tf_DenseNet
Tensorflow implementation of DenseNet for general classification problems
Usage
Initialization
"""
Args:
sess :
TensorFlow session
c :
An integer.
The number of class to classify.
Optional args:
is_train :
A bool.
Whether training mode or not.
The default is True.
im_shape :
A 3-length list.
The defualt is [224, 224, 3]
k :
An integer.
Growth rate of dense blocks.
l :
An integer or a 4-length list.
How many layers for each block.
The default is [6, 12, 24, 16].
keep_prop :
Keep proportion of DropOut.
The default option is 'not using DropOut'
optim :
TensorFlow Optimizer
The default is AdamOptimizer(learning_rate=1e-4)
"""
with tf.Session as sess:
...
densenet = DenseNet(sess, c)
...Train & Prediction
"""
Args :
image :
An np.float32 numpy-array with shape [N, H, W, C].
[H, W, C] should be same with the given im_shape.
label :
An np.int64 numpy-array with shape [N].
It represents the class of the image from 0 to c-1.
Return :
loss :
The mean cross-entropy loss of the input batch.
acc :
The mean accurary of the input batch.
label :
An np.uint64 numpy-array with shape [N]
Predicted integer values from 0 to c-1.
"""
loss, acc = densenet.fit(image, label)
label = densenet.predict(image)Test
MNIST example
python test.py