The attempted training of the U-Net on the Jetson NX has been a bit slow, making odd progress over 2 nights, and I’m not sure if it’s working. I’ve had to reduce batch size to 1, and the filter size, which has reduced the number of parameters by about a factor of 10, and still, loading the NN into memory sometimes dies on a concatenation call. The number of images per batch can also crash it, so perhaps some memory can be saved with a better image loading process.
Anyway, projects under an official NVIDIA repo are suggesting that we should be able to train smaller networks like resnet18, with 11 million parameters, on the Jetson. So maybe we can still avoid the cloud.
But judging by the NVIDIA TLT info, any training of resnet50s or 100s are going to need serious GPUs and memory and space for training.
After looking at Google, Amazon and Microsoft offerings, the AWS g4dn.xlarge instance looks like it might be the best option, at $0.526/hr, or Google’s got a T4 based compute engine for only $0.35/hr. These are good options, if 16GB of video ram will be enough. It should be, because we’re working with like 5GB on the Jetson.
Microsoft has the NC6 option, which looks good for a much more beefy GPU and memory, at $0.90/hr.
We’re just looking at Pay-as-you-go prices, as the 1-year and 3-year commitments will end up being expensive.
I’m still keen to try train on the Jetson, but the cloud is becoming more and more probable. In Sweden, visiting Miranda, we’re unable to order a Jetson AGX Xavier, the 32GB version. Arrow won’t ship here without a VAT number, and SiliconHighway is out of stock.
So, attempting Cloud GPUs. If you want to cut to the chase, read this one backwards. So many problems. In the end, it turned out setting it up yourself is practically impossible, but there is an ‘AI Platform’ section that works.
Amazon AWS. Tried to log in to AWS. “Authentication failed because your account has been suspended.” Tells me to create a new account. But then brings me back to the same failure screen. Ok, sending email to their accounts department. Next.
Google Cloud. I tried to create a VM and add a T4 GPU, but none of the regions have them. So I need to download the Gcloud SDK and CLI tool first, to run a command to describe the regions, according to the ‘Before you begin‘ instructions..
Ok, GPUs will only run on N1 and A2 VMs. The A2 VMs are only for A100s, so I need an N1 VM in one of these regions, and we add a T4 GPU.
There’s an option to load a specific docker, and unfortunately they don’t seem to have one with both Pytorch and TF2. Let’s start with TF2 gcr.io/deeplearning-platform-release/tf2-gpu.2-4
So this looks like a good enough VM. 30GB RAM, 8 cpus. For europe-west3, the cost is about 50 cents / hr for the VM and 41 cents / hr for the GPU.
| n1-standard-8 | 8 | 30GB | $0.4896 | $0.09840 |
| 1 GPU | 16 GB GDDR6 | $0.41 per GPU |
So let’s round up to about $1/hour. I ended up picking the n1-standard-4 (4 cpus, 15 gb ram).
At these prices I’ll want to get things up and running asap. So I am going to prep a bit, before I click the Create VM button.
I had to try a few things to find a cloud instance with a gpu, because the official list didn’t really work. I eventually got one with a T4 GPU from europe-west4-c.
It seems like Google Drive isn’t really part of the google cloud platform ecosystem, so I started a storage bucket with 50GB of space, and am uploading the chicken images to it.
The instance doesn’t have pip or jupyter installed. So let’s do that…
ok so when I sudo’ed, I got this error
Jul 20 14:45:01 chicken-vm konlet-startup[1665]: {"errorDetail":{"message":"write /var/lib/docker/tmp/GetImageBlob362062711: no space left on device"},"error":"write /var/lib/docker/tmp/GetImageBl
Jul 20 14:45:01 chicken-vm konlet-startup[1665]: ).
Jul 20 14:45:01 chicken-vm konlet-startup[1665]: 2021/07/20 14:43:04 No containers created by previous runs of Konlet found.
Jul 20 14:45:01 chicken-vm konlet-startup[1665]: 2021/07/20 14:43:04 Found 0 volume mounts in container chicken-vm declaration.
Jul 20 14:45:01 chicken-vm konlet-startup[1665]: 2021/07/20 14:43:04 Error: Failed to start container: Error: No such image: gcr.io/deeplearning-platform-release/tf2-gpu.2-4
Jul 20 14:45:01 chicken-vm konlet-startup[1665]: 2021/07/20 14:43:04 Saving welcome script to profile.d
So 10GB wasn’t enough to load gcr.io/deeplearning-platform-release/tf2-gpu.2-4 , I guess.
Ok deleting the VM. Next time, bigger hard drive. I’m now adding a cloud storage bucket and uploading the chicken images, so I can copy them to the VM’s drive later. It’s taking forever. Wow. Ok.
Now I am trying to spin up a VM again, and it’s practically impossible. I’ve tried every region and zone possible. Ok europe-west1-c. Finally. I also upped my ‘quota’ of gpus, under IAM->Quotas, in case that is a reason I couldn’t find a GPU VM. They reviewed and approved it in about 15 minutes.
+------------------+--------+-----------------+
| Name | Region | Requested Limit |
+------------------+--------+-----------------+
| GPUS_ALL_REGIONS | GLOBAL | 1 |
+------------------+--------+-----------------+
So after like 10 minutes of nothing, I see the docker container started up.
68ee22bf268f gcr.io/deeplearning-platform-release/tf2-gpu.2-4 "/entrypoint.sh /run…" 5 minutes ago Up 4 minutes klt-chicken-vm-template-1-ursn
I’ve enabled tcp:8080 port in the firewall settings, but the external ip and new port don’t seem to connect. https://35.195.66.139:8080/ Ah ha. http. We’re in!
Jupyter Lab starting up.
So I tried to download the gcloud tools to get gsutil to access my storage bucket, but was getting ‘Permission denied’, even as root. I chown’ed it to my user, but still no.
I had to go out, so I stopped the VM. Seems you can’t suspend a VM with a GPU. I also saw when I typed ‘sudo -i’ to switch user to root, it said to ‘docker attach’ to my container. But the container is just like a tty printing out logs, so you can get stuck in the docker, and need to ssh in again.
I think the issue was just that I need to be inside the docker to do things. The VM you log into is just a minimal container running environment. So I think that was my issue. Next time I install gsutil, I’ll run ‘docker exec -it 68ee22bf268f bash’ to get into the docker first.
Ok fired up the VM again. This time I exec’ed into the docker, and gsutil was already installed. gsutil cp -r gs://chicken-drive . is copying the files now. It’s slow, and it says to try with -m, for parallel copying, but I’m just going to let it carry on for now. It’s slow, but I can do some other stuff for now. So far our gcloud bill is $1.80.
Ok, /opt/jupyter/chicken-drive has my data now. But according to /opt/jupyter/.jupyter/jupyter_notebook_config.py, I need to move it under /home/jupyter.
Hmm. No space left on drive. What? 26GB all full. But it wasn’t full a second ago. How can moving files cause this? I guess the mv operation must copy and then delete. Ok, so deleting the new one. Let’s try again, one folder at a time. Oh boy. This is something a bit off about the google process. I didn’t start my container, and if I did, I’d probably map a volume. But the host is sort of read only. Anyway. We’re in. I can see the files in Jupyter Lab.
So now we’re training U-Net binary classification using keras-unet, by karolzak, based on the kz-isbi-chanllenge.ipynb notebook.
But now I’m getting this error when it’s clearly there…
FileNotFoundError: [Errno 2] No such file or directory: '/OID/v6/images/Chicken/train/'
Ok well I can’t work it out but changing it to a path relative to the notebook worked. base_dir = “../../../”
Ok first test round of training, binary classification: chicken, not-chicken. Just 173 image/mask pairs, 10 epochs of 40 steps.
Now let’s try with the training set. 1989 chickens this time. 50/50 split. 30 epochs of 50 steps. Ok second round… hmm, not so good. Pretty much all black.
Ok I’m changing the parameters of the network, fixing some code, and starting again.
I see that the pngs were loading float values, whereas in the example, they were loading ints. I fixed it by adding a m = m.convert(‘L’) to the mask (png) loading code. I think previously, it was training with the float values from 0 to 1, divided by 255, whereas the original example had int values from 0 to 255, divided by 255.
So I’m also resetting the parameters, to make this a larger network, since we’re training in the cloud. 512×512 instead of 256×256. Batch size of 3. Horizontal flip augmentation. 64 filters. 10 epochs of 100 steps. Go go go. Ok, out of memory. Batch size of 1. Still out of memory. Back to test set of 173 chickens. Ok it’s only maxing at 40% RAM now. I’ll let it run.
Ok, honestly I don’t know anymore. What is it even doing? Looks like it’s inversing black and white. That’s not very useful.
Ok before giving up, I’m going to make some changes.
The next day, I’m starting up the VM. Total cost so far, $8.84. The files are all missing, so I’m recopying, though using the gsutil -m cp -R gs://chicken-drive . option, and yes it is a lot faster. Though it slows down.
I think the current setup is maybe failing because we’re using 173 images with one kind of augmentation. Instead of 10 epochs of 100 steps of the same shit, let’s rather swap out the training images.
First problem is that Keras is basically broken, in this regard. I’ve immediately discovered that saving and loading a checkpoint does not save and load the metrics, and so it keeps evaluating against a loss of infinity, instead of what your saved model achieved. Very annoying.
Now, after stopping and restarting the VM, and enabling all cloud APIs, I’m having a new problem. gsutil no longer works. After 4% copied, network throughput drops to 0.0B/s. I tried reconnecting and now get:
Connection via Cloud Identity-Aware Proxy Failed Code: 4003 Reason: failed to connect to backend You may be able to connect without using the Cloud Identity-Aware Proxy.
I’ve switched back to ‘Allow default access’. Still getting 4003.
Ok, I’ve deleted the instance. Trying again. Started it up. It’s not installing the docker I asked for, after 22 minutes. Something is wrong. Let’s try again. Stopping VM. I’m ticking the ‘Run as priviliged’ box this time.
Ok now it’s working again. It even started up with the docker ready. I’m trying with the multiprocess copying again, and it slowed down at 55%, but is still going. Phew. Ok.
I changed to using the TF2 SavedModel format. Still restarts the ‘best’ metric. What a piece of shit. I can’t actually believe it. Ok I wrote my own code for finding the best, by saving all weights with the val_loss in the filename, and then loading the best weights for the next epoch. It’s still not perfect, but it’s better than Keras overwriting the best weights every time.
Interestingly, it seems like maybe my training on the Jetson was actually working, because the same weird little vignette-ing is occurring.
Ok we’re up to $20 billing, on gcloud. It’s adding up, but not too badly yet. Nothing seems to be beating a round of training from like 4 hours ago, so to keep things more exploratory, I added a 50/50 chance to pick from the saved weights at random, rather than loading the winner every time.
Something seems to be happening. The vignette is shrinking, but some chicken border action, maybe.
I left it running overnight, and this morning, we’re up to $33 spent, and today, we can’t log into the VM again. Pretty annoying. Of the 3 reasons for ‘Permission denied’, only one makes sense, Your key expired and Compute Engine deleted your ~/.ssh/authorized_keys file.
Same story if I run the gcloud commands: gcloud beta compute ssh –zone “europe-west4-c” “chicken-vm-template-1” –project “gpu-ggr”
So I apparently need to add a new public key to the Metadata section. I just know something is going to go wrong. Yeah, so I did everything I know I’m supposed to do, and it didn’t work. I generated an OpenSSH private/public key pair in PuttyGen, I changed the permissions on the private key so that only I have access, I updated the SSH Keys in the VM instance metadata, and the metadata for good measure. And ssh -i opensshprivate daniel_brownell@34.91.21.245 -v just ends up with Permission denied (publickey).
ssh-keygen -t rsa -f ~/.ssh/gcloud_instance1 -C daniel_brownell
Ok and then print the public key, and copy paste it to the VM Instance ‘Edit…’ / SSH Keys… and connect with PuTTY with the private key and… nope. Permission denied (publickey).. Ok I need to go through these answers and find one that works. Same error with windows cmd line ssh, except also complains that the openssh key is an invalid format. Try again later.
Fuck you gcloud. Ok I’m stopping and deleting the VM. $43 used so far.
Also, the training through the night didn’t improve on the val_loss score. Something’s fucked.
Ok I’ve started it up again a few days later. I was wondering about the warnings at the beginning of my training that carious CUDA things were not installed. So apparently I need:
cos-extensions install gpu
and… no space left on device
Ok so more space.
/dev/sda1 31G 22G 9.2G 70% /mnt/stateful_partition
So I increased the boot disk to 35GB and called ‘ cos-extensions install gpu’ again, after cd’ing into /mnt/stateful-partition and it worked a bit better. Still has ‘ERROR: Unable to load the kernel module 'nvidia.ko'.‘ in the logs though. But install logs at ./mnt/stateful_partition/var/lib/nvidia/nvidia-installer.log say its ok…
So the error now is ‘Could not load dynamic library ‘libcuda.so.1′; dlerror: libcuda.so.1: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/nvidia/lib:/usr/local/nvidia/lib64’
And so we need to modify the docker container run command, something like the example in the instructions.
Ok so our container is… gcr.io/deeplearning-platform-release/tf2-gpu.2-4
According to this stackoverflow answer, this already has everything installed. Ok but the host needs the drivers installed.
tf.config.list_physical_devices('GPU')
[]
So yeah, i think i need to install the cos crap, and restart the container with those volume and device bits.
docker stop klt-chicken-vm-template-1-ursn
docker run \
--volume /var/lib/nvidia/lib64:/usr/local/nvidia/lib64 \
--volume /var/lib/nvidia/bin:/usr/local/nvidia/bin \
--device /dev/nvidia0:/dev/nvidia0 \
--device /dev/nvidia-uvm:/dev/nvidia-uvm \
--device /dev/nvidiactl:/dev/nvidiactl \
gcr.io/deeplearning-platform-release/tf2-gpu.2-4
...
[I 14:54:49.167 LabApp] Jupyter Notebook 6.3.0 is running at:
[I 14:54:49.168 LabApp] http://46fce08b5770:8080/
[I 14:54:49.168 LabApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation).
^C^C^C^C^C^C^C^C^C^CNot so good. Ok can’t access it either. -p 8080:8080 fixes that. It didn’t like --gpus all.
“Unable to determine GPU information”. Container optimised shit.
Ok I’m going to delete the VM again. Going to check out these nvidia cloud containers. There’s 21.07-tf2-py3 and NGC stuff.
So I can’t pull the dockers cause there’s no space, and even after attaching a persistent disk, not, because things are stored on the boot disk. Ok but I can tell docker to store stuff on a persistent disk.
/etc/docker/daemon.json:
{
"data-root": "/mnt/x/y/docker_data"
}root@nvidia-ngc-tensorflow-test-b-1-vm:/mnt/disks/disk# docker run --gpus all --rm -it -p 8080:8080 -p 6006:6006 nvcr.io/nvidia/tensorflow:21.07-tf2-py3 docker: Error response from daemon: OCI runtime create failed: container_linux.go:380: starting container process caused: process_linux.go:545: container init caused: Running hook #0:: error running hook: exit status 1, stdout: , stderr: nvidia-container-cli: initialization error: nvml error: driver not loaded: unknown.
Followed the ubuntu 20.04 driver installation,
cuda : Depends: cuda-11-4 (>= 11.4.1) but it is not going to be installed
E: Unable to correct problems, you have held broken packages.
Oh boy. Ok so I used this trick to make some /tmp space:
mount --bind /path/to/dir/with/plenty/of/space /tmpand then as per this answer and the nvidia instructions:
wget https://developer.download.nvidia.com/compute/cuda/11.1.0/local_installers/cuda_11.1.0_455.23.05_linux.run
chmod +x cuda_11.1.0_455.23.05_linux.run
sudo ./cuda_11.1.0_455.23.05_linux.run
or some newer version:
wget https://developer.download.nvidia.com/compute/cuda/11.4.1/local_installers/cuda_11.4.1_470.57.02_linux.run
sudo sh cuda_11.4.1_470.57.02_linux.run‘boost::filesystem::filesystem_error’
Ok using all the space again. 32GB. Not enough. Fuck this. I’m deleting the VM again. 64GB. SSD persistent disk. Ok installed driver. Running docker…
And…
FFS. Something is compromised. In the time it took to install CUDA and run docker on an Ubuntu VM, an army of Indian hackers managed to delete my root user.
Ok. Maybe it’s time to consider AWS again for GPUs. I think I can officially count GCP GPU as unusable. Learned a few useful things, but overall, yeesh.
I think maybe I’ll just run the training on a cheap non-GPU VM on GCP for now, so that I’m not paying for a GPU that I’m not using.
docker run -d -p 8080:8080 -v /home/daniel_brownell:/home/jupyter gcr.io/deeplearning-platform-release/tf2-cpu.2-4Ok wow so now with the cpu version, the loss is improving like crazy. It went from 0.28 to 0.24 in 10 epochs (10 minutes or so). That sort of improvement was not happening after like 10 hours on the ‘gpu’.
So yeah, amazing. The code now does a sort of population based training, by picking a random previous set of weights instead of the best weights, half of the time. Overall it slows things down, but should result in a bit more variation in the end.
What finally worked
Ok there’s also an ‘AI platform – notebook’ option. I might try that too.
Ok the instance started up. But it failed to start 4 cron services: nscd, unscd, crond, sshd. CPU use goes to zero. Nothing. Ok so I need to ssh tunnel apparently.
gcloud compute ssh --project gpu-ggr --zone europe-west1-b notebook -- -L 8080:localhost:8080
Ok that was easy. Let’s try this.
Successfully opened dynamic library libcudart.so.11.0
‘ModelCheckpoint’ object has no attribute ‘_implements_train_batch_hooks’
Ok, needed to change all keras.* etc. to tensorflow.keras.*
Ok fuck me that’s a lot faster than CPU.
Permission denied: ‘weights-0.2439.hdf5’
Ok, let’s sudo it.
Ok there she goes. It’s like 20 times faster maybe. Strangely isn’t doing much better than the CPU though. But I’ll let it run for a bit. It’s only been a minute. I think maybe the CPU doing well was just good luck. Perhaps we trained them too well on the original set of like 173 images, and it was getting good results on those original images.
Ok now it’s been an hour or so, and it’s not beating the CPU. I’ve changed the train / validation set to 50/50 now, and the learning rate is randomly chosen between 0.001 and 0.0003. And I’m upping the epochs to 30. And the filters to 64. batch_size=4, use_batch_norm=True.
We’re down to 23.3 after an hour and a half. 21 now… 3 hours maybe now
Ok 5 hours, lets check:
Holy shit it’s working. That’s great. I’ll leave it running overnight. The overnight results didn’t improve much for some reason.
(TODO: learn about focal loss / dice loss / jaccard distance as possible change to loss function.? less necessary now.)
So it’s cool but it’s 364MB. We need it 1/4 size to run it on the Jetson NX I think.
So, retraining, with filters=32. We’re already down to 0.24 after an hour. Ok I stopped at 0.2104 after a few hours.
So yeah. Good enough for now.
There’s some other things to train, too.
The eggs in simulation: generate views, save images to disk. save segmentation images to disk.
Train the walking again with the gripper.
Eggs in the real world. Use augmentation to place real egg pics in scenes. Possibly use Mask-RCNN/YOLACT code with COCO, instead of continuing in Keras.
The now-working U-net binary chicken segmentation is in Keras, so there will be some tricks required, to run a multi-class segmentation detector, or multiple binary classifiers. Advice for multi-class segmentation is here and the multiple binary classifier advice is here.
When we finally try running it all on a Jetson, we will maybe need to shrink the neural network further. But that can be done last minute. It looks like we can save the h5fs file to TF2’s SavedModel format with model.save(model_fname) and convert to frozen graph, to import into TensorRT, the NVIDIA format. Similar to this. TensorRT shrinks neurons to single bytes, I believe.