The "Red Logs"

The exam loves troubleshooting questions. You see an error message. What is the root cause?

---

1. Out of Memory (OOM)

• Symptom: ResourceExhaustedError: OOM when allocating tensor...
• Cause: The batch size is too large for the GPU VRAM.
• Fix:
  1. Decrease Batch Size.
  2. Use Gradient Accumulation (Simulate larger batch).
  3. Upgrade GPU (T4 -> A100).

---

2. Loss is NaN (Not a Number)

• Symptom: Loss: NaN after Step 100.
• Cause: Gradient Explosion. The numbers became too big for float32.
• Fix:
  1. Clip Gradients (clipnorm=1.0).
  2. Decrease Learning Rate.
  3. Check for dirty data (e.g., dividing by zero).

---

3. Permission Denied (403)

• Symptom: google.api_core.exceptions.PermissionDenied: 403 Access Not Configured
• Cause: The Service Account running the training job does not have permission to read the GCS bucket or write logs.
• Fix: Grant Storage Object Admin to the Compute Engine Service Account (or Custom SA).

---

4. Slow Training (Starvation)

• Symptom: Training takes 1 week instead of 1 day. GPU usage is low.
• Cause: Input Pipeline bottleneck.
• Fix: Use tf.data.Dataset.prefetch() and cache().

---

Knowledge Check