def _train_ppo_epoch(self, full_input):
total_obs = self.n_steps * self.envs.n_envs
shuffle_idx = np.random.permutation(total_obs)
batches = dict_of_lists_to_list_of_dicst({
k: np.split(v[shuffle_idx], total_obs // self.ppo_par.batch_size)
for k, v in full_input.items()
})
for b in batches:
self.agent.train(b)
def _train_ppo_recurrent_epoch(self, full_input, rnn_state):
# HE SHUFFLES SO BE CAREFUL!!! RECHECK IT: rnn_state might need to get in the full_input
total_obs = self.n_steps * self.envs.num_envs
shuffle_idx = np.random.permutation(total_obs)
batches = dict_of_lists_to_list_of_dicst({
k: np.split(v[shuffle_idx], total_obs // self.ppo_par.batch_size)
for k, v in full_input.items()
})
for b in batches:
self.agent.train_recurrent(b, rnn_state) # IMPORTANT : όταν κανεις training δεν χρειαζεσαι την rnn_State, ξεκινας απο το 0 και αθτη παιρνη την μορφή πουπρεπει να εχει
def _train_ppo_epoch(self, full_input):
total_obs = self.n_steps * self.envs.num_envs
shuffle_idx = np.random.permutation(total_obs)
batches = dict_of_lists_to_list_of_dicst({
k: np.split(v[shuffle_idx], total_obs // self.ppo_par.batch_size)
for k, v in full_input.items()
})
if self.policy_type == MetaPolicy: # We take out the if from the loop so you choose trainer BEFORE getting into the batch loop
for b in batches:
self.agent.train_recurrent(b)
else:
for b in batches:
self.agent.train(b)
def test_dict_list_transpose(self):
x = {
"a": [1, 2, 3],
"b": [np.array([5, 6]), np.array([7, 8]), np.array([90, 100])]
}
result = dict_of_lists_to_list_of_dicst(x)
expected = [
{'a': 1, 'b': np.array([5, 6])},
{'a': 2, 'b': np.array([7, 8])},
{'a': 3, 'b': np.array([90, 100])}
]
assert len(result) == len(expected)
for r, e in zip(expected, result):
assert r.keys() == e.keys()
assert r["a"] == e["a"]
self.assertAllEqual(r["b"], e["b"])