TfmsManager: Tune Your Transforms

What is that

TfmsManager is an easy to use tool that helps you tune and debug complex groups and chains of transformations.
It comes with a two use cases for Image and Audio transformation.

IMPORTANT: you can preview your transformation before creating a “data bunch”, having only some samples (Image or Audio data).

Features

  • Preview: see the extent of modifications made by your transform.
  • Diff: visually compare your original data with transformed one.
  • tfms: the “TFMS” tab shows you the exact list of transformation applied on each step.
  • Shape: see the shape of resulting tensor.

Integration with fast.ai

Integrating the TfmsManager into fast.ai DataBlock is very easy: you need only to call tm.get_tfms() and pass the resulting tfms array to your transform step (where tm is a TfmsManager).

   .transform(tm.get_tfms())
# Simple integration with DataBlock api
data = (AudioList.from_df(src, path, cols=['SampleAndSr'])
        .split_from_df('is_valid')
        .label_from_df('WRD', classes=classes)
        .transform(tm.get_tfms())
        .databunch(bs=64)) # NOTE: 

Image

The following code shows you how to preview all the tansformations you’ve choosen, having instant feedback from the new parateters without

#Show all results
tfms = get_transforms(
    do_flip = False, 
    flip_vert = False,
    max_rotate = 1,
    max_zoom = 1.0  
)
tm=ImageTfmsManager.get_from_default_tfms(tfms)
tm.try_train_tfms(image,'Sample with single RGB',repeats=4);

Audio

The following (advanced) example shows you how to preview a custom multi-spectrogram transform.

tm=AudioTfmsManager.get_audio_tfms_manager(
                         spectro=True, n_mels=256, ws=300, n_fft=3200, to_db_scale=True,
                         white_noise=True, noise_scl=.0052,
                         modulate_volume=True,
                         random_cutout=True,
                         pad_with_silence=False,
                         pitch_warp=False,
                         down_and_up=False,
                         mx_to_pad=32000)

#Replace the spectrogram transform with tfm_spectro_stft
filan_spec_tfm = partial(tfm_multiSpectrumSlide, max_duration=1000, n_windows=8)
tm.train_tfmsg[-1]=[filan_spec_tfm]
tm.valid_tfmsg[-1]=[filan_spec_tfm]

#Try!
yOut = tm.try_train_tfms(sampleSound,'test')
yOut.shape

Source code and complete examples

AUDIO: https://github.com/zcaceres/fastai-audio/blob/master/AudioTransformsManager.ipynb
IMAGE: https://github.com/zcaceres/fastai-audio/blob/master/ImageTransformsManager.ipynb
MULTI SPEC.: https://github.com/zcaceres/fastai-audio/blob/master/example_multi_spectrogram_classificaiton_phoneme-sliding.ipynb

22 Likes

Does it group all affine transforms in order to avoid unnecessary blur?

Sorry @Kaspar what you mean for “afgivne”?

It seems to be quite an useful tool, thanks for sharing, @ste! :wink:

1 Like

This looks awesome man, can’t wait to try it out tomorrow! You’re doing a really great job with the audio stuff lately, keep up the good work :slight_smile:

2 Likes

Not worked in FastAi ver 2?