File size: 8,039 Bytes
b34d1d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
208
209
210
211
212
213
214
215
216
217
import torch
import torch.nn as nn
from torch.nn import functional as F

try:
    import torch.distributed.nn
    from torch import distributed as dist

    has_distributed = True
except ImportError:
    has_distributed = False

try:
    import horovod.torch as hvd
except ImportError:
    hvd = None


def gather_features(
        image_features,
        text_features,
        local_loss=False,
        gather_with_grad=False,
        rank=0,
        world_size=1,
        use_horovod=False
):
    assert has_distributed, 'torch.distributed did not import correctly, please use a PyTorch version with support.'
    if use_horovod:
        assert hvd is not None, 'Please install horovod'
        if gather_with_grad:
            all_image_features = hvd.allgather(image_features)
            all_text_features = hvd.allgather(text_features)
        else:
            with torch.no_grad():
                all_image_features = hvd.allgather(image_features)
                all_text_features = hvd.allgather(text_features)
            if not local_loss:
                # ensure grads for local rank when all_* features don't have a gradient
                gathered_image_features = list(all_image_features.chunk(world_size, dim=0))
                gathered_text_features = list(all_text_features.chunk(world_size, dim=0))
                gathered_image_features[rank] = image_features
                gathered_text_features[rank] = text_features
                all_image_features = torch.cat(gathered_image_features, dim=0)
                all_text_features = torch.cat(gathered_text_features, dim=0)
    else:
        # We gather tensors from all gpus
        if gather_with_grad:
            all_image_features = torch.cat(torch.distributed.nn.all_gather(image_features), dim=0)
            all_text_features = torch.cat(torch.distributed.nn.all_gather(text_features), dim=0)
        else:
            gathered_image_features = [torch.zeros_like(image_features) for _ in range(world_size)]
            gathered_text_features = [torch.zeros_like(text_features) for _ in range(world_size)]
            dist.all_gather(gathered_image_features, image_features)
            dist.all_gather(gathered_text_features, text_features)
            if not local_loss:
                # ensure grads for local rank when all_* features don't have a gradient
                gathered_image_features[rank] = image_features
                gathered_text_features[rank] = text_features
            all_image_features = torch.cat(gathered_image_features, dim=0)
            all_text_features = torch.cat(gathered_text_features, dim=0)

    return all_image_features, all_text_features


class ClipLoss(nn.Module):

    def __init__(
            self,
            local_loss=False,
            gather_with_grad=False,
            cache_labels=False,
            rank=0,
            world_size=1,
            use_horovod=False,
    ):
        super().__init__()
        self.local_loss = local_loss
        self.gather_with_grad = gather_with_grad
        self.cache_labels = cache_labels
        self.rank = rank
        self.world_size = world_size
        self.use_horovod = use_horovod

        # cache state
        self.prev_num_logits = 0
        self.labels = {}

    def get_ground_truth(self, device, num_logits) -> torch.Tensor:
        # calculated ground-truth and cache if enabled
        if self.prev_num_logits != num_logits or device not in self.labels:
            labels = torch.arange(num_logits, device=device, dtype=torch.long)
            if self.world_size > 1 and self.local_loss:
                labels = labels + num_logits * self.rank
            if self.cache_labels:
                self.labels[device] = labels
                self.prev_num_logits = num_logits
        else:
            labels = self.labels[device]
        return labels

    def get_logits(self, image_features, text_features, logit_scale):
        if self.world_size > 1:
            all_image_features, all_text_features = gather_features(
                image_features, text_features,
                self.local_loss, self.gather_with_grad, self.rank, self.world_size, self.use_horovod)

            if self.local_loss:
                logits_per_image = logit_scale * image_features @ all_text_features.T
                logits_per_text = logit_scale * text_features @ all_image_features.T
            else:
                logits_per_image = logit_scale * all_image_features @ all_text_features.T
                logits_per_text = logits_per_image.T
        else:
            logits_per_image = logit_scale * image_features @ text_features.T
            logits_per_text = logit_scale * text_features @ image_features.T
        
        return logits_per_image, logits_per_text

    def forward(self, image_features, text_features, logit_scale, output_dict=False):
        device = image_features.device
        logits_per_image, logits_per_text = self.get_logits(image_features, text_features, logit_scale)

        labels = self.get_ground_truth(device, logits_per_image.shape[0])

        total_loss = (
            F.cross_entropy(logits_per_image, labels) +
            F.cross_entropy(logits_per_text, labels)
        ) / 2

        return {"contrastive_loss": total_loss} if output_dict else total_loss


class CoCaLoss(ClipLoss):
    def __init__(
            self,
            caption_loss_weight,
            clip_loss_weight,
            pad_id=0,  # pad_token for open_clip custom tokenizer
            local_loss=False,
            gather_with_grad=False,
            cache_labels=False,
            rank=0,
            world_size=1,
            use_horovod=False,
    ):
        super().__init__(
            local_loss=local_loss,
            gather_with_grad=gather_with_grad,
            cache_labels=cache_labels,
            rank=rank,
            world_size=world_size,
            use_horovod=use_horovod
        )

        self.clip_loss_weight = clip_loss_weight
        self.caption_loss_weight = caption_loss_weight
        self.caption_loss = nn.CrossEntropyLoss(ignore_index=pad_id)

    def forward(self, image_features, text_features, logits, labels, logit_scale, output_dict=False):
        
        clip_loss = torch.tensor(0)
        
        if self.clip_loss_weight:
            clip_loss = super().forward(image_features, text_features, logit_scale)
            clip_loss = self.clip_loss_weight * clip_loss

        caption_loss = self.caption_loss(
            logits.permute(0, 2, 1),
            labels,
        )
        caption_loss = caption_loss * self.caption_loss_weight

        if output_dict:
            return {"contrastive_loss": clip_loss, "caption_loss": caption_loss}

        return clip_loss, caption_loss


class DistillClipLoss(ClipLoss):

    def dist_loss(self, teacher_logits, student_logits):
        return -(teacher_logits.softmax(dim=1) * student_logits.log_softmax(dim=1)).sum(dim=1).mean(dim=0)

    def forward(
            self,
            image_features,
            text_features,
            logit_scale,
            dist_image_features,
            dist_text_features,
            dist_logit_scale,
            output_dict=False,
    ):
        logits_per_image, logits_per_text = \
            self.get_logits(image_features, text_features, logit_scale)

        dist_logits_per_image, dist_logits_per_text = \
            self.get_logits(dist_image_features, dist_text_features, dist_logit_scale)

        labels = self.get_ground_truth(image_features.device, logits_per_image.shape[0])

        contrastive_loss = (
            F.cross_entropy(logits_per_image, labels) +
            F.cross_entropy(logits_per_text, labels)
        ) / 2

        distill_loss = (
            self.dist_loss(dist_logits_per_image, logits_per_image) +
            self.dist_loss(dist_logits_per_text, logits_per_text)
        ) / 2

        if output_dict:
            return {"contrastive_loss": contrastive_loss, "distill_loss": distill_loss}

        return contrastive_loss, distill_loss