電腦硬體
處理器(CPU)：Intel® Core™ i5-10300H CPU @ 2.50GHz
顯示晶片(GPU)：NVIDIA® GeForce GTX 1650 Ti
Python程式環境
- Tensorflow-gpu-1.15.0
- Keras 2.3.1
MNIST手寫數據集
mnist為手寫數字0-9的黑白圖片數據集。
圖片與對應標籤方面，數量為training data：55000筆，validation data：5000; 僅有圖片的test data：10000筆; 每張黑白圖片大小為[1,28,28,1]（張,寬,長,通道），有784個像素(pixels)，每個像素值都是數字0-255（0黑255白）。
標籤與圖片的手寫數字對應，如圖片內為手寫數字5，標籤就設5，0-9種手寫圖片就有10個標籤，可用list儲存，如圖片1,5,8,6,4對應標籤就是label=[1,5,8,6,4]
虛擬碼
- 下載mnist數據集，分成training data 跟test data
- 先處理圖片，將其轉換成列表形式，並透過標準化將所有像素值變成0-1之間
- 標籤處理成one hot encoding 形式
- 建立CNN網路
- 編譯模型
- 訓練模型
- 衡量模型訓練誤差
CNN網路架構
問題
因為下載的是Tensorflow-gpu-1.15.0，相對的版本使用Keras 2.3.1，若直接跑Keras寫的CNN程式碼，會產生以下錯誤(部份節錄)
Epoch 1/1
2020-08-14 18:01:49.576948: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcublas.so.10.0
2020-08-14 18:01:49.775817: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudnn.so.7
2020-08-14 18:01:50.056149: E tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR
2020-08-14 18:01:50.062654: E tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR
Traceback (most recent call last):
File “test.py", line 49, in
model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs, verbose=1)
File “/home/leodflag5/.local/lib/python3.6/site-packages/keras/engine/training.py", line 1239, in fit
validation_freq=validation_freq)
File “/home/leodflag5/.local/lib/python3.6/site-packages/keras/engine/training_arrays.py", line 196, in fit_loop
outs = fit_function(ins_batch)
File “/home/leodflag5/.local/lib/python3.6/site-packages/tensorflow_core/python/keras/backend.py", line 3476, in call
run_metadata=self.run_metadata)
File “/home/leodflag5/.local/lib/python3.6/site-packages/tensorflow_core/python/client/session.py", line 1472, in call
run_metadata_ptr)
tensorflow.python.framework.errors_impl.UnknownError: 2 root error(s) found.
(0) Unknown: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.
[[{{node conv2d_1/convolution}}]]
(1) Unknown: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.
[[{{node conv2d_1/convolution}}]]
[[metrics/accuracy/Identity/_91]]
0 successful operations.
0 derived errors ignored.
關閉GPU，使用CPU運算
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"使用GPU運算
import tensorflow as tf
config = tf.ConfigProto()
config.gpu_options.allow_growth=True
session=tf.InteractiveSession(config=config)CPU運算時間
GPU 運算時間
大約差4倍，GPU還有使用CUDA加速運算
LEODFLAG's Blog 