提交 da613567 编写于 作者: W wangmeng28

Refine SE-ResNeXt

上级 024ffd16
import os
import math
import random
import functools
import numpy as np
......@@ -7,10 +8,6 @@ from PIL import Image, ImageEnhance
random.seed(0)
_R_MEAN = 123.0
_G_MEAN = 117.0
_B_MEAN = 104.0
DATA_DIM = 224
THREAD = 8
......@@ -20,7 +17,8 @@ DATA_DIR = 'ILSVRC2012'
TRAIN_LIST = 'ILSVRC2012/train_list.txt'
TEST_LIST = 'ILSVRC2012/test_list.txt'
img_mean = np.array([_R_MEAN, _G_MEAN, _B_MEAN]).reshape((3, 1, 1))
img_mean = np.array([0.485, 0.456, 0.406]).reshape((3, 1, 1))
img_std = np.array([0.229, 0.224, 0.225]).reshape((3, 1, 1))
def resize_short(img, target_size):
......@@ -46,6 +44,30 @@ def crop_image(img, target_size, center):
return img
def random_crop(img, size, scale=[0.08, 1.0], ratio=[3. / 4., 4. / 3.]):
aspect_ratio = math.sqrt(random.uniform(*ratio))
w = 1. * aspect_ratio
h = 1. / aspect_ratio
bound = min((float(img.size[0]) / img.size[1]) / (w**2),
(float(img.size[1]) / img.size[0]) / (h**2))
scale_max = min(scale[1], bound)
scale_min = min(scale[0], bound)
target_area = img.size[0] * img.size[1] * random.uniform(scale_min,
scale_max)
target_size = math.sqrt(target_area)
w = int(target_size * w)
h = int(target_size * h)
i = random.randint(0, img.size[0] - w)
j = random.randint(0, img.size[1] - h)
img = img.crop((i, j, i + w, j + h))
img = img.resize((size, size), Image.LANCZOS)
return img
def rotate_image(img):
angle = random.randint(-10, 10)
img = img.rotate(angle)
......@@ -75,26 +97,28 @@ def distort_color(img):
return img
def process_image(sample, mode):
def process_image(sample, mode, color_jitter, rotate):
img_path = sample[0]
img = Image.open(img_path)
if mode == 'train':
img = resize_short(img, DATA_DIM + 32)
img = rotate_image(img)
if rotate: img = rotate_image(img)
img = random_crop(img, DATA_DIM)
else:
img = resize_short(img, DATA_DIM)
img = crop_image(img, target_size=DATA_DIM, center=(mode != 'train'))
img = crop_image(img, target_size=DATA_DIM, center=True)
if mode == 'train':
img = distort_color(img)
if color_jitter:
img = distort_color(img)
if random.randint(0, 1) == 1:
img = img.transpose(Image.FLIP_LEFT_RIGHT)
if img.mode != 'RGB':
img = img.convert('RGB')
img = np.array(img).astype('float32').transpose((2, 0, 1))
img = np.array(img).astype('float32').transpose((2, 0, 1)) / 255
img -= img_mean
img /= img_std
if mode == 'train' or mode == 'test':
return img, sample[1]
......@@ -102,7 +126,11 @@ def process_image(sample, mode):
return img
def _reader_creator(file_list, mode, shuffle=False):
def _reader_creator(file_list,
mode,
shuffle=False,
color_jitter=False,
rotate=False):
def reader():
with open(file_list) as flist:
lines = [line.strip() for line in flist]
......@@ -117,13 +145,15 @@ def _reader_creator(file_list, mode, shuffle=False):
img_path = os.path.join(DATA_DIR, line)
yield [img_path]
mapper = functools.partial(process_image, mode=mode)
mapper = functools.partial(
process_image, mode=mode, color_jitter=color_jitter, rotate=rotate)
return paddle.reader.xmap_readers(mapper, reader, THREAD, BUF_SIZE)
def train():
return _reader_creator(TRAIN_LIST, 'train', shuffle=True)
return _reader_creator(
TRAIN_LIST, 'train', shuffle=True, color_jitter=True, rotate=True)
def test():
......
......@@ -35,7 +35,11 @@ def squeeze_excitation(input, num_channels, reduction_ratio):
def shortcut(input, ch_out, stride):
ch_in = input.shape[1]
if ch_in != ch_out:
return conv_bn_layer(input, ch_out, 3, stride)
if stride == 1:
filter_size = 1
else:
filter_size = 3
return conv_bn_layer(input, ch_out, filter_size, stride)
else:
return input
......@@ -109,9 +113,9 @@ def train(learning_rate, batch_size, num_passes, model_save_dir='model'):
avg_cost = fluid.layers.mean(x=cost)
optimizer = fluid.optimizer.Momentum(
learning_rate=learning_rate / batch_size,
learning_rate=learning_rate,
momentum=0.9,
regularization=fluid.regularizer.L2Decay(1e-4 * batch_size))
regularization=fluid.regularizer.L2Decay(1e-4))
opts = optimizer.minimize(avg_cost)
accuracy = fluid.evaluator.Accuracy(input=out, label=label)
......@@ -125,8 +129,8 @@ def train(learning_rate, batch_size, num_passes, model_save_dir='model'):
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
train_reader = paddle.batch(datareader.train(), batch_size=batch_size)
test_reader = paddle.batch(datareader.test(), batch_size=batch_size)
train_reader = paddle.batch(reader.train(), batch_size=batch_size)
test_reader = paddle.batch(reader.test(), batch_size=batch_size)
feeder = fluid.DataFeeder(place=place, feed_list=[image, label])
for pass_id in range(num_passes):
......@@ -153,4 +157,4 @@ def train(learning_rate, batch_size, num_passes, model_save_dir='model'):
if __name__ == '__main__':
train(learning_rate=0.1, batch_size=7, num_passes=100)
train(learning_rate=0.1, batch_size=8, num_passes=100)
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册