def valid_acc(sess, valid_x, y):
div = y.shape[0]
count = 0
for i in range(div):
temp_x = valid_x[i]
#print (temp_x.shape)
rotate90 = np.array(op.rotate18(temp_x, 90))
rotate180 = np.array(op.rotate18(temp_x, 180))
rotate270 = np.array(op.rotate18(temp_x, 270))
fin_x = np.concatenate([temp_x, rotate90, rotate180, rotate270],
axis=0)
fin_x = fin_x.reshape(-1, 32, 32, 18)
#print (fin_x.shape)
y_ = sess.run(pred_number,
feed_dict={
x: fin_x,
keep_prob: 1.0,
is_training: False
})
y_ = np.argmax(np.bincount(y_))
label = np.argmax(y[i])
if (y_ == label):
count = count + 1
return count / div
y_ = graph.get_tensor_by_name("conv/dense_2/BiasAdd:0")
loss = graph.get_tensor_by_name("conv_1/Mean:0")
#train=graph.get_operation_by_name("conv_1/Adam:0")
pred_number = graph.get_tensor_by_name("ArgMax:0")
#data IO
path = '..\\dataset\\round1_test_a_20181109.h5'
h = h5py.File(path, 'r')
sen2 = h['sen2'].value
num = sen2.shape[0]
ans = np.zeros((num, 17))
for i in range(num):
img = sen2[i]
img_90 = np.array(op.rotate18(img, 90, channle=10))
img_180 = np.array(op.rotate18(img, 180, channle=10))
img_270 = np.array(op.rotate18(img, 270, channle=10))
fin_img = np.concatenate([img, img_90, img_180, img_270], axis=0)
fin_img = fin_img.reshape(-1, 32, 32, 10)
y_ = sess.run(pred_number, feed_dict={x: fin_img, keep_prob: 1.0})
y_ = np.argmax(np.bincount(y_))
ans[i][y_] = 1
np.savetxt("round1_test_a_20181109.csv", ans, delimiter=',')