fix(wyh): add test for rl_utils ppo and td (opendilab#89)

* fix(wyh):test rl_utils code

* fix(wyh):modify rl utils bug ppo adv batch B,A

* fix(wyh):style

* fix(wyh):fix bug

ding/rl_utils/tests/test_ppo.py

@@ -42,3 +42,51 @@ def test_ppo(use_value_clip, dual_clip, weight):
     total_loss.backward()
     assert isinstance(logit_new.grad, torch.Tensor)
     assert isinstance(value_new.grad, torch.Tensor)
+
+
+@pytest.mark.unittest
+def test_mappo():
+    B, A, N = 4, 8, 32
+    logit_new = torch.randn(B, A, N).requires_grad_(True)
+    logit_old = logit_new + torch.rand_like(logit_new) * 0.1
+    action = torch.randint(0, N, size=(B, A))
+    value_new = torch.randn(B, A).requires_grad_(True)
+    value_old = value_new + torch.rand_like(value_new) * 0.1
+    adv = torch.rand(B, A)
+    return_ = torch.randn(B, A) * 2
+    data = ppo_data(logit_new, logit_old, action, value_new, value_old, adv, return_, None)
+    loss, info = ppo_error(data)
+    assert all([l.shape == tuple() for l in loss])
+    assert all([np.isscalar(i) for i in info])
+    assert logit_new.grad is None
+    assert value_new.grad is None
+    total_loss = sum(loss)
+    total_loss.backward()
+    assert isinstance(logit_new.grad, torch.Tensor)
+    assert isinstance(value_new.grad, torch.Tensor)
+
+
+@pytest.mark.unittest
+@pytest.mark.parametrize('use_value_clip, dual_clip, weight', args)
+def test_ppo_error_continous(use_value_clip, dual_clip, weight):
+    B, N = 4, 6
+    mu_sigma_new = [torch.randn(B, N).requires_grad_(True), torch.randn(B, N).requires_grad_(True)]
+    mu_sigma_old = [
+        mu_sigma_new[0] + torch.rand_like(mu_sigma_new[0]) * 0.1,
+        mu_sigma_new[1] + torch.rand_like(mu_sigma_new[1]) * 0.1
+    ]
+    action = torch.rand(B, N)
+    value_new = torch.randn(B).requires_grad_(True)
+    value_old = value_new + torch.rand_like(value_new) * 0.1
+    adv = torch.rand(B)
+    return_ = torch.randn(B) * 2
+    data = ppo_data(mu_sigma_new, mu_sigma_old, action, value_new, value_old, adv, return_, weight)
+    loss, info = ppo_error_continuous(data, use_value_clip=use_value_clip, dual_clip=dual_clip)
+    assert all([l.shape == tuple() for l in loss])
+    assert all([np.isscalar(i) for i in info])
+    assert mu_sigma_new[0].grad is None
+    assert value_new.grad is None
+    total_loss = sum(loss)
+    total_loss.backward()
+    assert isinstance(mu_sigma_new[0].grad, torch.Tensor)
+    assert isinstance(value_new.grad, torch.Tensor)

ding/rl_utils/tests/test_td.py

@@ -2,7 +2,9 @@
 import torch
 from ding.rl_utils import q_nstep_td_data, q_nstep_td_error, q_1step_td_data, q_1step_td_error, td_lambda_data,\
     td_lambda_error, q_nstep_td_error_with_rescale, dist_1step_td_data, dist_1step_td_error, dist_nstep_td_data, \
-    dist_nstep_td_error, v_1step_td_data, v_1step_td_error, v_nstep_td_data, v_nstep_td_error
+    dist_nstep_td_error, v_1step_td_data, v_1step_td_error, v_nstep_td_data, v_nstep_td_error, q_nstep_sql_td_error, \
+    iqn_nstep_td_data, iqn_nstep_td_error, qrdqn_nstep_td_data, qrdqn_nstep_td_error
+from ding.rl_utils.td import shape_fn_dntd, shape_fn_qntd, shape_fn_td_lambda, shape_fn_qntd_rescale
 
 
 @pytest.mark.unittest
@@ -23,7 +25,13 @@ def test_q_nstep_td():
         assert q.grad is None
         loss.backward()
         assert isinstance(q.grad, torch.Tensor)
-        print(loss)
+        data = q_nstep_td_data(q, next_q, action, next_action, reward, done, None)
+        loss, td_error_per_sample = q_nstep_td_error(data, 0.95, nstep=nstep, cum_reward=True)
+        value_gamma = torch.tensor(0.9)
+        data = q_nstep_td_data(q, next_q, action, next_action, reward, done, None)
+        loss, td_error_per_sample = q_nstep_td_error(data, 0.95, nstep=nstep, cum_reward=True, value_gamma=value_gamma)
+        loss.backward()
+        assert isinstance(q.grad, torch.Tensor)
 
 
 @pytest.mark.unittest
@@ -84,6 +92,13 @@ def test_dist_nstep_td():
     assert dist.grad is None
     loss.backward()
     assert isinstance(dist.grad, torch.Tensor)
+    weight = torch.tensor([0.9])
+    value_gamma = torch.tensor(0.9)
+    data = dist_nstep_td_data(dist, next_n_dist, action, next_action, reward, done, weight)
+    loss, _ = dist_nstep_td_error(data, 0.95, v_min, v_max, n_atom, nstep, value_gamma)
+    assert loss.shape == ()
+    loss.backward()
+    assert isinstance(dist.grad, torch.Tensor)
 
 
 @pytest.mark.unittest
@@ -106,6 +121,29 @@ def test_q_nstep_td_with_rescale():
         print(loss)
 
 
+@pytest.mark.unittest
+def test_qrdqn_nstep_td():
+    batch_size = 4
+    action_dim = 3
+    tau = 3
+    next_q = torch.randn(batch_size, action_dim, tau)
+    done = torch.randn(batch_size)
+    action = torch.randint(0, action_dim, size=(batch_size, ))
+    next_action = torch.randint(0, action_dim, size=(batch_size, ))
+    for nstep in range(1, 10):
+        q = torch.randn(batch_size, action_dim, tau).requires_grad_(True)
+        reward = torch.rand(nstep, batch_size)
+        data = qrdqn_nstep_td_data(q, next_q, action, next_action, reward, done, tau, None)
+        loss, td_error_per_sample = qrdqn_nstep_td_error(data, 0.95, nstep=nstep)
+        assert td_error_per_sample.shape == (batch_size, )
+        assert loss.shape == ()
+        assert q.grad is None
+        loss.backward()
+        assert isinstance(q.grad, torch.Tensor)
+        loss, td_error_per_sample = qrdqn_nstep_td_error(data, 0.95, nstep=nstep, value_gamma=torch.tensor(0.9))
+        assert td_error_per_sample.shape == (batch_size, )
+
+
 @pytest.mark.unittest
 def test_dist_1step_compatible():
     batch_size = 4
@@ -153,6 +191,8 @@ def test_v_1step_td():
     assert isinstance(v.grad, torch.Tensor)
     data = v_1step_td_data(v, next_v, reward, None, None)
     loss, td_error_per_sample = v_1step_td_error(data, 0.99)
+    loss.backward()
+    assert isinstance(v.grad, torch.Tensor)
 
 
 @pytest.mark.unittest
@@ -168,5 +208,143 @@ def test_v_nstep_td():
     assert v.grad is None
     loss.backward()
     assert isinstance(v.grad, torch.Tensor)
-    data = v_nstep_td_data(v, next_v, reward, done, 0.9, 0.99)
+    data = v_nstep_td_data(v, next_v, reward, done, None, 0.99)
     loss, td_error_per_sample = v_nstep_td_error(data, 0.99, 5)
+    loss.backward()
+    assert isinstance(v.grad, torch.Tensor)
+
+
+@pytest.mark.unittest
+def test_q_nstep_sql_td():
+    batch_size = 4
+    action_dim = 3
+    next_q = torch.randn(batch_size, action_dim)
+    done = torch.randn(batch_size)
+    action = torch.randint(0, action_dim, size=(batch_size, ))
+    next_action = torch.randint(0, action_dim, size=(batch_size, ))
+    for nstep in range(1, 10):
+        q = torch.randn(batch_size, action_dim).requires_grad_(True)
+        reward = torch.rand(nstep, batch_size)
+        data = q_nstep_td_data(q, next_q, action, next_action, reward, done, None)
+        loss, td_error_per_sample, record_target_v = q_nstep_sql_td_error(data, 0.95, 1.0, nstep=nstep)
+        assert td_error_per_sample.shape == (batch_size, )
+        assert loss.shape == ()
+        assert q.grad is None
+        loss.backward()
+        assert isinstance(q.grad, torch.Tensor)
+        data = q_nstep_td_data(q, next_q, action, next_action, reward, done, None)
+        loss, td_error_per_sample, record_target_v = q_nstep_sql_td_error(data, 0.95, 0.5, nstep=nstep, cum_reward=True)
+        value_gamma = torch.tensor(0.9)
+        data = q_nstep_td_data(q, next_q, action, next_action, reward, done, None)
+        loss, td_error_per_sample, record_target_v = q_nstep_sql_td_error(
+            data, 0.95, 0.5, nstep=nstep, cum_reward=True, value_gamma=value_gamma
+        )
+        loss.backward()
+        assert isinstance(q.grad, torch.Tensor)
+
+
+@pytest.mark.unittest
+def test_iqn_nstep_td():
+    batch_size = 4
+    action_dim = 3
+    tau = 3
+    next_q = torch.randn(tau, batch_size, action_dim)
+    done = torch.randn(batch_size)
+    action = torch.randint(0, action_dim, size=(batch_size, ))
+    next_action = torch.randint(0, action_dim, size=(batch_size, ))
+    for nstep in range(1, 10):
+        q = torch.randn(tau, batch_size, action_dim).requires_grad_(True)
+        replay_quantile = torch.randn([tau, batch_size, 1])
+        reward = torch.rand(nstep, batch_size)
+        data = iqn_nstep_td_data(q, next_q, action, next_action, reward, done, replay_quantile, None)
+        loss, td_error_per_sample = iqn_nstep_td_error(data, 0.95, nstep=nstep)
+        assert td_error_per_sample.shape == (batch_size, )
+        assert loss.shape == ()
+        assert q.grad is None
+        loss.backward()
+        assert isinstance(q.grad, torch.Tensor)
+        loss, td_error_per_sample = iqn_nstep_td_error(data, 0.95, nstep=nstep, value_gamma=torch.tensor(0.9))
+        assert td_error_per_sample.shape == (batch_size, )
+
+
+@pytest.mark.unittest
+def test_shape_fn_qntd():
+    batch_size = 4
+    action_dim = 3
+    next_q = torch.randn(batch_size, action_dim)
+    done = torch.randn(batch_size)
+    action = torch.randint(0, action_dim, size=(batch_size, ))
+    next_action = torch.randint(0, action_dim, size=(batch_size, ))
+    for nstep in range(1, 10):
+        q = torch.randn(batch_size, action_dim).requires_grad_(True)
+        reward = torch.rand(nstep, batch_size)
+        data = q_nstep_td_data(q, next_q, action, next_action, reward, done, None)
+        tmp = shape_fn_qntd([data, 0.95, 1], {})
+        assert tmp[0] == reward.shape[0]
+        assert tmp[1] == q.shape[0]
+        assert tmp[2] == q.shape[1]
+        tmp = shape_fn_qntd([], {'gamma': 0.95, 'nstep': 1, 'data': data})
+        assert tmp[0] == reward.shape[0]
+        assert tmp[1] == q.shape[0]
+        assert tmp[2] == q.shape[1]
+
+
+@pytest.mark.unittest
+def test_shape_fn_dntd():
+    batch_size = 4
+    action_dim = 3
+    n_atom = 51
+    v_min = -10.0
+    v_max = 10.0
+    nstep = 5
+    dist = torch.randn(batch_size, action_dim, n_atom).abs().requires_grad_(True)
+    next_n_dist = torch.randn(batch_size, action_dim, n_atom).abs()
+    done = torch.randn(batch_size)
+    action = torch.randint(0, action_dim, size=(batch_size, ))
+    next_action = torch.randint(0, action_dim, size=(batch_size, ))
+    reward = torch.randn(nstep, batch_size)
+    data = dist_nstep_td_data(dist, next_n_dist, action, next_action, reward, done, None)
+    tmp = shape_fn_dntd([data, 0.9, v_min, v_max, n_atom, nstep], {})
+    assert tmp[0] == reward.shape[0]
+    assert tmp[1] == dist.shape[0]
+    assert tmp[2] == dist.shape[1]
+    assert tmp[3] == n_atom
+    tmp = shape_fn_dntd([], {'data': data, 'gamma': 0.9, 'v_min': v_min, 'v_max': v_max, 'n_atom': n_atom, 'nstep': 5})
+    assert tmp[0] == reward.shape[0]
+    assert tmp[1] == dist.shape[0]
+    assert tmp[2] == dist.shape[1]
+    assert tmp[3] == n_atom
+
+
+@pytest.mark.unittest
+def test_shape_fn_qntd_rescale():
+    batch_size = 4
+    action_dim = 3
+    next_q = torch.randn(batch_size, action_dim)
+    done = torch.randn(batch_size)
+    action = torch.randint(0, action_dim, size=(batch_size, ))
+    next_action = torch.randint(0, action_dim, size=(batch_size, ))
+    for nstep in range(1, 10):
+        q = torch.randn(batch_size, action_dim).requires_grad_(True)
+        reward = torch.rand(nstep, batch_size)
+        data = q_nstep_td_data(q, next_q, action, next_action, reward, done, None)
+        tmp = shape_fn_qntd_rescale([data, 0.95, 1], {})
+        assert tmp[0] == reward.shape[0]
+        assert tmp[1] == q.shape[0]
+        assert tmp[2] == q.shape[1]
+        tmp = shape_fn_qntd_rescale([], {'gamma': 0.95, 'nstep': 1, 'data': data})
+        assert tmp[0] == reward.shape[0]
+        assert tmp[1] == q.shape[0]
+        assert tmp[2] == q.shape[1]
+
+
+@pytest.mark.unittest
+def test_fn_td_lambda():
+    T, B = 8, 4
+    value = torch.randn(T + 1, B).requires_grad_(True)
+    reward = torch.rand(T, B)
+    data = td_lambda_data(value, reward, None)
+    tmp = shape_fn_td_lambda([], {'data': data})
+    assert tmp == reward.shape[0]
+    tmp = shape_fn_td_lambda([data], {})
+    assert tmp == reward.shape