def dense_block(x, size):
input_size, output_size = size
w = tf.Variable(init_matrix(size=size, init='alexnet'), dtype=tf.float32)
b = tf.Variable(np.zeros(shape=output_size), dtype=tf.float32)
fc = tf.matmul(x, w) + b
return fc
def dense(x, size, name):
input_size, output_size = size
w = tf.Variable(init_matrix(size=size, init='alexnet'),
dtype=tf.float32,
name=name)
b = tf.Variable(np.zeros(shape=output_size),
dtype=tf.float32,
name=name + '_bias')
fc = tf.matmul(x, w) + b
return fc
def __init__(self, isize, osize):
self.layer_id = layer.layer_id
layer.layer_id += 1
self.isize = isize
self.osize = osize
self.w = tf.Variable(init_matrix(size=(self.isize, self.osize),
init='glorot_uniform'),
dtype=tf.float32)
self.b = tf.Variable(np.zeros(shape=(self.osize)),
dtype=tf.float32,
trainable=False)
def dense(x, size, weights, name):
_, output_size = size
if weights:
w = tf.Variable(weights[name], dtype=tf.float32)
b = tf.Variable(weights[name + '_bias'], dtype=tf.float32)
assert (np.shape(w) == size)
assert (np.shape(b)[0] == output_size)
else:
w = tf.Variable(init_matrix(size=size, init='glorot_uniform'),
dtype=tf.float32)
b = tf.Variable(np.zeros(output_size), dtype=tf.float32)
dot = tf.matmul(x, w) + b
return dot
def dense(x, size, w, name):
input_size, output_size = size
trainable = (w == None)
if w is not None:
print('loading %s | trainable %d ' % (name, trainable))
weights_np = w[name]
bias_np = w[name + '_bias']
else:
print('making %s | trainable %d ' % (name, trainable))
weights_np = init_matrix(size=size, init='glorot_uniform')
bias_np = np.zeros(shape=output_size)
w = tf.Variable(weights_np, dtype=tf.float32, trainable=trainable)
b = tf.Variable(bias_np, dtype=tf.float32, trainable=trainable)
out = tf.matmul(x, w) + b
return out
############################################################### dropout_rate = tf.placeholder(tf.float32, shape=()) learning_rate = tf.placeholder(tf.float32, shape=()) block1 = block(features, 3, 64, 2, 'block1') # 64 block2 = block(block1, 64, 128, 2, 'block2') # 32 block3 = block(block2, 128, 256, 2, 'block3') # 16 block4 = block(block3, 256, 512, 2, 'block4') # 8 block5 = block(block4, 512, 1024, 1, 'block5') # 4 pool = tf.nn.avg_pool(block5, ksize=[1,4,4,1], strides=[1,4,4,1], padding='SAME') # 1 flat = tf.reshape(pool, [args.batch_size, 1024]) mat1 = tf.Variable(init_matrix(size=(1024, 1000), init='alexnet'), dtype=tf.float32, name='fc1') bias1 = tf.Variable(np.zeros(shape=1000), dtype=tf.float32, name='fc1_bias') fc1 = tf.matmul(flat, mat1) + bias1 ############################################################### loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=fc1, labels=labels)) correct = tf.equal(tf.argmax(fc1, axis=1), tf.argmax(labels, 1)) total_correct = tf.reduce_sum(tf.cast(correct, tf.float32)) train = tf.train.AdamOptimizer(learning_rate=learning_rate, epsilon=args.eps).minimize(loss) ############################################################### config = tf.ConfigProto(allow_soft_placement=True) config.gpu_options.allow_growth=True
