Added some graph tools

scripts/loulou.py

@@ -26,7 +26,7 @@ def feed_forward(X_input: np.ndarray, weights: list, activation_fn: list) -> np.
 
     # Forward loop
     for id, w in enumerate(weights):
-        # Weighted average `z = w^T · x`
+        # Weighted average `z = w · x`
         z = x[-1].dot(w)
         # Activation function `y = g(x)`
         y = activation_fn[id](z)
@@ -85,10 +85,14 @@ def train(weights: list, trainX: np.ndarray, trainY: np.ndarray, testX: np.ndarr
     accuracy = np.mean(prediction == np.argmax(testY, axis=1))
     accuracy_table.append(accuracy)
 
+    initial_cost = 1/len(testY) * np.sum((prediction -
+                                          np.argmax(testY, axis=1)) ** 2)
+    average_cost_table.append(initial_cost)
+
     if reduce_output <= 1:
-        print('Accuracy at epoch 0 :', accuracy)
+        print('Accuracy at epoch 0 :', accuracy, ' cost =', initial_cost)
     elif reduce_output == 2:
-        print(0, accuracy)
+        print(0, accuracy, initial_cost)
 
     if epochs < 0:
         epochs = 99999999999
@@ -135,7 +139,7 @@ def train(weights: list, trainX: np.ndarray, trainY: np.ndarray, testX: np.ndarr
             print('Accuracy at epoch', i+1, ':',
                   accuracy, ' cost =', average_cost)
         if reduce_output == 2:
-            print(i+1, accuracy)
+            print(i+1, accuracy, average_cost)
 
         # Save temp file if set so
         if filename:
@@ -178,7 +182,7 @@ def train(weights: list, trainX: np.ndarray, trainY: np.ndarray, testX: np.ndarr
         utils.save(weights, activations_fn, filename,
                    no_infos, infos, reduce_output)
 
-    return accuracy_table
+    return (accuracy_table, average_cost_table), weights
 
 
 def runTrain(params: dict, architecture: list, file=None) -> dict:
@@ -199,11 +203,13 @@ def runTrain(params: dict, architecture: list, file=None) -> dict:
             architecture, activations_arch, epochs, batch, learning_rate)
 
     # Load data
-    #   TODO do not load arbitrary data
     trX, trY, teX, teY = mnist.load_data()
 
     # Init weights
     weights = [np.random.randn(*w) * 0.1 for w in architecture]
 
     # Train network
-    return train(weights, trX, trY, teX, teY, activations_arch, primes_arch, file, epochs, batch, learning_rate, save_timeout, graph, no_infos, reduce_output)
+    tr, weights = train(weights, trX, trY, teX, teY, activations_arch, primes_arch, file,
+                        epochs, batch, learning_rate, save_timeout, graph, no_infos, reduce_output)
+
+    return tr

scripts/train.py

@@ -78,10 +78,11 @@
 
     params = json.dumps(params)
     try:
-        accuracy = runTrain(params, architecture, file=filename)
+        accuracy, cost = runTrain(params, architecture, file=filename)
     except KeyboardInterrupt:
         print('\nTraining stopped by user')
         exit()
 
     if args.return_json:
         print(accuracy)
+        print(cost)