| /external/tensorflow/tensorflow/contrib/lite/toco/tflite/ |
| D | import.cc | 73 auto quantization = input_tensor->quantization(); in ImportTensors() local
74 if (quantization) { in ImportTensors()
77 if (quantization->min() && quantization->max()) { in ImportTensors()
78 CHECK_EQ(1, quantization->min()->Length()); in ImportTensors()
79 CHECK_EQ(1, quantization->max()->Length()); in ImportTensors()
81 minmax.min = quantization->min()->Get(0); in ImportTensors()
82 minmax.max = quantization->max()->Get(0); in ImportTensors()
84 if (quantization->scale() && quantization->zero_point()) { in ImportTensors()
85 CHECK_EQ(1, quantization->scale()->Length()); in ImportTensors()
86 CHECK_EQ(1, quantization->zero_point()->Length()); in ImportTensors()
[all …]
|
| /external/gemmlowp/doc/ |
| D | public.md | 15 rationale for a specific quantization paradigm is given in
16 [quantization.md](quantization.md). That specific quantization paradigm is
97 quantization paradigm explained in [quantization.md](quantization.md) that
102 pipeline (see [output.md](output.md)). This is the part of the quantization
103 paradigm explained in [quantization.md](quantization.md) that needs to be
146 quantization", whence the PC suffix. This has been useful in some settings where
149 the need for per-channel quantization. For that reason, the long-term usefulness
155 section of [low-precision.md](low-precision.md) on the legacy quantization
|
| D | quantization.md | 1 # Building a quantization paradigm from first principles
12 quantization paradigm affects the calculations that gemmlowp itself needs to
25 naturally at some specific quantization paradigm, and how that can be
28 We also aim to show how that differs from the older, legacy quantization
30 newer quantization paradigm described in this document was useful as far as some
75 ## The final form of the quantization equation
78 representable — means in either quantization equations, (1) and (2).
111 In other words, `D = -zero_point`. This suggests rewriting the quantization
122 With this quantization equation (3), the condition that 0 be exactly
148 above equation (3), with some already-known quantization parameters `lhs_scale`,
[all …]
|
| D | output.md | 10 quantization paradigms. See [low-precision.md](low-precision.md) and
11 [quantization.md](quantization.md).
13 Besides implementing a quantization paradigm, the other thing that output
51 quantized matrix multiplication with a sounds quantization paradigm, is here:
|
| D | low-precision.md | 31 Refer to [quantization.md](quantization.md) for details of how one gets from
62 matrix of uint8 quantized values - the following int32 "quantization
86 [quantization.md](quantization.md) for how reasoning from first principles, one
87 arrives to a substantially different quantization paradigm.
|
| /external/tensorflow/tensorflow/contrib/quantization/python/ |
| D | __init__.py | 22 from tensorflow.contrib.quantization.python.array_ops import *
23 from tensorflow.contrib.quantization.python.math_ops import *
24 from tensorflow.contrib.quantization.python.nn_ops import *
|
| /external/tensorflow/tensorflow/contrib/quantization/ |
| D | __init__.py | 23 from tensorflow.contrib.quantization.python import array_ops as quantized_array_ops
24 from tensorflow.contrib.quantization.python.math_ops import *
25 from tensorflow.contrib.quantization.python.nn_ops import *
|
| /external/tensorflow/tensorflow/core/api_def/base_api/ |
| D | api_def_QuantizeAndDequantizeV2.pbtxt | 26 If the quantization is signed or unsigned.
32 The bitwidth of the quantization.
45 quantization method when it is used in inference.
52 quantization), so that 0.0 maps to 0.
62 Next, we choose our fixed-point quantization buckets, [min_fixed, max_fixed].
|
| D | api_def_FakeQuantWithMinMaxArgs.pbtxt | 6 `inputs` values are quantized into the quantization range (`[0; 2^num_bits - 1]`
9 `num_bits` is the bitwidth of the quantization; between 2 and 8, inclusive.
|
| D | api_def_FakeQuantWithMinMaxVars.pbtxt | 8 `inputs` values are quantized into the quantization range (`[0; 2^num_bits - 1]`
11 `num_bits` is the bitwidth of the quantization; between 2 and 8, inclusive.
|
| D | api_def_FakeQuantWithMinMaxVarsPerChannel.pbtxt | 9 `inputs` values are quantized into the quantization range (`[0; 2^num_bits - 1]`
12 `num_bits` is the bitwidth of the quantization; between 2 and 8, inclusive.
|
| D | api_def_Dequantize.pbtxt | 52 `SCALED` mode matches the quantization approach used in
57 -127 to 127, not -128 to 127 for signed 8 bit quantization), so that 0.0 maps to
68 Next, we choose our fixed-point quantization buckets, `[min_fixed, max_fixed]`.
|
| D | api_def_QuantizeV2.pbtxt | 80 `SCALED` mode matches the quantization approach used in
85 -127 to 127, not -128 to 127 for signed 8 bit quantization), so that 0.0 maps to
96 Next, we choose our fixed-point quantization buckets, `[min_fixed, max_fixed]`.
120 requested minimum and maximum values slightly during the quantization process,
|
| /external/webp/src/enc/ |
| D | predictor_enc.c | 151 uint8_t boundary, int quantization) { in NearLosslessComponent() argument
154 const int lower = residual & ~(quantization - 1); in NearLosslessComponent()
155 const int upper = lower + quantization; in NearLosslessComponent()
165 return lower + (quantization >> 1); in NearLosslessComponent()
174 return lower + (quantization >> 1); in NearLosslessComponent()
189 int quantization; in NearLossless() local
196 quantization = max_quantization; in NearLossless()
197 while (quantization >= max_diff) { in NearLossless()
198 quantization >>= 1; in NearLossless()
204 a = NearLosslessComponent(value >> 24, predict >> 24, 0xff, quantization); in NearLossless()
[all …]
|
| /external/libjpeg-turbo/ |
| D | wizard.txt | 21 jcparam.c. At very low quality settings, some quantization table entries
22 can get scaled up to values exceeding 255. Although 2-byte quantization
26 quantization values to no more than 255 by giving the -baseline switch.
30 You can substitute a different set of quantization values by using the
33 -qtables file Use the quantization tables given in the named file.
35 The specified file should be a text file containing decimal quantization
44 duplicates the default quantization tables:
72 the quantization values are constrained to the range 1-255.
74 By default, cjpeg will use quantization table 0 for luminance components and
78 -qslots N[,...] Select which quantization table to use for
[all …]
|
| D | usage.txt | 73 -quality N[,...] Scale quantization tables to adjust image quality.
113 -quality 100 will generate a quantization table of all 1's, minimizing loss
114 in the quantization step (but there is still information loss in subsampling,
124 quantization tables, which are considered optional in the JPEG standard.
131 separate settings for every quantization table slot.) The principle is the
143 quantization table slots. If there are more q-table slots than parameters,
147 customized) quantization tables can be set with the -qtables option and
236 -baseline Force baseline-compatible quantization tables to be
237 generated. This clamps quantization values to 8 bits
243 -qtables file Use the quantization tables given in the specified
[all …]
|
| /external/tensorflow/tensorflow/contrib/lite/toco/ |
| D | toco_flags.proto | 48 // quantization of input arrays, separately from other arrays.
60 // the uint8 values are interpreted as real numbers, and the quantization
66 // the representation (quantization) of real numbers in the output file,
74 // file to choose a different real-numbers representation (quantization)
91 // with quantization of models. Normally, quantization requires the input
99 // allowing for quantization to proceed.
109 // generate plain float code without fake-quantization from a quantized
|
| /external/tensorflow/tensorflow/contrib/lite/toco/g3doc/ |
| D | cmdline_reference.md | 52 control their quantization or dequantization, effectively switching
62 * Some transformation flags allow to carry on with quantization when the
131 the (de-)quantization parameters of the input array, when it is quantized.
143 performed by the inference code; however, the quantization parameters of
166 output file, that is, controls the representation (quantization) of real
175 to choose a different real-numbers representation (quantization) from what
191 allows to control specifically the quantization of input arrays, separately
204 are interpreted as real numbers, and the quantization parameters used for
209 These flags enable what is called "dummy quantization". If defined, their
212 allowing to proceed with quantization of non-quantized or
[all …]
|
| /external/tensorflow/tensorflow/contrib/quantize/ |
| D | README.md | 2 model quantization of weights, biases and activations during both training and
4 [fake quantization op]
10 quantization operation during training in both the forward and backward passes.
11 The fake quantization operator achieves this by modeling the quantizer as a pass
|
| /external/tensorflow/tensorflow/contrib/lite/ |
| D | model.cc | 638 TfLiteQuantizationParams quantization; in ParseTensors() local
639 quantization.scale = 0; in ParseTensors()
640 quantization.zero_point = 0; in ParseTensors()
641 auto* q_params = tensor->quantization(); in ParseTensors()
647 if (q_params->scale()) quantization.scale = q_params->scale()->Get(0); in ParseTensors()
649 quantization.zero_point = q_params->zero_point()->Get(0); in ParseTensors()
706 i, type, get_name(tensor), dims, quantization, buffer_ptr, in ParseTensors()
714 i, type, get_name(tensor), dims, quantization) != kTfLiteOk) { in ParseTensors()
|
| D | context.c | 68 TfLiteQuantizationParams quantization, char* buffer, in TfLiteTensorReset() argument
75 tensor->params = quantization; in TfLiteTensorReset()
|
| /external/ImageMagick/config/ |
| D | Makefile.am | 42 config/quantization-table.xml \
66 config/quantization-table.xml \
|
| /external/glide/third_party/gif_encoder/ |
| D | LICENSE | 13 NEUQUANT Neural-Net quantization algorithm by Anthony Dekker, 1994. See
14 "Kohonen neural networks for optimal colour quantization" in "Network:
|
| /external/tensorflow/tensorflow/docs_src/api_guides/python/ |
| D | array_ops.md | 79 ## Fake quantization
80 Operations used to help train for better quantization accuracy.
|
| /external/tensorflow/tensorflow/docs_src/performance/ |
| D | index.md | 41 * @{$quantization$How to Quantize Neural Networks with TensorFlow}, which
42 can explains how to use quantization to reduce model size, both in storage
|