1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
|
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, settings
dyndep.InitOpsLibrary("//caffe2/caffe2/quantization/server:dnnlowp_ops")
workspace.GlobalInit(["caffe2", "--caffe2_omp_num_threads=11"])
class DNNLowPLSTMUnitOpTest(hu.HypothesisTestCase):
@given(
N=st.integers(0, 64),
D=st.integers(4, 64),
forget_bias=st.integers(0, 4),
**hu.gcs_cpu_only
)
@settings(max_examples=10, deadline=None)
def test_dnnlowp_lstm_unit(self, N, D, forget_bias, gc, dc):
# X has scale 1, so exactly represented after quantization
H_in = np.clip(np.random.randn(1, N, D), -1, 1).astype(np.float32)
C_in = np.clip(np.random.randn(1, N, D), -1, 1).astype(np.float32)
G = np.clip(np.random.randn(1, N, 4 * D), -1, 1).astype(np.float32)
seq_lengths = np.round(np.random.rand(N)).astype(np.int32)
# seq_lengths.fill(0)
t = np.array([5]).astype(np.int32)
Output = collections.namedtuple("Output", ["H_out", "C_out", "engine"])
outputs = []
engine_list = ["", "DNNLOWP"]
for engine in engine_list:
net = core.Net("test_net")
if engine == "DNNLOWP":
quantize_H_in = core.CreateOperator(
"Quantize", ["H_in"], ["H_in_q"], engine=engine, device_option=gc
)
quantize_C_in = core.CreateOperator(
"Quantize", ["C_in"], ["C_in_q"], engine=engine, device_option=gc
)
quantize_G = core.CreateOperator(
"Quantize", ["G"], ["G_q"], engine=engine, device_option=gc
)
net.Proto().op.extend([quantize_H_in, quantize_C_in, quantize_G])
lstm = core.CreateOperator(
"LSTMUnit",
[
"H_in_q" if engine == "DNNLOWP" else "H_in",
"C_in_q" if engine == "DNNLOWP" else "C_in",
"G_q" if engine == "DNNLOWP" else "G",
"seq_lengths",
"t",
],
[
"H_out_q" if engine == "DNNLOWP" else "H_out",
"C_out_q" if engine == "DNNLOWP" else "C_out",
],
engine=engine,
device_option=gc,
axis=0,
)
net.Proto().op.extend([lstm])
if engine == "DNNLOWP":
dequantize_H_out = core.CreateOperator(
"Dequantize",
["H_out_q"],
["H_out"],
engine=engine,
device_option=gc,
)
dequantize_C_out = core.CreateOperator(
"Dequantize",
["C_out_q"],
["C_out"],
engine=engine,
device_option=gc,
)
net.Proto().op.extend([dequantize_H_out, dequantize_C_out])
self.ws.create_blob("H_in").feed(H_in, device_option=gc)
self.ws.create_blob("C_in").feed(C_in, device_option=gc)
self.ws.create_blob("G").feed(G, device_option=gc)
self.ws.create_blob("seq_lengths").feed(seq_lengths, device_option=gc)
self.ws.create_blob("t").feed(t, device_option=gc)
self.ws.run(net)
outputs.append(
Output(
H_out=self.ws.blobs["H_out"].fetch(),
C_out=self.ws.blobs["C_out"].fetch(),
engine=engine,
)
)
for o in outputs:
np.testing.assert_allclose(o.C_out, outputs[0].C_out, atol=0.1, rtol=0.2)
np.testing.assert_allclose(o.H_out, outputs[0].H_out, atol=0.1, rtol=0.2)
|
