✨ Add some debug of the input frames

arcades/experiment/BaseExperiment.lua

@@ -155,6 +155,7 @@ function class:report()
   _ = self.output.report.text and self:_text_report()
   _ = self.output.report.torch and self:_torch_report()
   _ = self.output.report.graphical and self:_graphical_report()
+  _ = self.output.report.input > 0 and self:_inputs_report()
   return self
 end
 
@@ -509,6 +510,30 @@ function class:_plot_f1_score()
   return self
 end
 
+function class:_inputs_report()
+  self.logger:debug("Saving input frames.")
+  if #self.loop.train.inputs > 0 then
+    image.savePNG(
+      paths.concat(self.save_path, "train_inputs.png"),
+      image.toDisplayTensor({
+        input = self.loop.train.inputs,
+        padding = 6
+      })
+    )
+  end
+  if #self.loop.test.inputs > 0 then
+    image.savePNG(
+      paths.concat(self.save_path, "test_inputs.png"),
+      image.toDisplayTensor({
+        input = self.loop.test.inputs,
+        padding = 6
+      })
+    )
+  end
+  return self
+end
+
+
 --- Save a graphical representation of agent network filters
 -- @return self
 function class:_plot_network_filters()
@@ -587,6 +612,7 @@ function class:_train(steps)
     confusion_matrix = interactions.confusion_matrix,
     interactions = interactions.interactions,
     classification_metrics = classification_metrics,
+    inputs = interactions.inputs,
   }
 end
 
@@ -624,6 +650,7 @@ function class:_test(steps)
     confusion_matrix = interactions.confusion_matrix,
     interactions = interactions.interactions,
     classification_metrics = classification_metrics,
+    inputs = interactions.inputs,
   }
 end
 
@@ -646,6 +673,7 @@ function class:_interact(environment, steps)
   }
   local env_num_actions = #(environment:actions())
   local confusion_matrix = torch.zeros(env_num_actions, env_num_actions)
+  local inputs = {}
 
   for step = 1,steps do
     num_it = step
@@ -672,6 +700,10 @@ function class:_interact(environment, steps)
       confusion_matrix[true_action][action] = confusion_matrix[true_action][action] + 1
     end
 
+    if step > steps - self.output.report.input then
+      table.insert(inputs, state.observation)
+    end
+
     if state.terminal then
       assert(action == 0)
       environment:reset()
@@ -682,6 +714,7 @@ function class:_interact(environment, steps)
     num_it = num_it,
     interactions = interactions,
     confusion_matrix = confusion_matrix,
+    inputs = inputs
   }
 end
 

arcades/utils/argparse.lua

@@ -42,6 +42,7 @@ function class:__init()
             {"field", "text", true, "Produce text/log report", "boolean"},
             {"field", "graphical", true, "Produce graphical report", "boolean"},
             {"field", "torch", true, "Produce torch object report", "boolean"},
+            {"field", "input", 3, "Output some encountered input frames", "number"},
           },
         },
         {"field", "metrics", "", "Which kind of metrics to report", "table",

arcades/utils/setup.lua

@@ -122,18 +122,28 @@ local function _environment_setup(args)
     envs.test = envs.train
   end
 
-  if envs.train and envs.test then
-    local training_feature = envs.train:get_observable_state().observation
-    local testing_feature = envs.test:get_observable_state().observation
-    assert(
-      training_feature:isSameSizeAs(testing_feature),
-      "Training and testing environments must have the same features sizes... Aborting"
-    )
-    assert(
-      #envs.train:actions() == #envs.test:actions(),
-      "Training and testing environments must accept the same range of actions... Aborting"
-    )
-  end
+  local training_feature = envs.train:get_observable_state().observation
+  local testing_feature = envs.test:get_observable_state().observation
+  assert(
+  training_feature:isSameSizeAs(testing_feature),
+  "Training and testing environments must have the same features sizes... Aborting"
+  )
+  assert(
+  #envs.train:actions() == #envs.test:actions(),
+  "Training and testing environments must accept the same range of actions... Aborting"
+  )
+  require('image').savePNG(
+    paths.concat(
+      args.output.path,
+      'training_features.png'
+    ),
+    training_feature)
+  require('image').savePNG(
+    paths.concat(
+      args.output.path,
+      'testing_features.png'
+    ),
+    testing_feature)
 
   return envs
 end