Activitiez...

// keras // Object arrays cannot be loaded when allow_pickle=False

mindrefill — Wed, 4 Sep 2019 18:26:55 +0900

imdb.load_data 실행시에 아래와 같은 에러가 난다면,

Object arrays cannot be loaded when allow_pickle=False

아래 세가지 중에 한가지로 개선됨.

1. Numpy 버전을 1.16.1이나 1.16.2로 다운그레이드

2. 또는 아래와 같이 데이터 로딩할때만 defualt parameter를 수정

##############################################################################

import numpy as np

#####   Fix Numpy Error
np_load_old = np.load    # save np.load
np.load = lambda *a,**k: np_load_old(*a, allow_pickle=True, **k)    # modify the default parameters of np.load
(input_train, y_train), (input_test, y_test) = imdb.load_data(num_words = max_features)
np.load = np_load_old    # restore np.load for future normal usage
del(np_load_old)
#####   Fix Numpy Error

##############################################################################

thanks to https://stackoverflow.com/questions/55890813/how-to-fix-object-arrays-cannot-be-loaded-when-allow-pickle-false-for-imdb-loa

How to fix 'Object arrays cannot be loaded when allow_pickle=False' for imdb.load_data() function?

I'm trying to implement the binary classification example using the IMDb dataset in Google Colab. I have implemented this model before. But when I tried to do it again after a few days, it returned a

stackoverflow.com

3. keras 2.2.5 버전에서는 개선 됨

// 책 // 허교수의 ARM Mbed 프로그래밍 입문

mindrefill — Fri, 30 Aug 2019 13:27:12 +0900

Mbed로 넘어가려고 이것저것 끄적대기한 하다가 딥러닝에 정신이 팔려있었다...

오랫만에 다시 Mbed를 만져봐야겠다.

안그래도 카페에서 책 출판한다고 해서 기다리고 있던차에 꼼꼼하게 한번 봐야지...

// Keras // MNIST를 이용한 계산시간 비교

mindrefill — Thu, 8 Aug 2019 11:01:45 +0900

프랑소와 숄레의 Deep Learning with Python에 나와있는 MNIST 예제를 PC와 Colab으로 계산시간 비교

Source Code https://github.com/keras-team/keras/blob/master/examples/mnist_cnn.py

keras-team/keras

Deep Learning for humans. Contribute to keras-team/keras development by creating an account on GitHub.

github.com

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77

'''Trains a simple convnet on the MNIST dataset.
Gets to 99.25% test accuracy after 12 epochs
(there is still a lot of margin for parameter tuning).
16 seconds per epoch on a GRID K520 GPU.
'''
 
from __future__ import print_function
import keras
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten
from keras.layers import Conv2D, MaxPooling2D
from keras import backend as K
 
batch_size = 128
num_classes = 10
epochs = 12
 
# input image dimensions
img_rows, img_cols = 28, 28
 
# the data, split between train and test sets
(x_train, y_train), (x_test, y_test) = mnist.load_data()
 
if K.image_data_format() == 'channels_first':
    x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
    x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
    input_shape = (1, img_rows, img_cols)
else:
    x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
    x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
    input_shape = (img_rows, img_cols, 1)
 
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
 
# convert class vectors to binary class matrices
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
 
model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3),
                 activation='relu',
                 input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))
 
model.compile(loss=keras.losses.categorical_crossentropy,
              optimizer=keras.optimizers.Adadelta(),
              metrics=['accuracy'])
 
model.summary()
 
import time
start = time.time()  # 시작 시간 저장
 
model.fit(x_train, y_train,
          batch_size=batch_size,
          epochs=epochs,
          verbose=1,
          validation_data=(x_test, y_test))
 
score = model.evaluate(x_test, y_test, verbose=0)
 
print("time :", time.time() - start)  # 현재시각 - 시작시간 = 실행 시간
print('Test loss:', score[0])
print('Test accuracy:', score[1])
Colored by Color Scripter

cs

PC i7-7700 @ 3.60GHz, 32GB x64 (CPU)

Train on 60000 samples, validate on 10000 samples
Epoch 1/12
60000/60000 [==============================] - 71s 1ms/step - loss: 0.2752 - acc: 0.9160 - val_loss: 0.0547 - val_acc: 0.9826
Epoch 2/12
60000/60000 [==============================] - 69s 1ms/step - loss: 0.0909 - acc: 0.9730 - val_loss: 0.0396 - val_acc: 0.9859
Epoch 3/12
60000/60000 [==============================] - 69s 1ms/step - loss: 0.0676 - acc: 0.9797 - val_loss: 0.0325 - val_acc: 0.9887
Epoch 4/12
60000/60000 [==============================] - 68s 1ms/step - loss: 0.0550 - acc: 0.9838 - val_loss: 0.0300 - val_acc: 0.9901
Epoch 5/12
60000/60000 [==============================] - 68s 1ms/step - loss: 0.0476 - acc: 0.9854 - val_loss: 0.0299 - val_acc: 0.9900
Epoch 6/12
60000/60000 [==============================] - 68s 1ms/step - loss: 0.0417 - acc: 0.9876 - val_loss: 0.0273 - val_acc: 0.9911
Epoch 7/12
60000/60000 [==============================] - 69s 1ms/step - loss: 0.0365 - acc: 0.9887 - val_loss: 0.0262 - val_acc: 0.9915
Epoch 8/12
60000/60000 [==============================] - 71s 1ms/step - loss: 0.0349 - acc: 0.9889 - val_loss: 0.0333 - val_acc: 0.9885
Epoch 9/12
60000/60000 [==============================] - 69s 1ms/step - loss: 0.0327 - acc: 0.9899 - val_loss: 0.0300 - val_acc: 0.9907
Epoch 10/12
60000/60000 [==============================] - 69s 1ms/step - loss: 0.0292 - acc: 0.9907 - val_loss: 0.0271 - val_acc: 0.9911
Epoch 11/12
60000/60000 [==============================] - 71s 1ms/step - loss: 0.0279 - acc: 0.9913 - val_loss: 0.0286 - val_acc: 0.9905
Epoch 12/12
60000/60000 [==============================] - 73s 1ms/step - loss: 0.0270 - acc: 0.9920 - val_loss: 0.0277 - val_acc: 0.9912
time : 839.7007076740265
Test loss: 0.02767643234600473
Test accuracy: 0.9912

Colab (GPU)

Train on 60000 samples, validate on 10000 samples
Epoch 1/12
60000/60000 [==============================] - 5s 79us/step - loss: 0.0246 - acc: 0.9924 - val_loss: 0.0320 - val_acc: 0.9903
Epoch 2/12
60000/60000 [==============================] - 5s 76us/step - loss: 0.0238 - acc: 0.9922 - val_loss: 0.0251 - val_acc: 0.9920
Epoch 3/12
60000/60000 [==============================] - 5s 76us/step - loss: 0.0232 - acc: 0.9931 - val_loss: 0.0276 - val_acc: 0.9916
Epoch 4/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0226 - acc: 0.9931 - val_loss: 0.0293 - val_acc: 0.9917
Epoch 5/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0197 - acc: 0.9939 - val_loss: 0.0318 - val_acc: 0.9907
Epoch 6/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0199 - acc: 0.9937 - val_loss: 0.0269 - val_acc: 0.9917
Epoch 7/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0200 - acc: 0.9935 - val_loss: 0.0339 - val_acc: 0.9903
Epoch 8/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0189 - acc: 0.9938 - val_loss: 0.0279 - val_acc: 0.9922
Epoch 9/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0203 - acc: 0.9935 - val_loss: 0.0294 - val_acc: 0.9919
Epoch 10/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0177 - acc: 0.9948 - val_loss: 0.0270 - val_acc: 0.9923
Epoch 11/12
60000/60000 [==============================] - 5s 75us/step - loss: 0.0183 - acc: 0.9942 - val_loss: 0.0307 - val_acc: 0.9915
Epoch 12/12
60000/60000 [==============================] - 5s 76us/step - loss: 0.0184 - acc: 0.9941 - val_loss: 0.0261 - val_acc: 0.9919
time : 55.00918984413147
Test loss: 0.02614340115247869
Test accuracy: 0.9919

* PC 그래픽카드가 허접한거라 CPU와 Colab GPU를 비교한거지만... 확실히 GPU가 빠르긴하네...

* Colab TPU도 그닥 빠르지는 않은 듯?

// keras // Ubuntu에서 Tensorflow 설치하기

mindrefill — Wed, 7 Aug 2019 18:09:56 +0900

출처 : 유튜브 CS With James님

https://www.youtube.com/watch?v=LkpNEkQe7N0&list=PL6kYwK3mqV2tCcRdXTkM-ahjLgLUkhQXx&index=2

1. 우분투를 설치한다.

2. 설치한 우분투를 최신 업데이트 한다.

$ sudo apt-get update && sudo apt-get upgrade

3. 터미널에서

1) $ python3

2) $ sudo apt-get python3-pip // PIP 설치

3) $ pip3 // 설치된 패키지 확인

4) $ sudo pip3 install tensorflow // CPU 버전 tensorflow 설치

4. 설치가 완료되면, python3로 들어가서 import tensorflow로 설치확인

5. Jupyter Notebook 사용을 위해 설치

$ sudo pip3 jupyter

6. Jupyter Notebook 실행

$ jupyter notebook

7. Tensorflow GPU 버전 설치

1) $ sudo apt-cache search nvidia // nvidia 드라이버 최신 검색

2) $ sudo apt-get install nvidia-384

3) $ sudo chmod +x install_cuda.sh // cuda 설치 스크립트 실행

4) $ sudo ./install_cuda.sh

install_cuda.sh

0.00MB

5) cuda toolkit 9.0과 cuDNN을 자동으로 다운받아 설치해 준다

6) sudo pip3 uninstall tensorflow // tensorflow CPU용을 삭제한다

7) sudo pip3 install tensorflow-gpu // tensorflow GPU를 설치한다

Arduino IDE digitalWrite를 이용한 속도비교

mindrefill — Thu, 27 Dec 2018 13:34:50 +0900

Arduino IDE digitalWrite HIGH/LOW를 이용한 속도비교

Saleae Logic Analyser로 비교 함.

Arduino Uno : 145.5kHz

STM32duino 72MHz : 705.9kHz

STM32duino 128MHz : 1.263MHz

STM32duino 128MHz가 Uno 대비해서 8.7배, 72Mhz 대비 1.8배 빠르네~

포트를 직접제어하면 훨씬 빠르겠네.. 어쨋던 속도비교는 됨.

단적으로 모든 성능을 비교할 수는 없지만, ARM M3의 특징에 클럭까지 더해져 많이 빠른듯~

GY-61 ADXL335 Acelerometro 3-Axis Analog Output Accelerometer Module Angular Transducer 3V-5V

mindrefill — Wed, 26 Dec 2018 14:47:04 +0900

GY-61 ADXL335 Acelerometro 3-Axis Analog Output Accelerometer Module Angular Transducer 3V-5V

출력측 Bandwidth를 세팅하기 위해 Cx, Cy, Cz를 선정함.

Break-Out Board에 0.1nF이 장착되어 있음(BW 약 50kHz), 필요시 변경요함!!!

Self-Test

신호는 잘 들어온다...

ST-Link로 업로드 할 때 시리얼모니터 문제

mindrefill — Tue, 18 Dec 2018 21:40:24 +0900

// ST-Link로 프로그래밍 시에는 시리얼 모니터를 닫고 업로드 한 후, 업로드 끝나고 나면 시리얼 모니터를 열 것

// Arduino IDE에서 CPU Speed는 Overclocking 하지 말 것

#define LED_BUILTIN PC13

#define SerialMonitor Serial

int i = 0;

void setup() {

SerialMonitor.begin(115200);

pinMode(LED_BUILTIN, OUTPUT);

digitalWrite(LED_BUILTIN, HIGH);

delay(1000);

}

void loop() {

SerialMonitor.print(i++);

SerialMonitor.print(" ");

SerialMonitor.println(digitalRead(LED_BUILTIN));

digitalWrite(LED_BUILTIN, !(digitalRead(LED_BUILTIN)));

delay(500);

}

아두이노 스피드 업

mindrefill — Tue, 18 Dec 2018 11:28:21 +0900

https://stackabuse.com/speeding-up-arduino/

https://www.brightdevelopers.com/speed-arduino-code/

https://www.instructables.com/id/Arduino-is-Slow-and-how-to-fix-it/

http://forum.arduino.cc/index.php?topic=377891.0

https://www.arduino.cc/en/Reference/PortManipulation

http://skpang.co.uk/blog/archives/323

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

STM32

http://www.stm32duino.com/viewtopic.php?f=18&t=2888&p=37578#p37593

//port/gpio oriented macros

#define IO_SET(port, pins) port->regs->ODR |= (pins) //set bits on port

#define IO_CLR(port, pins) port->regs->ODR &=~(pins) //clear bits on port

//fast routines through BRR/BSRR registers

#define FIO_SET(port, pins) port->regs->BSRR = (pins)

#define FIO_CLR(port, pins) port->regs->BRR = (pins)

IO_SET/CLR 570ms (118ns/IO) 8.5cycles

FIO_SET/CLR 417ms (42ns/FIO) 3.02cycles

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

Fast Library

https://github.com/NicksonYap/digitalWriteFast

http://masteringarduino.blogspot.com/2013/10/fastest-and-smallest-digitalread-and.html

STM32F103C8T6(Blue Pill, Black Pill)를 ST-Link V2로 만들기

mindrefill — Mon, 17 Dec 2018 20:23:42 +0900

STM32F103C8T6(Blue Pill, Black Pill)를 ST-Link V2로 만들기

참조 사이트

https://microcontrollerelectronics.com/turn-an-stm32f103c8t6-blueplll-into-an-stlink-programmer/

http://slemi.info/2018/08/14/making-your-own-st-link-v2/

STM32F103C8T6에 USB-UART를 연결하고 (Vcc, GND, Rx,Tx) st.com의 Flash Downloader에서

아래 첨부파일을 STM32F103C8T6에 업로드 해 준다.

STLinkV2.J16.S4.bin

PC의 USB를 STM32F103C8T6에 연결 해주고, ST-Link 윈도우 드라이버를 깔아주면 정상적으로 인식한다.

이후, ST-Link Utility에서 ST-Link Firmware Update를 해주면 된다.

사용은 아래와 같이 하면 된다(오른쪽이 ST-Link로 사용하는 STM32F103C8T6)

중요:

1) 저항 : 220Ohm 2ea

2) PB12와 PB14를 점퍼 해줘야 한다.

STM32 Arduino IDE OverClocking Test

mindrefill — Thu, 13 Dec 2018 17:46:29 +0900

STM32 Arduino IDE OverClocking Test를 하기 위해서는

1. Arduuino - 툴 - CPU Speed - Overclocking 128MHz.. 를 선택한다.

2. Arduuino - 툴 - Upload Method는 Serial

(STM32에 부트로더에서는 안된다)

3. UPLOAD를 위해서는 USB-TTL을 이용하여 수동으로 리셋버튼을 누르며 해줘야 한다.