-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
207 lines (157 loc) · 6.29 KB
/
train.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
import torch
import glob
from datetime import datetime
import random
import numpy as np
from torch import nn
import torch.nn.functional as F
import time
import os
import argparse
from tqdm import tqdm
## models
## copy-pasted from paper appendix A
# D x, D y: latent dimension of unimodal encoders
# D s: latent dimension of shared space
# depth x, depth y: number of blocks for each adapter
# expansion factor: expansion factor hyperparameter
# dropout: dropout hyperparameter
D_x = 768
D_y = 1536
D_s = 512
# block depth
depth_x, depth_y = 2, 2
class Block(nn.Module):
def __init__(self, dim, expansion_factor=4, dropout=0.2):
super().__init__()
self.fn = nn.Sequential(
nn.Linear(dim, int(expansion_factor * dim)),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(int(expansion_factor * dim), dim),
)
self.ln = nn.LayerNorm(dim)
# self.apply(self.init_weights)
def init_weights(self, m):
pass
def forward(self, x):
return x + self.fn(self.ln(x))
class FusionAdapter(nn.Module):
def __init__(self, dim, dim_out, depth, n_paths=1, expansion_factor=4, dropout=0.2):
super().__init__()
if depth > 0:
self.enc = nn.Sequential(
nn.Linear(dim, dim),
nn.GELU(),
# nn.Linear(dim, dim),
# nn.GELU(),
nn.Dropout(dropout),
)
self.fns = nn.ModuleList([
nn.Sequential(
*[Block(dim, expansion_factor=expansion_factor, dropout=dropout) for _ in range(depth)],
nn.GELU(),
nn.Linear(dim, dim_out),
)
for _ in range(n_paths)
])
self.tail = nn.Sequential(
nn.GELU(),
nn.Dropout(dropout),
nn.LayerNorm(dim_out),
nn.Linear(dim_out, dim_out)
)
# self.apply(self.init_weights)
def forward(self, x):
if depth < 1:
return x
enc = self.enc(x)
sum = torch.stack([fn(enc) for fn in self.fns], dim=0).sum(dim=0)
return self.tail(sum)
def train(hhX, hhY, STEPS=1e5, B=1000, savedir=None):
STEPS = int(STEPS)
B = int(B)
all_img_emb = torch.zeros((82600, 768), dtype=torch.float32, device="cuda")
all_txt_emb = torch.zeros((82600, 1536), dtype=torch.float32, device="cuda")
all_ids = []
emb_paths = glob.glob("coco_embeddings_3500/*.pt")
cum_ind = 0
for path_idx, path in enumerate(emb_paths):
batch, text_batch, ids = torch.load(path)
all_img_emb[cum_ind:cum_ind+batch.shape[0], :] = torch.tensor(batch).cuda()
all_txt_emb[cum_ind:cum_ind+batch.shape[0], :] = torch.tensor(text_batch).cuda()
all_ids.extend(ids)
cum_ind += batch.shape[0]
print("UP TO", cum_ind)
t = nn.Parameter(0.15 * torch.ones([], requires_grad=True).to("cuda"))
params = list(hhX.parameters()) + list(hhY.parameters())
optimizer = torch.optim.AdamW(params, lr=3e-4)
print_every = STEPS // 100
save_every = STEPS // 10
last_time = time.time()
losses = []
# symmetric alignment loss
labels = torch.arange(B, device="cuda", dtype=torch.long)
for i in range(STEPS):
optimizer.zero_grad()
temp_optim.zero_grad()
ind = torch.randint(0, 82600, (2*B,)).cuda()
# text_ind = torch.randint(0, 5, (2*B,))
z_x = all_img_emb[ind, :]
z_y = all_txt_emb[ind, :]
# print(z_x[0, 0], z_y[0, 0])
z_x1, z_x2 = torch.chunk(z_x, 2) # B x D x
z_y1, z_y2 = torch.chunk(z_y, 2) # B x D y
lam = random.random()
z_x = lam * z_x1 + (1 - lam) * z_x2
z_y = lam * z_y1 + (1 - lam) * z_y2
# joint space and normalize
s_x = F.normalize(hhX(z_x), dim=-1) # B x D s
s_y = F.normalize(hhY(z_y), dim=-1) # B x D s
# pairwise cosine similarity w/ temperature
logits_xy = (s_x @ s_y.T) / t # B x B
logits_yx = (s_y @ s_x.T) / t # B x B
loss_xy = F.cross_entropy(logits_xy, labels)
loss_yx = F.cross_entropy(logits_yx, labels)
loss = (loss_xy + loss_yx) / 2.0
# optimize
loss.backward()
optimizer.step()
losses.append(loss.item())
if i % save_every == 0 and savedir:
os.makedirs(savedir, exist_ok=True)
torch.save({"model": (hhX, hhY, t), "loss_history": losses, "opt": optimizer}, os.path.join(savedir, f"step_{i}.pt"))
if i % print_every == 0:
print(i, loss.item(), time.time() - last_time)
total_norm = 0
max_norm = 0
for p in params:
if (p is None) or (p.grad is None):
continue
param_norm = p.grad.detach().data.norm(2)
max_norm = max(max_norm, param_norm)
total_norm += param_norm.item() ** 2
total_norm = total_norm ** 0.5
print(f"GRAD NORM: {total_norm:.08f}, MAX NORM: {max_norm:.08f}, LR: {optimizer.param_groups[0]['lr']:.08f}, T: {t.item():.04f}", flush=True)
last_time = time.time()
# ind = (ind + 2*B) % (2*B)
return optimizer, losses, t
def main():
random.seed(42)
np.random.seed(42)
torch.manual_seed(42)
# torch.use_deterministic_algorithms(True)
parser = argparse.ArgumentParser()
parser.add_argument("--steps", type=int, default=1e5)
parser.add_argument("--batch", type=int, default=1e3)
parser.add_argument("--output")
args = parser.parse_args()
adapterx = FusionAdapter(D_x, D_s, depth_x, n_paths=1, dropout=0.4)
adaptery = FusionAdapter(D_y, D_s, depth_y, n_paths=1, dropout=0.4)
adapterx.cuda().train()
adaptery.cuda().train()
path = args.output if args.output else f"model_{datetime.utcnow().isoformat()}/"
opt, losses, t = train(adapterx, adaptery, STEPS=args.steps, B=args.batch, savedir=path)
torch.save({"model": (adapterx, adaptery, t), "loss_history": losses, "opt": opt}, os.path.join(path, "model.pt"))
if __name__ == "__main__":
main()