File size: 8,767 Bytes
9bf9e42
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
218
import torch
import torch.nn as nn
from ..utils.rewards import get_scores, get_self_cider_scores

class RewardCriterion(nn.Module):
    def __init__(self):
        super(RewardCriterion, self).__init__()

    def forward(self, input, seq, reward):
        input = input.gather(2, seq.unsqueeze(2)).squeeze(2)
        
        input = input.reshape(-1)
        reward = reward.reshape(-1)
        mask = (seq>0).to(input)
        mask = torch.cat([mask.new(mask.size(0), 1).fill_(1), mask[:, :-1]], 1).reshape(-1)
        output = - input * reward * mask
        output = torch.sum(output) / torch.sum(mask)

        return output

class StructureLosses(nn.Module):
    """
    This loss is inspired by Classical Structured Prediction Losses for Sequence to Sequence Learning (Edunov et al., 2018).
    """
    def __init__(self, opt):
        super(StructureLosses, self).__init__()
        self.opt = opt
        self.loss_type = opt.structure_loss_type

    def forward(self, input, seq, data_gts):
        """
        Input is either logits or log softmax
        """
        out = {}

        batch_size = input.size(0)# batch_size = sample_size * seq_per_img
        seq_per_img = batch_size // len(data_gts)

        assert seq_per_img == self.opt.train_sample_n, seq_per_img

        mask = (seq>0).to(input)
        mask = torch.cat([mask.new_full((mask.size(0), 1), 1), mask[:, :-1]], 1)
        
        scores = get_scores(data_gts, seq, self.opt)
        scores = torch.from_numpy(scores).type_as(input).view(-1, seq_per_img)
        out['reward'] = scores #.mean()
        if self.opt.entropy_reward_weight > 0:
            entropy = - (F.softmax(input, dim=2) * F.log_softmax(input, dim=2)).sum(2).data
            entropy = (entropy * mask).sum(1) / mask.sum(1)
            print('entropy', entropy.mean().item())
            scores = scores + self.opt.entropy_reward_weight * entropy.view(-1, seq_per_img)
        # rescale cost to [0,1]
        costs = - scores
        if self.loss_type == 'risk' or self.loss_type == 'softmax_margin': 
            costs = costs - costs.min(1, keepdim=True)[0]
            costs = costs / costs.max(1, keepdim=True)[0]
        # in principle
        # Only risk need such rescale
        # margin should be alright; Let's try.

        # Gather input: BxTxD -> BxT
        input = input.gather(2, seq.unsqueeze(2)).squeeze(2)

        if self.loss_type == 'seqnll':
            # input is logsoftmax
            input = input * mask
            input = input.sum(1) / mask.sum(1)
            input = input.view(-1, seq_per_img)

            target = costs.min(1)[1]
            output = F.cross_entropy(input, target)
        elif self.loss_type == 'risk':
            # input is logsoftmax
            input = input * mask
            input = input.sum(1)
            input = input.view(-1, seq_per_img)

            output = (F.softmax(input.exp()) * costs).sum(1).mean()

            # test
            # avg_scores = input
            # probs = F.softmax(avg_scores.exp_())
            # loss = (probs * costs.type_as(probs)).sum() / input.size(0)
            # print(output.item(), loss.item())            

        elif self.loss_type == 'max_margin':
            # input is logits
            input = input * mask
            input = input.sum(1) / mask.sum(1)
            input = input.view(-1, seq_per_img)
            _, __ = costs.min(1, keepdim=True)
            costs_star = _
            input_star = input.gather(1, __)
            output = F.relu(costs - costs_star - input_star + input).max(1)[0] / 2
            output = output.mean()

            # sanity test
            # avg_scores = input + costs
            # scores_with_high_target = avg_scores.clone()
            # scores_with_high_target.scatter_(1, costs.min(1)[1].view(-1, 1), 1e10)

            # target_and_offender_index = scores_with_high_target.sort(1, True)[1][:, 0:2]
            # avg_scores = avg_scores.gather(1, target_and_offender_index)
            # target_index = avg_scores.new_zeros(avg_scores.size(0), dtype=torch.long)
            # loss = F.multi_margin_loss(avg_scores, target_index, size_average=True, margin=0)
            # print(loss.item() * 2, output.item()) 

        elif self.loss_type == 'multi_margin':
            # input is logits
            input = input * mask
            input = input.sum(1) / mask.sum(1)
            input = input.view(-1, seq_per_img)
            _, __ = costs.min(1, keepdim=True)
            costs_star = _
            input_star = input.gather(1, __)
            output = F.relu(costs - costs_star - input_star + input)
            output = output.mean()

            # sanity test
            # avg_scores = input + costs
            # loss = F.multi_margin_loss(avg_scores, costs.min(1)[1], margin=0)
            # print(output, loss)

        elif self.loss_type == 'softmax_margin':
            # input is logsoftmax
            input = input * mask
            input = input.sum(1) / mask.sum(1)
            input = input.view(-1, seq_per_img)

            input = input + costs
            target = costs.min(1)[1]
            output = F.cross_entropy(input, target)

        elif self.loss_type == 'real_softmax_margin':
            # input is logits
            # This is what originally defined in Kevin's paper
            # The result should be equivalent to softmax_margin
            input = input * mask
            input = input.sum(1) / mask.sum(1)
            input = input.view(-1, seq_per_img)

            input = input + costs
            target = costs.min(1)[1]
            output = F.cross_entropy(input, target)

        elif self.loss_type == 'new_self_critical':
            """
            A different self critical
            Self critical uses greedy decoding score as baseline;
            This setting uses the average score of the rest samples as baseline
            (suppose c1...cn n samples, reward1 = score1 - 1/(n-1)(score2+..+scoren) )
            """
            baseline = (scores.sum(1, keepdim=True) - scores) / (scores.shape[1] - 1)
            scores = scores - baseline
            # self cider used as reward to promote diversity (not working that much in this way)
            if getattr(self.opt, 'self_cider_reward_weight', 0) > 0:
                _scores = get_self_cider_scores(data_gts, seq, self.opt)
                _scores = torch.from_numpy(_scores).type_as(scores).view(-1, 1)
                _scores = _scores.expand_as(scores - 1)
                scores += self.opt.self_cider_reward_weight * _scores
            output = - input * mask * scores.view(-1, 1)
            output = torch.sum(output) / torch.sum(mask)

        out['loss'] = output
        return out

class LanguageModelCriterion(nn.Module):
    def __init__(self):
        super(LanguageModelCriterion, self).__init__()

    def forward(self, input, target, mask):
        if target.ndim == 3:
            target = target.reshape(-1, target.shape[2])
            mask = mask.reshape(-1, mask.shape[2])
        # truncate to the same size
        target = target[:, :input.size(1)]
        mask = mask[:, :input.size(1)].to(input)

        output = -input.gather(2, target.unsqueeze(2)).squeeze(2) * mask
        # Average over each token
        output = torch.sum(output) / torch.sum(mask)

        return output

class LabelSmoothing(nn.Module):
    "Implement label smoothing."
    def __init__(self, size=0, padding_idx=0, smoothing=0.0):
        super(LabelSmoothing, self).__init__()
        self.criterion = nn.KLDivLoss(size_average=False, reduce=False)
        # self.padding_idx = padding_idx
        self.confidence = 1.0 - smoothing
        self.smoothing = smoothing
        # self.size = size
        self.true_dist = None
        
    def forward(self, input, target, mask):
        if target.ndim == 3:
            target = target.reshape(-1, target.shape[2])
            mask = mask.reshape(-1, mask.shape[2])
        # truncate to the same size
        target = target[:, :input.size(1)]
        mask =  mask[:, :input.size(1)]

        input = input.reshape(-1, input.size(-1))
        target = target.reshape(-1)
        mask = mask.reshape(-1).to(input)

        # assert x.size(1) == self.size
        self.size = input.size(1)
        # true_dist = x.data.clone()
        true_dist = input.data.clone()
        # true_dist.fill_(self.smoothing / (self.size - 2))
        true_dist.fill_(self.smoothing / (self.size - 1))
        true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence)
        # true_dist[:, self.padding_idx] = 0
        # mask = torch.nonzero(target.data == self.padding_idx)
        # self.true_dist = true_dist
        return (self.criterion(input, true_dist).sum(1) * mask).sum() / mask.sum()