torch-snippets 0.321

One line functions for common tasks

Pytorch utilities for simple pytorch needs

Logging and plotting metrics

from torch_snippets import Report

n_epochs = 5
log = Report(n_epochs)

for epoch in range(n_epochs):
    # No need to --> train_epoch_losses, train_epoch_accuracies = [], []
    N = len(trn_dl)
    for ix, batch in enumerate(iter(trn_dl)):
        ...
        pos = (epoch + (ix+1)/N) # a float between 0 - n_epochs
        # give any number of kwargs that need to be reported and stored.
        # args should be float
        log.record(pos=pos, train_acc=np.mean(is_correct), train_loss=batch_loss, end='\r') # impersistent log

    N = len(val_dl)
    for ix, batch in enumerate(iter(val_dl)):
        ...
        pos = (epoch + (ix+1)/N) # a float between 0 - n_epochs
        log.record(pos=pos, val_loss=batch_loss, end='\r') # impersistent log
    log.report_avgs(epoch+1) # persist the report

• Auto calculates time remaining
• No need to preinitialize empty lists
• Automatically stores metrics as collection of key words
• Persistent vs Transient logging (use end='\r')

>>> print(log.logged) # get logged metric names
# ['train_loss', 'train_acc', 'val_loss', 'val_acc']
>>> log.plot() # plot all the logged metrics

• Auto calculates average of all metrics in an epoch

• Plot entire training history with one command
  

• selectively plot logged metrics using regex

>>> log.plot('*_loss')
# or you can list out the metrics that need plotting
# >>> log.plot(['train_acc','val_acc'])

Auxiliary Functions

There are simple functions that are overloaded to take inputs and perform repetitive tasks that usually take a few lines to write

Images

show, inspect, Glob, read

FilePaths

stem, Glob, parent, name

Dump and load python objects

loaddill,dumpdill

Misc

Tqdm, Timer, randint, unique, diff

Pytorch Modules

Reshape and Permute (nn.Modules)

and many more...

Install

pip install torch_snippets

Usage

import pytorch_snippets
dir(pytorch_snippets)