[SHARING] DFL 2.0 ReadyToTrain model

NoWayMan · Jan 11, 2022

Hello there, its my first try to deepfacelive. I want to be sure that i have understand.

I want to swap my face (A) in streaming with an italian celeb (B).
A (me) = dst
B (celeb) = src

For do a RTM to export in DeepfaceLive i have to do theses step?

1) do as faceset like i do with DFL "classic"
2) download a RTM pretrained model with res i need (256?)
3) continue to pretain model and then, train following this guide

<a href=""><img src="" alt="Screenshot-1143" border="0"></a>

4) export as DFM model in DFLive and use it

Its correct?

DeliveryBoy said:
Pretraining Model & Ready to Merge Model (256 resolution WF Liau-ud)

Updated: 2 files in 1 (Pretrained 2,000,000 iterations + ReadToMerge 2,000,000 iterations.

Hellok many thanks for your share.

Im confused me too for wich of two I have to use.
I need to do a RTM for DeepFaceLIVE using my face as DST and an italian celeb as SRC. What i have to do?

DeliveryBoy · Jan 16, 2022

ThisIsLive said:
in the RTM workflow, at the end, "Extra training and reuse of LIAE RTM models:", you mention "3. To create a new RTM model for different celebrity delete inter_ab file, replace SRC with new one, add random DST set and proceed with the same workflow as when creating new RTM model."

Just for information, I'm using 's 256 RTM LIAE-UD model, and have never deleted the inter_ab file when changing the src/celeb.

When starting a new workspace / celeb, I just copy the original model to the workspace and start training it immediately, without deleting files in the model directory.

It seems to work for me.. is there an upside in deleting the inter_ab file ?

Your way is correct, if you delete inter_b on a higher resolution model all you will see is a black screen, it will not learn anything especially on a 320 and higher model.

I was training on a higher resolution model 320 after deleting the inter_b it didn't learn anything, all i got was just a black screen in the preview window (the src will show on preview window but the dst stays black forever and not even a sign of face shape ever after 10K iteration).

I do the exact same thing you do wihtout deleting inter_b and still get src like result after training on new dst.

When you delete inter_b or inter_ab, then it is no longer "ready to merge" because you are training a new dst or src all over again.

When you don't delete anything and you just train your new dst and src, it only takes short time for it to be ready to merge.

NoWayMan · Jan 18, 2022

Fast question.

RTM can be used only with LIAE UD, right? I have created a RTM (export SAEHD as dfm.BAT) from SAEHD model. I can use it? How?

dsyrock · Jan 20, 2022

DeliveryBoy said:
If someone donate to me 224 or 256 RTM Liae-udt pretrained model, i will make those RTM for you.

I have already made a 256 RTM Liae-ud if you want to use that it's on page 8. I will probably make a Liae-udt version in feb 2022 when someone is willing to share a pretrained model for the 256 Liae-udt pretrained model.

If you want to use my 256 RTM Liae-UD here is the link

I am pretraining this model, do you need it?

Code:

================= Model Summary =================
==                                             ==
==            Model name: pre_SAEHD            ==
==                                             ==
==     Current iteration: 821417               ==
==                                             ==
==--------------- Model Options ---------------==
==                                             ==
==            resolution: 256                  ==
==             face_type: wf                   ==
==     models_opt_on_gpu: True                 ==
==                 archi: liae-udt             ==
==               ae_dims: 256                  ==
==                e_dims: 64                   ==
==                d_dims: 64                   ==
==           d_mask_dims: 22                   ==
==       masked_training: True                 ==
==       eyes_mouth_prio: False                ==
==           uniform_yaw: False                ==
==         blur_out_mask: False                ==
==             adabelief: True                 ==
==            lr_dropout: False                ==
==           random_warp: False                ==
==      random_hsv_power: 0                    ==
==       true_face_power: 0.0                  ==
==      face_style_power: 0.0                  ==
==        bg_style_power: 0.0                  ==
==               ct_mode: rct                  ==
==              clipgrad: False                ==
==              pretrain: False                ==
==       autobackup_hour: 0                    ==
== write_preview_history: False                ==
==           target_iter: 0                    ==
==       random_src_flip: False                ==
==       random_dst_flip: False                ==
==            batch_size: 8                    ==
==             gan_power: 0.0                  ==
==        gan_patch_size: 36                   ==
==              gan_dims: 16                   ==
==                                             ==
==---------------- Running On -----------------==
==                                             ==
==          Device index: 0                    ==
==                  Name: Tesla P100-PCIE-16GB ==
==                  VRAM: 14.61GB              ==
==                                             ==
=================================================

DeliveryBoy · Jan 21, 2022

dsyrock said:

DeliveryBoy said:
If someone donate to me 224 or 256 RTM Liae-udt pretrained model, i will make those RTM for you.

I have already made a 256 RTM Liae-ud if you want to use that it's on page 8. I will probably make a Liae-udt version in feb 2022 when someone is willing to share a pretrained model for the 256 Liae-udt pretrained model.

If you want to use my 256 RTM Liae-UD here is the link

I am pretraining this model, do you need it?

Code:

================= Model Summary =================
==                                             ==
==            Model name: pre_SAEHD            ==
==                                             ==
==     Current iteration: 821417               ==
==                                             ==
==--------------- Model Options ---------------==
==                                             ==
==            resolution: 256                  ==
==             face_type: wf                   ==
==     models_opt_on_gpu: True                 ==
==                 archi: liae-udt             ==
==               ae_dims: 256                  ==
==                e_dims: 64                   ==
==                d_dims: 64                   ==
==           d_mask_dims: 22                   ==
==       masked_training: True                 ==
==       eyes_mouth_prio: False                ==
==           uniform_yaw: False                ==
==         blur_out_mask: False                ==
==             adabelief: True                 ==
==            lr_dropout: False                ==
==           random_warp: False                ==
==      random_hsv_power: 0                    ==
==       true_face_power: 0.0                  ==
==      face_style_power: 0.0                  ==
==        bg_style_power: 0.0                  ==
==               ct_mode: rct                  ==
==              clipgrad: False                ==
==              pretrain: False                ==
==       autobackup_hour: 0                    ==
== write_preview_history: False                ==
==           target_iter: 0                    ==
==       random_src_flip: False                ==
==       random_dst_flip: False                ==
==            batch_size: 8                    ==
==             gan_power: 0.0                  ==
==        gan_patch_size: 36                   ==
==              gan_dims: 16                   ==
==                                             ==
==---------------- Running On -----------------==
==                                             ==
==          Device index: 0                    ==
==                  Name: Tesla P100-PCIE-16GB ==
==                  VRAM: 14.61GB              ==
==                                             ==
=================================================

I have already pretrained 256 Liae-UDT with 1KK iterations. I have already shared it on another thread

Code:

================== Model Summary ===================
==                                                ==
==            Model name: 256RTMLiaeUDT_SAEHD     ==
==                                                ==
==     Current iteration: 1044253                 ==
==                                                ==
==---------------- Model Options -----------------==
==                                                ==
==            resolution: 256                     ==
==             face_type: wf                      ==
==     models_opt_on_gpu: True                    ==
==                 archi: liae-udt                ==
==               ae_dims: 256                     ==
==                e_dims: 64                      ==
==                d_dims: 64                      ==
==           d_mask_dims: 22                      ==
==       masked_training: True                    ==
==       eyes_mouth_prio: True                    ==
==           uniform_yaw: True                    ==
==         blur_out_mask: True                    ==
==             adabelief: True                    ==
==            lr_dropout: n                       ==
==           random_warp: False                   ==
==      random_hsv_power: 0.0                     ==
==       true_face_power: 0.0                     ==
==      face_style_power: 0.0                     ==
==        bg_style_power: 0.0                     ==
==               ct_mode: lct                     ==
==              clipgrad: False                   ==
==              pretrain: True                    ==
==       autobackup_hour: 0                       ==
== write_preview_history: False                   ==
==           target_iter: 0                       ==
==       random_src_flip: False                   ==
==       random_dst_flip: False                   ==
==            batch_size: 8                       ==
==             gan_power: 0.0                     ==
==        gan_patch_size: 32                      ==
==              gan_dims: 16                      ==
==                                                ==
==------------------ Running On ------------------==
==                                                ==
==          Device index: 0                       ==
==                  Name: NVIDIA GeForce RTX 3090 ==
==                  VRAM: 21.17GB                 ==
==                                                ==
====================================================

AdamW · Jan 26, 2022

Hi, I have read the guide on the first page and now I want to make a rtm model of one of my src.

I have a question: which dst should I use to generate the rtm model that covers each angle? Is there a link to download it?

shaiir · Feb 3, 2022

DeliveryBoy said:

dsyrock said:

DeliveryBoy said:
If someone donate to me 224 or 256 RTM Liae-udt pretrained model, i will make those RTM for you.

I have already made a 256 RTM Liae-ud if you want to use that it's on page 8. I will probably make a Liae-udt version in feb 2022 when someone is willing to share a pretrained model for the 256 Liae-udt pretrained model.

If you want to use my 256 RTM Liae-UD here is the link

I am pretraining this model, do you need it?

Code:

================= Model Summary =================
==                                             ==
==            Model name: pre_SAEHD            ==
==                                             ==
==     Current iteration: 821417               ==
==                                             ==
==--------------- Model Options ---------------==
==                                             ==
==            resolution: 256                  ==
==             face_type: wf                   ==
==     models_opt_on_gpu: True                 ==
==                 archi: liae-udt             ==
==               ae_dims: 256                  ==
==                e_dims: 64                   ==
==                d_dims: 64                   ==
==           d_mask_dims: 22                   ==
==       masked_training: True                 ==
==       eyes_mouth_prio: False                ==
==           uniform_yaw: False                ==
==         blur_out_mask: False                ==
==             adabelief: True                 ==
==            lr_dropout: False                ==
==           random_warp: False                ==
==      random_hsv_power: 0                    ==
==       true_face_power: 0.0                  ==
==      face_style_power: 0.0                  ==
==        bg_style_power: 0.0                  ==
==               ct_mode: rct                  ==
==              clipgrad: False                ==
==              pretrain: False                ==
==       autobackup_hour: 0                    ==
== write_preview_history: False                ==
==           target_iter: 0                    ==
==       random_src_flip: False                ==
==       random_dst_flip: False                ==
==            batch_size: 8                    ==
==             gan_power: 0.0                  ==
==        gan_patch_size: 36                   ==
==              gan_dims: 16                   ==
==                                             ==
==---------------- Running On -----------------==
==                                             ==
==          Device index: 0                    ==
==                  Name: Tesla P100-PCIE-16GB ==
==                  VRAM: 14.61GB              ==
==                                             ==
=================================================

I have already pretrained 256 Liae-UDT with 1KK iterations. I have already shared it on another thread

Code:

================== Model Summary ===================
==                                                ==
==            Model name: 256RTMLiaeUDT_SAEHD     ==
==                                                ==
==     Current iteration: 1044253                 ==
==                                                ==
==---------------- Model Options -----------------==
==                                                ==
==            resolution: 256                     ==
==             face_type: wf                      ==
==     models_opt_on_gpu: True                    ==
==                 archi: liae-udt                ==
==               ae_dims: 256                     ==
==                e_dims: 64                      ==
==                d_dims: 64                      ==
==           d_mask_dims: 22                      ==
==       masked_training: True                    ==
==       eyes_mouth_prio: True                    ==
==           uniform_yaw: True                    ==
==         blur_out_mask: True                    ==
==             adabelief: True                    ==
==            lr_dropout: n                       ==
==           random_warp: False                   ==
==      random_hsv_power: 0.0                     ==
==       true_face_power: 0.0                     ==
==      face_style_power: 0.0                     ==
==        bg_style_power: 0.0                     ==
==               ct_mode: lct                     ==
==              clipgrad: False                   ==
==              pretrain: True                    ==
==       autobackup_hour: 0                       ==
== write_preview_history: False                   ==
==           target_iter: 0                       ==
==       random_src_flip: False                   ==
==       random_dst_flip: False                   ==
==            batch_size: 8                       ==
==             gan_power: 0.0                     ==
==        gan_patch_size: 32                      ==
==              gan_dims: 16                      ==
==                                                ==
==------------------ Running On ------------------==
==                                                ==
==          Device index: 0                       ==
==                  Name: NVIDIA GeForce RTX 3090 ==
==                  VRAM: 21.17GB                 ==
==                                                ==
====================================================

where can i download the 256 Liae-UDT? in which thread did u shared it?

DeliveryBoy · Feb 7, 2022

In the trained model section

//deep.whitecatchel.ru/literotica/forums/thre...b-2-0-saehd-models-sharing?pid=41208#pid41208

iperov · Mar 3, 2022

main post update !

bomxart · Mar 4, 2022

Is there any problem if I change the Batch size?

lythat · Mar 5, 2022

place RTM WF Faceset to workspace/data_dst/aligned

2. place RTM WF Faceset to workspace/data_src/aligned

If I do, what will happen?

kkdlux · Mar 5, 2022

DeliveryBoy said:
In the trained model section

//deep.whitecatchel.ru/literotica/forums/thre...b-2-0-saehd-models-sharing?pid=41208#pid41208

Used your model to train a Wemma Atson DFM for Deepfacelive, if someone wants to try it it works fine in my computer with good quality overall.

Maybe "iperov" found it good enough to be in the deepfacelive model download if he wants.

DeliveryBoy · Mar 10, 2022

ThisIsLive said:
DeliveryBoy said:

You're welcome it should be finished in a couple of days when i reach 3,800,000 iterations.

For those that can't wait for the 320 RTM to reach 3,800,000 iterations here is the download link for the 320 RTM model that is very close to 3,700,000 iterations.

Download Link:

Click to expand...

Just for information, did you ever finish the 320 RTM model, or is it best to use your latest one (3615968 iterations) ?

Thanks again for all your work!

i put it on hold as i had to move house which has cost me financially - however i am training the 256 RTM UDT because most people cant afford high end cards with all the scam that's going on from NVIDIA factory and Taiwan.

My training is almost done but for some reasons i cant see any difference after using GAN option for 200K iterations.

Maybe someone could point out where i am going wrong with GAN?

Code:

================== Model Summary ===================
==                                                ==
==            Model name: 256RTMLiaeUDT_SAEHD     ==
==                                                ==
==     Current iteration: 3776566                 ==
==                                                ==
==---------------- Model Options -----------------==
==                                                ==
==            resolution: 256                     ==
==             face_type: wf                      ==
==     models_opt_on_gpu: True                    ==
==                 archi: liae-udt                ==
==               ae_dims: 256                     ==
==                e_dims: 64                      ==
==                d_dims: 64                      ==
==           d_mask_dims: 22                      ==
==       masked_training: True                    ==
==       eyes_mouth_prio: True                    ==
==           uniform_yaw: False                   ==
==         blur_out_mask: True                    ==
==             adabelief: True                    ==
==            lr_dropout: y                       ==
==           random_warp: False                   ==
==      random_hsv_power: 0.0                     ==
==       true_face_power: 0.0                     ==
==      face_style_power: 0.0                     ==
==        bg_style_power: 0.0                     ==
==               ct_mode: lct                     ==
==              clipgrad: False                   ==
==              pretrain: False                   ==
==       autobackup_hour: 0                       ==
== write_preview_history: True                    ==
==           target_iter: 0                       ==
==       random_src_flip: False                   ==
==       random_dst_flip: True                    ==
==            batch_size: 8                       ==
==             gan_power: 0.1                     ==
==        gan_patch_size: 32                      ==
==              gan_dims: 16                      ==
==                                                ==
==------------------ Running On ------------------==
==                                                ==
==          Device index: 0                       ==
==                  Name: NVIDIA GeForce RTX 3090 ==
==                  VRAM: 21.17GB                 ==
==                                                ==
====================================================

As you can see in the 2 pics after adding more 70K GAN iterations (all together GAN trained = 200K) I still cant spot any difference between a prior to gan and a gan training

DeliveryBoy · Mar 10, 2022

iperov said:
This is heavily pretrained model to train your own Ready-To-Merge model for 224 res using RTM WF Faceset.

How to use this model.

1. place RTM WF Faceset to workspace/data_dst/aligned

2. place your celeb to workspace/data_src/aligned

3. place this model folder to workspace/model

4. do not change settings, train +500.000 iters

5. random_warp OFF, train +500.000, periodically disable masked training for 5.000 iters and enable again

6. GAN 0.1 power, patch size 28, dims 32 train +500.000

Thank you so very much Good Sir.

iperov · Mar 13, 2022

6. GAN 0.1 power, patch size 28, gan_dims:32. Train until the src loss value has not increased in the last 12 hours.

7. finalize model by disabling masked training for 100-200 (not thousand) iterations

BATHORY93 · Mar 15, 2022

kkdlux said:
DeliveryBoy said:

In the trained model section

//deep.whitecatchel.ru/literotica/forums/thre...b-2-0-saehd-models-sharing?pid=41208#pid41208

Click to expand...

Used your model to train a Wemma Atson DFM for Deepfacelive, if someone wants to try it it works fine in my computer with good quality overall.

Maybe "iperov" found it good enough to be in the deepfacelive model download if he wants.

Anyway you can reupload that? It got removed.

Grrkin · Mar 16, 2022

BATHORY93 said:
kkdlux said:

DeliveryBoy said:

In the trained model section

//deep.whitecatchel.ru/literotica/forums/thre...b-2-0-saehd-models-sharing?pid=41208#pid41208

Click to expand...

Used your model to train a Wemma Atson DFM for Deepfacelive, if someone wants to try it it works fine in my computer with good quality overall.

Maybe "iperov" found it good enough to be in the deepfacelive model download if he wants.

Click to expand...

Anyway you can reupload that? It got removed.

Works for me.

BATHORY93 · Mar 17, 2022

Grrkin said:
BATHORY93 said:

kkdlux said:

DeliveryBoy said:

In the trained model section

//deep.whitecatchel.ru/literotica/forums/thre...b-2-0-saehd-models-sharing?pid=41208#pid41208

Click to expand...

Used your model to train a Wemma Atson DFM for Deepfacelive, if someone wants to try it it works fine in my computer with good quality overall.

Maybe "iperov" found it good enough to be in the deepfacelive model download if he wants.

Click to expand...

Anyway you can reupload that? It got removed.

Click to expand...

Works for me.

Strange. When using Firefox it says it has been removed, but in Jdownloader it's there. Sorry about that and thanks.

viddy · Mar 23, 2022

iperov said:
This is heavily pretrained model to train your own Ready-To-Merge model for 224 res using RTM WF Faceset.

How to use this model.

1. place RTM WF Faceset to workspace/data_dst/aligned

2. place your celeb to workspace/data_src/aligned

3. place this model folder to workspace/model

4. do not change settings, train +500.000 iters

5. random_warp OFF, train +500.000, periodically disable masked training for 5.000 iters and enable again

6. GAN 0.1 power, patch size 28, gan_dims:32. Train until the src loss value has not increased in the last 12 hours

7. finalize model by disabling masked training for 100-200 (not thousand) iterations

Apologies. Stupid question from an idiot who doesn't have a better GPU to do this on his own computer. I'm guessing this isn't usable on google colab because of filesize? The faceset provided is over 9 gigs and pushes the zip file to something that can't really be saved. Again- certified idiot alert so apologies again.

computera14 · Mar 29, 2022

iperov said:
6. GAN 0.1 power, patch size 28, gan_dims:32. Train until the src loss value has not increased in the last 12 hours.

7. finalize model by disabling masked training for 100-200 (not thousand) iterations

What is your mean (src loss value has not increased)?

[SHARING] DFL 2.0 ReadyToTrain model

New member

Member

New member

New member

Member

New member

New member

Member

Member

New member

New member

New member

Member

Member

Member

New member

New member

New member

New member

New member