image classification benchmark and finetune

[jacquesqiao]

project: #6024

[jacquesqiao]

Info

1 pass

Paddle fluid:

Total examples: 1120, total time: 162.26692
6.90221 examples/sec, 5.07084 sec/batch
pass_acc: [ 0.07985038]

Tensorflow

Total examples: 1120, total time: 69.05409
16.21917 examples/sec, 2.15794 sec/batch

[jacquesqiao]

Info

4 pass

PaddlePaddle fluid

pass_acc: [ 0.27012524]
Total examples: 1120, total time: 608.13095
1.84171 examples/sec, 19.00409 sec/batch

Tensorflow

pass=3, batch=192, loss=1.980749, acc=0.400000

Total examples: 1120, total time: 236.70867
4.73155 examples/sec, 7.39715 sec/batch

[qingqing01]

@jacquesqiao I tested PaddlePaddle fluid in TITAN X (Pascal) by using cuDNN and removing the stream synchronization between each operator. The total time of one pass of PaddlePaddle fluid is 70.47459 sec. From your testing, the time of TensorFlow is 69.05409 sec.

Config:

• 1 pass
• batch_size: 32
• Singe GPU
• cuDNN 5.1

Time:

Total examples: 6149, total time: 70.47459
87.25131 examples/sec, 0.36515 sec/batch

[jacquesqiao]

@qingqing01 cool! I will also have a try~

[dzhwinter]

If we use the Finish implement the synchronized training. That's really hurt performance a lot when doing the training. How about put Finish barrier at the executor finish line?

[jacquesqiao]

condition:

• use a fixed numpy as data.
• drop the first 10 iteration

result:

• Fluid:
  Total examples: 2240, total time: 16.79173
  133.39901 examples/sec, 0.23988 sec/batch

• Tensorflow:
  Total examples: 2240, total time: 15.99317
  140.05980 examples/sec, 0.22847 sec/batch