|
| 1 | +import tensorgraph as tg |
| 2 | +from tensorgraph.layers.backbones import * |
| 3 | +from tensorgraph.layers import Softmax, Flatten, Linear, MaxPooling |
| 4 | +import tensorflow as tf |
| 5 | +import os |
| 6 | +from tensorgraph.trainobject import train as mytrain |
| 7 | +os.environ['CUDA_VISIBLE_DEVICES'] = '-1' |
| 8 | + |
| 9 | + |
| 10 | +X_train = np.random.rand(100, 96, 96, 3) |
| 11 | +y_train = np.random.rand(100, 10) |
| 12 | +_, h, w, c = X_train.shape |
| 13 | +_, nclass = y_train.shape |
| 14 | +X_ph = tf.placeholder('float32', [None, h, w, c]) |
| 15 | +y_ph = tf.placeholder('float32', [None, nclass]) |
| 16 | + |
| 17 | + |
| 18 | +def train(seq): |
| 19 | + y_train_sb = seq.train_fprop(X_ph) |
| 20 | + y_test_sb = seq.test_fprop(X_ph) |
| 21 | + train_cost_sb = tg.cost.entropy(y_ph, y_train_sb) |
| 22 | + optimizer = tf.train.AdamOptimizer(0.0001) |
| 23 | + test_accu_sb = tg.cost.accuracy(y_ph, y_test_sb) |
| 24 | + with tf.Session() as sess: |
| 25 | + mytrain(session=sess, |
| 26 | + feed_dict={X_ph:X_train, y_ph:y_train}, |
| 27 | + train_cost_sb=train_cost_sb, |
| 28 | + valid_cost_sb=-test_accu_sb, |
| 29 | + optimizer=optimizer, |
| 30 | + epoch_look_back=5, max_epoch=1, |
| 31 | + percent_decrease=0, train_valid_ratio=[5,1], |
| 32 | + batchsize=64, randomize_split=False) |
| 33 | + |
| 34 | + |
| 35 | +def test_VGG16(): |
| 36 | + seq = tg.Sequential() |
| 37 | + vgg = VGG16(input_channels=c, input_shape=(h, w)) |
| 38 | + print('output channels:', vgg.output_channels) |
| 39 | + print('output shape:', vgg.output_shape) |
| 40 | + out_dim = np.prod(vgg.output_shape) * vgg.output_channels |
| 41 | + seq.add(vgg) |
| 42 | + seq.add(Flatten()) |
| 43 | + seq.add(Linear(int(out_dim), nclass)) |
| 44 | + seq.add(Softmax()) |
| 45 | + train(seq) |
| 46 | + |
| 47 | + |
| 48 | +def test_VGG19(): |
| 49 | + seq = tg.Sequential() |
| 50 | + vgg = VGG19(input_channels=c, input_shape=(h, w)) |
| 51 | + print('output channels:', vgg.output_channels) |
| 52 | + print('output shape:', vgg.output_shape) |
| 53 | + out_dim = np.prod(vgg.output_shape) * vgg.output_channels |
| 54 | + seq.add(vgg) |
| 55 | + seq.add(Flatten()) |
| 56 | + seq.add(Linear(int(out_dim), nclass)) |
| 57 | + seq.add(Softmax()) |
| 58 | + train(seq) |
| 59 | + |
| 60 | + |
| 61 | +def test_ResNetSmall(): |
| 62 | + seq = tg.Sequential() |
| 63 | + model = ResNetSmall(input_channels=c, input_shape=(h, w), config=[1,1]) |
| 64 | + model = ResNetBase(input_channels=c, input_shape=(h, w), config=[1,1,1,1]) |
| 65 | + print('output channels:', model.output_channels) |
| 66 | + print('output shape:', model.output_shape) |
| 67 | + seq.add(model) |
| 68 | + seq.add(MaxPooling(poolsize=tuple(model.output_shape), stride=(1,1), padding='VALID')) |
| 69 | + outshape = valid_nd(model.output_shape, kernel_size=model.output_shape, stride=(1,1)) |
| 70 | + print(outshape) |
| 71 | + out_dim = model.output_channels |
| 72 | + seq.add(Flatten()) |
| 73 | + seq.add(Linear(int(out_dim), nclass)) |
| 74 | + seq.add(Softmax()) |
| 75 | + train(seq) |
| 76 | + |
| 77 | + |
| 78 | +def test_ResNetBase(): |
| 79 | + seq = tg.Sequential() |
| 80 | + model = ResNetBase(input_channels=c, input_shape=(h, w), config=[1,1,1,1]) |
| 81 | + print('output channels:', model.output_channels) |
| 82 | + print('output shape:', model.output_shape) |
| 83 | + seq.add(model) |
| 84 | + seq.add(MaxPooling(poolsize=tuple(model.output_shape), stride=(1,1), padding='VALID')) |
| 85 | + |
| 86 | + outshape = valid_nd(model.output_shape, kernel_size=model.output_shape, stride=(1,1)) |
| 87 | + print(outshape) |
| 88 | + out_dim = model.output_channels |
| 89 | + seq.add(Flatten()) |
| 90 | + seq.add(Linear(int(out_dim), nclass)) |
| 91 | + seq.add(Softmax()) |
| 92 | + train(seq) |
| 93 | + |
| 94 | + |
| 95 | +def test_DenseNet(): |
| 96 | + seq = tg.Sequential() |
| 97 | + model = DenseNet(input_channels=c, input_shape=(h, w), ndense=1, growth_rate=1, nlayer1blk=1) |
| 98 | + print('output channels:', model.output_channels) |
| 99 | + print('output shape:', model.output_shape) |
| 100 | + out_dim = np.prod(model.output_shape) * model.output_channels |
| 101 | + seq.add(model) |
| 102 | + seq.add(MaxPooling(poolsize=tuple(model.output_shape), stride=(1,1), padding='VALID')) |
| 103 | + seq.add(Flatten()) |
| 104 | + seq.add(Linear(model.output_channels, nclass)) |
| 105 | + seq.add(Softmax()) |
| 106 | + train(seq) |
| 107 | + |
| 108 | + |
| 109 | +def test_UNet(): |
| 110 | + seq = tg.Sequential() |
| 111 | + model = UNet(input_channels=c, input_shape=(h, w)) |
| 112 | + print('output channels:', model.output_channels) |
| 113 | + print('output shape:', model.output_shape) |
| 114 | + out_dim = np.prod(model.output_shape) * model.output_channels |
| 115 | + seq.add(model) |
| 116 | + seq.add(MaxPooling(poolsize=tuple(model.output_shape), stride=(1,1), padding='VALID')) |
| 117 | + seq.add(Flatten()) |
| 118 | + seq.add(Linear(model.output_channels, nclass)) |
| 119 | + seq.add(Softmax()) |
| 120 | + train(seq) |
| 121 | + |
| 122 | + |
| 123 | +if __name__ == '__main__': |
| 124 | + print('runtime test') |
| 125 | + test_VGG16() |
| 126 | + print('..VGG16 running test done') |
| 127 | + test_VGG19() |
| 128 | + print('..VGG19 running test done') |
| 129 | + test_ResNetSmall() |
| 130 | + print('..ResNetSmall running test done') |
| 131 | + test_ResNetBase() |
| 132 | + print('..ResNetBase running test done') |
| 133 | + test_DenseNet() |
| 134 | + print('..DenseNet running test done') |
| 135 | + test_UNet() |
| 136 | + print('..UNet running test done') |
0 commit comments