benchmark: GRU+SVM for MNIST dataset about fashion-mnist

[AFAgarap]

Using the same GRU+SVM model, but now, with Dropout, here's the result on MNIST dataset:

Epoch : 0 completed out of 10, loss : 183.0938720703125, accuracy : 0.87109375
Epoch : 1 completed out of 10, loss : 83.48553466796875, accuracy : 0.953125
Epoch : 2 completed out of 10, loss : 96.95372009277344, accuracy : 0.94921875
Epoch : 3 completed out of 10, loss : 36.80517578125, accuracy : 0.984375
Epoch : 4 completed out of 10, loss : 57.431053161621094, accuracy : 0.97265625
Epoch : 5 completed out of 10, loss : 31.22872543334961, accuracy : 0.98046875
Epoch : 6 completed out of 10, loss : 33.04032516479492, accuracy : 0.98046875
Epoch : 7 completed out of 10, loss : 31.262266159057617, accuracy : 0.984375
Epoch : 8 completed out of 10, loss : 20.89887046813965, accuracy : 0.98828125
Epoch : 9 completed out of 10, loss : 26.805370330810547, accuracy : 0.98828125
Accuracy : 0.9761001467704773

The following were the hyper-parameters used:

BATCH_SIZE = 256
CELL_SIZE = 256
DROPOUT_P_KEEP = 0.85
EPOCHS = 10
LEARNING_RATE = 1e-3
SVM_C = 1

Trained using tf.train.AdamOptimizer(), and used tf.nn.dynamic_rnn(). The source may be found here.