def test_softmax(self):
self.assertCountEqual(common_tool.softmax([1, 2, 3, 4, 5]),
self._softmax([1, 2, 3, 4, 5]))
self.assertCountEqual(common_tool.softmax([0, 0, 0, 0, 0]),
self._softmax([0, 0, 0, 0, 0]))
self.assertCountEqual(common_tool.softmax([-1, -2, -3, -4, -5]),
self._softmax([-1, -2, -3, -4, -5]))
def test_softmax_error(self):
self.assertRaises(ValueError, lambda: common_tool.softmax([]))
self.assertRaises(ValueError, lambda: common_tool.softmax(()))
self.assertRaises(TypeError, lambda: common_tool.softmax({}))
self.assertRaises(TypeError, lambda: common_tool.softmax(None))
self.assertRaises(TypeError, lambda: common_tool.softmax(""))
def infer(model, infer_data, ckpt_file):
tf.reset_default_graph()
with tf.Session(graph=model.graph) as sess_predict:
model.saver.restore(sess_predict, ckpt_file)
output = sess_predict.run([model.y], feed_dict={model.x: infer_data})
classification = common.softmax(output)
y = common.onehot_decoding(classification)
return y
def test_infer_c(self):
dll = conv_infer_c.init(self.so_lib_path, self.input_info,
self.output_info)
res = []
for i in prepare_infer_dataset():
output = numpy.zeros(dtype=numpy.float32, shape=(1, 10))
conv_infer_c.infer_c(dll,
numpy.expand_dims(i, 0).astype(numpy.float32),
output)
classification = common_tool.softmax(output)
y = common_tool.onehot_decoding(classification)
res.append(y[0])
self.assertCountEqual(res, [7, 5, 1, 0, 4, 1, 4, 9, 2, 9])
# prepare the infer date 28px * 28px image
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_test = x_test.reshape(10000, 784) / 255.
image_data = x_test[:10, ...]
# model path
so_lib_path = path.join(path.dirname(__file__), 'out_c', 'qumico.so')
# Load
input_info = np.ctypeslib.ndpointer(dtype=np.float32,
ndim=2,
shape=(1, 784),
flags='CONTIGUOUS')
output_info = np.ctypeslib.ndpointer(dtype=np.float32,
ndim=2,
shape=(1, 10),
flags='CONTIGUOUS')
dll = init(so_lib_path, input_info, output_info)
# infer
res = []
for i in image_data:
output = np.zeros(dtype=np.float32, shape=(1, 10))
infer_c(dll, np.expand_dims(i, 0).astype(np.float32), output)
classification = common.softmax(output)
y = common.onehot_decoding(classification)
res.append(y[0])
print('Predict Index', res) # [7, 2, 1, 0, 4, 1, 4, 9, 5, 9]