benchmark: update on GRU+SVM with Dropout

[AFAgarap]

Hey @hanxiao , it's me again. I saw an update in the dataset, regarding duplicate samples. I did another training using my GRU+SVM (with Dropout) model (from #8 ) on the updated dataset. Here's the result:

Epoch : 0 completed out of 100, loss : 316.9036560058594, accuracy : 0.734375
Epoch : 1 completed out of 100, loss : 201.2646026611328, accuracy : 0.83984375
Epoch : 2 completed out of 100, loss : 253.3709259033203, accuracy : 0.796875
Epoch : 3 completed out of 100, loss : 257.7744140625, accuracy : 0.8359375
Epoch : 4 completed out of 100, loss : 179.52682495117188, accuracy : 0.8671875
Epoch : 5 completed out of 100, loss : 224.97421264648438, accuracy : 0.83984375
Epoch : 6 completed out of 100, loss : 212.19381713867188, accuracy : 0.859375
Epoch : 7 completed out of 100, loss : 200.80978393554688, accuracy : 0.859375
Epoch : 8 completed out of 100, loss : 187.77052307128906, accuracy : 0.85546875
Epoch : 9 completed out of 100, loss : 190.96389770507812, accuracy : 0.86328125
Epoch : 10 completed out of 100, loss : 185.72314453125, accuracy : 0.85546875
Epoch : 11 completed out of 100, loss : 189.3765411376953, accuracy : 0.8515625
Epoch : 12 completed out of 100, loss : 130.086669921875, accuracy : 0.89453125
Epoch : 13 completed out of 100, loss : 151.38232421875, accuracy : 0.8828125
Epoch : 14 completed out of 100, loss : 159.71595764160156, accuracy : 0.88671875
Epoch : 15 completed out of 100, loss : 218.80592346191406, accuracy : 0.84375
Epoch : 16 completed out of 100, loss : 131.5895233154297, accuracy : 0.9140625
Epoch : 17 completed out of 100, loss : 162.96995544433594, accuracy : 0.8671875
Epoch : 18 completed out of 100, loss : 155.52630615234375, accuracy : 0.890625
Epoch : 19 completed out of 100, loss : 159.76901245117188, accuracy : 0.88671875
Epoch : 20 completed out of 100, loss : 137.74642944335938, accuracy : 0.890625
Epoch : 21 completed out of 100, loss : 162.48875427246094, accuracy : 0.890625
Epoch : 22 completed out of 100, loss : 179.6526336669922, accuracy : 0.8828125
Epoch : 23 completed out of 100, loss : 127.58981323242188, accuracy : 0.8984375
Epoch : 24 completed out of 100, loss : 185.6982421875, accuracy : 0.8671875
Epoch : 25 completed out of 100, loss : 159.8983612060547, accuracy : 0.8828125
Epoch : 26 completed out of 100, loss : 160.69525146484375, accuracy : 0.89453125
Epoch : 27 completed out of 100, loss : 173.42813110351562, accuracy : 0.859375
Epoch : 28 completed out of 100, loss : 166.0702667236328, accuracy : 0.87890625
Epoch : 29 completed out of 100, loss : 157.59085083007812, accuracy : 0.87109375
Epoch : 30 completed out of 100, loss : 127.72993469238281, accuracy : 0.9140625
Epoch : 31 completed out of 100, loss : 136.65415954589844, accuracy : 0.90234375
Epoch : 32 completed out of 100, loss : 172.4806365966797, accuracy : 0.8515625
Epoch : 33 completed out of 100, loss : 139.81488037109375, accuracy : 0.8984375
Epoch : 34 completed out of 100, loss : 144.55099487304688, accuracy : 0.85546875
Epoch : 35 completed out of 100, loss : 122.90949249267578, accuracy : 0.8984375
Epoch : 36 completed out of 100, loss : 150.0441131591797, accuracy : 0.890625
Epoch : 37 completed out of 100, loss : 153.2085723876953, accuracy : 0.88671875
Epoch : 38 completed out of 100, loss : 143.91455078125, accuracy : 0.8984375
Epoch : 39 completed out of 100, loss : 117.63712310791016, accuracy : 0.91796875
Epoch : 40 completed out of 100, loss : 93.80998229980469, accuracy : 0.92578125
Epoch : 41 completed out of 100, loss : 136.52537536621094, accuracy : 0.87109375
Epoch : 42 completed out of 100, loss : 137.24530029296875, accuracy : 0.90625
Epoch : 43 completed out of 100, loss : 108.73893737792969, accuracy : 0.921875
Epoch : 44 completed out of 100, loss : 106.48686218261719, accuracy : 0.9296875
Epoch : 45 completed out of 100, loss : 104.41219329833984, accuracy : 0.92578125
Epoch : 46 completed out of 100, loss : 101.19454956054688, accuracy : 0.94140625
Epoch : 47 completed out of 100, loss : 127.536376953125, accuracy : 0.91015625
Epoch : 48 completed out of 100, loss : 109.94172668457031, accuracy : 0.9296875
Epoch : 49 completed out of 100, loss : 85.25288391113281, accuracy : 0.94140625
Epoch : 50 completed out of 100, loss : 112.01800537109375, accuracy : 0.91796875
Epoch : 51 completed out of 100, loss : 107.6760482788086, accuracy : 0.91015625
Epoch : 52 completed out of 100, loss : 121.9848403930664, accuracy : 0.921875
Epoch : 53 completed out of 100, loss : 101.01953887939453, accuracy : 0.9375
Epoch : 54 completed out of 100, loss : 69.95838165283203, accuracy : 0.94921875
Epoch : 55 completed out of 100, loss : 119.3257827758789, accuracy : 0.91796875
Epoch : 56 completed out of 100, loss : 102.73481750488281, accuracy : 0.921875
Epoch : 57 completed out of 100, loss : 89.11821746826172, accuracy : 0.94921875
Epoch : 58 completed out of 100, loss : 110.71992492675781, accuracy : 0.9140625
Epoch : 59 completed out of 100, loss : 105.85194396972656, accuracy : 0.9375
Epoch : 60 completed out of 100, loss : 114.6805648803711, accuracy : 0.921875
Epoch : 61 completed out of 100, loss : 99.33323669433594, accuracy : 0.92578125
Epoch : 62 completed out of 100, loss : 128.26809692382812, accuracy : 0.90625
Epoch : 63 completed out of 100, loss : 117.59638214111328, accuracy : 0.9140625
Epoch : 64 completed out of 100, loss : 86.27313995361328, accuracy : 0.9453125
Epoch : 65 completed out of 100, loss : 114.16581726074219, accuracy : 0.92578125
Epoch : 66 completed out of 100, loss : 102.78227233886719, accuracy : 0.94921875
Epoch : 67 completed out of 100, loss : 88.23193359375, accuracy : 0.9375
Epoch : 68 completed out of 100, loss : 60.24769592285156, accuracy : 0.953125
Epoch : 69 completed out of 100, loss : 97.67103576660156, accuracy : 0.94140625
Epoch : 70 completed out of 100, loss : 86.58494567871094, accuracy : 0.91796875
Epoch : 71 completed out of 100, loss : 98.33272552490234, accuracy : 0.921875
Epoch : 72 completed out of 100, loss : 77.44849395751953, accuracy : 0.94921875
Epoch : 73 completed out of 100, loss : 114.52888488769531, accuracy : 0.9296875
Epoch : 74 completed out of 100, loss : 94.6647720336914, accuracy : 0.9453125
Epoch : 75 completed out of 100, loss : 106.62199401855469, accuracy : 0.921875
Epoch : 76 completed out of 100, loss : 116.0970230102539, accuracy : 0.91015625
Epoch : 77 completed out of 100, loss : 78.5435791015625, accuracy : 0.953125
Epoch : 78 completed out of 100, loss : 125.43787384033203, accuracy : 0.91796875
Epoch : 79 completed out of 100, loss : 112.84344482421875, accuracy : 0.9296875
Epoch : 80 completed out of 100, loss : 65.7440185546875, accuracy : 0.95703125
Epoch : 81 completed out of 100, loss : 115.66653442382812, accuracy : 0.91796875
Epoch : 82 completed out of 100, loss : 76.14566040039062, accuracy : 0.9375
Epoch : 83 completed out of 100, loss : 72.91943359375, accuracy : 0.95703125
Epoch : 84 completed out of 100, loss : 56.55884552001953, accuracy : 0.95703125
Epoch : 85 completed out of 100, loss : 87.09599304199219, accuracy : 0.93359375
Epoch : 86 completed out of 100, loss : 80.97771453857422, accuracy : 0.93359375
Epoch : 87 completed out of 100, loss : 94.14187622070312, accuracy : 0.9453125
Epoch : 88 completed out of 100, loss : 80.44708251953125, accuracy : 0.94140625
Epoch : 89 completed out of 100, loss : 52.18363952636719, accuracy : 0.96875
Epoch : 90 completed out of 100, loss : 93.15214538574219, accuracy : 0.9296875
Epoch : 91 completed out of 100, loss : 97.51387023925781, accuracy : 0.9296875
Epoch : 92 completed out of 100, loss : 82.44243621826172, accuracy : 0.9375
Epoch : 93 completed out of 100, loss : 60.52445983886719, accuracy : 0.96484375
Epoch : 94 completed out of 100, loss : 57.100406646728516, accuracy : 0.96484375
Epoch : 95 completed out of 100, loss : 89.62207794189453, accuracy : 0.94140625
Epoch : 96 completed out of 100, loss : 86.14447784423828, accuracy : 0.9375
Epoch : 97 completed out of 100, loss : 75.90823364257812, accuracy : 0.953125
Epoch : 98 completed out of 100, loss : 65.80587768554688, accuracy : 0.9609375
Epoch : 99 completed out of 100, loss : 114.98580169677734, accuracy : 0.92578125
Accuracy : 0.897300124168396

The hyper-parameters used were as follows:

BATCH_SIZE = 256
CELL_SIZE = 256
DROPOUT_P_KEEP = 0.85
EPOCHS = 100
LEARNING_RATE = 1e-3
SVM_C = 1

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

The graph from TensorBoard, tracking the training (accuracy at the top, loss at the bottom):

The improved accuracy may not be too much, but I suppose it's still a considerable difference, i.e. ~85.5% v. ~89.7%.

[hanxiao]

Thanks again for contributing the benchmark, I saw three PRs but they all look same to me.

[AFAgarap]

Thank you as well. :) I updated the READMEs myself.