Do burn-in for PPO

com.unity.ml-agents/CHANGELOG.md

 ### Bug Fixes
 #### com.unity.ml-agents / com.unity.ml-agents.extensions (C#)
 #### ml-agents / ml-agents-envs / gym-unity (Python)
-
+- Fixed an issue with LSTM when used with POCA and `sequence_length` < `time_horizon`. Also improved behavior of LSTMs slightly. (#5206)
 
 ## [1.9.0-preview] - 2021-03-17
 ### Major Changes

ml-agents/mlagents/trainers/optimizer/torch_optimizer.py

 from typing import Dict, Optional, Tuple, List
 from mlagents.torch_utils import torch
 import numpy as np
-import math
+from collections import defaultdict
 
 from mlagents.trainers.buffer import AgentBuffer, AgentBufferField
 from mlagents.trainers.trajectory import ObsUtil
         """
         num_experiences = tensor_obs[0].shape[0]
         all_next_memories = AgentBufferField()
-        # In the buffer, the 1st sequence are the ones that are padded. So if seq_len = 3 and
-        # trajectory is of length 10, the 1st sequence is [pad,pad,obs].
-        # Compute the number of elements in this padded seq.
-        leftover = num_experiences % self.policy.sequence_length
-
-        # Compute values for the potentially truncated initial sequence
-        seq_obs = []
-
-        first_seq_len = leftover if leftover > 0 else self.policy.sequence_length
-        for _obs in tensor_obs:
-            first_seq_obs = _obs[0:first_seq_len]
-            seq_obs.append(first_seq_obs)
-
-        # For the first sequence, the initial memory should be the one at the
-        # beginning of this trajectory.
-        for _ in range(first_seq_len):
-            all_next_memories.append(ModelUtils.to_numpy(initial_memory.squeeze()))
+        # When using LSTM, we need to divide the trajectory into sequences of even length. Sometimes,
+        # that division isn't even, and we must pad the leftover sequence.
+        # When it is added to the buffer, the last sequence will be padded. So if seq_len = 3 and
+        # trajectory is of length 10, the last sequence is [obs,pad,pad] once it is added to the buffer.
+        # Compute the number of elements in this sequence that will end up being padded.
+        leftover_seq_len = num_experiences % self.policy.sequence_length
-        init_values, _mem = self.critic.critic_pass(
-            seq_obs, initial_memory, sequence_length=first_seq_len
-        )
-        all_values = {
-            signal_name: [init_values[signal_name]]
-            for signal_name in init_values.keys()
-        }
-
+        all_values: Dict[str, List[np.ndarray]] = defaultdict(list)
+        _mem = initial_memory
-        for seq_num in range(
-            1, math.ceil((num_experiences) / (self.policy.sequence_length))
-        ):
+        for seq_num in range(num_experiences // (self.policy.sequence_length)):
-            start = seq_num * self.policy.sequence_length - (
-                self.policy.sequence_length - leftover
-            )
-            end = (seq_num + 1) * self.policy.sequence_length - (
-                self.policy.sequence_length - leftover
-            )
+            start = seq_num * self.policy.sequence_length
+            end = (seq_num + 1) * self.policy.sequence_length
+
             for _obs in tensor_obs:
                 seq_obs.append(_obs[start:end])
             values, _mem = self.critic.critic_pass(
                 all_values[signal_name].append(_val)
+
+        # Compute values for the potentially truncated last sequence. Note that this
+        # sequence isn't padded yet, but will be.
+        seq_obs = []
+
+        if leftover_seq_len > 0:
+            for _obs in tensor_obs:
+                last_seq_obs = _obs[-leftover_seq_len:]
+                seq_obs.append(last_seq_obs)
+
+            # For the last sequence, the initial memory should be the one at the
+            # end of this trajectory.
+            for _ in range(leftover_seq_len):
+                all_next_memories.append(ModelUtils.to_numpy(_mem.squeeze()))
+
+            last_values, _mem = self.critic.critic_pass(
+                seq_obs, _mem, sequence_length=leftover_seq_len
+            )
+            for signal_name, _val in last_values.items():
+                all_values[signal_name].append(_val)
+
         # Create one tensor per reward signal
         all_value_tensors = {
             signal_name: torch.cat(value_list, dim=0)

ml-agents/mlagents/trainers/buffer.py

                 else:
                     # We want to duplicate the last value in the array, multiplied by the padding_value.
                     padding = np.array(self[-1], dtype=np.float32) * self.padding_value
-                return [padding] * (training_length - leftover) + self[:]
+                return self[:] + [padding] * (training_length - leftover)
 
             else:
                 return self[len(self) - batch_size * training_length :]

ml-agents/mlagents/trainers/poca/optimizer_torch.py

 from typing import Dict, cast, List, Tuple, Optional
+from collections import defaultdict
-import math
 from mlagents.torch_utils import torch, default_device
 
 from mlagents.trainers.buffer import (
         num_experiences = self_obs[0].shape[0]
         all_next_value_mem = AgentBufferField()
         all_next_baseline_mem = AgentBufferField()
-        # In the buffer, the 1st sequence are the ones that are padded. So if seq_len = 3 and
-        # trajectory is of length 10, the 1st sequence is [pad,pad,obs].
+
+        # When using LSTM, we need to divide the trajectory into sequences of even length. Sometimes,
+        # that division isn't even, and we must pad the leftover sequence.
+        # In the buffer, the last sequence are the ones that are padded. So if seq_len = 3 and
+        # trajectory is of length 10, the last sequence is [obs,pad,pad].
-        leftover = num_experiences % self.policy.sequence_length
+        leftover_seq_len = num_experiences % self.policy.sequence_length
-        # Compute values for the potentially truncated initial sequence
-
-        first_seq_len = leftover if leftover > 0 else self.policy.sequence_length
-
-        self_seq_obs = []
-        groupmate_seq_obs = []
-        groupmate_seq_act = []
-        seq_obs = []
-        for _self_obs in self_obs:
-            first_seq_obs = _self_obs[0:first_seq_len]
-            seq_obs.append(first_seq_obs)
-        self_seq_obs.append(seq_obs)
-
-        for groupmate_obs, groupmate_action in zip(obs, actions):
-            seq_obs = []
-            for _obs in groupmate_obs:
-                first_seq_obs = _obs[0:first_seq_len]
-                seq_obs.append(first_seq_obs)
-            groupmate_seq_obs.append(seq_obs)
-            _act = groupmate_action.slice(0, first_seq_len)
-            groupmate_seq_act.append(_act)
-
-        # For the first sequence, the initial memory should be the one at the
-        # beginning of this trajectory.
-        for _ in range(first_seq_len):
-            all_next_value_mem.append(ModelUtils.to_numpy(init_value_mem.squeeze()))
-            all_next_baseline_mem.append(
-                ModelUtils.to_numpy(init_baseline_mem.squeeze())
-            )
-
-        all_seq_obs = self_seq_obs + groupmate_seq_obs
-        init_values, _value_mem = self.critic.critic_pass(
-            all_seq_obs, init_value_mem, sequence_length=first_seq_len
-        )
-        all_values = {
-            signal_name: [init_values[signal_name]]
-            for signal_name in init_values.keys()
-        }
-
-        groupmate_obs_and_actions = (groupmate_seq_obs, groupmate_seq_act)
-        init_baseline, _baseline_mem = self.critic.baseline(
-            self_seq_obs[0],
-            groupmate_obs_and_actions,
-            init_baseline_mem,
-            sequence_length=first_seq_len,
-        )
-        all_baseline = {
-            signal_name: [init_baseline[signal_name]]
-            for signal_name in init_baseline.keys()
-        }
+        all_values: Dict[str, List[np.ndarray]] = defaultdict(list)
+        all_baseline: Dict[str, List[np.ndarray]] = defaultdict(list)
+        _baseline_mem = init_baseline_mem
+        _value_mem = init_value_mem
-        for seq_num in range(
-            1, math.ceil((num_experiences) / (self.policy.sequence_length))
-        ):
+        for seq_num in range(num_experiences // self.policy.sequence_length):
             for _ in range(self.policy.sequence_length):
                 all_next_value_mem.append(ModelUtils.to_numpy(_value_mem.squeeze()))
                 all_next_baseline_mem.append(
-            start = seq_num * self.policy.sequence_length - (
-                self.policy.sequence_length - leftover
-            )
-            end = (seq_num + 1) * self.policy.sequence_length - (
-                self.policy.sequence_length - leftover
-            )
+            start = seq_num * self.policy.sequence_length
+            end = (seq_num + 1) * self.policy.sequence_length
 
             self_seq_obs = []
             groupmate_seq_obs = []
-                seq_obs.append(_obs[start:end])
+                seq_obs.append(_self_obs[start:end])
-            for groupmate_obs, team_action in zip(obs, actions):
+            for groupmate_obs, groupmate_action in zip(obs, actions):
-                for (_obs,) in groupmate_obs:
-                    first_seq_obs = _obs[start:end]
-                    seq_obs.append(first_seq_obs)
+                for _obs in groupmate_obs:
+                    sliced_seq_obs = _obs[start:end]
+                    seq_obs.append(sliced_seq_obs)
-                _act = team_action.slice(start, end)
+                _act = groupmate_action.slice(start, end)
                 groupmate_seq_act.append(_act)
 
             all_seq_obs = self_seq_obs + groupmate_seq_obs
-            all_values = {
-                signal_name: [init_values[signal_name]] for signal_name in values.keys()
-            }
+            for signal_name, _val in values.items():
+                all_values[signal_name].append(_val)
 
             groupmate_obs_and_actions = (groupmate_seq_obs, groupmate_seq_act)
             baselines, _baseline_mem = self.critic.baseline(
-                sequence_length=first_seq_len,
+                sequence_length=self.policy.sequence_length,
-            all_baseline = {
-                signal_name: [baselines[signal_name]]
-                for signal_name in baselines.keys()
-            }
+            for signal_name, _val in baselines.items():
+                all_baseline[signal_name].append(_val)
+
+        # Compute values for the potentially truncated initial sequence
+        if leftover_seq_len > 0:
+            self_seq_obs = []
+            groupmate_seq_obs = []
+            groupmate_seq_act = []
+            seq_obs = []
+            for _self_obs in self_obs:
+                last_seq_obs = _self_obs[-leftover_seq_len:]
+                seq_obs.append(last_seq_obs)
+            self_seq_obs.append(seq_obs)
+
+            for groupmate_obs, groupmate_action in zip(obs, actions):
+                seq_obs = []
+                for _obs in groupmate_obs:
+                    last_seq_obs = _obs[-leftover_seq_len:]
+                    seq_obs.append(last_seq_obs)
+                groupmate_seq_obs.append(seq_obs)
+                _act = groupmate_action.slice(len(_obs) - leftover_seq_len, len(_obs))
+                groupmate_seq_act.append(_act)
+
+            # For the last sequence, the initial memory should be the one at the
+            # beginning of this trajectory.
+            seq_obs = []
+            for _ in range(leftover_seq_len):
+                all_next_value_mem.append(ModelUtils.to_numpy(_value_mem.squeeze()))
+                all_next_baseline_mem.append(
+                    ModelUtils.to_numpy(_baseline_mem.squeeze())
+                )
+
+            all_seq_obs = self_seq_obs + groupmate_seq_obs
+            last_values, _value_mem = self.critic.critic_pass(
+                all_seq_obs, _value_mem, sequence_length=leftover_seq_len
+            )
+            for signal_name, _val in last_values.items():
+                all_values[signal_name].append(_val)
+            groupmate_obs_and_actions = (groupmate_seq_obs, groupmate_seq_act)
+            last_baseline, _baseline_mem = self.critic.baseline(
+                self_seq_obs[0],
+                groupmate_obs_and_actions,
+                _baseline_mem,
+                sequence_length=leftover_seq_len,
+            )
+            for signal_name, _val in last_baseline.items():
+                all_baseline[signal_name].append(_val)
         # Create one tensor per reward signal
         all_value_tensors = {
             signal_name: torch.cat(value_list, dim=0)

ml-agents/mlagents/trainers/tests/torch/test_ppo.py

     optimizer = create_test_ppo_optimizer(
         dummy_config, use_rnn=rnn, use_discrete=discrete, use_visual=visual
     )
-    time_horizon = 15
+    # Time horizon is longer than sequence length, make sure to test
+    # process trajectory on multiple sequences in trajectory + some padding
+    time_horizon = 30
     trajectory = make_fake_trajectory(
         length=time_horizon,
         observation_specs=optimizer.policy.behavior_spec.observation_specs,
 
     for key, val in run_out.items():
         assert type(key) is str
-        assert len(val) == 15
+        assert len(val) == time_horizon
-        assert len(all_memories) == 15
+        assert len(all_memories) == time_horizon
 
     run_out, final_value_out, _ = optimizer.get_trajectory_value_estimates(
         trajectory.to_agentbuffer(), trajectory.next_obs, done=True

ml-agents/mlagents/trainers/tests/torch/test_poca.py

     }
 
     trainer_settings.network_settings.memory = (
-        NetworkSettings.MemorySettings(sequence_length=16, memory_size=10)
+        NetworkSettings.MemorySettings(sequence_length=8, memory_size=10)
         if use_rnn
         else None
     )
     optimizer = create_test_poca_optimizer(
         dummy_config, use_rnn=rnn, use_discrete=discrete, use_visual=visual
     )
-    time_horizon = 15
+    time_horizon = 30
     trajectory = make_fake_trajectory(
         length=time_horizon,
         observation_specs=optimizer.policy.behavior_spec.observation_specs,
     )
     for key, val in value_estimates.items():
         assert type(key) is str
-        assert len(val) == 15
+        assert len(val) == time_horizon
-        assert len(val) == 15
+        assert len(val) == time_horizon
-        assert len(value_memories) == 15
-        assert len(baseline_memories) == 15
+        assert len(value_memories) == time_horizon
+        assert len(baseline_memories) == time_horizon
 
     (
         value_estimates,

ml-agents/mlagents/trainers/ppo/optimizer_torch.py

         old_log_probs = ActionLogProbs.from_buffer(batch).flatten()
         log_probs = log_probs.flatten()
         loss_masks = ModelUtils.list_to_tensor(batch[BufferKey.MASKS], dtype=torch.bool)
+        seq_len = self.policy.sequence_length
+        if seq_len > 1:
+            # Do burn-in
+            _burn_in_percent = 0.2
+            burn_in_mask = torch.ones_like(loss_masks)
+            burn_in_amt = int(seq_len * _burn_in_percent)
+            for i in range(batch.num_experiences // seq_len):
+                burn_in_mask[seq_len * i : seq_len * i + burn_in_amt] = 0.0
+            loss_masks = loss_masks * burn_in_mask
         value_loss = ModelUtils.trust_region_value_loss(
             values, old_values, returns, decay_eps, loss_masks
         )