Upgrade supported paddle version

Dockerfile

@@ -1,5 +1,5 @@
 # 0. set args
-ARG PPTAG=2.4.1  # tags refer to https://hub.docker.com/r/paddlepaddle/paddle/tags
+ARG PPTAG=2.5.1  # tags refer to https://hub.docker.com/r/paddlepaddle/paddle/tags
 
 # 1. pull base image
 FROM paddlepaddle/paddle:${PPTAG}

docs/docker_cn.md

@@ -4,16 +4,16 @@
 
 PaddleRS提供`Dockerfile`，可构建基础镜像用于开发或部署。在镜像构建过程中，默认将拉取PaddleRS develop分支内容，并存放在`/opt/PaddleRS`。在构建镜像时可以通过`PPTAG`参数指定要使用的PaddlePaddle版本，例如：
 
-- 安装CPU版本的PaddlePaddle-2.4.1，未指定`PPTAG`的情况下将默认安装此版本：
+- 安装CPU版本的PaddlePaddle-2.5.1，未指定`PPTAG`的情况下将默认安装此版本：
 
 ```shell
 docker build -t paddlers:latest -f Dockerfile .
 ```
 
-- 安装GPU版本PaddlePaddle-2.4.1，使用CUDA 11.7、cuDNN 8.4以及TensorRT 8.4：
+- 安装GPU版本PaddlePaddle-2.5.1，使用CUDA 11.7、cuDNN 8.4以及TensorRT 8.4：
 
 ```shell
-docker build -t paddlers:latest -f Dockerfile . --build-arg PPTAG=2.4.1-gpu-cuda11.7-cudnn8.4-trt8.4
+docker build -t paddlers:latest -f Dockerfile . --build-arg PPTAG=2.5.1-gpu-cuda11.7-cudnn8.4-trt8.4
 ```
 
 其他环境的`PPTAG`可以参考[此处](https://hub.docker.com/r/paddlepaddle/paddle/tags)。请注意，如果需要安装GPU版本的PaddlePaddle，请确保Docker版本>=19。

docs/docker_en.md

@@ -4,16 +4,16 @@
 
 PaddleRS provides `Dockerfile` to build a base Docker image for development/deployment. By default the develop branch of PaddleRS is fetched and stored in `/opts/PaddleRS` during image build. The `PPTAG` argument can be specified to the PaddlePaddle version you want to install. For example,
 
-- To install CPU version of PaddlePaddle-2.4.1 (which is installed when `docker build` does not receive a `PPTAG` argument), run:
+- To install CPU version of PaddlePaddle-2.5.1 (which is installed when `docker build` does not receive a `PPTAG` argument), run:
 
 ```shell
 docker build -t paddlers:latest -f Dockerfile .
 ```
 
-- To install GPU version of PaddlePaddle-2.4.1, with CUDA 11.8, cuDNN 8.4, and TensorRT 8.4, run:
+- To install GPU version of PaddlePaddle-2.5.1, with CUDA 11.8, cuDNN 8.4, and TensorRT 8.4, run:
 
 ```shell
-docker build -t paddlers:latest -f Dockerfile . --build-arg PPTAG=2.4.1-gpu-cuda11.7-cudnn8.4-trt8.4
+docker build -t paddlers:latest -f Dockerfile . --build-arg PPTAG=2.5.1-gpu-cuda11.7-cudnn8.4-trt8.4
 ```
 
 You can find a full list of available PaddlePaddle versions [here](https://hub.docker.com/r/paddlepaddle/paddle/tags). Please note that if a GPU version of PaddlePaddle is to be used, the version of Docker should >=19.

docs/quick_start_cn.md

@@ -5,14 +5,13 @@
 ## 环境准备
 
 1. [安装PaddlePaddle](https://www.paddlepaddle.org.cn/install/quick)
-  - 版本要求：PaddlePaddle>=2.2.0
+  - 版本要求：PaddlePaddle>=2.5.0
 
 2. 安装PaddleRS
 
 如果希望获取更加稳定的体验，请下载安装[PaddleRS发行版](https://github.com/PaddlePaddle/PaddleRS/releases)。
 
 ```shell
-pip install -r requirements.txt
 pip install .
 ```
 
@@ -22,7 +21,6 @@ PaddleRS代码会跟随开发进度不断更新，如果希望使用最新功能
 git clone https://github.com/PaddlePaddle/PaddleRS
 cd PaddleRS
 git checkout develop
-pip install -r requirements.txt
 pip install .
 ```
 

docs/quick_start_en.md

@@ -5,7 +5,7 @@
 ## Prerequisites
 
 1. [Install PaddlePaddle](https://www.paddlepaddle.org.cn/install/quick)
-  - Version requirements: PaddlePaddle>=2.2.0
+  - Version requirements: PaddlePaddle>=2.5.0
 
 2. Install PaddleRS
 

paddlers/datasets/base.py

@@ -15,7 +15,7 @@
 import copy
 
 from paddle.io import Dataset
-from paddle.fluid.dataloader.collate import default_collate_fn
+from paddle.io.dataloader.collate import default_collate_fn
 
 from paddlers.utils import get_num_workers
 import paddlers.utils.logging as logging

paddlers/models/paddleseg/__init__.py

@@ -12,6 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from . import models, datasets, transforms
+from . import models, datasets, transforms, optimizers
 
-__version__ = '2.7.0'
+__version__ = '2.8.0'

paddlers/models/paddleseg/core/infer.py

@@ -95,8 +95,8 @@ def slide_inference(model, im, crop_size, stride):
     w_crop, h_crop = crop_size
     w_stride, h_stride = stride
     # calculate the crop nums
-    rows = np.int(np.ceil(1.0 * (h_im - h_crop) / h_stride)) + 1
-    cols = np.int(np.ceil(1.0 * (w_im - w_crop) / w_stride)) + 1
+    rows = int(np.ceil(1.0 * (h_im - h_crop) / h_stride)) + 1
+    cols = int(np.ceil(1.0 * (w_im - w_crop) / w_stride)) + 1
     # prevent negative sliding rounds when imgs after scaling << crop_size
     rows = 1 if h_im <= h_crop else rows
     cols = 1 if w_im <= w_crop else cols
@@ -208,12 +208,13 @@ def aug_inference(model,
     final_logit = 0
     h_input, w_input = im.shape[-2], im.shape[-1]
     flip_comb = flip_combination(flip_horizontal, flip_vertical)
+    num_augs = len(scales) * len(flip_comb)
     for scale in scales:
         h = int(h_input * scale + 0.5)
         w = int(w_input * scale + 0.5)
-        im = F.interpolate(im, [h, w], mode='bilinear')
+        im_scale = F.interpolate(im, [h, w], mode='bilinear')
         for flip in flip_comb:
-            im_flip = tensor_flip(im, flip)
+            im_flip = tensor_flip(im_scale, flip)
             logit = inference(
                 model,
                 im_flip,
@@ -222,10 +223,11 @@ def aug_inference(model,
                 stride=stride)
             logit = tensor_flip(logit, flip)
             logit = F.interpolate(logit, [h_input, w_input], mode='bilinear')
-
-            logit = F.softmax(logit, axis=1)
-            final_logit = final_logit + logit
-
+            # Accumulate final logits in place
+            final_logit += logit
+    # We average the accumulated logits to make the numeric values of `final_logit`
+    # comparable to single-scale logits
+    final_logit /= num_augs
     final_logit = reverse_transform(final_logit, trans_info, mode='bilinear')
     pred = paddle.argmax(final_logit, axis=1, keepdim=True, dtype='int32')
 

paddlers/models/paddleseg/core/predict.py

@@ -145,3 +145,6 @@ def predict(model,
             pred_mask.save(pred_saved_path)
 
             progbar_pred.update(i + 1)
+
+    logger.info("Predicted images are saved in {} and {} .".format(
+        added_saved_dir, pred_saved_dir))

paddlers/models/paddleseg/core/train.py

@@ -16,12 +16,14 @@
 import time
 from collections import deque
 import shutil
+from copy import deepcopy
 
 import paddle
 import paddle.nn.functional as F
 
 from paddlers.models.paddleseg.utils import (TimeAverager, calculate_eta, resume, logger,
-                             worker_init_fn, train_profiler, op_flops_funs)
+                             worker_init_fn, train_profiler, op_flops_funs,
+                             init_ema_params, update_ema_model)
 from paddlers.models.paddleseg.core.val import evaluate
 
 
@@ -68,6 +70,7 @@ def train(model,
           log_iters=10,
           num_workers=0,
           use_vdl=False,
+          use_ema=False,
           losses=None,
           keep_checkpoint_max=5,
           test_config=None,
@@ -102,6 +105,13 @@ def train(model,
         profiler_options (str, optional): The option of train profiler.
         to_static_training (bool, optional): Whether to use @to_static for training.
     """
+
+    if use_ema:
+        ema_model = deepcopy(model)
+        ema_model.eval()
+        for param in ema_model.parameters():
+            param.stop_gradient = True
+
     model.train()
     nranks = paddle.distributed.ParallelEnv().nranks
     local_rank = paddle.distributed.ParallelEnv().local_rank
@@ -154,14 +164,16 @@ def train(model,
     avg_loss_list = []
     iters_per_epoch = len(batch_sampler)
     best_mean_iou = -1.0
+    best_ema_mean_iou = -1.0
     best_model_iter = -1
     reader_cost_averager = TimeAverager()
     batch_cost_averager = TimeAverager()
     save_models = deque()
     batch_start = time.time()
-
     iter = start_iter
     while iter < iters:
+        if iter == start_iter and use_ema:
+            init_ema_params(ema_model, model)
         for data in loader:
             iter += 1
             if iter > iters:
@@ -176,6 +188,7 @@ def train(model,
             edges = None
             if 'edge' in data.keys():
                 edges = data['edge'].astype('int64')
+
             if hasattr(model, 'data_format') and model.data_format == 'NHWC':
                 images = images.transpose((0, 2, 3, 1))
 
@@ -189,11 +202,19 @@ def train(model,
                         custom_black_list={'bilinear_interp_v2'}):
                     logits_list = ddp_model(images) if nranks > 1 else model(
                         images)
-                    loss_list = loss_computation(
-                        logits_list=logits_list,
-                        labels=labels,
-                        edges=edges,
-                        losses=losses)
+                    if nranks > 1 and hasattr(ddp_model._layers,
+                                              'loss_computation'):
+                        loss_list = ddp_model._layers.loss_computation(
+                            logits_list, losses, data)
+                    elif nranks == 1 and hasattr(model, 'loss_computation'):
+                        loss_list = model.loss_computation(logits_list, losses,
+                                                           data)
+                    else:
+                        loss_list = loss_computation(
+                            logits_list=logits_list,
+                            labels=labels,
+                            edges=edges,
+                            losses=losses)
                     loss = sum(loss_list)
 
                 scaled = scaler.scale(loss)  # scale the loss
@@ -204,18 +225,23 @@ def train(model,
                     scaler.minimize(optimizer, scaled)  # update parameters
             else:
                 logits_list = ddp_model(images) if nranks > 1 else model(images)
-                loss_list = loss_computation(
-                    logits_list=logits_list,
-                    labels=labels,
-                    edges=edges,
-                    losses=losses)
+
+                if nranks > 1 and hasattr(ddp_model._layers,
+                                          'loss_computation'):
+                    loss_list = ddp_model._layers.loss_computation(logits_list,
+                                                                   losses, data)
+                elif nranks == 1 and hasattr(model, 'loss_computation'):
+                    loss_list = model.loss_computation(logits_list, losses,
+                                                       data)
+                else:
+                    loss_list = loss_computation(
+                        logits_list=logits_list,
+                        labels=labels,
+                        edges=edges,
+                        losses=losses)
                 loss = sum(loss_list)
                 loss.backward()
-                # if the optimizer is ReduceOnPlateau, the loss is the one which has been pass into step.
-                if isinstance(optimizer, paddle.optimizer.lr.ReduceOnPlateau):
-                    optimizer.step(loss)
-                else:
-                    optimizer.step()
+                optimizer.step()
 
             lr = optimizer.get_lr()
 
@@ -225,7 +251,10 @@ def train(model,
             else:
                 lr_sche = optimizer._learning_rate
             if isinstance(lr_sche, paddle.optimizer.lr.LRScheduler):
-                lr_sche.step()
+                if isinstance(lr_sche, paddle.optimizer.lr.ReduceOnPlateau):
+                    lr_sche.step(loss)
+                else:
+                    lr_sche.step()
 
             train_profiler.add_profiler_step(profiler_options)
 
@@ -273,6 +302,9 @@ def train(model,
                 reader_cost_averager.reset()
                 batch_cost_averager.reset()
 
+            if use_ema:
+                update_ema_model(ema_model, model, step=iter)
+
             if (iter % save_interval == 0 or
                     iter == iters) and (val_dataset is not None):
                 num_workers = 1 if num_workers > 0 else 0
@@ -288,6 +320,15 @@ def train(model,
                     amp_level=amp_level,
                     **test_config)
 
+                if use_ema:
+                    ema_mean_iou, ema_acc, _, _, _ = evaluate(
+                        ema_model,
+                        val_dataset,
+                        num_workers=num_workers,
+                        precision=precision,
+                        amp_level=amp_level,
+                        **test_config)
+
                 model.train()
 
             if (iter % save_interval == 0 or iter == iters) and local_rank == 0:
@@ -299,6 +340,12 @@ def train(model,
                             os.path.join(current_save_dir, 'model.pdparams'))
                 paddle.save(optimizer.state_dict(),
                             os.path.join(current_save_dir, 'model.pdopt'))
+
+                if use_ema:
+                    paddle.save(
+                        ema_model.state_dict(),
+                        os.path.join(current_save_dir, 'ema_model.pdparams'))
+
                 save_models.append(current_save_dir)
                 if len(save_models) > keep_checkpoint_max > 0:
                     model_to_remove = save_models.popleft()
@@ -315,10 +362,28 @@ def train(model,
                     logger.info(
                         '[EVAL] The model with the best validation mIoU ({:.4f}) was saved at iter {}.'
                         .format(best_mean_iou, best_model_iter))
+                    if use_ema:
+                        if ema_mean_iou > best_ema_mean_iou:
+                            best_ema_mean_iou = ema_mean_iou
+                            best_ema_model_iter = iter
+                            best_ema_model_dir = os.path.join(save_dir,
+                                                              "ema_best_model")
+                            paddle.save(ema_model.state_dict(),
+                                        os.path.join(best_ema_model_dir,
+                                                     'ema_model.pdparams'))
+                        logger.info(
+                            '[EVAL] The EMA model with the best validation mIoU ({:.4f}) was saved at iter {}.'
+                            .format(best_ema_mean_iou, best_ema_model_iter))
 
                     if use_vdl:
                         log_writer.add_scalar('Evaluate/mIoU', mean_iou, iter)
                         log_writer.add_scalar('Evaluate/Acc', acc, iter)
+
+                        if use_ema:
+                            log_writer.add_scalar('Evaluate/Ema_mIoU',
+                                                  ema_mean_iou, iter)
+                            log_writer.add_scalar('Evaluate/Ema_Acc', ema_acc,
+                                                  iter)
             batch_start = time.time()
 
     # Calculate flops.

paddlers/models/paddleseg/cvlibs/__init__.py

@@ -15,3 +15,4 @@
 from . import manager
 from . import param_init
 from .config import Config
+from .builder import Builder, SegBuilder