VSAG/vision_sjtu_19
3
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

计算速度优化，摄像头读取逻辑改变，分类器更新。

Browse Source
This commit is contained in:
xinyang
2019-07-28 15:59:01 +08:00
parent a5259cbd1f
commit 5a0fdb30af
24 changed files with 17667 additions and 115529 deletions
Download Patch File Download Diff File Expand all files Collapse all files

90
tools/TrainCNN/backward.py
View File

@@ -1,6 +1,7 @@
#!/usr/bin/python3
print("Preparing...")
import tensorflow as tf
import os
from tqdm import tqdm
import generate
import forward
@@ -35,28 +36,22 @@ def save_bias(fp, val):
print(val[i], file=fp)
def save_para(folder, paras):
with open(folder + "/conv1_w", "w") as fp:
save_kernal(fp, paras[0])
with open(folder + "/conv1_b", "w") as fp:
save_bias(fp, paras[1])
with open(folder + "/conv2_w", "w") as fp:
save_kernal(fp, paras[2])
with open(folder + "/conv2_b", "w") as fp:
save_bias(fp, paras[3])
with open(folder + "/fc1_w", "w") as fp:
save_weight_mat(fp, paras[4])
with open(folder + "/fc1_b", "w") as fp:
save_bias(fp, paras[5])
with open(folder + "/fc2_w", "w") as fp:
save_weight_mat(fp, paras[6])
with open(folder + "/fc2_b", "w") as fp:
save_bias(fp, paras[7])
def save_para(folder, paras, names, info):
os.system("mkdir %s/%s" % (folder, info))
for para, name in zip(paras, names):
fp = open("%s/%s/%s" % (folder, info, name), "w")
if name[-1:] == "b":
save_bias(fp, para)
elif name[:2] == "fc":
save_weight_mat(fp, para)
elif name[:4] == "conv":
save_kernal(fp, para)
fp.close()
STEPS = 60000
BATCH = 50
LEARNING_RATE_BASE = 0.001
STEPS = 50000
BATCH = 30
LEARNING_RATE_BASE = 0.0005
LEARNING_RATE_DECAY = 0.99
MOVING_AVERAGE_DECAY = 0.99
@@ -65,7 +60,7 @@ def train(dataset, show_bar=False):
x = tf.placeholder(tf.float32, [None, generate.SRC_ROWS, generate.SRC_COLS, generate.SRC_CHANNELS])
y_= tf.placeholder(tf.float32, [None, forward.OUTPUT_NODES])
keep_rate = tf.placeholder(tf.float32)
nodes, vars = forward.forward(x, 0.01)
nodes, vars, vars_name = forward.forward(x, 0.01)
y = nodes[-1]
ce = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=tf.argmax(y_, 1))
@@ -101,35 +96,38 @@ def train(dataset, show_bar=False):
_, loss_value, step = sess.run(
[train_op, loss, global_step],
feed_dict={x: images_samples, y_: labels_samples, keep_rate:0.2}
feed_dict={x: images_samples, y_: labels_samples, keep_rate:0.4}
)
if (i-1) % 100 == 0:
if (i-1) % 500 == 0:
test_images, test_labels = dataset.sample_test_sets(5000)
test_acc, output = sess.run([accuracy, y], feed_dict={x: test_images, y_: test_labels, keep_rate:1.0})
output = np.argmax(output, axis=1)
real = np.argmax(test_labels, axis=1)
print("=============test-set===============")
for n in range(forward.OUTPUT_NODES):
print("label: %d, precise: %f, recall: %f" %
(n, np.mean(real[output==n]==n), np.mean(output[real==n]==n)))
train_images, train_labels = dataset.sample_train_sets(5000)
train_acc, output = sess.run([accuracy, y], feed_dict={x: train_images, y_: train_labels, keep_rate:1.0})
output = np.argmax(output, axis=1)
real = np.argmax(train_labels, axis=1)
print("=============train-set===============")
for n in range(forward.OUTPUT_NODES):
print("label: %d, precise: %f, recall: %f" %
(n, np.mean(real[output==n]==n), np.mean(output[real==n]==n)))
print("\n")
if step % 500 == 0:
test_images, test_labels = dataset.sample_test_sets(6000)
test_acc, output = sess.run([accuracy, y], feed_dict={x: test_images, y_: test_labels, keep_rate:1.0})
output = np.argmax(output, axis=1)
real = np.argmax(test_labels, axis=1)
print("=============test-set===============")
for n in range(forward.OUTPUT_NODES):
print("label: %d, precise: %f, recall: %f" %
(n, np.mean(real[output==n]==n), np.mean(output[real==n]==n)))
train_images, train_labels = dataset.sample_train_sets(6000)
train_acc, output = sess.run([accuracy, y], feed_dict={x: train_images, y_: train_labels, keep_rate:1.0})
output = np.argmax(output, axis=1)
real = np.argmax(train_labels, axis=1)
print("=============train-set===============")
for n in range(forward.OUTPUT_NODES):
print("label: %d, precise: %f, recall: %f" %
(n, np.mean(real[output==n]==n), np.mean(output[real==n]==n)))
print("\n")
if train_acc >= 0.99 and test_acc >= 0.99:
vars_val = sess.run(vars)
save_para(
"model",
vars_val,
vars_name,
"steps:%d-train_acc:%f-test_acc:%f" % (step, train_acc, test_acc)
)
bar.set_postfix({"loss": loss_value, "train_acc": train_acc, "test_acc": test_acc})
vars_val = sess.run(vars)
save_para("/home/xinyang/Workspace/RM_auto-aim/tools/para", vars_val)
print("save done!")
# print("save done!")
# pred = sess.run(y, feed_dict={x: test_images, keep_rate:1.0})

44
tools/TrainCNN/forward.py
View File

@@ -29,16 +29,22 @@ def max_pool_2x2(x):
CONV1_KERNAL_SIZE = 5
# 第一层卷积输出通道数
CONV1_OUTPUT_CHANNELS = 8
CONV1_OUTPUT_CHANNELS = 4
# 第二层卷积核大小
CONV2_KERNAL_SIZE = 3
# 第二层卷积输出通道数
CONV2_OUTPUT_CHANNELS = 16
CONV2_OUTPUT_CHANNELS = 6
# 第三层卷积核大小
CONV3_KERNAL_SIZE = 3
# 第三层卷积输出通道数
CONV3_OUTPUT_CHANNELS = 8
# 第一层全连接宽度
FC1_OUTPUT_NODES = 100
FC1_OUTPUT_NODES = 50
# 第二层全连接宽度（输出标签类型数）
FC2_OUTPUT_NODES = 15
@@ -49,6 +55,7 @@ OUTPUT_NODES = FC2_OUTPUT_NODES
def forward(x, regularizer=None, keep_rate=tf.constant(1.0)):
vars = []
vars_name = []
nodes = []
conv1_w = get_weight(
@@ -57,7 +64,10 @@ def forward(x, regularizer=None, keep_rate=tf.constant(1.0)):
conv1_b = get_bias([CONV1_OUTPUT_CHANNELS])
conv1 = tf.nn.relu(tf.nn.bias_add(conv2d(x, conv1_w), conv1_b))
pool1 = avg_pool_2x2(conv1)
print("conv1: ", conv1.shape)
print("pool1: ", pool1.shape)
vars.extend([conv1_w, conv1_b])
vars_name.extend(["conv1_w", "conv1_b"])
nodes.extend([conv1, pool1])
conv2_w = get_weight(
@@ -66,27 +76,41 @@ def forward(x, regularizer=None, keep_rate=tf.constant(1.0)):
conv2_b = get_bias([CONV2_OUTPUT_CHANNELS])
conv2 = tf.nn.relu(tf.nn.bias_add(conv2d(pool1, conv2_w), conv2_b))
pool2 = avg_pool_2x2(conv2)
print("conv2: ", conv2.shape)
print("pool2: ", pool2.shape)
vars.extend([conv2_w, conv2_b])
vars_name.extend(["conv2_w", "conv2_b"])
nodes.extend([conv2, pool2])
pool_shape = pool2.get_shape().as_list()
node = pool_shape[1] * pool_shape[2] * pool_shape[3]
reshaped = tf.reshape(pool2, [-1, node])
conv3_w = get_weight(
[CONV3_KERNAL_SIZE, CONV3_KERNAL_SIZE, CONV2_OUTPUT_CHANNELS, CONV3_OUTPUT_CHANNELS]
)
conv3_b = get_bias([CONV3_OUTPUT_CHANNELS])
conv3 = tf.nn.relu(tf.nn.bias_add(conv2d(pool2, conv3_w), conv3_b))
print("conv3: ", conv3.shape)
vars.extend([conv3_w, conv3_b])
vars_name.extend(["conv3_w", "conv3_b"])
nodes.extend([conv3])
conv_shape = conv3.get_shape().as_list()
node = conv_shape[1] * conv_shape[2] * conv_shape[3]
reshaped = tf.reshape(conv3, [-1, node])
reshaped = tf.nn.dropout(reshaped, keep_rate)
print("reshaped: ", reshaped.shape)
fc1_w = get_weight([node, FC1_OUTPUT_NODES], regularizer)
fc1_b = get_bias([FC1_OUTPUT_NODES])
fc1 = tf.nn.relu(tf.matmul(reshaped, fc1_w) + fc1_b)
fc1 = tf.nn.dropout(fc1, keep_rate)
vars.extend([fc1_w, fc1_b])
vars_name.extend(["fc1_w", "fc1_b"])
nodes.extend([fc1])
fc2_w = get_weight([FC1_OUTPUT_NODES, FC2_OUTPUT_NODES], regularizer)
fc2_b = get_bias([FC2_OUTPUT_NODES])
# fc2 = tf.nn.softmax(tf.matmul(fc1, fc2_w) + fc2_b)
fc2 = tf.matmul(fc1, fc2_w) + fc2_b
fc2 = tf.matmul(fc1, fc2_w) + fc2_b
vars.extend([fc2_w, fc2_b])
vars_name.extend(["fc2_w", "fc2_b"])
nodes.extend([fc2])
return nodes, vars
return nodes, vars, vars_name