Update to PyTorch 1.3

Author: Omegastick

.gitmodules

@@ -1,7 +1,7 @@
-[submodule "lib/spdlog"]
+[submodule "example/lib/spdlog"]
 	path = example/lib/spdlog
 	url = [email protected]:gabime/spdlog.git
-[submodule "lib/msgpack-c"]
+[submodule "example/lib/msgpack-c"]
 	path = example/lib/msgpack-c
 	url = [email protected]:msgpack/msgpack-c.git
 [submodule "example/lib/libzmq"]

example/gym_client.cpp

@@ -182,7 +182,7 @@ int main(int argc, char *argv[])
                                          storage.get_masks()[step]);
             }
             auto actions_tensor = act_result[1].cpu().to(torch::kFloat);
-            float *actions_array = actions_tensor.data<float>();
+            float *actions_array = actions_tensor.data_ptr<float>();
             std::vector<std::vector<float>> actions(num_envs);
             for (int i = 0; i < num_envs; ++i)
             {
@@ -218,7 +218,7 @@ int main(int argc, char *argv[])
                 returns_rms->update(returns);
                 reward_tensor = torch::clamp(reward_tensor / torch::sqrt(returns_rms->get_variance() + 1e-8),
                                              -reward_clip_value, reward_clip_value);
-                rewards = std::vector<float>(reward_tensor.data<float>(), reward_tensor.data<float>() + reward_tensor.numel());
+                rewards = std::vector<float>(reward_tensor.data_ptr<float>(), reward_tensor.data_ptr<float>() + reward_tensor.numel());
                 real_rewards = flatten_vector(step_result->real_reward);
                 dones_vec = step_result->done;
             }
@@ -233,7 +233,7 @@ int main(int argc, char *argv[])
                 returns_rms->update(returns);
                 reward_tensor = torch::clamp(reward_tensor / torch::sqrt(returns_rms->get_variance() + 1e-8),
                                              -reward_clip_value, reward_clip_value);
-                rewards = std::vector<float>(reward_tensor.data<float>(), reward_tensor.data<float>() + reward_tensor.numel());
+                rewards = std::vector<float>(reward_tensor.data_ptr<float>(), reward_tensor.data_ptr<float>() + reward_tensor.numel());
                 real_rewards = flatten_vector(step_result->real_reward);
                 dones_vec = step_result->done;
             }

src/algorithms/a2c.cpp

@@ -36,7 +36,7 @@ A2C::A2C(Policy &policy,
 std::vector<UpdateDatum> A2C::update(RolloutStorage &rollouts, float decay_level)
 {
     // Decay learning rate
-    optimizer->options.learning_rate_ = original_learning_rate * decay_level;
+    optimizer->options.learning_rate(original_learning_rate * decay_level);
 
     // Prep work
     auto full_obs_shape = rollouts.get_observations().sizes();

src/algorithms/ppo.cpp

@@ -45,7 +45,7 @@ std::vector<UpdateDatum> PPO::update(RolloutStorage &rollouts, float decay_level
 {
     // Decay lr and clip parameter
     float clip_param = original_clip_param * decay_level;
-    optimizer->options.learning_rate_ = original_learning_rate * decay_level;
+    optimizer->options.learning_rate(original_learning_rate * decay_level);
 
     // Calculate advantages
     auto returns = rollouts.get_returns();

src/model/nn_base.cpp

@@ -81,8 +81,8 @@ std::vector<torch::Tensor> NNBase::forward_gru(torch::Tensor x,
         // has_zeros = [0] + has_zeros + [timesteps]
         has_zeros = has_zeros.contiguous().to(torch::kInt);
         std::vector<int> has_zeros_vec(
-            has_zeros.data<int>(),
-            has_zeros.data<int>() + has_zeros.numel());
+            has_zeros.data_ptr<int>(),
+            has_zeros.data_ptr<int>() + has_zeros.numel());
         has_zeros_vec.insert(has_zeros_vec.begin(), {0});
         has_zeros_vec.push_back(timesteps);
 

src/observation_normalizer.cpp

@@ -76,13 +76,13 @@ torch::Tensor ObservationNormalizerImpl::process_observation(torch::Tensor obser
 std::vector<float> ObservationNormalizerImpl::get_mean() const
 {
     auto mean = rms->get_mean();
-    return std::vector<float>(mean.data<float>(), mean.data<float>() + mean.numel());
+    return std::vector<float>(mean.data_ptr<float>(), mean.data_ptr<float>() + mean.numel());
 }
 
 std::vector<float> ObservationNormalizerImpl::get_variance() const
 {
     auto variance = rms->get_variance();
-    return std::vector<float>(variance.data<float>(), variance.data<float>() + variance.numel());
+    return std::vector<float>(variance.data_ptr<float>(), variance.data_ptr<float>() + variance.numel());
 }
 
 void ObservationNormalizerImpl::update(torch::Tensor observations)