Conditional GANs with Auxiliary Discriminative Classifier — Reproduction & Extensions

This repository contains the full implementation, experiments, and analysis for my CS436/536 final project:

“Conditional GANs with Auxiliary Discriminative Classifier: Reproduction and Early Extensions”
(Crystal Sembhi, Mudit Golchha, Binghamton University)

This project reproduces the ADC-GAN (Hou et al., ICML 2022) paper and compares it against AC-GAN and PD-GAN under a unified BigGAN-CIFAR10 framework.
It also introduces a small extension γ-ADC-GAN, which scales the auxiliary discriminative classifier loss by a factor γ.

---

1. Overview

Conditional GANs aim to generate class-specific images.
The commonly used AC-GAN uses an auxiliary classifier trained only on real images, which causes:

• low intra-class diversity
• mode collapse tendencies
• training instability

ADC-GAN introduces a discriminative classifier trained on both real and fake samples, and outputs:

Real class k → 2k
Fake class k → 2k + 1

This improves stability, diversity, and early FID/IS performance.

In this project:

1. Reproduce ADC-GAN on CIFAR-10 using the BigGAN backbone
2. Implement and compare AC-GAN and PD-GAN using identical training settings
3. Develop γ-ADC-GAN, a lightweight extension using γ ∈ {1.0, 0.5}
4. Evaluate all models on CIFAR-10 and a custom 1D synthetic task

---

2. Repository Structure

Conditional-GANs-with-Auxiliary-Discriminative-Classifier/
|
├── Custom_py
│   ├── custom_losses.py
│   └── custom_train_fns.py
├── experiments
│   ├── acgan_12k_20251205-000221
│   │   ├── metrics.csv
│   │   └── notes.md
│   ├── acgan_cifar10_biggan_20251203-170409
│   │   ├── metrics.csv
│   │   └── notes.md
│   ├── adcgan_cifar10_biggan_20251116-065557
│   │   ├── metrics.csv
│   │   └── notes.md
│   ├── adcgan_gamma05_12k_20251205-002946
│   │   ├── metrics.csv
│   │   └── notes.md
│   ├── adcgan_gamma05_20251204-004734
│   │   ├── metrics.csv
│   │   └── notes.md
│   ├── pdgan_12k_20251204-232757
│   │   ├── metrics.csv
│   │   └── notes.md
│   ├── pdgan_cifar10_biggan_20251202-203055
│   │   └── notes.md
│   └── pdgan_cifar10_biggan_20251203-163507
│       ├── metrics.csv
│       └── notes.md
├── Graph and Plots
│   ├── acgan_synthetic_1d.png
│   ├── adcgan_fid_cifar10.png
│   ├── adcgan_is_cifar10.png
│   ├── adcgan_synthetic_1d.png
│   ├── combine_losses.py
│   ├── combine_samples.py
│   ├── dloss_1d.png
│   ├── fid_0_12k_comparison.png
│   ├── fid_adc_pd_ac.png
│   ├── gamma_adcgan_synthetic_1d.png
│   ├── gloss_1d.png
│   ├── is_0_12k_comparison.png
│   ├── is_adc_pd_ac.png
│   ├── pdgan_synthetic_1d.png
│   ├── synthetic_1d_losses.png
│   └── synthetic_1d_samples.png
├── Main Notebook
│   └── MAIN_CODEBASE.ipynb
└── README.md

---

3. Running Experiments

3.1 Install Dependencies

pip install torch torchvision tqdm numpy matplotlib

---

3.2 Training ADC-GAN / AC-GAN / PD-GAN

Inside BigGAN-PyTorch/:

ADC-GAN

python train.py   --loss adcgan   --dataset C10   --use_ema   --batch_size 50   --num_D_steps 4   --save_every 2000   --test_every 2000   --experiment_name adcgan_run

AC-GAN

python train.py   --loss acgan   --dataset C10   --use_ema   --experiment_name acgan_run

PD-GAN

python train.py   --loss hinge   --dataset C10   --projection   --use_ema   --experiment_name pdgan_run

---

3.3 γ-ADC-GAN (New Hypothesis)

Set γ inside custom_losses.py / custom_train_fns.py:

G_lambda = gamma
D_lambda = gamma

Run:

python train.py   --loss adcgan   --experiment_name gamma05_adcgan   --use_ema

---

4. 1D Synthetic Experiment

The synthetic dataset is a 3-mode Gaussian mixture with means at -4, 0, and +4.

Run:

python train_1d.py

This produces:

• acgan_synthetic_1d.png
• pdgan_synthetic_1d.png
• adcgan_synthetic_1d.png
• gamma_adcgan_synthetic_1d.png
• dloss_1d.png
• gloss_1d.png

These visualizations help compare diversity and training stability.

---

5. Results Summary

CIFAR-10 Early FID (0–12k iterations)

Model	FID ↓	Behavior
AC-GAN	Slow improvement	Unstable classifier
PD-GAN	Moderate	Noisy projection term
ADC-GAN	Best early FID	Stable and diverse
γ-ADC-GAN (0.5)	Slight FID improvement at several points	Smoothest overall training

---

1D Synthetic Results

Model	Mode Coverage	Stability
AC-GAN	Collapses to 1–2 modes	Unstable
PD-GAN	Covers all modes	Noisy gradients
ADC-GAN	Best accuracy	Smooth discriminator signals
γ-ADC-GAN (0.5)	Preserves modes	Smoothest training curve

---

6. Contribution

Crystal Sembhi:

• Trained the ADC-GAN and PD-GAN with Cifar-10.
• Trained the AC-GAN for 1D Dataset.
• Ran the Evaluation Metrics.

Mudit Golchha:

• Trained the γ-ADC-GAN.
• Created Graph and plots for all the models.
• Ran the Evaluation Metrics.

---