def inference():
# build model
model = cnn_model.model(5, FLAGS.charset_size)
# 初始化
sess = tf.Session()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
ckpt = tf.train.latest_checkpoint(FLAGS.model_path)
saver.restore(sess, ckpt)
# label对应字符
char_dict = {}
with open("char_dict2", "rb") as f:
char_dict = pickle.load(f)
img = cv2.imread(FLAGS.test_pic, 0)
img = 255 - img
img = pre_process(img)
cv2.imshow("", img)
cv2.waitKey(0)
img = np.reshape(img, (1, FLAGS.image_size, FLAGS.image_size, 1))
feed_dict = {model['images']: img, model['training_or_not']: False}
out_labels, probs = sess.run([model['index_top_k'], model['val_top_k']],
feed_dict=feed_dict)
print('-' * 30)
print('Top k result')
for i in range(len(out_labels[0])):
print('%d: Predicted val %s, Probality %3f' %
(i + 1, char_dict[out_labels[0][i]], probs[0][i]))
sess.close()
def variableFromId(Id):
image_list = imageListFromId(Id)
if len(image_list) == 0:
a = torch.autograd.Variable(torch.zeros(1, 1, 4096))
b = maxPool(a)
print(Id)
if use_cuda:
return b.cuda(), True
else:
return b, True
features_vectors = []
n = 0
count = 0
for img in image_list:
if n%5 == 0:
img_tensor = preprocess(img)
img_tensor.unsqueeze_(0)
img_variable = Variable(img_tensor, requires_grad=False)
if use_cuda:
img_variable = img_variable.cuda(async=True)
vector = model(img_variable)
features_vectors.append(vector)
count += 1
n += 1
if count > 4:
break
a = torch.stack(features_vectors)
b = maxPool(a)
# print(b.size())
if use_cuda:
return b.cuda(), False
else:
return b, False
def extract_weight():
model = cnn_model.model(5, 3755)
# 初始化
sess = tf.Session()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
ckpt = tf.train.latest_checkpoint(MODEL_PATH)
saver.restore(sess, ckpt)
v = tf.all_variables()
d = {}
for i in range(16):
a = v[i]
print(a.name)
value = sess.run(a)
d[a.name] = value
with open("test.bin", 'wb') as f:
pickle.dump(d, f)
sess.close()
import numpy as np
np.random.seed(10)
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
tf.reset_default_graph()
(train_images, train_labels), (test_images, test_labels)= tf.keras.datasets.mnist.load_data()
train_images, train_labels= helper.reshape_and_onehot(train_images, train_labels)
test_images, test_labels= helper.reshape_and_onehot(test_images, test_labels)
with tf.variable_scope('Input'):
X= tf.placeholder(name= 'x', shape= [None, train_images.shape[1], train_images.shape[2], train_images.shape[3]], dtype= tf.float32)
Y= tf.placeholder(name= 'y', shape= [None, config.num_classes], dtype= tf.float32)
output_logits= cnn_model.model(X)
with tf.variable_scope('Train'):
with tf.variable_scope('Loss'):
loss= tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels= Y, logits= output_logits), name= 'loss')
tf.summary.scalar('loss_summary', loss)
with tf.variable_scope('Optimizer'):
optimizer= tf.train.AdamOptimizer(learning_rate= config.learning_rate, name= 'AdamOpt').minimize(loss)
with tf.variable_scope('Accuracy'):
correct_predictions= tf.equal(tf.argmax(output_logits, 1), tf.argmax(Y, 1), name= 'correct_prediction')
accuracy= (tf.reduce_mean(tf.cast(correct_predictions, tf.float32), name= 'accuracy'))
tf.summary.scalar('accuracy_summary', accuracy)
gridsize = 6
train_b_sets, train_solution_sets = mg_system_data.gen_data(gridsize, 1)
train_b_sets = train_b_sets.reshape((1, gridsize, 1))
train_solution_sets = train_solution_sets.reshape((1, gridsize, 1))
test_b_sets, test_solution_sets = mg_system_data.gen_data(gridsize, 1)
#test_b_sets = test_b_sets.reshape((1, 1, gridsize))
#test_solution_sets = test_solution_sets.reshape((1, 1, gridsize))
print(train_b_sets.shape, test_b_sets.shape)
print('training b set = ', train_b_sets)
print('training u set = ', train_solution_sets)
print('A = ', A.shape)
#print (np.matmul(train_solution_sets, A))
print(train_solution_sets.shape, test_solution_sets.shape)
b, u_, output = cnn_model.model(6)
output = tf.Print(output, [output])
loss = tf.losses.absolute_difference(output, u_)
#optimizer = tf.train.GradientDescentOptimizer(learning_rate=1e-4)
optimizer = tf.train.RMSPropOptimizer(learning_rate=1e-4)
train = optimizer.minimize(loss)
correct = tf.equal(output, u_)
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
sess = tf.Session()
def network_training(b_vals, actual_u_outputs, epoch_num):
for e in range(epoch_num):
def validation_test():
val_data_reader = DataReader(FLAGS.val_data_dir, 1, 1, False)
image_batch, label_batch = val_data_reader.input()
model = cnn_model.model(5, FLAGS.charset_size)
# The op for initializing the variables.
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
# 初始化
sess = tf.Session()
sess.run(init_op)
# Start input enqueue threads.
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# modele saver
saver = tf.train.Saver()
ckpt = tf.train.latest_checkpoint(FLAGS.model_path)
saver.restore(sess, ckpt)
# 字典
char_dict = {}
with open("char_dict2", "rb") as f:
char_dict = pickle.load(f)
# 保存输出结果，错误的情况
out_file = open('wrong_condition.txt', 'w')
step = 0
sum1 = 0
sumk = 0
wrong_num = 1
try:
while not coord.should_stop():
# 图像和标签数据
val_images_batch, val_labels_batch = sess.run(
[image_batch, label_batch])
feed_dict = {
model['images']: val_images_batch,
model['labels']: val_labels_batch,
model['training_or_not']: False
}
out_labels, probs = sess.run(
[model['index_top_k'], model['val_top_k']],
feed_dict=feed_dict)
if out_labels[0][0] != val_labels_batch[0]:
out_file.write('true:' + char_dict[val_labels_batch[0]] + '\n')
out_file.write('topk')
for i in range(len(out_labels[0])):
out_file.write(char_dict[out_labels[0][i]] +
str(probs[0][i]))
out_file.write('\n')
val_images_batch[0] = (val_images_batch[0] + 0.5) * 255
cv2.imwrite('./wrong_pic/' + str(wrong_num) + '.bmp',
val_images_batch[0])
wrong_num += 1
step += 1
if step % 100 == 0:
print('Step %d' % step)
except tf.errors.OutOfRangeError:
print('Done!')
finally:
# When done, ask the threads to stop.
coord.request_stop()
# Wait for threads to finish.
coord.join(threads)
sess.close()
out_file.close()
print('Average: accuracy = %.4f, top k accuracy = %.4f' %
(sum1 / step, sumk / step))
# coding=utf-8
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
from cnn_model import model
#1. 准备model
x_data, y_true, y_predict, loss, train_op, accuracy = model()
#1.1 准备数据
mnist = input_data.read_data_sets("./data",one_hot=True)
#2. 开启会话执行
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
image_batch,label_batch = mnist.train.next_batch(batch_size=50)
for i in range(2000):
_, _loss, _accuracy = sess.run([train_op,loss,accuracy],feed_dict={x_data:image_batch,y_true:label_batch})
print("i:",i," loss:",_loss," acc:",_accuracy)