def test_conv_module():
img_height = 4
img_width = 4
x = tf.constant(np.array([[1, 2, 3, 4],
[4, 3, 2, 1],
[5, 6, 7, 8],
[1, 2, 3, 4]]), dtype='float32')
# 1 * img_height * img_width * 1
x = tf.expand_dims(x, 0)
filter_size = 2
depth = 2
stride = 2
conv = modeling.conv(x, filter_size, depth, stride)
conv_shape = conv.get_shape().as_list()
assert conv_shape == [1,
math.ceil(float(img_height) / stride),
math.ceil(float(img_width) / stride),
depth]
init()
conv_val = sess.run(conv)
assert conv_val.shape == (1,
math.ceil(float(img_height) / stride),
math.ceil(float(img_width) / stride),
depth)
def logits(x):
hidden_sizes = [500, 350, 250, 230]
with tf.name_scope('model'):
x = tf.reshape(x, [data.BATCH_SIZE, data.IMG_HEIGHT, data.IMG_WIDTH])
conv1 = modeling.conv(x, 5, 32, 2, name='conv')
conv_shape = conv1.get_shape().as_list()
# Flatten for fully-connected layers
input_t = tf.reshape(
conv1,
[conv_shape[0],
conv_shape[1] * conv_shape[2] * conv_shape[3]])
for i, hsize in enumerate(hidden_sizes):
layer_no = i + 1
input_t = modeling.hidden_layer('hidden%d' % layer_no,
input_t, hsize)
last_hidden = input_t
logits = modeling.linear_softmax(last_hidden, NUM_CLASSES)
logits = tf.identity(logits, name='logits')
return logits
