-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_hash_embedding.py
237 lines (176 loc) · 8.15 KB
/
test_hash_embedding.py
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
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
import logging
import math
import time
import torch
from pytorch_ps.nn.hash_embedding import HashEmbeddingLayer
from pytorch_ps.nn.sparse_adam import HashSparseAdam
from pytorch_ps.ps.client import PSClient
def test_one():
inputs1 = torch.tensor([2, 4, 5], dtype=torch.long, requires_grad=False)
inputs2 = torch.tensor([10, 11, 12], dtype=torch.long, requires_grad=False)
labels = torch.tensor([0.20, 0.44, 0.60], dtype=torch.float32, requires_grad=False)
ps_client = PSClient()
emb_layer1 = HashEmbeddingLayer(ps_client,"sparse_test1", 10_000_000_000, 4,[0.1,0.2,0.3,0.4])
emb_layer2 = HashEmbeddingLayer(ps_client,"sparse_test2", 10_000_000_000, 8,[0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8])
params_name = []
params_v = []
for n, v in emb_layer1.named_parameters():
params_name.append(n)
params_v.append(v)
for n, v in emb_layer2.named_parameters():
params_name.append(n)
params_v.append(v)
output1: torch.Tensor = emb_layer1(inputs1)
output2: torch.Tensor = emb_layer2(inputs2)
hidden_inputs = torch.concat([output1, output2], dim=1)
print(f"hidden_inputs:%s" % (hidden_inputs))
loss = torch.mean(hidden_inputs)
loss.backward()
def test_lr():
inputs1 = torch.tensor([2, 4, 5], dtype=torch.long, requires_grad=False)
inputs2 = torch.tensor([10, 11, 12], dtype=torch.long, requires_grad=False)
labels = torch.tensor([0.20, 0.44, 0.60], dtype=torch.float32, requires_grad=False)
ps_client = PSClient()
emb_layer1 = HashEmbeddingLayer(ps_client, "sparse_test1", 10_000_000_000, 4, [0.1, 0.2, 0.3, 0.4])
emb_layer2 = HashEmbeddingLayer(ps_client, "sparse_test2", 10_000_000_000, 8,
[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8])
params_name = []
params_v = []
mlp = torch.nn.Sequential(
torch.nn.Linear(12, 4),
torch.nn.ReLU(),
torch.nn.Linear(4, 1)
)
for n, v in emb_layer1.named_parameters():
params_name.append(n)
params_v.append(v)
for n, v in emb_layer2.named_parameters():
params_name.append(n)
params_v.append(v)
sparse_optim = HashSparseAdam(ps_client,params_name, params_v, lr=0.002)
dense_optim = torch.optim.Adam(mlp.parameters(), lr=0.002)
lossfn = torch.nn.L1Loss()
for epoch in range(1000):
output1: torch.Tensor = emb_layer1(inputs1)
output2: torch.Tensor = emb_layer2(inputs2)
hidden_inputs = torch.concat([output1, output2], dim=1)
#print(f"hidden_inputs:%s" % (hidden_inputs))
pred_y = mlp(hidden_inputs)
print("pred_y", pred_y)
loss = lossfn(pred_y.squeeze(), labels)
loss.backward()
sparse_optim.step()
dense_optim.zero_grad()
sparse_optim.zero_grad()
dense_optim.zero_grad()
print(f"************%dloss:%s" % (epoch, loss))
if loss < 0.05:
break
time.sleep(1)
def _compare_grad1(init_data:list,init_data2:list,init_data3:list):
ps_client = PSClient()
hash_emb_layer = HashEmbeddingLayer(ps_client,"test1",10,4,[0.0,0.0,0.0,0.0])
params_name = []
params_v = []
for n, v in hash_emb_layer.named_parameters():
params_name.append(n)
params_v.append(v)
sparse_optim = HashSparseAdam(ps_client, params_name, params_v, lr=0.002)
mlp1 = torch.nn.Sequential(
TestLinear(len(init_data[0]), len(init_data), _weight=init_data),
torch.nn.ReLU(),
TestLinear(len(init_data2[0]), len(init_data2), _weight=init_data2),
torch.nn.Sigmoid()
)
dense_optim1 = torch.optim.Adam(mlp1.parameters(), lr=0.002)
lossfn1 = torch.nn.BCEWithLogitsLoss()
inputs1 = torch.tensor([2, 4, 5], dtype=torch.long, requires_grad=False)
labels1 = torch.tensor([0.5, 1, 0.9], dtype=torch.float32, requires_grad=False)
inputs1_other = torch.tensor(init_data3,dtype=torch.float32)
for epoch in range(10):
hidden_output1 = hash_emb_layer(inputs1)
pred_y1 = mlp1(torch.concat([hidden_output1,inputs1_other],dim=-1))
loss1 = lossfn1(pred_y1.squeeze(), labels1)
loss1.backward()
sparse_optim.step()
dense_optim1.step()
sparse_optim.zero_grad()
dense_optim1.zero_grad()
yield hash_emb_layer.get_hash_parameter()
def _compare_grad2(init_data:list,init_data2:list,init_data3:list):
torch_emb_layer = torch.nn.Embedding(10, 4, _weight=torch.zeros((10, 4), dtype=torch.float32))
emb_optim2 = torch.optim.Adam(torch_emb_layer.parameters(), lr=0.002)
mlp2 = torch.nn.Sequential(
TestLinear(len(init_data[0]), len(init_data), _weight=init_data),
torch.nn.ReLU(),
TestLinear(len(init_data2[0]), len(init_data2), _weight=init_data2),
torch.nn.Sigmoid()
)
dense_optim2 = torch.optim.Adam(mlp2.parameters(), lr=0.002)
lossfn2 = torch.nn.BCEWithLogitsLoss()
inputs2 = torch.tensor([2, 4, 5], dtype=torch.long, requires_grad=False)
labels2 = torch.tensor([0.5, 1, 0.9], dtype=torch.float32, requires_grad=False)
inputs2_other = torch.tensor(init_data3, dtype=torch.float32)
for epoch in range(10):
hidden_output2 = torch_emb_layer(inputs2)
pred_y2 = mlp2(torch.concat([hidden_output2,inputs2_other],dim=-1))
loss2 = lossfn2(pred_y2.squeeze(), labels2)
loss2.backward()
emb_optim2.step()
dense_optim2.step()
emb_optim2.zero_grad()
dense_optim2.zero_grad()
yield torch_emb_layer.parameters().__next__()
def run_compare_grad():
input_dim = 16
d1 = torch.randn((input_dim,32),dtype=torch.float32).t()
d2 = torch.randn((32, 1), dtype=torch.float32).t()
d3 = torch.randn((3, input_dim-4), dtype=torch.float32)
for v1,v2 in zip(_compare_grad1(d1.tolist(),d2.tolist(),d3.tolist()),_compare_grad2(d1.tolist(),d2.tolist(),d3.tolist())):
print("hash embedding:", torch.index_select(input=v1,dim=0,index=torch.tensor([2,4,5])))
print("torch embedding:", torch.index_select(input=v1,dim=0,index=torch.tensor([2,4,5])))
print(f"===================equal items:{torch.sum((v1.sub(v2)<0.00001).to(torch.int)).item()}====================")
class TestLinear(torch.nn.Module):
"""
copy from torch.nn.Linear
fix parameter for testing
"""
__constants__ = ['in_features', 'out_features']
in_features: int
out_features: int
weight: torch.Tensor
def __init__(self, in_features: int, out_features: int,_weight:list = None, bias: bool = True,
device=None, dtype=None) -> None:
factory_kwargs = {'device': device, 'dtype': dtype}
super().__init__()
self.in_features = in_features
self.out_features = out_features
if _weight is not None:
self.weight = torch.nn.Parameter(torch.tensor(_weight))
self.bias = torch.nn.Parameter(torch.zeros(out_features)+0.1)
else:
self.weight = torch.nn.Parameter(torch.empty((out_features, in_features), **factory_kwargs))
if bias:
self.bias = torch.nn.Parameter(torch.empty(out_features, **factory_kwargs))
else:
self.register_parameter('bias', None)
self.reset_parameters()
def reset_parameters(self) -> None:
# Setting a=sqrt(5) in kaiming_uniform is the same as initializing with
# uniform(-1/sqrt(in_features), 1/sqrt(in_features)). For details, see
# https://github.com/pytorch/pytorch/issues/57109
torch.nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))
if self.bias is not None:
fan_in, _ = torch.nn.init._calculate_fan_in_and_fan_out(self.weight)
bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
torch.nn.init.uniform_(self.bias, -bound, bound)
def forward(self, input: torch.Tensor) -> torch.Tensor:
return torch.nn.functional.linear(input, self.weight, self.bias)
def extra_repr(self) -> str:
return f'in_features={self.in_features}, out_features={self.out_features}, bias={self.bias is not None}'
if __name__=="__main__":
logging.getLogger().setLevel(logging.DEBUG)
logging.getLogger().setLevel(logging.ERROR)
test_one()
test_lr()
run_compare_grad()