Comparsion of TensorFlow Wrappers
Run Keras, TensorLayer and Tflearn with same model and data on a same GPU machine.
The parameter initialization may have slightly different, but would not effect the speed.
Feel free to PUSH !
Speed of MLP
GPU: GTX980
TensorFlow: r0.10
Data: MNIST train:50k val:10k test:10k
Model: 784-800-800-10
Num of epochs: 200
Batch size: 500
Keras: 282.475250s = 1.41 s/epoch
TensorLayer: 116.670947s = 0.58 s/epoch
Tflearn:
Arcyfelix's test
Setup
GPU: GTX970
Driver Version: 375.39
TensorFlow: 1.0.1
Data: MNIST train: 50k val: 10k test: 10k
Speed of MLP
Num of epochs: 200 Batch size: 500 FC-x = Fully Connected / Dense with Relu activation with x number of neurons DP = Dropout
| Architecture / Library | Keras | TFLearn | TensorLayer |
|---|---|---|---|
| INPUT + FC-800 + DP + FC-800 + DP + OUTPUT | 173.825s | 337.312s | To be tested |
| INPUT + FC-2000 + DP + FC-2000 + DP + OUTPUT | 377.443s | 477.034s | To be tested |
| INPUT + FC-4000 + DP + FC-4000 + DP + OUTPUT | 1007.613s | 872.662s | To be tested |
| INPUT + FC-4000 + DP + FC-4000 + FC-4000 + DP + OUTPUT | 1715.068s | 1313.363s | To be tested |
Speed of CNN
Num of epochs: 20 Batch size: 100 Conv2d[kernel-x, kernel-y]-filters = Convolutional layer with padding = 'same'
| Architecture / Library | Keras | TFLearn | TensorLayer |
|---|---|---|---|
| INPUT + Conv2d[3,3]-8 + Conv2d[3,3]-8 + FC-100 + DP + FC-100 + DP + OUTPUT | 79.999s | 84.487s | To be tested |
| INPUT + Conv2d[3,3]-32 + Conv2d[3,3]-32 + FC-100 + DP + FC-100 + DP + OUTPUT | 132.741s | 125.306s | To be tested |
| INPUT + Conv2d[3,3]-64 + Conv2d[3,3]-64 + FC-100 + DP + FC-100 + DP + OUTPUT | 230.574s | 204.685s | To be tested |
| INPUT + Conv2d[3,3]-128 + Conv2d[3,3]-128 + FC-100 + DP + FC-100 + DP + OUTPUT | 477.009s | 407.489s | To be tested |
| INPUT + Conv2d[3,3]-256 + Conv2d[3,3]-256 + FC-100 + DP + FC-100 + DP + OUTPUT | 1186.775s | 1037.954s | To be tested |