polarathene
commented
2 years ago I had a lengthy response but lost it to a system crash. Still I'll post changes required here, looking at PR activity and maintainer interaction lacking on open PRs, I'm hesitant to create a new PR.
This KeyError occured due to the media stream for b.mkv not having a duration property, but the format object sibling of streams object array does have the expected duration value. You could use that instead, otherwise perhaps it's better as a fallback?:
https://github.com/JNoDuq/videobench/blob/ea447031bfcc0315a4ae7d793896f051cc8bc2c3/videobench.py#L83
try:
duration = ffprobe_json['streams'][0]['duration']
except KeyError:
duration = ffprobe_json['format']['duration']That won't work though as format object is not available unless we also query that here:
It needs -show_format:
cmd = "{0} ffprobe -loglevel {3} -print_format json -show_format -show_streams -select_streams v -i {1}Then the next failure is here:
Terminal output logs deprecated API calls and advises using the newer ones, which are also documented here for ffmpeg vmaf. psnr=1 and model_path need to be changed:
cmd = (''' {DOCKER_CMD} ffmpeg -y -loglevel {LOGLEVEL} -stats -i {CONTAINER_TMP_PATH}{REF_FILENAME} -i {CONTAINER_TMP_PATH}{INPUT_FILENAME} ''' \
'''-lavfi "[0]{REF_DEINT}[refdeint];[refdeint]{REF_SCALE_FILTER}[ref];[1]setpts=PTS{SYNC_TIME}/TB[b];[b]{SCALE_FILTER}[c];[c][ref]libvmaf='n_threads={N_THREADS}:log_fmt=json:feature='name=psnr':model='path={VMAF_MODEL}':n_subsample={N_SUBSAMPLE}:log_path={CONTAINER_TMP_PATH}quality_{INPUT_NAME}.json'" ''' \
''' -t {DURATION} -f null - ''').format(Now the quality_<input name>.json is properly generated..
Next failure is with psnr metrics. Presumably in the past this was a single value, inspecting the quality json file has related metrics listed like this:
"metrics": {
"psnr_y": 57.247755,
"psnr_cb": 57.478840,
"psnr_cr": 57.988193,These are values related to the individual channel components of the YCbCr colorspace, I don't know too much about this, and assume the previous single value according to the wikipedia article would have been the Y channel value. (EDIT: It is the correct one)
This should avoid the following error when an empty array is otherwise the current result:
input_obj.psnr_avg = round(sum(psnr_values)/len(psnr_values), 2)
ZeroDivisionError: division by zeroRelevant line:
Update to:
psnr_values.append(data_json['frames'][i]["metrics"]["psnr_y"]) ################################# get psnr values (libvmaf)Now the processing completes successfully and actual graph results on the UI are drawn :)
150481878
The progress reached 40% but failed after processing for a while with the following: