Conditional Deep Convolutional Generative Adversarial Network (cDCGAN) | on the MNIST DataSet

---

Training

Training Loss Analysis

1. D trained every iteration:

• D performs too well (low D loss).
• G can’t keep up (high G loss).

2. D trained 2× less frequently:

• The generator still struggles to catch up (G loss is higher than D loss).

3. D trained 3× less frequently:

• Possibly a much better result (G loss is lower than D's).
• However, the trend suggests a role reversal, which is not desirable.

(alternative - decreese D's learning rate)

More traning data with D trained 3x less frequently:

The plot flips here, indicating the G started to fail. I think that best performance was likely around 30 iterations (see second plot).

Summary:

Training the D too often hinders the G’s laearning. Training it too rarely weakens its ability to distinguish. The best results come from balancing the learning pace of both networks.

📈 Legend

Symbol	Meaning
🟥 Red line	Discriminator loss
🟦 Blue line	Generator loss
D	Discriminator
G	Generator
📊	Average loss per epoch

---

Evaluating

Generated numbers by the trained model (0-9):

Inerpolation

between class 1 and 9: