-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathrun_perf_tests.py
247 lines (216 loc) · 9.77 KB
/
run_perf_tests.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
238
239
240
241
242
243
244
245
246
247
import numpy as np
import time
import torch
import networkx as nx
from hot_pytorch.batch.sparse import make_batch
from hot_pytorch.batch.dense import Batch as D
import hot_pytorch.models
@torch.no_grad()
def get_batched_data(n, bsize, dim, sparse, seed, device='cuda'):
tic = time.time()
adj_list = []
for _ in range(bsize):
graph = nx.barabasi_albert_graph(n, 5, seed)
adj = nx.adjacency_matrix(graph).tocoo()
adj_list.append(adj)
print(f'Graph init done in \t{time.time() - tic:.3f}sec')
tic = time.time()
# initialize features
assert dim % 2 == 0
init_list = []
for adj in adj_list:
edge_indices = torch.tensor(np.vstack((adj.row, adj.col)), dtype=torch.long, device=device) # [2, |E|]
e = edge_indices.size(1)
node_feat = torch.randn(n, dim // 2, device=device) # [N, D/2]
edge_feat = torch.randn(e, dim // 2, device=device) # [|E|, D/2]
init_list.append((edge_indices, node_feat, edge_feat))
if sparse:
# get sparse batch
edge_indices, node_features, edge_features = zip(*init_list)
batch = make_batch(node_features, edge_indices, edge_features)
else:
# get dense batch
A_list = []
for edge_indices, node_feat, edge_feat in init_list:
edge_feat = torch.sparse_coo_tensor(edge_indices, edge_feat, size=(n, n, dim // 2)).to_dense() # [N, N, D/2]
node_feat = node_feat[None, ...] * torch.eye(n, device=device)[..., None] # [N, N, D/2]
A = torch.cat([node_feat, edge_feat], dim=-1) # [N, N, D]
A_list.append(A)
# setup batch
A = torch.stack(A_list, dim=0) # [B, N, N, D]
n_nodes = [n] * bsize
batch = D(A, n_nodes)
print(f'Batch init done in \t{time.time() - tic:.3f}sec')
return batch
def get_peak_mem_and_reset():
stats = torch.cuda.memory_stats()
peak_bytes_requirement = stats["allocated_bytes.all.peak"]
torch.cuda.reset_peak_memory_stats()
torch.cuda.empty_cache()
return peak_bytes_requirement / 1024 ** 3 # unit: GB
def measure(model, G):
torch.cuda.reset_peak_memory_stats()
torch.cuda.empty_cache()
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
start.record()
out = model(G) # [B, D]
end.record()
# Waits for everything to finish running
torch.cuda.synchronize()
forward_t = start.elapsed_time(end) / 1000 # unit: milliseconds
out = out.sum()
start.record()
out.backward()
end.record()
# Waits for everything to finish running
torch.cuda.synchronize()
backward_t = start.elapsed_time(end) / 1000 # unit: milliseconds
peak_mem = get_peak_mem_and_reset()
return forward_t, backward_t, peak_mem
def main_routine(repeat, n, bsize, n_layers, dim, dim_qk, dim_v, n_heads, dim_ff, readout_dim_qk, readout_dim_v, readout_n_heads):
print(f'\n\nn = {n}')
result = {}
print("DL")
try:
forward_t, backward_t, peak_mem = [], [], []
model = hot_pytorch.models.MLP(2, 0, [2] * n_layers, dim, dim, dim, sparse=False).to('cuda')
for i in range(repeat):
G = get_batched_data(n, bsize, dim, sparse=False, seed=i)
ft, bt, pm = measure(model, G)
forward_t.append(ft)
backward_t.append(bt)
peak_mem.append(pm)
result['dl_forward_t'] = (np.mean(np.array(forward_t)), np.std(np.array(forward_t)))
result['dl_backward_t'] = (np.mean(np.array(backward_t)), np.std(np.array(backward_t)))
result['dl_peak_mem'] = (np.mean(np.array(peak_mem)), np.std(np.array(peak_mem)))
except RuntimeError as e:
result['dl_forward_t'] = 'OOM'
result['dl_backward_t'] = 'OOM'
result['dl_peak_mem'] = 'OOM'
print(e)
print("DA")
try:
forward_t, backward_t, peak_mem = [], [], []
model = hot_pytorch.models.Encoder(2, 0, [2] * n_layers, dim, dim, dim, dim_qk, dim_v, dim_ff, n_heads,
readout_dim_qk, readout_dim_v, readout_n_heads, 'default', 'default',
drop_input=0., dropout=0., drop_mu=0., sparse=False).to('cuda')
for i in range(repeat):
G = get_batched_data(n, bsize, dim, sparse=False, seed=i)
ft, bt, pm = measure(model, G)
forward_t.append(ft)
backward_t.append(bt)
peak_mem.append(pm)
result['da_forward_t'] = (np.mean(np.array(forward_t)), np.std(np.array(forward_t)))
result['da_backward_t'] = (np.mean(np.array(backward_t)), np.std(np.array(backward_t)))
result['da_peak_mem'] = (np.mean(np.array(peak_mem)), np.std(np.array(peak_mem)))
except RuntimeError as e:
result['da_forward_t'] = 'OOM'
result['da_backward_t'] = 'OOM'
result['da_peak_mem'] = 'OOM'
print("DK")
try:
forward_t, backward_t, peak_mem = [], [], []
model = hot_pytorch.models.Encoder(2, 0, [2] * n_layers, dim, dim, dim, dim_qk, dim_v, dim_ff, n_heads,
readout_dim_qk, readout_dim_v, readout_n_heads, 'default', 'generalized_kernel',
drop_input=0., dropout=0., drop_mu=0., sparse=False).to('cuda')
model.skip_redraw_projections = True
for i in range(repeat):
G = get_batched_data(n, bsize, dim, sparse=False, seed=i)
ft, bt, pm = measure(model, G)
forward_t.append(ft)
backward_t.append(bt)
peak_mem.append(pm)
result['dk_forward_t'] = (np.mean(np.array(forward_t)), np.std(np.array(forward_t)))
result['dk_backward_t'] = (np.mean(np.array(backward_t)), np.std(np.array(backward_t)))
result['dk_peak_mem'] = (np.mean(np.array(peak_mem)), np.std(np.array(peak_mem)))
except RuntimeError as e:
result['dk_forward_t'] = 'OOM'
result['dk_backward_t'] = 'OOM'
result['dk_peak_mem'] = 'OOM'
print(e)
print("SL")
try:
forward_t, backward_t, peak_mem = [], [], []
model = hot_pytorch.models.MLP(2, 0, [2] * n_layers, dim, dim, dim, sparse=True).to('cuda')
for i in range(repeat):
G = get_batched_data(n, bsize, dim, sparse=True, seed=i)
ft, bt, pm = measure(model, G)
forward_t.append(ft)
backward_t.append(bt)
peak_mem.append(pm)
result['sl_forward_t'] = (np.mean(np.array(forward_t)), np.std(np.array(forward_t)))
result['sl_backward_t'] = (np.mean(np.array(backward_t)), np.std(np.array(backward_t)))
result['sl_peak_mem'] = (np.mean(np.array(peak_mem)), np.std(np.array(peak_mem)))
except RuntimeError as e:
result['sl_forward_t'] = 'OOM'
result['sl_backward_t'] = 'OOM'
result['sl_peak_mem'] = 'OOM'
print(e)
print("SA")
try:
forward_t, backward_t, peak_mem = [], [], []
model = hot_pytorch.models.Encoder(2, 0, [2] * n_layers, dim, dim, dim, dim_qk, dim_v, dim_ff, n_heads,
readout_dim_qk, readout_dim_v, readout_n_heads, 'default', 'default',
drop_input=0., dropout=0., drop_mu=0., sparse=True).to('cuda')
for i in range(repeat):
G = get_batched_data(n, bsize, dim, sparse=True, seed=i)
ft, bt, pm = measure(model, G)
forward_t.append(ft)
backward_t.append(bt)
peak_mem.append(pm)
result['sa_forward_t'] = (np.mean(np.array(forward_t)), np.std(np.array(forward_t)))
result['sa_backward_t'] = (np.mean(np.array(backward_t)), np.std(np.array(backward_t)))
result['sa_peak_mem'] = (np.mean(np.array(peak_mem)), np.std(np.array(peak_mem)))
except RuntimeError as e:
result['sa_forward_t'] = 'OOM'
result['sa_backward_t'] = 'OOM'
result['sa_peak_mem'] = 'OOM'
print(e)
print("SK")
try:
forward_t, backward_t, peak_mem = [], [], []
model = hot_pytorch.models.Encoder(2, 0, [2] * n_layers, dim, dim, dim, dim_qk, dim_v, dim_ff, n_heads,
readout_dim_qk, readout_dim_v, readout_n_heads, 'default', 'generalized_kernel',
drop_input=0., dropout=0., drop_mu=0., sparse=True).to('cuda')
model.skip_redraw_projections = True
for i in range(repeat):
G = get_batched_data(n, bsize, dim, sparse=True, seed=i)
ft, bt, pm = measure(model, G)
forward_t.append(ft)
backward_t.append(bt)
peak_mem.append(pm)
result['sk_forward_t'] = (np.mean(np.array(forward_t)), np.std(np.array(forward_t)))
result['sk_backward_t'] = (np.mean(np.array(backward_t)), np.std(np.array(backward_t)))
result['sk_peak_mem'] = (np.mean(np.array(peak_mem)), np.std(np.array(peak_mem)))
except RuntimeError as e:
result['sk_forward_t'] = 'OOM'
result['sk_backward_t'] = 'OOM'
result['sk_peak_mem'] = 'OOM'
print(e)
return result
def main():
repeat = 10
bsize = 1
n_layers = 4
dim = 32
dim_qk = 32
dim_v = 32
n_heads = 4
dim_ff = 32
readout_dim_qk = 32
readout_dim_v = 32
readout_n_heads = 4
result = {}
n_list = list((2 ** np.linspace(5, 18, 27, endpoint=True)).astype(int) // 5) # for log-scale plot
for n in n_list:
start = time.time()
n_result = main_routine(repeat, n, bsize, n_layers, dim, dim_qk, dim_v, n_heads, dim_ff, readout_dim_qk, readout_dim_v, readout_n_heads)
print(f"{n}: done after {(time.time() - start):.2f} sec")
print(f"result_{n} = {n_result}")
result[n] = n_result
if n_result['sk_forward_t'] == 'OOM':
break
print(result)
if __name__ == '__main__':
main()