Fixed recently-introduced crashing bug on save. Added ability to unsh…

…are convolutional units.

git-svn-id: http://cuda-convnet.googlecode.com/svn/trunk@391 bc73d74b-de6f-cad7-86a4-6ddd3d5d919c

convnet.py

@@ -30,7 +30,7 @@
 from gpumodel import *
 import sys
 import math as m
-from layer import LayerParser, LayerParsingError
+import layer as lay
 from convdata import *
 from os import linesep as NL
 
@@ -47,14 +47,20 @@ def init_model_lib(self):
     def init_model_state(self):
         ms = self.model_state
         if self.load_file:
-            ms['layers'] = LayerParser.parse_layers(self.layer_def, self.layer_params, self, ms['layers'])
+            ms['layers'] = lay.LayerParser.parse_layers(self.layer_def, self.layer_params, self, ms['layers'])
         else:
-            ms['layers'] = LayerParser.parse_layers(self.layer_def, self.layer_params, self)
+            ms['layers'] = lay.LayerParser.parse_layers(self.layer_def, self.layer_params, self)
 
         logreg_name = self.op.get_value('logreg_name')
         if logreg_name:
             self.logreg_idx = self.get_layer_idx(logreg_name, check_type='cost.logreg')
-
+
+        # Convert convolutional layers to local
+        if self.op.get_value('unshare_conv'):
+            for layer in ms['layers']:
+                if layer['type'] == 'conv':
+                    lay.LocalLayerParser.conv_to_local(layer)
+
     def get_layer_idx(self, layer_name, check_type=None):
         try:
             layer_idx = [l['name'] for l in self.model_state['layers']].index(layer_name)
@@ -145,6 +151,7 @@ def get_options_parser(cls):
         op.add_option("multiview-test", "multiview_test", BooleanOptionParser, "Cropped DP: test on multiple patches?", default=0, requires=['logreg_name'])
         op.add_option("crop-border", "crop_border", IntegerOptionParser, "Cropped DP: crop border size", default=4, set_once=True)
         op.add_option("logreg-name", "logreg_name", StringOptionParser, "Cropped DP: logreg layer name", default="")
+        op.add_option("unshare-conv", "unshare_conv", BooleanOptionParser, "Convert all convolutional layers to unshared locally-connected?", default=False)
 
         op.delete_option('max_test_err')
         op.options["max_filesize_mb"].default = 0

include/common/matrix.h

@@ -197,7 +197,7 @@ class Matrix {
      * Use transpose() if you want to get the transpose of this matrix.
      */
     inline void setTrans(bool trans) {
-        assert(!isView());
+        assert(isTrans() == trans || !isView());
         _trans = trans ? CblasTrans : CblasNoTrans;
     }
 

layer.py

@@ -288,6 +288,16 @@ class LocalLayerParser(LayerWithInputParser):
     def __init__(self):
         LayerWithInputParser.__init__(self, num_inputs=1)
 
+    @staticmethod
+    def conv_to_local(layer):
+        for w in ('weights', 'weightsInc'):
+            layer[w] = n.require(n.reshape(n.tile(n.reshape(layer[w], (1, n.prod(layer[w].shape))), (layer['modules'], 1)), (layer['modules'] * layer['channels'] * layer['filterPixels'], layer['filters'])), requirements='C')
+        if layer['sharedBiases']:
+            for b in ('biases', 'biasesInc'):
+                layer[b] = n.require(n.tile(layer[b], (layer['modules'], 1)), requirements='C')
+        layer['type'] = 'local'
+        return layer
+
     def requires_params(self):
         return True
 

shownet.py

@@ -74,7 +74,8 @@ def plot_cost(self):
         test_errors = numpy.row_stack(test_errors)
         test_errors = numpy.tile(test_errors, (1, self.testing_freq))
         test_errors = list(test_errors.flatten())
-        test_errors += [test_errors[-1]] * (len(train_errors) - len(test_errors))
+        test_errors += [test_errors[-1]] * max(0,len(train_errors) - len(test_errors))
+        test_errors = test_errors[:len(train_errors)]
 
         numepochs = len(train_errors) / float(numbatches)
         pl.figure(1)