Open-Sora: Democratizing Efficient Video Production for All
Python
train dataset support new format
Zangwei Zheng13423c57a7bc
diff --git a/opensora/datasets/datasets.py b/opensora/datasets/datasets.pyindex 7e4baf5..3f2b71c 100644--- a/opensora/datasets/datasets.py+++ b/opensora/datasets/datasets.py@@ -1,7 +1,7 @@-import csvimport osimport numpy as np+import pandas as pdimport torchimport torchvisionfrom torchvision.datasets.folder import IMG_EXTENSIONS, pil_loader@@ -28,11 +28,9 @@ class DatasetFromCSV(torch.utils.data.Dataset):root=None,):self.csv_path = csv_path- with open(csv_path, "r") as f:- reader = csv.reader(f)- self.samples = list(reader)+ self.data = pd.read_csv(csv_path)- ext = self.samples[0][0].split(".")[-1]+ ext = self.data["path"][0].split(".")[-1]if ext.lower() in VID_EXTENSIONS:self.is_video = Trueelse:@@ -40,18 +38,17 @@ class DatasetFromCSV(torch.utils.data.Dataset):self.is_video = Falseself.transform = transform-self.num_frames = num_framesself.frame_interval = frame_intervalself.temporal_sample = video_transforms.TemporalRandomCrop(num_frames * frame_interval)self.root = rootdef getitem(self, index):- sample = self.samples[index]- path = sample[0]+ sample = self.data.iloc[index]+ path = sample["path"]if self.root:path = os.path.join(self.root, path)- text = sample[1]+ text = sample["text"]if self.is_video:vframes, aframes, info = torchvision.io.read_video(filename=path, pts_unit="sec", output_format="TCHW")@@ -86,4 +83,4 @@ class DatasetFromCSV(torch.utils.data.Dataset):raise RuntimeError("Too many bad data.")def __len__(self):- return len(self.samples)+ return len(self.data)diff --git a/tools/datasets/csvutil.py b/tools/datasets/csvutil.pyindex 44cb61b..813fc1b 100644--- a/tools/datasets/csvutil.py+++ b/tools/datasets/csvutil.py@@ -166,7 +166,7 @@ def main(args):data.append(pd.read_csv(input_path))input_name += os.path.basename(input_path).split(".")[0]if i != len(args.input) - 1:- input_name += "|"+ input_name += "+"print(f"Loaded {len(data[-1])} samples from {input_path}.")data = pd.concat(data, ignore_index=True, sort=False)print(f"Total number of samples: {len(data)}.")