Skip to main content
Python Software Foundation 20th Year Anniversary Fundraiser  Donate today!

A set of scripts to make the Neural Network training with pytorch faster

Project description

Scorch: utilities for network training with PyTorch

This package is a set of codes that may be reused quite often for different networks trainings.


You will need Python 3 to use this package.

You will need the following packages installed:

  • pytorch
  • numpy
  • scikit-image
  • scikit-learn

To use the notebooks for testing the model and the dataset you will need Jupyter Notebook or JupyterLab installed.


Here is the minimal command to run to train the model specified in MODEL_FILE with a dataset specified in DATASET_FILE, the data is located in the DATASET_PATH.

python --model MODEL_FILE --dataset DATASET_FILE --dataset-path DATASET_PATH

Here is a list of parameters of the script (it will be soon updated):

  -b BATCH_SIZE, --batch-size BATCH_SIZE
                        Batch size to train or validate your model
  -w WORKERS, --workers WORKERS
                        Number of workers in a dataloader
  --pretraining         Pretraining mode
  -lr LEARNING_RATE, --learning-rate LEARNING_RATE
                        Learning rate
  -d DUMP_PERIOD, --dump-period DUMP_PERIOD
                        Dump period
  -e EPOCHS, --epochs EPOCHS
                        Number of epochs to perform
  -c CHECKPOINT, --checkpoint CHECKPOINT
                        Checkpoint to load from
  --use-cuda            Use cuda for training
  --validate-on-train   Flag showing that you want to perform validation on training dataset 
                        along with the validation on the validation set

  --model MODEL         File with a model specification
  --dataset DATASET     File with a dataset sepcification
  --max-train-iterations MAX_TRAIN_ITERATIONS
                        Maximum training iterations
  --max-valid-iterations MAX_VALID_ITERATIONS
                        Maximum validation iterations
  -dp DATASET_PATH, --dataset-path DATASET_PATH
                        Path to the dataset
  -v VERBOSITY, --verbosity VERBOSITY
                        -1 for no output, 0 for epoch output, positive number
                        is printout frequency during the training
                        Prefix to the checkpoint name

Model module syntax

The syntax for the model file is the following:

class Network(torch.nn.Module):
    def __init__(self):
        super(Network, self).__init__()

    def forward(self, input):
        return [output1, output2]

    def __call__(self, input):
        return self.forward(input)

class Socket:
    def __init__(self, model):
        self.model = model

    def criterion(self, pred, target):

    def metrics(self, pred, target):

The requirements are as follows:

  • Network should have (obviously) the constructor
  • Network should have forward function which:
    • Takes as input a list of inputs.
    • Outputs a list of outputs.
  • There should be also the __call__ function specified that is a proxy for the forward function.
  • There should be a Socket class defined in order to specify how to handle the model, it should contain:
    • criterion method that takes as inputs a list of tensors with predictions and a list of tensors with targets. The output should be a number.
    • metrics method that specifies the metrics which are of the interest for your experiment. It should take as inputs a list of tensors with predictions and a list of tensors with targets and return a list of metrics which.

The reason there are lists everywhere is the following: the network may have more than one input and more than one output. We have to deal with this fact smart enough to reuse the code. Thus, the best way to do things is to pass the values of interests in lists.

Dataset module syntax

Here is a syntax for the Dataset module:

class DataSetIndex():
    def __init__(self, path):

class DataSet():
    def __init__(self, ds_index, mode='train'):
        self.ds_index = ds_index

    def __len__(self):
        if self.mode == 'test':

        elif self.mode == 'valid':


    def __getitem__(self, index):
        img = None
        target = None

        if self.mode == 'test':

        elif self.mode == 'valid':


        return [img1, img2], [target1, target2]

The dataset script should have at least the class DataSet which should have the following specified:

  • __init__, the constructor that defines all three parts of the dataset. The mode of the dataset should be defined here.
  • __len__ function that returns the length of the dataset
  • __getitem__ function that returns a list of input tensors and a list of target tensors

Although it is enough to have only the DataSet specified, it is recommended to specify also the DataSetIndex class that contains the information about the dataset's data. It is recommended to share one instance of the DataSetIndex between all the instances of the DataSet with different modes to avoid doubling or tripling the memory used to store this index and also to avoid collecting the dataset index several times.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for pytorch-scorch, version 0.0.7
Filename, size File type Python version Upload date Hashes
Filename, size pytorch_scorch-0.0.7-py3-none-any.whl (17.7 kB) File type Wheel Python version py3 Upload date Hashes View
Filename, size pytorch-scorch-0.0.7.tar.gz (18.5 kB) File type Source Python version None Upload date Hashes View

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring DigiCert DigiCert EV certificate Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page