detector object out of memory

[mltrinhammer]

Hello,

I am using detector.detect_video() so process videos with 4500 frames (3 min duration). When the object is done processing each frame, it performs some additional computations which are extremely memory intensive, requiring up to 82 GB. Is it possible to turn the computation of these additional steps off? I just need the AUs. Thank you Martin

[Allexxann]

Or at least do them in a slow, but memory-possible way? I have videos about 20 minutes each, by that logic they'll require several Tb at least.

[cyrellejohn]

Or at least do them in a slow, but memory-possible way? I have videos about 20 minutes each, by that logic they'll require several Tb at least.

how many video you have? and how many hours your vid being processed? its 12 minutes for me but it estimated of 3 hours @Allexxann

[Allexxann]

Or at least do them in a slow, but memory-possible way? I have videos about 20 minutes each, by that logic they'll require several Tb at least.

how many video you have? and how many hours your vid being processed? its 12 minutes for me but it estimated of 3 hours @Allexxann

Hundreds of recordings. The short ones are about 12-14 minutes, the long ones are up to 30 (they are usually bundled with each other).

Though I use a modified algorhitm - it adds extra features, but quadruples the time, so I have to snap a frame every second (or even two) and analyze these (and still it took me a literal week of raw computing time to process a sample from a single experimental series).

[giracle]

Hello! Great work. By the way, does it get stuck during video detection? How did you solve the problem when detecting Action Units (AU)? My 12 - second video keeps getting stuck and can't proceed. Thank you.

[mltrinhammer]

Hello! Great work. By the way, does it get stuck during video detection? How did you solve the problem when detecting Action Units (AU)? My 12 - second video keeps getting stuck and can't proceed. Thank you.

What do you mean getting stuck? Can you perhaps supply more details?

[giracle]

Just like this, the progress bar keeps getting stuck here. I had a 12-second video that could be stuck for 24 hours without moving. I have no idea what the reason is. this is my issues

[pywugate]

@mlt94 did you solve it?

I solve it by changing some codes in detector.py file

1. I comment the line 710 "batch_output.compute_identities(...)" to stop calculations
2. I give identity_model = None for Detecor initiation, by doing so, idntity_detector is also None
3. I also comment line 434 because it creates a 512-element data in torch if idntity_detector is None
4. To prevent error when creating result csv without identity data, I also comment other three parts line 485-487 feat_identities variable creation line 506 feat_identities Fex pd concatenation line 692 identity_columns = FEAT_IDENTITY_COLUMNS[1:] assignment

And the output does not contain 512 identity columns

[mltrinhammer]

Thank you for this neat solution @pywugate! I didnt move further with it, but I am happy to know of your solution should I come back to it