00:00TensorBoard is your MRI for neural nets. See gradients, weights, activations live.
00:06I found a dead layer in 30 seconds. Would have taken days without it.
00:12Log everything, loss, accuracy, weights, gradients.
00:16I'll add callbacks and show you a dying gradient live.
00:21GradCam shows exactly where your model is looking. Is it the cat or the carpet?
00:25My model was predicting cat because of the carpet.
00:30GradCam caught it instantly.
00:32Essential for computer vision debugging.
00:35I'll run GradCAM on your day 85 image model live.
00:40Tef data bugs are silent. Wrong shuffling. No prefetch caching disasters.
00:46I trained for 12 hours because prefetch was off. Data starvation.
00:52Always inspect your pipeline with .as underscore numpy underscore iterator.
00:56I'll show a broken pipeline fix with prefetch and cache.
01:02One command shows if your model is built wrong. Input, output shapes, param count.
01:07I once had a model with zero parameters. Summary. Caught it.
01:12Always run model.summary after building.
01:15I'll break the day 85 model shape live. Then fix it.
01:20Add asserts to catch impossible values. Nan, negative probabilities, wrong shapes.
01:26I once had negative probabilities. Assert. Saved me.
01:32Fail fast. Fail loud.
01:34I'll add asserts to day 85. Watch them scream when I break it.
01:40Weights and biases logs everything. Compare 50 runs in one click.
01:45I found my best model from three weeks ago. Wand B remembered.
01:50Free tier is amazing.
01:53I'll log your day 85 training live.
01:55Your training is slow. Profiler shows exactly where. Data, GPU, CPU.
02:03I was blaming the model. It was tf.data the whole time.
02:09Built into TensorBoard.
02:12I'll profile your day 85 app live.
02:15Shapp tells you exactly which words made it positive or negative.
02:20My model hated the word but. Shapp showed me.
02:24Explainable AI gold standard.
02:28I'll explain a wrong prediction live.
02:31Write tests for your pre-processing, model output, predictions.
02:36Never break silently again.
02:38I added tests after a disaster.
02:41Never going back.
02:44CI slash CD for AI.
02:47I'll write five tests for day 85 app live.
02:49Automatically validate every new batch.
02:53Catch corrupted data before training.
02:56I once trained on all NAN images.
02:59Great expectations would have saved me.
03:02Data tests are as important as code tests.
03:06I'll add three expectations live.
03:09Your complete debugging checklist.
03:12Print it, save it, live by it.
03:13I have this taped above my monitor.
03:16Saved me hundreds of hours.
03:19This plus TensorBoard plus the seven steps equals unstoppable.
03:24Download link in description.
03:26Free forever.
03:29Every single one of us has lost days to these bugs.
03:32Now you're immune.
03:34I still have the Slack message where I cried for six hours straight.
03:38This is what separates juniors from seniors.
03:43You're now in the 1% who actually debug fast.
Comments