我一直在玩TensorFlow Cifar10演示（https://www.tensorflow.org/versions/master/tutorials/deep_cnn/index.html#convolutional-neural-networks），当我在Google内部实施批量大小为1的AlexNet（227x227图像）时，我遇到了以下有趣的输出。演示代码：
2016-01-12 13:19:28.042847: step 0, loss = 208.40 (0.0 examples/sec; 28.081 sec/batch)
2016-01-12 13:19:37.991020: step 10, loss = 106981280.00 (2.3 examples/sec; 0.440 sec/batch)
2016-01-12 13:19:42.422207: step 20, loss = 318574152697306415104.00 (2.2 examples/sec; 0.446 sec/batch)
2016-01-12 13:19:46.824720: step 30, loss = 316031624417051148288.00 (2.3 examples/sec; 0.439 sec/batch)
2016-01-12 13:19:51.238086: step 40, loss = 313512247453630332928.00 (2.2 examples/sec; 0.449 sec/batch)
2016-01-12 13:19:55.676777: step 50, loss = 311012714476067618816.00 (2.3 examples/sec; 0.436 sec/batch)
2016-01-12 13:20:00.056848: step 60, loss = 308535664312269668352.00 (2.3 examples/sec; 0.438 sec/batch)
2016-01-12 13:20:04.442751: step 70, loss = 306075115618981380096.00 (2.3 examples/sec; 0.436 sec/batch)
为什么几批后损失价值如此巨大？网络也似乎没有收敛到任何随机的东西（我训练它以确定照片中的色彩平衡有多好）。我只会发布我更改的代码 - 除了小目录名称更改和图像大小参数外，其余部分可在TensorFlow网站上找到。如果您需要查看更多代码，请告诉我。
def inference(images):
"""Build the CIFAR-10 model.
Args:
images: Images returned from distorted_inputs() or inputs().
Returns:
Logits.
"""
# We instantiate all variables using tf.get_variable() instead of
# tf.Variable() in order to share variables across multiple GPU training runs.
# If we only ran this model on a single GPU, we could simplify this function
# by replacing all instances of tf.get_variable() with tf.Variable().
#
# conv1
with tf.variable_scope('conv1') as scope:
kernel = _variable_with_weight_decay('weights', shape=[11, 11, 3, 96],
stddev=1e-4, wd=0.0)
conv = tf.nn.conv2d(images, kernel, [1, 4, 4, 1], padding='SAME')
biases = _variable_on_cpu('biases', [96], tf.constant_initializer(0.0))
bias = tf.nn.bias_add(conv, biases)
conv1 = tf.nn.relu6(bias, name=scope.name)
_activation_summary(conv1)
# norm1
norm1 = tf.nn.lrn(conv1, 4, bias=1.0, alpha=0.0001, beta=0.75,
name='norm1')
# pool1
pool1 = tf.nn.max_pool(norm1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1],
padding='SAME', name='pool1')
# conv2
with tf.variable_scope('conv2') as scope:
kernel = _variable_with_weight_decay('weights', shape=[5, 5, 96, 256],
stddev=1e-4, wd=0.0)
conv = tf.nn.conv2d(pool1, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [256], tf.constant_initializer(0.1))
bias = tf.nn.bias_add(conv, biases)
conv2 = tf.nn.relu6(bias, name=scope.name)
_activation_summary(conv2)
# norm2
norm2 = tf.nn.lrn(conv2, 4, bias=1.0, alpha=0.0001, beta=0.75,
name='norm2')
# pool2
pool2 = tf.nn.max_pool(norm2, ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool2')
# conv3
with tf.variable_scope('conv3') as scope:
kernel = _variable_with_weight_decay('weights', shape=[3, 3, 256, 384],
stddev=1e-4, wd=0.0)
conv = tf.nn.conv2d(pool2, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [384], tf.constant_initializer(0.1))
bias = tf.nn.bias_add(conv, biases)
conv3 = tf.nn.relu6(bias, name=scope.name)
_activation_summary(conv3)
# conv4
with tf.variable_scope('conv4') as scope:
kernel = _variable_with_weight_decay('weights', shape=[3, 3, 384, 384],
stddev=1e-4, wd=0.0)
conv = tf.nn.conv2d(conv3, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [384], tf.constant_initializer(0.1))
bias = tf.nn.bias_add(conv, biases)
conv4 = tf.nn.relu6(bias, name=scope.name)
_activation_summary(conv4)
# conv5
with tf.variable_scope('conv5') as scope:
kernel = _variable_with_weight_decay('weights', shape=[3, 3, 384, 256],
stddev=1e-4, wd=0.0)
conv = tf.nn.conv2d(conv4, kernel, [1, 1, 1, 1], padding='SAME')
biases = _variable_on_cpu('biases', [256], tf.constant_initializer(0.1))
bias = tf.nn.bias_add(conv, biases)
conv5 = tf.nn.relu6(bias, name=scope.name)
_activation_summary(conv5)
# pool5
pool5 = tf.nn.max_pool(conv5, ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1], padding='SAME', name='pool5')
# local6
with tf.variable_scope('local6') as scope:
# Move everything into depth so we can perform a single matrix multiply.
dim = 1
for d in pool5.get_shape()[1:].as_list():
dim *= d
reshape = tf.reshape(pool5, [FLAGS.batch_size, dim])
weights = _variable_with_weight_decay('weights', shape=[dim, 4096],
stddev=0.04, wd=0.004)
biases = _variable_on_cpu('biases', [4096], tf.constant_initializer(0.1))
local6 = tf.nn.relu_layer(reshape, weights, biases, name=scope.name)
_activation_summary(local6)
# local7
with tf.variable_scope('local7') as scope:
weights = _variable_with_weight_decay('weights', shape=[4096, 4096],
stddev=0.04, wd=0.004)
biases = _variable_on_cpu('biases', [4096], tf.constant_initializer(0.1))
local7 = tf.nn.relu_layer(local6, weights, biases, name=scope.name)
_activation_summary(local7)
# softmax, i.e. softmax(WX + b)
with tf.variable_scope('softmax_linear') as scope:
weights = _variable_with_weight_decay('weights', [4096, NUM_CLASSES],
stddev=1/4096.0, wd=0.0)
biases = _variable_on_cpu('biases', [NUM_CLASSES],
tf.constant_initializer(0.0))
softmax_linear = tf.nn.xw_plus_b(local7, weights, biases, name=scope.name)
_activation_summary(softmax_linear)
return softmax_linear