follow comment

recognize_digits/evaluate.py

@@ -30,8 +30,6 @@ def get_best_pass(filename):
 
 filename = sys.argv[1]
 log = get_best_pass(filename)
-predict_error = math.sqrt(float(log[0])) / 2
 classification_accuracy = (1 - float(log[1])) * 100
-print 'Best pass is %s, error is %s, which means predict get error as %f' % (
-    log[2], log[0], predict_error)
+print 'Best pass is %s, testing Avgcost is %s' % (log[2], log[0])
 print 'The classification accuracy is %.2f%%' % classification_accuracy

recognize_digits/load_data.py

@@ -17,24 +17,17 @@
 
 
 def read_data(path, filename):
-    imgf = path + filename + "-images-idx3-ubyte"
-    labelf = path + filename + "-labels-idx1-ubyte"
-    f = open(imgf, "rb")
-    l = open(labelf, "rb")
-
-    f.read(16)
-    l.read(8)
-
-    # Define number of samples for train/test
-    n = 60000 if "train" in filename else 10000
-
-    rows = 28
-    cols = 28
-
-    images = np.fromfile(
-        f, 'ubyte',
-        count=n * rows * cols).reshape(n, rows, cols).astype('float32')
-    labels = np.fromfile(l, 'ubyte', count=n).astype("int")
+    with open(path + filename + "-images-idx3-ubyte",
+              "rb") as f:  # open picture file
+        magic, n, rows, cols = struct.upack(">IIII", f.read(16))
+        images = np.fromfile(
+            f, 'ubyte',
+            count=n * rows * cols).reshape(n, rows, cols).astype('float32')
+
+    with open(path + filename + "-labels-idx1-ubyte",
+              "rb") as l:  # open label file
+        magic, n = struct.upack(">II", l.read(8))
+        labels = np.fromfile(l, 'ubyte', count=n).astype("int")
 
     return images, labels
 

recognize_digits/mlp_mnist.py → recognize_digits/mnist_model.py

@@ -39,13 +39,53 @@
 label_size = 10
 img = data_layer(name='pixel', size=data_size)
 
-# The first fully-connected layer
-hidden1 = fc_layer(input=img, size=128, act=ReluActivation())
-# The second fully-connected layer and the according activation function
-hidden2 = fc_layer(input=hidden1, size=64, act=ReluActivation())
-# The thrid fully-connected layer, note that the hidden size should be 10,
-# which is the number of unique digits
-predict = fc_layer(input=hidden2, size=10, act=SoftmaxActivation())
+
+def softmax_regression(img):
+    predict = fc_layer(input=img, size=10, act=SoftmaxActivation())
+    return predict
+
+
+def multilayer_perceptron(img):
+    # The first fully-connected layer
+    hidden1 = fc_layer(input=img, size=128, act=ReluActivation())
+    # The second fully-connected layer and the according activation function
+    hidden2 = fc_layer(input=hidden1, size=64, act=ReluActivation())
+    # The thrid fully-connected layer, note that the hidden size should be 10,
+    # which is the number of unique digits
+    predict = fc_layer(input=hidden2, size=10, act=SoftmaxActivation())
+    return predict
+
+
+def convolutional_neural_network(img):
+    # first conv layer
+    conv_pool_1 = simple_img_conv_pool(
+        input=img,
+        filter_size=5,
+        num_filters=20,
+        num_channel=1,
+        pool_size=2,
+        pool_stride=2,
+        act=TanhActivation())
+    # second conv layer
+    conv_pool_2 = simple_img_conv_pool(
+        input=conv_pool_1,
+        filter_size=5,
+        num_filters=50,
+        num_channel=20,
+        pool_size=2,
+        pool_stride=2,
+        act=TanhActivation())
+    # The first fully-connected layer
+    fc1 = fc_layer(input=conv_pool_2, size=128, act=TanhActivation())
+    # The softmax layer, note that the hidden size should be 10,
+    # which is the number of unique digits
+    predict = fc_layer(input=fc1, size=10, act=SoftmaxActivation())
+    return predict
+
+
+predict = softmax_regression(img)
+#predict = multilayer_perceptron(img)
+#predict = convolutional_neural_network(img)
 
 if not is_predict:
     lbl = data_layer(name="label", size=label_size)

recognize_digits/mnist_provider.py

@@ -19,30 +19,17 @@
 # Define a py data provider
 @provider(
     input_types={'pixel': dense_vector(28 * 28),
-                 'label': integer_value(10)},
-    cache=CacheType.CACHE_PASS_IN_MEM)
+                 'label': integer_value(10)})
 def process(settings, filename):  # settings is not used currently.
-    imgf = filename + "-images-idx3-ubyte"
-    labelf = filename + "-labels-idx1-ubyte"
-    f = open(imgf, "rb")
-    l = open(labelf, "rb")
+    with open(filename + "-images-idx3-ubyte", "rb") as f:  # open picture file
+        magic, n, rows, cols = struct.upack(">IIII", f.read(16))
+        images = np.fromfile(
+            f, 'ubyte',
+            count=n * rows * cols).reshape(n, rows, cols).astype('float32')
 
-    f.read(16)
-    l.read(8)
-
-    # Define number of samples for train/test
-    if "train" in filename:
-        n = 60000
-    else:
-        n = 10000
-
-    images = numpy.fromfile(
-        f, 'ubyte', count=n * 28 * 28).reshape((n, 28 * 28)).astype('float32')
-    images = images / 255.0 * 2.0 - 1.0
-    labels = numpy.fromfile(l, 'ubyte', count=n).astype("int")
+    with open(filename + "-labels-idx1-ubyte", "rb") as l:  # open label file
+        magic, n = struct.upack(">II", l.read(8))
+        labels = np.fromfile(l, 'ubyte', count=n).astype("int")
 
     for i in xrange(n):
         yield {"pixel": images[i, :], 'label': labels[i]}
-
-    f.close()
-    l.close()

recognize_digits/plot_error.py

@@ -20,15 +20,33 @@
 def plot_log(filename):
     with open(filename, 'r') as f:
         text = f.read()
-        pattern = re.compile('Test.*? cost=([0-9]+\.[0-9]+).*?pass-([0-9]+)',
-                             re.S)
+        pattern = re.compile(
+            'AvgCost=([0-9]+\.[0-9]+).*?Test.*? cost=([0-9]+\.[0-9]+).*?pass-([0-9]+)',
+            re.S)
         results = re.findall(pattern, text)
-        cost, pass_ = zip(*results)
-        cost_float = map(float, cost)
+        train_cost, test_cost, pass_ = zip(*results)
+        train_cost_float = map(float, train_cost)
+        test_cost_float = map(float, test_cost)
         pass_int = map(int, pass_)
-        plt.plot(pass_int, cost_float, 'bo', pass_, cost_float, 'k')
+        plt.plot(pass_int, train_cost_float, 'red', label='Train Error')
+        plt.plot(pass_int, test_cost_float, 'green', label='Test Error')
         plt.ylabel('AvgCost')
         plt.xlabel('epoch')
+
+        # Now add the legend with some customizations.
+        legend = plt.legend(loc='upper center', shadow=True)
+
+        # The frame is matplotlib.patches.Rectangle instance surrounding the legend.
+        frame = legend.get_frame()
+        frame.set_facecolor('0.90')
+
+        # Set the fontsize
+        for label in legend.get_texts():
+            label.set_fontsize('large')
+
+        for label in legend.get_lines():
+            label.set_linewidth(1.5)  # the legend line width
+
         plt.show()
 
 

recognize_digits/predict.py

@@ -57,8 +57,10 @@ def predict(self, index):
         output = self.network.forwardTest(input)
         prob = output[0]["value"]
         predict = np.argsort(-prob)
+        print "Predicted probability of each digit:"
         print prob
-        print predict[0][0], self.labels[index]
+        print "Predict Number: %d" % predict[0][0]
+        print "Actual Number: %d" % self.labels[index]
 
 
 def main():

recognize_digits/train.sh

@@ -14,9 +14,9 @@
 # limitations under the License.
 set -e
 
-config=cnn_mnist.py
-output=./cnn_mnist_model
-log=cnn_train.log
+config=mnist_model.py
+output=./softmax_mnist_model
+log=softmax_train.log