Collgon

command_anaconda_prompt

@@ -0,0 +1 @@
+%windir%\System32\cmd.exe "/K" C:\ProgramData\Anaconda3\Scripts\activate.bat C:\ProgramData\Anaconda3

reaver/agents/random.py

@@ -9,6 +9,8 @@ def __init__(self, act_spec, n_envs):
 
     def get_action(self, obs):
         function_id = [np.random.choice(np.argwhere(obs[2][i] > 0).flatten()) for i in range(self.n_envs)]
+        # Va pour les n_env selectionner vraisemblablement? a chaque fois un carré parmi les 16 ou il y a une unité.
         args = [[[np.random.randint(0, size) for size in arg.shape] for _ in range(self.n_envs)]
                 for arg in self.act_spec.spaces[1:]]
         return [function_id] + args
+

reaver/envs/base/abc.py

@@ -11,6 +11,10 @@ class Env(ABC):
 
     Note: observation / action specs contain a list of spaces,
           this is implicitly assumed across all Reaver components
+
+
+    Abstract est un cahier des charges, si à l'intérieur de SC2Env il n'y a pas de définition des fonctions start, step etc, il y aura un message d'erreur
+    SC2Env hérite de Env, Env étant la class cahier des charges définie ici.
     """
     def __init__(self, _id: str, render=False, reset_done=True, max_ep_len=None):
         self.id = _id

reaver/envs/sc2.py

@@ -27,8 +27,8 @@ def __init__(
         render=False,
         reset_done=True,
         max_ep_len=None,
-        spatial_dim=16,
-        step_mul=8,
+        spatial_dim=32, #bitmap size
+        step_mul=8, #frame rate
         obs_features=None,
         action_ids=ACTIONS_MINIGAMES
     ):
@@ -239,6 +239,7 @@ def __call__(self, action):
                 args.append([defaults[arg_name]])
 
         return [actions.FunctionCall(fn_id, args)]
+        #renvoie fn id et arg pour le fn_id utilisés
 
     def make_spec(self, spec):
         spec = spec[0]
@@ -270,6 +271,7 @@ def __init__(self, func_ids, args):
         for fn_id in func_ids:
             fn_id_args = [arg_type.name for arg_type in actions.FUNCTIONS[fn_id].args]
             self.args_mask.append([arg in fn_id_args for arg in args])
+            #ajoute arg à args_mask si arg est dans fn_id_args. Vecteur de type [[False, ...],...,[True,False..]]
 
 
 def get_spatial_dims(feat_names, feats):

reaver/models/base/policy.py

@@ -14,14 +14,20 @@ def __init__(self, act_spec, logits):
         self.logli = sum([dist.log_prob(act) for dist, act in zip(self.dists, self.inputs)])
 
         self.sample = [dist.sample() for dist in self.dists]
-
+        # Sample 1 à 1 les différentes éléments qui constitueront le vecteur action en output via self.dists
+        #
     @staticmethod
     def make_dist(space, logits):
+        """Va permettre de déterminer les lois de probabilités dans le cas discret / continu pour les utiliser ensuite
+        dans self.sample"""
+
         # tfp is really heavy on init, better to lazy load
         import tensorflow_probability as tfp
 
         if space.is_continuous():
             mu, logstd = tf.split(logits, 2, axis=-1)
+            #Le split permet de couper en deux vu qu'on suppose que la matrice de covariance est diagonale.
+            #Mu et logstd ont donc même taille et logits peut être séparé en deux vecteurs de meme taille via split.
             return tfp.distributions.MultivariateNormalDiag(mu, tf.exp(logstd))
         else:
             return tfp.distributions.Categorical(logits)

reaver/models/sc2/fully_conv.py

@@ -3,7 +3,7 @@
 import tensorflow as tf
 from tensorflow.keras import Model
 from tensorflow.keras.initializers import VarianceScaling
-from tensorflow.keras.layers import Input, Concatenate, Dense, Embedding, Conv2D, Flatten, Lambda
+from tensorflow.keras.layers import Input, Concatenate, Dense, Embedding, Conv2D, Flatten, Lambda, LSTM
 from reaver.models.base.layers import Squeeze, Split, Transpose, Log, Broadcast2D
 
 
@@ -29,12 +29,25 @@ def build_fully_conv(obs_spec, act_spec, data_format='channels_first', broadcast
     value = Squeeze(axis=-1)(value)
 
     logits = []
+    lstm = LSTM(1024)
     for space in act_spec:
-        if space.is_spatial():
-            logits.append(Conv2D(1, 1, **conv_cfg(data_format, scale=0.1))(state))
-            logits[-1] = Flatten()(logits[-1])
-        else:
-            logits.append(Dense(space.size(), **dense_cfg(scale=0.1))(fc))
+        print('space: {}, spatial: {}'.format(space, space.is_spatial()))
+        with tf.name_scope(space.name):
+            if space.is_spatial():
+                logits.append(Conv2D(1, 1, **conv_cfg(data_format, scale=0.1))(state))
+                logits[-1] = Flatten()(logits[-1])
+                #print(logits[-1].shape)
+                from tensorflow.keras.layers import Reshape
+                logits[-1] = Reshape((1,logits[-1].shape[1]))(logits[-1])
+                #logits[-1] = Dense(1024, **dense_cfg(scale=0.1))(logits[-1])
+
+                #state = lstm.zero_state(1, dtype=tf.float32)
+                #output, state = lstm(logits[-1], state)
+                #logits[-1]=output
+                logits[-1] = LSTM(1024)(logits[-1])
+                print(logits[-1].shape)
+            else:
+                logits.append(Dense(space.size(), **dense_cfg(scale=0.1))(fc))
 
     mask_actions = Lambda(
         lambda x: tf.where(non_spatial_inputs[0] > 0, x, -1000 * tf.ones_like(x)),