Data streaming using chunks #787
Conversation
|
@ygnn123 |
|
Please, put minimized version of |
|
Files |
|
You need to add cvat-data to a Dockerfile. |
Dockerfile
Outdated
| @@ -142,6 +142,7 @@ COPY ssh ${HOME}/.ssh | |||
| COPY utils ${HOME}/utils | |||
| COPY cvat/ ${HOME}/cvat | |||
| COPY cvat-core/ ${HOME}/cvat-core | |||
| COPY cvat-core/ ${HOME}/cvat-data | |||
| throw new ArgumentError( | ||
| `Invalid mode is specified ${mode}`, | ||
| ); | ||
| } | ||
|
|
||
| frameDataCache[taskID][frame] = new FrameData(size.width, size.height, taskID, frame); |
There was a problem hiding this comment.
Maybe is it better to cache these instances like before? It isn't exactly frames. They are only wrappers which don't have binary information about frames.
There was a problem hiding this comment.
No!!! I realy need cache in cvat-data due to worker using.
|
@azhavoro , please merge the latest develop branch. Need to have a working solution this week. Next week we need to start testing otherwise the feature will be delivered only next year and it is below our expectations. |
| @@ -1,4 +1,5 @@ | |||
| exclude_paths: | |||
| - '**/3rdparty/**' | |||
| - '**/engine/js/cvat-core.min.js' | |||
| - '**/engine/js/unzip_imgs.js' | |||
There was a problem hiding this comment.
What is a reason to exclude the file from scanning?
There was a problem hiding this comment.
this is bundled and minified version of cvat-data/src/js/unzip_imgs.js which included to scanning.
| from cvat.apps.engine.mime_types import mimetypes | ||
|
|
||
|
|
||
| class FrameProvider(): |
There was a problem hiding this comment.
Just a comment for future discussion. I generally like the class but prefer to change its interface:
- QUALITY_ORIGINAL=100, QUALITY_COMPRESSED=0
- get_preview(self)
- get_chunk(self, n, quality=QUALITY_ORIGINAL)
- get_frame(self, n, quality=QUALITY_ORIGINAL)
- get_frames(self, quality=QUALITY_ORIGINAL)
The main idea behind the change is to keep the API stable if we generate compressed frames on the fly with a quality. Thus we will be able to call get_frame(self, n, 75).
Known issues in this PR:
There are no get frame REST API methods.
Crash browser tab on 4k video after several navigation events (Possible solution: change the size of the decoded frames buffer dynamically).
Some number of fast clicks (>10-20) on the navigation bar leads to the corresponding number of chunk request. Decoding begins after all requests have finished. Possible solution: implement the same logic as for cvat-data (process only the last request).
Build instruction for avc.wasm and Decoder.js as first step and dockerized build as final solution.
There are no access to
original qualitychunks and frames.Need to check the creation task time for both case (images and video) and compare with current version.
Estimate network traffic for new and old version on the same data.
Add small-size preview for the navigation slider.
Video playback is not smooth - player freezes and skips some frames after that. Partially fixed. Need to implement a buffering on the player side to make playback more smoothly.
Player issue: now player is requiring frames every 1/25s. It may leads to the following case: the frame which requested later will be received by the player earlier. (If decoder works slow on the large frame resolution).
Decoder potentially may has performance problems and decoding speed may be not enough to play the video smoothly, especially on the high resolution videos. Solution: use downscaled images on the backend and upscale on the frontend. Implemented https://github.com/ygnn123/cvat/pull/7/files
Don't create decoder worker threads for each chunk, it leads to extra server requests, etc. (Checked: h264 decoder works fine).
Access to the original data for old tasks issue (data migration)
Cleanup code, fix codacy issues.
auto_segmentation, reid and datumaro are broken