Saving ipynb file corrupts the JSON format - Unexpected end of JSON input

[Rane90]

Hi,

I've been working with vs code via SSH for quite a few years now and encountered and new bug yesterday that makes working with .ipynb files almost impossible.

What happens is - After saving the file the JSON format is corrupted and it cannot be accessed again. I need to recreate it from scratch.
There are some strange behaviors for this bug:

1. Often the file saving happens regardless of me saving or not.
2. Sometimes vs-code explicitly asks me if it's ok to save stating it needs to overwrite changes.
3. Looking into the JSON file with the text editor, it seems as though it simply "cuts out" the last part of the file. See example:

  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "for epoch in range(num_epochs):\n",
    "    losses = []\n",
    "    for batch_images, batch_labels in tqdm(train_loader):\n",
    "        optimizer.zero_grad()\n",
    "        batch_images = batch_images.to(device).cuda().fl

*** Notice the string doesn't end properly.

I assume this has to do with the print output of the code cell (I'm using Python's tqdm library)

Does this issue occur when all extensions are disabled?: Yes/No

All extensions are active.

• VS Code Version: 1.77.3
• OS Version: Linux:

NAME="CentOS Stream"
VERSION="8"
ID="centos"
ID_LIKE="rhel fedora"
VERSION_ID="8"
PLATFORM_ID="platform:el8"
PRETTY_NAME="CentOS Stream 8"
ANSI_COLOR="0;31"
CPE_NAME="cpe:/o:centos:centos:8"
HOME_URL="https://centos.org/"
BUG_REPORT_URL="https://bugzilla.redhat.com/"
REDHAT_SUPPORT_PRODUCT="Red Hat Enterprise Linux 8"
REDHAT_SUPPORT_PRODUCT_VERSION="CentOS Stream"

Steps to Reproduce:

1. This is the cell causing the issue:

for epoch in range(num_epochs):
    losses = []
    for batch_images, batch_labels in tqdm(train_loader):
        optimizer.zero_grad()
        batch_images = batch_images.to(device).cuda().double().reshape(batch_images.shape[0], 3, 224, 224)
        output = model(batch_images)
        loss = nn.MSELoss()(batch_images, output)
        loss.backword()
        losses.append(loss.item())
        optimizer.step()
    if epoch % 1 == 0:
        print("Epoch [{}/{}], MSE Loss = {:.8}".format(epoch, num_epochs, np.mean(losses)))
    if epoch % 1 == 0:
        umap_validation(model, train_loader, device, epoch)

(Let me know if you wish me to upload the entire notebook)
2. The dataset I'm using is very large - The COCO dataset (~18GB)

[Rane90]

Of course, I can add these lines to the JSON raw text but this is just a workaround:

               " # End of text
            ]
        }
    ]
}

[rebornix]

@Rane90 Thanks for the detailed report, it's a critical bug and thank you for sharing with us.

Often the file saving happens regardless of me saving or not

Do you have auto save turned on?

Sometimes vs-code explicitly asks me if it's ok to save stating it needs to overwrite changes.

This usually means that there is another process attempting to modify the file, and when that happens and you attempt to save, the file on disk is newer so we will ask if you want to override.

2. The dataset I'm using is very large - The COCO dataset (~18GB)

This made me wonder if it's possible that the extension host crashed when we attempt to save. Is the dataset loaded into the kernel fully?

[Rane90]

Thank you for your response.

Autosave is not turned on.
And yes, a large portion of the dataset is loaded into the kernel. I'm working on a remote server and have even been notified that I'm using too much disk space. Although it's very logical that these two are related I cannot see how to avoid this behavior.

[atapley]

Commenting to say I also just ran into this issue. I was attempting to save a dataset and I ran out of room in my container. After this happened, the notebook got corrupted and cut off a portion of my file that cannot be re-added.