Merge branch 'develop'

learning/run.sh

@@ -1,15 +1,15 @@
 #!/bin/bash
 
-cd ../src
-make
-make preprocess
-cd ../learning
+# cd ../src
+# make
+# make preprocess
+# cd ../learning
 
-python make_data.py $1
-if [ $? -ne 0 ]
-then
-    exit
-fi
+# python make_data.py $1
+# if [ $? -ne 0 ]
+# then
+#     exit
+# fi
 
 python train_cnn.py
 if [ $? -ne 0 ]

learning/train_cnn.py

@@ -32,7 +32,9 @@ def load_data(input_dirs, nb_dim, pad=3):
         for name in os.listdir(input_dir):
             if not name.endswith('.png'):
                 continue
-            img = cv.imread(input_dir + name, 0)
+            img = cv.imread(os.path.join(input_dir, name), 0)
+            if img is None:
+                continue
 
             # Thresholding
             img = cv.threshold(img, 255 * 0.7, 255, cv.THRESH_BINARY_INV)[1]
@@ -92,7 +94,7 @@ def dump_weights(filename, model, unique_label):
     for label in unique_label:
         b += struct.pack('i', len(label))
         for c in label:
-            b += struct.pack('c', c)
+            b += struct.pack('c', bytes(c, 'utf-8'))
     for layer in model.layers:
         for w in layer.get_weights():
             for v in w.astype(np.float32).reshape(-1):
@@ -148,24 +150,26 @@ def eraser(input_img):
     parser.add_argument('--nb_epoch', type=int, default=200)
     args = parser.parse_args()
 
-    print 'Load data'
+    print ('Load data')
     X, y, unique_label = load_data(args.train_dirs, args.nb_dim)
     n = X.shape[0]
 
-    print 'Split data into train set and validation set'
+    print ('Split data into train set and validation set')
     idx = np.random.permutation(n)
-    n_train = int(n * 0.9) / args.batch_size * args.batch_size
+    n_train = int(int(n * 0.9) / args.batch_size) * args.batch_size
     X_train, y_train = X[idx[:n_train]], y[idx[:n_train]]
     X_valid, y_valid = X[idx[n_train:]], y[idx[n_train:]]
     steps_per_epoch = n_train / args.batch_size
 
-    print 'Build model'
+    print ('Build model')
+    # set image data format to "channels, conv_dim1, conv_dim2, conv_dim3".
+    keras.backend.set_image_data_format('channels_first')
     model = build_model(args.nb_dim, len(unique_label))
     opt = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08)
     model.compile(loss='categorical_crossentropy', optimizer=opt,
                 metrics=['accuracy'])
 
-    print 'Fit'
+    print ('Fit')
     earlystopping = EarlyStopping(monitor='val_loss', patience=1000)
     checkpointer = ModelCheckpoint(filepath=args.dump_prefix + 'weights.hdf5',
                                    verbose=0, save_best_only=True)
@@ -184,13 +188,13 @@ def eraser(input_img):
                         epochs=args.nb_epoch, verbose=1)
     model.load_weights(args.dump_prefix + 'weights.hdf5')
 
-    print 'Dump results to binary'
+    print ('Dump results to binary')
     dump_weights(args.dump_prefix + 'cnn-result.bin', model, unique_label)
 
-    print 'Testing on validation set:', (model.predict_classes(X_valid) == y_valid.argmax(axis=1)).mean()
+    print ('Testing on validation set:', (model.predict_classes(X_valid) == y_valid.argmax(axis=1)).mean())
     for test_dir in args.test_dirs:
         X_test, y_test, unique_label_test = load_data([test_dir], args.nb_dim)
         y_test = unique_label_test[y_test.argmax(axis=1)]
         y_pred = unique_label[model.predict_classes(X_test)]
-        print 'Testing on {}: {}'.format(test_dir, (y_test == y_pred).mean())
+        print ('Testing on {}: {}'.format(test_dir, (y_test == y_pred).mean()))
 

src/main.cpp

@@ -103,7 +103,7 @@ usage()
     printf(" kocr\tweights-file target\t\tRecognize characters in target\n");
     printf("\n");
 
-    printf("\tweights-file: *.txt\n");
+    printf("\tweights-file: *.bin\n");
 #else
     printf(" kocr\timage-list\t\tCreates a database file\n");
     printf("\tdatabase-file\t\tEvaluates a database file\n");