About to make a few modifications to DRY

kastnerkyle · Jan 28, 2014 · 35abf19 · 35abf19
1 parent 9192428
commit 35abf19
Show file tree

Hide file tree

Showing 4 changed files with 8 additions and 10 deletions.
diff --git a/convertscript.sh b/convertscript.sh
@@ -1,3 +1,3 @@
 #!/bin/bash
 #Make sure the convert program is installed sudo apt-get install imagemagick
-for i in `ls *.jpg`; do convert $i -resize 512x512\! "${i%.jpg}.png"; done
+for i in `ls *.jpg`; do convert $i -resize 221x221\! "${i%.jpg}.png"; done
diff --git a/kaggle_dataset.py → kaggle_dataset_decaf.py b/kaggle_dataset.py → kaggle_dataset_decaf.py
diff --git a/kaggle_train.py b/kaggle_train.py
@@ -50,14 +50,14 @@ def to_one_hot(l):
 
 output = mlp.HingeLoss(n_classes=2,
                        layer_name='y',
-                       irange=.0001)
+                       irange=.001)
 
 layers = [l1, l2, l3, l4, output]
 
 mdl = mlp.MLP(layers,
               input_space=in_space)
 
-lr = .001
+lr = .01
 epochs = 100
 trainer = sgd.SGD(learning_rate=lr,
                   batch_size=128,
@@ -82,12 +82,8 @@ def to_one_hot(l):
 
 decay = sgd.LinearDecayOverEpoch(start=1,
                                  saturate=100,
-                                 decay_factor=.05 * lr)
+                                 decay_factor=lr*.05)
 
-win = window_flip.WindowAndFlipC01B(pad_randomized=8,
-                                    window_shape=(32, 32),
-                                    randomize=[trn],
-                                    center=[tst])
 experiment = Train(dataset=trn,
                    model=mdl,
                    algorithm=trainer,

diff --git a/kaggle_train_full.py b/kaggle_train_full.py
@@ -53,13 +53,15 @@ def to_one_hot(l):
 mdl = mlp.MLP(layers,
               input_space=in_space)
 
+lr = .01
+epochs = 100
 trainer = sgd.SGD(learning_rate=.01,
                   batch_size=128,
                   learning_rule=learning_rule.Momentum(.5),
                   # Remember, default dropout is .5
                   cost=Dropout(input_include_probs={'l1': .8},
                                input_scales={'l1': 1.}),
-                  termination_criterion=EpochCounter(100),
+                  termination_criterion=EpochCounter(epochs),
                   monitoring_dataset={'train': full})
 
 watcher = best_params.MonitorBasedSaveBest(
@@ -72,7 +74,7 @@ def to_one_hot(l):
 
 decay = sgd.LinearDecayOverEpoch(start=1,
                                  saturate=250,
-                                 decay_factor=.0005)
+                                 decay_factor=lr*.05)
 experiment = Train(dataset=full,
                    model=mdl,
                    algorithm=trainer,