-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathzs_spur_correlation.py
138 lines (107 loc) · 5.97 KB
/
zs_spur_correlation.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
from itertools import chain
import torch
from torchvision import transforms as T
import numpy as np
from torch.utils.data import DataLoader
from datasets.clean_waterbirds import *
from helpers.linear_decompose import *
from helpers.inspect_utils import *
from helpers.utils import *
from helpers.model_utils import *
from helpers.decompose_utils import *
from helpers.interpret_utils import *
set_seed(0)
model_keys = ["DeiT", "CLIP", "DINO", "DINOv2", "SWIN", "MaxVit"]
dataset_name = 'waterbirds'
correlated_waterbirds_dataset = None # replace with path to correlated dataset
uncorrelated_waterbirds_dataset = None # replace with path to uncorrelated dataset
num_comps = 10
save_path = f'./saved_plots/zs_spur_correlation_{dataset_name}.csv'
def get_spur_inds(embeds_decomp, spur_embeds, core_embeds, temp=1.0, method='var'):
with torch.no_grad():
spur_vars = variance_attributed(clip_aligner_head(embeds_decomp), spur_embeds)
core_vars = variance_attributed(clip_aligner_head(embeds_decomp), core_embeds)
spur_to_core_ratio = spur_vars - core_vars
sorted_inds = torch.argsort(spur_to_core_ratio, descending=True)
return sorted_inds, spur_to_core_ratio
def mitigate_spur(embeds_decomp, sorted_inds, topk=10):
spur_inds = sorted_inds[:topk]
abl_embeds_decomp = embeds_decomp.clone()
abl_embeds_decomp[spur_inds] = abl_embeds_decomp[spur_inds].mean(dim=1, keepdims=True)
return abl_embeds_decomp
def get_group_accuracies(pred_head, embeds, labels, group_dict):
acc_dict = dict()
for grp_desc, grp_labels in group_dict.items():
embeds_grp = embeds[grp_labels]
labels_grp = labels[grp_labels]
preds = pred_head(embeds_grp)
acc_dict[grp_desc] = (preds.argmax(dim=-1) == labels_grp).float().mean().item()
return acc_dict
with open("./templates.txt", "r") as fp:
templates = [x.strip() for x in fp.readlines()]
num_workers = 4*torch.cuda.device_count()
gpu_size = 512*torch.cuda.device_count()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
for model_key in model_keys:
model, model_descr, batch_size, pred_head = load_model(model_key, device, pred_head_type=None)
if model_key == "SWIN":
detach_block, end_block = (2, 14), (3,2)
elif model_key == "MaxVit":
detach_block, end_block = (2,3), (3,2)
else:
detach_block, end_block = 7, 12
model.detach_from_res(detach_block, end_block)
model.freeze_blocks(0, detach_block)
model.to(device)
_=model.eval()
waterbirds_classes = ['landbird', 'waterbird']
classes = waterbirds_classes
super_dataset = CleanWaterbirdsDataset(correlated_waterbirds_dataset, transform=model.preprocess)
train_dataset, val_dataset = torch.utils.data.random_split(super_dataset, [0.8, 0.2])
test_dataset = CleanWaterbirdsDataset(uncorrelated_waterbirds_dataset, transform=model.preprocess)
train_dataloader = DataLoader(train_dataset, batch_size=64, shuffle=False, num_workers=num_workers)
val_dataloader = DataLoader(val_dataset, batch_size=16, shuffle=False, num_workers=num_workers)
test_dataloader = DataLoader(test_dataset, batch_size=16, shuffle=False, num_workers=num_workers)
if 'clip' in model_descr:
pred_head = get_clip_text_embeds(model, waterbirds_classes, templates, device,
load_file=f"./saved_outputs/{model_descr}_wb_zeroshot_classes_head.pt")
else:
pred_head = get_head(model, train_dataloader, len(classes), device, bias=True,
load_file=f"./saved_outputs/{model_descr}_wb_trained_head.pt",
epochs=10)
pred_head.cpu()
comp_names, embeds_decomp, labels = \
get_decomposed_embeds(model, chain(val_dataloader, test_dataloader),
len(val_dataloader) + len(test_dataloader), device,
load_file=f'./saved_outputs/{model_descr}_waterbirds_decomposed_embeds.pt')
labels, bg_labels = labels[:,0], labels[:,1]
spur_feats_desc = ['water', 'ocean', 'lake', 'sky', 'land', 'forest', 'trees', 'bamboo', 'leaves']
core_feats_desc = ['bird', 'finch', 'hummingbird', 'crow', 'sparrow', 'swan', 'albatross', 'pelican']
group_dict = {
'waterbird on water': (labels == 1) & (bg_labels == 1),
'waterbird on land': (labels == 1) & (bg_labels == 0),
'landbird on water': (labels == 0) & (bg_labels == 1),
'landbird on land': (labels == 0) & (bg_labels == 0)
}
clip_model, clip_model_descr, clip_aligner_head = get_clip_and_aligner(model, model_descr, device)
spur_feat_embeds = get_clip_text_embeds(clip_model, spur_feats_desc, templates, device).weight.data.cpu()
core_feat_embeds = get_clip_text_embeds(clip_model, core_feats_desc, templates, device).weight.data.cpu()
orig_grp_acc_dict = get_group_accuracies(pred_head, embeds_decomp.sum(0), labels, group_dict)
sorted_inds, _ = get_spur_inds(embeds_decomp, spur_feat_embeds, core_feat_embeds, temp=1, method='var')
new_embeds_decomp = mitigate_spur(embeds_decomp, sorted_inds, topk=num_comps)
new_grp_acc_dict = get_group_accuracies(pred_head, new_embeds_decomp.sum(0), labels, group_dict)
if not os.path.exists(save_path):
# Write header
with open(save_path, 'w') as fp:
fp.write('model_key,num_comps')
for k in orig_grp_acc_dict.keys():
fp.write(f",{k}")
fp.write(f", worst group accuracy, average group accuracy")
fp.write('\n')
with open(save_path, 'a') as fp:
fp.write(f"{model_key},{num_comps}")
for k, v in orig_grp_acc_dict.items():
fp.write(f", {v:.3f} $\\rightarrow$ {new_grp_acc_dict[k]:.3f}")
fp.write(f", {min(orig_grp_acc_dict.values()):.3f} $\\rightarrow$ {min(new_grp_acc_dict.values()):.3f},"
f"{np.mean(list(orig_grp_acc_dict.values())):.3f} $\\rightarrow$ {np.mean(list(new_grp_acc_dict.values())):.3f}")
fp.write('\n')