From 90bde33e60a3ae0d18a767eba9a2092eab204788 Mon Sep 17 00:00:00 2001 From: Rishabh Jain <56974688+jainris@users.noreply.github.com> Date: Tue, 4 Aug 2020 07:38:13 +0530 Subject: [PATCH] [TFLite] Implemented PADV2 Operator for TFLite and added support for constant values in PAD. (#6167) --- python/tvm/relay/frontend/tflite.py | 38 +++++++++--- tests/python/frontend/tflite/test_forward.py | 91 ++++++++++++++++++++++++++++ 2 files changed, 122 insertions(+), 7 deletions(-) diff --git a/python/tvm/relay/frontend/tflite.py b/python/tvm/relay/frontend/tflite.py index 67a29e7..6e032b1 100644 --- a/python/tvm/relay/frontend/tflite.py +++ b/python/tvm/relay/frontend/tflite.py @@ -116,6 +116,7 @@ class OperatorConverter(object): 'NOT_EQUAL': self.convert_not_equal, 'PACK': self.convert_pack, 'PAD': self.convert_pad, + 'PADV2': self.convert_pad, 'POW': self.convert_pow, 'PRELU': self.convert_prelu, 'RANGE': self.convert_range, @@ -2306,35 +2307,58 @@ class OperatorConverter(object): return out def convert_pad(self, op): - """Convert TFLite PAD""" + """Convert TFLite PAD/PADV2 \ + TFLite treats PAD and PADV2 operators identically""" + input_tensors = self.get_input_tensors(op) - assert len(input_tensors) == 2, "input tensors length should be 2" - # TFLite PAD only support CONSTANT mode and does not support constant_values parameter. - # tensor + # TFLite PAD/PADV2 only supports CONSTANT mode + assert (len(input_tensors) == 2 or len(input_tensors) == 3), \ + "input tensor's length should be 2 for PAD and 3 for PADV2" + + if len(input_tensors) == 3: + assert input_tensors[0].tensor.Type() == input_tensors[2].tensor.Type(), \ + "constant_values tensor must be of same type as input tensor" + input_tensor = input_tensors[0] in_expr = self.get_expr(input_tensor.tensor_idx) # paddings pad_list = self.get_tensor_value(input_tensors[1]) + # convert list of lists to tuple of tuples paddings = tuple(tuple(l) for l in pad_list) - # Set the pad value + # Set the pad value, by default 0, unless constant_values parameter is provided pad_value = 0 + if input_tensor.qnn_params: # Check that input and output tensor have same qnn params. output_tensors = self.get_output_tensors(op) output_tensor = output_tensors[0] assert self.has_same_qnn_params(input_tensor, output_tensor), \ - "TFLite reshape requires input and output scale and zero points to be equal" + "TFLite PADV2 requires input and output scale and zero points to be equal" - # The pad value for quantized pad is the input zero point. + # The pad value for quantized pad is the input zero point by default. pad_value = float(input_tensor.qnn_params['zero_point'].data.asnumpy()) + if len(input_tensors) == 3: + pad_value = self.get_tensor_value(input_tensors[2]) + if isinstance(pad_value, np.ndarray): + pad_value = pad_value.tolist() + if isinstance(pad_value, list): + assert len(pad_value) == 1, "Only one constant value is expected." + pad_value = pad_value[0] + if input_tensor.qnn_params: + # Check that input tensor and constant_values have same qnn params. + assert self.has_same_qnn_params(input_tensor, input_tensors[2]), \ + "TFLite PADV2 requires input and constant_values tensors' \ + scale and zero points to be equal" + out = _op.nn.pad(in_expr, pad_width=paddings, pad_value=pad_value) return out + def convert_floor_div(self, op): """Convert TFLite FLOOR_DIV""" if self.is_quantized(op): diff --git a/tests/python/frontend/tflite/test_forward.py b/tests/python/frontend/tflite/test_forward.py index 69268f3..603eb11 100644 --- a/tests/python/frontend/tflite/test_forward.py +++ b/tests/python/frontend/tflite/test_forward.py @@ -1925,6 +1925,97 @@ def test_forward_pad(): ####################################################################### +# PADV2 +# ----- + +def _test_padv2(data, mode="CONSTANT", quantized=False): + """ One iteration of PADV2 """ + + assert (len(data) == 2 or len(data) == 3) + + with_constant_values = len(data) == 3 + + # Test with tensor and constant + with tf.Graph().as_default(): + in_data = [array_ops.placeholder(shape=data[0].shape, dtype='float32', name='in')] + + if quantized: + # fake_quant will keep the tensors in float32 until the conversion in the session + input_range = {'inq_0': (-100, 100)} + inq_data = [tf.quantization.fake_quant_with_min_max_args(in_data[0], + min=-100, + max=100, + name="inq_0")] + if with_constant_values: + in_constant_values = constant_op.constant(data[2], shape=data[2].shape, dtype='float32', name='in_constant_values') + inq_constant_values = tf.quantization.fake_quant_with_min_max_args(in_constant_values, + min=-100, + max=100, + name='inq_constant_values') + out = array_ops.pad_v2(inq_data[0], + ops.convert_to_tensor(data[1], dtype=data[1].dtype), + constant_values=inq_constant_values, + mode=mode) + out = tf.quantization.fake_quant_with_min_max_args(out, min=-100, max=100, name="out") + else: + out = array_ops.pad_v2(inq_data[0], ops.convert_to_tensor(data[1], dtype=data[1].dtype), mode=mode) + compare_tflite_with_tvm([data[0]], ['inq_0:0'], inq_data, [out], quantized=True, input_range=input_range) + else: + if with_constant_values: + out = array_ops.pad_v2(in_data[0], + ops.convert_to_tensor(data[1], dtype=data[1].dtype), + constant_values= ops.convert_to_tensor(data[2], dtype=data[2].dtype), + mode=mode) + else: + out = array_ops.pad_v2(in_data[0], ops.convert_to_tensor(data[1], dtype=data[1].dtype), mode=mode) + compare_tflite_with_tvm([data[0]], ['in:0'], in_data, [out]) + + +def test_forward_padv2(): + """ PADV2 """ + # Tests without Constant_values + _test_padv2([np.arange(1.0, 7.0, dtype=np.float32).reshape((2, 1, 1, 3)), + np.array([[1, 1], [2, 2], [1, 1], [2, 2]], dtype=np.int32)]) + _test_padv2([np.arange(1.0, 7.0, dtype=np.float32).reshape((2, 1, 3)), + np.array([[2, 2], [1, 1], [1, 1]], dtype=np.int32)]) + _test_padv2([np.arange(1.0, 7.0, dtype=np.float32).reshape((2, 3)), + np.array([[1, 1], [2, 2]], dtype=np.int32)]) + _test_padv2([np.arange(1.0, 4.0, dtype=np.float32).reshape((1, 3)), + np.array([[1, 1], [2, 2]], dtype=np.int32)]) + _test_padv2([np.arange(1.0, 7.0, dtype=np.float32).reshape((2, 3)), + np.array([[1, 1], [2, 2]], dtype=np.int32)], mode="REFLECT") + _test_padv2([np.arange(1.0, 7.0, dtype=np.float32).reshape((2, 3)), + np.array([[1, 1], [2, 2]], dtype=np.int32)], mode="SYMMETRIC") + _test_padv2([np.arange(0, 256, dtype=np.uint8).reshape((1, 256)), + np.array([[1, 1], [2, 2]], dtype=np.int32)], quantized=True) + + # Tests with Constant_values + _test_padv2([np.arange(1.0, 7.0, dtype=np.float32).reshape((2, 1, 1, 3)), + np.array([[1, 1], [2, 2], [1, 1], [2, 2]], dtype=np.int32), + np.array([2], dtype=np.float32)]) + _test_padv2([np.arange(1.0, 7.0, dtype=np.float32).reshape((2, 1, 3)), + np.array([[2, 2], [1, 1], [1, 1]], dtype=np.int32), + np.array([1], dtype=np.float32)]) + _test_padv2([np.arange(1.0, 7.0, dtype=np.float32).reshape((2, 3)), + np.array([[1, 1], [2, 2]], dtype=np.int32), + np.array([-1], dtype=np.float32)]) + _test_padv2([np.arange(1.0, 4.0, dtype=np.float32).reshape((1, 3)), + np.array([[1, 1], [2, 2]], dtype=np.int32), + np.array([2], dtype=np.float32)]) + _test_padv2([np.arange(0, 256, dtype=np.uint8).reshape((1, 256)), + np.array([[1, 1], [2, 2]], dtype=np.int32), + np.array([2], dtype=np.uint8)], quantized=True) + + # Constant Values input can be scalar + _test_padv2([np.arange(1.0, 7.0, dtype=np.float32).reshape((2, 1, 1, 3)), + np.array([[1, 1], [2, 2], [1, 1], [2, 2]], dtype=np.int32), + np.float32(2)]) + _test_padv2([np.arange(0, 256, dtype=np.uint8).reshape((1, 256)), + np.array([[1, 1], [2, 2]], dtype=np.int32), + np.uint8(10)], quantized=True) + + +####################################################################### # Pack # ---- -- 2.7.4