Repaired paths.

Author: tomassykora

eval_results.py

@@ -68,6 +68,6 @@ def euclidean_distance(vects):
 
 model = Model([input_anchor, input_positive, input_negative], [net_anchor,net_positive, net_negative], name='gen')
 
-model.load_weights('./model.h5')
+model.load_weights('./model_weights.h5')
 
 evaluate.test(base_model)

instructions

@@ -7,13 +7,25 @@ Then run
 	./train_test_split.sh
 to split the dataset into training and testing parts
 
-Alternatively, you can skip previous steps and download prepared toy data from http://www.stud.fit.vutbr.cz/~xjonjo00/data_small.tar.gz
+Alternatively, you can skip previous steps and download prepared toy data (small size) from http://www.stud.fit.vutbr.cz/~xjonjo00/data_small.tar.gz
 
+Now you can run training:
 
-Now you can launch training with
-	python3 resnet.py
+To train with hard mining (web api available):
 
-after the model is trained and saved, you can run the sample website:
-	FLASK_APP=web_pova.py flask run
+	launch training with
+		python3 resnet.py
+
+	after the model is trained and saved, you can run the sample website:
+		FLASK_APP=web_pova.py flask run
+
+To train without hard mining (web api not available):
+
+	launch training with
+		python3 model_noHardMin.py
+
+To evaluate saved trained model run:
+
+	python3 eval_results.py
 
 Requirements: Keras, tensorflow, Flask

model_noHardMin.py

@@ -11,6 +11,8 @@
 from triplets_generator_noHardMin import DataGenerator
 import evaluate
 
+datapath="./"
+
 def l2Norm(x):
 	return  K.l2_normalize(x, axis=-1)
 
@@ -87,14 +89,14 @@ def mean_neg_dist(_, y_pred):
 
 """ Training """
 batch_size = 5
-training_generator = DataGenerator(dim_x=224, dim_y=224, batch_size=batch_size, dataset_path='./20_classes').generate()
+training_generator = DataGenerator(dim_x=224, dim_y=224, batch_size=batch_size, dataset_path=os.path.join(datapath,'data_train5')).generate()
 #validation_generator = DataGenerator(dim_x = 224, dim_y = 224, batch_size = batch_size, dataset_path = './places365-dataset/20_classes').generate()
 
 opt = optimizers.Adam(lr=0.0005)
 model.compile(loss=triplet_loss, optimizer=opt, metrics=[accuracy, mean_pos_dist, mean_neg_dist])
 
 model.fit_generator(generator = training_generator,
-					steps_per_epoch = 248300//batch_size,
+					steps_per_epoch = 24500//batch_size,
 					epochs = 1)
 
 model.save_weights("model_1epoch.h5")
@@ -105,14 +107,14 @@ def mean_neg_dist(_, y_pred):
 opt = optimizers.Adam(lr=0.00004)
 model.compile(loss=triplet_loss, optimizer=opt, metrics=[accuracy, mean_pos_dist, mean_neg_dist])
 model.fit_generator(generator = training_generator,
-					steps_per_epoch = 248300//batch_size,
+					steps_per_epoch = 24500//batch_size,
 					epochs = 1)
 
 # serialize model to JSON
 model_json = model.to_json()
 with open("model.json", "w") as json_file:
     json_file.write(model_json)
 # serialize weights to HDF5
-model.save_weights("model.h5")
+model.save_weights("model_weights.h5")
 print("Saved model to disk")
 

resnet.py

@@ -137,6 +137,7 @@ def fake_loss(__,_):
     # serialize weights to HDF5
     base_model.save(os.path.join(datapath,"model.h5"))
     model.save(os.path.join(datapath,"model_triplets.h5"))
+    model.save_weights('model_weights.h5')
 
     print("Saved model to disk")
 

triplets_generator_noHardMin.py

@@ -3,8 +3,6 @@
 from keras.preprocessing import image
 from keras.applications.resnet50 import preprocess_input
 
-import matplotlib.pyplot as plt
-
 class DataGenerator(object):
 	'Generates data for Keras'
 	def __init__(self, dim_x = 224, dim_y = 224, batch_size = 10, dataset_path = './places365-dataset/20_classes'):