[dev] add amp eval (#6400)

cast dtype in load_pretrain_weight

ppdet/engine/trainer.py

@@ -65,6 +65,8 @@ class Trainer(object):
         self.mode = mode.lower()
         self.optimizer = None
         self.is_loaded_weights = False
+        self.use_amp = self.cfg.get('amp', False)
+        self.amp_level = self.cfg.get('amp_level', 'O1')
 
         # build data loader
         capital_mode = self.mode.capitalize()
@@ -124,17 +126,6 @@ class Trainer(object):
         else:
             self.model.load_meanstd(cfg['TestReader']['sample_transforms'])
 
-        self.use_ema = ('use_ema' in cfg and cfg['use_ema'])
-        if self.use_ema:
-            ema_decay = self.cfg.get('ema_decay', 0.9998)
-            cycle_epoch = self.cfg.get('cycle_epoch', -1)
-            ema_decay_type = self.cfg.get('ema_decay_type', 'threshold')
-            self.ema = ModelEMA(
-                self.model,
-                decay=ema_decay,
-                ema_decay_type=ema_decay_type,
-                cycle_epoch=cycle_epoch)
-
         # EvalDataset build with BatchSampler to evaluate in single device
         # TODO: multi-device evaluate
         if self.mode == 'eval':
@@ -162,6 +153,20 @@ class Trainer(object):
                 self.pruner = create('UnstructuredPruner')(self.model,
                                                            steps_per_epoch)
 
+        if self.use_amp and self.amp_level == 'O2':
+            self.model = paddle.amp.decorate(
+                models=self.model, level=self.amp_level)
+        self.use_ema = ('use_ema' in cfg and cfg['use_ema'])
+        if self.use_ema:
+            ema_decay = self.cfg.get('ema_decay', 0.9998)
+            cycle_epoch = self.cfg.get('cycle_epoch', -1)
+            ema_decay_type = self.cfg.get('ema_decay_type', 'threshold')
+            self.ema = ModelEMA(
+                self.model,
+                decay=ema_decay,
+                ema_decay_type=ema_decay_type,
+                cycle_epoch=cycle_epoch)
+
         self._nranks = dist.get_world_size()
         self._local_rank = dist.get_rank()
 
@@ -387,13 +392,10 @@ class Trainer(object):
             model = paddle.nn.SyncBatchNorm.convert_sync_batchnorm(model)
 
         # enabel auto mixed precision mode
-        use_amp = self.cfg.get('amp', False)
-        amp_level = self.cfg.get('amp_level', 'O1')
-        if use_amp:
+        if self.use_amp:
             scaler = paddle.amp.GradScaler(
                 enable=self.cfg.use_gpu or self.cfg.use_npu,
                 init_loss_scaling=self.cfg.get('init_loss_scaling', 1024))
-            model = paddle.amp.decorate(models=model, level=amp_level)
         # get distributed model
         if self.cfg.get('fleet', False):
             model = fleet.distributed_model(model)
@@ -438,9 +440,9 @@ class Trainer(object):
                 self._compose_callback.on_step_begin(self.status)
                 data['epoch_id'] = epoch_id
 
-                if use_amp:
+                if self.use_amp:
                     with paddle.amp.auto_cast(
-                            enable=self.cfg.use_gpu, level=amp_level):
+                            enable=self.cfg.use_gpu, level=self.amp_level):
                         # model forward
                         outputs = model(data)
                         loss = outputs['loss']
@@ -532,7 +534,12 @@ class Trainer(object):
             self.status['step_id'] = step_id
             self._compose_callback.on_step_begin(self.status)
             # forward
-            outs = self.model(data)
+            if self.use_amp:
+                with paddle.amp.auto_cast(
+                        enable=self.cfg.use_gpu, level=self.amp_level):
+                    outs = self.model(data)
+            else:
+                outs = self.model(data)
 
             # update metrics
             for metric in self._metrics:

ppdet/modeling/assigners/utils.py

@@ -176,7 +176,8 @@ def compute_max_iou_gt(ious):
 def generate_anchors_for_grid_cell(feats,
                                    fpn_strides,
                                    grid_cell_size=5.0,
-                                   grid_cell_offset=0.5):
+                                   grid_cell_offset=0.5,
+                                   dtype='float32'):
     r"""
     Like ATSS, generate anchors based on grid size.
     Args:
@@ -206,16 +207,15 @@ def generate_anchors_for_grid_cell(feats,
                 shift_x - cell_half_size, shift_y - cell_half_size,
                 shift_x + cell_half_size, shift_y + cell_half_size
             ],
-            axis=-1).astype(feat.dtype)
-        anchor_point = paddle.stack(
-            [shift_x, shift_y], axis=-1).astype(feat.dtype)
+            axis=-1).astype(dtype)
+        anchor_point = paddle.stack([shift_x, shift_y], axis=-1).astype(dtype)
 
         anchors.append(anchor.reshape([-1, 4]))
         anchor_points.append(anchor_point.reshape([-1, 2]))
         num_anchors_list.append(len(anchors[-1]))
         stride_tensor.append(
             paddle.full(
-                [num_anchors_list[-1], 1], stride, dtype=feat.dtype))
+                [num_anchors_list[-1], 1], stride, dtype=dtype))
     anchors = paddle.concat(anchors)
     anchors.stop_gradient = True
     anchor_points = paddle.concat(anchor_points)

ppdet/modeling/heads/ppyoloe_head.py

@@ -160,7 +160,7 @@ class PPYOLOEHead(nn.Layer):
             num_anchors_list, stride_tensor
         ], targets)
 
-    def _generate_anchors(self, feats=None):
+    def _generate_anchors(self, feats=None, dtype='float32'):
         # just use in eval time
         anchor_points = []
         stride_tensor = []
@@ -175,11 +175,9 @@ class PPYOLOEHead(nn.Layer):
             shift_y, shift_x = paddle.meshgrid(shift_y, shift_x)
             anchor_point = paddle.cast(
                 paddle.stack(
-                    [shift_x, shift_y], axis=-1), dtype='float32')
+                    [shift_x, shift_y], axis=-1), dtype=dtype)
             anchor_points.append(anchor_point.reshape([-1, 2]))
-            stride_tensor.append(
-                paddle.full(
-                    [h * w, 1], stride, dtype='float32'))
+            stride_tensor.append(paddle.full([h * w, 1], stride, dtype=dtype))
         anchor_points = paddle.concat(anchor_points)
         stride_tensor = paddle.concat(stride_tensor)
         return anchor_points, stride_tensor

tools/eval.py

@@ -77,6 +77,12 @@ def parse_args():
         default=False,
         help='Whether to save the evaluation results only')
 
+    parser.add_argument(
+        "--amp",
+        action='store_true',
+        default=False,
+        help="Enable auto mixed precision eval.")
+
     args = parser.parse_args()
     return args