pad input to use tensor core (#4911)

PaddleCV/image_classification/dali.py

@@ -43,7 +43,8 @@ class HybridTrainPipe(Pipeline):
                  num_shards=1,
                  random_shuffle=True,
                  num_threads=4,
-                 seed=42):
+                 seed=42,
+                 pad_output=False):
         super(HybridTrainPipe, self).__init__(
             batch_size, num_threads, device_id, seed=seed)
         self.input = ops.FileReader(
@@ -73,7 +74,8 @@ class HybridTrainPipe(Pipeline):
             crop=(crop, crop),
             image_type=types.RGB,
             mean=mean,
-            std=std)
+            std=std,
+            pad_output=pad_output)
         self.coin = ops.CoinFlip(probability=0.5)
         self.to_int64 = ops.Cast(dtype=types.INT64, device="gpu")
 
@@ -104,7 +106,8 @@ class HybridValPipe(Pipeline):
                  num_shards=1,
                  random_shuffle=False,
                  num_threads=4,
-                 seed=42):
+                 seed=42,
+                 pad_output=False):
         super(HybridValPipe, self).__init__(
             batch_size, num_threads, device_id, seed=seed)
         self.input = ops.FileReader(
@@ -123,7 +126,8 @@ class HybridValPipe(Pipeline):
             crop=(crop, crop),
             image_type=types.RGB,
             mean=mean,
-            std=std)
+            std=std,
+            pad_output=pad_output)
         self.to_int64 = ops.Cast(dtype=types.INT64, device="gpu")
 
     def define_graph(self):
@@ -169,6 +173,9 @@ def build(settings, mode='train'):
     }
     assert interp in interp_map, "interpolation method not supported by DALI"
     interp = interp_map[interp]
+    pad_output = False
+    if settings.image_shape[0] == 4:
+        pad_output = True
 
     if mode != 'train':
         p = fluid.framework.cuda_places()[0]
@@ -188,7 +195,8 @@ def build(settings, mode='train'):
             interp,
             mean,
             std,
-            device_id=device_id)
+            device_id=device_id,
+            pad_output=pad_output)
         pipe.build()
         return DALIGenericIterator(
             pipe, ['feed_image', 'feed_label'],
@@ -221,7 +229,8 @@ def build(settings, mode='train'):
             device_id,
             shard_id,
             num_shards,
-            seed=42 + shard_id)
+            seed=42 + shard_id,
+            pad_output=pad_output)
         pipe.build()
         pipelines = [pipe]
         sample_per_shard = len(pipe) // num_shards
@@ -248,7 +257,8 @@ def build(settings, mode='train'):
                 device_id,
                 idx,
                 num_shards,
-                seed=42 + idx)
+                seed=42 + idx,
+                pad_output=pad_output)
             pipe.build()
             pipelines.append(pipe)
         sample_per_shard = len(pipelines[0])

PaddleCV/image_classification/reader.py

@@ -245,6 +245,9 @@ def process_image(sample, settings, mode, color_jitter, rotate):
     img_std = np.array(std).reshape((3, 1, 1))
     img -= img_mean
     img /= img_std
+    if settings.image_shape[0] == 4:
+        pad0 = np.zeros((1, img.shape[1], img.shape[2]))
+        img = np.concatenate((img, pad0), axis=0)
     # doing training (train.py)
     if mode == 'train' or (mode == 'val' and
                            not hasattr(settings, 'save_json_path')):

PaddleCV/image_classification/scripts/train/ResNet50_fp16.sh

@@ -19,7 +19,7 @@ python train.py \
        --data_dir=${DATA_DIR} \
        --batch_size=256 \
        --total_images=1281167 \
-       --image_shape 3 224 224 \
+       --image_shape 4 224 224 \
        --class_dim=1000 \
        --print_step=10 \
        --model_save_dir=output/ \