TensorFlow saved my life last night
When I realized what it was I swore and jumped backward #parody
Like any good story, this one begins in the middle of the night on Sunday. I can't recall the exact time, but it was early, around 3-4am. I rolled out of bed, thinking I needed to attempt a long run. Twenty miles seemed like the goal I was aiming for, and I felt ready.
I started running, it was cold, and the full moon was out. I almost didn’t need my headlamp but I had it anyway, climbing higher up the hill in the dark, snowy terrain. Unfortunately—or fortunately—no one had come up here. The only tracks I saw were from animals, and I found myself sinking up to my waist in the snow. I could hardly call this running anymore; it was more like post-holing in running shoes and shorts.
Occasionally, I see eyes glowing in the dense trees above and below me, reflecting my headlamp back at me. This is common during night runs, but it doesn't make it any less eerie. Most of them are deer—if I shout and they run, they are deer. If there are pairs of eyes, they were usually deer. The more eyes the better because it is definitely deer then. If there is a single set and it doesn’t run away when I yell, then I have to slow down to assess what I was dealing with. Deer, alright, good—run on. One of these times I’ll turn over the not-deer card, it’s a numbers game.
I have a decent knife and a small .380 in my backpack, with nothing in the chamber. I've never taken either of them on a long run, but I find comfort in knowing I could. Additionally, I have trekking poles as an extra buffer which has worked well against stray dogs that are much meaner than their owners ever remember them being.
I often see deer bones on the side of the trail, especially higher up. They serve as a reminder that I'm not alone up here, with frequent sightings of coyotes and the occasional mountain lion or bobcat.
As I round the corner, I encounter a fork in the path—left or right. I choose right, and then I see it. It takes me a few seconds to fully grasp the gravity of what I'm witnessing.
A full-sized deer lies on the trail... and it hasn't been eaten yet...
IT HASN’T BEEN EATEN YET!!
A wave of emotion sweeps over me. I curse, jump back, and look around. Ripping off my headphones—which I know I shouldn't be wearing—I survey my surroundings. I then start scrambling to find my gun. It always takes longer than expected, I should have practiced more at home. More cursing: "Damn it! Where's my gun?!?" Finally, I find it, rack a round into the chamber, and feel a bit of relief. Although, I'm not out of the woods yet. I don't see anything, but I'm certain something is watching me, the deer is a valuable prize indeed.
The emotions are like an emotional grenade. I experience the kind of fear that makes you want to escape—either by leaving the theater during a horror movie or waking up from a nightmare. It's the kind that makes you feel truly present, as your brain demands your full attention and floods your body with adrenaline. I also feel an odd sense of sadness, an emotion I can't quite understand—empathy for the deer, perhaps, or sorrow for my current situation. Additionally, I'm struck by a long-forgotten feeling: the desire or need for someone more experienced to be there—another human, someone more adult, stronger, and smarter than me. Someone I can rely on to get me out of this predicament. It feels like I'm having flashbacks to my childhood, clutching a knife in the snow, waiting for the end.
I need to muster the mental strength to continue, to halt the spiral of emotion and panic. Panicking doesn't help anyone; making the best decisions requires a clear mind. Emotions can cloud your judgment. Emotions can even make you run, which is the worst possible reaction in this situation.
I find the strength I need in the most unlikely of places.
I realize I've faced worse, much much worse.
I've dealt with TensorFlow.
That’s right! I've been paid to take TensorFlow into production, a task no human should ever attempt. I've compiled TensorFlow with Bazel. I've tried to compile TensorFlow to accept the latest Nvidia drivers. I've run TensorFlow from official blogs by its creators, only to find it nonfunctional despite matching exact version numbers. I've attempted multi-GPU processing in TensorFlow. I’ve reacted to dozens of unhandled exceptions with little to no help on the internet. I've lived in a world where TensorFlow continues to persist, shrieking onward, even as applied professionals have turned to PyTorch.
Deep Mind? PyTorch.
OpenAI? PyTorch.
Amazon? Used MxNet (probably more PyTorch now).
Meta? PyTorch.
If you look at Google trends we can now see that PyTorch has taken over with community interest:
Maybe by the year 2025 we will finally send it to the same place Theano and Caffe went and keep them company. In the meantime, I thank TensorFlow for giving me the strength to run on! I don’t know how I’ll ever repay you… despite being grateful I still have no plans to ever use you again. Friends encourage friends to try PyTorch 2.0 instead.