Implementation of resizing codec

ocrolib/lstm.py

@@ -27,7 +27,7 @@
 from __future__ import print_function
 
 import common as ocrolib
-from numpy import (amax, amin, argmax, arange, array, clip, concatenate, dot,
+from numpy import (amax, amin, argmax, arange, array, clip, concatenate, delete, dot,
                    exp, isnan, log, maximum, mean, nan, ones, outer, roll, sum,
                    tanh, tile, vstack, zeros)
 from pylab import (clf, cm, figure, ginput, imshow, newaxis, rand, subplot,
@@ -288,6 +288,19 @@ def __init__(self,Nh,No,initial_range=initial_range,rand=rand):
         self.No = No
         self.W2 = randu(No,Nh+1)*initial_range
         self.DW2 = zeros((No,Nh+1))
+        self.initial_range = initial_range
+    def resizeoutput(self,No, deleted_positions):
+        """resize all matrices to the new codec created by a given charset"""
+        # delete rows for chars that are not necessary
+        W2_temp = delete(self.W2, deleted_positions, axis=0)
+        # enlarge output and weights for extra chars
+        W2 = randu(No, self.Nh + 1) * initial_range
+        # use the trained weights (if --load was used)
+        W2[: len(W2_temp)] = W2_temp
+        self.W2 = W2
+        self.DW2 = zeros((No, self.Nh + 1))
+        self.No = No
+        self.deltas = None
     def ninputs(self):
         return self.Nh
     def noutputs(self):
@@ -578,6 +591,8 @@ def backward(self,deltas):
                     self.DWGI,self.DWGF,self.DWGO,self.DWCI,
                     self.DWIP,self.DWFP,self.DWOP)
         return [s[1:1+ni] for s in self.sourceerr[:n]]
+    def resizeoutput(self, No, deleted_positions):
+        pass
 
 ################################################################
 # combination classifiers
@@ -621,6 +636,9 @@ def weights(self):
         for i,net in enumerate(self.nets):
             for w,dw,n in net.weights():
                 yield w,dw,"Stacked%d/%s"%(i,n)
+    def resizeoutput(self, nout, deleted_positions):
+        self.nets[-1].resizeoutput(nout, deleted_positions)
+        self.deltas = None
 
 class Reversed(Network):
     """Run a network on the time-reversed input."""
@@ -645,6 +663,8 @@ def states(self):
     def weights(self):
         for w,dw,n in self.net.weights():
             yield w,dw,"Reversed/%s"%n
+    def resizeoutput(self, no, deleted_positions):
+        self.net.resizeOutput(no, deleted_positions)
 
 class Parallel(Network):
     """Run multiple networks in parallel on the same input."""
@@ -679,6 +699,9 @@ def weights(self):
         for i,net in enumerate(self.nets):
             for w,dw,n in net.weights():
                 yield w,dw,"Parallel%d/%s"%(i,n)
+    def resizeoutput(self, no, deleted_positions):
+        for net in self.nets:
+            net.resizeoutput(no, deleted_positions)
 
 def MLP1(Ni,Ns,No):
     """An MLP implementation by stacking two `Logreg` networks on top
@@ -839,7 +862,7 @@ def ctc_align_targets(outputs,targets,threshold=100.0,verbose=0,debug=0,lo=1e-5)
     return aligned
 
 def normalize_nfkc(s):
-    return unicodedata.normalize('NFKC',s)
+    return unicodedata.normalize('NFC',s)
 
 def add_training_info(network):
     return network
@@ -933,6 +956,17 @@ def predictString(self,xs):
         "Predict output as a string. This uses codec and normalizer."
         cs = self.predictSequence(xs)
         return self.l2s(cs)
+    def extendCodec(self, codec):
+        """create a codec that exactly fits to ground truth/given codec as parameter"""
+        print("# creating a codec thas fits to the given charset")
+        # add all unknown and new chars to the codec
+        self.codec.extend(codec)
+        # search for chars that should not be in the codec anymore
+        deleted_positions = self.codec.shrink(codec)
+        # let the output fit to the new defined codec
+        self.lstm.resizeoutput(self.codec.size(), deleted_positions)
+        self.No = self.codec.size()
+        return self.codec
 
 class Codec:
     """Translate between integer codes and characters."""
@@ -957,6 +991,33 @@ def decode(self,l):
         "Decode a code sequence into a string."
         s = [self.code2char.get(c,"~") for c in l]
         return s
+    def extend(self, codec):
+        charset = self.code2char.values()
+        size = self.size()
+        counter = 0
+        for c in codec.code2char.values():
+            if not c in charset: # append chars that doesn't appear in the codec
+                self.code2char[size] = c
+                self.char2code[c] = size
+                size += 1
+                counter += 1
+        print("#", counter, " extra chars added")
+    def shrink(self, codec):
+        deleted_positions = []
+        positions = []
+        for number, char in self.code2char.iteritems():
+            if not char in codec.char2code and char != "~":
+                deleted_positions.append(number)
+            else:
+                positions.append(number)
+        charset = [self.code2char[c] for c in sorted(positions)]
+        self.code2char = {}
+        self.char2code = {}
+        for code, char in enumerate(charset):
+            self.code2char[code] = char
+            self.char2code[char] = code
+        print("#", len(deleted_positions), " unnecessary chars deleted")
+        return deleted_positions
 
 ascii_labels = [""," ","~"] + [unichr(x) for x in range(33,126)]
 

ocropus-rtrain

@@ -207,6 +207,11 @@ network.upgrade()
 if network.last_trial%100==99: network.last_trial += 1
 print("# last_trial", network.last_trial)
 
+# required if model was loaded
+if args.codec != []:
+    codec = network.extendCodec(codec)
+else:
+    codec = network.codec
 
 # set up the learning rate
 
@@ -296,7 +301,8 @@ for trial in range(start,args.ntrain):
     except FloatingPointError as e:
         print("# oops, got FloatingPointError", e)
         traceback.print_exc()
-        network = load_lstm(last_save)
+        # TODO: adjust codec if an error occurred
+        # network = load_lstm(last_save)
         continue
     except lstm.RangeError as e:
         continue