File size: 13,981 Bytes
8014d08
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
from torch.utils.data import Dataset
import numpy as np
import copy
from torch.utils.data import Dataset

class DummyDataset(Dataset):
    def __init__(self, buffer):
        self.buffer = buffer

    def __len__(self):
        return len(self.buffer)

    def __getitem__(self, idx):
        return self.buffer[idx]


class Transition:
    def __init__(self, observation, action, reward, next_observation, done, **kwargs):
        self.observation       = observation
        self.action            = action
        self.reward            = np.single(reward)
        self.next_observation  = next_observation

        if isinstance(done, bool):
            self.done          = done
        elif done == 'False':
            self.done          = False
        elif done == 'True':
            self.done          = True
        else: 
            raise ValueError
        
        # internal, to see how many times a certain transition was sampled
        self.times_was_sampled = 0
    def as_dict(self, as_string = False):
        return {
            "observation":      self.observation,
            "action":           self.action,
            "reward":           self.reward if as_string == False else str(self.reward),
            "next_observation": self.next_observation,
            "done":             self.done if as_string == False else str(self.done)
        }
    def __str__(self):
        printout = '\n'
        for key in self.as_dict():
            printout += "\n" + key + ':'
            printout += '\n' + str(self.as_dict()[key])
        return printout

class Trajectory:
    def __init__(self):
        self.transitions = []
        self.info        = {}
    def __len__(self):
        return len(self.transitions)
    def check_consistency(self):
        assert(any([transition.done for transition in self.transitions[:-1]]) == False) # should not be done until the end
        assert(self.transitions[-1].done == True )
        for t in range(1,len(self.transitions)):
            prior_transition   = self.transitions[t-1]
            current_transition = self.transitions[t]
            assert(prior_transition.next_observation == current_transition.observation)
    def get_rewards(self):
        return [transition.reward for transition in self.transitions]
    def get_return(self):
        return sum([transition.reward for transition in self.transitions])
    def append(self, transition):
        assert(self.transitions == [] or self.transitions[-1].done == False)
        self.transitions.append(Transition(**transition))
    def __str__(self):
        printout = '\n*** Trajectory Begins *** \n'
        printout += "\nTrajectory Length: " + str(len(self))
        for idx, transition in enumerate(self.transitions):
            printout += "\nTransition: " + str(idx)
            printout += "\n" + transition.__str__()
        if self.info != None:
            printout += "\nFound Special Items"  
            printout += str(self.info)
        printout += '\n *** Trajectory Ends **** \n'
        return printout

class TrajectoryDataset(Dataset):
    def __init__(self):
        self.trajectories              = []
        self.samples                   = [] # pointer list for fast sampling
        self._last_oar                 = None # Last (observation, action, reward) for sequential addition
    def __len__(self):
        return len(self.samples)
    def __getitem__(self, idx): 
        self.samples[idx].times_was_sampled += 1
        ### Must return a copy to avoid issues if further processing is done
        return copy.deepcopy(self.samples[idx].as_dict())
    def append_trajectory(self, trajectory: Trajectory):
        trajectory.check_consistency()
        assert(self.last_trajectory_reached_end())
        for transition in trajectory.transitions:
            self.append_sample_sequentially(copy.deepcopy(transition.as_dict()))
        self.trajectories[-1].info = copy.deepcopy(trajectory.info)
        self.trajectories[-1].check_consistency()
    def append_observation_action_reward(self, observation, action, reward):
        if self._last_oar != None:
            self.append_sample_sequentially({"observation": self._last_oar["observation"],
                                             "action": self._last_oar["action"],
                                             "reward": self._last_oar["reward"],
                                             "next_observation": observation,
                                             "done": False })
        self._last_oar = {"observation": observation,
                          "action": action,
                          "reward": reward}
    def append_terminal_observation(self, observation, trajectory_info = None):
        assert self._last_oar != None
        self.append_sample_sequentially({"observation": self._last_oar["observation"],
                                         "action": self._last_oar["action"],
                                         "reward": self._last_oar["reward"],
                                         "next_observation": observation,
                                         "done": True })
        self._last_oar = None
        if trajectory_info != None:
            self.trajectories[-1].info = trajectory_info
        self.trajectories[-1].check_consistency()

    def last_trajectory_reached_end(self):  
        return (self.trajectories == [] or self.trajectories[-1].transitions[-1].done)
         
    def append_sample_sequentially(self, transition):
        ### is the trajectory new?
        if self.last_trajectory_reached_end():
            self.trajectories.append(Trajectory())
        self.trajectories[-1].transitions.append(Transition(**transition))
        self.samples.append(self.trajectories[-1].transitions[-1])
    def nTrajectories(self):
        return len(self.trajectories)
    def get_all_trajectory_returns(self):
        return np.asarray([trajectory.get_return() for trajectory in self.trajectories])
    def check_consistency(self):
        assert (sum([len(trajectory) for trajectory in self.trajectories]) == len(self.samples))
        for trajectory in self.trajectories:
            trajectory.check_consistency()
    def sample(self, batch_size=None):
        if batch_size is None:
            batch_size = self.batch_size
        rand_indices = np.random.randint(0, len(self.samples), size=(batch_size,))
        # rand_indices = [np.random.randint(0, len(self.samples)) for _ in range(batch_size)]
        for idx in rand_indices:
            self.samples[idx].times_was_sampled += 1 
        return {
            "observation":      [self.samples[idx].observation            for idx in rand_indices],
            "action":           [self.samples[idx].action                 for idx in rand_indices],
            "reward":           [self.samples[idx].reward                 for idx in rand_indices],
            "next_observation": [self.samples[idx].next_observation       for idx in rand_indices],
            "done":             [self.samples[idx].done                   for idx in rand_indices],
        }
    def mean_trajectory_return(self):
        return np.mean(self.get_all_trajectory_returns())
    def std_trajectory_return(self):
        return np.std(self.get_all_trajectory_returns())
    def merge(self, dataset):
        self.check_consistency()
        dataset.check_consistency()
        for trajectory in dataset.trajectories:
            for transition in trajectory.transitions:
                self.append_sample_sequentially(transition.as_dict())
            self.trajectories[-1].info = copy.deepcopy(trajectory.info)
        self.check_consistency()
        # assert(self.batch_size == dataset.batch_size)
    def __str__(self):
        printout = '\n \n '
        printout += '\n ************************ '
        printout += '\n *** Printing Dataset *** '
        printout += '\n ************************ '
        printout += '\n \n '
        printout += '\n Number of Samples    : ' + str(len(self))
        printout += '\n Dataset Trajectories : ' + str(self.nTrajectories()) + '\n'
        for idx, trajectory in enumerate(self.trajectories):
            printout += "\n >>> Trajectory id: " + str(idx) + '\n'
            printout += trajectory.__str__()
        if self._last_oar != None:
            printout += "\n !!! Found incomplete transition !!! \n"
            for key in self._last_oar:
                printout += key + '\n'
                printout += str(self._last_oar[key]) + "\n"
        printout += '\n ************************ '
        printout += '\n *** Dataset  Printed *** '
        printout += '\n ************************ '
        return printout
    
    def keep_top_fraction_of_trajectories(self, fraction: float, from_high_to_low = True):
        self.sort(from_high_to_low=from_high_to_low)
        trajectories = self.trajectories
        import math
        nTraj_to_keep = int(fraction * self.nTrajectories())
        self.__init__()
        for i in range(nTraj_to_keep):
            self.append_trajectory(trajectories[i])
        print("*** Kept ", self.nTrajectories(), " trajectories")

    def keep_bottom_fraction_of_trajectories(self, fraction: float):
        self.keep_top_fraction_of_trajectories(fraction=fraction, from_high_to_low=False)


    def max_trajectory_return(self):
        return max(self.get_all_trajectory_returns())
    
    def argmax_trajectory_return(self):
        return np.argmax(self.get_all_trajectory_returns())
    
    def min_trajectory_return(self):
        return min(self.get_all_trajectory_returns())
    
    def argmin_trajectory_return(self):
        return np.argmin(self.get_all_trajectory_returns())
    
    def sort(self, from_high_to_low):
        print("Warning: new dataset created!")
        returns = [trajectory.get_return() for trajectory in self.trajectories]  
        sorted_trajectories = sort_list(self.trajectories, returns, from_high_to_low)
        self.__init__()
        for traj in sorted_trajectories:
            self.append_trajectory(traj)
        
    # useful to set all rewards to eg -1 and encourage reaching the goal faster
    def set_all_rewards_to_value(self, value):
        for sample in self.samples:
            sample.reward = np.single(value)
    
    def scale_all_rewards_by_value(self, value):
        for sample in self.samples:
            sample.reward *= np.single(value)

    def add_value_to_all_rewards(self, value):
        for sample in self.samples:
            sample.reward += np.single(value)

    def increase_final_reward_by_value(self, value):
        for trajectory in self.trajectories:
            trajectory.transitions[-1].reward  += np.single(value)
            
    def append_eos_token_to_all_actions(self, eos_token):
        for sample in self.samples:
            sample.action += eos_token
        
    def push_all_rewards_at_the_end_of_the_trajectory(self):
        for trajectory in self.trajectories:
            trajectory.transitions[-1].reward  = np.single(trajectory.get_return())
            for transition in trajectory.transitions[:-1]:
                transition.reward = np.single(0)
            assert(- len(trajectory) == trajectory.get_return() == trajectory.transitions[-1].reward)

    def save(self, filename):
        import json 
        self.check_consistency()
        with open(filename, "w") as final:
            json.dump([sample.as_dict(as_string = True) for sample in self.samples], final)

    def load(self, filename):
        import json 
        with open(filename, "r") as final:
            data = json.load(final)
            for sample in data:
                self.append_sample_sequentially(sample)
                
    def times_was_sampled(self):
        return [sample.times_was_sampled for sample in self.samples]
    
    def keep_only_trajectories_with_exact_key_and_value(self, key, value):
        trajectories = self.trajectories
        new_dataset = TrajectoryDataset()
        for trajectory in trajectories:
            if trajectory.info[key] == value:
                new_dataset.append_trajectory(trajectory)
        return new_dataset
    
    def construct_tabular_state_action_space(self):
        self.state_space        = Counter()
        self.action_space       = Counter()
        self.state_action_space = Counter()
        for sample in self.samples:
            self.state_space.add(sample.observation)
            self.action_space.add(sample.action)
            self.state_action_space.add((sample.observation, sample.action))

    def assert_deterministic(self):
        successor_states = {}
        rewards          = {}
        for sample in self.samples:
            sa = (sample.observation, sample.action)
          
            if sa not in rewards:
                rewards[sa] = sample.reward

            else:
                assert(rewards[sa] == sample.reward)
            
            if sample.done: # end transition may be ill-defined
                continue

            if sa not in successor_states:
                successor_states[sa] = sample.next_observation
            else:
                assert(successor_states[sa] == sample.next_observation)
        
class Counter():
    def __init__(self):
        self.register = {}
    def add(self, item):
        if item not in self.register:
            self.register[item] = 1
        else:
            self.register[item] += 1
    def contains(self, item):
        return item in self.register

    def n_samples(self, item):
        return self.register[item]
    
class EmptyDataset():
    def __init__(self, length):
        self.length     = length

    def __len__(self):
        return self.length

    def __getitem__(self, idx):
        return [0]
    
def sort_list(list1, list2, from_high_to_low):
    # Sorting the List1 based on List2
    return [val for (_, val) in sorted(zip(list2, list1), key=lambda x: x[0], reverse=from_high_to_low)]