/python/keras/models.py", line 25, in <module>
    from tensorflow.python.keras.engine import training_v1
  File "/home/minimonk/.local/lib/python3.10/site-packages/tensorflow/python/keras/engine/training_v1.py", line 46, in <module>
    from tensorflow.python.keras.engine import training_arrays_v1
  File "/home/minimonk/.local/lib/python3.10/site-packages/tensorflow/python/keras/engine/training_arrays_v1.py", line 37, in <module>
    from scipy.sparse import issparse  # pylint: disable=g-import-not-at-top
  File "/usr/lib/python3/dist-packages/scipy/sparse/__init__.py", line 267, in <module>
    from ._csr import *
  File "/usr/lib/python3/dist-packages/scipy/sparse/_csr.py", line 10, in <module>
    from ._sparsetools import (csr_tocsc, csr_tobsr, csr_count_blocks,
AttributeError: _ARRAY_API not found

[링크 : https://mhui123.tistory.com/143]

 

그런데 mobilenet이 ssd가 없으면 classification만 되는 놈이었나?

from keras.applications import MobileNet
from keras.layers import Dense,Dropout,Activation, Flatten, GlobalAveragePooling2D
from keras.layers import Conv2D, MaxPooling2D, ZeroPadding2D
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from keras.optimizers import RMSprop, Adam
from keras.callbacks import ModelCheckpoint, EarlyStopping, ReduceLROnPlateau

img_rows,img_cols = 224,224
MobileNet = MobileNet(weights='imagenet', include_top=False, input_shape=(img_rows, img_cols, 3))
2025-09-15 16:00:19.852870: E external/local_xla/xla/stream_executor/cuda/cuda_platform.cc:51] failed call to cuInit: INTERNAL: CUDA error: Failed call to cuInit: UNKNOWN ERROR (303)
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/mobilenet/mobilenet_1_0_224_tf_no_top.h5
17225924/17225924 ━━━━━━━━━━━━━━━━━━━━ 1s 0us/step 

>>> for layer in MobileNet.layers:
...   layer.trainable = True
... 
>>> for (i, layer) in enumerate(MobileNet.layers):
...   print(str(i), layer.__class__.__name__, layer.trainable)
... 
0 InputLayer True
1 Conv2D True
2 BatchNormalization True
3 ReLU True
4 DepthwiseConv2D True
5 BatchNormalization True
6 ReLU True
7 Conv2D True
8 BatchNormalization True
9 ReLU True
10 ZeroPadding2D True
11 DepthwiseConv2D True
12 BatchNormalization True
13 ReLU True
14 Conv2D True
15 BatchNormalization True
16 ReLU True
17 DepthwiseConv2D True
18 BatchNormalization True
19 ReLU True
20 Conv2D True
21 BatchNormalization True
22 ReLU True
23 ZeroPadding2D True
24 DepthwiseConv2D True
25 BatchNormalization True
26 ReLU True
27 Conv2D True
28 BatchNormalization True
29 ReLU True
30 DepthwiseConv2D True
31 BatchNormalization True
32 ReLU True
33 Conv2D True
34 BatchNormalization True
35 ReLU True
36 ZeroPadding2D True
37 DepthwiseConv2D True
38 BatchNormalization True
39 ReLU True
40 Conv2D True
41 BatchNormalization True
42 ReLU True
43 DepthwiseConv2D True
44 BatchNormalization True
45 ReLU True
46 Conv2D True
47 BatchNormalization True
48 ReLU True
49 DepthwiseConv2D True
50 BatchNormalization True
51 ReLU True
52 Conv2D True
53 BatchNormalization True
54 ReLU True
55 DepthwiseConv2D True
56 BatchNormalization True
57 ReLU True
58 Conv2D True
59 BatchNormalization True
60 ReLU True
61 DepthwiseConv2D True
62 BatchNormalization True
63 ReLU True
64 Conv2D True
65 BatchNormalization True
66 ReLU True
67 DepthwiseConv2D True
68 BatchNormalization True
69 ReLU True
70 Conv2D True
71 BatchNormalization True
72 ReLU True
73 ZeroPadding2D True
74 DepthwiseConv2D True
75 BatchNormalization True
76 ReLU True
77 Conv2D True
78 BatchNormalization True
79 ReLU True
80 DepthwiseConv2D True
81 BatchNormalization True
82 ReLU True
83 Conv2D True
84 BatchNormalization True
85 ReLU True

>>> MobileNet.output
<KerasTensor shape=(None, 7, 7, 1024), dtype=float32, sparse=False, ragged=False, name=keras_tensor_85>
>>> MobileNet.input
<KerasTensor shape=(None, 224, 224, 3), dtype=float32, sparse=False, ragged=False, name=keras_tensor>
>>> MobileNet.summary()
Model: "mobilenet_1.00_224"
┏--------------------------------------┳-----------------------------┳-----------------┓
┃ Layer (type)                         ┃ Output Shape                ┃         Param # ┃
┡--------------------------------------╇-----------------------------╇-----------------┩
│ input_layer (InputLayer)             │ (None, 224, 224, 3)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv1 (Conv2D)                       │ (None, 112, 112, 32)        │             864 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv1_bn (BatchNormalization)        │ (None, 112, 112, 32)        │             128 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv1_relu (ReLU)                    │ (None, 112, 112, 32)        │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_1 (DepthwiseConv2D)          │ (None, 112, 112, 32)        │             288 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_1_bn (BatchNormalization)    │ (None, 112, 112, 32)        │             128 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_1_relu (ReLU)                │ (None, 112, 112, 32)        │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_1 (Conv2D)                   │ (None, 112, 112, 64)        │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_1_bn (BatchNormalization)    │ (None, 112, 112, 64)        │             256 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_1_relu (ReLU)                │ (None, 112, 112, 64)        │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pad_2 (ZeroPadding2D)           │ (None, 113, 113, 64)        │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_2 (DepthwiseConv2D)          │ (None, 56, 56, 64)          │             576 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_2_bn (BatchNormalization)    │ (None, 56, 56, 64)          │             256 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_2_relu (ReLU)                │ (None, 56, 56, 64)          │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_2 (Conv2D)                   │ (None, 56, 56, 128)         │           8,192 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_2_bn (BatchNormalization)    │ (None, 56, 56, 128)         │             512 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_2_relu (ReLU)                │ (None, 56, 56, 128)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_3 (DepthwiseConv2D)          │ (None, 56, 56, 128)         │           1,152 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_3_bn (BatchNormalization)    │ (None, 56, 56, 128)         │             512 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_3_relu (ReLU)                │ (None, 56, 56, 128)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_3 (Conv2D)                   │ (None, 56, 56, 128)         │          16,384 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_3_bn (BatchNormalization)    │ (None, 56, 56, 128)         │             512 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_3_relu (ReLU)                │ (None, 56, 56, 128)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pad_4 (ZeroPadding2D)           │ (None, 57, 57, 128)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_4 (DepthwiseConv2D)          │ (None, 28, 28, 128)         │           1,152 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_4_bn (BatchNormalization)    │ (None, 28, 28, 128)         │             512 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_4_relu (ReLU)                │ (None, 28, 28, 128)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_4 (Conv2D)                   │ (None, 28, 28, 256)         │          32,768 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_4_bn (BatchNormalization)    │ (None, 28, 28, 256)         │           1,024 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_4_relu (ReLU)                │ (None, 28, 28, 256)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_5 (DepthwiseConv2D)          │ (None, 28, 28, 256)         │           2,304 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_5_bn (BatchNormalization)    │ (None, 28, 28, 256)         │           1,024 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_5_relu (ReLU)                │ (None, 28, 28, 256)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_5 (Conv2D)                   │ (None, 28, 28, 256)         │          65,536 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_5_bn (BatchNormalization)    │ (None, 28, 28, 256)         │           1,024 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_5_relu (ReLU)                │ (None, 28, 28, 256)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pad_6 (ZeroPadding2D)           │ (None, 29, 29, 256)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_6 (DepthwiseConv2D)          │ (None, 14, 14, 256)         │           2,304 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_6_bn (BatchNormalization)    │ (None, 14, 14, 256)         │           1,024 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_6_relu (ReLU)                │ (None, 14, 14, 256)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_6 (Conv2D)                   │ (None, 14, 14, 512)         │         131,072 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_6_bn (BatchNormalization)    │ (None, 14, 14, 512)         │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_6_relu (ReLU)                │ (None, 14, 14, 512)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_7 (DepthwiseConv2D)          │ (None, 14, 14, 512)         │           4,608 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_7_bn (BatchNormalization)    │ (None, 14, 14, 512)         │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_7_relu (ReLU)                │ (None, 14, 14, 512)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_7 (Conv2D)                   │ (None, 14, 14, 512)         │         262,144 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_7_bn (BatchNormalization)    │ (None, 14, 14, 512)         │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_7_relu (ReLU)                │ (None, 14, 14, 512)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_8 (DepthwiseConv2D)          │ (None, 14, 14, 512)         │           4,608 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_8_bn (BatchNormalization)    │ (None, 14, 14, 512)         │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_8_relu (ReLU)                │ (None, 14, 14, 512)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_8 (Conv2D)                   │ (None, 14, 14, 512)         │         262,144 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_8_bn (BatchNormalization)    │ (None, 14, 14, 512)         │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_8_relu (ReLU)                │ (None, 14, 14, 512)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_9 (DepthwiseConv2D)          │ (None, 14, 14, 512)         │           4,608 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_9_bn (BatchNormalization)    │ (None, 14, 14, 512)         │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_9_relu (ReLU)                │ (None, 14, 14, 512)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_9 (Conv2D)                   │ (None, 14, 14, 512)         │         262,144 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_9_bn (BatchNormalization)    │ (None, 14, 14, 512)         │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_9_relu (ReLU)                │ (None, 14, 14, 512)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_10 (DepthwiseConv2D)         │ (None, 14, 14, 512)         │           4,608 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_10_bn (BatchNormalization)   │ (None, 14, 14, 512)         │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_10_relu (ReLU)               │ (None, 14, 14, 512)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_10 (Conv2D)                  │ (None, 14, 14, 512)         │         262,144 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_10_bn (BatchNormalization)   │ (None, 14, 14, 512)         │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_10_relu (ReLU)               │ (None, 14, 14, 512)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_11 (DepthwiseConv2D)         │ (None, 14, 14, 512)         │           4,608 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_11_bn (BatchNormalization)   │ (None, 14, 14, 512)         │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_11_relu (ReLU)               │ (None, 14, 14, 512)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_11 (Conv2D)                  │ (None, 14, 14, 512)         │         262,144 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_11_bn (BatchNormalization)   │ (None, 14, 14, 512)         │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_11_relu (ReLU)               │ (None, 14, 14, 512)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pad_12 (ZeroPadding2D)          │ (None, 15, 15, 512)         │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_12 (DepthwiseConv2D)         │ (None, 7, 7, 512)           │           4,608 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_12_bn (BatchNormalization)   │ (None, 7, 7, 512)           │           2,048 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_12_relu (ReLU)               │ (None, 7, 7, 512)           │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_12 (Conv2D)                  │ (None, 7, 7, 1024)          │         524,288 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_12_bn (BatchNormalization)   │ (None, 7, 7, 1024)          │           4,096 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_12_relu (ReLU)               │ (None, 7, 7, 1024)          │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_13 (DepthwiseConv2D)         │ (None, 7, 7, 1024)          │           9,216 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_13_bn (BatchNormalization)   │ (None, 7, 7, 1024)          │           4,096 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_dw_13_relu (ReLU)               │ (None, 7, 7, 1024)          │               0 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_13 (Conv2D)                  │ (None, 7, 7, 1024)          │       1,048,576 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_13_bn (BatchNormalization)   │ (None, 7, 7, 1024)          │           4,096 │
├--------------------------------------┼-----------------------------┼-----------------┤
│ conv_pw_13_relu (ReLU)               │ (None, 7, 7, 1024)          │               0 │
└--------------------------------------┴-----------------------------┴-----------------┘
 Total params: 3,228,864 (12.32 MB)
 Trainable params: 3,206,976 (12.23 MB)
 Non-trainable params: 21,888 (85.50 KB)

[링크 : https://kau-deeperent.tistory.com/m/59]

 

# from keras.preprocessing.image import ImageDataGenerator #  에러났음
from tensorflow.keras.preprocessing.image import ImageDataGenerator

[링크 : https://sugyeong0425.tistory.com/151]

 

voc2012 데이터셋설명

[링크 : https://bo-10000.tistory.com/38]

[링크 : https://velog.io/@kyungmin1029/CV-OpenCV]

 

2024.8 월 이니 한번 시도해볼 만할 듯?

[링크 : https://velog.io/@choonsik_mom/MobileNet-SSD-object-detector-커스텀-데이터-학습하기-m3j5d0xh]

저작자표시 (새창열림)

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

keras - transfer learning / fine tuning  (0) 2025.09.17
ssd mobilenet  (0) 2025.09.11
LLM 시각화  (0) 2025.09.06
모델 학습  (0) 2025.09.05
mean Average Precision(mAP)  (0) 2025.09.05
Posted by 구차니
프로그램 사용/yolo_tensorflow2025. 9. 11. 19:09

ssd mobilenet

최종 목표는 ssd-mobilenet 인데

어째 걸려온은건 mobilenet을 이용한 classification 같냐..

 

[링크 : https://keras.io/api/applications/mobilenet/]

[링크 : https://keras.io/api/models/model_training_apis/]

 

[링크 : https://kau-deeperent.tistory.com/m/59] << mobilenet을 이용한 감정 탐지

[링크 : https://wikidocs.net/193031] keras model.fit_generator() - deprecated -> model.fit()

 

전이학습 - 학습된 모델에 추가로 내가 원하는 클래스 추가하기

 FC(Fully conencted) layer

[링크 : https://recipesds.tistory.com/entry/Transfer-Learning을-해-보자-mobilenet을-활용해서요-실습]

 

tensorflow의 python 스크립트로 학습하기

[링크 : https://prod.velog.io/@choonsik_mom/MobileNet-SSD-object-detector-커스텀-데이터-학습하기-m3j5d0xh]

   [링크 : https://github.com/tensorflow/models/tree/master/research/object_detection]

저작자표시 (새창열림)

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

keras - transfer learning / fine tuning  (0) 2025.09.17
mobilenet 학습시키기 with keras, tensorflow  (0) 2025.09.15
LLM 시각화  (0) 2025.09.06
모델 학습  (0) 2025.09.05
mean Average Precision(mAP)  (0) 2025.09.05
Posted by 구차니
프로그램 사용/yolo_tensorflow2025. 9. 6. 23:43

LLM 시각화

보기에는 멋진데.. 이걸 어떻게 해야해야 하냐 -_ㅠ

netron 에서 보이는 신경망 하나하가 오른쪽에 보이는 그것 같은데

곱해지면서 어떻게 결과가 나오는지 이해는 여전히 안간다..(!)

 

[링크 : https://bbycroft.net/llm]

   [링크 : https://news.hada.io/topic?id=22925]

저작자표시 (새창열림)

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

mobilenet 학습시키기 with keras, tensorflow  (0) 2025.09.15
ssd mobilenet  (0) 2025.09.11
모델 학습  (0) 2025.09.05
mean Average Precision(mAP)  (0) 2025.09.05
내장 그래픽으로 ROCm?  (0) 2025.09.05
Posted by 구차니
프로그램 사용/yolo_tensorflow2025. 9. 5. 18:05

모델 학습

파이썬으로 학습을 하던 멀 하던.. 방법을 일단 근래걸로 찾아봐야겠다

 

[링크 : https://insight8094.tistory.com/entry/파이썬-딥러닝-모델-구축하는-방법-TensorFlowPyTorch]

 

생(?)으로 keras 부터 올려야 하나..

[링크 : https://kau-deeperent.tistory.com/59]

[링크 : https://keras.io/api/applications/mobilenet/]

저작자표시 (새창열림)

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

ssd mobilenet  (0) 2025.09.11
LLM 시각화  (0) 2025.09.06
mean Average Precision(mAP)  (0) 2025.09.05
내장 그래픽으로 ROCm?  (0) 2025.09.05
ai 모델들 조사  (0) 2025.09.04
Posted by 구차니
프로그램 사용/yolo_tensorflow2025. 9. 5. 12:46

mean Average Precision(mAP)

eIQ에서 학습하는데 train mean Average Precision 라고 써있어서 먼가 했는데

학습시에 얼마나 잘 배우고 있냐 라는 지표중에 하나라고 한다.

ROI가 얼마나 정확하게 모든 이미지에 대해서 측정하고 평균내면 그게 mAP 라고

 

[링크 : https://ctkim.tistory.com/entry/mAPMean-Average-Precision-정리]

[링크 : https://wikidocs.net/162466]

[링크 : https://littlefoxdiary.tistory.com/126]

 

그런데 학습시키면서 0.2 정도에서 수렴을 하면서 학습이 중단되는데 머가 잘못설정되서 그런걸까?

몇개글을 찾아봐도 eIQ에서 설정 가능한 항목들이 아닌것 같은데

python으로 직접 학습시켜봐야하나...

[링크 : https://github.com/ultralytics/ultralytics/issues/10226]

[링크 : https://stackoverflow.com/questions/78558728/yolo-training-with-low-precision-and-low-map]

[링크 : https://stackoverflow.com/questions/48595240/map-decreasing-with-training-tensorflow-object-detection-ssd]

저작자표시 (새창열림)

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

LLM 시각화  (0) 2025.09.06
모델 학습  (0) 2025.09.05
내장 그래픽으로 ROCm?  (0) 2025.09.05
ai 모델들 조사  (0) 2025.09.04
facenet  (0) 2025.08.22
Posted by 구차니
프로그램 사용/yolo_tensorflow2025. 9. 5. 12:23

내장 그래픽으로 ROCm?

tensorflow 보면 대개 nvidia만 쓰게 되어있는데, 어떻게 tensorflow에서 AMD 쓸수 있대서 찾아보는 중

[링크 : https://yhu0409.tistory.com/14]

 

nartmada
on Mar 22, 2024
Collaborator
Hi @CardLin, unfortunately 5700G APU is not a supported HW and the workaround environment flags will have performance impact. Sorry, there will be no fix in future ROCm release for this Cezanne APU issue. Closing the ticket.

[링크 : https://github.com/ROCm/ROCm/issues/2964]

 

AMD "Vega" Architecture Gets No More ROCm Updates After Release 5.6
by btarunr Jul 3rd, 2023 14:55 Discuss (42 Comments)

[링크 : https://www.techpowerup.com/310764/amd-vega-architecture-gets-no-more-rocm-updates-after-release-5-6]

 

우분투에서 ROCm 지원하다가 vega 에 대해서는 추가 지원이 없다고 한다.

현재 최신이 6.4.3인데 5.6까지 밖에 지원안한다라..(혹은 더이상의 버그 픽스만 없다면 쓸수는 있다는건가?)

[링크 : https://rocm.docs.amd.com/en/latest/release/versions.html]

 

그럼.. 이걸 해보려면 노트북을 밀어야겠...네?!

윈도우 11도 마음에 안드는데 이 기회에?!?! ㅋㅋ

저작자표시 (새창열림)

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

모델 학습  (0) 2025.09.05
mean Average Precision(mAP)  (0) 2025.09.05
ai 모델들 조사  (0) 2025.09.04
facenet  (0) 2025.08.22
NMS, soft-NMS  (0) 2024.01.15
Posted by 구차니
프로그램 사용/yolo_tensorflow2025. 9. 4. 20:13

ai 모델들 조사

1060 6GB 가 내가 가진 가장 좋은(?) 녀석이라

이것에서 돌릴만한 LLM 이나 모델들 찾아보는 중.

 

3070 6GB로 돌리는 GAN 모델

[링크 : https://marcus-story.tistory.com/170]

 

gtx1060 6GB로 돌릴만한 LLM 모델

[링크 : https://www.virtualizationhowto.com/2025/05/still-worth-it-running-ai-workloads-on-a-gtx-1060-in-2025/]

 

 

저작자표시 (새창열림)

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

mean Average Precision(mAP)  (0) 2025.09.05
내장 그래픽으로 ROCm?  (0) 2025.09.05
facenet  (0) 2025.08.22
NMS, soft-NMS  (0) 2024.01.15
VGG-16 / VGG-19  (0) 2024.01.11
Posted by 구차니
프로그램 사용/yolo_tensorflow2025. 8. 22. 15:31

facenet

코드분석하다 보니 신경망 모델중에

facenet_int_quantized.tflite 이런파일이 보여서 imagenet 처럼 머하는건가 찾아보는데

128 차원 임베딩 벡터란걸 학습시켜 비교를 한다는데

일단 코드 상으로는

결과물이 128 개의 배열로 나온다.

그리고 그걸 코사인 유사로 계산을 통해서 두개의 백터가 유사하다면 동일 인물로 추정하는 것 같은데..

 

 

[링크 : https://cake.tistory.com/44]

[링크 : https://hwangtoemat.github.io/paper-review/2020-04-02-FaceNet-%EB%82%B4%EC%9A%A9/]

[링크 : https://wikidocs.net/151188]

[링크 : https://hyunah-home.tistory.com/entry/Facenet-얼굴-인식-모델-Fine-tuning-하기]

[링크 : https://hayunjong83.tistory.com/50]

저작자표시 (새창열림)

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

내장 그래픽으로 ROCm?  (0) 2025.09.05
ai 모델들 조사  (0) 2025.09.04
NMS, soft-NMS  (0) 2024.01.15
VGG-16 / VGG-19  (0) 2024.01.11
MobileNetV2 SSD FPN-Lite  (0) 2024.01.11
Posted by 구차니
프로그램 사용/yolo_tensorflow2024. 1. 15. 11:18

NMS, soft-NMS

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]

 

저작자표시 (새창열림)

'프로그램 사용 > yolo_tensorflow' 카테고리의 다른 글

ai 모델들 조사  (0) 2025.09.04
facenet  (0) 2025.08.22
VGG-16 / VGG-19  (0) 2024.01.11
MobileNetV2 SSD FPN-Lite  (0) 2024.01.11
mobilenet v2 ssd  (0) 2024.01.11
Posted by 구차니

티스토리툴바