Write samples to parquet files incrementally with a buffer

Maintain a counter with a capping to avoid memory issues

Detects cuts in input videos and returns contiguous segments in a video as metadata. expects: - cut_detection_mode to be either "longest

Adjusts the frame rate of the videos to the specified frame rate. Args: frame_rate (int): Target frame rate of the videos.

Cuts videos up into segments according to the 'clips' metadata expects: - clips to be sorted in increasing order and non-overlapping

Adjusts the frame rate of the videos to the specified frame rate. Args: frame_rate (int): Target frame rate of the videos.

Adjusts the resolution of the videos to the specified height and width. Args: video_size (int): Target resolution of the videos.

Log performance metrics

Video data reader provide data for a video

Downloader class for mp4 links

Downloader class for yt-dlp links

Filters the dataset based on aesthethics

randomly initialize N queries input: x: [B, n_token, C (context_dim)] as K and V output: queries: [B, n_query, d_query (d_model)]

Adds cuts to video sample

The downloader class gets calls with shards, download them then call the writer to write them down

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

Vision Transformer with support for patch or hybrid CNN input stage

Vision Transformer with support for patch or hybrid CNN input stage

The reader class reads an url list and returns shards It provides an iter method It provides attributes: - column_list: the list of c

Filters the dataset based on the key

Decoder which allows you to pass through some keys

Filters the dataset based on the language

Logger process that reads stats files regularly, aggregates and send to wandb / print to terminal

The reader class reads a shard list and returns shards It provides an iter method It provides attributes: - shard_list: a list of sh

https://arxiv.org/abs/2212.00794

https://arxiv.org/abs/2212.00794

Image to Patch Embedding

Image to Patch Embedding

Log status table to W&B, up to `max_status` most frequent items

The loader class reads the shards, then the selected data is chosen and writen by the writer

Filters the dataset based on the probability a sample is unsafe

Resizes frames to specified height and width

Does not write

FilesSampleWriter is a caption+video writer to files

Subclass torch's LayerNorm (with cast back to input dtype).