Skip to content

P
PaddleSeg

PaddlePaddle / PaddleSeg

通知 285
Star 8
Fork 1
P
PaddleSeg
未验证 提交 3f658a36 编写于 作者: M michaelowenliu 提交者： GitHub
浏览文件
操作

Merge pull request #396 from michaelowenliu/develop 

redesign deeplab model
上级 23d69271 7dac6524
隐藏空白更改
内联 并排
Showing with 412 addition and 338 deletion
dygraph/paddleseg/core/seg_train.py
浏览文件 @ 3f658a36
......@@ -87,7 +87,8 @@ def seg_train(model,
out_labels = ["loss", "reader_cost", "batch_cost"]
base_logger = callbacks.BaseLogger(period=log_iters)
train_logger = callbacks.TrainLogger(log_freq=log_iters)
model_ckpt = callbacks.ModelCheckpoint(save_dir, save_params_only=False, period=save_interval_iters)
model_ckpt = callbacks.ModelCheckpoint(
save_dir, save_params_only=False, period=save_interval_iters)
vdl = callbacks.VisualDL(log_dir=os.path.join(save_dir, "log"))
cbks_list = [base_logger, train_logger, model_ckpt, vdl]
......@@ -120,7 +121,7 @@ def seg_train(model,
iter += 1
if iter > iters:
break
logs["reader_cost"] = timer.elapsed_time()
############## 2 ################
cbks.on_iter_begin(iter, logs)
......@@ -136,7 +137,7 @@ def seg_train(model,
loss = ddp_model.scale_loss(loss)
loss.backward()
ddp_model.apply_collective_grads()
else:
logits = model(images)
loss = loss_computation(logits, labels, losses)
......@@ -148,7 +149,7 @@ def seg_train(model,
model.clear_gradients()
logs['loss'] = loss.numpy()[0]
logs["batch_cost"] = timer.elapsed_time()
############## 3 ################
......@@ -159,4 +160,6 @@ def seg_train(model,
############### 4 ###############
cbks.on_train_end(logs)
#################################
\ No newline at end of file
#################################
dygraph/paddleseg/core/val.py
浏览文件 @ 3f658a36
......@@ -67,7 +67,7 @@ def evaluate(model,
pred = pred[np.newaxis, :, :, np.newaxis]
pred = pred.astype('int64')
mask = label != ignore_index
# To-DO Test Execution Time
conf_mat.calculate(pred=pred, label=label, ignore=mask)
_, iou = conf_mat.mean_iou()
......
dygraph/paddleseg/cvlibs/callbacks.py
浏览文件 @ 3f658a36
......@@ -13,7 +13,6 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import time
......@@ -24,6 +23,7 @@ from visualdl import LogWriter
from paddleseg.utils.progbar import Progbar
import paddleseg.utils.logger as logger
class CallbackList(object):
"""Container abstracting a list of callbacks.
# Arguments
......@@ -44,7 +44,7 @@ class CallbackList(object):
def set_model(self, model):
for callback in self.callbacks:
callback.set_model(model)
def set_optimizer(self, optimizer):
for callback in self.callbacks:
callback.set_optimizer(optimizer)
......@@ -82,6 +82,7 @@ class CallbackList(object):
def __iter__(self):
return iter(self.callbacks)
class Callback(object):
"""Abstract base class used to build new callbacks.
"""
......@@ -94,7 +95,7 @@ class Callback(object):
def set_model(self, model):
self.model = model
def set_optimizer(self, optimizer):
self.optimizer = optimizer
......@@ -110,18 +111,18 @@ class Callback(object):
def on_train_end(self, logs=None):
pass
class BaseLogger(Callback):
class BaseLogger(Callback):
def __init__(self, period=10):
super(BaseLogger, self).__init__()
self.period = period
def _reset(self):
self.totals = {}
def on_train_begin(self, logs=None):
self.totals = {}
def on_iter_end(self, iter, logs=None):
logs = logs or {}
#(iter - 1) // iters_per_epoch + 1
......@@ -132,13 +133,13 @@ class BaseLogger(Callback):
self.totals[k] = v
if iter % self.period == 0 and ParallelEnv().local_rank == 0:
for k in self.totals:
logs[k] = self.totals[k] / self.period
self._reset()
class TrainLogger(Callback):
class TrainLogger(Callback):
def __init__(self, log_freq=10):
self.log_freq = log_freq
......@@ -154,7 +155,7 @@ class TrainLogger(Callback):
return result.format(*arr)
def on_iter_end(self, iter, logs=None):
if iter % self.log_freq == 0 and ParallelEnv().local_rank == 0:
total_iters = self.params["total_iters"]
iters_per_epoch = self.params["iters_per_epoch"]
......@@ -167,49 +168,50 @@ class TrainLogger(Callback):
reader_cost = logs["reader_cost"]
logger.info(
"[TRAIN] epoch={}, iter={}/{}, loss={:.4f}, lr={:.6f}, batch_cost={:.4f}, reader_cost={:.4f} | ETA {}".
format(current_epoch, iter, total_iters,
loss, lr, batch_cost, reader_cost, eta))
"[TRAIN] epoch={}, iter={}/{}, loss={:.4f}, lr={:.6f}, batch_cost={:.4f}, reader_cost={:.4f} | ETA {}"
.format(current_epoch, iter, total_iters, loss, lr, batch_cost,
reader_cost, eta))
class ProgbarLogger(Callback):
class ProgbarLogger(Callback):
def __init__(self):
super(ProgbarLogger, self).__init__()
def on_train_begin(self, logs=None):
self.verbose = self.params["verbose"]
self.total_iters = self.params["total_iters"]
self.target = self.params["total_iters"]
self.target = self.params["total_iters"]
self.progbar = Progbar(target=self.target, verbose=self.verbose)
self.seen = 0
self.log_values = []
def on_iter_begin(self, iter, logs=None):
#self.seen = 0
if self.seen < self.target:
self.log_values = []
def on_iter_end(self, iter, logs=None):
logs = logs or {}
self.seen += 1
for k in self.params['metrics']:
if k in logs:
self.log_values.append((k, logs[k]))
#if self.verbose and self.seen < self.target and ParallelEnv.local_rank == 0:
#print(self.log_values)
#print(self.log_values)
if self.seen < self.target:
self.progbar.update(self.seen, self.log_values)
class ModelCheckpoint(Callback):
def __init__(self,
save_dir,
monitor="miou",
save_best_only=False,
save_params_only=True,
mode="max",
period=1):
def __init__(self, save_dir, monitor="miou",
save_best_only=False, save_params_only=True,
mode="max", period=1):
super(ModelCheckpoint, self).__init__()
self.monitor = monitor
self.save_dir = save_dir
......@@ -241,7 +243,7 @@ class ModelCheckpoint(Callback):
current_save_dir = os.path.join(self.save_dir, "iter_{}".format(iter))
current_save_dir = os.path.abspath(current_save_dir)
#if self.iters_since_last_save % self.period and ParallelEnv().local_rank == 0:
#self.iters_since_last_save = 0
#self.iters_since_last_save = 0
if iter % self.period == 0 and ParallelEnv().local_rank == 0:
if self.verbose > 0:
print("iter {iter_num}: saving model to {path}".format(
......@@ -252,11 +254,9 @@ class ModelCheckpoint(Callback):
if not self.save_params_only:
paddle.save(self.optimizer.state_dict(), filepath)
class VisualDL(Callback):
def __init__(self, log_dir="./log", freq=1):
super(VisualDL, self).__init__()
self.log_dir = log_dir
......@@ -274,4 +274,4 @@ class VisualDL(Callback):
self.writer.flush()
def on_train_end(self, logs=None):
self.writer.close()
\ No newline at end of file
self.writer.close()
dygraph/paddleseg/models/ann.py
浏览文件 @ 3f658a36
......@@ -28,7 +28,7 @@ class ANN(nn.Layer):
"""
The ANN implementation based on PaddlePaddle.
The orginal artile refers to
The original article refers to
Zhen, Zhu, et al. "Asymmetric Non-local Neural Networks for Semantic Segmentation."
(https://arxiv.org/pdf/1908.07678.pdf)
......@@ -37,8 +37,8 @@ class ANN(nn.Layer):
Args:
num_classes (int): the unique number of target classes.
backbone (Paddle.nn.Layer): backbone network, currently support Resnet50/101.
model_pretrained (str): the path of pretrained model. Defaullt to None.
backbone_indices (tuple): two values in the tuple indicte the indices of output of backbone.
model_pretrained (str): the path of pretrained model. Default to None.
backbone_indices (tuple): two values in the tuple indicate the indices of output of backbone.
the first index will be taken as low-level features; the second one will be
taken as high-level features in AFNB module. Usually backbone consists of four
downsampling stage, and return an output of each stage, so we set default (2, 3),
......@@ -48,7 +48,7 @@ class ANN(nn.Layer):
Default to 256.
inter_channels (int): both input and output channels of APNB modules.
psp_size (tuple): the out size of pooled feature maps. Default to (1, 3, 6, 8).
enable_auxiliary_loss (bool): a bool values indictes whether adding auxiliary loss. Default to True.
enable_auxiliary_loss (bool): a bool values indicates whether adding auxiliary loss. Default to True.
"""
def __init__(self,
......@@ -79,7 +79,7 @@ class ANN(nn.Layer):
psp_size=psp_size)
self.context = nn.Sequential(
layer_libs.ConvBnRelu(
layer_libs.ConvBNReLU(
in_channels=high_in_channels,
out_channels=inter_channels,
kernel_size=3,
......@@ -94,9 +94,7 @@ class ANN(nn.Layer):
psp_size=psp_size))
self.cls = nn.Conv2d(
in_channels=inter_channels,
out_channels=num_classes,
kernel_size=1)
in_channels=inter_channels, out_channels=num_classes, kernel_size=1)
self.auxlayer = layer_libs.AuxLayer(
in_channels=low_in_channels,
inter_channels=low_in_channels // 2,
......@@ -122,7 +120,8 @@ class ANN(nn.Layer):
if self.enable_auxiliary_loss:
auxiliary_logit = self.auxlayer(low_level_x)
auxiliary_logit = F.resize_bilinear(auxiliary_logit, input.shape[2:])
auxiliary_logit = F.resize_bilinear(auxiliary_logit,
input.shape[2:])
logit_list.append(auxiliary_logit)
return logit_list
......@@ -219,7 +218,7 @@ class APNB(nn.Layer):
SelfAttentionBlock_APNB(in_channels, out_channels, key_channels,
value_channels, size) for size in sizes
])
self.conv_bn = layer_libs.ConvBnRelu(
self.conv_bn = layer_libs.ConvBNReLU(
in_channels=in_channels * 2,
out_channels=out_channels,
kernel_size=1)
......@@ -280,11 +279,11 @@ class SelfAttentionBlock_AFNB(nn.Layer):
if out_channels == None:
self.out_channels = high_in_channels
self.pool = nn.Pool2D(pool_size=(scale, scale), pool_type="max")
self.f_key = layer_libs.ConvBnRelu(
self.f_key = layer_libs.ConvBNReLU(
in_channels=low_in_channels,
out_channels=key_channels,
kernel_size=1)
self.f_query = layer_libs.ConvBnRelu(
self.f_query = layer_libs.ConvBNReLU(
in_channels=high_in_channels,
out_channels=key_channels,
kernel_size=1)
......@@ -315,7 +314,7 @@ class SelfAttentionBlock_AFNB(nn.Layer):
key = _pp_module(key, self.psp_size)
sim_map = paddle.matmul(query, key)
sim_map = (self.key_channels ** -.5) * sim_map
sim_map = (self.key_channels**-.5) * sim_map
sim_map = F.softmax(sim_map, axis=-1)
context = paddle.matmul(sim_map, value)
......@@ -358,7 +357,7 @@ class SelfAttentionBlock_APNB(nn.Layer):
self.value_channels = value_channels
self.pool = nn.Pool2D(pool_size=(scale, scale), pool_type="max")
self.f_key = layer_libs.ConvBnRelu(
self.f_key = layer_libs.ConvBNReLU(
in_channels=self.in_channels,
out_channels=self.key_channels,
kernel_size=1)
......@@ -384,15 +383,14 @@ class SelfAttentionBlock_APNB(nn.Layer):
value = paddle.transpose(value, perm=(0, 2, 1))
query = self.f_query(x)
query = paddle.reshape(
query, shape=(batch_size, self.key_channels, -1))
query = paddle.reshape(query, shape=(batch_size, self.key_channels, -1))
query = paddle.transpose(query, perm=(0, 2, 1))
key = self.f_key(x)
key = _pp_module(key, self.psp_size)
sim_map = paddle.matmul(query, key)
sim_map = (self.key_channels ** -.5) * sim_map
sim_map = (self.key_channels**-.5) * sim_map
sim_map = F.softmax(sim_map, axis=-1)
context = paddle.matmul(sim_map, value)
......
dygraph/paddleseg/models/backbones/resnet_vd.py
浏览文件 @ 3f658a36
......@@ -133,8 +133,9 @@ class BottleneckBlock(nn.Layer):
# If given dilation rate > 1, using corresponding padding
if self.dilation > 1:
padding = self.dilation
y = F.pad(y, [0, 0, 0, 0, padding, padding, padding, padding])
y = F.pad(y, [padding, padding, padding, padding])
#####################################################################
conv1 = self.conv1(y)
conv2 = self.conv2(conv1)
......@@ -197,11 +198,10 @@ class BasicBlock(nn.Layer):
class ResNet_vd(nn.Layer):
def __init__(self,
backbone_pretrained=None,
layers=50,
class_dim=1000,
output_stride=None,
multi_grid=(1, 1, 1)):
multi_grid=(1, 1, 1),
pretrained=None):
super(ResNet_vd, self).__init__()
self.layers = layers
......@@ -224,6 +224,10 @@ class ResNet_vd(nn.Layer):
] if layers >= 50 else [64, 64, 128, 256]
num_filters = [64, 128, 256, 512]
# for channels of returned stage
self.backbone_channels = [c * 4 for c in num_filters
] if layers >= 50 else num_filters
dilation_dict = None
if output_stride == 8:
dilation_dict = {2: 2, 3: 4}
......@@ -315,6 +319,8 @@ class ResNet_vd(nn.Layer):
block_list.append(basic_block)
shortcut = True
self.stage_list.append(block_list)
utils.load_pretrained_model(self, pretrained)
def forward(self, inputs):
y = self.conv1_1(inputs)
......@@ -324,12 +330,14 @@ class ResNet_vd(nn.Layer):
# A feature list saves the output feature map of each stage.
feat_list = []
for i, stage in enumerate(self.stage_list):
for j, block in enumerate(stage):
for stage in self.stage_list:
for block in stage:
y = block(y)
feat_list.append(y)
return feat_list
@manager.BACKBONES.add_component
......
dygraph/paddleseg/models/common/pyramid_pool.py
浏览文件 @ 3f658a36
......@@ -13,7 +13,6 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle
from paddle import nn
import paddle.nn.functional as F
......@@ -34,11 +33,11 @@ class ASPPModule(nn.Layer):
image_pooling: if augmented with image-level features.
"""
def __init__(self,
aspp_ratios,
in_channels,
out_channels,
sep_conv=False,
def __init__(self,
aspp_ratios,
in_channels,
out_channels,
sep_conv=False,
image_pooling=False):
super(ASPPModule, self).__init__()
......@@ -47,42 +46,41 @@ class ASPPModule(nn.Layer):
for ratio in aspp_ratios:
if sep_conv and ratio > 1:
conv_func = layer_libs.DepthwiseConvBnRelu
conv_func = layer_libs.DepthwiseConvBNReLU
else:
conv_func = layer_libs.ConvBnRelu
conv_func = layer_libs.ConvBNReLU
block = conv_func(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1 if ratio == 1 else 3,
dilation=ratio,
padding=0 if ratio == 1 else ratio
)
padding=0 if ratio == 1 else ratio)
self.aspp_blocks.append(block)
out_size = len(self.aspp_blocks)
if image_pooling:
self.global_avg_pool = nn.Sequential(
nn.AdaptiveAvgPool2d(output_size=(1, 1)),
layer_libs.ConvBnRelu(in_channels, out_channels, kernel_size=1, bias_attr=False)
)
layer_libs.ConvBNReLU(
in_channels, out_channels, kernel_size=1, bias_attr=False))
out_size += 1
self.image_pooling = image_pooling
self.conv_bn_relu = layer_libs.ConvBnRelu(
in_channels=out_channels * out_size,
out_channels=out_channels,
self.conv_bn_relu = layer_libs.ConvBNReLU(
in_channels=out_channels * out_size,
out_channels=out_channels,
kernel_size=1)
self.dropout = nn.Dropout(p=0.1) # drop rate
self.dropout = nn.Dropout(p=0.1) # drop rate
def forward(self, x):
outputs = []
for block in self.aspp_blocks:
outputs.append(block(x))
if self.image_pooling:
img_avg = self.global_avg_pool(x)
img_avg = F.resize_bilinear(img_avg, out_shape=x.shape[2:])
......@@ -93,17 +91,17 @@ class ASPPModule(nn.Layer):
x = self.dropout(x)
return x
class PPModule(nn.Layer):
"""
Pyramid pooling module orginally in PSPNet
Pyramid pooling module originally in PSPNet
Args:
in_channels (int): the number of intput channels to pyramid pooling module.
out_channels (int): the number of output channels after pyramid pooling module.
bin_sizes (tuple): the out size of pooled feature maps. Default to (1,2,3,6).
dim_reduction (bool): a bool value represent if reduing dimention after pooling. Default to True.
dim_reduction (bool): a bool value represent if reducing dimension after pooling. Default to True.
"""
def __init__(self,
......@@ -125,7 +123,7 @@ class PPModule(nn.Layer):
for size in bin_sizes
])
self.conv_bn_relu2 = layer_libs.ConvBnRelu(
self.conv_bn_relu2 = layer_libs.ConvBNReLU(
in_channels=in_channels + inter_channels * len(bin_sizes),
out_channels=out_channels,
kernel_size=3,
......@@ -135,7 +133,7 @@ class PPModule(nn.Layer):
"""
Create one pooling layer.
In our implementation, we adopt the same dimention reduction as the original paper that might be
In our implementation, we adopt the same dimension reduction as the original paper that might be
slightly different with other implementations.
After pooling, the channels are reduced to 1/len(bin_sizes) immediately, while some other implementations
......@@ -151,7 +149,7 @@ class PPModule(nn.Layer):
"""
prior = nn.AdaptiveAvgPool2d(output_size=(size, size))
conv = layer_libs.ConvBnRelu(
conv = layer_libs.ConvBNReLU(
in_channels=in_channels, out_channels=out_channels, kernel_size=1)
return nn.Sequential(prior, conv)
......@@ -167,4 +165,4 @@ class PPModule(nn.Layer):
cat = paddle.concat(cat_layers, axis=1)
out = self.conv_bn_relu2(cat)
return out
\ No newline at end of file
return out
dygraph/paddleseg/models/deeplab.py
浏览文件 @ 3f658a36
......@@ -29,140 +29,193 @@ class DeepLabV3P(nn.Layer):
"""
The DeepLabV3Plus implementation based on PaddlePaddle.
The orginal artile refers to
"Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation"
Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, Hartwig Adam.
The original article refers to
Liang-Chieh Chen, et, al. "Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation"
(https://arxiv.org/abs/1802.02611)
The DeepLabV3P consists of three main components, Backbone, ASPP and Decoder.
Args:
num_classes (int): the unique number of target classes.
backbone (paddle.nn.Layer): backbone network, currently support Xception65, Resnet101_vd.
model_pretrained (str): the path of pretrained model.
aspp_ratios (tuple): the dilation rate using in ASSP module.
if output_stride=16, aspp_ratios should be set as (1, 6, 12, 18).
if output_stride=8, aspp_ratios is (1, 12, 24, 36).
backbone_indices (tuple): two values in the tuple indicte the indices of output of backbone.
the first index will be taken as a low-level feature in Deconder component;
backbone (paddle.nn.Layer): backbone network, currently support Resnet50_vd/Resnet101_vd/Xception65.
backbone_indices (tuple): two values in the tuple indicate the indices of output of backbone.
the first index will be taken as a low-level feature in Decoder component;
the second one will be taken as input of ASPP component.
Usually backbone consists of four downsampling stage, and return an output of
each stage, so we set default (0, 3), which means taking feature map of the first
stage in backbone as low-level feature used in Decoder, and feature map of the fourth
stage as input of ASPP.
backbone_channels (tuple): the same length with "backbone_indices". It indicates the channels of corresponding index.
aspp_ratios (tuple): the dilation rate using in ASSP module.
if output_stride=16, aspp_ratios should be set as (1, 6, 12, 18).
if output_stride=8, aspp_ratios is (1, 12, 24, 36).
aspp_out_channels (int): the output channels of ASPP module.
pretrained (str): the path of pretrained model for fine tuning.
"""
def __init__(self,
num_classes,
backbone,
backbone_pretrained=None,
model_pretrained=None,
backbone_indices=(0, 3),
backbone_channels=(256, 2048),
aspp_ratios=(1, 6, 12, 18),
aspp_out_channels=256):
aspp_out_channels=256,
pretrained=None):
super(DeepLabV3P, self).__init__()
self.backbone = backbone
self.backbone_pretrained = backbone_pretrained
self.model_pretrained = model_pretrained
backbone_channels = backbone.backbone_channels
self.head = DeepLabV3PHead(
num_classes,
backbone_indices,
backbone_channels,
aspp_ratios,
aspp_out_channels)
utils.load_entire_model(self, pretrained)
def forward(self, input):
feat_list = self.backbone(input)
logit_list = self.head(feat_list)
return [
F.resize_bilinear(logit, input.shape[2:]) for logit in logit_list
]
class DeepLabV3PHead(nn.Layer):
"""
The DeepLabV3PHead implementation based on PaddlePaddle.
Args:
num_classes (int): the unique number of target classes.
backbone_indices (tuple): two values in the tuple indicate the indices of output of backbone.
the first index will be taken as a low-level feature in Decoder component;
the second one will be taken as input of ASPP component.
Usually backbone consists of four downsampling stage, and return an output of
each stage, so we set default (0, 3), which means taking feature map of the first
stage in backbone as low-level feature used in Decoder, and feature map of the fourth
stage as input of ASPP.
backbone_channels (tuple): returned channels of backbone
aspp_ratios (tuple): the dilation rate using in ASSP module.
if output_stride=16, aspp_ratios should be set as (1, 6, 12, 18).
if output_stride=8, aspp_ratios is (1, 12, 24, 36).
aspp_out_channels (int): the output channels of ASPP module.
"""
def __init__(self,
num_classes,
backbone_indices,
backbone_channels,
aspp_ratios=(1, 6, 12, 18),
aspp_out_channels=256):
super(DeepLabV3PHead, self).__init__()
self.aspp = pyramid_pool.ASPPModule(
aspp_ratios, backbone_channels[1], aspp_out_channels, sep_conv=True, image_pooling=True)
self.decoder = Decoder(num_classes, backbone_channels[0])
aspp_ratios,
backbone_channels[backbone_indices[1]],
aspp_out_channels,
sep_conv=True,
image_pooling=True)
self.decoder = Decoder(num_classes, backbone_channels[backbone_indices[0]])
self.backbone_indices = backbone_indices
self.init_weight()
def forward(self, input, label=None):
def forward(self, feat_list):
logit_list = []
_, feat_list = self.backbone(input)
low_level_feat = feat_list[self.backbone_indices[0]]
x = feat_list[self.backbone_indices[1]]
x = self.aspp(x)
logit = self.decoder(x, low_level_feat)
logit = F.resize_bilinear(logit, input.shape[2:])
logit_list.append(logit)
return logit_list
def init_weight(self):
"""
Initialize the parameters of model parts.
Args:
pretrained_model ([str], optional): the path of pretrained model. Defaults to None.
"""
if self.model_pretrained is not None:
utils.load_pretrained_model(self, self.model_pretrained)
elif self.backbone_pretrained is not None:
utils.load_pretrained_model(self.backbone, self.backbone_pretrained)
pass
@manager.MODELS.add_component
class DeepLabV3(nn.Layer):
"""
The DeepLabV3 implementation based on PaddlePaddle.
The orginal article refers to
"Rethinking Atrous Convolution for Semantic Image Segmentation"
Liang-Chieh Chen, George Papandreou, Florian Schroff, Hartwig Adam.
The original article refers to
Liang-Chieh Chen, et, al. "Rethinking Atrous Convolution for Semantic Image Segmentation"
(https://arxiv.org/pdf/1706.05587.pdf)
Args:
Refer to DeepLabV3P above
"""
def __init__(self,
num_classes,
backbone,
backbone_pretrained=None,
model_pretrained=None,
backbone_indices=(3,),
backbone_channels=(2048,),
pretrained=None,
backbone_indices=(3, ),
aspp_ratios=(1, 6, 12, 18),
aspp_out_channels=256):
super(DeepLabV3, self).__init__()
self.backbone = backbone
backbone_channels = backbone.backbone_channels
self.head = DeepLabV3Head(
num_classes,
backbone_indices,
backbone_channels,
aspp_ratios,
aspp_out_channels)
utils.load_entire_model(self, pretrained)
def forward(self, input):
feat_list = self.backbone(input)
logit_list = self.head(feat_list)
return [
F.resize_bilinear(logit, input.shape[2:]) for logit in logit_list
]
class DeepLabV3Head(nn.Layer):
def __init__(self,
num_classes,
backbone_indices=(3, ),
backbone_channels=(2048, ),
aspp_ratios=(1, 6, 12, 18),
aspp_out_channels=256):
super(DeepLabV3Head, self).__init__()
self.aspp = pyramid_pool.ASPPModule(
aspp_ratios, backbone_channels[0], aspp_out_channels,
sep_conv=False, image_pooling=True)
aspp_ratios,
backbone_channels[backbone_indices[0]],
aspp_out_channels,
sep_conv=False,
image_pooling=True)
self.cls = nn.Conv2d(
in_channels=backbone_channels[0],
in_channels=backbone_channels[backbone_indices[0]],
out_channels=num_classes,
kernel_size=1)
self.backbone_indices = backbone_indices
self.init_weight(model_pretrained)
self.init_weight()
def forward(self, input, label=None):
def forward(self, feat_list):
logit_list = []
_, feat_list = self.backbone(input)
x = feat_list[self.backbone_indices[0]]
logit = self.cls(x)
logit = F.resize_bilinear(logit, input.shape[2:])
logit_list.append(logit)
return logit_list
def init_weight(self, pretrained_model=None):
"""
Initialize the parameters of model parts.
Args:
pretrained_model ([str], optional): the path of pretrained model. Defaults to None.
"""
if pretrained_model is not None:
if os.path.exists(pretrained_model):
utils.load_pretrained_model(self, pretrained_model)
else:
raise Exception('Pretrained model is not found: {}'.format(
pretrained_model))
def init_weight(self):
pass
class Decoder(nn.Layer):
......@@ -178,12 +231,12 @@ class Decoder(nn.Layer):
def __init__(self, num_classes, in_channels):
super(Decoder, self).__init__()
self.conv_bn_relu1 = layer_libs.ConvBnRelu(
self.conv_bn_relu1 = layer_libs.ConvBNReLU(
in_channels=in_channels, out_channels=48, kernel_size=1)
self.conv_bn_relu2 = layer_libs.DepthwiseConvBnRelu(
self.conv_bn_relu2 = layer_libs.DepthwiseConvBNReLU(
in_channels=304, out_channels=256, kernel_size=3, padding=1)
self.conv_bn_relu3 = layer_libs.DepthwiseConvBnRelu(
self.conv_bn_relu3 = layer_libs.DepthwiseConvBNReLU(
in_channels=256, out_channels=256, kernel_size=3, padding=1)
self.conv = nn.Conv2d(
in_channels=256, out_channels=num_classes, kernel_size=1)
......
dygraph/paddleseg/models/fast_scnn.py
浏览文件 @ 3f658a36
......@@ -26,15 +26,15 @@ class FastSCNN(nn.Layer):
As mentioned in the original paper, FastSCNN is a real-time segmentation algorithm (123.5fps)
even for high resolution images (1024x2048).
The orginal artile refers to
The original article refers to
Poudel, Rudra PK, et al. "Fast-scnn: Fast semantic segmentation network."
(https://arxiv.org/pdf/1902.04502.pdf)
Args:
num_classes (int): the unique number of target classes. Default to 2.
model_pretrained (str): the path of pretrained model. Defaullt to None.
enable_auxiliary_loss (bool): a bool values indictes whether adding auxiliary loss.
model_pretrained (str): the path of pretrained model. Default to None.
enable_auxiliary_loss (bool): a bool values indicates whether adding auxiliary loss.
if true, auxiliary loss will be added after LearningToDownsample module, where the weight is 0.4. Default to False.
"""
......@@ -105,15 +105,15 @@ class LearningToDownsample(nn.Layer):
def __init__(self, dw_channels1=32, dw_channels2=48, out_channels=64):
super(LearningToDownsample, self).__init__()
self.conv_bn_relu = layer_libs.ConvBnRelu(
self.conv_bn_relu = layer_libs.ConvBNReLU(
in_channels=3, out_channels=dw_channels1, kernel_size=3, stride=2)
self.dsconv_bn_relu1 = layer_libs.DepthwiseConvBnRelu(
self.dsconv_bn_relu1 = layer_libs.DepthwiseConvBNReLU(
in_channels=dw_channels1,
out_channels=dw_channels2,
kernel_size=3,
stride=2,
padding=1)
self.dsconv_bn_relu2 = layer_libs.DepthwiseConvBnRelu(
self.dsconv_bn_relu2 = layer_libs.DepthwiseConvBNReLU(
in_channels=dw_channels2,
out_channels=out_channels,
kernel_size=3,
......@@ -208,13 +208,13 @@ class LinearBottleneck(nn.Layer):
expand_channels = in_channels * expansion
self.block = nn.Sequential(
# pw
layer_libs.ConvBnRelu(
layer_libs.ConvBNReLU(
in_channels=in_channels,
out_channels=expand_channels,
kernel_size=1,
bias_attr=False),
# dw
layer_libs.ConvBnRelu(
layer_libs.ConvBNReLU(
in_channels=expand_channels,
out_channels=expand_channels,
kernel_size=3,
......@@ -239,7 +239,7 @@ class LinearBottleneck(nn.Layer):
class FeatureFusionModule(nn.Layer):
"""
Feature Fusion Module Implememtation.
Feature Fusion Module Implementation.
This module fuses high-resolution feature and low-resolution feature.
......@@ -253,7 +253,7 @@ class FeatureFusionModule(nn.Layer):
super(FeatureFusionModule, self).__init__()
# There only depth-wise conv is used WITHOUT point-wise conv
self.dwconv = layer_libs.ConvBnRelu(
self.dwconv = layer_libs.ConvBNReLU(
in_channels=low_in_channels,
out_channels=out_channels,
kernel_size=3,
......@@ -301,13 +301,13 @@ class Classifier(nn.Layer):
def __init__(self, input_channels, num_classes):
super(Classifier, self).__init__()
self.dsconv1 = layer_libs.DepthwiseConvBnRelu(
self.dsconv1 = layer_libs.DepthwiseConvBNReLU(
in_channels=input_channels,
out_channels=input_channels,
kernel_size=3,
padding=1)
self.dsconv2 = layer_libs.DepthwiseConvBnRelu(
self.dsconv2 = layer_libs.DepthwiseConvBNReLU(
in_channels=input_channels,
out_channels=input_channels,
kernel_size=3,
......
dygraph/paddleseg/models/gcnet.py
浏览文件 @ 3f658a36
......@@ -27,15 +27,15 @@ class GCNet(nn.Layer):
"""
The GCNet implementation based on PaddlePaddle.
The orginal artile refers to
The original article refers to
Cao, Yue, et al. "GCnet: Non-local networks meet squeeze-excitation networks and beyond."
(https://arxiv.org/pdf/1904.11492.pdf)
Args:
num_classes (int): the unique number of target classes.
backbone (Paddle.nn.Layer): backbone network, currently support Resnet50/101.
model_pretrained (str): the path of pretrained model. Defaullt to None.
backbone_indices (tuple): two values in the tuple indicte the indices of output of backbone.
model_pretrained (str): the path of pretrained model. Default to None.
backbone_indices (tuple): two values in the tuple indicate the indices of output of backbone.
the first index will be taken as a deep-supervision feature in auxiliary layer;
the second one will be taken as input of GlobalContextBlock. Usually backbone
consists of four downsampling stage, and return an output of each stage, so we
......@@ -43,8 +43,8 @@ class GCNet(nn.Layer):
and the fourth stage (res5c) in backbone.
backbone_channels (tuple): the same length with "backbone_indices". It indicates the channels of corresponding index.
gc_channels (int): input channels to Global Context Block. Default to 512.
ratio (float): it indictes the ratio of attention channels and gc_channels. Default to 1/4.
enable_auxiliary_loss (bool): a bool values indictes whether adding auxiliary loss. Default to True.
ratio (float): it indicates the ratio of attention channels and gc_channels. Default to 1/4.
enable_auxiliary_loss (bool): a bool values indicates whether adding auxiliary loss. Default to True.
"""
def __init__(self,
......@@ -63,7 +63,7 @@ class GCNet(nn.Layer):
self.backbone = backbone
in_channels = backbone_channels[1]
self.conv_bn_relu1 = layer_libs.ConvBnRelu(
self.conv_bn_relu1 = layer_libs.ConvBNReLU(
in_channels=in_channels,
out_channels=gc_channels,
kernel_size=3,
......@@ -71,13 +71,13 @@ class GCNet(nn.Layer):
self.gc_block = GlobalContextBlock(in_channels=gc_channels, ratio=ratio)
self.conv_bn_relu2 = layer_libs.ConvBnRelu(
self.conv_bn_relu2 = layer_libs.ConvBNReLU(
in_channels=gc_channels,
out_channels=gc_channels,
kernel_size=3,
padding=1)
self.conv_bn_relu3 = layer_libs.ConvBnRelu(
self.conv_bn_relu3 = layer_libs.ConvBNReLU(
in_channels=in_channels + gc_channels,
out_channels=gc_channels,
kernel_size=3,
......@@ -154,7 +154,7 @@ class GlobalContextBlock(nn.Layer):
in_channels=in_channels, out_channels=1, kernel_size=1)
self.softmax = nn.Softmax(axis=2)
inter_channels = int(in_channels * ratio)
self.channel_add_conv = nn.Sequential(
nn.Conv2d(
......
dygraph/paddleseg/models/ocrnet.py
浏览文件 @ 3f658a36
......@@ -124,7 +124,6 @@ class ObjectAttentionBlock(nn.Layer):
class OCRHead(nn.Layer):
"""
The Object contextual representation head.
Args:
num_classes(int): the unique number of target classes.
in_channels(tuple): the number of input channels.
......@@ -179,11 +178,9 @@ class OCRHead(nn.Layer):
class OCRNet(nn.Layer):
"""
The OCRNet implementation based on PaddlePaddle.
The original article refers to
Yuan, Yuhui, et al. "Object-Contextual Representations for Semantic Segmentation"
(https://arxiv.org/pdf/1909.11065.pdf)
Args:
num_classes(int): the unique number of target classes.
backbone(Paddle.nn.Layer): backbone network.
......@@ -234,4 +231,4 @@ class OCRNet(nn.Layer):
utils.load_pretrained_model(self, pretrained)
else:
raise Exception(
'Pretrained model is not found: {}'.format(pretrained))
'Pretrained model is not found: {}'.format(pretrained))
\ No newline at end of file
dygraph/paddleseg/models/pspnet.py
浏览文件 @ 3f658a36
......@@ -26,7 +26,7 @@ class PSPNet(nn.Layer):
"""
The PSPNet implementation based on PaddlePaddle.
The orginal artile refers to
The original article refers to
Zhao, Hengshuang, et al. "Pyramid scene parsing network."
Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.
(https://openaccess.thecvf.com/content_cvpr_2017/papers/Zhao_Pyramid_Scene_Parsing_CVPR_2017_paper.pdf)
......@@ -34,8 +34,8 @@ class PSPNet(nn.Layer):
Args:
num_classes (int): the unique number of target classes.
backbone (Paddle.nn.Layer): backbone network, currently support Resnet50/101.
model_pretrained (str): the path of pretrained model. Defaullt to None.
backbone_indices (tuple): two values in the tuple indicte the indices of output of backbone.
model_pretrained (str): the path of pretrained model. Default to None.
backbone_indices (tuple): two values in the tuple indicate the indices of output of backbone.
the first index will be taken as a deep-supervision feature in auxiliary layer;
the second one will be taken as input of Pyramid Pooling Module (PPModule).
Usually backbone consists of four downsampling stage, and return an output of
......@@ -44,7 +44,7 @@ class PSPNet(nn.Layer):
backbone_channels (tuple): the same length with "backbone_indices". It indicates the channels of corresponding index.
pp_out_channels (int): output channels after Pyramid Pooling Module. Default to 1024.
bin_sizes (tuple): the out size of pooled feature maps. Default to (1,2,3,6).
enable_auxiliary_loss (bool): a bool values indictes whether adding auxiliary loss. Default to True.
enable_auxiliary_loss (bool): a bool values indicates whether adding auxiliary loss. Default to True.
"""
def __init__(self,
......@@ -107,6 +107,7 @@ class PSPNet(nn.Layer):
def init_weight(self, pretrained_model=None):
"""
Initialize the parameters of model parts.
Args:
pretrained_model ([str], optional): the path of pretrained model. Defaults to None.
"""
......
dygraph/paddleseg/utils/metrics.py
浏览文件 @ 3f658a36
......@@ -41,7 +41,7 @@ class ConfusionMatrix(object):
label = np.asarray(label)[mask]
pred = np.asarray(pred)[mask]
one = np.ones_like(pred)
# Accumuate ([row=label, col=pred], 1) into sparse matrix
# Accumuate ([row=label, col=pred], 1) into sparse
spm = csr_matrix((one, (label, pred)),
shape=(self.num_classes, self.num_classes))
spm = spm.todense()
......
dygraph/paddleseg/utils/progbar.py
浏览文件 @ 3f658a36
......@@ -17,8 +17,9 @@ import time
import numpy as np
class Progbar(object):
"""Displays a progress bar.
"""Displays a progress bar.
refers to https://github.com/keras-team/keras/blob/keras-2/keras/utils/generic_utils.py
Arguments:
target: Total number of steps expected, None if unknown.
......@@ -31,39 +32,39 @@ class Progbar(object):
unit_name: Display name for step counts (usually "step" or "sample").
"""
def __init__(self,
target,
width=30,
verbose=1,
interval=0.05,
stateful_metrics=None,
unit_name='step'):
self.target = target
self.width = width
self.verbose = verbose
self.interval = interval
self.unit_name = unit_name
if stateful_metrics:
self.stateful_metrics = set(stateful_metrics)
else:
self.stateful_metrics = set()
self._dynamic_display = ((hasattr(sys.stdout, 'isatty') and
sys.stdout.isatty()) or
'ipykernel' in sys.modules or
'posix' in sys.modules or
'PYCHARM_HOSTED' in os.environ)
self._total_width = 0
self._seen_so_far = 0
# We use a dict + list to avoid garbage collection
# issues found in OrderedDict
self._values = {}
self._values_order = []
self._start = time.time()
self._last_update = 0
def update(self, current, values=None, finalize=None):
"""Updates the progress bar.
def __init__(self,
target,
width=30,
verbose=1,
interval=0.05,
stateful_metrics=None,
unit_name='step'):
self.target = target
self.width = width
self.verbose = verbose
self.interval = interval
self.unit_name = unit_name
if stateful_metrics:
self.stateful_metrics = set(stateful_metrics)
else:
self.stateful_metrics = set()
self._dynamic_display = ((hasattr(sys.stdout, 'isatty')
and sys.stdout.isatty())
or 'ipykernel' in sys.modules
or 'posix' in sys.modules
or 'PYCHARM_HOSTED' in os.environ)
self._total_width = 0
self._seen_so_far = 0
# We use a dict + list to avoid garbage collection
# issues found in OrderedDict
self._values = {}
self._values_order = []
self._start = time.time()
self._last_update = 0
def update(self, current, values=None, finalize=None):
"""Updates the progress bar.
Arguments:
current: Index of current step.
values: List of tuples: `(name, value_for_last_step)`. If `name` is in
......@@ -72,129 +73,131 @@ class Progbar(object):
finalize: Whether this is the last update for the progress bar. If
`None`, defaults to `current >= self.target`.
"""
if finalize is None:
if self.target is None:
finalize = False
else:
finalize = current >= self.target
values = values or []
for k, v in values:
if k not in self._values_order:
self._values_order.append(k)
if k not in self.stateful_metrics:
# In the case that progress bar doesn't have a target value in the first
# epoch, both on_batch_end and on_epoch_end will be called, which will
# cause 'current' and 'self._seen_so_far' to have the same value. Force
# the minimal value to 1 here, otherwise stateful_metric will be 0s.
value_base = max(current - self._seen_so_far, 1)
if k not in self._values:
self._values[k] = [v * value_base, value_base]
else:
self._values[k][0] += v * value_base
self._values[k][1] += value_base
else:
# Stateful metrics output a numeric value. This representation
# means "take an average from a single value" but keeps the
# numeric formatting.
self._values[k] = [v, 1]
self._seen_so_far = current
now = time.time()
info = ' - %.0fs' % (now - self._start)
if self.verbose == 1:
if now - self._last_update < self.interval and not finalize:
return
prev_total_width = self._total_width
if self._dynamic_display:
sys.stdout.write('\b' * prev_total_width)
sys.stdout.write('\r')
else:
sys.stdout.write('\n')
if self.target is not None:
numdigits = int(np.log10(self.target)) + 1
bar = ('%' + str(numdigits) + 'd/%d [') % (current, self.target)
prog = float(current) / self.target
prog_width = int(self.width * prog)
if prog_width > 0:
bar += ('=' * (prog_width - 1))
if current < self.target:
bar += '>'
else:
bar += '='
bar += ('.' * (self.width - prog_width))
bar += ']'
else:
bar = '%7d/Unknown' % current
self._total_width = len(bar)
sys.stdout.write(bar)
if current:
time_per_unit = (now - self._start) / current
else:
time_per_unit = 0
if self.target is None or finalize:
if time_per_unit >= 1 or time_per_unit == 0:
info += ' %.0fs/%s' % (time_per_unit, self.unit_name)
elif time_per_unit >= 1e-3:
info += ' %.0fms/%s' % (time_per_unit * 1e3, self.unit_name)
else:
info += ' %.0fus/%s' % (time_per_unit * 1e6, self.unit_name)
else:
eta = time_per_unit * (self.target - current)
if eta > 3600:
eta_format = '%d:%02d:%02d' % (eta // 3600,
(eta % 3600) // 60, eta % 60)
elif eta > 60:
eta_format = '%d:%02d' % (eta // 60, eta % 60)
else:
eta_format = '%ds' % eta
info = ' - ETA: %s' % eta_format
for k in self._values_order:
info += ' - %s:' % k
if isinstance(self._values[k], list):
avg = np.mean(self._values[k][0] / max(1, self._values[k][1]))
if abs(avg) > 1e-3:
info += ' %.4f' % avg
else:
info += ' %.4e' % avg
else:
info += ' %s' % self._values[k]
self._total_width += len(info)
if prev_total_width > self._total_width:
info += (' ' * (prev_total_width - self._total_width))
if finalize:
info += '\n'
sys.stdout.write(info)
sys.stdout.flush()
elif self.verbose == 2:
if finalize:
numdigits = int(np.log10(self.target)) + 1
count = ('%' + str(numdigits) + 'd/%d') % (current, self.target)
info = count + info
for k in self._values_order:
info += ' - %s:' % k
avg = np.mean(self._values[k][0] / max(1, self._values[k][1]))
if avg > 1e-3:
info += ' %.4f' % avg
else:
info += ' %.4e' % avg
info += '\n'
sys.stdout.write(info)
sys.stdout.flush()
self._last_update = now
def add(self, n, values=None):
self.update(self._seen_so_far + n, values)
\ No newline at end of file
if finalize is None:
if self.target is None:
finalize = False
else:
finalize = current >= self.target
values = values or []
for k, v in values:
if k not in self._values_order:
self._values_order.append(k)
if k not in self.stateful_metrics:
# In the case that progress bar doesn't have a target value in the first
# epoch, both on_batch_end and on_epoch_end will be called, which will
# cause 'current' and 'self._seen_so_far' to have the same value. Force
# the minimal value to 1 here, otherwise stateful_metric will be 0s.
value_base = max(current - self._seen_so_far, 1)
if k not in self._values:
self._values[k] = [v * value_base, value_base]
else:
self._values[k][0] += v * value_base
self._values[k][1] += value_base
else:
# Stateful metrics output a numeric value. This representation
# means "take an average from a single value" but keeps the
# numeric formatting.
self._values[k] = [v, 1]
self._seen_so_far = current
now = time.time()
info = ' - %.0fs' % (now - self._start)
if self.verbose == 1:
if now - self._last_update < self.interval and not finalize:
return
prev_total_width = self._total_width
if self._dynamic_display:
sys.stdout.write('\b' * prev_total_width)
sys.stdout.write('\r')
else:
sys.stdout.write('\n')
if self.target is not None:
numdigits = int(np.log10(self.target)) + 1
bar = ('%' + str(numdigits) + 'd/%d [') % (current, self.target)
prog = float(current) / self.target
prog_width = int(self.width * prog)
if prog_width > 0:
bar += ('=' * (prog_width - 1))
if current < self.target:
bar += '>'
else:
bar += '='
bar += ('.' * (self.width - prog_width))
bar += ']'
else:
bar = '%7d/Unknown' % current
self._total_width = len(bar)
sys.stdout.write(bar)
if current:
time_per_unit = (now - self._start) / current
else:
time_per_unit = 0
if self.target is None or finalize:
if time_per_unit >= 1 or time_per_unit == 0:
info += ' %.0fs/%s' % (time_per_unit, self.unit_name)
elif time_per_unit >= 1e-3:
info += ' %.0fms/%s' % (time_per_unit * 1e3, self.unit_name)
else:
info += ' %.0fus/%s' % (time_per_unit * 1e6, self.unit_name)
else:
eta = time_per_unit * (self.target - current)
if eta > 3600:
eta_format = '%d:%02d:%02d' % (eta // 3600,
(eta % 3600) // 60, eta % 60)
elif eta > 60:
eta_format = '%d:%02d' % (eta // 60, eta % 60)
else:
eta_format = '%ds' % eta
info = ' - ETA: %s' % eta_format
for k in self._values_order:
info += ' - %s:' % k
if isinstance(self._values[k], list):
avg = np.mean(
self._values[k][0] / max(1, self._values[k][1]))
if abs(avg) > 1e-3:
info += ' %.4f' % avg
else:
info += ' %.4e' % avg
else:
info += ' %s' % self._values[k]
self._total_width += len(info)
if prev_total_width > self._total_width:
info += (' ' * (prev_total_width - self._total_width))
if finalize:
info += '\n'
sys.stdout.write(info)
sys.stdout.flush()
elif self.verbose == 2:
if finalize:
numdigits = int(np.log10(self.target)) + 1
count = ('%' + str(numdigits) + 'd/%d') % (current, self.target)
info = count + info
for k in self._values_order:
info += ' - %s:' % k
avg = np.mean(
self._values[k][0] / max(1, self._values[k][1]))
if avg > 1e-3:
info += ' %.4f' % avg
else:
info += ' %.4e' % avg
info += '\n'
sys.stdout.write(info)
sys.stdout.flush()
self._last_update = now
def add(self, n, values=None):
self.update(self._seen_so_far + n, values)
dygraph/paddleseg/utils/utils.py
浏览文件 @ 3f658a36
......@@ -44,6 +44,19 @@ def seconds_to_hms(seconds):
return hms_str
def load_entire_model(model, pretrained):
if pretrained is not None:
if os.path.exists(pretrained):
load_pretrained_model(model, pretrained)
else:
raise Exception('Pretrained model is not found: {}'.format(
pretrained))
else:
logger.warning('Not all pretrained params of {} to load, '\
'training from scratch or a pretrained backbone'.format(model.__class__.__name__))
def load_pretrained_model(model, pretrained_model):
if pretrained_model is not None:
logger.info('Load pretrained model from {}'.format(pretrained_model))
......@@ -82,7 +95,7 @@ def load_pretrained_model(model, pretrained_model):
model_state_dict[k] = para_state_dict[k]
num_params_loaded += 1
model.set_dict(model_state_dict)
logger.info("There are {}/{} varaibles are loaded.".format(
logger.info("There are {}/{} variables are loaded.".format(
num_params_loaded, len(model_state_dict)))
else:
......
Markdown is supported
0% 或 .
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 或