can't debug anymore

need to tweak my caffeine dose... just until this model is fully trained, i promise.

17:50 just woke up, gonna start logging stuff here as i seen hitori and lelouch do it, it seems rather fun

plan for tonight:

• going to bypass the usual data cleaning—diving straight into hyperparameter optimization (feeling brave, maybe some bayesian approaches).
• if the gpu overheats past midnight, i’ll shift to writing some documentation for the neural net architecture.

20:58 not totally sure how i’ll conquer the convergence issues, but i have a couple of papers bookmarked that might shine some light.

for epoch in range(num_epochs):
    adjust_learning_rate(optimizer, epoch, init_lr=0.01)
    train()
    validate()

• thinking about dropout rates and learning rates adjustments
• might revisit a few chapters from “deep learning” by goodfellow et al. for inspiration

00:03 woke up after a 3-hour power nap, need to wrap up some model evaluations

04:19 still up, model’s training accuracy is stuck at 87% - been tweaking the loss function.

def custom_loss(output, target):
    loss = torch.mean((output - target) ** 2)
    return loss

wondering if i’ll ever see the bed tonight, will queue up some long overdue model tests and a batch of simulations before i crash.

06:47 observing some interesting patterns in the training data with the new visualization tool i coded up last week.

it’s intriguing but expected; using mixed precision training, i’m seeing a slight increase in efficiency. wonder why more people don’t share these little hacks—though the fear of giving away competitive edges makes sense.

08:30 feeling really shitty right now, might just rest for a bit

09:30 Noticed a minor memory leak during the test runs. Quick fix with a few lines:

void sadly_this_works(resource *res) {
    if (!res->is_active) {
        free(res);
    }
}

10:50 Reviewed a friend’s patch; they missed a semaphore release. Fixed it and pushed it upstream

11:20 Shutting down