Revert "update batch evaluate (#154)"

This reverts commit b5c716c.

paddlers/tasks/change_detector.py

@@ -29,7 +29,7 @@
 import paddlers.utils.logging as logging
 from paddlers.models import seg_losses
 from paddlers.transforms import Resize, decode_image, construct_sample
-from paddlers.utils import to_data_parallel
+from paddlers.utils import get_single_card_bs
 from paddlers.utils.checkpoint import cd_pretrain_weights_dict
 from .base import BaseModel
 from .utils import seg_metrics as metrics
@@ -447,22 +447,25 @@ def evaluate(self, eval_dataset, batch_size=1, return_details=False):
         """
 
         self._check_transforms(eval_dataset.transforms)
-        net = self.net
-        net.eval()
 
-        # XXX: Hard-coding
+        self.net.eval()
         nranks = paddle.distributed.get_world_size()
+        local_rank = paddle.distributed.get_rank()
         if nranks > 1:
             # Initialize parallel environment if not done.
-            if not paddle.distributed.parallel.parallel_helper._is_parallel_ctx_initialized(
-            ):
+            if not (paddle.distributed.parallel.parallel_helper.
+                    _is_parallel_ctx_initialized()):
                 paddle.distributed.init_parallel_env()
-                net = to_data_parallel(
-                    net, find_unused_parameters=self.find_unused_parameters)
-            else:
-                net = to_data_parallel(
-                    net, find_unused_parameters=self.find_unused_parameters)
 
+        batch_size_each_card = get_single_card_bs(batch_size)
+        if batch_size_each_card > 1:
+            batch_size_each_card = 1
+            batch_size = batch_size_each_card * paddlers.env_info['num']
+            logging.warning(
+                "ChangeDetector only supports batch_size=1 for each gpu/cpu card " \
+                "during evaluation, so batch_size " \
+                "is forcibly set to {}.".format(batch_size)
+            )
         self.eval_data_loader = self.build_data_loader(
             eval_dataset, batch_size=batch_size, mode='eval')
 
@@ -482,9 +485,9 @@ def evaluate(self, eval_dataset, batch_size=1, return_details=False):
                             enable=True,
                             custom_white_list=self.custom_white_list,
                             custom_black_list=self.custom_black_list):
-                        outputs = self.run(net, data, 'eval')
+                        outputs = self.run(self.net, data, 'eval')
                 else:
-                    outputs = self.run(net, data, 'eval')
+                    outputs = self.run(self.net, data, 'eval')
                 pred_area = outputs['pred_area']
                 label_area = outputs['label_area']
                 intersect_area = outputs['intersect_area']
@@ -691,8 +694,6 @@ def postprocess(self, batch_pred, batch_restore_list):
                 else:
                     raise RuntimeError
             results.append(pred)
-        if len(results) > 1:
-            results = [paddle.concat(results, axis=0)]
         return results
 
     def _infer_postprocess(self, batch_label_map, batch_score_map,

paddlers/tasks/classifier.py

@@ -25,7 +25,6 @@
 import paddlers.rs_models.clas as cmcls
 import paddlers.utils.logging as logging
 from paddlers.models.ppcls.metric import build_metrics
-from paddlers.utils import to_data_parallel
 from paddlers.models import clas_losses
 from paddlers.models.ppcls.data.postprocess import build_postprocess
 from paddlers.utils.checkpoint import cls_pretrain_weights_dict
@@ -403,67 +402,54 @@ def evaluate(self, eval_dataset, batch_size=1, return_details=False):
         """
 
         self._check_transforms(eval_dataset.transforms)
-        net = self.net
-        net.eval()
 
-        # XXX: Hard-coding
+        self.net.eval()
         nranks = paddle.distributed.get_world_size()
+        local_rank = paddle.distributed.get_rank()
         if nranks > 1:
             # Initialize parallel environment if not done.
             if not paddle.distributed.parallel.parallel_helper._is_parallel_ctx_initialized(
             ):
                 paddle.distributed.init_parallel_env()
-                net = to_data_parallel(
-                    net, find_unused_parameters=self.find_unused_parameters)
-            else:
-                net = to_data_parallel(
-                    net, find_unused_parameters=self.find_unused_parameters)
-
-        self.eval_data_loader = self.build_data_loader(
-            eval_dataset, batch_size=batch_size, mode='eval')
-        logging.info("Start to evaluate (total_samples={}, total_steps={})...".
-                     format(eval_dataset.num_samples, eval_dataset.num_samples))
-
-        top1s = []
-        top5s = []
-        with paddle.no_grad():
-            for step, data in enumerate(self.eval_data_loader):
-                if self.precision == 'fp16':
-                    with paddle.amp.auto_cast(
-                            level=self.amp_level,
-                            enable=True,
-                            custom_white_list=self.custom_white_list,
-                            custom_black_list=self.custom_black_list):
-                        outputs = self.run(net, data, 'eval')
-                else:
-                    outputs = self.run(net, data, 'eval')
-                if nranks > 1:
-                    t1 = outputs["top1"]
-                    t5 = outputs["top5"]
-                    t1s = []
-                    t5s = []
-                    paddle.distributed.all_gather(t1s, t1)
-                    paddle.distributed.all_gather(t5s, t5)
-                    for rank_id in range(nranks):
-                        top1 = t1s[rank_id]
-                        top5 = t5s[rank_id]
-                        for i in range(data['image'].shape[0]):
-                            top1s.append(top1)
-                            top5s.append(top5)
-                else:
-                    for i in range(data['image'].shape[0]):
-                        top1s.append(outputs["top1"])
-                        top5s.append(outputs["top5"])
-
-        top1 = np.mean(top1s)
-        top5 = np.mean(top5s)
-        eval_metrics = OrderedDict(zip(['top1', 'top5'], [top1, top5]))
-
-        if return_details:
-            # TODO: Add details
-            return eval_metrics, None
 
-        return eval_metrics
+        if batch_size > 1:
+            logging.warning(
+                "Classifier only supports single card evaluation with batch_size=1 "
+                "during evaluation, so batch_size is forcibly set to 1.")
+            batch_size = 1
+
+        if nranks < 2 or local_rank == 0:
+            self.eval_data_loader = self.build_data_loader(
+                eval_dataset, batch_size=batch_size, mode='eval')
+            logging.info(
+                "Start to evaluate (total_samples={}, total_steps={})...".
+                format(eval_dataset.num_samples, eval_dataset.num_samples))
+
+            top1s = []
+            top5s = []
+            with paddle.no_grad():
+                for step, data in enumerate(self.eval_data_loader):
+                    if self.precision == 'fp16':
+                        with paddle.amp.auto_cast(
+                                level=self.amp_level,
+                                enable=True,
+                                custom_white_list=self.custom_white_list,
+                                custom_black_list=self.custom_black_list):
+                            outputs = self.run(self.net, data, 'eval')
+                    else:
+                        outputs = self.run(self.net, data, 'eval')
+                    top1s.append(outputs["top1"])
+                    top5s.append(outputs["top5"])
+
+            top1 = np.mean(top1s)
+            top5 = np.mean(top5s)
+            eval_metrics = OrderedDict(zip(['top1', 'top5'], [top1, top5]))
+
+            if return_details:
+                # TODO: Add details
+                return eval_metrics, None
+
+            return eval_metrics
 
     @paddle.no_grad()
     def predict(self, img_file, transforms=None):