smallish bug fixes

ocrolib/lstm.py

@@ -1,3 +1,4 @@
+import common as ocrolib
 import pdb
 from pylab import *
 import sys
@@ -368,7 +369,8 @@ def forward(self,xs):
         ni,ns,na = self.dims
         assert len(xs[0])==ni
         n = len(xs)
-        if n>len(self.gi): raise RangeError("input too large")
+        if n>len(self.gi): 
+	    raise ocrolib.RecognitionError("input too large for LSTM model")
         self.last_n = n
         self.reset(n)
         for t in range(n):
@@ -398,7 +400,8 @@ def forward(self,xs):
     def backward(self,deltas):
         """Perform backward propagation of deltas."""
         n = len(deltas)
-        if n>len(self.gi): raise RangeError("input too large")
+        if n>len(self.gi): 
+	    raise ocrolib.RecognitionError("input too large")
         assert n==self.last_n
         ni,ns,na = self.dims
         for t in reversed(range(n)):
@@ -595,7 +598,8 @@ def log_mul(x,y):
 
 def log_add(x,y):
     "Perform addition in the log domain."
-    return where(abs(x-y)>10,maximum(x,y),log(exp(x-y)+1)+y)
+    #return where(abs(x-y)>10,maximum(x,y),log(exp(x-y)+1)+y)
+    return where(abs(x-y)>10,maximum(x,y),log(exp(clip(x-y,-20,20))+1)+y)
 
 def forward_algorithm(match,skip=-5.0):
     """Apply the forward algorithm to an array of log state
@@ -631,9 +635,11 @@ def ctc_align_targets(outputs,targets,threshold=100.0,verbose=0,debug=0,lo=1e-5)
     assert not isnan(lmatch).any()
     both = forwardbackward(lmatch)
     epath = exp(both-amax(both))
-    epath /= sum(epath,axis=0)[newaxis,:]
+    l = sum(epath,axis=0)[newaxis,:]
+    epath /= where(l==0.0,1e-9,l)
     aligned = maximum(lo,dot(epath,targets))
-    aligned /= sum(aligned,axis=1)[:,newaxis]
+    l = sum(aligned,axis=1)[:,newaxis]
+    aligned /= where(l==0.0,1e-9,l)
     if debug:
         subplot(413); imshow(epath.T,cmap=cm.hot,interpolation='nearest')
         subplot(414); imshow(aligned.T,cmap=cm.hot,interpolation='nearest')

ocropus-gtedit

@@ -33,7 +33,7 @@ p_org.add_argument('files',nargs='+')
 
 p_html = subparsers.add_parser("html",help="generate html")
 p_html.add_argument('-d','--debug',action="store_true")
-p_html.add_argument('-o','--output',default="correct.html")
+p_html.add_argument('-o','--output',default="correction.html")
 p_html.add_argument('-x','--extension',default='.txt')
 p_html.add_argument('-f','--fontsize',default=14,type=int)
 p_html.add_argument('-H','--height',default=24,type=int)

ocropus-hocr

@@ -59,11 +59,13 @@ dirs = [ocrolib.allsplitext(name)[0] for name in args.files]
 xhfiles = python.sum([glob.glob(d+"/??????.xheight") for d in dirs],[])
 if len(xhfiles)>5:
     xheights = [float(ocrolib.read_text(f)) for f in xhfiles]
-    median_xheight = median(xheights)
+    if len(xheights)>0:
+        median_xheight = median(xheights)
 else:    
     lfiles = python.sum([glob.glob(d+"/??????.bin.png") for d in dirs],[])
     pyrandom.shuffle(lfiles)
-    median_xheight = 0.5*median([imread(f).shape[0] for f in lfiles[:100]])
+    if len(lfiles)>0:
+        median_xheight = 0.5*median([imread(f).shape[0] for f in lfiles[:100]])
 E("median_xheight",median_xheight)
 
 P(hocr.header())

ocropus-rpred

@@ -180,6 +180,9 @@ def process1(arg):
 def safe_process1(arg):
     try:
         return process1(arg)
+    except ocrolib.RecognitionError as e:
+        print e
+        return None
     except:
         traceback.print_exc()
         return None

ocropus-rtrain

@@ -7,7 +7,6 @@ import os.path
 import glob
 import ocrolib
 import argparse
-import lstm
 import scipy
 import matplotlib
 import numpy
@@ -218,6 +217,9 @@ for trial in range(start,args.ntrain):
     else:
         assert "dew.png" in fname,"input must already be dewarped"
 
+    if line.size<10 or amax(line)==amin(line):
+        print "EMPTY-INPUT"
+        continue
     line = line * 1.0/amax(line)
     line = amax(line)-line
     line = line.T