The model must use a single boolean attribute only

[hma02]

Hi, thanks for providing the source code of the paper.

I just tried training on the default settings except I ran preprocess.py to generate 128_128 images.

I see the default settings of train.py have Smiling and Male set as attributes, which I assume are the attributes to separate entanglement between during training. (OR maybe I understand it incorrectly, the code always automatically disentangles between all attributes available in list_attr_celeba.txt no matter what is set for --attr)

parser.add_argument("--attr", type=attr_flag, default="Smiling,Male",
                    help="Attributes to classify")

fadernetworks $ python train.py
INFO - 03/19/18 13:50:56 - 0:00:00 - ============ Initialized logger ============
INFO - 03/19/18 13:50:56 - 0:00:00 - ae_optimizer: adam,lr=0.0002
                                     ae_reload:
                                     attr: [('Male', 2), ('Smiling', 2)]
                                     batch_size: 32
                                     clf_dis_reload:
                                     clip_grad_norm: 5
                                     debug: False
                                     dec_dropout: 0.0
                                     deconv_method: convtranspose
                                     dis_optimizer: adam,lr=0.0002
                                     dump_path: ./models/default/aq8phwtw2o
                                     epoch_size: 50000
                                     eval_clf: models/classifier128.pth
                                     h_flip: True
                                     hid_dim: 512
                                     img_fm: 3 
                                     img_sz: 128
                                     init_fm: 32
                                     instance_norm: False
                                     lambda_ae: 1
                                     lambda_clf_dis: 0
                                     lambda_lat_dis: 0.0001
                                     lambda_ptc_dis: 0
                                     lambda_schedule: 500000
                                     lat_dis_dropout: 0.3
                                     lat_dis_reload:
                                     max_fm: 512
                                     n_attr: 4 
                                     n_clf_dis: 0
                                     n_epochs: 1000
                                     n_lat_dis: 1
                                     n_layers: 6
                                     n_ptc_dis: 0
                                     n_skip: 0 
                                     name: default
                                     ptc_dis_reload:
                                     smooth_label: 0.2
                                     v_flip: False

Then I used my saved snapshots after training, male_smiling.pth, to interpolate images with the default settings of interpolate.py and got the following error. Any ideas how to run this interpolate.py, or what should I change?

models $ cp default/aq8phwtw2o/best_rec_ae.pth ./male_smiling.pth
models $ cd ..
fadernetworks $ python interpolate.py 
Traceback (most recent call last):
  File "interpolate.py", line 60, in <module>
    raise Exception("The model must use a single boolean attribute only.")
Exception: The model must use a single boolean attribute only.

[domarps]

A bit late but I reproduced your issue and a couple of observations :

Regarding the attribute disentanglement, the parsing happens in https://github.com/facebookresearch/FaderNetworks/blob/master/src/utils.py#L60. The params.attr is prepared in the check_attr function here.

The interpolate code works well for single-attribute only, since you have specified two attributes -- you will have to scale up alphas accordingly (one pair for each attribute).

I have written a simple notebook which summarizes the interpolation steps for gender and smile here