def basic_mnist_input_corpus(choose_randomly=False, data_dir="/tmp/mnist"):
"""Returns the first image and label from MNIST.
Args:
choose_randomly: a boolean indicating whether to choose randomly.
data_dir: a string giving the location of the original MNIST data.
Returns:
A single image and a single label.
"""
dataset = mnist.train(data_dir)
dataset = dataset.cache().shuffle(buffer_size=50000).batch(100).repeat()
iterator = dataset.make_one_shot_iterator()
images, integer_labels = iterator.get_next()
images = tf.reshape(images, [-1, 28, 28, 1])
# labels = tf.one_hot(integer_labels, 10)
labels = integer_labels
with tf.train.MonitoredTrainingSession() as sess:
image_batch, label_batch = sess.run([images, labels])
if choose_randomly:
idx = random.choice(range(image_batch.shape[0]))
else:
idx = 0
tf.logging.info("Seeding corpus with element at idx: %s", idx)
return image_batch[idx], label_batch[idx]
def main(_):
"""Trains the unstable model."""
dataset = mnist.train(FLAGS.data_dir)
dataset = dataset.cache().shuffle(buffer_size=50000).batch(
100).repeat() # 维持一个50000大小的shuffle buffer，每轮去除100个数据
iterator = dataset.make_one_shot_iterator() # 通过迭代器读取数据，数据输出一次后就丢弃了
images, integer_labels = iterator.get_next() # 获得图片和标签
images = tf.reshape(images, [-1, 28, 28, 1]) # 改变形状，-1表示不知道大小
# x_, integer_labels = iterator.get_next()
# x_ = tf.reshape(x_, [-1, 28, 28, 1])
label_input_tensor = tf.identity(integer_labels) # 标签输入张量
labels = tf.one_hot(label_input_tensor, 10) # 使用独热编码对标签编码
# input_layer, hidden_layer_1, hidden_layer_2, logits, image_input_tensor = classifier(images, init_func) # 分类,得到logits和图片输入张量
logits = output_layer
equality = tf.equal(tf.argmax(logits, 1),
tf.argmax(labels, 1)) # 比较是否相等，argmax返回最大值的索引号
accuracy = tf.reduce_mean(tf.to_float(equality)) # 计算正确率,先将
# This will NaN if abs of any logit >= 88.
bad_softmax = unsafe_softmax(logits)
# This will NaN if max_logit - min_logit >= 88.
bad_cross_entropies = unsafe_cross_entropy(bad_softmax, labels)
loss = tf.reduce_mean(bad_cross_entropies) # 损失函数
# loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits))
optimizer = tf.train.GradientDescentOptimizer(0.01) # 优化器
tf.add_to_collection("input_tensors", image_input_tensor) # 添加输入图片张量
tf.add_to_collection("input_tensors", label_input_tensor) # 添加标签张量
tf.add_to_collection("coverage_tensors", logits) # 输出层的输出
tf.add_to_collection("metadata_tensors",
bad_softmax) # 输出层经过unsafe_softmax后的输出
tf.add_to_collection("metadata_tensors",
bad_cross_entropies) # bad_softmax和labels的交叉熵
tf.add_to_collection("metadata_tensors", logits) # 输出层的输出
train_op = optimizer.minimize(loss) # 训练
saver = tf.train.Saver(keep_checkpoint_every_n_hours=1)
sess = tf.Session()
sess.run(tf.initialize_all_tables())
sess.run(tf.global_variables_initializer())
# train classifier on these images and labels
for idx in range(FLAGS.training_steps):
sess.run(train_op, feed_dict={x_: images.eval()})
if idx % 1000 == 0:
loss_val, accuracy_val = sess.run([loss, accuracy])
print(idx,
":loss: {}, accuracy: {}".format(loss_val, accuracy_val))
# print(len(input_values[0]))
# print(os.path)
saver.save(
sess,
os.path.join(FLAGS.checkpoint_dir, "fuzz_checkpoint"),
global_step=idx,
) # 保存模型
def main(_):
"""Trains the unstable model."""
dataset = mnist.train(FLAGS.data_dir)
dataset = dataset.cache().shuffle(buffer_size=50000).batch(100).repeat()
iterator = tf.data.make_one_shot_iterator(dataset)
images, integer_labels = iterator.get_next()
images = tf.reshape(images, [-1, 28, 28, 1])
label_input_tensor = tf.identity(integer_labels)
labels = tf.one_hot(label_input_tensor, 10)
init_func = tf.random_uniform_initializer(-FLAGS.init_scale,
FLAGS.init_scale)
logits, image_input_tensor = classifier(images, init_func)
equality = tf.math.equal(tf.argmax(logits, 1), tf.argmax(labels, 1))
accuracy = tf.math.reduce_mean(tf.cast(equality, tf.float32))
# This will NaN if abs of any logit >= 88.
bad_softmax = unsafe_softmax(logits)
# This will NaN if max_logit - min_logit >= 88.
bad_cross_entropies = unsafe_cross_entropy(bad_softmax, labels)
loss = tf.compat.v1.reduce_mean(bad_cross_entropies)
optimizer = tf.train.GradientDescentOptimizer(0.01)
tf.compat.v1.add_to_collection("input_tensors", image_input_tensor)
tf.compat.v1.add_to_collection("input_tensors", label_input_tensor)
tf.compat.v1.add_to_collection("coverage_tensors", logits)
tf.compat.v1.add_to_collection("metadata_tensors", bad_softmax)
tf.compat.v1.add_to_collection("metadata_tensors", bad_cross_entropies)
tf.compat.v1.add_to_collection("metadata_tensors", logits)
print('loss: {0}'.format(type(loss)))
train_op = optimizer.minimize(loss)
saver = tf.train.Saver(keep_checkpoint_every_n_hours=1)
sess = tf.compat.v1.Session()
sess.run(tf.compat.v1.initialize_all_tables())
sess.run(tf.compat.v1.global_variables_initializer())
# train classifier on these images and labels
for idx in range(FLAGS.training_steps):
sess.run(train_op)
if idx % 1000 == 0:
loss_val, accuracy_val = sess.run([loss, accuracy])
print("loss: {}, accuracy: {}".format(loss_val, accuracy_val))
saver.save(
sess,
os.path.join(FLAGS.checkpoint_dir, "fuzz_checkpoint"),
global_step=idx,
)
def main(_):
"""Train a model and a sort-of-quantized version."""
dataset = mnist.train(FLAGS.data_dir)
dataset = dataset.cache().shuffle(buffer_size=50000).batch(100).repeat()
iterator = dataset.make_one_shot_iterator()
images, integer_labels = iterator.get_next()
images = tf.reshape(images, [-1, 28, 28, 1])
images = tf.identity(
images) # Return a tensor with the same shape and contents as input.
# Now we construct the model in kind of a goofy way, because this makes （愚蠢的方式）
# quantization easier?
# Sizes of hidden layers
h_0 = 784
h_1 = 200
h_2 = 100
h_3 = 10
# Declaring the weight variables
images_flattened = tf.reshape(images, [-1, h_0]) # 展平图片
w_fc1 = weight_variable([h_0, h_1])
b_fc1 = bias_variable([h_1])
w_fc2 = weight_variable([h_1, h_2])
b_fc2 = bias_variable([h_2])
w_fc3 = weight_variable([h_2, h_3])
b_fc3 = bias_variable([h_3])
# Constructing the classifier from the weight variables
h_fc1 = tf.nn.relu(tf.matmul(images_flattened, w_fc1) + b_fc1)
h_fc2 = tf.nn.relu(tf.matmul(h_fc1, w_fc2) + b_fc2)
h_fc3 = tf.matmul(h_fc2, w_fc3) + b_fc3
logits = h_fc3
# Now I want to construct another classifier w/ quantized weights
images_quantized = tf.cast(images_flattened, tf.float16) # 转化数据格式
w_fc1_quantized = tf.cast(w_fc1, tf.float16)
b_fc1_quantized = tf.cast(b_fc1, tf.float16)
w_fc2_quantized = tf.cast(w_fc2, tf.float16)
b_fc2_quantized = tf.cast(b_fc2, tf.float16)
w_fc3_quantized = tf.cast(w_fc3, tf.float16)
b_fc3_quantized = tf.cast(b_fc3, tf.float16)
# Constructing the classifier from the weight variables
h_fc1_quantized = tf.nn.relu(
tf.matmul(images_quantized, w_fc1_quantized) + b_fc1_quantized)
h_fc2_quantized = tf.nn.relu(
tf.matmul(h_fc1_quantized, w_fc2_quantized) + b_fc2_quantized)
h_fc3_quantized = (tf.matmul(h_fc2_quantized, w_fc3_quantized) +
b_fc3_quantized)
logits_quantized = h_fc3_quantized
labels = tf.one_hot(integer_labels, 10)
equality = tf.equal(tf.argmax(logits, 1), tf.argmax(labels, 1))
accuracy = tf.reduce_mean(tf.to_float(equality))
cross_entropies = tf.nn.softmax_cross_entropy_with_logits(logits=logits,
labels=labels)
loss = tf.reduce_mean(cross_entropies)
optimizer = tf.train.GradientDescentOptimizer(0.01)
tf.add_to_collection("input_tensors", images)
tf.add_to_collection("coverage_tensors", logits)
tf.add_to_collection("coverage_tensors", logits_quantized)
tf.add_to_collection("metadata_tensors", logits)
tf.add_to_collection("metadata_tensors", logits_quantized)
train_op = optimizer.minimize(loss)
saver = tf.train.Saver(keep_checkpoint_every_n_hours=1)
sess = tf.Session()
sess.run(tf.initialize_all_tables())
sess.run(tf.global_variables_initializer())
# train classifier on these images and labels
for idx in range(FLAGS.training_steps):
sess.run(train_op)
if idx % 100 == 0:
loss_val, accuracy_val = sess.run([loss, accuracy])
print("loss: {}, accuracy: {}".format(loss_val, accuracy_val))
saver.save(
sess,
os.path.join(FLAGS.checkpoint_dir, "fuzz_checkpoint"),
global_step=idx,
)