In L173 you have var name "test", but you use features['train'] here. Looks like a mistake. Am I wrong?

Dear authors,
in your paper, you describe the regressor $g(\cdot): \mathbb{R}^{N} \rightarrow [-1, 1]$ architecture as composed of three fully connected layers (multi-layer perceptron). See the images below for reference.

However, in this repository, the regressor is implemented using convolutions with kernel size 1.

Fully connected layers are usually implemented using a linear layer. Do you perform some slicing/viewing of the tensors, which makes the implementation equivalent to a linear layer? Or do you use convolutions here by mistake? Can you please elaborate on this?

Jakub

I try to train the regressor only with pretrained gol-weights encoder(freezed) on the same datasets, but it didn't converge to a desirable performance.
To my limited understanding, gol is a way to train the encoder to map the images to a certain feature embedding space where the features extracted from images of people of certain ages would follow those 3 constraints mentioned in gol paper. Therefore, it was supposed to work that use mwr stategy to replace the knn strategy in gol post-process.
Is it because my training procedure went wrong?Or should I try combine these model to train together?

I tried to run op.py with the next command:

python op.py --dataset utk --regression global --experiment_setting coral --im_path UTKFace

But I have a problem with inhomogeneous part of sequence.
The full error text:

  File "/home/vid/hdd/projects/PycharmProjects/MWR2/code/op.py", line 62, in <module>
    global_regression(args, train_data, test_data, sampling=sampling, sample_rate=sample_rate)
  File "/home/vid/hdd/projects/PycharmProjects/MWR2/code/Prediction.py", line 99, in global_regression
    refer_idx, refer_idx_pair = select_reference_global_regression(train_data, pair_idx, np.array(loss_total), limit=1)
  File "/home/vid/hdd/projects/PycharmProjects/MWR2/code/Utils.py", line 98, in select_reference_global_regression
    y_1_idx_tmp, y_2_idx_tmp = idx[row], np.array(pair_idx)[idx[0]][column]
ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (6582,) + inhomogeneous part.

How to fix it?

First thanks for your great work. I have a question that how can you evaluate how good your model while training ? I mean that how did you the valid set to evaluate the training model?

Thanks for your excellent work. Is the training two-stage for global regressor and local regressor respectively?

Thank you very much for your code's contribution to the field of ordinal regression. Your code is poetic.Could you please make public the Lower bounds and Upper bounds results you obtained in CLAP2015, and make public the Testing... [1/1436]
Iteration 1:
y1_age, y2_age, pred_mwr
Iteration 2:
y1_age, y2_age, pred_mwr

Because I don't know if the Lower bounds and Upper bounds I get with my own data set are correct, I get the following result：Testing... [1/1436]
Iteration 1:
y1_age: [33.2 31.2 38.2 47.2 44.5]
y2_age: [33.2 31.2 38.2 47.2 44.5]
pred_mwr: 38.85999999999999

Iteration 2:
y1_age: [33.2 31.2 38.2 47.2 44.5]
y2_age: [33.2 31.2 38.2 47.2 44.5]
pred_mwr: 38.85999999999999

Would you please provide pipeline for visualization CLAP2015 dataset (maybe without code, just describing specific steps)? Am I need to use additional 7x7 avg pool after 5'th stage VGG16 for this? Does facial alignment have an impact here?

Could you please also provide a demo python file? I would like to run the model on my own test images.

I can't run the code without errors. Can you make it with notebook in Colab?

Hi, I read your paper and great work!

I have some issues running the models on my images. Given that you provide the model pretrained I expected to use them more or less in this manner

global_model = Network.Global_Regressor()
global_model.load_state_dict(torch.load('global.pth'))
local_models = []
for x in range(num_locals):
    local_m = Network.Local_Regressor()
    local_models.append(local_m.load_state_dict(torch.load(f'local{x}.pth')))

my_image = Image.open(open('face.jpg', 'rb'))
# Preprocess image
my_image = torch.transforms.ToTensor(my_image)

global_output = global_model(my_image)
if global_output == something:
    age = local_models[3][my_image]
elif ....  #other cases 

But your code is super complicated and I don't understand it fully. In particular why the datasets are required if the models are pretrained? And why that's not always the case based on the "sampling" variable (I guess)?

So, basically, I just want to use your approach with the model pretrained on morph (or another dataset) to do inference on my data without ever having to download morph (or any other dataset). Is it possible? If not what do you mean with pretrained models?

Thanks

First congratulate for your great work? I have my own dataset and I want to train with your model.
But I have a question about the initial estimation when training. In your paper you wrote that "in the feature space, we find the K NNs to
x among all training instances in terms of the Euclidean distances"
I don't know the feature space your made from the beginning before training model or you got from one batchsize when training ?
Hope you can answer this question for me?

Hi, how to fix this issue, I have train with utk dataset.

(base) test3@ai3:~/Documents/test/ai/ag/ag-v3$ python3.6 code/op.py --dataset utk --regression global --experiment_setting train --im_path /mnt/test3/age
Namespace(backbone='Global_Regressor', ckpt_dir='ckpt', data_path='datalist', dataset='utk', experiment_setting='train', experiment_title='global/utk/train', fold=0, gpu=0, im_path='/mnt/test3/age', regression='global', tau=0.2)
Global Regression
Age range partition:  [[21, 28], [24, 33], [29, 40], [34, 49], [41, 60]]
cuda:0
Get Global_Regressor
ckpt/global/UTK/train/fold0/utk_train.pth
Traceback (most recent call last):
  File "code/op.py", line 62, in <module>
    global_regression(args, train_data, test_data, sampling=sampling, sample_rate=sample_rate)
  File "/home/test3/Documents/test/ai/ag/ag-v3/code/Prediction.py", line 22, in global_regression
    checkpoint = torch.load(initial_model, map_location=device)
  File "/home/test3/.local/lib/python3.6/site-packages/torch/serialization.py", line 593, in load
    return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)
  File "/home/test3/.local/lib/python3.6/site-packages/torch/serialization.py", line 763, in _legacy_load
    magic_number = pickle_module.load(f, **pickle_load_args)
EOFError: Ran out of input

@nhshin-mcl can you please indicate how to train the model on any of the 4 datasets?

would you please share the preprocessing code for testing set so that we can verify the testing result reported in the paper?

Is there any protocol with UTKface dataset's train valid spilt?