Preparing Dataset *a modification in readme file

[HOMGH]

Just use img_align_celebA_png.7z I think.
I suggest you read this part of code carefully (containing the aligning code)

3D-Face-GCNs/utils.py

Line 217 in 2c4459c

def load_images(filenames, img_size, alpha, landmark, cropper=None, gray=False):

Originally posted by @deepmo24 in #10 (comment)

[HOMGH]

Prepare dataset:
1)For training this model, you need to use "img_celeba" folder in celebA dataset.
2)This data should be aligned using 5 landmarks from "list_landmarks_align_celeba.tx using preproces function in https://github.com/microsoft/Deep3DFaceReconstruction/blob/4b27ddf4227c687b953caa7aa3f3e7acf87cc786/preprocess_img.py#L52
list_landmarks_align_celeba.txt
3) Segment the aligned dataset
4)Make RGBA using the following code:
def save_RGBA_face(im, parsing_anno, img_size, save_path='vis_results/parsing_map_on_im.jpg'):

vis_parsing_anno = parsing_anno.copy().astype(np.uint8)
face_mask = np.zeros((vis_parsing_anno.shape[0], vis_parsing_anno.shape[1]))

num_of_class = np.max(vis_parsing_anno)
for pi in range(1, num_of_class + 1):
    index = np.where(vis_parsing_anno == pi)
    if pi in [1,2,3,4,5,10,12,13]:
        face_mask[index[0], index[1]] = 255.0

im = np.array(im)
img = im.copy().astype(np.uint8)
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)

seg_img = face_mask.astype(np.uint8)

img = cv2.resize(img, (img_size, img_size))
seg_img = cv2.resize(seg_img, (img_size, img_size))
seg_img = seg_img[:,:,None]

BGRA_img = np.concatenate((img, seg_img), axis=2)

cv2.imwrite(save_path, BGRA_img)

5. create .bin files using create_bin.py

Done!

Now the Dataset is ready to be used for training the model...

P.S, Thank @deepmo24 for her contribution.

[ZechCal]

Great, but how to segment the dataset? Is there any good public algorithm available?

[HOMGH]

Great, but how to segment the dataset? Is there any good public algorithm available?

#2 (comment)
Here is the segmentation algorithm suggested by the author.

[ZechCal]

Thx a lot!

[ZechCal]

Hello, what is the size of your bin dataset? I may faced with some troubles....(152GB)

[HOMGH]

Hello, what is the size of your bin dataset? I may faced with some troubles....(152GB)

Hi, Mine is 156 GB.

[ZechCal]

Hello, what is the size of your bin dataset? I may faced with some troubles....(152GB)

Hi, Mine is 156 GB.

Thx a lot!

[ghost]

@HOMGH
can you share the dataset?

[cyjouc]

Prepare dataset:
1)For training this model, you need to use "img_celeba" folder in celebA dataset.
2)This data should be aligned using 5 landmarks from "list_landmarks_align_celeba.tx using preproces function in https://github.com/microsoft/Deep3DFaceReconstruction/blob/4b27ddf4227c687b953caa7aa3f3e7acf87cc786/preprocess_img.py#L52
list_landmarks_align_celeba.txt
3) Segment the aligned dataset
4)Make RGBA using the following code:
def save_RGBA_face(im, parsing_anno, img_size, save_path='vis_results/parsing_map_on_im.jpg'):

vis_parsing_anno = parsing_anno.copy().astype(np.uint8)
face_mask = np.zeros((vis_parsing_anno.shape[0], vis_parsing_anno.shape[1]))

num_of_class = np.max(vis_parsing_anno)
for pi in range(1, num_of_class + 1):
    index = np.where(vis_parsing_anno == pi)
    if pi in [1,2,3,4,5,10,12,13]:
        face_mask[index[0], index[1]] = 255.0

im = np.array(im)
img = im.copy().astype(np.uint8)
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)

seg_img = face_mask.astype(np.uint8)

img = cv2.resize(img, (img_size, img_size))
seg_img = cv2.resize(seg_img, (img_size, img_size))
seg_img = seg_img[:,:,None]

BGRA_img = np.concatenate((img, seg_img), axis=2)

cv2.imwrite(save_path, BGRA_img)

1. create .bin files using create_bin.py

Done!

Now the Dataset is ready to be used for training the model...

P.S, Thank @deepmo24 for her contribution.

hi,Can you success to the train?Could you ask me some questions？(hicaicaihi for my wetchat)

[cyjouc]

how do you obtain the the fllowing three data?
train_txt = '{}_train.txt'.format(prefix)
val_txt = '{}_val.txt'.format(prefix)
test_txt = '{}

[cyjouc]

Segment the aligned dataset

Segment the aligned dataset?link only speak decode process,how split face and no_face data?

[cyjouc]

@HOMGH
can you share the dataset?

Hi,do you prepare the dataset for training?would you like share with me?

[cyjouc]

Hello, what is the size of your bin dataset? I may faced with some troubles....(152GB)

Hi, Mine is 156 GB.

Thx a lot!

can you share the dataset?

[Msj5]

Hi, do you train with main.py? have you suffer out of memory?Could you share your train.py with me?plz reply to [email protected]
thx. @HOMGH

[biga-bomb]

Prepare dataset:
1)For training this model, you need to use "img_celeba" folder in celebA dataset.
2)This data should be aligned using 5 landmarks from "list_landmarks_align_celeba.tx using preproces function in https://github.com/microsoft/Deep3DFaceReconstruction/blob/4b27ddf4227c687b953caa7aa3f3e7acf87cc786/preprocess_img.py#L52
list_landmarks_align_celeba.txt
3) Segment the aligned dataset
4)Make RGBA using the following code:
def save_RGBA_face(im, parsing_anno, img_size, save_path='vis_results/parsing_map_on_im.jpg'):

vis_parsing_anno = parsing_anno.copy().astype(np.uint8)
face_mask = np.zeros((vis_parsing_anno.shape[0], vis_parsing_anno.shape[1]))

num_of_class = np.max(vis_parsing_anno)
for pi in range(1, num_of_class + 1):
    index = np.where(vis_parsing_anno == pi)
    if pi in [1,2,3,4,5,10,12,13]:
        face_mask[index[0], index[1]] = 255.0

im = np.array(im)
img = im.copy().astype(np.uint8)
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)

seg_img = face_mask.astype(np.uint8)

img = cv2.resize(img, (img_size, img_size))
seg_img = cv2.resize(seg_img, (img_size, img_size))
seg_img = seg_img[:,:,None]

BGRA_img = np.concatenate((img, seg_img), axis=2)

cv2.imwrite(save_path, BGRA_img)

1. create .bin files using create_bin.py

Done!

Now the Dataset is ready to be used for training the model...

P.S, Thank @deepmo24 for her contribution.

Hi, can you train the model correctly? If you can do that, could you please share your pre-model? Thank you very much! [email protected]

[Rafael8830]

you obtain the the fllowing three data?

how do you obtain the the fllowing three data? train_txt = '{}_train.txt'.format(prefix) val_txt = '{}_val.txt'.format(prefix) test_txt = '{}

I have the same problem, did you solve it?