File size: 6,416 Bytes
ed00004
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import datetime
import time
from pathlib import Path

import einops
import torch
import torch.nn.functional as F

from src.test.webvid_covr import eval_recall
from src.tools.files import json_dump


class TestWebVidCoVRTextOnly:
    def __init__(self, remove_self_similarity=True):
        self.remove_self_similarity = remove_self_similarity

    @torch.no_grad()
    def __call__(self, model, data_loader, fabric):
        model.eval()

        fabric.print("Computing features for evaluation...")
        start_time = time.time()

        tar_img_feats = []
        query_feats = []
        captions = []
        pair_ids = []

        for _, tar_feat, caption, pair_id, *_ in data_loader:
            pair_ids.extend(pair_id.cpu().numpy().tolist())
            captions.extend(caption)

            device = pair_id.device

            text = model.tokenizer(
                caption,
                padding="longest",
                truncation=True,
                max_length=64,
                return_tensors="pt",
            ).to(device)

            # Shift encoder
            query_embs = model.text_encoder(
                text.input_ids,
                attention_mask=text.attention_mask,
                return_dict=True,
                mode="text",
            )
            query_feat = query_embs.last_hidden_state[:, 0, :]
            query_feat = F.normalize(model.text_proj(query_feat), dim=-1)
            query_feats.append(query_feat.cpu())

            # Encode the target image
            tar_img_feats.append(tar_feat.cpu())

        query_feats = torch.cat(query_feats, dim=0)
        tar_img_feats = torch.cat(tar_img_feats, dim=0)

        query_feats = F.normalize(query_feats, dim=-1)
        tar_img_feats = F.normalize(tar_img_feats, dim=-1)

        ref_img_ids = [data_loader.dataset.pairid2ref[pair_id] for pair_id in pair_ids]
        tar_img_ids = [data_loader.dataset.pairid2tar[pair_id] for pair_id in pair_ids]

        ref_img_ids = torch.tensor(ref_img_ids, dtype=torch.long)
        tar_img_ids = torch.tensor(tar_img_ids, dtype=torch.long)

        if fabric.world_size > 1:
            # Gather tensors from every process
            query_feats = fabric.all_gather(query_feats)
            tar_img_feats = fabric.all_gather(tar_img_feats)
            ref_img_ids = fabric.all_gather(ref_img_ids)
            tar_img_ids = fabric.all_gather(tar_img_ids)

            query_feats = einops.rearrange(query_feats, "d b e -> (d b) e")
            tar_img_feats = einops.rearrange(tar_img_feats, "d b e -> (d b) e")
            ref_img_ids = einops.rearrange(ref_img_ids, "d b -> (d b)")
            tar_img_ids = einops.rearrange(tar_img_ids, "d b -> (d b)")

        if fabric.global_rank == 0:
            sim_q2t = (query_feats @ tar_img_feats.t()).cpu().numpy()

            if self.remove_self_similarity:
                for i in range(len(ref_img_ids)):
                    for j in range(len(tar_img_ids)):
                        if ref_img_ids[i] == tar_img_ids[j]:
                            sim_q2t[i][j] = -10

            total_time = time.time() - start_time
            total_time_str = str(datetime.timedelta(seconds=int(total_time)))
            print("Evaluation time {}".format(total_time_str))

            recalls = eval_recall(sim_q2t)
            recalls["annotation"] = Path(data_loader.dataset.annotation_pth).name
            fabric.print(recalls)

            # Save results
            self_sim = "" if self.remove_self_similarity else "_ss"
            json_dump(recalls, f"recalls_covr_txt{self_sim}.json")

            print(f"Recalls saved in {Path.cwd()} as recalls_covr_txt{self_sim}.json")

        fabric.barrier()


class TestWebVidCoVRVisualOnly:
    def __init__(self):
        pass

    @staticmethod
    @torch.no_grad()
    def __call__(model, data_loader, fabric):
        model.eval()

        fabric.print("Computing features for evaluation...")
        start_time = time.time()

        tar_img_feats = []
        query_feats = []
        pair_ids = []

        for ref_img, tar_feat, _, pair_id, *_ in data_loader:
            pair_ids.extend(pair_id.cpu().numpy().tolist())

            ref_img_embs = model.visual_encoder(ref_img)
            query_feat = F.normalize(model.vision_proj(ref_img_embs[:, 0, :]), dim=-1)
            query_feats.append(query_feat.cpu())

            # Encode the target image
            tar_img_feats.append(tar_feat.cpu())

        query_feats = torch.cat(query_feats, dim=0)
        tar_img_feats = torch.cat(tar_img_feats, dim=0)

        query_feats = F.normalize(query_feats, dim=-1)
        tar_img_feats = F.normalize(tar_img_feats, dim=-1)

        ref_img_ids = [data_loader.dataset.pairid2ref[pair_id] for pair_id in pair_ids]
        tar_img_ids = [data_loader.dataset.pairid2tar[pair_id] for pair_id in pair_ids]

        ref_img_ids = torch.tensor(ref_img_ids, dtype=torch.long)
        tar_img_ids = torch.tensor(tar_img_ids, dtype=torch.long)

        if fabric.world_size > 1:
            # Gather tensors from every process
            query_feats = fabric.all_gather(query_feats)
            tar_img_feats = fabric.all_gather(tar_img_feats)
            ref_img_ids = fabric.all_gather(ref_img_ids)
            tar_img_ids = fabric.all_gather(tar_img_ids)

            query_feats = einops.rearrange(query_feats, "d b e -> (d b) e")
            tar_img_feats = einops.rearrange(tar_img_feats, "d b e -> (d b) e")
            ref_img_ids = einops.rearrange(ref_img_ids, "d b -> (d b)")
            tar_img_ids = einops.rearrange(tar_img_ids, "d b -> (d b)")

        if fabric.global_rank == 0:
            sim_q2t = (query_feats @ tar_img_feats.t()).cpu().numpy()

            # Add zeros where ref_img_id == tar_img_id
            for i in range(len(ref_img_ids)):
                for j in range(len(tar_img_ids)):
                    if ref_img_ids[i] == tar_img_ids[j]:
                        sim_q2t[i][j] = -10

            total_time = time.time() - start_time
            total_time_str = str(datetime.timedelta(seconds=int(total_time)))
            print("Evaluation time {}".format(total_time_str))

            recalls = eval_recall(sim_q2t)
            fabric.print(recalls)

            # Save results
            json_dump(recalls, "recalls_covr.json")

            print(f"Recalls saved in {Path.cwd()} as recalls_covr.json")

        fabric.barrier()