-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtest_model.py
165 lines (135 loc) · 5.56 KB
/
test_model.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
from model import (
TokenDropout,
filip_similarity_score,
contrastive_loss,
)
import torch
import torch.nn.functional as F
from einops import rearrange
from itertools import product
def test_decoupled_contrastive_loss():
""" I can't think of a good way to test the DCL portion TBH """
logits = torch.randn((64, 64))
assert torch.allclose(
F.cross_entropy(logits, torch.arange(logits.shape[0]).to(logits.device)),
contrastive_loss(logits, use_dcl=False).mean(),
)
assert torch.allclose(
F.cross_entropy(logits.T, torch.arange(logits.shape[0]).to(logits.device)),
contrastive_loss(logits.T, use_dcl=False).mean(),
)
# assert torch.isclose(default_clip_loss(hA, hB), decoupled_contrastive_loss(hA, hB))
def test_token_dropout():
B, T, D = 16, 128, 512
x = torch.randn(B, T, D)
token_dropout, _ = TokenDropout(0.5)
x = token_dropout(x)
assert x.shape == (B, T // 2, D)
def test_similarity_masking():
torch.manual_seed(0)
m, n = 3, 1
bA, bB = 32, 64
Ta, Tb = 128, 256
maskA = torch.rand(m, bA, Ta) > 0.5
maskB = torch.rand(n, bB, Tb) > 0.5
maskA = rearrange(maskA, "m bA tA -> m 1 bA 1 tA 1")
maskB = rearrange(maskB, "n bB tB -> 1 n 1 bB 1 tB")
combined = maskA * maskB
for group in range(m):
for batch in range(bA):
for token in range(Ta):
if not maskA[group, 0, batch, 0, token, 0].item():
assert torch.all(~combined[group, :, batch, :, token, :]).item()
def test_similarity_score_single_batch():
"""Compare similarity scoring from filip to loopy version."""
torch.manual_seed(0)
Ta, Tb = 128, 256 # sequence steps
d = 50 # embedding dimension
hA = torch.randn(Ta, d)
hB = torch.randn(Tb, d)
maskA = torch.rand(Ta) > 0.2
maskB = torch.rand(Tb) > 0.2
temperature = torch.randn(1).item()
sim_scores_A_to_B = torch.empty(Ta)
sim_scores_B_to_A = torch.empty(Tb)
for k in range(Ta):
max_sim_score = -float("inf")
sim_score = (hA[k] @ hB.T) / temperature
for tb in range(Tb):
if (sim_score[tb] > max_sim_score).item() and maskB[tb]:
max_sim_score = sim_score[tb]
sim_scores_A_to_B[k] = max_sim_score
for k in range(Tb):
max_sim_score = -float("inf")
sim_score = hB[k] @ hA.T / temperature
for ta in range(Ta):
if (sim_score[ta] > max_sim_score).item() and maskA[ta]:
max_sim_score = sim_score[ta]
sim_scores_B_to_A[k] = max_sim_score
sim_score_A = torch.sum(sim_scores_A_to_B * maskA) / torch.sum(maskA)
sim_score_B = torch.sum(sim_scores_B_to_A * maskB) / torch.sum(maskB)
ssA_main, ssB_main = filip_similarity_score(
hA[None, None],
hB[None, None],
maskA[None, None],
maskB[None, None],
temperature,
)
assert torch.isclose(ssA_main, sim_score_A.squeeze())
assert torch.isclose(ssB_main, sim_score_B.squeeze())
def test_similarity_score_multi_batch():
"""Compare similarity scoring from filip to loopy version.
It may actually make sense to collapse the group and the batch into a single dimension,
but I don't think it makes a huge efficiency difference and I want to avoid bugs from
that mixing.
"""
torch.manual_seed(0)
m, n = 2, 3 # number of groups
bA, bB = 4, 8 # batch size
Ta, Tb = 32, 32 # sequence steps
d = 16 # embedding dimension
hA = torch.randn(m, bA, Ta, d)
hB = torch.randn(n, bB, Tb, d)
maskA = torch.rand(m, bA, Ta) > 0.2
maskB = torch.rand(n, bB, Tb) > 0.2
temperature = torch.randn(1).item()
sim_scores_A_to_B = torch.empty(m, n, bA, bB, Ta)
sim_scores_B_to_A = torch.empty(m, n, bA, bB, Tb)
for groupA, groupB in product(range(m), range(n)):
for batchA, batchB in product(range(bA), range(bB)):
for k in range(Ta):
max_sim_score = -float("inf")
sim_score = (hA[groupA, batchA, k] @ hB[groupB, batchB].T) / temperature
for tb in range(Tb):
if (sim_score[tb] > max_sim_score).item() and maskB[
groupB, batchB, tb
]:
max_sim_score = sim_score[tb]
sim_scores_A_to_B[groupA, groupB, batchA, batchB, k] = max_sim_score
for groupA, groupB in product(range(m), range(n)):
for batchA, batchB in product(range(bA), range(bB)):
for k in range(Tb):
max_sim_score = -float("inf")
sim_score = hB[groupB, batchB, k] @ hA[groupA, batchA].T / temperature
for ta in range(Ta):
if (sim_score[ta] > max_sim_score).item() and maskA[
groupA, batchA, ta
]:
max_sim_score = sim_score[ta]
sim_scores_B_to_A[groupA, groupB, batchA, batchB, k] = max_sim_score
sim_score_A = (
torch.sum(sim_scores_A_to_B * maskA[:, None, :, None, :], dim=-1)
/ torch.sum(maskA, dim=-1)[:, None, :, None]
)
sim_score_B = (
torch.sum(sim_scores_B_to_A * maskB[None, :, None, :, :], dim=-1)
/ torch.sum(maskB, dim=-1)[None, :, None, :]
)
ssA_main, ssB_main = filip_similarity_score(hA, hB, maskA, maskB, temperature)
assert torch.allclose(sim_score_A, ssA_main)
assert torch.allclose(sim_score_B, ssB_main)
if __name__ == "__main__":
test_decoupled_contrastive_loss()
# test_token_dropout()
# test_similarity_score_single_batch()
#test_similarity_score_multi_batch()