def test_ssd_mobilenet_v2_model():
model = DLRModel(MODEL_PATH.as_posix())
data = np.load(DATA_PATH)
assert model.get_input_names() == ['image_tensor']
assert model.get_output_names() == [
'detection_scores:0', 'detection_classes:0', 'num_detections:0'
]
assert model.get_input_dtypes() == ['uint8']
assert model.get_output_dtypes() == ['float32', 'float32', 'float32']
outputs = model.run({"image_tensor": data})
assert outputs[0].shape == (1, 100, 4)
assert outputs[1].shape == (1, 100)
assert outputs[2].shape == (1, 100)
detections = np.multiply(np.ceil(outputs[1]), outputs[2])
expected = np.zeros(detections.shape)
expected[:, :6] = np.array([[1., 1., 1., 2., 3., 1]])
comparison = detections == expected
assert comparison.all()
def test_resnet():
# Load the model
model_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
'resnet18_v1')
device = 'cpu'
model = DLRModel(model_path, device)
# Run the model
image = np.load(os.path.join(os.path.dirname(os.path.abspath(__file__)), 'dog.npy')).astype(np.float32)
#flatten within a input array
input_data = {'data': image}
print('Testing inference on resnet18...')
probabilities = model.run(input_data) #need to be a list of input arrays matching input names
assert probabilities[0].argmax() == 151
assert model.get_input_names() == ["data"]
assert model.get_input_dtypes() == ["float32"]
assert model.get_output_dtypes() == ["float32"]
assert model.get_input_dtype(0) == "float32"
assert model.get_output_dtype(0) == "float32"
def test_ssd_mobilenet_v2_model():
model = DLRModel(MODEL_PATH.as_posix())
data = np.load(DATA_PATH)
assert model.get_input_names() == ['image_tensor']
assert model.get_output_names() == [
'detection_classes:0', 'num_detections:0', 'detection_boxes:0',
'detection_scores:0'
]
assert model.get_input_dtypes() == ['uint8']
assert model.get_output_dtypes() == [
'float32', 'float32', 'float32', 'float32'
]
outputs = model.run({"image_tensor": data})
assert outputs[0].shape == (1, 100)
assert outputs[1].shape == (1, )
assert outputs[2].shape == (1, 100, 4)
assert outputs[3].shape == (1, 100)
detections = np.multiply(np.ceil(outputs[3]), outputs[0])
expected = np.zeros(detections.shape)
assert np.count_nonzero(detections) == outputs[1][0]
def test_mobilenet_v1_0_75_224_quant():
# Load the model
model_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
'mobilenet_v1_0.75_224_quant')
device = 'cpu'
model = DLRModel(model_path, device)
# load image (dtype: uint8)
image = np.load(
os.path.join(os.path.dirname(os.path.abspath(__file__)),
'cat_224_uint8.npy'))
print('Testing inference on mobilenet_v1_0.75_224_quant...')
probabilities = model.run({'input': image})
assert probabilities[0].argmax() == 282
assert model.get_input_names() == ["input"]
assert model.get_input_dtypes() == ["uint8"]
assert model.get_output_dtypes() == ["uint8"]
assert model.get_input_dtype(0) == "uint8"
assert model.get_output_dtype(0) == "uint8"
input2 = model.get_input("input")
assert input2.dtype == 'uint8'
assert input2.shape == (1, 224, 224, 3)
assert (input2 == image).all()
def test_tf_model(dev_type=None, dev_id=None):
_generate_frozen_graph()
model = DLRModel(FROZEN_GRAPH_PATH, dev_type, dev_id)
inp_names = model.get_input_names()
assert inp_names == ['import/input1:0', 'import/input2:0']
out_names = model.get_output_names()
assert out_names == [
'import/preproc/output1:0', 'import/preproc/output2:0'
]
inp1 = [[4., 1.], [3., 2.]]
inp2 = [[0., 1.], [1., 0.]]
res = model.run({'import/input1:0': inp1, 'import/input2:0': inp2})
assert res is not None
assert len(res) == 2
assert np.alltrue(res[0] == [[36., 361.], [49., 324.]])
assert res[1] == 1
m_inp1 = model.get_input('import/input1:0')
m_inp2 = model.get_input('import/input2:0')
assert np.alltrue(m_inp1 == inp1)
assert np.alltrue(m_inp2 == inp2)
def test_tflite_model():
_generate_tflite_file()
m = DLRModel(TFLITE_FILE_PATH)
inp_names = m.get_input_names()
assert sorted(inp_names) == ['input1', 'input2']
out_names = m.get_output_names()
assert sorted(out_names) == ['preproc/output1', 'preproc/output2']
inp1 = np.array([[4., 1.], [3., 2.]]).astype("float32")
inp2 = np.array([[0., 1.], [1., 0.]]).astype("float32")
res = m.run({'input1': inp1, 'input2': inp2})
assert res is not None
assert len(res) == 2
exp_out0 = np.array([[36., 361.], [49., 324.]]).astype("float32")
assert np.alltrue(res[0] == exp_out0)
assert res[1] == 1
m_inp1 = m.get_input('input1')
m_inp2 = m.get_input('input2')
assert np.alltrue(m_inp1 == inp1)
assert np.alltrue(m_inp2 == inp2)