custom_torch_ppg.py 17.9 KB
Newer Older
Dipam Chakraborty's avatar
Dipam Chakraborty committed
1
2
3
4
5
6
7
8
9
10
11
from ray.rllib.policy.torch_policy import TorchPolicy
import numpy as np
from ray.rllib.utils.torch_ops import convert_to_non_torch_type, convert_to_torch_tensor
from ray.rllib.utils import try_import_torch
from ray.rllib.models import ModelCatalog
from ray.rllib.utils.annotations import override
from collections import deque
from .utils import *
import time

torch, nn = try_import_torch()
12
import torch.distributions as td
Dipam Chakraborty's avatar
Dipam Chakraborty committed
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

class CustomTorchPolicy(TorchPolicy):
    """Example of a random policy
    If you are using tensorflow/pytorch to build custom policies,
    you might find `build_tf_policy` and `build_torch_policy` to
    be useful.
    Adopted from examples from https://docs.ray.io/en/master/rllib-concepts.html
    """

    def __init__(self, observation_space, action_space, config):
        self.config = config

        self.device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
        dist_class, logit_dim = ModelCatalog.get_action_dist(
            action_space, self.config["model"], framework="torch")
        self.model = ModelCatalog.get_model_v2(
                        obs_space=observation_space,
                        action_space=action_space,
                        num_outputs=logit_dim,
                        model_config=self.config["model"],
                        framework="torch",
                        device=self.device,
                     )

        TorchPolicy.__init__(
            self,
            observation_space=observation_space,
            action_space=action_space,
            config=config,
            model=self.model,
            loss=None,
            action_distribution_class=dist_class,
        )

        self.framework = "torch"
Dipam Chakraborty's avatar
Dipam Chakraborty committed
48
49
        self.optimizer = torch.optim.Adam(self.model.parameters(), lr=5e-4)
        self.aux_optimizer = torch.optim.Adam(self.model.parameters(), lr=5e-4)
Dipam Chakraborty's avatar
Dipam Chakraborty committed
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
        self.max_reward = self.config['env_config']['return_max']
        self.rewnorm = RewardNormalizer(cliprew=self.max_reward) ## TODO: Might need to go to custom state
        self.reward_deque = deque(maxlen=100)
        self.best_reward = -np.inf
        self.best_weights = None
        self.time_elapsed = 0
        self.batch_end_time = time.time()
        self.timesteps_total = 0
        self.best_rew_tsteps = 0
        
        nw = self.config['num_workers'] if self.config['num_workers'] > 0 else 1
        self.nbatch = nw * self.config['num_envs_per_worker'] * self.config['rollout_fragment_length']
        self.actual_batch_size = self.nbatch // self.config['updates_per_batch']
        self.accumulate_train_batches = int(np.ceil( self.actual_batch_size / self.config['max_minibatch_size'] ))
        self.mem_limited_batch_size = self.actual_batch_size // self.accumulate_train_batches
        if self.nbatch % self.actual_batch_size != 0 or self.nbatch % self.mem_limited_batch_size != 0:
            print("#################################################")
            print("WARNING: MEMORY LIMITED BATCHING NOT SET PROPERLY")
            print("#################################################")
        self.retune_selector = RetuneSelector(self.nbatch, observation_space, action_space, 
                                              skips = self.config['retune_skips'], 
                                              replay_size = self.config['retune_replay_size'], 
                                              num_retunes = self.config['num_retunes'])
        
74
75
76
        self.exp_replay = np.zeros((self.retune_selector.replay_size, *observation_space.shape), dtype=np.uint8)
        self.vtarg_replay = np.zeros((self.retune_selector.replay_size), dtype=np.float32)
        self.save_success = 0
Dipam Chakraborty's avatar
Dipam Chakraborty committed
77
78
79
80
81
82
83
84
85
86
87
        self.target_timesteps = 8_000_000
        self.buffer_time = 20 # TODO: Could try to do a median or mean time step check instead
        self.max_time = 7200
        self.maxrewep_lenbuf = deque(maxlen=100)
        self.gamma = self.config['gamma']
        self.adaptive_discount_tuner = AdaptiveDiscountTuner(self.gamma, momentum=0.98, eplenmult=3)
        
        self.lr = config['lr']
        self.ent_coef = config['entropy_coeff']
        
        self.last_dones = np.zeros((nw * self.config['num_envs_per_worker'],))
88
        self.make_distr = dist_build(action_space)
Dipam Chakraborty's avatar
Dipam Chakraborty committed
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
        
    def to_tensor(self, arr):
        return torch.from_numpy(arr).to(self.device)
        
    @override(TorchPolicy)
    def learn_on_batch(self, samples):
        """Fused compute gradients and apply gradients call.
        Either this or the combination of compute/apply grads must be
        implemented by subclasses.
        Returns:
            grad_info: dictionary of extra metadata from compute_gradients().
        Examples:
            >>> batch = ev.sample()
            >>> ev.learn_on_batch(samples)
        Reference: https://github.com/ray-project/ray/blob/master/rllib/policy/policy.py#L279-L316
        """
        ## Config data values
        nbatch = self.nbatch
        nbatch_train = self.mem_limited_batch_size 
        gamma, lam = self.gamma, self.config['lambda']
        nsteps = self.config['rollout_fragment_length']
        nenvs = nbatch//nsteps
        ts = (nenvs, nsteps)
        mb_dones = unroll(samples['dones'], ts)
        
        ## Reward Normalization - No reward norm works well for many envs
        if self.config['standardize_rewards']:
            mb_origrewards = unroll(samples['rewards'], ts)
            mb_rewards =  np.zeros_like(mb_origrewards)
            mb_rewards[0] = self.rewnorm.normalize(mb_origrewards[0], self.last_dones)
            for ii in range(1, nsteps):
                mb_rewards[ii] = self.rewnorm.normalize(mb_origrewards[ii], mb_dones[ii-1])
            self.last_dones = mb_dones[-1]
        else:
            mb_rewards = unroll(samples['rewards'], ts)
        
        should_skip_train_step = self.best_reward_model_select(samples)
        if should_skip_train_step:
            self.update_batch_time()
            return {} # Not doing last optimization step - This is intentional due to noisy gradients
          
        obs = samples['obs']

        ## Value prediction
        next_obs = unroll(samples['new_obs'], ts)[-1]
        last_values, _ = self.model.vf_pi(next_obs, ret_numpy=True, no_grad=True, to_torch=True)
        values = np.empty((nbatch,), dtype=np.float32)
        for start in range(0, nbatch, nbatch_train): # Causes OOM up if trying to do all at once
            end = start + nbatch_train
            values[start:end], _ = self.model.vf_pi(samples['obs'][start:end], ret_numpy=True, no_grad=True, to_torch=True)
        
        ## GAE
        mb_values = unroll(values, ts)
        mb_returns = np.zeros_like(mb_rewards)
        mb_advs = np.zeros_like(mb_rewards)
        lastgaelam = 0
        for t in reversed(range(nsteps)):
            if t == nsteps - 1:
                nextvalues = last_values
            else:
                nextvalues = mb_values[t+1]
            nextnonterminal = 1.0 - mb_dones[t]
            delta = mb_rewards[t] + gamma * nextvalues * nextnonterminal - mb_values[t]
            mb_advs[t] = lastgaelam = delta + gamma * lam * nextnonterminal * lastgaelam
        mb_returns = mb_advs + mb_values
        
        ## Data from config
        cliprange, vfcliprange = self.config['clip_param'], self.config['vf_clip_param']
        lrnow = self.lr
        max_grad_norm = self.config['grad_clip']
        ent_coef, vf_coef = self.ent_coef, self.config['vf_loss_coeff']
        
        neglogpacs = -samples['action_logp'] ## np.isclose seems to be True always, otherwise compute again if needed
        noptepochs = self.config['num_sgd_iter']
        actions = samples['actions']
        returns = roll(mb_returns)
        
166
167
168
        advs = returns - values
        normalized_advs = (advs - np.mean(advs)) / (np.std(advs) + 1e-8) 
        
Dipam Chakraborty's avatar
Dipam Chakraborty committed
169
170
171
172
173
174
175
176
177
178
179
180
181
182
        ## Train multiple epochs
        optim_count = 0
        inds = np.arange(nbatch)
        for _ in range(noptepochs):
            np.random.shuffle(inds)
            for start in range(0, nbatch, nbatch_train):
                end = start + nbatch_train
                mbinds = inds[start:end]
                slices = (self.to_tensor(arr[mbinds]) for arr in (obs, returns, actions, values, neglogpacs, normalized_advs))
                optim_count += 1
                apply_grad = (optim_count % self.accumulate_train_batches) == 0
                self._batch_train(apply_grad, self.accumulate_train_batches,
                                  lrnow, cliprange, vfcliprange, max_grad_norm, ent_coef, vf_coef, *slices)

Dipam Chakraborty's avatar
Dipam Chakraborty committed
183
        ## Distill with aux head
184
        should_retune = self.retune_selector.update(obs, returns, self.exp_replay, self.vtarg_replay)
Dipam Chakraborty's avatar
Dipam Chakraborty committed
185
186
187
188
189
        if should_retune:
            self.aux_train()
            self.update_batch_time()
            return {}
        
Dipam Chakraborty's avatar
Dipam Chakraborty committed
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
        self.update_gamma(samples)
        self.update_lr()
        self.update_ent_coef()
            
        self.update_batch_time()
        return {}
    
    def update_batch_time(self):
        self.time_elapsed += time.time() - self.batch_end_time
        self.batch_end_time = time.time()
        
    def _batch_train(self, apply_grad, num_accumulate, 
                     lr, cliprange, vfcliprange, max_grad_norm,
                     ent_coef, vf_coef,
                     obs, returns, actions, values, neglogpac_old, advs):
        
        for g in self.optimizer.param_groups:
            g['lr'] = lr
        vpred, pi_logits = self.model.vf_pi(obs, ret_numpy=False, no_grad=False, to_torch=False)
209
210
211
        pd = self.make_distr(pi_logits)
        neglogpac = -pd.log_prob(actions[...,None]).squeeze(1)
        entropy = torch.mean(pd.entropy())
Dipam Chakraborty's avatar
Dipam Chakraborty committed
212

Dipam Chakraborty's avatar
Dipam Chakraborty committed
213
        vf_loss = .5 * torch.mean(torch.pow((vpred - returns), 2))
Dipam Chakraborty's avatar
Dipam Chakraborty committed
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229

        ratio = torch.exp(neglogpac_old - neglogpac)
        pg_losses1 = -advs * ratio
        pg_losses2 = -advs * torch.clamp(ratio, 1-cliprange, 1+cliprange)
        pg_loss = torch.mean(torch.max(pg_losses1, pg_losses2))

        loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
        
        loss = loss / num_accumulate

        loss.backward()
        if apply_grad:
            self.optimizer.step()
            self.optimizer.zero_grad()

        
Dipam Chakraborty's avatar
Dipam Chakraborty committed
230
231
232
    def aux_train(self):
        for g in self.aux_optimizer.param_groups:
            g['lr'] = self.lr
Dipam Chakraborty's avatar
Dipam Chakraborty committed
233
        nbatch_train = self.mem_limited_batch_size 
Dipam Chakraborty's avatar
Dipam Chakraborty committed
234
        aux_nbatch_train = self.config['aux_mbsize'] 
Dipam Chakraborty's avatar
Dipam Chakraborty committed
235
236
237
238
239
240
241
        retune_epochs = self.config['retune_epochs']
        replay_size = self.retune_selector.replay_size
        replay_pi = np.empty((replay_size, self.retune_selector.ac_space.n), dtype=np.float32)

        # Store current value function and policy logits
        for start in range(0, replay_size, nbatch_train):
            end = start + nbatch_train
242
            replay_batch = self.exp_replay[start:end]
Dipam Chakraborty's avatar
Dipam Chakraborty committed
243
244
            _, replay_pi[start:end] = self.model.vf_pi(replay_batch, 
                                                       ret_numpy=True, no_grad=True, to_torch=True)
Dipam Chakraborty's avatar
Dipam Chakraborty committed
245
246
247
        
        optim_count = 0
        # Tune vf and pi heads to older predictions with augmented observations
248
        inds = np.arange(len(self.exp_replay))
Dipam Chakraborty's avatar
Dipam Chakraborty committed
249
250
        for ep in range(retune_epochs):
            np.random.shuffle(inds)
Dipam Chakraborty's avatar
Dipam Chakraborty committed
251
252
            for start in range(0, replay_size, aux_nbatch_train):
                end = start + aux_nbatch_train
Dipam Chakraborty's avatar
Dipam Chakraborty committed
253
254
                mbinds = inds[start:end]
                optim_count += 1
255
256
                slices = [self.exp_replay[mbinds], 
                          self.to_tensor(self.vtarg_replay[mbinds]), 
Dipam Chakraborty's avatar
Dipam Chakraborty committed
257
                          self.to_tensor(replay_pi[mbinds])]
Dipam Chakraborty's avatar
Dipam Chakraborty committed
258
                self.tune_policy(*slices)
Dipam Chakraborty's avatar
Dipam Chakraborty committed
259
260
261

        self.retune_selector.retune_done()
 
Dipam Chakraborty's avatar
Dipam Chakraborty committed
262
    def tune_policy(self, obs, target_vf, target_pi):
263
264
265
266
267
268
269
270
        obs_aug = np.empty(obs.shape, obs.dtype)
        aug_idx = np.random.randint(6, size=len(obs))
        obs_aug[aug_idx == 0] = pad_and_random_crop(obs[aug_idx == 0], 64, 10)
        obs_aug[aug_idx == 1] = random_cutout_color(obs[aug_idx == 1], 10, 30)
        obs_aug[aug_idx >= 2] = obs[aug_idx >= 2]
        obs_aug = self.to_tensor(obs_aug)
        
        vpred, pi_logits = self.model.vf_pi(obs_aug, ret_numpy=False, no_grad=False, to_torch=False)
Dipam Chakraborty's avatar
Dipam Chakraborty committed
271
        aux_vpred = self.model.aux_value_function()
Dipam Chakraborty's avatar
Dipam Chakraborty committed
272
        vf_loss = .5 * torch.mean(torch.pow(vpred - target_vf, 2))
Dipam Chakraborty's avatar
Dipam Chakraborty committed
273
        aux_loss = .5 * torch.mean(torch.pow(aux_vpred - target_vf, 2))
Dipam Chakraborty's avatar
Dipam Chakraborty committed
274
        
275
276
277
278
        target_pd = self.make_distr(target_pi)
        pd = self.make_distr(pi_logits)
        pi_loss = td.kl_divergence(target_pd, pd).mean()
        
Dipam Chakraborty's avatar
Dipam Chakraborty committed
279
        loss = vf_loss + pi_loss + aux_loss
Dipam Chakraborty's avatar
Dipam Chakraborty committed
280
281
        
        loss.backward()
Dipam Chakraborty's avatar
Dipam Chakraborty committed
282
283
        self.aux_optimizer.step()
        self.aux_optimizer.zero_grad()
Dipam Chakraborty's avatar
Dipam Chakraborty committed
284
285
        
    def best_reward_model_select(self, samples):
286
        self.timesteps_total += len(samples['dones'])
Dipam Chakraborty's avatar
Dipam Chakraborty committed
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
330
331
332
333
334
335
336
337
        
        ## Best reward model selection
        eprews = [info['episode']['r'] for info in samples['infos'] if 'episode' in info]
        self.reward_deque.extend(eprews)
        mean_reward = safe_mean(eprews) if len(eprews) >= 100 else safe_mean(self.reward_deque)
        if self.best_reward < mean_reward:
            self.best_reward = mean_reward
            self.best_weights = self.get_weights()["current_weights"]
            self.best_rew_tsteps = self.timesteps_total
           
        if self.timesteps_total > self.target_timesteps or (self.time_elapsed + self.buffer_time) > self.max_time:
            if self.best_weights is not None:
                self.set_model_weights(self.best_weights)
                return True
            
        return False
    
    def update_lr(self):
        if self.config['lr_schedule'] == 'linear':
            self.lr = linear_schedule(initial_val=self.config['lr'],
                                      final_val=self.config['final_lr'],
                                      current_steps=self.timesteps_total,
                                      total_steps=self.target_timesteps)
            
        elif self.config['lr_schedule'] == 'exponential':
            self.lr = 0.997 * self.lr 

    
    def update_ent_coef(self):
        if self.config['entropy_schedule']:
            self.ent_coef = linear_schedule(initial_val=self.config['entropy_coeff'], 
                                            final_val=self.config['final_entropy_coeff'], 
                                            current_steps=self.timesteps_total, 
                                            total_steps=self.target_timesteps)
    
    def update_gamma(self, samples):
        if self.config['adaptive_gamma']:
            epinfobuf = [info['episode'] for info in samples['infos'] if 'episode' in info]
            self.maxrewep_lenbuf.extend([epinfo['l'] for epinfo in epinfobuf if epinfo['r'] >= self.max_reward])
            sorted_nth = lambda buf, n: np.nan if len(buf) < 100 else sorted(self.maxrewep_lenbuf.copy())[n]
            target_horizon = sorted_nth(self.maxrewep_lenbuf, 80)
            self.gamma = self.adaptive_discount_tuner.update(target_horizon)

        
    def get_custom_state_vars(self):
        return {
            "time_elapsed": self.time_elapsed,
            "timesteps_total": self.timesteps_total,
            "best_weights": self.best_weights,
            "reward_deque": self.reward_deque,
            "batch_end_time": self.batch_end_time,
338
            "retune_selector": self.retune_selector,
Dipam Chakraborty's avatar
Dipam Chakraborty committed
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
            "gamma": self.gamma,
            "maxrewep_lenbuf": self.maxrewep_lenbuf,
            "lr": self.lr,
            "ent_coef": self.ent_coef,
            "rewnorm": self.rewnorm,
            "best_rew_tsteps": self.best_rew_tsteps,
            "best_reward": self.best_reward,
            "last_dones": self.last_dones,
        }
    
    def set_custom_state_vars(self, custom_state_vars):
        self.time_elapsed = custom_state_vars["time_elapsed"]
        self.timesteps_total = custom_state_vars["timesteps_total"]
        self.best_weights = custom_state_vars["best_weights"]
        self.reward_deque = custom_state_vars["reward_deque"]
        self.batch_end_time = custom_state_vars["batch_end_time"]
355
        self.retune_selector = custom_state_vars["retune_selector"]
Dipam Chakraborty's avatar
Dipam Chakraborty committed
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
        self.gamma = self.adaptive_discount_tuner.gamma = custom_state_vars["gamma"]
        self.maxrewep_lenbuf = custom_state_vars["maxrewep_lenbuf"]
        self.lr =custom_state_vars["lr"]
        self.ent_coef = custom_state_vars["ent_coef"]
        self.rewnorm = custom_state_vars["rewnorm"]
        self.best_rew_tsteps = custom_state_vars["best_rew_tsteps"]
        self.best_reward = custom_state_vars["best_reward"]
        self.last_dones = custom_state_vars["last_dones"]
    
    @override(TorchPolicy)
    def get_weights(self):
        weights = {}
        weights["current_weights"] = {
            k: v.cpu().detach().numpy()
            for k, v in self.model.state_dict().items()
        }
372
373
374
375
376
377
378
379
380
#         weights["optimizer_state"] = {
#             k: v
#             for k, v in self.optimizer.state_dict().items()
#         }
#         weights["aux_optimizer_state"] = {
#             k: v
#             for k, v in self.aux_optimizer.state_dict().items()
#         }
#         weights["custom_state_vars"] = self.get_custom_state_vars()
Dipam Chakraborty's avatar
Dipam Chakraborty committed
381
382
383
384
385
386
        return weights
        
    
    @override(TorchPolicy)
    def set_weights(self, weights):
        self.set_model_weights(weights["current_weights"])
387
388
389
#         self.set_optimizer_state(weights["optimizer_state"])
#         self.set_aux_optimizer_state(weights["aux_optimizer_state"])
#         self.set_custom_state_vars(weights["custom_state_vars"])
Dipam Chakraborty's avatar
Dipam Chakraborty committed
390
        
Dipam Chakraborty's avatar
Dipam Chakraborty committed
391
392
393
394
    def set_aux_optimizer_state(self, aux_optimizer_state):
        aux_optimizer_state = convert_to_torch_tensor(aux_optimizer_state, device=self.device)
        self.aux_optimizer.load_state_dict(aux_optimizer_state)
        
Dipam Chakraborty's avatar
Dipam Chakraborty committed
395
396
397
398
399
400
401
    def set_optimizer_state(self, optimizer_state):
        optimizer_state = convert_to_torch_tensor(optimizer_state, device=self.device)
        self.optimizer.load_state_dict(optimizer_state)
        
    def set_model_weights(self, model_weights):
        model_weights = convert_to_torch_tensor(model_weights, device=self.device)
        self.model.load_state_dict(model_weights)