GenZSL/train_clf.py at main · shiming-chen/GenZSL · GitHub

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

import h5py
import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import TensorDataset, DataLoader

from config import parse_args
from utils.helper_func import *
from utils.myLoader import MyDataloader
from model.myModels import Encoder, Decoder, Myclassifier


def generate_syn_data(generator, classes, embedding, in_package, args):
    syn_num = args.syn_num
    class_nums = classes.size(0)
    label2Idx = {value.item(): idx for idx, value in enumerate(classes)}
    syn_f, syn_l = [], []

    topk_l = in_package["topk_l"]
    seenlabel2Idx = in_package["sl2I"]
    label2features = in_package["l2f"]
    VAE_E = in_package["VAE_E"]
    for i in range(class_nums):
        iclass = classes[i]
        iclass_embedding = embedding[label2Idx[iclass.item()]]
        text_features = iclass_embedding.repeat(syn_num, 1)
        """
            - topk
        """
        topk_labels = topk_l[label2Idx[iclass.item()]]
        topk_idxes = torch.tensor([seenlabel2Idx[label.item()] for label in topk_labels])
        mix_topk_features = torch.zeros(text_features.size())
        weights = args.weights
        for i in range(args.topK):
            topi_features = label2features[topk_idxes[i]]
            n_samples = topi_features.size(0)
            n_rest = syn_num % n_samples
            features = topi_features.repeat(syn_num // n_samples, 1)
            if n_rest != 0:
                random_idxes = torch.randint(0, topi_features.size(0), (n_rest,))
                extra_features = topi_features[random_idxes]
                features = torch.cat((features, extra_features), dim=0)
            mix_topk_features += weights[i] * features.cpu()
        syn_input = mix_topk_features.to(args.device)

        with torch.no_grad():
           # ---------- VAE Encoder ---------- #
            mean, log_var = VAE_E(syn_input, text_features)
            std = torch.exp(0.5 * log_var)
            noise = torch.randn(mean.shape).to(args.device)
            z = std * noise + mean
            # ---------- VAE Decoder ---------- #
            syn_features = generator(torch.cat((z, text_features), dim=1))
            syn_features = syn_features / syn_features.norm(dim=-1, keepdim=True)
        syn_f.append(syn_features.cpu())
        syn_l.append(iclass.repeat(syn_num, 1))
    syn_f = np.concatenate(syn_f, axis=0)
    syn_l = np.concatenate(syn_l, axis=0)
    return syn_f, syn_l


args = parse_args()
print('args:', args)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)

# ======================================== Prepare dataset ======================================== #
myloader = MyDataloader(args)
seenlabel2Idx = {value.item(): idx for idx, value in enumerate(myloader.seenclasses)}
seenidx2Label = {idx: value.item() for idx, value in enumerate(myloader.seenclasses)}
# ======================================== CLIP features ======================================== #
print(" ==> Load extracted features.")
CLIP_feature_path = args.image_root + f"/{args.dataset}/CLIP_feature.hdf5"
hf = h5py.File(CLIP_feature_path, 'r')
train_f, train_l = np.array(hf.get('train_f')), np.array(hf.get('train_l'))
testSeen_f, testSeen_l = np.array(hf.get('testSeen_f')), np.array(hf.get('testSeen_l'))
testUnseen_f, testUnseen_l = np.array(hf.get('testUnseen_f')), np.array(hf.get('testUnseen_l'))
# ---------- text description data ---------- #
seen_embeddings = np.array(hf.get('seen_embeddings'))
unseen_embeddings = np.array(hf.get('unseen_embeddings'))
print(" ====> Feature loaded.")
train_f = torch.from_numpy(train_f).float().to(args.device)
train_l = torch.from_numpy(train_l).float().to(args.device)
testSeen_f = torch.from_numpy(testSeen_f).float().to(args.device)
testSeen_l = torch.from_numpy(testSeen_l).float().to(args.device)
testUnseen_f = torch.from_numpy(testUnseen_f).float().to(args.device)
testUnseen_l = torch.from_numpy(testUnseen_l).float().to(args.device)
# ---------- text description data ---------- #
seen_embeddings = torch.from_numpy(seen_embeddings).float().to(args.device)
unseen_embeddings = torch.from_numpy(unseen_embeddings).float().to(args.device)

all_embeddings = torch.cat([seen_embeddings, unseen_embeddings], dim=0)
U, _, _ = np.linalg.svd(all_embeddings.cpu())
_U = U[:, 1:]
P_text = np.dot(_U, _U.T)
all_embeddings = torch.from_numpy(np.dot(P_text, all_embeddings.cpu())).float().to(args.device)
seen_embeddings = all_embeddings[:seen_embeddings.size(0)]
unseen_embeddings = all_embeddings[seen_embeddings.size(0):]

"""
    - calculate top k similarity
"""
# seen topk similarity
seenSimilarity = F.cosine_similarity(seen_embeddings.unsqueeze(1), seen_embeddings.unsqueeze(0), dim=2)
seentopk = np.argsort(-seenSimilarity.cpu(), axis=1)[:, 1:1+args.topK]  # except itself
for i in range(len(seentopk)):
    for j, idx in enumerate(seentopk[i]):
        seentopk[i][j] = seenidx2Label[idx.item()]
# unseen topk similarity
unseenSimilarity = F.cosine_similarity(unseen_embeddings.unsqueeze(1), seen_embeddings.unsqueeze(0), dim=2)
unseentopk = np.argsort(-unseenSimilarity.cpu(), axis=1)[:, :args.topK]
for i in range(len(unseentopk)):
    for j, idx in enumerate(unseentopk[i]):
        unseentopk[i][j] = seenidx2Label[idx.item()]
"""
    - for random get topk features from train_seen
    - label2features -> a list of features
"""
label_idxes = [torch.nonzero(train_l == label).squeeze() for label in myloader.seenclasses]
label2features = [train_f[idxes, :] for idxes in label_idxes]

# ======================================== Config models ======================================== #
netE_weight = f"{args.dataset}_netE_bestGZSL.pth"
netD_weight = f"{args.dataset}_netD_bestGZSL.pth"
netE = Encoder().to(args.device)
netD = Decoder().to(args.device)
netE.load_state_dict(torch.load(netE_weight, map_location='cpu'))
netD.load_state_dict(torch.load(netD_weight, map_location='cpu'))

netCLF = Myclassifier(seen_embeddings.size(1), len(myloader.allclasses_name)).to(args.device)
optimizerCLF = torch.optim.Adam(netCLF.parameters(), lr=1e-3, betas=(0.5, 0.999))
criterion = nn.CrossEntropyLoss().to(args.device)  # for classifier
# ======================================== Config dataloader ======================================== #
# ---------- trainloader ---------- #
trainset = TensorDataset(train_f, train_l)
trainloader = DataLoader(trainset, batch_size=args.batch_size, shuffle=False)
# ======================================== Test Stage ======================================== #

# ---------- synthesize unseen classes features ---------- #
syn_unseen_f = []
syn_unseen_l = []
with torch.no_grad():
    netD.eval()
    in_package = {"topk_l": unseentopk, "l2f": label2features, "VAE_E": netE, "sl2I": seenlabel2Idx}
    syn_unseen_f, syn_unseen_l = generate_syn_data(
        netD, myloader.unseenclasses, unseen_embeddings, in_package, args)
syn_unseen_f = torch.from_numpy(syn_unseen_f).float().to(args.device)
syn_unseen_l = torch.from_numpy(syn_unseen_l).squeeze(1).long().to(args.device)
print(" ==> feature synthesized.")

# ---------- train classifier ---------- #
new_train_f = torch.cat((train_f, syn_unseen_f), dim=0)
new_train_l = torch.cat((train_l, syn_unseen_l), dim=0)
new_trainset = TensorDataset(new_train_f, new_train_l)
new_trainloader = DataLoader(new_trainset, batch_size=args.batch_size, shuffle=False)
record = [0.0, 0.0, 0.0]

for clf_epoch in range(args.clf_nepoch):
    loss_clf = []
    for batch_features, batch_labels in new_trainloader:
        netCLF.train()
        netCLF.zero_grad()
        batch_features, batch_labels = batch_features.to(args.device), batch_labels.long().to(args.device)
        output = netCLF(batch_features)
        loss = criterion(output, batch_labels)
        loss_clf.append(loss.item())
        loss.backward()
        optimizerCLF.step()

    testloader = {"test_seen_f": testSeen_f, "test_seen_l": testSeen_l,
                    "test_unseen_f": testUnseen_f, "test_unseen_l": testUnseen_l}
    acc_train = eval_train_acc(myloader, new_train_f, new_train_l, netCLF, args.device)
    S, U, H, _ = eval_zs_gzsl(myloader, testloader, netCLF, args.device)
    print(f" ====> clf_epoch:{clf_epoch}; Loss:{sum(loss_clf)/len(loss_clf)}; Acc_train: {acc_train}")
    print(f" ====> S:{S}; U:{U}; H:{H}")

    if H > record[2]:
        record[:] = [S, U, H]


print('Best GZSL|CZSL => S:{:.6f}; U:{:.6f}; H:{:.6f}'.format(record[0], record[1], record[2]))