lychrel / tensorflow-cifar-10

Cifar-10 CNN implementation using TensorFlow library with 24% error.

tensorflow-cifar-10

Cifar-10 convolutional network implementation example using TensorFlow library.

Requirement

Library	Version
Python	^3.5
Tensorflow	^1.0.1
Numpy	^1.12.0
Pickle	*

Usage

Download code:

git clone https://github.com/exelban/tensorflow-cifar-10

cd tensorflow-cifar-10

Train cnn:

Batch size: 128

After every 1000 iteration making prediction on testing batch.

10000 iteration take about 50min on NVIDIA K10 GPU (g2.2xlarge) or 30min on NVIDIA K80 (p2.xlarge).

python3 train.py

Example output:

Trying to restore last checkpoint ...
Restored checkpoint from: ./tensorboard/cifar-10/-20000
Global Step:  9910, accuracy: 100.0%, loss = 0.04 (928.6 examples/sec, 0.09 sec/batch)
Global Step:  9920, accuracy: 100.0%, loss = 0.02 (931.4 examples/sec, 0.09 sec/batch)
Global Step:  9930, accuracy: 100.0%, loss = 0.01 (928.0 examples/sec, 0.09 sec/batch)
Global Step:  9940, accuracy:  98.4%, loss = 0.04 (927.3 examples/sec, 0.09 sec/batch)
Global Step:  9950, accuracy:  98.4%, loss = 0.01 (930.1 examples/sec, 0.09 sec/batch)
Global Step:  9960, accuracy: 100.0%, loss = 0.02 (941.0 examples/sec, 0.10 sec/batch)
Global Step:  9970, accuracy: 100.0%, loss = 0.01 (936.6 examples/sec, 0.10 sec/batch)
Global Step:  9980, accuracy:  98.4%, loss = 0.05 (928.1 examples/sec, 0.09 sec/batch)
Global Step:  9990, accuracy:  99.2%, loss = 0.01 (928.4 examples/sec, 0.09 sec/batch)
Global Step:  10000, accuracy: 100.0%, loss = 0.00 (926.6 examples/sec, 0.09 sec/batch)
Accuracy on Test-Set: 76.23% (7623 / 10000)
Saved checkpoint.

Make prediction:

python3 predict.py

Example output:

Trying to restore last checkpoint ...
Restored checkpoint from: ./tensorboard/cifar-10/-20000
Accuracy on Test-Set: 75.73% (7573 / 10000)
[848   9  42  12  16   3   8   8  38  16] (0) airplane
[ 21 841   7   6   1   8   5   1  35  75] (1) automobile
[ 55   2 720  47  78  29  26  26   6  11] (2) bird
[ 33  10  83 587  74 118  47  24   8  16] (3) cat
[ 18   0  89  56 755  16  18  40   7   1] (4) deer
[ 18   5  77 194  58 581  15  40   4   8] (5) dog
[ 15   4  65  69  39  18 771   6   8   5] (6) frog
[ 23   0  36  36  75  30   3 789   1   7] (7) horse
[ 61  18  10   9   8   6   6   2 858  22] (8) ship
[ 41  70  10  14   3   4   2   6  27 823] (9) truck
 (0) (1) (2) (3) (4) (5) (6) (7) (8) (9)

Tensorboard

tensorboard --logdir tensorboard

Model

Convolution layer 1
Conv_2d
ReLu
MaxPool
LRN
Convolution layer 2
Conv_2d
ReLu
LRN
MaxPool
Convolution layer 3
Conv_2d
ReLu
Convolution layer 4
Conv_2d
ReLu
Convolution layer 5
Conv_2d
ReLu
LRN
MaxPool
Fully connected 1
Fully connected 2
Softmax_linear

What's new

v0.0.1

- Make tests on AWS instances;
- Model fixes;
- Remove cifar-100 dataset;

v0.0.0

- First release

License

Apache License 2.0