def test_encoding_analyzer_with_numpy_interface(self):
"""
compare qat asymmetric quantization with qc quantize implementation
:return:
"""
np.random.seed(10)
random_input = 5 * (np.random.normal(size=[1, 3, 224, 224])) + 2
# Full range min, max (no scaling input)
x_min = np.min([0., random_input.min()])
x_max = np.max([0., random_input.max()])
delta = (x_max - x_min) / 255
offset = np.round(x_min / delta)
x_min = offset * delta
x_max = x_min + 255 * delta
enc_analyzer = libpymo.EncodingAnalyzerForPython(
libpymo.QuantizationMode.QUANTIZATION_TF)
enc_analyzer.updateStats(random_input, False)
encoding, is_valid = enc_analyzer.computeEncoding(
8, False, False, False)
print("Encoding.min=", encoding.min)
print("Encoding.max=", encoding.max)
self.assertTrue(is_valid)
self.assertAlmostEqual(x_min, encoding.min, places=5)
self.assertAlmostEqual(x_max, encoding.max, places=5)
def compute_encoding_for_given_bitwidth(data: np.ndarray, bitwidth: int, quant_scheme: QuantScheme,
is_symmetric: bool) -> Dict:
"""
Return encoding dictionary for given bitwidth
:param data: Numpy data
:param bitwidth: bitwidth (4-31) to use for quantizing data
:param quant_scheme: Quantization scheme
:param is_symmetric: True if symmetric encodings is used, False otherwise
:return: Encoding Dictionary
"""
# Create Encodings Analyzer and collect statistical data to compute encodings
# Since the data is numpy array and on CPU memory, useCuda is False
encoding_analyzer = libpymo.EncodingAnalyzerForPython(quant_scheme)
encoding_analyzer.updateStats(data, False)
encoding, is_encoding_valid = encoding_analyzer.computeEncoding(bitwidth, is_symmetric, False, False)
if is_encoding_valid:
return {'min': encoding.min,
'max': encoding.max,
'scale': encoding.delta,
'offset': encoding.offset,
'bitwidth': encoding.bw,
'is_symmetric': str(is_symmetric)}
return {}
def _get_quantized_weights(weight_tensor, quant_params):
"""
helper function to get quantized dequantized weights
:param weight_tensor: weight tensor
:param quant_params: quantization params such as mode, rounding etc
:return: quantized de-quantized weight tensor
"""
q_wt_tensor = weight_tensor
quant_mode = libpymo.QuantizationMode.QUANTIZATION_TF_ENHANCED
if quant_params.quant_mode == QuantScheme.post_training_tf or quant_params.quant_mode == 'tf':
quant_mode = libpymo.QuantizationMode.QUANTIZATION_TF
round_mode = libpymo.RoundingMode.ROUND_NEAREST
if quant_params.round_mode == 'stochastic':
round_mode = libpymo.RoundingMode.ROUND_STOCHASTIC
bitwidth = 8
# use tensorQuantizerForPython to get quantizeDequantize weights
encoding_analyzer = libpymo.EncodingAnalyzerForPython(quant_mode)
encoding_analyzer.updateStats(weight_tensor, quant_params.use_cuda)
encoding, is_encoding_valid = encoding_analyzer.computeEncoding(
bitwidth, False, False, False)
if is_encoding_valid:
tensor_quantizer = libpymo.TensorQuantizationSimForPython()
q_wt_tensor = tensor_quantizer.quantizeDequantize(
weight_tensor, encoding, round_mode, quant_params.use_cuda)
return q_wt_tensor