Add debug output and original implementation

config/value_dice.json5

@@ -9,12 +9,12 @@
         deepsets: {
             input_dim: 6, // number of relative state vars for others
             phi: {
-                hidden_n: 2,
+                hidden_n: 1,
                 hidden_dim: 20,
             },
             latent_dim: 20,
             rho: {
-                hidden_n: 2,
+                hidden_n: 1,
                 hidden_dim: 10,
             },
             output_dim: 10
@@ -43,12 +43,12 @@
         deepsets: {
             input_dim: 6, // number of relative state vars for others
             phi: {
-                hidden_n: 2,
+                hidden_n: 1,
                 hidden_dim: 20,
             },
             latent_dim: 20,
             rho: {
-                hidden_n: 2,
+                hidden_n: 1,
                 hidden_dim: 10,
             },
             output_dim: 10
@@ -70,13 +70,13 @@
     },
     policy_optim: {
             optimizer: 'adam',
-            lr: 1e-3,
+            lr: 1e-0,
-            weight_decay: 0.1,
+            weight_decay: 0.01,
     },
     value_optim: {
             optimizer: 'adam',
-            lr: 1e-3,
+            lr: 1e-6,
-            weight_decay: 0.1,
+            weight_decay: 0.01,
     },
     train_epochs: 200,
     train_batch_size: 32,

scratch/johannes/valuedice.py

@@ -0,0 +1,121 @@
+def weighted_softmax(x, weights, axis=0):
+  x = x - tf.reduce_max(x, axis=axis)
+  return weights * tf.exp(x) / tf.reduce_sum(
+      weights * tf.exp(x), axis=axis, keepdims=True)
+
+
+@tf.function
+  def update(self,
+             expert_dataset_iter,
+             policy_dataset_iter,
+             discount,
+             replay_regularization=0.05,
+             nu_reg=10.0):
+    """A function that updates nu network.
+    When replay regularization is non-zero, it learns
+    (d_pi * (1 - replay_regularization) + d_rb * replay_regulazation) /
+    (d_expert * (1 - replay_regularization) + d_rb * replay_regulazation)
+    instead.
+    Args:
+      expert_dataset_iter: An tensorflow graph iteratable over expert data.
+      policy_dataset_iter: An tensorflow graph iteratable over training policy
+        data, used for regularization.
+      discount: An MDP discount.
+      replay_regularization: A fraction of samples to add from a replay buffer.
+      nu_reg: A grad penalty regularization coefficient.
+    """
+
+    (expert_states, expert_actions,
+     expert_next_states) = expert_dataset_iter.get_next()
+
+    expert_initial_states = expert_states
+
+    rb_states, rb_actions, rb_next_states, _, _ = policy_dataset_iter.get_next(
+    )[0]
+
+    with tf.GradientTape(
+        watch_accessed_variables=False, persistent=True) as tape:
+      tape.watch(self.actor.variables)
+      tape.watch(self.nu_net.variables)
+
+      _, policy_next_actions, _ = self.actor(expert_next_states)
+    #   _, rb_next_actions, rb_log_prob = self.actor(rb_next_states)
+
+      _, policy_initial_actions, _ = self.actor(expert_initial_states)
+
+      Inputs for the linear part of DualDICE loss.
+      expert_init_inputs = tf.concat(
+          [expert_initial_states, policy_initial_actions], 1)
+
+      expert_inputs = tf.concat([expert_states, expert_actions], 1)
+      expert_next_inputs = tf.concat([expert_next_states, policy_next_actions],
+                                     1)
+
+      rb_inputs = tf.concat([rb_states, rb_actions], 1)
+      rb_next_inputs = tf.concat([rb_next_states, rb_next_actions], 1)
+
+      expert_nu_0 = self.nu_net(expert_init_inputs)
+      expert_nu = self.nu_net(expert_inputs)
+      expert_nu_next = self.nu_net(expert_next_inputs)
+
+      rb_nu = self.nu_net(rb_inputs)
+      rb_nu_next = self.nu_net(rb_next_inputs)
+
+      expert_diff = expert_nu - discount * expert_nu_next
+      rb_diff = rb_nu - discount * rb_nu_next
+
+      linear_loss_expert = tf.reduce_mean(expert_nu_0 * (1 - discount))
+
+      linear_loss_rb = tf.reduce_mean(rb_diff)
+
+      rb_expert_diff = tf.concat([expert_diff, rb_diff], 0)
+      rb_expert_weights = tf.concat([
+          tf.ones(expert_diff.shape) * (1 - replay_regularization),
+          tf.ones(rb_diff.shape) * replay_regularization
+      ], 0)
+
+      rb_expert_weights /= tf.reduce_sum(rb_expert_weights)
+      non_linear_loss = tf.reduce_sum(
+          tf.stop_gradient(
+              weighted_softmax(rb_expert_diff, rb_expert_weights, axis=0)) *
+          rb_expert_diff)
+
+      linear_loss = (
+          linear_loss_expert * (1 - replay_regularization) +
+          linear_loss_rb * replay_regularization)
+
+      loss = (non_linear_loss - linear_loss)
+
+      alpha = tf.random.uniform(shape=(expert_inputs.shape[0], 1))
+
+      nu_inter = alpha * expert_inputs + (1 - alpha) * rb_inputs
+      nu_next_inter = alpha * expert_next_inputs + (1 - alpha) * rb_next_inputs
+
+      nu_inter = tf.concat([nu_inter, nu_next_inter], 0)
+
+      with tf.GradientTape(watch_accessed_variables=False) as tape2:
+        tape2.watch(nu_inter)
+        nu_output = self.nu_net(nu_inter)
+      nu_grad = tape2.gradient(nu_output, [nu_inter])[0] + EPS
+      nu_grad_penalty = tf.reduce_mean(
+          tf.square(tf.norm(nu_grad, axis=-1, keepdims=True) - 1))
+
+      nu_loss = loss + nu_grad_penalty * nu_reg
+      pi_loss = -loss + keras_utils.orthogonal_regularization(self.actor.trunk)
+
+    nu_grads = tape.gradient(nu_loss, self.nu_net.variables)
+    pi_grads = tape.gradient(pi_loss, self.actor.variables)
+
+    self.nu_optimizer.apply_gradients(zip(nu_grads, self.nu_net.variables))
+    self.actor_optimizer.apply_gradients(zip(pi_grads, self.actor.variables))
+
+    del tape
+
+    self.avg_nu_expert(expert_nu)
+    self.avg_nu_rb(rb_nu)
+
+    self.nu_reg_metric(nu_grad_penalty)
+    self.avg_loss(loss)
+
+    self.avg_actor_loss(pi_loss)
+    self.avg_actor_entropy(-rb_log_prob)

src/value_dice/value_dice.py

@@ -215,9 +215,16 @@ def train(config, policy, train_dataset, cv_dataset, filestr, **kwargs):
 
         # nonlinear loss
         value_diff = value - discount * value_next
-        nonlinear_loss = torch.logsumexp(value_diff, dim=0) - np.log(len(value_diff))
+        # print(value_diff)
+        # nonlinear_loss = torch.logsumexp(value_diff, dim=0) #- np.log(len(value_diff))
+        nonlinear_loss = torch.log(torch.mean(torch.exp(value_diff), dim=0))
 
         loss = nonlinear_loss - linear_loss
+        print("Loss report:")
+        print("Linear: {}".format(linear_loss.item()))
+        print("Nonlinear: {}".format(nonlinear_loss.item()))
+        print("Total: {}".format(loss.item()))
+
         return loss
 
 
@@ -237,7 +244,6 @@ def train(config, policy, train_dataset, cv_dataset, filestr, **kwargs):
         # train
         epoch_loss = 0
         for (batch_idx, batch) in enumerate(training_loader):
-
             loss = f_value_dice_loss(batch)
 
             # In original implementation policy is regularized with orthogonal regularization, 
@@ -260,12 +266,22 @@ def train(config, policy, train_dataset, cv_dataset, filestr, **kwargs):
             if batch_idx % 2 == 0:
                 policy_optimizer.zero_grad()
                 policy_loss.backward()
-                clip_grad_norm_(policy.policy.parameters(), clip_grad_norm)
+                # clip_grad_norm_(policy.policy.parameters(), clip_grad_norm)
                 policy_optimizer.step()
+
+                grad_list = torch.cat([torch.flatten(p.grad) for p in policy.policy.parameters()])
+                torch.mean(grad_list)
+                print("gradient stats:")
+                print(torch.mean(grad_list))
+                print(torch.std(grad_list))
+                print(torch.min(grad_list))
+                print(torch.max(grad_list))
+                # print(policy.policy.head.layers[0].weight.grad)
+                # print(policy.policy.head.layers[0].bias.grad)
             else:
                 value_optimizer.zero_grad()
                 value_loss.backward()
-                clip_grad_norm_(policy.value.parameters(), clip_grad_norm)
+                # clip_grad_norm_(policy.value.parameters(), clip_grad_norm)
                 value_optimizer.step()
 
             epoch_loss += loss.item() / len(train_dataset)