Merge pull request #99 from Paeti/Adjustments_Code

Adjustments code(model, trainer, test), Docs written

docs/OurModel.md

@@ -0,0 +1,34 @@
+The OurModel class builds and compiles the convolutional networks for gender or age estimation.
+Depending on the identifier in the construcor either a gender or a age estimation network will be build.
+
+The model from the OurModel class consists of a vgg16 base model with the pre-trained weights from 'imagenet' and 5 additional layers.
+
+So the structure is as follows:
+
+    16 Layer from the vgg16 base model    
+    1 GlobalAveragePooling2D layer
+    2 Dense layer
+
+    1 Dense layer with sigmoid for age estimation
+        or
+    1 Dense layer with softmax for gender estimation
+
+The structure of this model is inspired by the following paper.
+    https://www.vision.ee.ethz.ch/en/publications/papers/proceedings/eth_biwi_01229.pdf
+
+As optimizer the GradientDescentOptimizer from tf.train is use.
+This optimizer makes it possible to load and save the trained model as saved_model file effortlessly. 
+    https://www.tensorflow.org/api_docs/python/tf/train/GradientDescentOptimizer
+
+
+The load_model method offers the possibility to load a saved_model file and compile it allready. 
+So after the model is loaded with this method, it's ready for training.
+
+
+
+
+
+
+
+
+

docs/Train.md

@@ -0,0 +1,12 @@
+This is the main method of the training. Here the training will be started, models can be loaded and/or evaluated.
+
+Requirement for training:
+    Make sure to have the dataset be available in subfolders sorted by their use. So the structure is as follows:    
+        ../dataset/Train
+        ../dataset/Valid
+        ../dataset/Test    
+
+    https://medium.com/@vijayabhaskar96/tutorial-image-classification-with-keras-flow-from-directory-and-generators-95f75ebe5720
+
+
+

docs/Trainer.md

@@ -0,0 +1,27 @@
+The Trainer class does the training for gender or age estimation depending on the identifier given in the constructor of this class.
+
+Loading and preprocessing of the dataset will be done by the ImageDataGenerator from keras.
+All the preprocessing parameters are adjusted by the preprocess_input from keras. These are the parameters which were used for the imagent weights.
+
+The Trainer class takes use of the Cback class, which consits of the callbacks for the training. 
+These callbacks are in use: 
+    EarlyStopping
+    batch_print_callback
+    json_logging_callback
+    cleanup_callback
+
+For further reading:
+https://keras.io/callbacks/
+
+The train method returns the trained model and saves the model in the Trainer instance.
+The trained model will be saved per default. This will be default directory for the saved model:
+
+    /IPNeuronaleNetze/models/GenderWeights for gender
+        or
+    /IPNeuronaleNetze/models/AgeWeights" for age
+
+A .csv file will be created by the evaluate method. Use this .csv file in combination with the parser scripts to interpret the results of the training.
+
+
+
+

docs/parsers.md

@@ -0,0 +1,11 @@
+The files parse_age.py and parse_gender.py are for evualtion of the .csv file created in training.
+    parse_age: returns the average difference in years from test dataset
+    parse_gender: returns the correct predictions in percent from test dataset
+
+Requirement:
+    python 3
+
+How to:
+    python3 parse_age csv.file_name
+    python3 parse_gender csv.file_name
+

docs/tests.md

@@ -0,0 +1,37 @@
+The modelTest class offers a few methods to confirm the proper functionality of the convolutional network you want to train.
+It is advisable to take a small dataset with just a couple of hundreds of pictures. Due to the test environment each single step
+of training will take more ressources of your computer as well as time than it would normally do.
+
+This testclass tests the correct structure of the OurModel class, which will be used for training.
+Also it tests the training itself, to confirm the right functionality within the training. 
+
+Test cases:
+    1. Test whether saved model file will be created after training
+        test_model_get_saved
+
+    2. Test whether model will be saved and loaded properly
+        test_model_get_saved_and_loaded_correctly
+
+    3. Test whether the weigts are changed after one step of training
+        test_one_training_step
+
+    4. Test whether all of the layers of a model are trainable, except the input layers
+        test_model_layers_allTrainable
+
+    5. Test whether Gendermodel is in expected shape
+        test_Gendermodel_layerLength 
+
+    6. Test whether Agemodel is in expected shape
+        test_Agemodel_layerLength
+
+Recommendation:
+    Use the test.py before the actual training.
+
+Note: 
+    Test case 1 and 2 will create a folder with a saved_model file.
+    You might want to delete these folders before the actual training. 
+
+Inspiration for the test cases:
+    https://medium.com/@keeper6928/how-to-unit-test-machine-learning-code-57cf6fd81765
+
+

models/OurModel.py

@@ -1,67 +1,57 @@
-import sys, os
-parent_dir = os.getcwd()
-sys.path.append("/home/ip/IPNeuronaleNetze")
-sys.path.append("/home/ip/IPNeuronaleNetze/trainers")
+import sys
+sys.path.append("/IPNeuronaleNetze")
+sys.path.append("/IPNeuronaleNetze/trainers")
 import tensorflow as tf
-
-from tensorflow.python.keras.models import Sequential
 from tensorflow.python.keras.applications.vgg16 import VGG16
 from tensorflow.python.keras.models import Model
-from tensorflow.python.keras import models
-from tensorflow.python.keras.layers import Flatten, Dense, Dropout, GlobalAveragePooling2D, Input
+from tensorflow.python.keras.layers import Flatten, Dense, GlobalAveragePooling2D, Input
 from tensorflow.python.keras.optimizers import SGD
-import numpy as np
-from models.optimizer.LR_SGD import LR_SGD
 
-from keras.utils import plot_model
+
 
 class OurModel:
     def __init__(self, identifier):
-        self.model = self.buildModel(identifier)              
-
+        self.model = self.build_model(identifier)
+
+    def build_model(self, identifier):        
+        # Load VGG16                    
+        base_model = VGG16(weights='imagenet', include_top=False)
+        fyipg = base_model.output
 
-    def buildModel(self, identifier):   
-        # LOAD VGG16		        
-	base_model = VGG16(weights='imagenet', include_top=False)	
-     	fyipg = base_model.output
-
-        # add a global spatial average pooling layer
-        fyipg = GlobalAveragePooling2D()(fyipg)	        
-        # let's add a fully-connected layer
-	fyipg = Flatten(name ='Flatten1')(fyipg)
+        # Add a global spatial average pooling layer
+        fyipg = GlobalAveragePooling2D()(fyipg)
+        # Let's add a fully-connected layer
+        fyipg = Flatten(name ='Flatten1')(fyipg)
         fyipg = Dense(4096, activation='relu', name='AdditianlLayer1')(fyipg)
         fyipg = Dense(4096, activation='relu', name='AdditianlLayer2')(fyipg)
         if identifier == 1:
             fyipg = Dense(1, activation='sigmoid', name='Predictions')(fyipg)
         else:
             fyipg = Dense(101, activation='softmax', name='Predictions')(fyipg)
-
-        # this is the model we will train       
-	model = Model(inputs = base_model.input , outputs =  fyipg)       
-
-        # Setting the Learning rate multipliers
-        LR_mult_dict = {}
-    	LR_mult_dict['Flatten1'] = 100
-        LR_mult_dict['AdditianLayer1'] = 100
-        LR_mult_dict['AdditianLayer2'] = 100   
-        LR_mult_dict['Predictions'] = 100 
-
-        # Setting optimizer for model        
-        optimizer = LR_SGD(lr=0.0001, momentum=0.9, decay=0.0005, nesterov=True, multipliers = LR_mult_dict)
 
+        # This is the model we will train       
+        model = Model(inputs = base_model.input , outputs =  fyipg)
+        # Setting optimizer for model                
+        optimizer =tf.train.GradientDescentOptimizer(learning_rate = 0.0001)
         # Optimize model for gender- and agemodel
         if identifier == 1:
             model.compile(optimizer=optimizer,
-                                loss='binary_crossentropy', metrics=['mae'])
+                                loss='binary_crossentropy', metrics=['mae','acc'])
         else:
             model.compile(optimizer= optimizer,
-                                loss='categorical_crossentropy', metrics=['mae'])  
-
-        return model
-
-
-    def loadModel(self, filepath):
-	self.model = tf.contrib.saved_model.load_keras_model(filepath)
-	return self.model
-
-
+                                loss='categorical_crossentropy', metrics=['mae', 'acc'])
+        return model                                                                                                                                 
+
+    def load_model(self, filepath, identifier):
+        # Load the save_model file
+        self.model = tf.contrib.saved_model.load_keras_model(filepath)                                                        
+        # Setting optimizer for model                
+        optimizer =tf.train.GradientDescentOptimizer(learning_rate = 0.0001)                                                                                                                              
+        # Optimize model for gender- and agemodel
+        if identifier == 1:
+            self.model.compile(optimizer=optimizer,                                                                                
+                                loss='binary_crossentropy', metrics=['mae','acc'])
+        else:
+            self.model.compile(optimizer= optimizer,                                                                               
+                                loss='categorical_crossentropy', metrics=['mae', 'acc'])                                                                                                                              
+        return self.model