def test_replay_with_hindsight_relabel(self):
self.max_length = 8
torch.manual_seed(0)
configs = [
"hindsight_relabel_fn.her_proportion=0.8",
'hindsight_relabel_fn.achieved_goal_field="o.a"',
'hindsight_relabel_fn.desired_goal_field="o.g"',
"ReplayBuffer.postprocess_exp_fn=@hindsight_relabel_fn",
]
gin.parse_config_files_and_bindings("", configs)
replay_buffer = ReplayBuffer(data_spec=self.data_spec,
num_environments=2,
max_length=self.max_length,
keep_episodic_info=True,
step_type_field="t",
with_replacement=True)
steps = [
[
ds.StepType.FIRST, # will be overwritten
ds.StepType.MID, # idx == 1 in buffer
ds.StepType.LAST,
ds.StepType.FIRST,
ds.StepType.MID,
ds.StepType.MID,
ds.StepType.LAST,
ds.StepType.FIRST,
ds.StepType.MID # idx == 0
],
[
ds.StepType.FIRST, # will be overwritten in RingBuffer
ds.StepType.LAST, # idx == 1 in RingBuffer
ds.StepType.FIRST,
ds.StepType.MID,
ds.StepType.MID,
ds.StepType.LAST,
ds.StepType.FIRST,
ds.StepType.MID,
ds.StepType.MID # idx == 0
]
]
# insert data that will be overwritten later
for b, t in list(itertools.product(range(2), range(8))):
batch = get_batch([b], self.dim, t=steps[b][t], x=0.1 * t + b)
replay_buffer.add_batch(batch, batch.env_id)
# insert data
for b, t in list(itertools.product(range(2), range(9))):
batch = get_batch([b], self.dim, t=steps[b][t], x=0.1 * t + b)
replay_buffer.add_batch(batch, batch.env_id)
# Test padding
idx = torch.tensor([[7, 0, 0, 6, 3, 3, 3, 0], [6, 0, 5, 2, 2, 2, 0,
6]])
pos = replay_buffer._pad(idx, torch.tensor([[0] * 8, [1] * 8]))
self.assertTrue(
torch.equal(
pos,
torch.tensor([[15, 16, 16, 14, 11, 11, 11, 16],
[14, 16, 13, 10, 10, 10, 16, 14]])))
# Verify _index is built correctly.
# Note, the _index_pos 8 represents headless timesteps, which are
# outdated and not the same as the result of padding: 16.
pos = torch.tensor([[15, 8, 8, 14, 11, 11, 11, 16],
[14, 8, 13, 10, 10, 10, 16, 14]])
self.assertTrue(torch.equal(replay_buffer._indexed_pos, pos))
self.assertTrue(
torch.equal(replay_buffer._headless_indexed_pos,
torch.tensor([10, 9])))
# Save original exp for later testing.
g_orig = replay_buffer._buffer.o["g"].clone()
r_orig = replay_buffer._buffer.reward.clone()
# HER selects indices [0, 2, 3, 4] to relabel, from all 5:
# env_ids: [[0, 0], [1, 1], [0, 0], [1, 1], [0, 0]]
# pos: [[6, 7], [1, 2], [1, 2], [3, 4], [5, 6]] + 8
# selected: x x x x
# future: [ 7 2 2 4 6 ] + 8
# g [[.7,.7],[0, 0], [.2,.2],[1.4,1.4],[.6,.6]] # 0.1 * t + b with default 0
# reward: [[-1,0], [-1,-1],[-1,0], [-1,0], [-1,0]] # recomputed with default -1
env_ids = torch.tensor([0, 0, 1, 0])
dist = replay_buffer.steps_to_episode_end(
replay_buffer._pad(torch.tensor([7, 2, 4, 6]), env_ids), env_ids)
self.assertEqual(list(dist), [1, 0, 1, 0])
# Test HER relabeled experiences
res = replay_buffer.get_batch(5, 2)[0]
self.assertEqual(list(res.o["g"].shape), [5, 2])
# Test relabeling doesn't change original experience
self.assertTrue(torch.allclose(r_orig, replay_buffer._buffer.reward))
self.assertTrue(torch.allclose(g_orig, replay_buffer._buffer.o["g"]))
# test relabeled goals
g = torch.tensor([0.7, 0., .2, 1.4, .6]).unsqueeze(1).expand(5, 2)
self.assertTrue(torch.allclose(res.o["g"], g))
# test relabeled rewards
r = torch.tensor([[-1., 0.], [-1., -1.], [-1., 0.], [-1., 0.],
[-1., 0.]])
self.assertTrue(torch.allclose(res.reward, r))
def test_compute_her_future_step_distance(self, end_prob):
num_envs = 2
max_length = 100
torch.manual_seed(0)
configs = [
"hindsight_relabel_fn.her_proportion=0.8",
'hindsight_relabel_fn.achieved_goal_field="o.a"',
'hindsight_relabel_fn.desired_goal_field="o.g"',
"ReplayBuffer.postprocess_exp_fn=@hindsight_relabel_fn",
]
gin.parse_config_files_and_bindings("", configs)
replay_buffer = ReplayBuffer(data_spec=self.data_spec,
num_environments=num_envs,
max_length=max_length,
keep_episodic_info=True,
step_type_field="t")
# insert data
max_steps = 1000
# generate step_types with certain density of episode ends
steps = self.generate_step_types(num_envs,
max_steps,
end_prob=end_prob)
for t in range(max_steps):
for b in range(num_envs):
batch = get_batch([b],
self.dim,
t=steps[b * max_steps + t],
x=1. / max_steps * t + b)
replay_buffer.add_batch(batch, batch.env_id)
if t > 1:
sample_steps = min(t, max_length)
env_ids = torch.tensor([0] * sample_steps + [1] * sample_steps)
idx = torch.tensor(
list(range(sample_steps)) + list(range(sample_steps)))
gd = self.steps_to_episode_end(replay_buffer, env_ids, idx)
idx_orig = replay_buffer._indexed_pos.clone()
idx_headless_orig = replay_buffer._headless_indexed_pos.clone()
d = replay_buffer.steps_to_episode_end(
replay_buffer._pad(idx, env_ids), env_ids)
# Test distance to end computation
if not torch.equal(gd, d):
outs = [
"t: ", t, "\nenvids:\n", env_ids, "\nidx:\n", idx,
"\npos:\n",
replay_buffer._pad(idx, env_ids), "\nNot Equal: a:\n",
gd, "\nb:\n", d, "\nsteps:\n", replay_buffer._buffer.t,
"\nindexed_pos:\n", replay_buffer._indexed_pos,
"\nheadless_indexed_pos:\n",
replay_buffer._headless_indexed_pos
]
outs = [str(out) for out in outs]
assert False, "".join(outs)
# Save original exp for later testing.
g_orig = replay_buffer._buffer.o["g"].clone()
r_orig = replay_buffer._buffer.reward.clone()
# HER relabel experience
res = replay_buffer.get_batch(sample_steps, 2)[0]
self.assertEqual(list(res.o["g"].shape), [sample_steps, 2])
# Test relabeling doesn't change original experience
self.assertTrue(
torch.allclose(r_orig, replay_buffer._buffer.reward))
self.assertTrue(
torch.allclose(g_orig, replay_buffer._buffer.o["g"]))
self.assertTrue(
torch.all(idx_orig == replay_buffer._indexed_pos))
self.assertTrue(
torch.all(idx_headless_orig ==
replay_buffer._headless_indexed_pos))