add EDVR predictor dynamic (PaddlePaddle#315)

applications/tools/video-enhance.py

@@ -119,10 +119,8 @@
                                         weight_path=args.RealSR_weight)
             frames_path, temp_video_path = predictor.run(temp_video_path)
         elif order == 'EDVR':
-            paddle.enable_static()
             predictor = EDVRPredictor(args.output, weight_path=args.EDVR_weight)
             frames_path, temp_video_path = predictor.run(temp_video_path)
-            paddle.disable_static()
 
         print('Model {} output frames path:'.format(order), frames_path)
         print('Model {} output video path:'.format(order), temp_video_path)

ppgan/apps/edvr_predictor.py

@@ -19,12 +19,15 @@
 import numpy as np
 from tqdm import tqdm
 
+import paddle
+from paddle.io import Dataset, DataLoader
+
 from ppgan.utils.download import get_path_from_url
 from ppgan.utils.video import frames2video, video2frames
-
+from ppgan.models.generators import EDVRNet
 from .base_predictor import BasePredictor
 
-EDVR_WEIGHT_URL = 'https://paddlegan.bj.bcebos.com/applications/edvr_infer_model.tar'
+EDVR_WEIGHT_URL = 'https://paddlegan.bj.bcebos.com/models/EDVR_L_w_tsa_SRx4.pdparams'
 
 
 def get_img(pred):
@@ -110,7 +113,7 @@ def get_test_neighbor_frames(crt_i, N, max_n, padding='new_info'):
     return return_l
 
 
-class EDVRDataset:
+class EDVRDataset(Dataset):
     def __init__(self, frame_paths):
         self.frames = frame_paths
 
@@ -133,16 +136,15 @@ def __len__(self):
 
 
 class EDVRPredictor(BasePredictor):
-    def __init__(self, output='output', weight_path=None):
+    def __init__(self, output='output', weight_path=None, bs=1):
         self.input = input
         self.output = os.path.join(output, 'EDVR')
-
+        self.bs = bs
+        self.model = EDVRNet(nf=128, back_RBs=40)
         if weight_path is None:
             weight_path = get_path_from_url(EDVR_WEIGHT_URL)
-
-        self.weight_path = weight_path
-
-        self.build_inference_model()
+        self.model.set_dict(paddle.load(weight_path)['generator'])
+        self.model.eval()
 
     def run(self, video_path):
         vid = video_path
@@ -163,23 +165,23 @@ def run(self, video_path):
 
         frames = sorted(glob.glob(os.path.join(out_path, '*.png')))
 
-        dataset = EDVRDataset(frames)
+        test_dataset = EDVRDataset(frames)
+        dataset = DataLoader(test_dataset, batch_size=self.bs, num_workers=2)
 
         periods = []
         cur_time = time.time()
         for infer_iter, data in enumerate(tqdm(dataset)):
-            data_feed_in = [data[0]]
-
-            outs = self.base_forward(np.array(data_feed_in))
-
-            infer_result_list = [item for item in outs]
-
+            data_feed_in = paddle.to_tensor(data[0])
+            with paddle.no_grad():
+                outs = self.model(data_feed_in).numpy()
+            infer_result_list = [outs[i, :, :, :] for i in range(self.bs)]
             frame_path = data[1]
-
-            img_i = get_img(infer_result_list[0])
-            save_img(
-                img_i,
-                os.path.join(pred_frame_path, os.path.basename(frame_path)))
+            for i in range(self.bs):
+                img_i = get_img(infer_result_list[i])
+                save_img(
+                    img_i,
+                    os.path.join(pred_frame_path,
+                                 os.path.basename(frame_path[i])))
 
             prev_time = cur_time
             cur_time = time.time()