In Minecraft, torches are very important for mining. No one can get all the diamonds without a torch. So is data-mining, A special torch named PyTorch can help us get the diamonds in data. Minetorch is a tools collection for miners, to use PyTorch in a more convenient way.

• Logging of course.
• Play with Google Sheets.
• Visualization. built-in with Tensorboard or Matplotlib.
• Resumption. Default will resume from the last checkpoint.
• Many hook points.
• Training code. Or actually a namespace, used to easily sperate different training status.

pip install minetorch

Minetorch can easily work with Google Sheets. To step up, you should have a Google account and, if you are in China, a proxy.

This is the demo sheet that produced with Minetorch training on MNIST dataset.

Enable Google API

1. Sign in to https://console.developers.google.com with your Google account
2. Create a new Project
3. In the API Libarary, search drive and sheet, and enable them both. During this, Google should have alert you about creating the service account. If you have get the right credential file, ignore step 4.
4. Create a Google Service Account in at https://console.developers.google.com/iam-admin/iam .

After all these steps you should get a json credential file which looks like this:

{
  "type": "xxxxxx",
  "project_id": "xxxxxx",
  "private_key_id": "xxxxxx",
  "private_key": "-----BEGIN PRIVATE KEY-----\n .......... \n-----END PRIVATE KEY-----\n",
  "client_email": "xxxxxx@xxxxxx-xxxxxxxxxx.iam.gserviceaccount.com",
  "client_id": "xxxxxxxxxxxxxxxxxxxxx",
  "auth_uri": "https://accounts.google.com/o/oauth2/auth",
  "token_uri": "https://oauth2.googleapis.com/token",
  "auth_provider_x509_cert_url": "https://www.googleapis.com/oauth2/v1/certs",
  "client_x509_cert_url": "https://www.googleapis.com/robot/v1/metadata/x509/xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx.iam.gserviceaccount.com"
}

Create the spreadsheet

1. At http://drive.google.com/, create a new spreadsheet.
2. Share that spreadsheet with your service account, you can find your service account email address in your credential file.
3. Get the id of the spreadsheet, it's right in the middle of the URL.

Ready to go

In your code, just create a minetorch.spreadsheet.GoogleSheet instance and pass it to Miner.

miner = Miner(
    # ...
    sheet=GoogleSheet('Your sheet id', 'path-of-your-google-service-account-credensial-file.json')
    # ...
)

miner.train()

and that's it, your spreadsheet should be automatically updated when you start your training process.

Minetorch update the sheet in async way, so it will not slow down the training speed.

1. Clone this repo

git clone git@github.com:louis-she/minetorch.git

2. Goto examples and execute

python mnist.py

The mnist.py is pretty much like PyTorch official mnist example https://github.com/pytorch/examples/blob/master/mnist/main.py, with some minor changes to be adapted to Minetorch.

This will create an alchemistic_directory and data directory in the working directory.

3. Execute tensorboard to visualize

tensorboard ./log

or if you used matplotlib as the drawer, there would be 2 images under the path ${alchemistic_directory}/${code}/graph named loss.png and accuracy.png

loss.png	accuracy.png

Note: Since PyTorch 1.0 comes with CUDA10 but the newest version of Tensorboard just supported CUDA9, so if you are using Tensorboard as a drawer, make sure you got CUDA9 installed, or it's better to use matplotlib. Due to this, Minetorch default drawer has changed to matplotlib instead of Tensorboard.

4. Now let's change the loss function to cross_entropy(just follow some instructions in the comments of mnist.py), and train the mnist again.

Minetorch lets users focusing on the only necessary things, and will take care of the others. It's more like a skeleton and users should provide necessary components to make it work.

The parameters users must provide:

And that's it, minetorch will take care of other things like logging, resuming, visualization etc... The names of the components are actually the parameters of the Trainer class. The other optional parameters are:

Minetorch provided many hook points for users to control the training behaviors. All the hook functions receive current trainer instance as the arguments so the hook function can access all the functions and status of Trainer.

• More hook points.
• Number of x-axis of Drawer generated images are not changed after resuming from a checkpoint.
• Drawer DB Adapter. To persist drawer data to DB and then can be visualization by any tools.
• Graceful exists.
• Dockerize.