From aa0271ec3a4e78f648aca83fd771a5554b6af355 Mon Sep 17 00:00:00 2001 From: Dmitriy Smirnov Date: Fri, 14 Aug 2020 19:54:29 +0100 Subject: [PATCH] Added support for tflite quantized maximum and minimum (#6018) * Added support for tflite quantized maximum and minimum * Unit test simplified Bugfix in unit test. Unit test slightly simplified * re-trigger CI * renamed use_real_qnn to ignore_qnn_params --- python/tvm/relay/frontend/tflite.py | 42 +++++----- tests/python/frontend/tflite/test_forward.py | 118 ++++++++++++++------------- 2 files changed, 86 insertions(+), 74 deletions(-) diff --git a/python/tvm/relay/frontend/tflite.py b/python/tvm/relay/frontend/tflite.py index 11d6576..48c88d0 100644 --- a/python/tvm/relay/frontend/tflite.py +++ b/python/tvm/relay/frontend/tflite.py @@ -363,12 +363,16 @@ class OperatorConverter(object): rhs_scale = rhs_tensor.qnn_params['scale'] lhs_zero_point = lhs_tensor.qnn_params['zero_point'] rhs_zero_point = rhs_tensor.qnn_params['zero_point'] - lhs_scale_value = get_scalar_from_constant(lhs_scale) - rhs_scale_value = get_scalar_from_constant(rhs_scale) - lhs_zero_point_value = get_scalar_from_constant(lhs_zero_point) - rhs_zero_point_value = get_scalar_from_constant(rhs_zero_point) - return lhs_scale_value == rhs_scale_value and \ - lhs_zero_point_value == rhs_zero_point_value + # 0.1 + 0.2 != 0.3 + return np.allclose(lhs_scale.data.asnumpy(), + rhs_scale.data.asnumpy(), + rtol=1e-5, + atol=1e-5) \ + and \ + np.allclose(lhs_zero_point.data.asnumpy(), + rhs_zero_point.data.asnumpy(), + rtol=1e-5, + atol=1e-5) def is_quantized(self, op): """Check if an input tensor is quantized.""" @@ -1109,7 +1113,7 @@ class OperatorConverter(object): return out - def _convert_elemwise(self, relay_op, op): + def _convert_elemwise(self, relay_op, op, ignore_qnn_params=False): """Generic method to Convert TFLite elemwise""" try: from tflite.AddOptions import AddOptions @@ -1132,8 +1136,16 @@ class OperatorConverter(object): assert len(output_tensors) == 1, "output tensors length should be 1" output_tensor = output_tensors[0] + # TFLite format demands equal scale and zero_point tuple parameters for some operations + # to allow us to use non-quantized operation instead of quantized if ignore_qnn_params=True + if ignore_qnn_params: + assert lhs_tensor.qnn_params \ + and self.has_same_qnn_params(lhs_tensor, output_tensor) \ + and self.has_same_qnn_params(rhs_tensor, output_tensor), \ + "All tensors should be quantized with the same (scale,zero-point) tuple parameters" + # If quantized, extracts qnn params and call QNN add operator. - if lhs_tensor.qnn_params: + if not ignore_qnn_params and lhs_tensor.qnn_params: assert rhs_tensor.qnn_params, "Both tensors should be quantized." assert output_tensor.qnn_params, "Output tensor should be quantized." out = relay_op(lhs=lhs_expr, @@ -1164,7 +1176,7 @@ class OperatorConverter(object): fused_activation_fn = options.FusedActivationFunction() # Handle fused activations - if output_tensor.qnn_params: + if not ignore_qnn_params and output_tensor.qnn_params: scale_val = get_scalar_from_constant(output_tensor.qnn_params['scale']) zero_point_val = get_scalar_from_constant(output_tensor.qnn_params['zero_point']) output_tensor_type_str = self.get_tensor_type_str(output_tensor.tensor.Type()) @@ -1231,19 +1243,11 @@ class OperatorConverter(object): def convert_maximum(self, op): """Convert TFLite MAXIMUM""" - # Check if the input tensor is quantized, call QNN op - if self.is_quantized(op): - raise tvm.error.OpNotImplemented( - 'TFlite quantized MAXIMUM operator is not supported yet.') - return self._convert_elemwise(_op.maximum, op) + return self._convert_elemwise(_op.maximum, op, self.is_quantized(op)) def convert_minimum(self, op): """Convert TFLite MINIMUM""" - # Check if the input tensor is quantized, call QNN op - if self.is_quantized(op): - raise tvm.error.OpNotImplemented( - 'TFlite quantized MINIMUM operator is not supported yet.') - return self._convert_elemwise(_op.minimum, op) + return self._convert_elemwise(_op.minimum, op, self.is_quantized(op)) def convert_greater(self, op): """Convert TFLite GREATER""" diff --git a/tests/python/frontend/tflite/test_forward.py b/tests/python/frontend/tflite/test_forward.py index 33ac6d4..ebb4d77 100644 --- a/tests/python/frontend/tflite/test_forward.py +++ b/tests/python/frontend/tflite/test_forward.py @@ -29,10 +29,15 @@ from tvm import te from tvm import relay try: import tensorflow.compat.v1 as tf + # tensorflow.python.framework.ops module itself is not part of + # TensorFlow's public API: the precise contents of that module + # may vary from one version to the next + import tensorflow.compat.v1 as ops except ImportError: import tensorflow as tf + import tensorflow as ops from tensorflow.python.framework import constant_op -from tensorflow.python.framework import ops + from tensorflow.python.ops import math_ops from tensorflow.python.ops import nn_ops from tensorflow.python.ops import array_ops @@ -235,7 +240,8 @@ def run_tflite_graph(tflite_model_buf, input_data): def compare_tflite_with_tvm(in_data, in_name, input_tensors, output_tensors, init_global_variables=False, - out_names=None, quantized=False, input_range=None, mode='graph_runtime'): + out_names=None, quantized=False, input_range=None, + mode='graph_runtime', experimental_new_converter=False): """Generic function to generate and compare TFLite and TVM output""" in_data = convert_to_list(in_data) in_name = convert_to_list(in_name) @@ -250,7 +256,7 @@ def compare_tflite_with_tvm(in_data, in_name, input_tensors, # convert to tflite model converter = tf.lite.TFLiteConverter.from_session( sess, input_tensors, output_tensors) - + converter.experimental_new_converter=experimental_new_converter if quantized: converter.inference_type = tf.lite.constants.QUANTIZED_UINT8 input_arrays = converter.get_input_arrays() @@ -1300,70 +1306,68 @@ def test_all_unary_elemwise(): # Element-wise # ------------ -def _test_elemwise(math_op, data, fused_activation_function=None, quantized=False, qnn_op=None): +def _test_elemwise(math_op, data, fused_activation_function=None, quantized=False, qnn_op=None, same_qnn_params=False): """ One iteration of elemwise """ assert len(data) == 2 - # Test with two tensors - with tf.Graph().as_default(): - in_data = [array_ops.placeholder(shape=data[0].shape, dtype='float32', name='in_0'), - array_ops.placeholder(shape=data[1].shape, dtype='float32', name='in_1')] - + def __test_elemwise( in_data ): + assert 2 == len( in_data ) if quantized: - # fake_quant will keep the tensors in float32 until the conversion in the session - inq_data = [tf.quantization.fake_quant_with_min_max_args(in_data[0], min=-100, max=100, name="inq_0"), - tf.quantization.fake_quant_with_min_max_args(in_data[1], min=-50, max=50, name="inq_1")] - input_range = {'inq_0': (-100, 100), 'inq_1': (-50, 50)} - out = math_op(inq_data[0], inq_data[1]) - out = with_fused_activation_function(out, fused_activation_function) # set the fp32 output range with respect to the operation out_min, out_max = _test_elemwise_qnn_out_range(qnn_op) - out = tf.quantization.fake_quant_with_min_max_args(out, min=out_min, max=out_max, name="out") - compare_tflite_with_tvm(data, ['inq_0:0', 'inq_1:0'], inq_data, [out], - quantized=True, input_range=input_range) - else: - out = math_op(in_data[0], in_data[1]) - out = with_fused_activation_function(out, fused_activation_function) - compare_tflite_with_tvm(data, ['in_0:0', 'in_1:0'], in_data, [out]) + inq0_min, inq0_max = (-100, 100) + inq1_min, inq1_max = (-50, 50) - # Test with tensor and constant - with tf.Graph().as_default(): - in_data = [array_ops.placeholder(shape=data[0].shape, dtype='float32', name='in_0')] + # if requested use same quantization parameters provided by _test_elemwise_qnn_out_range + if same_qnn_params: + inq0_min, inq0_max = (out_min, out_max) + inq1_min, inq1_max = (out_min, out_max) - if quantized: - inq_data = [tf.quantization.fake_quant_with_min_max_args(in_data[0], min=-100, max=100, name="inq_0")] - inq_const = tf.quantization.fake_quant_with_min_max_args(data[1], min=-50, max=50, name="const_tensor") - input_range = {'inq_0': (-100, 100)} - # the 2nd tensor is treated as constant and directly added as part of the operation - out = math_op(inq_data, ops.convert_to_tensor(inq_const, dtype='float32', name='inq_const')) + # fake_quant will keep the tensors in float32 until the conversion in the session + inq_data = [tf.quantization.fake_quant_with_min_max_args(in_data[0], min=out_min, max=out_max, name="inq_0")\ + if None != in_data[0]\ + else tf.quantization.fake_quant_with_min_max_args(data[0], min=out_min, max=out_max, name="const_tensor0"), + tf.quantization.fake_quant_with_min_max_args(in_data[1], min=out_min, max=out_max, name="inq_1")\ + if None != in_data[1]\ + else tf.quantization.fake_quant_with_min_max_args(data[1], min=out_min, max=out_max, name="const_tensor1")] + + input_range = {x[1][0]:x[1][1] for x in zip(in_data, (('inq_0', (inq0_min, inq0_max)),\ + ('inq_1', (inq1_min, inq1_max)))) if None != x[0]} + + out = math_op(inq_data[0], inq_data[1]) out = with_fused_activation_function(out, fused_activation_function) - out_min, out_max = _test_elemwise_qnn_out_range(qnn_op) out = tf.quantization.fake_quant_with_min_max_args(out, min=out_min, max=out_max, name="out") - compare_tflite_with_tvm(data[0], ['inq_0:0'], inq_data, [out], quantized=True, input_range=input_range) + + # Note same_qnn_params uses experimental_new_converter as toco failed + compare_tflite_with_tvm([x[1] for x in zip(in_data, data) if None != x[0]], + [x + ":0" for x in input_range.keys()], + [x[1] for x in zip(in_data, inq_data) if None != x[0]], + [out], + quantized=True, + input_range=input_range, + experimental_new_converter=same_qnn_params) else: - out = math_op(in_data[0], ops.convert_to_tensor(data[1], dtype=data[1].dtype)) + out = math_op(in_data[0] if None != in_data[0] else ops.convert_to_tensor(data[0], dtype=data[0].dtype), + in_data[1] if None != in_data[1] else ops.convert_to_tensor(data[1], dtype=data[1].dtype)) out = with_fused_activation_function(out, fused_activation_function) - compare_tflite_with_tvm(data[0], ['in_0:0'], in_data, [out]) + compare_tflite_with_tvm([x[1] for x in zip( in_data, data ) if None != x[0]], + [x[1] for x in zip( in_data, ('in_0:0', 'in_1:0') ) if None != x[0]], + [x for x in in_data if None != x], + [out]) + # Test with two tensors + with tf.Graph().as_default(): + __test_elemwise( in_data = [array_ops.placeholder(shape=data[0].shape, dtype='float32', name='in_0'), + array_ops.placeholder(shape=data[1].shape, dtype='float32', name='in_1')]) + # Test with tensor and constant + with tf.Graph().as_default(): + __test_elemwise( in_data = [array_ops.placeholder(shape=data[0].shape, dtype='float32', name='in_0'), + None]) # Test with constant and tensor with tf.Graph().as_default(): - in_data = [array_ops.placeholder(shape=data[1].shape, dtype='float32', name='in_1')] - - if quantized: - inq_const = tf.quantization.fake_quant_with_min_max_args(data[0], min=-100, max=100, name="const_tensor") - inq_data = [tf.quantization.fake_quant_with_min_max_args(in_data[0], min=-50, max=50, name="inq_1")] - input_range = {'inq_1': (-50, 50)} - # the 1st tensor is treated as constant and directly added as part of the operation - out = math_op(ops.convert_to_tensor(inq_const, dtype='float32', name='inq_const'), inq_data) - out = with_fused_activation_function(out, fused_activation_function) - out_min, out_max = _test_elemwise_qnn_out_range(qnn_op) - out = tf.quantization.fake_quant_with_min_max_args(out, min=out_min, max=out_max, name="out") - compare_tflite_with_tvm(data[1], ['inq_1:0'], inq_data, [out], quantized=True, input_range=input_range) - else: - out = math_op(ops.convert_to_tensor(data[0], dtype=data[0].dtype), in_data[0]) - out = with_fused_activation_function(out, fused_activation_function) - compare_tflite_with_tvm(data[1], ['in_1:0'], in_data, [out]) + __test_elemwise( in_data = [None, + array_ops.placeholder(shape=data[1].shape, dtype='float32', name='in_1')]) ####################################################################### # Add @@ -1406,16 +1410,16 @@ def _test_pow(data): # Maximum # ------- -def _test_maximum(data): +def _test_maximum(data, fused_activation_function=None, quantized=False, qnn_op=None): """ One iteration of maximum """ - return _test_elemwise(math_ops.maximum, data) + return _test_elemwise(math_ops.maximum, data, fused_activation_function, quantized, qnn_op, same_qnn_params=True) ####################################################################### # Minimum # ------- -def _test_minimum(data): +def _test_minimum(data, fused_activation_function=None, quantized=False, qnn_op=None): """ One iteration of minimum """ - return _test_elemwise(math_ops.minimum, data) + return _test_elemwise(math_ops.minimum, data, fused_activation_function, quantized, qnn_op, same_qnn_params=True) ####################################################################### # Greater # ------- @@ -1501,6 +1505,8 @@ def _test_elemwise_qnn_out_range(qnn_op): _test_add: (-150, 150), _test_sub: (-150, 150), _test_mul: (-5e+3, 5e+3), + _test_maximum: (-112, 111), + _test_minimum: (-128, 127) } return qnn_out_range[qnn_op] @@ -1525,7 +1531,9 @@ def test_all_elemwise(): _test_forward_elemwise(partial(_test_div, fused_activation_function="RELU6")) _test_forward_elemwise(_test_pow) _test_forward_elemwise(_test_maximum) + _test_forward_elemwise_quantized(_test_maximum) _test_forward_elemwise(_test_minimum) + _test_forward_elemwise_quantized(_test_minimum) _test_forward_elemwise(_test_greater) _test_forward_elemwise(_test_squared_difference) _test_forward_elemwise(_test_greater_equal) -- 2.7.4