Author: Pannous

extensions.py

@@ -227,6 +227,16 @@ def is_dir(x, must_exist=True):
 #   print "\n"
 #   x
 
+
+def exists(x):
+	return os.path.isfile(x)
+
+def is_dir(x):
+	return os.path.isfile(x) and os.path.isdir(x)
+
+def is_file(x):#Is it a file, or a directory?
+	return os.path.isfile(x) and not os.path.isdir(x)
+
 def is_a(self, clazz):
 	if self is clazz: return True
 	try:

mouse_prediction.py

@@ -54,10 +54,10 @@ def get_mouse_position():
     # todo make model robust to extra text
     argmax = numpy.argmax(lines) # most white
     argmin = numpy.argmin(lines) # most black
-    if(argmax<argmin):
-      mat[:,:argmax,:]=1. # fill white above
-    if(argmin<argmax):
-      mat[:,argmax:,:]=1. # fill white below
+    # if(argmax<argmin):
+    #   mat[:,:argmax,:]=1. # fill white above
+    # if(argmin<argmax):
+    #   mat[:,argmax:,:]=1. # fill white below
     # todo: what if invert image!?
 
     tensor = mat
@@ -87,7 +87,7 @@ def get_mouse_position():
     histogram = numpy.histogram(mat, bins=10, range=None, normed=False, weights=None, density=None)
     print(argmax)
 
-    words = predict_tensor([tensor])
+    words = predict_tensor(tensor)
     if len(words) > 0:
       best = words[0]
     else:

text_recognizer.py

@@ -4,15 +4,16 @@
 import sys
 
 import numpy as np
+import keras.models
 from PIL import Image # python -m pip install --upgrade Pillow  # WTF
-from keras.models import load_model
 
 # weight_file = 'best_weights.h5'
 # weight_file = 'current_weights.h5'
 
 # weight_file = 'weights_ascii.h5' # learned on noisy data
 # weight_file = 'weights_ascii_easy.h5' # no freckles
-weight_file = 'weights_ascii_clean.h5' # pure text
+# weight_file = 'weights_ascii_clean.h5' # pure text
+model_file = 'model1000.h5'
 
 chars = u'abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZäöüÄÖÜß0123456789!@#$%^&*()[]{}-_=+\\|"\'`;:/.,?><~ '
 
@@ -22,7 +23,7 @@
 
 def load_model():
   global model
-  model = load_model(weight_file)
+  model = keras.models.load_model(model_file)
   # model.load_weights(weight_file, reshape=True, by_name=True)
 
 def predict_tensor(tensor):
@@ -33,8 +34,9 @@ def predict_tensor(tensor):
     tensor = tensor.transpose((2, 1, 0))  # 4*w*h
     tensor = tensor[:, :, :, np.newaxis]
 
+  print(tensor.shape)
   if not model: load_model()
-  prediction = model.predict(tensor, batch_size=1, verbose=1)
+  prediction = model.predict([tensor], batch_size=1, verbose=1)
   result = decode_results(prediction)
   return result
 
@@ -72,6 +74,5 @@ def decode_results(prediction):
   # image = image.transpose(Image.FLIP_TOP_BOTTOM)
   tensor = np.array(image) / 255.0  # RGBA: h*w*4
   print(tensor.shape)
-
-  words = predict_tensor([tensor])
+  words = predict_tensor(tensor)
   print(words)

train.py

@@ -565,6 +565,7 @@ def train(run_name, start_epoch, stop_epoch, img_w):
         # sgd = Adam()
 
     model = Model(inputs=[input_data, labels, input_length, label_length], outputs=loss_out)
+    model.save(os.path.join(MODEL_DIR, 'model%03d.h5' % (start_epoch + 1)))
 
     # for l in model.layers:
     #     if not "conv" in l.name and not "dense1" in l.name:
@@ -591,6 +592,7 @@ def train(run_name, start_epoch, stop_epoch, img_w):
 def last_epoch():
     maxi=0
     for date in os.listdir(MODEL_DIR):
+        if not os.path.isdir(date): continue
         for f in os.listdir(MODEL_DIR+"/"+date):
             if not f.startswith("weights"): continue
             if len(f)==12: