/
githubmirror
/
incubator-mxnet
Обзор
Документация
Войти
/
githubmirror
/
incubator-mxnet
Код
Запросы
0
Пакеты
0
Релизы
0
Аналитика
Безопасность
master
tests/python/quantization/test_quantization.py
1 596 строк
76 KB
Joe Evans
Port top-level-project updates from v1.x branch (#21162)
04 янв 2023, 15:09
Не верифицирован
04 янв 2023, 15:09
48d7f4a
Код
Авторство
О чём код?
# Licensed to the Apache Software Foundation (ASF) under one # or more contributor license agreements. See the NOTICE file # distributed with this work for additional information # regarding copyright ownership. The ASF licenses this file # to you under the Apache License, Version 2.0 (the # "License"); you may not use this file except in compliance # with the License. You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, # software distributed under the License is distributed on an # "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY # KIND, either express or implied. See the License for the # specific language governing permissions and limitations # under the License. """Some of the tests using CUDNN require a special GPU instruction called dp4a. Ref: http://images.nvidia.com/content/pdf/tesla/184457-Tesla-P4-Datasheet-NV-Final-Letter-Web.pdf """ import os import mxnet as mx import numpy as onp from mxnet import npx from mxnet.util import use_np from mxnet.gluon.model_zoo import vision from mxnet.test_utils import assert_almost_equal, assert_exception, rand_ndarray, rand_shape_nd, same, DummyIter from common import xfail_when_nonstandard_decimal_separator from mxnet.io import NDArrayIter import unittest import operator npx.reset_np() def collect_block_args_aux(block, sym): arg_params, aux_params = dict(), dict() for k, v in block.collect_params().items(): if k in sym.list_arguments(): arg_params[k]= v._reduce() elif k in sym.list_auxiliary_states(): aux_params[k]= v._reduce() return arg_params, aux_params def is_test_for_gpu(): return mx.current_device().device_type == 'gpu' def is_test_for_dnnl(): return (mx.current_device().device_type == 'cpu' and os.environ.get('ENABLE_ONEDNN_QUANTIZATION_TEST') == '1') def is_test_for_native_cpu(): return (mx.current_device().device_type == 'cpu' and os.environ.get('ENABLE_ONEDNN_QUANTIZATION_TEST') == None) def get_low_high(qtype): """ Return low and high value for given integer type as float number""" if qtype == 'uint8': return 0.0, 255.0 else: return -128.0, 127.0 def test_quantize_float32_to_int8(): shape = rand_shape_nd(4) data = rand_ndarray(shape, 'default', dtype='float32') min_range = mx.nd.min(data) max_range = mx.nd.max(data) qdata, min_val, max_val = mx.nd.contrib.quantize(data, min_range, max_range, out_type='int8') data_np = data.asnumpy() min_range = min_range.asscalar() max_range = max_range.asscalar() real_range = onp.maximum(onp.abs(min_range), onp.abs(max_range)) quantized_range = 127.0 scale = quantized_range / real_range assert qdata.dtype == onp.int8 assert min_val.dtype == onp.float32 assert max_val.dtype == onp.float32 assert same(min_val.asscalar(), -real_range) assert same(max_val.asscalar(), real_range) qdata_np = (onp.sign(data_np) * onp.minimum(onp.abs(data_np) * scale + 0.5, quantized_range)).astype(onp.int8) assert_almost_equal(qdata.asnumpy(), qdata_np, atol = 1) def test_calibrated_quantize_v2_bfloat16_to_int8(): shape = rand_shape_nd(4) data = mx.nd.random.normal(0, 1, shape).astype('bfloat16') min_range = mx.nd.min(data).asscalar() max_range = mx.nd.max(data).asscalar() qdata, min_val, max_val = mx.nd.contrib.quantize_v2(data, 'int8', min_range, max_range) data_np = data.asnumpy() real_range = onp.maximum(onp.abs(min_range), onp.abs(max_range)) quantized_range = 127.0 scale = quantized_range / real_range assert qdata.dtype == onp.int8 assert min_val.dtype == onp.float32 assert max_val.dtype == onp.float32 assert same(min_val.asscalar(), -real_range) assert same(max_val.asscalar(), real_range) qdata_np = (onp.sign(data_np) * onp.minimum(onp.abs(data_np) * scale + 0.5, quantized_range)).astype(onp.int8) assert_almost_equal(qdata.asnumpy(), qdata_np, atol=1) def test_dequantize_int8_to_float32(): def get_test_data(real_range, qdata_np): qdata = mx.nd.array(qdata_np, dtype=onp.int8) min_range = mx.nd.array([-real_range], dtype=onp.float32) max_range = mx.nd.array([real_range], dtype=onp.float32) return qdata, min_range, max_range def baseline_dequantization(qdata, real_range, qdata_np): quantized_range = 127.0 scale = real_range / quantized_range data_np = qdata_np * scale return data_np def test_nd_array_dequantization(qdata, min_range, max_range, expected_result): data = mx.nd.contrib.dequantize(qdata, min_range, max_range, out_type='float32') assert data.dtype == onp.float32 assert_almost_equal(data.asnumpy(), expected_result, atol = 1) def test_symbolic_api_dequantization(qdata, min_range, max_range, expected_result): sym_data = mx.sym.Variable('data') sym_min_range = mx.sym.Variable('min_range') sym_max_range = mx.sym.Variable('max_range') dequant = mx.sym.contrib.dequantize(sym_data, sym_min_range, sym_max_range, out_type='float32') out = dequant._bind(ctx=mx.current_device(), args={'data':qdata, 'min_range':min_range, 'max_range':max_range}) data = out.forward()[0] assert data.dtype == onp.float32 assert_almost_equal(data.asnumpy(), expected_result, atol = 1) real_range = 128 shape = rand_shape_nd(4) qdata_np = onp.random.uniform(low=-127, high=127, size=shape).astype(dtype=onp.int8) qdata, min_range, max_range = get_test_data(real_range, qdata_np) expected_result = baseline_dequantization(qdata, real_range, qdata_np) # test nd array implementation. test_nd_array_dequantization(qdata, min_range, max_range, expected_result) # test symbolic api implementaion. test_symbolic_api_dequantization(qdata, min_range, max_range, expected_result) def test_requantize_int32_to_int8(): def quantized_int32_to_float(qdata, min_range, max_range): assert qdata.dtype == 'int32' quantized_range = onp.iinfo('int32').max real_range = onp.maximum(onp.abs(min_range), onp.abs(max_range)) scale = float(real_range) / float(quantized_range) return qdata.astype('float32') * scale def float_to_quantized_int8(data, min_range, max_range): assert data.dtype == 'float32' real_range = onp.maximum(onp.abs(min_range), onp.abs(max_range)) quantized_range = onp.iinfo('int8').max scale = float(quantized_range) / float(real_range) return (onp.sign(data) * onp.minimum(onp.abs(data) * scale + 0.5, quantized_range)).astype('int8') def requantize(qdata, min_data, max_data, real_range): data = quantized_int32_to_float(qdata, min_data, max_data) output = float_to_quantized_int8(data, -real_range, real_range) return output, -real_range, real_range def requantize_baseline(qdata, min_data, max_data, min_calib_range=None, max_calib_range=None): if min_calib_range is not None and max_calib_range is not None: real_range = onp.maximum(onp.abs(min_calib_range), onp.abs(max_calib_range)) return requantize(qdata, min_data, max_data, real_range) else: min_range = quantized_int32_to_float(onp.min(qdata), min_data, max_data) max_range = quantized_int32_to_float(onp.max(qdata), min_data, max_data) return requantize(qdata, min_data, max_data, onp.maximum(onp.abs(min_range), onp.abs(max_range))) def check_requantize(shape, min_calib_range=None, max_calib_range=None): qdata = mx.nd.random.uniform(low=-1000.0, high=1000.0, shape=shape).astype('int32') min_range = mx.nd.array([-1010.0]) max_range = mx.nd.array([1020.0]) if min_calib_range is None or max_calib_range is None: qdata_int8, min_output, max_output = mx.nd.contrib.requantize(qdata, min_range, max_range) else: qdata_int8, min_output, max_output = mx.nd.contrib.requantize(qdata, min_range, max_range, min_calib_range=min_calib_range, max_calib_range=max_calib_range) qdata_int8_np, min_output_np, max_output_np = requantize_baseline(qdata.asnumpy(), min_range.asscalar(), max_range.asscalar(), min_calib_range=min_calib_range, max_calib_range=max_calib_range) assert_almost_equal(qdata_int8.asnumpy(), qdata_int8_np, atol = 1) assert_almost_equal(min_output.asnumpy(), onp.array([min_output_np])) assert_almost_equal(max_output.asnumpy(), onp.array([max_output_np])) @use_np def check_requantize_with_gluon(shape, min_calib_range=None, max_calib_range=None): qdata = mx.np.random.uniform(low=-1000.0, high=1000.0, size=shape).astype('int32') min_range = mx.np.array([-1010.0]) max_range = mx.np.array([1020.0]) class RequantizeBlock(mx.gluon.nn.HybridBlock): def __init__(self, min_calib_range=None, max_calib_range=None, **kwargs): super(RequantizeBlock, self).__init__(**kwargs) self.min_calib_range = min_calib_range self.max_calib_range = max_calib_range def forward(self, x, min_range, max_range): if self.min_calib_range is not None and self.max_calib_range is not None: out = npx.requantize(x, min_range, max_range, min_calib_range=self.min_calib_range, max_calib_range=self.max_calib_range) else: out = npx.requantize(x, min_range, max_range) return out requant = RequantizeBlock(min_calib_range, max_calib_range) # m*_calib_ranges can be None qdata_int8, min_output, max_output = requant(qdata, min_range, max_range) qdata_int8_np, min_output_np, max_output_np = requantize_baseline(qdata.asnumpy(), min_range.item(), max_range.item(), min_calib_range=min_calib_range, max_calib_range=max_calib_range) assert_almost_equal(qdata_int8.asnumpy(), qdata_int8_np, atol = 1) assert_almost_equal(min_output.asnumpy(), onp.array([min_output_np])) assert_almost_equal(max_output.asnumpy(), onp.array([max_output_np])) # test with gluon API. check_requantize_with_gluon((3, 4, 10, 10)) check_requantize_with_gluon((32, 3, 23, 23)) check_requantize_with_gluon((3, 4, 10, 10), min_calib_range=-1050.0, max_calib_range=1040.0) check_requantize_with_gluon((32, 3, 23, 23), min_calib_range=-134.349, max_calib_range=523.43) # Test with nd array API check_requantize((3, 4, 10, 10)) check_requantize((32, 3, 23, 23)) check_requantize((3, 4, 10, 10), min_calib_range=-1050.0, max_calib_range=1040.0) check_requantize((32, 3, 23, 23), min_calib_range=-134.349, max_calib_range=523.43) @use_np def test_quantized_conv(): def check_quantized_conv(data_shape, kernel, num_filter, pad, stride, dilate, use_bias, qdtype): if is_test_for_native_cpu(): print('skipped testing quantized_conv for native cpu since it is not supported yet') return elif is_test_for_dnnl(): # (TODO)Xinyu: https://github.com/apache/mxnet/issues/16830 print('skipped testing quantized_conv for oneDNN cpu since it is a flaky case') return elif qdtype == 'uint8' and is_test_for_gpu(): print('skipped testing quantized_conv for gpu uint8 since it is not supported yet') return elif is_test_for_gpu() and len(data_shape) != 4: print('skipped testing quantized_conv for gpu 5d layout since it is not supported yet') return # run fp32 conv if len(data_shape) == 4: convfp32 = mx.gluon.nn.Conv2D(channels=num_filter, kernel_size=kernel, strides=stride, padding=pad, dilation=dilate, use_bias=use_bias) elif len(data_shape) == 5: convfp32 = mx.gluon.nn.Conv3D(channels=num_filter, kernel_size=kernel, strides=stride, padding=pad, dilation=dilate, use_bias=use_bias) else: print('unsupported shape') assert False if qdtype == 'uint8': data_low = 0.0 data_high = 127.0 else: data_low = -127.0 data_high = 127.0 convfp32.initialize() input_data = mx.np.random.uniform(low=data_low, high=data_high, size=data_shape ).astype('int32').astype('float32') convfp32(input_data) # initialize params npx.waitall() fp32_params = convfp32.collect_params() weight_shape = fp32_params['weight'].shape new_args = dict() new_args['weight'] = mx.np.random.uniform(low=-127.0, high=127.0, size=weight_shape ).astype('int32').astype('float32') if use_bias: new_args['bias'] = mx.np.random.uniform(low=-127.0, high=127.0, size=fp32_params['bias'].shape ).astype('int32').astype('float32') convfp32.load_dict(new_args, cast_dtype=True, dtype_source='saved') output = convfp32(input_data) # run quantized conv class QuantConv(mx.gluon.nn.HybridBlock): def __init__(self, channels, kernel_size, strides=(1, 1), padding=(0, 0), dilation=(1, 1), use_bias=True, **kwargs): super(QuantConv, self).__init__(**kwargs) self.use_bias = use_bias self._kwargs = {'kernel': kernel_size, 'stride': strides, 'dilate': dilation, 'pad': padding, 'num_filter': channels, 'no_bias': not use_bias, 'num_group': 1, 'layout': 'NCHW'} self.min_data = mx.gluon.Parameter('min_data', dtype='float32', shape=(1), allow_deferred_init=True) self.max_data = mx.gluon.Parameter('max_data', dtype='float32', shape=(1), allow_deferred_init=True) self.weight = mx.gluon.Parameter('weight', dtype='int8', shape=weight_shape, allow_deferred_init=True) self.min_weight = mx.gluon.Parameter('min_weight', dtype='float32', shape=(1), allow_deferred_init=True) self.max_weight = mx.gluon.Parameter('max_weight', dtype='float32', shape=(1), allow_deferred_init=True) if use_bias: self.bias = mx.gluon.Parameter('bias', dtype='int8', shape=(num_filter,), allow_deferred_init=True) self.min_bias = mx.gluon.Parameter('min_bias', dtype='float32', shape=(1), allow_deferred_init=True) self.max_bias = mx.gluon.Parameter('max_bias', dtype='float32', shape=(1), allow_deferred_init=True) def forward(self, x): device = x.device weight = self.weight.data().to_device(device) bias = self.bias.data().to_device(device) if self.use_bias else None min_data = self.min_data.data().to_device(device) max_data = self.max_data.data().to_device(device) min_weight = self.min_weight.data().to_device(device) max_weight = self.max_weight.data().to_device(device) min_bias = self.min_bias.data().to_device(device) if self.use_bias else None max_bias = self.max_bias.data().to_device(device) if self.use_bias else None out = npx.quantized_conv(data=x, weight=weight, bias=bias, min_data=min_data, max_data=max_data, min_weight=min_weight, max_weight=max_weight, min_bias=min_bias, max_bias=max_bias, **self._kwargs) return out convint8 = QuantConv(channels=num_filter, kernel_size=kernel, strides=stride, padding=pad, dilation=dilate, use_bias=use_bias) quantized_range = 127.0 qargs = { 'weight': new_args['weight'].astype('int8'), 'min_data': mx.np.array([-quantized_range]), 'max_data': mx.np.array([quantized_range]), 'min_weight': mx.np.array([-quantized_range]), 'max_weight': mx.np.array([quantized_range]) } if use_bias: qargs.update({ 'bias': new_args['bias'].astype('int8'), 'min_bias': mx.np.array([-quantized_range]), 'max_bias': mx.np.array([quantized_range]), }) convint8.load_dict(qargs, cast_dtype=True, dtype_source='saved') qoutput, min_range, max_range = convint8(input_data.astype(qdtype)) if use_bias: # with adding bias, accuracy loss should not be greater than one diff = mx.np.abs(output - qoutput.astype(output.dtype)) cond = mx.np.less(2, diff).sum().item() assert cond == 0 else: assert_almost_equal(output.asnumpy(), qoutput.asnumpy(), atol = 1) for qdtype in ['int8', 'uint8']: check_quantized_conv((3, 4, 28, 28), (3, 3), 128, (1, 1), (1, 1), (1, 1), True, qdtype) check_quantized_conv((3, 4, 28, 28), (3, 3), 128, (1, 1), (1, 1), (1, 1), False, qdtype) check_quantized_conv((1, 3, 4, 28, 28), (1, 3, 3), 128, (1, 1, 1), (1, 1, 1), (1, 1, 1), False, qdtype) check_quantized_conv((1, 3, 4, 28, 28), (1, 3, 3), 128, (1, 1, 1), (1, 1, 1), (1, 1, 1), True, qdtype) check_quantized_conv((1, 3, 4, 28, 28), (1, 3, 3), 128, (1, 1, 1), (1, 1, 1), (2, 2, 2), False, qdtype) check_quantized_conv((1, 3, 4, 28, 28), (1, 3, 3), 128, (1, 1, 1), (1, 1, 1), (2, 2, 2), True, qdtype) @use_np def test_quantized_elemwise_add(): def check_quantized_elemwise_add(data_shape, qdtypeA, qdtypeB): if is_test_for_native_cpu(): print('skipped testing quantized_elemwise_add for native cpu since it is not supported yet') return elif (qdtypeA != 'uint8' and qdtypeA != 'int8') or (qdtypeB != 'uint8' and qdtypeB != 'int8'): print('skipped testing quantized_elemwise_add for not supported data type') return elif is_test_for_gpu(): print('skipped testing quantized_elemwise_add for gpu since it is not supported yet') return class ElemwiseSumBlock(mx.gluon.nn.HybridBlock): def __init__(self, **kwargs): super(ElemwiseSumBlock, self).__init__(**kwargs) def forward(self, dataA, dataB): return dataA + dataB class QuantElemwiseSumBlock(mx.gluon.nn.HybridBlock): def __init__(self, **kwargs): super(QuantElemwiseSumBlock, self).__init__(**kwargs) def forward(self, dataA, dataB, dataA_min, dataA_max, dataB_min, dataB_max): return npx.quantized_elemwise_add(dataA, dataB, dataA_min, dataA_max, dataB_min, dataB_max) elemwise_add_fp32 = ElemwiseSumBlock() dataA_low, dataA_high = get_low_high(qdtypeA) dataB_low, dataB_high = get_low_high(qdtypeB) dataA_val = mx.np.random.uniform(low=dataA_low, high=dataA_high, size=data_shape).astype('int32').astype('float32') dataB_val = mx.np.random.uniform(low=dataB_low, high=dataB_high, size=data_shape).astype('int32').astype('float32') output = elemwise_add_fp32(dataA_val, dataB_val) mx.nd.waitall() #run quantized quantized_elemwise_add = QuantElemwiseSumBlock() dataA_val_int8 = dataA_val.astype(qdtypeA) dataB_val_int8 = dataB_val.astype(qdtypeB) quantized_range = 127.0 min_dataA = mx.np.array([dataA_low]) max_dataA = mx.np.array([dataA_high]) min_dataB = mx.np.array([dataB_low]) max_dataB = mx.np.array([dataB_high]) qoutput, min_range, max_range = quantized_elemwise_add(dataA_val_int8, dataB_val_int8, min_dataA, max_dataA, min_dataB, max_dataB) int8_rslt = qoutput.astype(output.dtype) * max_range / 0x7fffffff diff = mx.np.abs(output - int8_rslt) cond = mx.np.less(2, diff).sum().item() assert cond == 0 check_quantized_elemwise_add((4, 6), 'uint8', 'int8') check_quantized_elemwise_add((13, 74, 52), 'uint8', 'uint8') check_quantized_elemwise_add((3, 4, 56, 56), 'int8', 'uint8') check_quantized_elemwise_add((32, 56, 64, 11), 'int8', 'int8') @use_np def test_quantized_npi_add(): def check_quantized_npi_add(data_shape, qdtypeA, qdtypeB, broadcast=None): if is_test_for_native_cpu(): print('skipped testing quantized_npi_add for native cpu since it is not supported yet') return elif (qdtypeA != 'uint8' and qdtypeA != 'int8') or (qdtypeB != 'uint8' and qdtypeB != 'int8'): print('skipped testing quantized_npi_add for not supported data type') return elif is_test_for_gpu(): print('skipped testing quantized_npi_add for gpu since it is not supported yet') return class ElemwiseSumBlock(mx.gluon.nn.HybridBlock): def __init__(self, **kwargs): super(ElemwiseSumBlock, self).__init__(**kwargs) def forward(self, dataA, dataB): return dataA + dataB class QuantElemwiseSumBlock(mx.gluon.nn.HybridBlock): def __init__(self, **kwargs): super(QuantElemwiseSumBlock, self).__init__(**kwargs) def forward(self, dataA, dataB, dataA_min, dataA_max, dataB_min, dataB_max): return npx.quantized_npi_add(dataA, dataB, dataA_min, dataA_max, dataB_min, dataB_max) elemwise_add_fp32 = ElemwiseSumBlock() dataA_low, dataA_high = get_low_high(qdtypeA) dataB_low, dataB_high = get_low_high(qdtypeB) data_shapeA = data_shape data_shapeB = data_shape if broadcast : if broadcast == 'A': data_shapeA = () for index in range(len(data_shape)): data_shapeA += (1,) else: data_shapeB = () for index in range(len(data_shape)): data_shapeB += (1,) dataA_val = mx.np.random.uniform(low=dataA_low, high=dataA_high, size=data_shapeA).astype('int32').astype('float32') dataB_val = mx.np.random.uniform(low=dataB_low, high=dataB_high, size=data_shapeB).astype('int32').astype('float32') output = elemwise_add_fp32(dataA_val, dataB_val) #run quantized quantized_elemwise_add = QuantElemwiseSumBlock() dataA_val_int8 = dataA_val.astype(qdtypeA) dataB_val_int8 = dataB_val.astype(qdtypeB) quantized_range = 127.0 min_dataA = mx.np.array([dataA_low]) max_dataA = mx.np.array([dataA_high]) min_dataB = mx.np.array([dataB_low]) max_dataB = mx.np.array([dataB_high]) qoutput, min_range, max_range = quantized_elemwise_add(dataA_val_int8, dataB_val_int8, min_dataA, max_dataA, min_dataB, max_dataB) int8_rslt = qoutput.astype(output.dtype) * max_range / 0x7fffffff diff = mx.np.abs(output - int8_rslt) cond = mx.np.less(2, diff).sum().item() assert cond == 0 check_quantized_npi_add((4, 6), 'uint8', 'int8') check_quantized_npi_add((13, 74, 52), 'uint8', 'uint8') check_quantized_npi_add((3, 4, 56, 56), 'int8', 'uint8') check_quantized_npi_add((32, 56, 64, 11), 'int8', 'int8') check_quantized_npi_add((4, 6), 'uint8', 'int8', 'A') check_quantized_npi_add((13, 74, 52), 'uint8', 'uint8', 'B') check_quantized_npi_add((3, 4, 56, 56), 'int8', 'uint8', 'A') check_quantized_npi_add((32, 56, 64, 11), 'int8', 'int8', 'B') @use_np def test_quantized_elemwise_mul(): def check_quantized_elemwise_mul(data_shape, qtype): if is_test_for_native_cpu(): print('skipped testing quantized_elemwise_mul for native cpu since it is not supported yet') return elif qtype != 'int8': print('skipped testing quantized_elemwise_mul for not supported data type') return elif is_test_for_gpu(): print('skipped testing quantized_elemwise_mul for gpu since it is not supported yet') return class ElemwiseMulBlock(mx.gluon.nn.HybridBlock): def __init__(self, **kwargs): super(ElemwiseMulBlock, self).__init__(**kwargs) def forward(self, dataA, dataB): return mx.np.multiply(dataA, dataB) class QuantElemwiseMulBlock(mx.gluon.nn.HybridBlock): def __init__(self, **kwargs): super(QuantElemwiseMulBlock, self).__init__(**kwargs) def forward(self, dataA, dataB, dataA_min, dataA_max, dataB_min, dataB_max): return npx.quantized_elemwise_mul(dataA, dataB, dataA_min, dataA_max, dataB_min, dataB_max) elemwise_mul_fp32 = ElemwiseMulBlock() data_low, data_high = get_low_high(qtype) dataA_val = mx.np.random.uniform(low=data_low, high=data_high, size=data_shape).astype('int32').astype('float32') dataB_val = mx.np.random.uniform(low=data_low, high=data_high, size=data_shape).astype('int32').astype('float32') output = elemwise_mul_fp32(dataA_val, dataB_val) quantized_elemwise_mul = QuantElemwiseMulBlock() dataA_val_int8 = dataA_val.astype(qtype) dataB_val_int8 = dataB_val.astype(qtype) quantized_range = 127.0 min_dataA = mx.np.array([data_low]) max_dataA = mx.np.array([data_high]) min_dataB = mx.np.array([data_low]) max_dataB = mx.np.array([data_high]) qoutput, min_range, max_range = quantized_elemwise_mul(dataA_val_int8, dataB_val_int8, min_dataA, max_dataA, min_dataB, max_dataB) fp32_rslt = output.asnumpy() int8_rslt = qoutput.astype(output.dtype) assert_almost_equal(fp32_rslt, int8_rslt, atol = 1e-4) for qtype in ['int8', 'uint8']: check_quantized_elemwise_mul((4, 6), qtype) check_quantized_elemwise_mul((13, 74, 52), qtype) check_quantized_elemwise_mul((3, 4, 56, 56), qtype) check_quantized_elemwise_mul((32, 56, 64, 11), qtype) @use_np def test_quantized_pooling(): def check_quantized_pooling(data_shape, kernel, pool_type, pad, stride, global_pool, qdtype, convention='valid'): if is_test_for_native_cpu(): print('skipped testing quantized_pooling for native cpu since it is not supported yet') return elif qdtype == 'uint8' and is_test_for_gpu(): print('skipped testing quantized_pooling for gpu uint8 since it is not supported yet') return elif is_test_for_gpu() and len(data_shape) != 4: print('skipped testing quantized_pooling for gpu 5d layout since it is not supported yet') return class PoolingBlock(mx.gluon.nn.HybridBlock): def __init__(self, kernel=kernel, pad=pad, stride=stride, pool_type=pool_type, global_pool=global_pool, cudnn_off=False, pooling_convention=convention): super(PoolingBlock, self).__init__() self._kwargs = {'kernel': kernel, 'pad': pad, 'stride': stride, 'pool_type': pool_type, 'global_pool': global_pool, 'cudnn_off': False, 'pooling_convention': convention} def forward(self, data): return npx.pooling(data, **self._kwargs) class QuantPoolingBlock(mx.gluon.nn.HybridBlock): def __init__(self, kernel=kernel, pad=pad, stride=stride, pool_type=pool_type, global_pool=global_pool, cudnn_off=False, pooling_convention=convention): super(QuantPoolingBlock, self).__init__() self._kwargs = {'kernel': kernel, 'pad': pad, 'stride': stride, 'pool_type': pool_type, 'global_pool': global_pool, 'cudnn_off': False, 'pooling_convention':convention} def forward(self, data, min_data, max_data): return npx.quantized_pooling(data, min_data, max_data, **self._kwargs) pooling_fp32 = PoolingBlock() if qdtype == 'uint8': data_low = 0.0 data_high = 127.0 else: data_low = -127.0 data_high = 127.0 input_data = mx.np.random.uniform(low=data_low, high=data_high, size=data_shape ).astype('int32').astype('float32') output = pooling_fp32(input_data) quantized_pooling = QuantPoolingBlock(kernel=kernel, pad=pad, stride=stride, pool_type=pool_type, global_pool=global_pool, pooling_convention=convention) int8_input_data = input_data.astype(qdtype) quantized_range = 127.0 min_data = mx.np.array([-quantized_range]) max_data = mx.np.array([quantized_range]) qoutput, min_range, max_range = quantized_pooling(int8_input_data, min_data, max_data) if pool_type == 'max': assert_almost_equal(output.asnumpy(), qoutput.asnumpy()) elif pool_type == 'avg': # for avg pooling, fp32 and int8 may be different due to rounding errors diff = mx.np.abs(output - qoutput.astype(output.dtype)) cond = mx.np.less(2, diff).sum().item() assert cond == 0 for qdtype in ['int8', 'uint8']: check_quantized_pooling((3, 4, 56, 56), (3, 3), 'max', (0, 0), (2, 2), False, qdtype) check_quantized_pooling((3, 4, 56, 56), (3, 3), 'max', (0, 0), (2, 2), True, qdtype) check_quantized_pooling((3, 512, 7, 7), (7, 7), 'avg', (0, 0), (1, 1), False, qdtype) check_quantized_pooling((3, 512, 7, 7), (7, 7), 'avg', (0, 0), (1, 1), True, qdtype) check_quantized_pooling((3, 4, 3, 56, 56), (1, 3, 3), 'max', (0, 0, 0), (1, 2, 2), False, qdtype) check_quantized_pooling((3, 4, 3, 56, 56), (1, 3, 3), 'max', (0, 0, 0), (1, 2, 2), True, qdtype) check_quantized_pooling((3, 512, 3, 7, 7), (1, 7, 7), 'avg', (0, 0, 0), (1, 2, 2), False, qdtype) check_quantized_pooling((3, 512, 3, 7, 7), (1, 7, 7), 'avg', (0, 0, 0), (1, 2, 2), True, qdtype) check_quantized_pooling((3, 4, 56, 56), (3, 3), 'max', (0, 0), (2, 2), False, qdtype, 'full') check_quantized_pooling((3, 4, 56, 56), (3, 3), 'max', (0, 0), (2, 2), True, qdtype, 'full') check_quantized_pooling((3, 512, 7, 7), (7, 7), 'avg', (0, 0), (1, 1), False, qdtype, 'full') check_quantized_pooling((3, 512, 7, 7), (7, 7), 'avg', (0, 0), (1, 1), True, qdtype, 'full') check_quantized_pooling((3, 4, 3, 56, 56), (1, 3, 3), 'max', (0, 0, 0), (1, 2, 2), False, qdtype, 'full') check_quantized_pooling((3, 4, 3, 56, 56), (1, 3, 3), 'max', (0, 0, 0), (1, 2, 2), True, qdtype, 'full') check_quantized_pooling((3, 512, 3, 7, 7), (1, 7, 7), 'avg', (0, 0, 0), (1, 2, 2), False, qdtype, 'full') check_quantized_pooling((3, 512, 3, 7, 7), (1, 7, 7), 'avg', (0, 0, 0), (1, 2, 2), True, qdtype, 'full') @use_np def test_quantized_fc(): def check_quantized_fc(data_shape, num_hidden, use_bias, qdtype, flatten=True): if is_test_for_native_cpu(): hasMKL = False for key in os.environ.keys(): if operator.eq(key, "BUILD_TAG"): if os.environ['BUILD_TAG'].find("MKL") != -1: hasMKL = True break if hasMKL == False: print('skipped testing quantized_fc on cpu since s8u8s32 is only supported by MKL BLAS library') return elif qdtype == 'uint8' and is_test_for_gpu(): print('skipped testing quantized_fc for gpu uint8 since it is not supported yet') return def maxabs(a, b): return mx.np.maximum(mx.np.abs(a), mx.np.abs(b)) int8_range = 127.0 if qdtype == 'uint8': data_low = 0.0 data_high = 63.0 quantized_range = 255.0 else: data_low = -63.0 data_high = 63.0 quantized_range = 127.0 data = mx.np.random.uniform(low=data_low, high=data_high, size=data_shape ).astype('int32').astype('float32') fc_fp32 = mx.gluon.nn.Dense(units=num_hidden, use_bias=use_bias, flatten=flatten) fc_fp32.initialize() fc_fp32(data) npx.waitall() fp32_params = fc_fp32.collect_params() weight_shape = fp32_params['weight'].shape new_args = dict() new_args['weight'] = mx.np.random.uniform(low=data_low, high=data_high, size=fp32_params['weight'].shape ).astype('int32').astype('float32') data_min = mx.np.min(data).astype('float32') data_max = mx.np.max(data).astype('float32') weight_min = mx.np.min(new_args['weight']).astype('float32') weight_max = mx.np.max(new_args['weight']).astype('float32') data_range = maxabs(data_min, data_max) weight_range = maxabs(weight_min, weight_max) if use_bias: bias = mx.np.random.uniform(low=data_low, high=data_high, size=fp32_params['bias'].shape ).astype('int32').astype('float32') bias_min = mx.np.min(bias).astype('float32') bias_max = mx.np.max(bias).astype('float32') bias_range = maxabs(bias_min, bias_max) bias_scale = int8_range / bias_range data_scale = quantized_range / data_range weight_scale = int8_range / weight_range bias_int32_rescale = data_scale * weight_scale / bias_scale new_bias = bias.astype('float32') * bias_int32_rescale new_args['bias'] = new_bias.astype('int32').astype('float32') fc_fp32.load_dict(new_args, cast_dtype=True, dtype_source='saved') output = fc_fp32(data) class QuantFC(mx.gluon.nn.HybridBlock): def __init__(self, num_hidden, use_bias, flatten, **kwargs): super(QuantFC, self).__init__(**kwargs) self.use_bias = use_bias self._kwargs = {'num_hidden': num_hidden, 'no_bias': not use_bias, 'flatten': flatten} self.min_data = mx.gluon.Parameter('min_data', dtype='float32', shape=(1), allow_deferred_init=True) self.max_data = mx.gluon.Parameter('max_data', dtype='float32', shape=(1), allow_deferred_init=True) self.weight = mx.gluon.Parameter('weight', dtype='int8', shape=weight_shape, allow_deferred_init=True) self.min_weight = mx.gluon.Parameter('min_weight', dtype='float32', shape=(1), allow_deferred_init=True) self.max_weight = mx.gluon.Parameter('max_weight', dtype='float32', shape=(1), allow_deferred_init=True) if use_bias: self.bias = mx.gluon.Parameter('bias', dtype='int8', shape=(num_hidden,), allow_deferred_init=True) self.min_bias = mx.gluon.Parameter('min_bias', dtype='float32', shape=(1), allow_deferred_init=True) self.max_bias = mx.gluon.Parameter('max_bias', dtype='float32', shape=(1), allow_deferred_init=True) def forward(self, x): device = x.device weight = self.weight.data().to_device(device) bias = self.bias.data().to_device(device) if self.use_bias else None min_data = self.min_data.data().to_device(device) max_data = self.max_data.data().to_device(device) min_weight = self.min_weight.data().to_device(device) max_weight = self.max_weight.data().to_device(device) min_bias = self.min_bias.data().to_device(device) if self.use_bias else None max_bias = self.max_bias.data().to_device(device) if self.use_bias else None out = npx.quantized_fully_connected(data=x, weight=weight, bias=bias, min_data=min_data, max_data=max_data, min_weight=min_weight, max_weight=max_weight, min_bias=min_bias, max_bias=max_bias, **self._kwargs) return out fc_int8 = QuantFC(num_hidden=num_hidden, use_bias=use_bias, flatten=flatten) qargs = { 'weight': new_args['weight'].astype('int8'), 'min_data': mx.np.array([-data_range]), 'max_data': mx.np.array([data_range]), 'min_weight': mx.np.array([-weight_range]), 'max_weight': mx.np.array([weight_range]) } if use_bias: qargs.update({ 'bias': bias.astype('int8'), 'min_bias': mx.np.array([-bias_range]), 'max_bias': mx.np.array([bias_range]), }) fc_int8.load_dict(qargs, cast_dtype=True, dtype_source='saved') qoutput, min_range, max_range = fc_int8(data.astype(qdtype)) if use_bias: # with adding bias, accuracy loss should not be greater than one diff = mx.np.abs(output - qoutput.astype(output.dtype)) cond = mx.np.less(2, diff).sum().item() assert cond == 0 else: assert_almost_equal(output.asnumpy(), qoutput.asnumpy()) for qdtype in ['int8', 'uint8']: if is_test_for_dnnl(): check_quantized_fc((32, 512, 2), 100, False, qdtype, flatten=False) check_quantized_fc((32, 512, 2), 100, True, qdtype, flatten=False) check_quantized_fc((32, 512, 2, 2), 100, False, qdtype, flatten=False) check_quantized_fc((32, 512, 2, 2), 100, True, qdtype, flatten=False) check_quantized_fc((32, 512, 2, 2), 100, False, qdtype) check_quantized_fc((32, 111, 2, 2), 100, False, qdtype) check_quantized_fc((32, 512, 2, 2), 100, True, qdtype) check_quantized_fc((32, 111, 2, 2), 100, True, qdtype) check_quantized_fc((256, 2048, 2, 2), 800, True, qdtype) check_quantized_fc((256, 111, 2, 2), 800, True, qdtype) check_quantized_fc((256, 2048, 2, 2), 800, False, qdtype) check_quantized_fc((256, 111, 2, 2), 800, False, qdtype) @use_np def test_quantized_transpose(): def check_quantized_transpose(shape, qdtype, axes): data_low, data_high = get_low_high(qdtype) data = mx.np.random.uniform(low=data_low, high=data_high, size=shape).astype(qdtype).astype('float32') min_data = mx.np.array([mx.np.min(data).astype('float32').item()]) max_data = mx.np.array([mx.np.max(data).astype('float32').item()]) qdata = data.astype(qdtype) output = mx.np.transpose(data, axes=axes) qoutput, min_output, max_output = npx.quantized_transpose(qdata, min_data, max_data, axes=axes) assert_almost_equal(output.asnumpy(), qoutput.asnumpy()) assert_almost_equal(min_output.item(), min_data.item()) assert_almost_equal(max_output.item(), max_data.item()) for qtype in ['int8', 'uint8']: check_quantized_transpose((), qtype, ()) check_quantized_transpose((2,3), qtype, (1,0)) check_quantized_transpose((8,21), qtype, (1,0)) check_quantized_transpose((7,3,9), qtype, (2,1,0)) check_quantized_transpose((5,3,6,8), qtype, (2,3,0,1)) @use_np def test_quantized_embedding(): def check_quantized_embedding(data_shape, input_dim, output_dim): if is_test_for_gpu(): print('skipped testing test_quantized_embedding for gpu since it is not supported yet') return def maxabs(a, b): return mx.np.maximum(mx.np.abs(a), mx.np.abs(b)) data = mx.np.random.uniform(low=0, high=input_dim, size=data_shape ).astype('int32').astype('float32') embedding_fp32 = mx.gluon.nn.Embedding(input_dim=input_dim, output_dim=output_dim) embedding_fp32.initialize() embedding_fp32(data) npx.waitall() fp32_params = embedding_fp32.collect_params() weight_shape = fp32_params['weight'].shape int8_range = 127.0 new_params = dict() weight = mx.np.random.uniform(low=-int8_range, high=int8_range, size=weight_shape ).astype('int32').astype('float32') new_params['weight'] = weight embedding_fp32.load_dict(new_params, cast_dtype=True, dtype_source='saved') output = embedding_fp32(data) weight_min = mx.np.min(weight).astype('float32') weight_max = mx.np.max(weight).astype('float32') weight_range = maxabs(weight_min, weight_max) class QuantEmbedding(mx.gluon.nn.HybridBlock): def __init__(self, input_dim=input_dim, output_dim=output_dim, **kwargs): super(QuantEmbedding, self).__init__(**kwargs) self._kwargs = {'input_dim': input_dim, 'output_dim': output_dim} self.weight = mx.gluon.Parameter('weight', dtype='float32', shape=weight_shape, allow_deferred_init=True) self.min_weight = mx.gluon.Parameter('min_weight', dtype='float32', shape=(1), allow_deferred_init=True) self.max_weight = mx.gluon.Parameter('max_weight', dtype='float32', shape=(1), allow_deferred_init=True) def forward(self, x): device = x.device weight = self.weight.data().to_device(device) min_weight = self.min_weight.data().to_device(device) max_weight = self.max_weight.data().to_device(device) out = npx.quantized_embedding(data=x, weight=weight, min_weight=min_weight, max_weight=max_weight, **self._kwargs) return out embedding_int8 = QuantEmbedding(input_dim=input_dim, output_dim=output_dim) qargs = { 'weight': weight.astype('int8'), 'min_weight': mx.np.array([-weight_range]), 'max_weight': mx.np.array([weight_range]) } embedding_int8.load_dict(qargs, cast_dtype=True, dtype_source='saved') qoutput, min_range, max_range = embedding_int8(data) assert_almost_equal(output.asnumpy(), qoutput.asnumpy()) check_quantized_embedding((1,), 1000, 256) check_quantized_embedding((1,), 1024, 512) check_quantized_embedding((32,), 1000, 256) check_quantized_embedding((32,), 1024, 512) @use_np def test_quantized_flatten(): def check_quantized_flatten(shape, qdtype): if qdtype == 'uint8': data_low = 0.0 data_high = 127.0 else: data_low = -127.0 data_high = 127.0 qdata = mx.np.random.uniform(low=data_low, high=data_high, size=shape).astype(qdtype) min_data = mx.np.array([-1023.343], dtype='float32') max_data = mx.np.array([2343.324275], dtype='float32') qoutput, min_output, max_output = npx.quantized_flatten(qdata, min_data, max_data) assert qoutput.ndim == 2 assert qoutput.shape[0] == qdata.shape[0] assert qoutput.shape[1] == onp.prod(qdata.shape[1:]) assert same(qdata.asnumpy().flatten(), qoutput.asnumpy().flatten()) assert same(min_data.asnumpy(), min_output.asnumpy()) assert same(max_data.asnumpy(), max_output.asnumpy()) for qdtype in ['int8', 'uint8']: check_quantized_flatten((10,), qdtype) check_quantized_flatten((10, 15), qdtype) check_quantized_flatten((10, 15, 18), qdtype) check_quantized_flatten((3, 4, 23, 23), qdtype) @use_np def test_quantized_act(): def check_quantized_act(data_shape, qdtype): if is_test_for_native_cpu(): print('skipped testing quantized_act for native cpu since it is not supported yet') return elif qdtype == 'int8' and is_test_for_dnnl(): print('skipped testing quantized_act for oneDNN cpu int8 since it is not supported yet') return elif is_test_for_gpu(): print('skipped testing quantized_act for gpu since it is not supported yet') return act_fp32 = mx.gluon.nn.Activation(activation='relu') if qdtype == 'uint8': data_low = 0.0 data_high = 127.0 else: data_low = -127.0 data_high = 127.0 data = mx.np.random.uniform(low=data_low, high=data_high, size=data_shape ).astype(qdtype).astype('float32') output = act_fp32(data) class QuantActivation(mx.gluon.nn.HybridBlock): def __init__(self, activation, **kwargs): super(QuantActivation, self).__init__(**kwargs) self._kwargs = {'act_type': activation} def forward(self, x, min_data, max_data): out = npx.quantized_act(data=x, min_data=min_data, max_data=max_data, **self._kwargs) return out quantized_act = QuantActivation(activation='relu') qdata = data.astype(qdtype) quantized_range_min = mx.np.array([mx.np.min(data).astype('float32').item()]) quantized_range_max = mx.np.array([mx.np.max(data).astype('float32').item()]) qoutput, min_range, max_range = quantized_act(qdata, quantized_range_min, quantized_range_max) assert_almost_equal(output.asnumpy(), qoutput.asnumpy()) assert_almost_equal(min_range.item(), quantized_range_min.item()) assert_almost_equal(max_range.item(), quantized_range_max.item()) for qdtype in ['int8', 'uint8']: check_quantized_act((10,), qdtype) check_quantized_act((10, 15), qdtype) check_quantized_act((10, 15, 18), qdtype) check_quantized_act((3, 4, 23, 23), qdtype) @use_np def test_quantized_bn(): def get_mean_var(data): axes = list(range(data.ndim)) del axes[1] mean = mx.np.mean(data, axis=axes) mean_broad = mx.np.expand_dims(mean, axis=0) mean_broad = mx.np.expand_dims(mean_broad, axis=2) mean_broad = mx.np.expand_dims(mean_broad, axis=3) mean_broad = mx.npx.broadcast_like(mean_broad, data) var = mx.np.multiply(data - mean_broad, data - mean_broad) axes = list(range(var.ndim)) del axes[1] var = mx.np.mean(var, axis=axes) return mean, var def check_quantized_bn(data_shape, qdtype): if is_test_for_native_cpu(): print('skipped testing quantize_bn for native cpu since it is not supported yet') return elif is_test_for_gpu(): print('skipped testing quantize_bn for gpu since it is not supported yet') return data_low, data_high = get_low_high(qdtype) # run fp32 bn bn_fp32 = mx.gluon.nn.BatchNorm(use_global_stats=True, scale=True) data = mx.np.random.uniform(low=data_low, high=data_high, size=data_shape) bn_fp32.initialize() bn_fp32.hybridize() bn_fp32(data) fp32_params = bn_fp32.collect_params() data = mx.np.random.uniform(low=data_low, high=data_high, size=data_shape) gamma = mx.np.random.uniform(low=data_low, high=data_high, size=fp32_params['gamma'].shape) beta = mx.np.random.uniform(low=data_low, high=data_high, size=fp32_params['beta'].shape) running_mean, running_var = get_mean_var(data) new_params = { 'gamma':gamma, 'beta':beta, 'running_mean': running_mean, 'running_var': running_var } bn_fp32.load_dict(new_params) output = bn_fp32(data) # generate int8 bn from fp32 bn calib_data = mx.gluon.data.DataLoader(data, batch_size=data_shape[0]) quant_bn = mx.contrib.quant.quantize_net(bn_fp32, quantized_dtype=qdtype, quantize_mode='full', calib_data=calib_data, calib_mode='naive', num_calib_batches=1, device=mx.current_device()) output_int8_to_fp32 = quant_bn(data) assert_almost_equal(output.asnumpy(), output_int8_to_fp32.asnumpy(), rtol=1e-1, atol=8) for qdtype in ['int8', 'uint8']: check_quantized_bn((32, 512, 4, 4), qdtype) check_quantized_bn((32, 1024, 8, 8), qdtype) check_quantized_bn((32, 3, 224, 224), qdtype) def test_quantized_reshape(): test_cases = [((2, 3, 5, 5), (-2, -1), False, (2, 75)), ((2, 3, 5, 5), (-2, -2, -1), False, (2, 3, 25)), ((5, 3, 4, 5), (-2, -1, -2), False, (5, 15, 4)), ((2, 3, 5, 4), (-1, -2, -2), False, (8, 3, 5)), ((2, 3, 5, 5), (-2, -2, -2, -2), False, (2, 3, 5, 5)), ((2, 1, 4, 5), (-2, -3, -2, -2), False, (2, 4, 5)), ((1, 1, 4, 1), (-3, -3, -2, -2), False, (4, 1)), ((1, 1, 1, 1), (-3, -3, -3, -3), False, ()), ((2, 4, 5, 3), (-1, 2, 2, 1), False, (30, 2, 2, 1)), ((2, 3, 5, 6), (-4,), False, (2, 3, 5, 6)), ((2, 3, 5, 6), (6, 1, -4), False, (6, 1, 5, 6)), ((2, 3, 5, 6), (-5, -5), False, (6, 30)), ((2, 3, 5, 6), (-5, -1), False, (6, 30)), ((64,), (-6, 16, 4), False, (16, 4)), ((64,), (-6, 16, -1), False, (16, 4)), ((64, 1, 2, 3), (-6, 16, -1, -4), False, (16, 4, 1, 2, 3)), ((8, 5, 4, 6), (-4, -1, 3, -6), True, (8, 5, 4, 2, 3))] def check_quantized_reshape(shape, qdtype, newshape, reverse, expected_ret_shape): data_low, data_high = get_low_high(qdtype) qdata = mx.np.random.uniform(low=data_low, high=data_high, size=shape).astype(qdtype) min_data = mx.np.array([-1023.343], dtype='float32') max_data = mx.np.array([2343.324275], dtype='float32') qoutput, min_output, max_output = npx.quantized_reshape(qdata, min_data, max_data, newshape=newshape, reverse=reverse) assert qoutput.shape == expected_ret_shape assert same(qdata.asnumpy().flatten(), qoutput.asnumpy().flatten()) assert same(min_data.asnumpy(), min_output.asnumpy()) assert same(max_data.asnumpy(), max_output.asnumpy()) for qdtype in ['int8', 'uint8']: for shape, newshape, reverse, expected_ret_shape in test_cases: check_quantized_reshape(shape, qdtype, newshape, reverse, expected_ret_shape) def test_quantize_params(): if is_test_for_native_cpu(): print('skipped testing quantized_params for native cpu since it is not supported yet') return data = mx.sym.Variable('data') conv = mx.sym.Convolution(data, kernel=(1, 1), num_filter=2048, name='conv') sym = mx.sym.BatchNorm(data=conv, eps=2e-05, fix_gamma=False, momentum=0.9, use_global_stats=False, name='bn') offline_params = [name for name in sym.list_arguments() if not name.startswith('data') and not name.endswith('label')] params = {} for name in offline_params: params[name] = mx.nd.uniform(shape=(2, 2)) qsym, _ = mx.contrib.quant._quantize_symbol(sym, device=mx.current_device(), offline_params=offline_params, quantize_mode='full') qparams = mx.contrib.quant._quantize_params(qsym, params, min_max_dict = {}) param_names = params.keys() qparam_names = qparams.keys() for name in qparam_names: if name.startswith('bn'): assert name in param_names elif name.startswith('conv'): assert name not in param_names assert name.find('quantize') != -1 class FP32Net(mx.gluon.nn.HybridBlock): def __init__(self, **kwargs): super(FP32Net, self).__init__(**kwargs) self.conv = mx.gluon.nn.Conv2D(channels=16, kernel_size=(1,1)) self.bn = mx.gluon.nn.BatchNorm(epsilon=2e-05, scale=True, momentum=0.9, use_global_stats=False) self.act = mx.gluon.nn.Activation(activation='relu') self.pool = mx.gluon.nn.AvgPool2D(pool_size=(4,4)) self.fc = mx.gluon.nn.Dense(units=10, flatten=True) def forward(self, x): out = self.conv(x) out = self.bn(out) out = self.act(out) out = self.pool(out) out = self.fc(out) return npx.softmax(out) class FP32MultipleOutputs(mx.gluon.nn.HybridBlock): def __init__(self, length, **kwargs): super(FP32MultipleOutputs, self).__init__(**kwargs) self.length = length self.convs = mx.gluon.nn.Conv2D(channels=16, kernel_size=(1,1)) self.fc = mx.gluon.nn.Dense(units=10, flatten=True) def forward(self, x): res = npx.slice_channel(x, num_outputs=self.length, axis=1, squeeze_axis=1) out = [] for i in range(self.length): out.append(self.convs(res[i])) out[i] = mx.np.expand_dims(out[i], axis=0) out = mx.np.concatenate(out) out = mx.np.reshape(out, ((self.length, -1))) out = self.fc(out) return npx.softmax(out) class FP32MultipleInputs(mx.gluon.nn.HybridBlock): def __init__(self, **kwargs): super(FP32MultipleInputs, self).__init__(**kwargs) self.conv1 = mx.gluon.nn.Conv2D(channels=64, kernel_size=(1,1), use_bias=False) self.bn1 = mx.gluon.nn.BatchNorm() self.conv2 = mx.gluon.nn.Conv2D(channels=64, kernel_size=(1,1), use_bias=False) self.bn2 = mx.gluon.nn.BatchNorm() def forward(self, data0, data1): out0 = self.conv1(data0) out0 = self.bn1(out0) out1 = self.conv2(data1) out1 = self.bn2(out1) return out1 + out0 @use_np @xfail_when_nonstandard_decimal_separator def test_quantize_model(): def check_params(params, qparams, qsym=None): if qsym is None: assert len(params) == len(qparams) for k, v in params.items(): assert k in qparams assert same(v.asnumpy(), qparams[k].asnumpy()) else: qparams_ground_truth = mx.contrib.quant._quantize_params(qsym, params, min_max_dict = {}) assert len(qparams) == len(qparams_ground_truth) for k, v in qparams_ground_truth.items(): assert k in qparams assert same(v.asnumpy(), qparams[k].asnumpy()) def check_qsym_calibrated(qsym): attrs = qsym.attr_dict() for k, v in attrs.items(): if k.find('requantize_') != -1: assert 'min_calib_range' in v assert 'max_calib_range' in v def check_qsym_qdtype(qsym, qdtype): attrs = qsym.attr_dict() for k, v in attrs.items(): if k.find('_quantize') != -1: assert 'out_type' in v assert v['out_type'] == qdtype def skip_not_supported(): if is_test_for_native_cpu(): print('skipped testing quantize_model for native cpu since it is not supported yet') return True elif qdtype == 'int8' and is_test_for_dnnl(): print('skipped testing quantize_model for oneDNN cpu int8 since it is not supported yet') return True elif qdtype == 'uint8' and is_test_for_gpu(): print('skipped testing quantize_model for gpu uint8 since it is not supported yet') return True return False def check_quantize_model(qdtype): if is_test_for_native_cpu(): print('skipped testing quantize_model for native cpu since it is not supported yet') return elif qdtype == 'int8' and is_test_for_dnnl(): print('skipped testing quantize_model for oneDNN cpu int8 since it is not supported yet') return elif qdtype == 'uint8' and is_test_for_gpu(): print('skipped testing quantize_model for gpu uint8 since it is not supported yet') return standard_net = FP32Net() standard_net.initialize() batch_size = 4 data_shape = (batch_size, 4, 10, 10) length = batch_size # specify num of outputs from split op multi_out_net = FP32MultipleOutputs(length) multi_out_net.initialize() multi_out_data_shape = (length, 4, 4, 10, 10) for net, dshape in zip((standard_net, multi_out_net), (data_shape, multi_out_data_shape)): data = mx.np.random.uniform(low=0, high=1, size=dshape) net.hybridize() net(data) sym, _ = net.export(None) arg_params, aux_params = collect_block_args_aux(net, sym) qsym, qarg_params, qaux_params = mx.contrib.quant.quantize_model(sym=sym, arg_params=arg_params, aux_params=aux_params, device=mx.current_device(), quantized_dtype=qdtype, calib_mode='none', quantize_mode='full') check_params(arg_params, qarg_params, qsym) check_params(aux_params, qaux_params) calib_data = mx.np.random.uniform(size=dshape) calib_data = mx.gluon.data.DataLoader(calib_data, batch_size=batch_size) qsym, qarg_params, qaux_params = mx.contrib.quant.quantize_model(sym=sym, arg_params=arg_params, aux_params=aux_params, device=mx.current_device(), quantized_dtype=qdtype, calib_mode='naive', calib_data=calib_data, num_calib_batches=1, quantize_mode='full') check_params(arg_params, qarg_params, qsym) check_params(aux_params, qaux_params) check_qsym_calibrated(qsym) check_qsym_qdtype(qsym, qdtype) def check_quantize_model_multiple_inputs(qdtype): if skip_not_supported(): return net = FP32MultipleInputs() net.initialize() net.hybridize() dshape = (64, 4, 10, 10) data = [mx.np.random.uniform(low=0, high=1, size=dshape), mx.np.random.uniform(low=0, high=1, size=dshape)] net(*data) sym, _ = net.export(None) arg_params, aux_params = collect_block_args_aux(net, sym) qsym, qarg_params, qaux_params = mx.contrib.quant.quantize_model(sym=sym, arg_params=arg_params, aux_params=aux_params, device=mx.current_device(), quantized_dtype=qdtype, calib_mode='none', quantize_mode='full') check_params(arg_params, qarg_params, qsym) check_params(aux_params, qaux_params) calib_data = [mx.np.random.uniform(size=dshape), mx.np.random.uniform(size=dshape)] calib_data = mx.gluon.data.DataLoader(mx.gluon.data.ArrayDataset(*calib_data), batch_size=4) qsym, qarg_params, qaux_params = mx.contrib.quant.quantize_model(sym=sym, arg_params=arg_params, aux_params=aux_params, device=mx.current_device(), quantized_dtype=qdtype, calib_mode='naive', calib_data=calib_data, data_names=["data0","data1"], num_calib_batches=1, quantize_mode='full') check_params(arg_params, qarg_params, qsym) check_params(aux_params, qaux_params) check_qsym_calibrated(qsym) check_qsym_qdtype(qsym, qdtype) for qdtype in ['int8', 'uint8']: check_quantize_model(qdtype) check_quantize_model_multiple_inputs(qdtype) @mx.util.use_np def test_quantize_gluon_with_forward(): def check_quantize_net(qdtype): if is_test_for_native_cpu(): print('skipped testing test_quantize_model_with_forward for native cpu since it is not supported yet') return elif is_test_for_gpu(): print('skipped testing test_quantize_model_with_forward for gpu uint8 since it is not supported yet') return data_shape = (32, 3, 224, 224) batch_size = 1 resnet18_v1 = vision.resnet18_v1(pretrained=True) resnet18_v1.reset_device(mx.current_device()) excluded_names_match = [] if mx.current_device() == mx.gpu(): excluded_names_match += ['activation', 'relu', 'conv0'] num_calib_batches = 1 random_data = mx.np.random.uniform(size=data_shape) calib_data = mx.gluon.data.DataLoader(random_data, batch_size=batch_size) quantized_resnet18_v1 = mx.contrib.quant.quantize_net(resnet18_v1, quantized_dtype=qdtype, exclude_layers=None, exclude_layers_match=excluded_names_match, calib_mode='none', data_shapes=[data_shape], device=mx.current_device()) quantized_resnet18_v1.hybridize(static_alloc=True, static_shape=True) quantized_resnet18_v1(random_data) for mode in ['naive', 'entropy']: for quantize_granularity in ['tensor-wise', 'channel-wise']: qdtype = qdtype if mode == 'naive' else 'auto' quantized_resnet18_v1 = mx.contrib.quant.quantize_net(resnet18_v1, quantized_dtype=qdtype, exclude_layers=None, exclude_layers_match=excluded_names_match, calib_data=calib_data, calib_mode=mode, quantize_granularity=quantize_granularity, num_calib_batches=num_calib_batches, device=mx.current_device()) quantized_resnet18_v1.hybridize(static_alloc=True, static_shape=True) quantized_resnet18_v1(random_data) for qdtype in ['int8', 'uint8']: check_quantize_net(qdtype) @xfail_when_nonstandard_decimal_separator def test_quantize_sym_with_calib(): if is_test_for_native_cpu(): print('skipped testing quantized_pooling for native cpu since it is not supported yet') return def get_fp32_sym(): data = mx.sym.Variable('data') conv = mx.sym.Convolution(data, kernel=(1, 1), num_filter=16, name='conv') bn = mx.sym.BatchNorm(data=conv, eps=2e-05, fix_gamma=False, momentum=0.9, use_global_stats=False, name='bn') act = mx.sym.Activation(data=bn, act_type='relu', name='relu') pool = mx.sym.Pooling(act, kernel=(4, 4), pool_type='avg', name='pool') fc = mx.sym.FullyConnected(pool, num_hidden=10, flatten=True, name='fc') sym = mx.sym.softmax(fc, name='softmax') return sym sym = get_fp32_sym() offline_params = [name for name in sym.list_arguments() if not name.startswith('data') and not name.endswith('label')] qsym, _ = mx.contrib.quant._quantize_symbol(sym, device=mx.current_device(), offline_params=offline_params, quantize_mode='full') requantize_op_names = ['requantize_conv', 'requantize_fc'] min_max_dict = {'conv_output': (onp.random.uniform(low=100.0, high=200.0), onp.random.uniform(low=100.0, high=200.0)), 'fc_output': (onp.random.uniform(low=100.0, high=200.0), onp.random.uniform(low=100.0, high=200.0))} op_name_to_th_name = {'requantize_conv': 'conv_output', 'requantize_fc': 'fc_output'} cqsym = mx.contrib.quant._calibrate_quantized_sym(qsym, min_max_dict) attr_dict = cqsym.attr_dict() for name in requantize_op_names: assert name in attr_dict lhs = float(attr_dict[name]['min_calib_range']) rhs = min_max_dict[op_name_to_th_name[name]][0] assert_almost_equal(onp.array([lhs]), onp.array([rhs])) lhs = float(attr_dict[name]['max_calib_range']) rhs = min_max_dict[op_name_to_th_name[name]][1] assert_almost_equal(onp.array([lhs]), onp.array([rhs]), rtol=1e-3, atol=1e-4) @use_np def test_quantization_net_with_different_data_inputs_options(): if is_test_for_native_cpu(): print('skipped testing test_quantization_net_with_different_data_inputs_options for native cpu since it is not supported yet') return elif is_test_for_gpu(): print('skipped testing test_quantization_net_with_different_data_inputs_options for gpu since it is not supported yet') return net = FP32Net() net.initialize() batch_size = 32 data_shape = (batch_size, 3, 224, 224) random_data = mx.np.random.uniform(size=data_shape) # pass data_shapes as list of tuples quantized_net = mx.contrib.quant.quantize_net(net, quantized_dtype='auto', data_shapes=[data_shape], device=mx.current_device()) out = quantized_net(random_data) out.wait_to_read() # pass data_shapes as list of DataDescs net2 = FP32Net() net2.initialize() data_desc = mx.io.DataDesc('data', data_shape) quantized_net2 = mx.contrib.quant.quantize_net(net2, quantized_dtype='auto', data_shapes=[data_desc], device=mx.current_device()) out2 = quantized_net2(random_data) out2.wait_to_read() # pass data as DataLoader net3 = FP32Net() net3.initialize() data_loader = mx.gluon.data.DataLoader(random_data, batch_size=batch_size) quantized_net3 = mx.contrib.quant.quantize_net(net3, quantized_dtype='auto', calib_data=data_loader, device=mx.current_device()) out3 = quantized_net3(random_data) out3.wait_to_read() def test_optimal_threshold_adversarial_case(): # The worst case for the optimal_threshold function is when the values are concentrated # at one edge: [0, 0, ..., 1000]. (histogram) # We want to make sure that the optimal threshold in this case is the max. hist = [] hist_edges = [] min_val = -2 max_val = 2 for _ in range(0, 998): hist.append(0) for i in range(0, 999): hist_edges.append((max_val - min_val) / 999 * i + min_val) hist.append(1000) hist_edges.append(max_val) hist_data = (hist, hist_edges, min_val, max_val, max_val) for dtype in ['uint8', 'int8', 'auto']: res = mx.contrib.quant._LayerHistogramCollector.get_optimal_threshold(hist_data, dtype, num_quantized_bins=5) # The threshold should be 2. print (res) assert abs(res[2] - 2) < 1e-5 def test_get_optimal_thresholds(): # Given an ndarray with elements following a uniform distribution, the optimal threshold # for quantizing the ndarray should be either abs(min(nd)) or abs(max(nd)). def get_threshold(nd): min_nd = mx.nd.min(nd) max_nd = mx.nd.max(nd) return mx.nd.maximum(mx.nd.abs(min_nd), mx.nd.abs(max_nd)).asnumpy() for dtype in ['uint8', 'int8', 'auto']: nd = mx.nd.uniform(low=-10.532, high=11.3432, shape=(8, 3, 23, 23), dtype=onp.float64) expected_threshold = get_threshold(nd) arr = nd.asnumpy() min_range = onp.min(arr) max_range = onp.max(arr) th = max(abs(min_range), abs(max_range)) hist, hist_edges = onp.histogram(arr, bins=8001, range=(-th, th)) hist_dict = {'layer1' : (hist, hist_edges, min_range, max_range, th)} min_max_dict = mx.contrib.quant._LayerHistogramCollector.get_optimal_thresholds(hist_dict, dtype) assert 'layer1' in min_max_dict assert_almost_equal(onp.array([min_max_dict['layer1'][1]]), expected_threshold, rtol=1e-2, atol=1e-4) @use_np def test_rnn_quantization(): data_low = -1 data_high = 1 def check_rnn_quantization(num_layers, bidirectional, seq_len, batch_size, input_dim, state_size): data_shape = (seq_len, batch_size, input_dim) rnn_fp32 = mx.gluon.rnn.LSTM(hidden_size=state_size, num_layers = num_layers, bidirectional=bidirectional) data = mx.np.random.uniform(low=data_low, high=data_high, size=data_shape) states_shape = (num_layers * 2 if bidirectional else num_layers, batch_size, state_size) states = [mx.np.zeros((states_shape)) for _ in range(batch_size)] rnn_fp32.initialize() rnn_fp32.hybridize() ref_out = rnn_fp32(data, states) class RNNDataLoader(mx.gluon.data.DataLoader): def __init__(self, data, states): super().__init__(mx.gluon.data.SimpleDataset([]), 1) self.data = data self.states = states def __iter__(self): return self def __next__(self): return [self.data, self.states] def __bool__(self): return bool(self.dataiter.iter_next()) calib_data = RNNDataLoader(data, states) quant_rnn = mx.contrib.quant.quantize_net(rnn_fp32, quantized_dtype='auto', quantize_mode='full', calib_data=calib_data, calib_mode='naive', num_calib_batches=1, device=mx.current_device()) qout = quant_rnn(data, states) qsym, _ = quant_rnn.export(None) assert qsym.tojson().find("quantized_rnn") != -1 ref_out = [ref_out[0], ref_out[1][0], ref_out[1][1]] for i in range(len(qout)): mse = onp.mean((ref_out[i].asnumpy() - qout[i].asnumpy())**2) assert mse < 0.001 check_rnn_quantization(1, False, 5, 2, 16, 16) check_rnn_quantization(1, True, 5, 2, 16, 16) @use_np def test_quantized_rnn(): def check_quantized_rnn(num_layers, bidirectional, seq_len, batch_size, input_dim, state_size): ndir = 2 if bidirectional else 1 size = ndir*state_size*4 first_lyr_param_size = (input_dim + state_size + 2) * size other_lyr_param_size = (state_size * ndir + state_size + 2) * size full_param_size = first_lyr_param_size + (num_layers - 1) * other_lyr_param_size data = mx.np.random.uniform(-1, 1, (seq_len, batch_size, input_dim)) state = mx.np.random.uniform(-1, 1, (num_layers*ndir, batch_size, state_size)) state_cell = mx.np.random.uniform(0, 1, (num_layers*ndir, batch_size, state_size)) params = mx.np.random.normal(0, 1, (full_param_size,)) out = npx.rnn(data=data, parameters=params, mode='lstm', state=state, state_size=state_size, state_cell=state_cell, num_layers=num_layers, bidirectional=bidirectional) data_min = mx.np.min(data) data_max = mx.np.max(data) data_scale = mx.np.array(128.0 / (data_max - data_min)).reshape((1,)) data_shift = mx.np.array(128.0 - data_max * data_scale).reshape((1,)) qdata = (data * data_scale + data_shift + 0.5).astype('uint8') qout = npx.contrib_quantized_rnn(data=qdata, parameters=params, mode='lstm', state=state, state_size=state_size, state_cell=state_cell, num_layers=num_layers, bidirectional=bidirectional, data_scale=data_scale, data_shift=data_shift) mse = onp.mean((out.asnumpy() - qout.asnumpy())**2) assert mse < 0.001 check_quantized_rnn(1, False, 5, 2, 16, 16) check_quantized_rnn(1, True, 5, 2, 16, 16)