Use MHingeGAN to do multi-label conditional generation on CLEVR dataset.
-
Loss Function:
Origin loss functions of the auxiliary classifier are:
New loss functions in this project are:
-
Initialization:
Initialize weights of conditional batch normalization layers in the generator with [nn.init.orthogonal_] instead of [nn.init.ones_].
All the training images are in ./image
-
./label/objects.json
This file is a dictionary file that contains the number of objects and the idexes. There are totally 24 objects in i-CLEVR datasets with 3 shapes and 8 colors. -
./label/train.json
The file is for training. The number off training data is 18012. train.json is a dictionary where keys are filenames and values are objects/ For example: {"CLEVR_train_001032_0.png": ["yellow sphere"], "CLEVR_train_001032_1.png": ["yellow sphere", "gray cylinder"], "CLEVR_train_001032_2.png": ["yellow sphere", "gray cylinder", "purple cube"], ... } One image can include objects from 1 to 3 -
./label/test.json
The file is for testing. The number of testing data is 32. test.json is a list where each element includes multiple objects For example: [['gray cube'], ['red cube'], ['blue cube'], ['blue cube', 'green cube'], ...]
(./label/test.json)
Accuracy evaluated by ./model/eval_model/classifier_weight.pth(Resnet18) = 0.83