better keras impl

Author: sethjuarez

LearnDigitz/keras_train.py

@@ -5,7 +5,7 @@
 from tensorflow.keras import Sequential
 from tensorflow.keras import backend as K
 from tensorflow.keras.utils import to_categorical
-from misc.helpers import print_info, print_args, check_dir
+from misc.helpers import print_info, print_args, check_dir, info
 from tensorflow.python.framework import graph_util, graph_io
 from tensorflow.keras.layers import Reshape, Flatten, Dense, Conv2D, MaxPooling2D
 
@@ -14,11 +14,10 @@ def save(model, model_dir):
   print('\nSaving h5 model to {}'.format(m))
   model.save(m)
   print('Saving pb model to {}'.format(os.path.join(model_dir, 'digits.pb')))
-
 
   input_node = model.input.name.split(':')[0]
   output_node = model.output.name.split(':')[0]
-  print("Input Tensor:", input_node)
+  print("\nInput Tensor:", input_node)
   print("Output Tensor:", output_node)
 
   K.set_learning_phase(0)
@@ -37,24 +36,18 @@ def load_digits(data_dir):
   return (x_train, y_train), (x_test, y_test)
 
 ###################################################################
-# Simple (W.T * X + b)                                            #
+# shapes                                                          #
 ###################################################################
 def linear():
   return Sequential([Dense(10)])
 
-###################################################################
-# Neural Network                                                  #
-###################################################################
 def mlp():
   return Sequential([
     Dense(512, activation='relu'),
     Dense(512, activation='relu'),
     Dense(10, activation='softmax')
   ])
 
-###################################################################
-# Convolutional Neural Network                                    #
-###################################################################
 def cnn():
   return Sequential([
     Reshape((28, 28, 1)),
@@ -67,28 +60,37 @@ def cnn():
     Dense(10, activation='softmax')
   ])
 
+
+@print_info
 def run(data_dir, model_dir, epochs):
   # get data
   (x_train, y_train), (x_test, y_test) = load_digits(data_dir)
-
+  
   # create model structure
   model = cnn()
-
+  
   # compile model
   model.compile(loss='mse', optimizer='adam', metrics=['accuracy'])
-
+  
   # run model
   model.fit(x_train, y_train, epochs=epochs)
   model.summary()
-  model.evaluate(x_test, y_test)
-
+  evaluation = model.evaluate(x_test, y_test)
+  
+  # save model
+  info('Output...')
   save(model, model_dir)
+
+  # metrics
+  info('Metrics...')
+  print('Loss:     {}'.format(evaluation[0]))
+  print('Accuracy: {}'.format(evaluation[1]))
 
 
 if __name__ == "__main__":
   data_dir = check_dir(os.path.abspath('data'))
   output_dir = os.path.abspath('output')
   unique = datetime.now().strftime('%m.%d_%H.%M')
   model_dir = check_dir(os.path.join(output_dir, 'models', 'model_{}'.format(unique)))
-  epochs = 5
+  epochs = 10
   run(data_dir, model_dir, epochs)

LearnDigitz/train.py

@@ -105,15 +105,14 @@ def linear_better(x, init=tf.zeros):
         return tf.identity(pred, name="prediction")
 
 @print_info
-def mlp_better(x, hidden=[512, 512]):
-    last_output = x
-    for i in range(len(hidden)):
-        # layer n
-        last_output = tf.layers.dense(inputs=last_output, units=hidden[i], activation=tf.nn.relu)
+def mlp_better(x):
+    # hidden layers
+    h1 = tf.layers.dense(inputs=x, units=512, activation=tf.nn.relu)
+    h2 = tf.layers.dense(inputs=h1, units=512, activation=tf.nn.relu)
 
     # output layer
     with tf.name_scope("Model"):
-        pred = tf.layers.dense(inputs=last_output, units=10, activation=tf.nn.softmax)
+        pred = tf.layers.dense(inputs=h2, units=10, activation=tf.nn.softmax)
         return tf.identity(pred, name="prediction")
 
 @print_info
@@ -136,7 +135,7 @@ def cnn_better(x):
     dense = tf.layers.dense(inputs=pool2_flat, units=1024, activation=tf.nn.relu)
 
     with tf.name_scope('Model'):
-        pred = tf.layers.dense(inputs=dense, units=10)
+        pred = tf.layers.dense(inputs=dense, units=10, activation=tf.nn.softmax)
         return tf.identity(pred, name="prediction")
 
 ###################################################################
@@ -168,18 +167,13 @@ def train_model(x, y, cost, optimizer, accuracy, learning_rate, batch_size, epoc
         acc = 0.
         info('Training')
         # epochs to run
-        with trange(epochs, desc="{:<10}".format("Training"), 
-            bar_format='{l_bar}{bar}|{postfix}', 
-            postfix=" acc: 0.0000") as t:
+        with trange(epochs, desc="{:<10}".format("Training"), bar_format='{l_bar}{bar}|{postfix}', postfix=" acc: 0.0000") as t:
             for epoch in t:
                 avg_cost = 0.
                 t.postfix = ' acc: {:.4f}'.format(acc)
                 t.update()
                 # loop over all batches
-                with tqdm(enumerate(digits), 
-                        total=digits.total, 
-                        desc="{:<10}".format("Epoch {}".format(epoch + 1)),
-                        bar_format='{l_bar}{bar}|{postfix}') as progress:
+                with tqdm(enumerate(digits), total=digits.total, desc="{:<10}".format("Epoch {}".format(epoch + 1)), bar_format='{l_bar}{bar}|{postfix}') as progress:
                     for i, (train_x, train_y) in progress:
                         # Run optimization, cost, and summary
                         _, c, summary = sess.run([optimizer, cost, merged_summary_op],
@@ -239,7 +233,7 @@ def main(settings):
 
     args.data = check_dir(os.path.abspath(args.data))
     args.output = os.path.abspath(args.output)
-    unique = datetime.now().strftime('%m.%d_%H.%M')    
+    unique = datetime.now().strftime('%m.%d_%H.%M')
     args.log = check_dir(os.path.join(args.output, 'logs', 'log_{}'.format(unique)))
     args.model = check_dir(os.path.join(args.output, 'models', 'model_{}'.format(unique)))