Model trained from blocks of size 16x16¶

Note

This model is trained using the data of size 16x16. But the evaluation results clearly proved: this model is applicable to size 32x32 and size 64x64 by using Bilinear Interpolation to do resizing for the larger blocks.

Statistics

1 global step 304857

2 batch size 128

3 samples each class 3075

4 number of classes 32

5 training time 11h 47m 50s

6 epoch 396

Epoch calculating

>>> 304857*128/3075/32.0
396.53125

Using blocks of size 16x16¶

Model Performance on Validating Dataset¶

Evaluation batch size 100, number of batches 192.

Using validating dataset, the details are documented below:

0.1 Name of dataset val_16x16.csv

0.2 Size of dataset 15.6 MB

0.3 Samples 600*32

0.4 Usage of dataset validation

1 top 5 accuracy 0.801

2 top 6 accuracy 0.842

3 top 7 accuracy 0.873

4 top 8 accuracy 0.895

5 top 9 accuracy 0.912

6 top 10 accuracy 0.924

7 top 11 accuracy 0.934

8 top 12 accuracy 0.942

9 top 16 accuracy 0.965

10 top 17 accuracy 0.970

11 top 18 accuracy 0.974

12 top 19 accuracy 0.977

13 top 20 accuracy 0.980

14 top 28 accuracy 0.995

Model Performance on Testing Dataset¶

Evaluation batch size 100, number of batches 192.

Using testing dataset, the details are documented below:

0.1 Name of dataset test_16x16.csv

0.2 Size of dataset 15.9 MB

0.3 Samples 600*32

0.4 Usage of dataset test

1 top 5 accuracy 0.739

2 top 6 accuracy 0.794

3 top 7 accuracy 0.829

4 top 8 accuracy 0.859

5 top 9 accuracy 0.877

6 top 10 accuracy 0.894

7 top 11 accuracy 0.911

8 top 12 accuracy 0.924

9 top 16 accuracy 0.958

10 top 17 accuracy 0.963

11 top 18 accuracy 0.970

12 top 19 accuracy 0.974

13 top 20 accuracy 0.977

14 top 28 accuracy 0.995

Using blocks of size 32x32¶

We have tried four resizing method:

Note

All the data for size 32x32 after pre-processing are used for evaluation. We just named it as val_32x32.csv, no need for anther test_32x32.csv.

Evaluation batch size 100, number of batches 192.

Performance with Bilinear Interpolation¶

Using validating dataset, with Bilinear interpolation, the details are documented below:

0.1 Name of dataset val_32x32.csv

0.2 Size of dataset 104.9 MB

0.3 Samples 872*32

0.4 Usage of dataset validate&test

1 top 5 accuracy 0.812

2 top 6 accuracy 0.855

3 top 7 accuracy 0.887

4 top 8 accuracy 0.908

5 top 9 accuracy 0.924

6 top 10 accuracy 0.936

7 top 11 accuracy 0.946

8 top 12 accuracy 0.954

9 top 16 accuracy 0.972

10 top 17 accuracy 0.976

11 top 18 accuracy 0.979

12 top 19 accuracy 0.982

13 top 20 accuracy 0.984

14 top 28 accuracy 0.996

Performance with Nearest Neighbor Interpolation¶

Almost the same performance as using Linear Interpolation! Omitted here for clarity.

Performance with Bicubic Interpolation¶

Almost the same performance as using Linear Interpolation! Omitted here for clarity.

Performance with Area Interpolation¶

Almost the same performance as using Linear Interpolation! Omitted here for clarity.

Using blocks of size 64x64¶

Based on the observations of the testing results of block size 32x32, we believe there should not be such differences among different interpolation method.

Here we only use Bilinear Interpolation.

Performance with Bilinear Interpolation¶

Using validating dataset, with Bilinear interpolation, the details are documented below:

Total samples: 1728

>>> 54*32
1728

Evaluation batch size 100, number of batches 17.

0.1 Name of dataset val_64x64.csv

0.2 Size of dataset 24.5 MB

0.3 Samples 54*32

0.4 Usage of dataset validate&test

1 top 5 accuracy 0.764

2 top 6 accuracy 0.821

3 top 7 accuracy 0.868

4 top 8 accuracy 0.892

5 top 9 accuracy 0.916

6 top 10 accuracy 0.932

7 top 11 accuracy 0.946

8 top 12 accuracy 0.956

9 top 16 accuracy 0.973

10 top 17 accuracy 0.979

11 top 18 accuracy 0.982

12 top 19 accuracy 0.984

13 top 20 accuracy 0.987

14 top 28 accuracy 0.994