FM / server

/usr/local/lib64/python3.6/site-packages/caffe2/quantization/server/

/usr/local/lib64/python3.6/site-packages/caffe2/quantization/server

Name	Size	Mode	Actions
__pycache__/	-	0755	rm
batch_matmul_dnnlowp_op_test.py	10152	0644	edit dl rm
batch_permutation_dnnlowp_op_test.py	1449	0644	edit dl rm
channel_shuffle_dnnlowp_op_test.py	3890	0644	edit dl rm
compute_equalization_scale_test.py	3154	0644	edit dl rm
concat_dnnlowp_op_test.py	3265	0644	edit dl rm
conv_depthwise_dnnlowp_op_test.py	10931	0644	edit dl rm
conv_dnnlowp_acc16_op_test.py	14202	0644	edit dl rm
conv_dnnlowp_op_test.py	17919	0644	edit dl rm
conv_groupwise_dnnlowp_acc16_op_test.py	11664	0644	edit dl rm
conv_groupwise_dnnlowp_op_test.py	9537	0644	edit dl rm
dequantize_dnnlowp_op_test.py	1743	0644	edit dl rm
dnnlowp_test_utils.py	14544	0644	edit dl rm
elementwise_add_dnnlowp_op_test.py	6662	0644	edit dl rm
elementwise_linear_dnnlowp_op_test.py	3101	0644	edit dl rm
elementwise_mul_dnnlowp_op_test.py	6315	0644	edit dl rm
elementwise_sum_dnnlowp_op_test.py	9682	0644	edit dl rm
fully_connected_dnnlowp_acc16_op_test.py	8324	0644	edit dl rm
fully_connected_dnnlowp_op_test.py	10463	0644	edit dl rm
fully_connected_fp16_test.py	2519	0644	edit dl rm
fully_connected_rowwise_dnnlowp_op_test.py	5434	0644	edit dl rm
gather_dnnlowp_op_test.py	2790	0644	edit dl rm
group_norm_dnnlowp_op_test.py	4566	0644	edit dl rm
int8_gen_quant_params_min_max_test.py	3491	0644	edit dl rm
int8_gen_quant_params_test.py	3550	0644	edit dl rm
int8_quant_scheme_blob_fill_test.py	1838	0644	edit dl rm
lstm_unit_dnnlowp_op_test.py	3975	0644	edit dl rm
observer_test.py	1011	0644	edit dl rm
pool_dnnlowp_op_test.py	6125	0644	edit dl rm
quantize_dnnlowp_op_test.py	2668	0644	edit dl rm
relu_dnnlowp_op_test.py	2418	0644	edit dl rm
resize_nearest_3d_dnnlowp_op_test.py	2289	0644	edit dl rm
resize_nearest_dnnlowp_op_test.py	1994	0644	edit dl rm
sigmoid_dnnlowp_op_test.py	2203	0644	edit dl rm
spatial_batch_norm_dnnlowp_op_test.py	4094	0644	edit dl rm
tanh_dnnlowp_op_test.py	2187	0644	edit dl rm
utils.py	16059	0644	edit dl rm
__init__.py	0	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/caffe2/quantization/server/quantize_dnnlowp_op_test.py (2668B)

dyndep.InitOpsLibrary("//caffe2/caffe2/quantization/server:dnnlowp_ops")
workspace.GlobalInit(["caffe2", "--caffe2_omp_num_threads=11"])

class DNNLowPQuantizeOpTest(hu.HypothesisTestCase):
    @given(size=st.integers(1024, 2048),
           is_empty=st.booleans(),
           absorb=st.booleans(),
           **hu.gcs_cpu_only)
    @settings(max_examples=10, deadline=None)
    def test_dnnlowp_quantize(self, size, is_empty, absorb, gc, dc):
        if is_empty:
            size = 0
        min_ = -10.0
        max_ = 20.0
        X = (np.random.rand(size) * (max_ - min_) + min_).astype(np.float32)
        X_min = 0 if X.size == 0 else X.min()
        X_max = 1 if X.size == 0 else X.max()
        X_scale = (max(X_max, 0) - min(X_min, 0)) / 255
        X_zero = np.round(-X_min / X_scale)

op_type_list = ["Quantize", "Int8Quantize"]
        engine = "DNNLOWP"

for op_type in op_type_list:
            net = core.Net("test_net")

quantize = core.CreateOperator(
                op_type, ["X"], ["X_q"], engine=engine, device_option=gc
            )
            net.Proto().op.extend([quantize])

dnnlowp_pybind11.CreateInt8QuantParamsBlob(
                "quant_param", float(X_scale), int(X_zero)
            )
            quantize_2 = core.CreateOperator(
                op_type,
                ["X", "quant_param"],
                ["X_q_2"],
                engine=engine,
                device_option=gc,
            )
            net.Proto().op.extend([quantize_2])
            if absorb:
                net_str = dnnlowp_pybind11.freeze_quantization_params(
                    net.Proto().SerializeToString())
                net.Proto().ParseFromString(net_str)
            workspace.FeedBlob("X", X, device_option=gc)
            workspace.RunNetOnce(net)
            X_q = workspace.FetchInt8Blob("X_q")[0]
            X_q_2 = workspace.FetchInt8Blob("X_q_2")[0]

# Dequantize results and measure quantization error against inputs
            X_dq = X_scale * (X_q - X_zero)
            X_dq_2 = X_scale * (X_q_2 - X_zero)

# should be divided by 2 in an exact math, but divide by 1.9 here
            # considering finite precision in floating-point numbers
            atol = X_scale / 1.9
            np.testing.assert_allclose(X_dq, X, atol=atol, rtol=0)
            np.testing.assert_allclose(X_dq_2, X, atol=atol, rtol=0)

import caffe2.python.hypothesis_test_util as hu import hypothesis.strategies as st import numpy as np from caffe2.python import core, dyndep, workspace from caffe2.quantization.server import dnnlowp_pybind11 from hypothesis import given, settings dyndep.InitOpsLibrary("//caffe2/caffe2/quantization/server:dnnlowp_ops") workspace.GlobalInit(["caffe2", "--caffe2_omp_num_threads=11"]) class DNNLowPQuantizeOpTest(hu.HypothesisTestCase): @given(size=st.integers(1024, 2048), is_empty=st.booleans(), absorb=st.booleans(), **hu.gcs_cpu_only) @settings(max_examples=10, deadline=None) def test_dnnlowp_quantize(self, size, is_empty, absorb, gc, dc): if is_empty: size = 0 min_ = -10.0 max_ = 20.0 X = (np.random.rand(size) * (max_ - min_) + min_).astype(np.float32) X_min = 0 if X.size == 0 else X.min() X_max = 1 if X.size == 0 else X.max() X_scale = (max(X_max, 0) - min(X_min, 0)) / 255 X_zero = np.round(-X_min / X_scale) op_type_list = ["Quantize", "Int8Quantize"] engine = "DNNLOWP" for op_type in op_type_list: net = core.Net("test_net") quantize = core.CreateOperator( op_type, ["X"], ["X_q"], engine=engine, device_option=gc ) net.Proto().op.extend([quantize]) dnnlowp_pybind11.CreateInt8QuantParamsBlob( "quant_param", float(X_scale), int(X_zero) ) quantize_2 = core.CreateOperator( op_type, ["X", "quant_param"], ["X_q_2"], engine=engine, device_option=gc, ) net.Proto().op.extend([quantize_2]) if absorb: net_str = dnnlowp_pybind11.freeze_quantization_params( net.Proto().SerializeToString()) net.Proto().ParseFromString(net_str) workspace.FeedBlob("X", X, device_option=gc) workspace.RunNetOnce(net) X_q = workspace.FetchInt8Blob("X_q")[0] X_q_2 = workspace.FetchInt8Blob("X_q_2")[0] # Dequantize results and measure quantization error against inputs X_dq = X_scale * (X_q - X_zero) X_dq_2 = X_scale * (X_q_2 - X_zero) # should be divided by 2 in an exact math, but divide by 1.9 here # considering finite precision in floating-point numbers atol = X_scale / 1.9 np.testing.assert_allclose(X_dq, X, atol=atol, rtol=0) np.testing.assert_allclose(X_dq_2, X, atol=atol, rtol=0)