Update model.

convert.py

@@ -2,7 +2,7 @@
 
 import tensorflow as tf
 from tensorflow.python.framework import graph_util
-from model import mnist as model
+import train
 
 # A workaround to fix an import issue
 # see https://github.com/tensorflow/tensorflow/issues/15410#issuecomment-352189481
@@ -20,9 +20,9 @@ def main():
 
 def convert(options):
     # Create a model to classify single image
-    x_single = tf.placeholder(tf.float32, [1, model.IMAGE_SIZE, model.IMAGE_SIZE, model.IMAGE_CHANNEL_NUM],
+    x_single = tf.placeholder(tf.float32, [1, train.IMAGE_SIZE, train.IMAGE_SIZE, train.IMAGE_CHANNEL_NUM],
                               name="input_single")
-    y_single = model.inference(x_single)
+    y_single = train.inference(x_single)
     output_single = tf.identity(tf.nn.softmax(y_single, axis=1), name="output_single")
 
     with tf.Session() as sess:

train.py

@@ -6,7 +6,17 @@
 from tensorflow.examples.tutorials.mnist import input_data
 from tensorflow.python.framework import graph_util
 
-from model import mnist as model
+
+IMAGE_SIZE = 28
+IMAGE_CHANNEL_NUM = 1
+CONV_1_SIZE = 6
+CONV_1_DEPTH = 6
+CONV_2_SIZE = 5
+CONV_2_DEPTH = 12
+CONV_3_SIZE = 4
+CONV_3_DEPTH = 24
+FC_SIZE = 200
+OUTPUT_SIZE = 10
 
 LEARNING_RATE_MAX = 0.003
 LEARNING_RATE_MIN = 0.0001
@@ -29,11 +39,57 @@ def main():
     train(mnist_data, options)
 
 
+def get_weight(shape):
+    return tf.get_variable("weight", shape, initializer=tf.truncated_normal_initializer(stddev=0.1))
+
+
+def get_bias(shape):
+    return tf.get_variable("bias", shape, initializer=tf.constant_initializer(0.0))
+
+
+def conv2d(input_tensor, weight, stride):
+    return tf.nn.conv2d(input_tensor, weight, strides=[1, stride, stride, 1], padding="SAME")
+
+
+def inference(input_tensor, regularizer=None):
+    with tf.variable_scope("layer_1_conv"):
+        conv_1_weight = get_weight([CONV_1_SIZE, CONV_1_SIZE, IMAGE_CHANNEL_NUM, CONV_1_DEPTH])
+        conv_1_bias = get_bias([CONV_1_DEPTH])
+        conv_1 = conv2d(input_tensor, conv_1_weight, stride=1)
+        conv_1_activation = tf.nn.relu(tf.nn.bias_add(conv_1, conv_1_bias))
+    with tf.variable_scope("layer_2_conv"):
+        conv_2_weight = get_weight([CONV_2_SIZE, CONV_2_SIZE, CONV_1_DEPTH, CONV_2_DEPTH])
+        conv_2_bias = get_bias([CONV_2_DEPTH])
+        conv_2 = conv2d(conv_1_activation, conv_2_weight, stride=2)
+        conv_2_activation = tf.nn.relu(tf.nn.bias_add(conv_2, conv_2_bias))
+    with tf.variable_scope("layer_3_conv"):
+        conv_3_weight = get_weight([CONV_3_SIZE, CONV_3_SIZE, CONV_2_DEPTH, CONV_3_DEPTH])
+        conv_3_bias = get_bias([CONV_3_DEPTH])
+        conv_3 = conv2d(conv_2_activation, conv_3_weight, stride=2)
+        conv_3_activation = tf.nn.relu(tf.nn.bias_add(conv_3, conv_3_bias))
+        shape = conv_3_activation.get_shape().as_list()
+        nodes = shape[1] * shape[2] * shape[3]
+        conv_3_activation_reshaped = tf.reshape(conv_3_activation, [-1, nodes])
+    with tf.variable_scope("layer_4_fc"):
+        w4 = get_weight([nodes, FC_SIZE])
+        if regularizer is not None:
+            tf.add_to_collection("losses", regularizer(w4))
+        b4 = get_bias([FC_SIZE])
+        a4 = tf.nn.relu(tf.matmul(conv_3_activation_reshaped, w4) + b4)
+    with tf.variable_scope("layer_5_fc"):
+        w5 = get_weight([FC_SIZE, OUTPUT_SIZE])
+        if regularizer is not None:
+            tf.add_to_collection("losses", regularizer(w5))
+        b5 = get_bias([OUTPUT_SIZE])
+        logits = tf.matmul(a4, w5) + b5
+    return logits
+
+
 def train(mnist_data, options):
     x = tf.placeholder(tf.float32, [None, 28, 28, 1], name='x')
     y_ = tf.placeholder(tf.float32, [None, 10], name='y_')
     regularizer = tf.contrib.layers.l2_regularizer(REGULARIZATION_RATE)
-    logits = model.inference(x, regularizer)
+    logits = inference(x, regularizer)
 
     cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=tf.argmax(y_, 1))
     loss = tf.reduce_mean(cross_entropy) + tf.add_n(tf.get_collection("losses"))