Add CondenseNet V2

paddlers/rs_models/clas/condensenetv2.py

@@ -20,7 +20,7 @@ Apache License [see LICENSE for details]
 import paddle
 import paddle.nn as nn
 
-__all__ = ["CondenseNetV2_a", "CondenseNetV2_b", "CondenseNetV2_c"]
+__all__ = ["CondenseNetV2_A", "CondenseNetV2_B", "CondenseNetV2_C"]
 
 
 class SELayer(nn.Layer):
@@ -394,7 +394,7 @@ class CondenseNetV2(nn.Layer):
                 nn.initializer.Constant(value=0.0)(m.bias)
 
 
-def CondenseNetV2_a(**kwargs):
+def CondenseNetV2_A(**kwargs):
     model = CondenseNetV2(
         stages=[1, 1, 4, 6, 8],
         growth=[8, 8, 16, 32, 64],
@@ -410,7 +410,7 @@ def CondenseNetV2_a(**kwargs):
     return model
 
 
-def CondenseNetV2_b(**kwargs):
+def CondenseNetV2_B(**kwargs):
     model = CondenseNetV2(
         stages=[2, 4, 6, 8, 6],
         growth=[6, 12, 24, 48, 96],
@@ -426,7 +426,7 @@ def CondenseNetV2_b(**kwargs):
     return model
 
 
-def CondenseNetV2_c(**kwargs):
+def CondenseNetV2_C(**kwargs):
     model = CondenseNetV2(
         stages=[4, 6, 8, 10, 8],
         growth=[8, 16, 32, 64, 128],

paddlers/tasks/classifier.py

@@ -34,9 +34,7 @@ from paddlers.utils.checkpoint import cls_pretrain_weights_dict
 from paddlers.transforms import Resize, decode_image
 from .base import BaseModel
 
-__all__ = [
+__all__ = ["ResNet50_vd", "MobileNetV3", "HRNet", "CondenseNetV2"]
-    "ResNet50_vd", "MobileNetV3_small_x1_0", "HRNet_W18_C", "CondenseNetV2_b"
-]
 
 
 class BaseClassifier(BaseModel):
@@ -600,13 +598,13 @@ class ResNet50_vd(BaseClassifier):
             **params)
 
 
-class MobileNetV3_small_x1_0(BaseClassifier):
+class MobileNetV3(BaseClassifier):
     def __init__(self,
                  num_classes=2,
                  use_mixed_loss=False,
                  losses=None,
                  **params):
-        super(MobileNetV3_small_x1_0, self).__init__(
+        super(MobileNetV3, self).__init__(
             model_name='MobileNetV3_small_x1_0',
             num_classes=num_classes,
             use_mixed_loss=use_mixed_loss,
@@ -614,13 +612,13 @@ class MobileNetV3_small_x1_0(BaseClassifier):
             **params)
 
 
-class HRNet_W18_C(BaseClassifier):
+class HRNet(BaseClassifier):
     def __init__(self,
                  num_classes=2,
                  use_mixed_loss=False,
                  losses=None,
                  **params):
-        super(HRNet_W18_C, self).__init__(
+        super(HRNet, self).__init__(
             model_name='HRNet_W18_C',
             num_classes=num_classes,
             use_mixed_loss=use_mixed_loss,
@@ -628,15 +626,21 @@ class HRNet_W18_C(BaseClassifier):
             **params)
 
 
-class CondenseNetV2_b(BaseClassifier):
+class CondenseNetV2(BaseClassifier):
     def __init__(self,
                  num_classes=2,
                  use_mixed_loss=False,
                  losses=None,
+                 in_chnanels=3,
+                 arch='A',
                  **params):
-        super(CondenseNetV2_b, self).__init__(
+        if arch not in ('A', 'B', 'C'):
-            model_name='CondenseNetV2_b',
+            raise ValueError("{} is not a supported architecture.".format(arch))
+        model_name = 'CondenseNetV2_' + arch
+        super(CondenseNetV2, self).__init__(
+            model_name=model_name,
             num_classes=num_classes,
             use_mixed_loss=use_mixed_loss,
             losses=losses,
+            in_channels=in_channels,
             **params)

test_tipc/configs/clas/condensenetv2/condensenetv2_ucmerced.yaml

@@ -0,0 +1,10 @@
+# Configurations of CondenseNet V2 with UCMerced dataset
+
+_base_: ../_base_/ucmerced.yaml
+
+save_dir: ./test_tipc/output/clas/condensenetv2/
+
+model: !Node
+    type: CondenseNetV2
+    args:
+        num_classes: 21

test_tipc/configs/clas/condensenetv2/train_infer_python.txt

@@ -0,0 +1,53 @@
+===========================train_params===========================
+model_name:clas:condensenetv2
+python:python
+gpu_list:0|0,1
+use_gpu:null|null
+--precision:null
+--num_epochs:lite_train_lite_infer=3|lite_train_whole_infer=3|whole_train_whole_infer=10
+--save_dir:adaptive
+--train_batch_size:lite_train_lite_infer=16|lite_train_whole_infer=16|whole_train_whole_infer=16
+--model_path:null
+--config:lite_train_lite_infer=./test_tipc/configs/clas/condensenetv2/condensenetv2_ucmerced.yaml|lite_train_whole_infer=./test_tipc/configs/clas/condensenetv2/condensenetv2_ucmerced.yaml|whole_train_whole_infer=./test_tipc/configs/clas/condensenetv2/condensenetv2_ucmerced.yaml
+train_model_name:best_model
+null:null
+##
+trainer:norm
+norm_train:test_tipc/run_task.py train clas
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params===========================
+eval:null
+null:null
+##
+===========================export_params===========================
+--save_dir:adaptive
+--model_dir:adaptive
+--fixed_input_shape:[-1,3,256,256]
+norm_export:deploy/export/export_model.py
+quant_export:null
+fpgm_export:null
+distill_export:null
+export1:null
+export2:null
+===========================infer_params===========================
+infer_model:null
+infer_export:null
+infer_quant:False
+inference:test_tipc/infer.py
+--device:cpu|gpu
+--enable_mkldnn:True
+--cpu_threads:6
+--batch_size:1
+--use_trt:False
+--precision:fp32
+--model_dir:null
+--config:null
+--save_log_path:null
+--benchmark:True
+--model_name:condensenetv2
+null:null

test_tipc/configs/clas/hrnet/hrnet.yaml

@@ -1,10 +0,0 @@
-# Basic configurations of HRNet
-
-_base_: ../_base_/ucmerced.yaml
-
-save_dir: ./test_tipc/output/clas/hrnet/
-
-model: !Node
-    type: HRNet_W18_C
-    args:
-        num_classes: 21

test_tipc/configs/clas/hrnet/hrnet_ucmerced.yaml

@@ -5,6 +5,6 @@ _base_: ../_base_/ucmerced.yaml
 save_dir: ./test_tipc/output/clas/hrnet/
 
 model: !Node
-    type: HRNet_W18_C
+    type: HRNet
     args:
         num_classes: 21

tests/rs_models/test_clas_models.py

@@ -18,7 +18,7 @@ from rs_models.test_model import TestModel
 __all__ = []
 
 
-class TestCDModel(TestModel):
+class TestClasModel(TestModel):
     DEFAULT_HW = (224, 224)
 
     def check_output(self, output, target):
@@ -36,3 +36,36 @@ class TestCDModel(TestModel):
     def set_targets(self):
         self.targets = [[self.DEFAULT_BATCH_SIZE, spec.get('num_classes', 2)]
                         for spec in self.specs]
+
+
+class TestCondenseNetV2AModel(TestClasModel):
+    MODEL_CLASS = paddlers.rs_models.clas.CondenseNetV2_A
+
+    def set_specs(self):
+        self.specs = [
+            dict(in_channels=3, num_classes=2),
+            dict(in_channels=10, num_classes=2),
+            dict(in_channels=3, num_classes=100)
+        ]   # yapf: disable
+
+
+class TestCondenseNetV2BModel(TestClasModel):
+    MODEL_CLASS = paddlers.rs_models.clas.CondenseNetV2_B
+
+    def set_specs(self):
+        self.specs = [
+            dict(in_channels=3, num_classes=2),
+            dict(in_channels=10, num_classes=2),
+            dict(in_channels=3, num_classes=100)
+        ]   # yapf: disable
+
+
+class TestCondenseNetV2CModel(TestClasModel):
+    MODEL_CLASS = paddlers.rs_models.clas.CondenseNetV2_C
+
+    def set_specs(self):
+        self.specs = [
+            dict(in_channels=3, num_classes=2),
+            dict(in_channels=10, num_classes=2),
+            dict(in_channels=3, num_classes=100)
+        ]   # yapf: disable

tutorials/train/classification/condensenetv2.py

@@ -0,0 +1,90 @@
+#!/usr/bin/env python
+
+# 场景分类模型CondenseNet V2训练示例脚本
+# 执行此脚本前，请确认已正确安装PaddleRS库
+
+import paddlers as pdrs
+from paddlers import transforms as T
+
+# 数据集存放目录
+DATA_DIR = './data/ucmerced/'
+# 训练集`file_list`文件路径
+TRAIN_FILE_LIST_PATH = './data/ucmerced/train.txt'
+# 验证集`file_list`文件路径
+EVAL_FILE_LIST_PATH = './data/ucmerced/val.txt'
+# 数据集类别信息文件路径
+LABEL_LIST_PATH = './data/ucmerced/labels.txt'
+# 实验目录，保存输出的模型权重和结果
+EXP_DIR = './output/hrnet/'
+
+# 下载和解压UC Merced数据集
+pdrs.utils.download_and_decompress(
+    'https://paddlers.bj.bcebos.com/datasets/ucmerced.zip', path='./data/')
+
+# 定义训练和验证时使用的数据变换（数据增强、预处理等）
+# 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
+# API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
+train_transforms = T.Compose([
+    # 读取影像
+    T.DecodeImg(),
+    # 将影像缩放到256x256大小
+    T.Resize(target_size=256),
+    # 以50%的概率实施随机水平翻转
+    T.RandomHorizontalFlip(prob=0.5),
+    # 以50%的概率实施随机垂直翻转
+    T.RandomVerticalFlip(prob=0.5),
+    # 将数据归一化到[-1,1]
+    T.Normalize(
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+    T.ArrangeClassifier('train')
+])
+
+eval_transforms = T.Compose([
+    T.DecodeImg(),
+    T.Resize(target_size=256),
+    # 验证阶段与训练阶段的数据归一化方式必须相同
+    T.Normalize(
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+    T.ArrangeClassifier('eval')
+])
+
+# 分别构建训练和验证所用的数据集
+train_dataset = pdrs.datasets.ClasDataset(
+    data_dir=DATA_DIR,
+    file_list=TRAIN_FILE_LIST_PATH,
+    label_list=LABEL_LIST_PATH,
+    transforms=train_transforms,
+    num_workers=0,
+    shuffle=True)
+
+eval_dataset = pdrs.datasets.ClasDataset(
+    data_dir=DATA_DIR,
+    file_list=EVAL_FILE_LIST_PATH,
+    label_list=LABEL_LIST_PATH,
+    transforms=eval_transforms,
+    num_workers=0,
+    shuffle=False)
+
+# 构建CondenseNet V2模型
+# 目前已支持的模型请参考：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/intro/model_zoo.md
+# 模型输入参数请参考：https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/tasks/classifier.py
+model = pdrs.tasks.clas.CondenseNetV2(num_classes=len(train_dataset.labels))
+
+# 执行模型训练
+model.train(
+    num_epochs=2,
+    train_dataset=train_dataset,
+    train_batch_size=16,
+    eval_dataset=eval_dataset,
+    save_interval_epochs=1,
+    # 每多少次迭代记录一次日志
+    log_interval_steps=50,
+    save_dir=EXP_DIR,
+    # 初始学习率大小
+    learning_rate=0.01,
+    # 是否使用early stopping策略，当精度不再改善时提前终止训练
+    early_stop=False,
+    # 是否启用VisualDL日志功能
+    use_vdl=True,
+    # 指定从某个检查点继续训练
+    resume_checkpoint=None)