File size: 9,713 Bytes
987ed1b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from diffusion_policy.sampler.metric import euclidean_distance, coverage_distance

import pdb
torch.set_printoptions(precision=2, sci_mode=False)

def contrastive_sampler(strong, weak, obs_dict, num_sample=10, num_mode=3, name='contrast'):
    """
    Sample an action by contrasting outputs from strong and weak policies.

    Args:
        strong: a strong policy to predict near-optimal sequences of actions
        weak: a weak policy to predict sub-optimal sequences of actions
        obs_dict: dictionary containing observations at the current time step
        num_sample (int, optional): number of samples to generate
        name (str, optional): type of samples ('contrast', 'positive', 'negative')
        num_mode (int, optional): Factor to determine the number of top samples to consider

    Returns:
        dict: A dictionary of actions sampled using the contrastive approach.
    """
    # pre-process
    B, OH, OD = obs_dict['obs'].shape
    obs_dict_batch = dict()
    obs_dict_batch = {key: val.unsqueeze(1).repeat(1, num_sample, 1, 1).view(B * num_sample, OH, OD) 
                      for key, val in obs_dict.items()}

    dist_avg_pos = 0.0
    dist_avg_neg = 0.0

    # positive samples
    action_strong_batch = strong.predict_action(obs_dict_batch)
    AH, PH, AD = action_strong_batch['action'].shape[1], action_strong_batch['action_pred'].shape[1], action_strong_batch['action_pred'].shape[2]

    action_strong_batch['action'] = action_strong_batch['action'].reshape(B, num_sample, AH, AD)
    action_strong_batch['action_pred'] = action_strong_batch['action_pred'].reshape(B, num_sample, PH, AD)
    if 'action_obs_pred' in action_strong_batch:
        action_strong_batch['action_obs_pred'] = action_strong_batch['action_obs_pred'].reshape(B, num_sample, AH, OD)
    if 'obs_pred' in action_strong_batch:
        action_strong_batch['obs_pred'] = action_strong_batch['obs_pred'].reshape(B, num_sample, PH, OD)

    src_expand = action_strong_batch['action_pred'].unsqueeze(1)
    tar_expand =  action_strong_batch['action_pred'].unsqueeze(2)
    dist_pos = euclidean_distance(src_expand, tar_expand).view(B, num_sample, num_sample)

    topk = num_sample // num_mode + 1
    values, _ = torch.topk(dist_pos, k=topk, largest=False, dim=-1)
    dist_avg_pos = values[:, :, 1:].mean(dim=-1)      # skip the self-distance first element 

    if name == "negative": dist_avg_pos.zero_()

    # negative samples
    if weak:
        action_weak_batch = weak.predict_action(obs_dict_batch)
        action_weak_batch['action'] = action_weak_batch['action'].reshape(B, num_sample, AH, AD)
        action_weak_batch['action_pred'] = action_weak_batch['action_pred'].reshape(B, num_sample, PH, AD)
        if 'action_obs_pred' in action_weak_batch:
            action_weak_batch['action_obs_pred'] = action_weak_batch['action_obs_pred'].reshape(B, num_sample, AH, OD)
        if 'obs_pred' in action_weak_batch:
            action_weak_batch['obs_pred'] = action_weak_batch['obs_pred'].reshape(B, num_sample, PH, OD)

        src_expand = action_strong_batch['action_pred'].unsqueeze(1)
        tar_expand = action_weak_batch['action_pred'].unsqueeze(2)
        dist_neg = euclidean_distance(src_expand, tar_expand).view(B, num_sample, num_sample)

        topk = num_sample // num_mode
        values, _ = torch.topk(dist_neg, k=topk, largest=False, dim=-1)
        dist_avg_neg = values[:, :, 0:].mean(dim=-1)

        if name == "positive": dist_avg_neg.zero_()

    # sample selection
    dist_avg = dist_avg_pos - dist_avg_neg
    index = dist_avg.argmin(dim=-1)

    # slicing
    action_dict = dict()
    range_tensor = torch.arange(B, device=index.device)
    for key in action_strong_batch.keys():
        action_dict[key] = action_strong_batch[key][range_tensor, index]

    return action_dict

def bidirectional_sampler(strong, weak, obs_dict, prior, num_sample=10, beta=0.99, num_mode=3):
    """
    Sample an action that preserves coherence with a prior and contrast outputs from strong and weak policies.
    Args:
        strong: a strong policy to predict near-optimal sequences of actions
        weak: a weak policy to predict sub-optimal sequences of actions
        prior: the prediction made in the previous time step
        obs_dict: dictionary containing observations at the current time step
        num_sample (int, optional): number of samples to generate
        beta (float, optional): weight decay factor for backward coherence
        num_mode (int, optional): Factor to determine the number of top samples to consider

    Returns:
        dict: A dictionary of actions sampled using the contrastive approach.
    """    
    # pre-process
    B, OH, OD = obs_dict['obs'].shape
    obs_dict_batch = dict()
    for key in obs_dict.keys():
        if key == 'prior':
            continue        
        obs_dict_batch[key] = obs_dict[key].unsqueeze(1).repeat(1, num_sample, 1, 1).view(B * num_sample, OH, OD)

    # predict
    action_strong_batch = strong.predict_action(obs_dict_batch)

    # post-process
    AH, PH, AD = action_strong_batch['action'].shape[1], action_strong_batch['action_pred'].shape[1], action_strong_batch['action_pred'].shape[2]

    action_strong_batch['action'] = action_strong_batch['action'].reshape(B, num_sample, AH, AD)
    action_strong_batch['action_pred'] = action_strong_batch['action_pred'].reshape(B, num_sample, PH, AD)
    if 'action_obs_pred' in action_strong_batch:
        action_strong_batch['action_obs_pred'] = action_strong_batch['action_obs_pred'].reshape(B, num_sample, AH, OD)
    if 'obs_pred' in action_strong_batch:
        action_strong_batch['obs_pred'] = action_strong_batch['obs_pred'].reshape(B, num_sample, PH, OD)

    if weak:
        action_weak_batch = weak.predict_action(obs_dict_batch)
        action_weak_batch['action'] = action_weak_batch['action'].reshape(B, num_sample, AH, AD)
        action_weak_batch['action_pred'] = action_weak_batch['action_pred'].reshape(B, num_sample, PH, AD)
        if 'action_obs_pred' in action_weak_batch:
            action_weak_batch['action_obs_pred'] = action_weak_batch['action_obs_pred'].reshape(B, num_sample, AH, OD)
        if 'obs_pred' in action_weak_batch:
            action_weak_batch['obs_pred'] = action_weak_batch['obs_pred'].reshape(B, num_sample, PH, OD)

    # backward
    if prior is not None:
        # distance measure
        start_overlap = strong.n_obs_steps - 1
        end_overlap = prior.shape[1]
        num_sample = num_sample // num_mode
        dist_raw = euclidean_distance(action_strong_batch['action_pred'][:, :, start_overlap:end_overlap], prior.unsqueeze(1)[:, :, start_overlap:], reduction='none')

        weights = torch.tensor([beta**i for i in range(end_overlap-start_overlap)]).to(dist_raw.device)
        weights = weights / weights.sum()
        dist_weighted = dist_raw * weights.view(1, 1, end_overlap-start_overlap)
        dist_strong_sum = dist_weighted.sum(dim=2)
        _, cross_index = dist_strong_sum.sort(descending=False)
        index = cross_index[:, 0:num_sample]

        # slicing
        action_dict = dict()
        range_tensor = torch.arange(B, device=index.device)
        for key in action_strong_batch.keys():
            action_dict[key] = action_strong_batch[key][range_tensor.unsqueeze(1), index]
        action_strong_batch = action_dict
        dist_avg_prior = dist_strong_sum[range_tensor.unsqueeze(1), index]

        if weak:
            # sample selection
            dist_weak = euclidean_distance(action_weak_batch['action_pred'][:, :, start_overlap:end_overlap], prior.unsqueeze(1)[:, :, start_overlap:], reduction='none')
            dist_weighted = dist_weak * weights.view(1, 1, end_overlap-start_overlap)
            dist_weak_sum = dist_weighted.sum(dim=2)
            _, cross_index = dist_weak_sum.sort(descending=False)
            index = cross_index[:, 0:num_sample]

            # slicing
            action_dict = dict()
            range_tensor = torch.arange(B, device=index.device)
            for key in action_weak_batch.keys():
                action_dict[key] = action_weak_batch[key][range_tensor.unsqueeze(1), index]
            action_weak_batch = action_dict

        # balance between backward and forward
        ratio = (PH * beta) ** 2 / ((PH * beta) ** 2 + AH ** 2)
    else:
        dist_avg_prior = 0.0
        ratio = 0.0

    # positive samples
    src_expand = action_strong_batch['action_pred'].unsqueeze(1)
    tar_expand =  action_strong_batch['action_pred'].unsqueeze(2)
    dist_pos = euclidean_distance(src_expand, tar_expand).view(B, num_sample, num_sample)

    # topk = num_sample
    topk = num_sample // 2 + 1
    values, _ = torch.topk(dist_pos, k=topk, largest=False, dim=-1)
    dist_avg_pos = values[:, :, 1:].mean(dim=-1)      # skip the self-distance first element 

    if weak:
        # negative samples
        src_expand = action_strong_batch['action_pred'].unsqueeze(1)
        tar_expand = action_weak_batch['action_pred'].unsqueeze(2)
        dist_neg = euclidean_distance(src_expand, tar_expand).view(B, num_sample, num_sample)

        topk = num_sample // 2
        values, _ = torch.topk(dist_neg, k=topk, largest=False, dim=-1)
        dist_avg_neg = values[:, :, 0:].mean(dim=-1)
    else:
        dist_avg_neg = 0

    # sample selection
    dist_avg = dist_avg_prior * ratio + (dist_avg_pos - dist_avg_neg) * (1 - ratio)
    _, index = dist_avg.min(dim=-1)

    # slicing
    action_dict = dict()
    range_tensor = torch.arange(B, device=index.device)
    for key in action_strong_batch.keys():
        action_dict[key] = action_strong_batch[key][range_tensor, index]

    return action_dict