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/fully_connected_fp16_test.py (2519B)

import collections

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 hypothesis import given

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

def mse(x, xh):
    d = (x - xh).reshape(-1)
    return 0 if len(d) == 0 else np.sqrt(np.matmul(d, d.transpose())) / len(d)

class FullyConnectedFP16Test(hu.HypothesisTestCase):
    @given(
        input_channels=st.integers(128, 256),
        output_channels=st.integers(128, 256),
        batch_size=st.integers(128, 256),
        empty_batch=st.booleans(),
        **hu.gcs_cpu_only
    )
    def test_fully_connected(self, input_channels, output_channels, batch_size, empty_batch, gc, dc):
        if empty_batch:
            batch_size = 0
        W = np.random.randn(output_channels, input_channels).astype(np.float32)
        X = np.random.randn(batch_size, input_channels).astype(np.float32)
        b = np.random.randn(output_channels).astype(np.float32)

Output = collections.namedtuple("Output", ["Y", "engine", "order"])

order = "NHWC"
        net = core.Net("test_net")
        engine = "FAKE_FP16"

fc = core.CreateOperator(
            "FC", ["X", "W", "b"], ["Y"], order=order, engine=engine, device_option=gc
        )
        net.Proto().op.extend([fc])

self.ws.create_blob("X").feed(X, device_option=gc)
        self.ws.create_blob("W").feed(W, device_option=gc)
        self.ws.create_blob("b").feed(b, device_option=gc)
        self.ws.run(net)
        output = Output(Y=self.ws.blobs["Y"].fetch(), engine=engine, order=order)

# Mimic the quantization in python
        Wh = W.astype(np.float16)
        Xh = X.astype(np.float16)
        bh = b.astype(np.float16)

bbh = np.outer(np.ones(batch_size, dtype=np.float16), bh)
        assert bbh.dtype == np.float16
        Yrefh = np.matmul(Xh, Wh.transpose()) + bbh
        assert Yrefh.dtype == np.float16

bb = np.outer(np.ones(batch_size, dtype=np.float32), b)
        Yref = np.matmul(X, W.transpose()) + bb
        assert Yref.dtype == np.float32

# The error between plain->quantized, and plain->python_quantized
        # should be very close
        mse_c2 = mse(Yref, output.Y)
        mse_py = mse(Yref, Yrefh)
        print(np.abs(mse_c2 - mse_py))
        assert np.isclose(mse_c2, mse_py, atol=1e-3), np.abs(mse_c2 - mse_py)

import collections 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 hypothesis import given dyndep.InitOpsLibrary("//caffe2/caffe2/quantization/server:dnnlowp_ops") workspace.GlobalInit(["caffe2", "--caffe2_omp_num_threads=11"]) def mse(x, xh): d = (x - xh).reshape(-1) return 0 if len(d) == 0 else np.sqrt(np.matmul(d, d.transpose())) / len(d) class FullyConnectedFP16Test(hu.HypothesisTestCase): @given( input_channels=st.integers(128, 256), output_channels=st.integers(128, 256), batch_size=st.integers(128, 256), empty_batch=st.booleans(), **hu.gcs_cpu_only ) def test_fully_connected(self, input_channels, output_channels, batch_size, empty_batch, gc, dc): if empty_batch: batch_size = 0 W = np.random.randn(output_channels, input_channels).astype(np.float32) X = np.random.randn(batch_size, input_channels).astype(np.float32) b = np.random.randn(output_channels).astype(np.float32) Output = collections.namedtuple("Output", ["Y", "engine", "order"]) order = "NHWC" net = core.Net("test_net") engine = "FAKE_FP16" fc = core.CreateOperator( "FC", ["X", "W", "b"], ["Y"], order=order, engine=engine, device_option=gc ) net.Proto().op.extend([fc]) self.ws.create_blob("X").feed(X, device_option=gc) self.ws.create_blob("W").feed(W, device_option=gc) self.ws.create_blob("b").feed(b, device_option=gc) self.ws.run(net) output = Output(Y=self.ws.blobs["Y"].fetch(), engine=engine, order=order) # Mimic the quantization in python Wh = W.astype(np.float16) Xh = X.astype(np.float16) bh = b.astype(np.float16) bbh = np.outer(np.ones(batch_size, dtype=np.float16), bh) assert bbh.dtype == np.float16 Yrefh = np.matmul(Xh, Wh.transpose()) + bbh assert Yrefh.dtype == np.float16 bb = np.outer(np.ones(batch_size, dtype=np.float32), b) Yref = np.matmul(X, W.transpose()) + bb assert Yref.dtype == np.float32 # The error between plain->quantized, and plain->python_quantized # should be very close mse_c2 = mse(Yref, output.Y) mse_py = mse(Yref, Yrefh) print(np.abs(mse_c2 - mse_py)) assert np.isclose(mse_c2, mse_py, atol=1e-3), np.abs(mse_c2 - mse_py)