def __init__(self, num_classes=10):
"""Declare all needed layers."""
self.num_classes = num_classes
try:
# weights_path =None
# weights_path = remote_helper.get_remote_date("https://www.flyai.com/m/v0.2|resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5")
weights_path = remote_helper.get_remote_data(
'https://www.flyai.com/m/v0.8|densenet121_weights_tf_dim_ordering_tf_kernels_notop.h5'
)
# weights_path = remote_helper.get_remote_date('https://www.flyai.com/m/v0.8|densenet201_weights_tf_dim_ordering_tf_kernels_notop.h5')
except OSError:
weights_path = 'imagenet'
# base_model = ResNet50(weights=None, input_shape=input_shape=(img_size[0], img_size[1], 3), include_top=False)
base_model = DenseNet121(weights=weights_path,
include_top=False,
input_shape=(img_size[0], img_size[1], 3))
Inp = Input(shape=(img_size[0], img_size[1], 3))
# x = Conv2D(256,3,
# activation='relu',
# padding='same',
# name='wangyi_conv1')(Inp)
# x = Conv2D(256,5,
# activation='relu',
# padding='same',
# name='wangyi_conv2')(x)
# x = MaxPooling2D((2, 2), strides=(1, 1), name='wangyi_pool')(x)
# x =Flatten()(x)
# x = Conv2D(3,7,
# activation='relu',
# padding='same',
# name='wangyi_conv3')(x)
# 增加定制层
x = base_model(Inp)
# x = base_model.output
# x = GlobalAveragePooling2D()(x)
# x = Flatten(name='flatten_1')(x)
# 冻结不打算训练的层。
# print('base_model.layers', len(base_model.layers))
# for i, layer in enumerate(base_model.layers):
# print(i, layer.name)
#
# for layer in base_model.layers[:]:
# layer.trainable = False
# print(layer)
x = GlobalAveragePooling2D()(x)
# x = Flatten(name='flatten_1')(x)
# x = Dense(2048, activation='relu' )(x)
predictions = Dense(num_classes, activation="softmax")(x)
# 创建最终模型
self.model_cnn = keras_model(inputs=Inp, outputs=predictions)
def get_pre_train_model_path(self):
"""FLYAI框架下获取预训练模型的PATH"""
path = remote_helper.get_remote_data('https://www.flyai.com/m/CDial-GPT2_LCCC-base.zip')
if platform.system() == "Linux":
path_list = path.split('/')
path = "/".join(path_list[:-1])
elif platform.system() == "Windows":
path_list = path.split('\\')
path = "\\".join(path_list[:-1])
logger.info("Model path: {}".format(path))
return path
def __init__(self, preTrained_file='sgns.weibo.bigram-char'):
wordVecURL = 'https://www.flyai.com/m/sgns.weibo.word.bz2'
path = remote_helper.get_remote_data(wordVecURL)
with open('./data/input/model/sgns', 'r') as f:
for line in f:
line = line.strip()
if not line:
continue
index = line.find(' ')
word = line[:index]
vector = np.array(line[index:].split(), dtype='float32')
self.embeddings[word] = vector
def train(self):
'''
训练模型，必须实现此方法
:return:
'''
dataset = FacialBeautyDataset()
model = osnet_x1_0(num_classes=1,
pretrained=True,
loss='smoothL1Loss',
use_gpu=True)
# load_pretrained_weights(model, './weights/pretrained/osnet_x1_0_imagenet.pth')
path = remote_helper.get_remote_data(
'https://www.flyai.com/m/osnet_x1_0_imagenet.pth')
load_pretrained_weights(model, path)
model = model.cuda()
optimizer = build_optimizer(model)
max_epoch = args.EPOCHS
batch_size = args.BATCH
scheduler = build_scheduler(optimizer,
lr_scheduler='cosine',
max_epoch=max_epoch)
criterion = nn.SmoothL1Loss()
model.train()
train_loader = DataLoader(dataset=dataset,
batch_size=batch_size,
shuffle=True)
cudnn.benchmark = True
for epoch in range(max_epoch):
for index, data in enumerate(train_loader):
im, label = data
im = im.cuda()
label = label.float().cuda()
optimizer.zero_grad()
out = model(im)
loss = criterion(out, label)
loss.backward()
optimizer.step()
if index % 100 == 0:
print("Epoch: [{}/{}][{}/{}] Loss {:.4f}".format(
epoch + 1, max_epoch, index + 1, len(train_loader),
loss))
scheduler.step()
torch.save(model.state_dict(), 'last.pth')
import sys
import os
from model import KERAS_MODEL_NAME
import WangyiUtilOnFlyai as wangyi
from flyai.utils import remote_helper
from time import clock
from processor import img_size
'''
设置项目的超级参数
'''
try:
# weights_path =None
# weights_path = remote_helper.get_remote_date("https://www.flyai.com/m/v0.2|resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5")
weights_path = remote_helper.get_remote_data(
'https://www.flyai.com/m/v0.8|densenet121_weights_tf_dim_ordering_tf_kernels_notop.h5'
)
# weights_path = remote_helper.get_remote_date('https://www.flyai.com/m/v0.8|densenet201_weights_tf_dim_ordering_tf_kernels_notop.h5')
except OSError:
weights_path = 'imagenet'
parser = argparse.ArgumentParser()
parser.add_argument("-e",
"--EPOCHS",
default=50,
type=int,
help="train epochs")
parser.add_argument("-b", "--BATCH", default=8, type=int, help="batch size")
args = parser.parse_args()
MODEL_PATH_FILE = os.path.join(MODEL_PATH, KERAS_MODEL_NAME)
def __init__(self):
path = remote_helper.get_remote_data(
"https://www.flyai.com/m/uncased_L-12_H-768_A-12.zip")
# path = remote_helper.get_remote_date("https://www.flyai.com/m/uncased_L-24_H-1024_A-16.zip")
data_root = os.path.splitext(path)[0]
bert_config_file = os.path.join(data_root, 'bert_config.json')
bert_config = modeling.BertConfig.from_json_file(bert_config_file)
init_checkpoint = os.path.join(data_root, 'bert_model.ckpt')
bert_vocab_file = os.path.join(data_root, 'vocab.txt')
self.input_ids = tf.placeholder(tf.int32,
shape=[None, None],
name='input_ids')
self.input_mask = tf.placeholder(tf.int32,
shape=[None, None],
name='input_masks')
self.segment_ids = tf.placeholder(tf.int32,
shape=[None, None],
name='segment_ids')
self.labels = tf.placeholder(tf.int32, shape=[
None,
], name='labels')
self.is_training = tf.placeholder_with_default(False,
shape=(),
name='is_training')
self.learning_rate = tf.placeholder_with_default(config.learning_rate,
shape=(),
name='learning_rate')
self.global_step = tf.Variable(0, trainable=False, name='global_step')
# 创建bert模型
with tf.name_scope('Bert'):
model = modeling.BertModel(
config=bert_config,
is_training=True,
input_ids=self.input_ids,
input_mask=self.input_mask,
token_type_ids=self.segment_ids,
# 这里如果使用TPU 设置为True，速度会快些。使用CPU 或GPU 设置为False ，速度会快些。
use_one_hot_embeddings=False)
# 这个获取每个token的output 输入数据[batch_size, seq_length, embedding_size] 如果做seq2seq 或者ner 用这个
# output_layer = model.get_sequence_output()
tvars = tf.trainable_variables()
# 加载BERT模型
(assignment_map, initialized_variable_names) = \
modeling.get_assignment_map_from_checkpoint(tvars, init_checkpoint)
tf.train.init_from_checkpoint(init_checkpoint, assignment_map)
output_layer = model.get_pooled_output() # 这个获取句子的output
hidden_size = output_layer.shape[-1].value # 获取输出的维度
# 构建W 和 b
output_weights = tf.get_variable(
"output_weights", [hidden_size, config.num_labels],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable("output_bias", [config.num_labels],
initializer=tf.zeros_initializer())
with tf.variable_scope("predict"):
if self.is_training is True:
# I.e., 0.1 dropout
output_layer = tf.nn.dropout(output_layer, keep_prob=0.5)
# logits = tf.matmul(output_layer, output_weights)
logits = tf.matmul(output_layer, output_weights)
logits = tf.nn.bias_add(logits, output_bias)
# probabilities = tf.nn.softmax(logits, axis=-1)
log_probs = tf.nn.log_softmax(logits, axis=-1)
self.pred = tf.argmax(log_probs, 1, name='pred')
with tf.name_scope("accuracy"):
# 准确率
correct_pred = tf.equal(self.labels, tf.cast(self.pred, tf.int32))
self.accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32),
name='acc')
with tf.name_scope("loss"):
# 将label进行onehot转化
one_hot_labels = tf.one_hot(self.labels,
depth=config.num_labels,
dtype=tf.float32)
# 构建损失函数
per_example_loss = -tf.reduce_sum(one_hot_labels * log_probs,
axis=-1)
self.loss = tf.reduce_mean(per_example_loss)
# # 优化器
# self.train_op = tf.train.AdamOptimizer(learning_rate=model_config.learning_rate).minimize(self.loss)
with tf.name_scope('optimize'):
optimizer = tf.train.AdamOptimizer(self.learning_rate)
gradients, variables = zip(*optimizer.compute_gradients(self.loss))
gradients, _ = tf.clip_by_global_norm(gradients, config.grad_clip)
self.train_op = optimizer.apply_gradients(
zip(gradients, variables), global_step=self.global_step)