Merge

pom.xml

@@ -8,7 +8,7 @@
 
   <groupId>de.monticore.lang.monticar</groupId>
   <artifactId>embedded-montiarc-emadl-generator</artifactId>
-  <version>0.3.3-SNAPSHOT</version>
+  <version>0.3.4-SNAPSHOT</version>
 
   <!-- == PROJECT DEPENDENCIES ============================================= -->
 

src/main/java/de/monticore/lang/monticar/emadl/generator/Backend.java

@@ -9,6 +9,7 @@ import de.monticore.lang.monticar.cnnarch.mxnetgenerator.CNNArch2MxNet;
 import de.monticore.lang.monticar.cnnarch.caffe2generator.CNNArch2Caffe2;
 import de.monticore.lang.monticar.cnnarch.mxnetgenerator.CNNTrain2MxNet;
 import de.monticore.lang.monticar.cnnarch.caffe2generator.CNNTrain2Caffe2;
+import de.monticore.lang.monticar.emadl.generator.reinforcementlearning.RewardFunctionCppGenerator;
 
 import java.util.Optional;
 

src/main/java/de/monticore/lang/monticar/emadl/generator/EMADLGenerator.java

@@ -34,6 +34,8 @@ import de.monticore.lang.monticar.cnnarch._symboltable.ArchitectureSymbol;
 import de.monticore.lang.monticar.cnnarch._symboltable.SerialCompositeElementSymbol;
 import de.monticore.lang.monticar.cnnarch.gluongenerator.CNNTrain2Gluon;
 import de.monticore.lang.monticar.cnnarch.gluongenerator.annotations.ArchitectureAdapter;
+import de.monticore.lang.monticar.cnntrain._cocos.CNNTrainCoCoChecker;
+import de.monticore.lang.monticar.cnntrain._cocos.CNNTrainCocos;
 import de.monticore.lang.monticar.cnntrain._symboltable.ConfigurationSymbol;
 import de.monticore.lang.monticar.emadl._cocos.EMADLCocos;
 import de.monticore.lang.monticar.generator.FileContent;
@@ -115,6 +117,18 @@ public class EMADLGenerator {
         processedArchitecture = new HashMap<>();
         setModelsPath( modelPath );
         TaggingResolver symtab = EMADLAbstractSymtab.createSymTabAndTaggingResolver(getModelsPath());
+        EMAComponentInstanceSymbol instance = resolveComponentInstanceSymbol(qualifiedName, symtab);
+
+
+        generateFiles(symtab, instance, symtab, pythonPath, forced);
+        
+        if (doCompile) {
+            compile();
+        }
+        processedArchitecture = null;
+    }
+
+    private EMAComponentInstanceSymbol resolveComponentInstanceSymbol(String qualifiedName, TaggingResolver symtab) {
         EMAComponentSymbol component = symtab.<EMAComponentSymbol>resolve(qualifiedName, EMAComponentSymbol.KIND).orElse(null);
 
         List<String> splitName = Splitters.DOT.splitToList(qualifiedName);
@@ -126,15 +140,7 @@ public class EMADLGenerator {
             System.exit(1);
         }
 
-        EMAComponentInstanceSymbol instance = component.getEnclosingScope().<EMAComponentInstanceSymbol>resolve(instanceName, EMAComponentInstanceSymbol.KIND).get();
-
-
-        generateFiles(symtab, instance, symtab, pythonPath, forced);
-        
-        if (doCompile) {
-            compile();
-        }
-        processedArchitecture = null;
+        return component.getEnclosingScope().<EMAComponentInstanceSymbol>resolve(instanceName, EMAComponentInstanceSymbol.KIND).get();
     }
 
     public void compile() throws IOException {
@@ -530,7 +536,32 @@ public class EMADLGenerator {
                 final String fullConfigName = String.join(".", names);
                 ArchitectureSymbol correspondingArchitecture = this.processedArchitecture.get(fullConfigName);
                 assert correspondingArchitecture != null : "No architecture found for train " + fullConfigName + " configuration!";
-                configuration.setTrainedArchitecture(new ArchitectureAdapter(correspondingArchitecture));
+                configuration.setTrainedArchitecture(
+                        new ArchitectureAdapter(correspondingArchitecture.getName(), correspondingArchitecture));
+                CNNTrainCocos.checkTrainedArchitectureCoCos(configuration);
+
+                // Resolve critic network if critic is present
+                if (configuration.getCriticName().isPresent()) {
+                    String fullCriticName = configuration.getCriticName().get();
+                    int indexOfFirstNameCharacter = fullCriticName.lastIndexOf('.') + 1;
+                    fullCriticName = fullCriticName.substring(0, indexOfFirstNameCharacter)
+                            + fullCriticName.substring(indexOfFirstNameCharacter, indexOfFirstNameCharacter + 1).toUpperCase()
+                            + fullCriticName.substring(indexOfFirstNameCharacter + 1);
+
+                    TaggingResolver symtab = EMADLAbstractSymtab.createSymTabAndTaggingResolver(getModelsPath());
+                    EMAComponentInstanceSymbol instanceSymbol = resolveComponentInstanceSymbol(fullCriticName, symtab);
+                    EMADLCocos.checkAll(instanceSymbol);
+                    Optional<ArchitectureSymbol> critic = instanceSymbol.getSpannedScope().resolve("", ArchitectureSymbol.KIND);
+                    if (!critic.isPresent()) {
+                        Log.error("During the resolving of critic component: Critic component "
+                                + fullCriticName + " does not have a CNN implementation but is required to have one");
+                        System.exit(-1);
+                    }
+                    critic.get().setComponentName(fullCriticName);
+                    configuration.setCriticNetwork(new ArchitectureAdapter(fullCriticName, critic.get()));
+                    CNNTrainCocos.checkCriticCocos(configuration);
+                }
+
 
                 cnnTrainGenerator.setInstanceName(componentInstance.getFullName().replaceAll("\\.", "_"));
                 Map<String, String> fileContentMap =  cnnTrainGenerator.generateStrings(configuration);

src/main/java/de/monticore/lang/monticar/emadl/generator/RewardFunctionCppGenerator.java → src/main/java/de/monticore/lang/monticar/emadl/generator/reinforcementlearning/RewardFunctionCppGenerator.java

-package de.monticore.lang.monticar.emadl.generator;
+package de.monticore.lang.monticar.emadl.generator.reinforcementlearning;
 
 import de.monticore.lang.embeddedmontiarc.embeddedmontiarc._symboltable.instanceStructure.EMAComponentInstanceSymbol;
 import de.monticore.lang.monticar.cnnarch.gluongenerator.reinforcement.RewardFunctionSourceGenerator;
+import de.monticore.lang.monticar.emadl.generator.EMADLAbstractSymtab;
 import de.monticore.lang.monticar.generator.cpp.GeneratorEMAMOpt2CPP;
 import de.monticore.lang.tagging._symboltable.TaggingResolver;
 import de.se_rwth.commons.logging.Log;
@@ -9,30 +10,49 @@ import de.se_rwth.commons.logging.Log;
 import java.io.IOException;
 import java.util.Optional;
 
-public class RewardFunctionCppGenerator implements RewardFunctionSourceGenerator {
+public class RewardFunctionCppGenerator implements RewardFunctionSourceGenerator{
     public RewardFunctionCppGenerator() {
     }
 
-    @Override
-    public void generate(String modelPath, String rootModel, String targetPath) {
-        GeneratorEMAMOpt2CPP generator = new GeneratorEMAMOpt2CPP();
-        generator.useArmadilloBackend();
 
-        TaggingResolver taggingResolver = EMADLAbstractSymtab.createSymTabAndTaggingResolver(modelPath);
+    @Override
+    public EMAComponentInstanceSymbol resolveSymbol(TaggingResolver taggingResolver, String rootModel) {
         Optional<EMAComponentInstanceSymbol> instanceSymbol = taggingResolver
                 .<EMAComponentInstanceSymbol>resolve(rootModel, EMAComponentInstanceSymbol.KIND);
 
         if (!instanceSymbol.isPresent()) {
             Log.error("Generation of reward function is not possible: Cannot resolve component instance "
-                + rootModel);
+                    + rootModel);
         }
 
+        return instanceSymbol.get();
+    }
+
+    @Override
+    public void generate(EMAComponentInstanceSymbol componentInstanceSymbol, TaggingResolver taggingResolver,
+                         String targetPath) {
+        GeneratorEMAMOpt2CPP generator = new GeneratorEMAMOpt2CPP();
+        generator.useArmadilloBackend();
+
         generator.setGenerationTargetPath(targetPath);
 
         try {
-            generator.generate(instanceSymbol.get(), taggingResolver);
+            generator.generate(componentInstanceSymbol, taggingResolver);
         } catch (IOException e) {
             Log.error("Generation of reward function is not possible: " + e.getMessage());
         }
+
+    }
+
+    @Override
+    public void generate(String modelPath, String rootModel, String targetPath) {
+        TaggingResolver taggingResolver = createTaggingResolver(modelPath);
+        EMAComponentInstanceSymbol instanceSymbol = resolveSymbol(taggingResolver, rootModel);
+        generate(instanceSymbol, taggingResolver, targetPath);
+    }
+
+    @Override
+    public TaggingResolver createTaggingResolver(final String modelPath) {
+        return EMADLAbstractSymtab.createSymTabAndTaggingResolver(modelPath);
     }
 }

src/test/java/de/monticore/lang/monticar/emadl/GenerationTest.java

@@ -275,8 +275,8 @@ public class GenerationTest extends AbstractSymtabTest {
                         "HelperA.h",
                         "start_training.sh",
                         "reinforcement_learning/__init__.py",
-                        "reinforcement_learning/CNNCreator_MountaincarCritic.py",
-                        "reinforcement_learning/CNNNet_MountaincarCritic.py",
+                        "reinforcement_learning/CNNCreator_mountaincar_agent_mountaincarCritic.py",
+                        "reinforcement_learning/CNNNet_mountaincar_agent_mountaincarCritic.py",
                         "reinforcement_learning/strategy.py",
                         "reinforcement_learning/agent.py",
                         "reinforcement_learning/environment.py",

src/test/resources/models/reinforcementModel/cartpole/agent/CartPoleDQN.cnnt

@@ -17,7 +17,7 @@ configuration CartPoleDQN {
 
     use_double_dqn : false
 
-    loss : euclidean
+    loss : huber
 
     replay_memory : buffer{
         memory_size : 10000

src/test/resources/models/reinforcementModel/mountaincar/agent/MountaincarCritic.cnna

-implementation Critic(state, action) {
-    (state ->
-        FullyConnected(units=400) ->
-        Relu() ->
-        FullyConnected(units=300)
-    |
-    action ->
-        FullyConnected(units=300)
-    ) ->
-    Add() ->
-    Relu();
-}
\ No newline at end of file

src/test/resources/models/reinforcementModel/mountaincar/agent/MountaincarCritic.emadl

+package mountaincar.agent;
+
+component MountaincarCritic {
+    ports
+        in Q^{2} state,
+        in Q(-1:1)^{1} action,
+        out Q(-oo:oo)^{1} qvalues;
+
+    implementation CNN {
+        (
+            state ->
+            FullyConnected(units=400) ->
+            Relu() ->
+            FullyConnected(units=300)
+        |
+            action ->
+            FullyConnected(units=300)
+        ) ->
+        Add() ->
+        Relu() ->
+        FullyConnected(units=1) ->
+        qvalues;
+    }
+}
\ No newline at end of file

src/test/resources/models/reinforcementModel/torcs/agent/dqn/TorcsDQN.cnnt

@@ -23,7 +23,7 @@ configuration TorcsDQN {
 
     use_double_dqn : true
 
-    loss : euclidean
+    loss : huber
 
     replay_memory : buffer{
         memory_size : 1000000

src/test/resources/target_code/gluon/reinforcementModel/cartpole/CNNDataLoader_cartpole_master_dqn.py

@@ -3,8 +3,9 @@ import h5py
 import mxnet as mx
 import logging
 import sys
+from mxnet import nd
 
-class cartpole_master_dqnDataLoader:
+class CNNDataLoader_cartpole_master_dqn:
     _input_names_ = ['state']
     _output_names_ = ['qvalues_label']
 
@@ -14,21 +15,38 @@ class cartpole_master_dqnDataLoader:
     def load_data(self, batch_size):
         train_h5, test_h5 = self.load_h5_files()
 
-        data_mean = train_h5[self._input_names_[0]][:].mean(axis=0)
-        data_std = train_h5[self._input_names_[0]][:].std(axis=0) + 1e-5
+        train_data = {}
+        data_mean = {}
+        data_std = {}
+
+        for input_name in self._input_names_:
+            train_data[input_name] = train_h5[input_name]
+            data_mean[input_name] = nd.array(train_h5[input_name][:].mean(axis=0))
+            data_std[input_name] = nd.array(train_h5[input_name][:].std(axis=0) + 1e-5)
+
+        train_label = {}
+        for output_name in self._output_names_:
+            train_label[output_name] = train_h5[output_name]
+
+        train_iter = mx.io.NDArrayIter(data=train_data,
+                                       label=train_label,
+                                       batch_size=batch_size)
 
-        train_iter = mx.io.NDArrayIter(train_h5[self._input_names_[0]],
-                                       train_h5[self._output_names_[0]],
-                                       batch_size=batch_size,
-                                       data_name=self._input_names_[0],
-                                       label_name=self._output_names_[0])
         test_iter = None
+
         if test_h5 != None:
-            test_iter = mx.io.NDArrayIter(test_h5[self._input_names_[0]],
-                                          test_h5[self._output_names_[0]],
-                                          batch_size=batch_size,
-                                          data_name=self._input_names_[0],
-                                          label_name=self._output_names_[0])
+            test_data = {}
+            for input_name in self._input_names_:
+                test_data[input_name] = test_h5[input_name]
+
+            test_label = {}
+            for output_name in self._output_names_:
+                test_label[output_name] = test_h5[output_name]
+
+            test_iter = mx.io.NDArrayIter(data=test_data,
+                                          label=test_label,
+                                          batch_size=batch_size)
+
         return train_iter, test_iter, data_mean, data_std
 
     def load_h5_files(self):
@@ -36,21 +54,39 @@ class cartpole_master_dqnDataLoader:
         test_h5 = None
         train_path = self._data_dir + "train.h5"
         test_path = self._data_dir + "test.h5"
+
         if os.path.isfile(train_path):
             train_h5 = h5py.File(train_path, 'r')
-            if not (self._input_names_[0] in train_h5 and self._output_names_[0] in train_h5):
-                logging.error("The HDF5 file '" + os.path.abspath(train_path) + "' has to contain the datasets: "
-                              + "'" + self._input_names_[0] + "', '" + self._output_names_[0] + "'")
-                sys.exit(1)
-            test_iter = None
+
+            for input_name in self._input_names_:
+                if not input_name in train_h5:
+                    logging.error("The HDF5 file '" + os.path.abspath(train_path) + "' has to contain the dataset "
+                                  + "'" + input_name + "'")
+                    sys.exit(1)
+
+            for output_name in self._output_names_:
+                if not output_name in train_h5:
+                    logging.error("The HDF5 file '" + os.path.abspath(train_path) + "' has to contain the dataset "
+                                  + "'" + output_name + "'")
+                    sys.exit(1)
+
             if os.path.isfile(test_path):
                 test_h5 = h5py.File(test_path, 'r')
-                if not (self._input_names_[0] in test_h5 and self._output_names_[0] in test_h5):
-                    logging.error("The HDF5 file '" + os.path.abspath(test_path) + "' has to contain the datasets: "
-                                  + "'" + self._input_names_[0] + "', '" + self._output_names_[0] + "'")
-                    sys.exit(1)
+
+                for input_name in self._input_names_:
+                    if not input_name in test_h5:
+                        logging.error("The HDF5 file '" + os.path.abspath(test_path) + "' has to contain the dataset "
+                                      + "'" + input_name + "'")
+                        sys.exit(1)
+
+                for output_name in self._output_names_:
+                    if not output_name in test_h5:
+                        logging.error("The HDF5 file '" + os.path.abspath(test_path) + "' has to contain the dataset "
+                                      + "'" + output_name + "'")
+                        sys.exit(1)
             else:
                 logging.warning("Couldn't load test set. File '" + os.path.abspath(test_path) + "' does not exist.")
+
             return train_h5, test_h5
         else:
             logging.error("Data loading failure. File '" + os.path.abspath(train_path) + "' does not exist.")

src/test/resources/target_code/gluon/reinforcementModel/cartpole/CNNNet_cartpole_master_dqn.py

@@ -101,7 +101,6 @@ class Net_0(gluon.HybridBlock):
             self.fc3_ = gluon.nn.Dense(units=2, use_bias=True)
             # fc3_, output shape: {[2,1,1]}
 
-        self.last_layers['qvalues'] = 'linear'
 
 
     def hybrid_forward(self, F, state):

src/test/resources/target_code/gluon/reinforcementModel/cartpole/CNNTrainer_cartpole_master_dqn.py

@@ -56,7 +56,7 @@ if __name__ == "__main__":
             'memory_size': 10000,
             'sample_size': 32,
             'state_dtype': 'float32',
-            'action_dtype': 'float32',
+            'action_dtype': 'uint8',
             'rewards_dtype': 'float32'
         },
         'strategy_params': {
@@ -78,10 +78,10 @@ if __name__ == "__main__":
         'snapshot_interval': 20,
         'max_episode_step': 250,
         'target_score': 185.5,
-        'qnet':qnet_creator.net,
+        'qnet':qnet_creator.networks[0],
         'use_fix_target': True,
         'target_update_interval': 200,
-        'loss_function': 'euclidean',
+        'loss_function': 'huber',
         'optimizer': 'rmsprop',
         'optimizer_params': {
             'learning_rate': 0.001        },
@@ -108,4 +108,4 @@ if __name__ == "__main__":
     train_successful = agent.train()
 
     if train_successful:
-        agent.save_best_network(qnet_creator._model_dir_ + qnet_creator._model_prefix_ + '_0_newest', epoch=0)
+        agent.export_best_network(path=qnet_creator._model_dir_ + qnet_creator._model_prefix_ + '_0_newest', epoch=0)

src/test/resources/target_code/gluon/reinforcementModel/cartpole/reinforcement_learning/strategy.py

@@ -13,18 +13,21 @@ class StrategyBuilder(object):
         epsilon_decay_method='no',
         epsilon_decay=0.0,
         epsilon_decay_start=0,
+        epsilon_decay_per_step=False,
         action_dim=None,
         action_low=None,
         action_high=None,
         mu=0.0,
         theta=0.5,
-        sigma=0.3
+        sigma=0.3,
+        noise_variance=0.1
     ):
 
         if epsilon_decay_method == 'linear':
             decay = LinearDecay(
                 eps_decay=epsilon_decay, min_eps=min_epsilon,
-                decay_start=epsilon_decay_start)
+                decay_start=epsilon_decay_start,
+                decay_per_step=epsilon_decay_per_step)
         else:
             decay = NoDecay()
 
@@ -44,6 +47,13 @@ class StrategyBuilder(object):
             return OrnsteinUhlenbeckStrategy(
                 action_dim, action_low, action_high, epsilon, mu, theta,
                 sigma, decay)
+        elif method == 'gaussian':
+            assert action_dim is not None
+            assert action_low is not None
+            assert action_high is not None
+            assert noise_variance is not None
+            return GaussianNoiseStrategy(action_dim, action_low, action_high,
+                                         epsilon, noise_variance, decay)
         else:
             assert action_dim is not None
             assert len(action_dim) == 1
@@ -70,17 +80,27 @@ class NoDecay(BaseDecay):
 
 
 class LinearDecay(BaseDecay):
-    def __init__(self, eps_decay, min_eps=0, decay_start=0):
+    def __init__(self, eps_decay, min_eps=0, decay_start=0, decay_per_step=False):
         super(LinearDecay, self).__init__()
         self.eps_decay = eps_decay
         self.min_eps = min_eps
         self.decay_start = decay_start
+        self.decay_per_step = decay_per_step
+        self.last_episode = -1
 
-    def decay(self, cur_eps, episode):
-        if episode < self.decay_start:
-            return cur_eps
+    def do_decay(self, episode):
+        if self.decay_per_step:
+            do = (episode >= self.decay_start)
         else:
+            do = ((self.last_episode != episode) and (episode >= self.decay_start))
+        self.last_episode = episode
+        return do
+
+    def decay(self, cur_eps, episode):
+        if self.do_decay(episode):
             return max(cur_eps - self.eps_decay, self.min_eps)
+        else:
+            return cur_eps
 
 
 class BaseStrategy(object):
@@ -168,5 +188,31 @@ class OrnsteinUhlenbeckStrategy(BaseStrategy):
 
     def select_action(self, values):
         noise = self._evolve_state()
-        action = values + (self.cur_eps * noise)
+        action = (1.0 - self.cur_eps) * values + (self.cur_eps * noise)
+        return np.clip(action, self._action_low, self._action_high)
+
+
+class GaussianNoiseStrategy(BaseStrategy):
+    def __init__(
+        self,
+        action_dim,
+        action_low,
+        action_high,
+        eps,
+        noise_variance,
+        decay=NoDecay()
+    ):
+        super(GaussianNoiseStrategy, self).__init__(decay)
+        self.eps = eps
+        self.cur_eps = eps
+
+        self._action_dim = action_dim
+        self._action_low = action_low
+        self._action_high = action_high
+
+        self._noise_variance = noise_variance
+
+    def select_action(self, values):
+        noise = np.random.normal(loc=0.0, scale=self._noise_variance, size=self._action_dim)
+        action = values + self.cur_eps * noise
         return np.clip(action, self._action_low, self._action_high)

src/test/resources/target_code/gluon/reinforcementModel/cartpole/reinforcement_learning/util.py

@@ -11,8 +11,8 @@ import cnnarch_logger
 
 LOSS_FUNCTIONS = {
         'l1': gluon.loss.L1Loss(),
-        'euclidean': gluon.loss.L2Loss(),
-        'huber_loss': gluon.loss.HuberLoss(),
+        'l2': gluon.loss.L2Loss(),
+        'huber': gluon.loss.HuberLoss(),
         'softmax_cross_entropy': gluon.loss.SoftmaxCrossEntropyLoss(),
         'sigmoid_cross_entropy': gluon.loss.SigmoidBinaryCrossEntropyLoss()}
 
@@ -127,13 +127,15 @@ class TrainingStats(object):
         else:
             return self._all_total_rewards[0]
 
-    def save(self, path):
-        np.save(os.path.join(path, 'total_rewards'), self._all_total_rewards)
-        np.save(os.path.join(path, 'eps'), self._all_eps)
-        np.save(os.path.join(path, 'time'), self._all_time)
+    def save(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+        np.save(os.path.join(path, 'total_rewards'), self._all_total_rewards[:episode])
+        np.save(os.path.join(path, 'eps'), self._all_eps[:episode])
+        np.save(os.path.join(path, 'time'), self._all_time[:episode])
         np.save(
             os.path.join(path, 'mean_reward'),
-            self._all_mean_reward_last_100_episodes)
+            self._all_mean_reward_last_100_episodes[:episode])
 
     def _log_episode(self, episode, start_time, training_steps, eps, reward):
         self.add_eps(episode, eps)
@@ -170,33 +172,43 @@ class DqnTrainingStats(TrainingStats):
         self._logger.info(info)
         return avg_reward
 
-    def save_stats(self, path):
+    def save_stats(self, path, episode=None):
+        if episode is None:
+            episode = self._max_episodes
+
+        all_total_rewards = self._all_total_rewards[:episode]
+        all_avg_loss = self._all_avg_loss[:episode]
+        all_eps = self._all_eps[:episode]
+        all_mean_reward_last_100_episodes = self._all_mean_reward_last_100_episodes[:episode]
+
         fig = plt.figure(figsize=(20, 20))
 
         sub_rewards = fig.add_subplot(221)
         sub_rewards.set_title('Total Rewards per episode')
         sub_rewards.plot(
-            np.arange(self._max_episodes), self._all_total_rewards)
+            np.arange(episode), all_total_rewards)
 
         sub_loss = fig.add_subplot(222)
         sub_loss.set_title('Avg. Loss per episode')
-        sub_loss.plot(np.arange(self._max_episodes), self._all_avg_loss)
+        sub_loss.plot(np.arange(episode), all_avg_loss)
 
         sub_eps = fig.add_subplot(223)
         sub_eps.set_title('Epsilon per episode')
-        sub_eps.plot(np.arange(self._max_episodes), self._all_eps)
+        sub_eps.plot(np.arange(episode), all_eps)
 
         sub_rewards = fig.add_subplot(224)
         sub_rewards.set_title('Avg. mean reward of last 100 episodes')
-        sub_rewards.plot(np.arange(self._max_episodes),
-                         self._all_mean_reward_last_100_episodes)
+        sub_rewards.plot(np.arange(episode),
+                         all_mean_reward_last_100_episodes)
 
-        self.save(path)
+        self.save(path, episode=episode)
         plt.savefig(os.path.join(path, 'stats.pdf'))