Merge pull request #740 from wanghaoshuang/model_avg

Add model average option for OCR CTC model

fluid/ocr_recognition/crnn_ctc_model.py

@@ -194,8 +194,19 @@ def ctc_train_net(images, label, args, num_classes):
     optimizer = fluid.optimizer.Momentum(
         learning_rate=args.learning_rate, momentum=args.momentum)
     _, params_grads = optimizer.minimize(sum_cost)
-
-    return sum_cost, error_evaluator, inference_program
+    model_average = None
+    if args.model_average:
+        model_average = fluid.optimizer.ModelAverage(
+            params_grads,
+            args.average_window,
+            min_average_window=args.min_average_window,
+            max_average_window=args.max_average_window)
+    decoded_out = fluid.layers.ctc_greedy_decoder(
+        input=fc_out, blank=num_classes)
+    casted_label = fluid.layers.cast(x=label, dtype='int64')
+    error_evaluator = fluid.evaluator.EditDistance(
+        input=decoded_out, label=casted_label)
+    return sum_cost, error_evaluator, inference_program, model_average
 
 
 def ctc_infer(images, num_classes):

fluid/ocr_recognition/ctc_train.py

@@ -23,6 +23,10 @@ add_arg('momentum',       float, 0.9,    "Momentum.")
 add_arg('rnn_hidden_size',int,   200,    "Hidden size of rnn layers.")
 add_arg('device',         int,   0,      "Device id.'-1' means running on CPU"
                                          "while '0' means GPU-0.")
+add_arg('model_average',     bool,   True,     "Whether to aevrage model for evaluation.")
+add_arg('min_average_window',     int,   10000,     "Min average window.")
+add_arg('max_average_window',     int,   15625,     "Max average window.")
+add_arg('average_window',     float,   0.15,     "Average window.")
 add_arg('parallel',     bool,   True,     "Whether use parallel training.")
 # yapf: disable
 
@@ -40,7 +44,7 @@ def train(args, data_reader=dummy_reader):
     # define network
     images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
     label = fluid.layers.data(name='label', shape=[1], dtype='int32', lod_level=1)
-    sum_cost, error_evaluator, inference_program = ctc_train_net(images, label, args, num_classes)
+    sum_cost, error_evaluator, inference_program, model_average = ctc_train_net(images, label, args, num_classes)
 
     # data reader
     train_reader = data_reader.train(args.batch_size)
@@ -75,10 +79,14 @@ def train(args, data_reader=dummy_reader):
                 sys.stdout.flush()
             batch_id += 1
 
+        with model_average.apply(exe):
             error_evaluator.reset(exe)
             for data in test_reader():
                 exe.run(inference_program, feed=get_feeder_data(data, place))
             _, test_seq_error = error_evaluator.eval(exe)
+            if model_average != None:
+                model_average.restore(exe)
+
             print "\nEnd pass[%d]; Test seq error: %s.\n" % (
                 pass_id, str(test_seq_error[0]))