def train(net, loss_func, optimizer, train_data, valid_data,
n_class, device, model_name, epochs=20):
"""
训练
:param net: AI网络
:param loss_func: loss function
:param optimizer: optimizer
:param train_data: train data set
:param valid_data: valid data set
:param n_class: n种分类
:param device: torch.device CPU or GPU
:param model_name: 用于保存模型权重
:param epochs: 训练多少个EPOCH
:return:
"""
for e in range(1, epochs + 1):
get_logger().info('Epoch: {:02d}'.format(e))
# 一个epoch训练
_epoch_train(net, loss_func, optimizer, train_data, n_class, device, e)
# 每个epoch的参数都保存
save_dir = save_weight(net, model_name, e)
get_logger().info(save_dir) # 日志记录
# 一个epoch验证
_epoch_valid(net, loss_func, valid_data, n_class, device, e)
pass
pass
def _epoch_valid(net, loss_func, data, n_class, device, i_epoch):
"""
一个epoch验证
:param net: AI网络
:param loss_func: loss function
:param data: valid data set
:param n_class: n种分类
:param device: torch.device CPU or GPU
:return: loss, miou
"""
net.to(device)
net.eval() # 验证
total_loss = 0. # 一个epoch验证的loss
total_cm = np.zeros((n_class, n_class)) # ndarray
total_batch_miou = 0.
with torch.no_grad(): # 验证阶段，不需要计算梯度，节省内存
bar_format = '{desc}{postfix}|{n_fmt}/{total_fmt}|{percentage:3.0f}%|{bar}|{elapsed}<{remaining}'
# {desc}{进度条百分比}[{当前/总数}{用时<剩余时间}{自己指定的后面显示的}]
tqdm_data = tqdm(data,
ncols=120, # 进度条宽120列，linux必须指定，否则按照terminal宽度80
bar_format=bar_format, # 进度条格式
desc='Epoch {:02d} Valid'.format(i_epoch)) # 进度条的{desc}
for i_batch, (im, lb) in enumerate(tqdm_data, start=1):
im = im.to(device) # [N,C,H,W] tensor 一个验证batch image
lb = lb.to(device) # [N,H,W] tensor 一个验证batch label
output = net(im) # [N,C,H,W] tensor 前向传播，计算一个验证batch的output
loss = loss_func(output, lb.type(torch.long)) # 计算一个验证batch的loss
batch_loss = loss.detach().item() # detach还是加上吧，
total_loss += batch_loss # 累加验证batch的loss
# 验证的时候不进行反向传播
pred = torch.argmax(F.softmax(output, dim=1), dim=1) # [N,H,W] tensor 将输出转化为dense prediction
batch_cm = get_confusion_matrix(pred.cpu().numpy(),
lb.cpu().numpy(),
n_class) # 计算混淆矩阵并累加
total_cm += batch_cm
batch_miou = get_metrics(batch_cm, metrics='mean_iou')
total_batch_miou += batch_miou
tqdm_str = 'Loss={:.4f}|mIoU={:.4f}|bat_mIoU={:.4f}' # 进度条
tqdm_data.set_postfix_str(
tqdm_str.format(total_loss / i_batch,
get_metrics(total_cm, metrics='mean_iou'),
total_batch_miou / i_batch))
pass
total_loss /= len(data) # 求取一个epoch验证的loss
mean_iou = get_metrics(total_cm, metrics='mean_iou') # float 求mIoU
total_batch_miou /= len(data)
# 记录Valid日志
log_str = ('Valid Loss: {:.4f}|'
'Valid mIoU: {:.4f}|'
'Valid bat_mIoU: {:.4f}')
log_str = log_str.format(total_loss, mean_iou, total_batch_miou)
get_logger().info(log_str)
return total_loss, mean_iou, total_batch_miou
def main():
args = tools.get_args(parser)
config = tools.get_config(args)
tools.init(config)
tb_logger, logger = tools.get_logger(config)
tools.check_dist_init(config, logger)
checkpoint = tools.get_checkpoint(config)
runner = tools.get_model(config, checkpoint)
loaders = tools.get_data_loader(config)
if dist.is_master():
logger.info(config)
if args.mode == 'train':
train(config, runner, loaders, checkpoint, tb_logger)
elif args.mode == 'evaluate':
evaluate(runner, loaders)
elif args.mode == 'calc_flops':
if dist.is_master():
flops = tools.get_model_flops(config, runner.get_model())
logger.info('flops: {}'.format(flops))
elif args.mode == 'calc_params':
if dist.is_master():
params = tools.get_model_parameters(runner.get_model())
logger.info('params: {}'.format(params))
else:
assert checkpoint is not None
from models.dmcp.utils import sample_model
sample_model(config, runner.get_model())
if dist.is_master():
logger.info('Done')
def get_model(model_type, in_channels, n_class, device, load_weight=None):
"""
获取AI网络
:param model_type: 网络类型
:param in_channels: 输入图像通道数
:param n_class: n种分类
:param device: torch.device GPU or CPU
:param load_weight: string已有权重文件的绝对路径，有就加载，默认没有
:return:
"""
if model_type == 'fcn8s':
# raise NotImplementedError
model = FCN8s(n_class)
elif model_type == 'unet_resnet152':
raise NotImplementedError
# model = unet_resnet('resnet152', in_channels, n_class, pretrained=True)
elif model_type == 'deeplabv3p_resnet50':
model = DeepLabV3P('resnet50', in_channels, n_class)
elif model_type == 'deeplabv3p_resnet101':
model = DeepLabV3P('resnet101', in_channels, n_class)
elif model_type == 'deeplabv3p_xception':
model = DeepLabV3P('xception', in_channels, n_class)
else:
raise ValueError('model name error!')
get_logger().info('-' * 32 + str(model_type) + '-' * 32)
model.to(device)
if load_weight is None:
get_logger().info('Load weight is not specified!')
elif os.path.exists(load_weight):
# 有训练好的模型就加载
get_logger().info(load_weight + ' exists! loading...')
wt = torch.load(load_weight, map_location=device)
model.load_state_dict(wt)
else:
get_logger().info(load_weight + ' can not be found!')
pass
return model
def test(net, data, device, resize_to=256, n_class=8, compare=False):
"""
测试
:param net: AI网络
:param data: test dataset
:param device: torch.device GPU or CPU
:param n_class: n种分类
:param compare: 是否生成对比图片
:return:
"""
net.to(device)
net.eval() # 测试
total_cm = np.zeros((n_class, n_class)) # 记录整个测试的混淆矩阵
total_batch_miou = 0. # 累加每张图像的mIoU
offset = 690 # 剪裁690x3384
pair_crop = PairCrop(offsets=(offset, None)) # 剪裁690x3384
pair_resize = PairResize(size=resize_to)
pair_norm_to_tensor = PairNormalizeToTensor(norm=True) # 归一化并正则化
with torch.no_grad(): # 测试阶段，不需要计算梯度，节省内存
bar_format = '{desc}{postfix}|{n_fmt}/{total_fmt}|{percentage:3.0f}%|{bar}|{elapsed}<{remaining}'
# {desc}{进度条百分比}[{当前/总数}{用时<剩余时间}{自己指定的后面显示的}]
tqdm_data = tqdm(data, ncols=120, bar_format=bar_format, desc='Test')
for i_batch, (im, lb) in enumerate(tqdm_data, start=1):
# if i_batch > 1:
# break
im_t, lb_t = pair_crop(im, lb) # PIL Image,PIL Image
im_t, lb_t = pair_resize(im_t, lb_t) # PIL Image,PIL Image
im_t, lb_t = pair_norm_to_tensor(im_t,
lb_t) # [C,H,W]tensor,[H,W]tensor
im_t = im_t.to(device) # [C,H,W]tensor装入GPU
im_t = im_t.unsqueeze(0) # 转换为[N,C,H,W] tensor
output = net(im_t) # 经过模型输出[N,C,H,W] tensor
pred = torch.argmax(F.softmax(output, dim=1),
dim=1) # [N,H,W] tensor
pred = pred.unsqueeze(
1) # [N,C,H,W] tensor, F.interpolate操作图像需要[N,C,H,W] tensor
pred = pred.type(
torch.float
) # 转为float数，F.interpolate只对float类型操作，int，long等都没有实现
pred = F.interpolate(pred,
size=(lb.size[1] - offset, lb.size[0]),
mode='nearest') # pred用nearest差值
pred = pred.type(torch.uint8) # 再转回int类型
pred = pred.squeeze(0).squeeze(0) # [H,W]tensor
pred = pred.cpu().numpy() # [H,W]ndarray
supplement = np.zeros((offset, lb.size[0]),
dtype=np.uint8) # [H,W]ndarray,补充成背景
pred = np.append(
supplement, pred,
axis=0) # 最终的估值，[H,W]ndarray,在H方向cat，给pred补充被剪裁的690x3384
batch_cm = get_confusion_matrix(pred, lb, n_class) # 本张图像的混淆矩阵
total_cm += batch_cm # 累加
if compare: # 生成对比图
fontsize = 16 # 图像文字字体大小
fig, ax = plt.subplots(2, 2, figsize=(20, 15)) # 画布
ax = ax.flatten()
ax[0].imshow(im) # 左上角显示原图
ax[0].set_title('Input Image', fontsize=fontsize) # 标题
ax[1].imshow(LaneSegDataset.decode_rgb(
np.asarray(lb))) # 右上角显示 Grand Truth
ax[1].set_title('Grand Truth', fontsize=fontsize) # 标题
batch_miou = get_metrics(batch_cm,
metrics='mean_iou') # 计算本张图像的mIoU
fig.suptitle('mIoU:{:.4f}'.format(batch_miou),
fontsize=fontsize) # 用mIoU作为大标题
total_batch_miou += batch_miou
mask = (pred != 0).astype(
np.uint8) * 255 # [H,W]ndarray,alpha融合的mask
pred = LaneSegDataset.decode_rgb(pred) # [H,W,C=3]ndarray RGB
ax[3].imshow(pred) # 右下角显示Pred
ax[3].set_title('Pred', fontsize=fontsize) # 标题
mask = mask[..., np.newaxis] # [H,W,C=1]ndarray
pred = np.append(pred, mask,
axis=2) # [H,W,C=4]ndarray，RGB+alpha变为RGBA
im = im.convert('RGBA')
pred = Image.fromarray(pred).convert('RGBA')
im_comp = Image.alpha_composite(im, pred) # alpha融合
ax[2].imshow(im_comp) # 左下角显示融合图像
ax[2].set_title('Pred over Input', fontsize=fontsize) # 标题
plt.subplots_adjust(left=0.01,
bottom=0.01,
right=0.99,
top=0.99,
wspace=0.01,
hspace=0.01) # 调整子图边距间距
plt.savefig('/home/mist/imfolder/pred-{:s}.jpg'.format(
now_str())) # 保存图像
plt.close(fig)
pass
tqdm_str = 'mIoU={:.4f}|bat_mIoU={:.4f}' # 进度条
tqdm_data.set_postfix_str(
tqdm_str.format(get_metrics(total_cm),
total_batch_miou / i_batch))
pass
mean_iou = get_metrics(total_cm) # 整个测试的mIoU
total_batch_miou /= len(data)
logger = get_logger()
msg = ('Test mIoU : {:.4f}|'
'Test bat_mIoU : {:.4f}').format(mean_iou, total_batch_miou)
logger.info(msg)
return mean_iou
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
if __name__ == '__main__':
# Read configures
conf = cp.ConfigParser()
# Project configures
conf.read('./CleverRec.properties', encoding='utf-8')
configs = dict(conf.items('default'))
recommender = configs['recommender']
# Model configures
conf.read(os.path.join('./conf/', recommender + '.properties'),
encoding='utf-8')
configs.update(dict(conf.items('parameters')))
# Get logger
logger = get_logger(configs['log.dir'], recommender)
logger.info('=' * 100)
logger.info('Current model: %s' % recommender)
# Read and preprocess data
if configs['model_type'] == 'ranking':
data = RankingPreprocess(configs, logger)
else:
data = RatingPreprocess(configs, logger)
# tf settings
if configs['gpu.is_gpu']:
os.environ['CUDA_VISIBLE_DEVICES'] = configs['gpu.id']
tf_conf = tf.ConfigProto()
tf_conf.gpu_options.per_process_gpu_memory_fraction = float(
configs['gpu.mem_frac'])
def _epoch_train(net, loss_func, optimizer, data, n_class, device, i_epoch):
"""
一个epoch训练
:param net: AI网络
:param loss_func: loss function
:param optimizer: optimizer
:param data: train data set
:param n_class: n种分类
:param device: torch.device CPU or GPU
:return: loss, miou
"""
net.to(device)
net.train() # 训练
total_loss = 0. # 一个epoch训练的loss
total_cm = np.zeros((n_class, n_class)) # ndarray 一个epoch的混淆矩阵
total_batch_miou = 0.
bar_format = '{desc}{postfix}|{n_fmt}/{total_fmt}|{percentage:3.0f}%|{bar}|{elapsed}<{remaining}'
# {desc}{进度条百分比}[{当前/总数}{用时<剩余时间}{自己指定的后面显示的}]
tqdm_data = tqdm(data,
ncols=120, # 进度条宽120列，linux必须指定，否则按照terminal宽度80
bar_format=bar_format, # 进度条格式
desc='Epoch {:02d} Train'.format(i_epoch)) # 进度条的{desc}
for i_batch, (im, lb) in enumerate(tqdm_data, start=1):
im = im.to(device) # [N,C,H,W] tensor 一个训练batch image
lb = lb.to(device) # [N,H,W] tensor 一个训练batch label
optimizer.zero_grad() # 清空梯度
output = net(im) # [N,C,H,W] tensor 前向传播，计算一个训练batch的output
loss = loss_func(output, lb.type(torch.long)) # 计算一个训练batch的loss
batch_loss = loss.detach().item() # train过程有gradient，必须detach才能读取
total_loss += batch_loss # 累加训练batch的loss
loss.backward() # 反向传播
optimizer.step() # 优化器迭代
pred = torch.argmax(F.softmax(output, dim=1), dim=1) # [N,H,W] tensor 将输出转化为dense prediction，减少一个C维度
batch_cm = get_confusion_matrix(pred.cpu().numpy(),
lb.cpu().numpy(),
n_class) # 计算混淆矩阵并累加
total_cm += batch_cm
batch_miou = get_metrics(batch_cm, metrics='mean_iou')
total_batch_miou += batch_miou
tqdm_str = 'Loss={:.4f}|mIoU={:.4f}|bat_mIoU={:.4f}' # 进度条
tqdm_data.set_postfix_str(
tqdm_str.format(total_loss / i_batch,
get_metrics(total_cm, metrics='mean_iou'),
total_batch_miou / i_batch))
pass
total_loss /= len(data) # float 求取一个epoch的loss
mean_iou = get_metrics(total_cm, metrics='mean_iou') # float 求mIoU
total_batch_miou /= len(data) # 计算所有batch的miou的平均
# 记录Train日志
log_str = ('Train Loss: {:.4f}|'
'Train mIoU: {:.4f}|'
'Train bat_mIoU: {:.4f}')
log_str = log_str.format(total_loss, mean_iou, total_batch_miou)
get_logger().info(log_str)
return total_loss, mean_iou, total_batch_miou
def worker(gpu, ngpus_per_node, args_in):
# init
args = copy.deepcopy(args_in)
jobid = os.environ["SLURM_JOBID"]
procid = int(os.environ["SLURM_PROCID"])
args.gpu = gpu
if args.gpu is not None:
logger_name = "{}.{}-{:d}-{:d}.search.log".format(
args.name, jobid, procid, gpu)
else:
logger_name = "{}.{}-{:d}-all.search.log".format(
args.name, jobid, procid)
logger = tools.get_logger(os.path.join(args.path, logger_name))
if args.dist_url == "env://" and args.rank == -1:
args.rank = int(os.environ["RANK"])
if args.mp_dist:
# For multiprocessing distributed training, rank needs to be the
# global rank among all the processes
args.rank = args.rank * ngpus_per_node + gpu
args.print_params(logger.info)
# get cuda device
device = torch.device('cuda', gpu)
# begin
logger.info("Logger is set - training start")
logger.info('back:{}, dist_url:{}, world_size:{}, rank:{}'.format(
args.dist_backend, args.dist_url, args.world_size, args.rank))
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size,
rank=args.rank)
data_parser = {
'ETTh1': {
'data': 'ETTh1.csv',
'T': 'OT',
'M': [7, 7, 7],
'S': [1, 1, 1],
'MS': [7, 7, 1]
},
'ETTh2': {
'data': 'ETTh2.csv',
'T': 'OT',
'M': [7, 7, 7],
'S': [1, 1, 1],
'MS': [7, 7, 1]
},
'ETTm1': {
'data': 'ETTm1.csv',
'T': 'OT',
'M': [7, 7, 7],
'S': [1, 1, 1],
'MS': [7, 7, 1]
},
'ETTm2': {
'data': 'ETTm2.csv',
'T': 'OT',
'M': [7, 7, 7],
'S': [1, 1, 1],
'MS': [7, 7, 1]
},
'WTH': {
'data': 'WTH.csv',
'T': 'WetBulbCelsius',
'M': [12, 12, 12],
'S': [1, 1, 1],
'MS': [12, 12, 1]
},
'ECL': {
'data': 'ECL.csv',
'T': 'MT_320',
'M': [321, 321, 321],
'S': [1, 1, 1],
'MS': [321, 321, 1]
},
'Solar': {
'data': 'solar_AL.csv',
'T': 'POWER_136',
'M': [137, 137, 137],
'S': [1, 1, 1],
'MS': [137, 137, 1]
},
}
if args.data in data_parser.keys():
data_info = data_parser[args.data]
args.data_path = data_info['data']
args.target = data_info['T']
args.enc_in, args.dec_in, args.c_out = data_info[args.features]
args.s_layers = [
int(s_l) for s_l in args.s_layers.replace(' ', '').split(',')
]
args.detail_freq = args.freq
args.freq = args.freq[-1:]
Exp = Exp_M_Informer
for ii in range(args.itr):
# setting record of experiments
setting = '{}_{}_ft{}_sl{}_ll{}_pl{}_dm{}_nh{}_el{}_dl{}_df{}_at{}_fc{}_eb{}_dt{}_mx{}_{}_{}'.format(
args.model, args.data, args.features, args.seq_len, args.label_len,
args.pred_len, args.d_model, args.n_heads, args.e_layers,
args.d_layers, args.d_ff, args.attn, args.factor, args.embed,
args.distil, args.mix, args.des, ii)
exp = Exp(args) # set experiments
logger.info(
'>>>>>>>start training : {}>>>>>>>>>>>>>>>>>>>>>>>>>>'.format(
setting))
exp.train(ii, logger)
logger.info(
'>>>>>>>testing : {}<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<'.format(
setting))
exp.test(setting, logger)
if args.do_predict:
logger.info(
'>>>>>>>predicting : {}<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<'.
format(setting))
exp.predict(setting, True)
torch.cuda.empty_cache()
