fastai-amalgam 0.0.26

A collection of various CV focused utilities for the fastai library. Includes interpretation, data augmentation, gradcam, loss-functions, model conversion utils, etc

FastAI2 Extensions

This library is a collection of utility functions for a variety of purposes that fit right into the fastai2 ecosystem. It's broadly divided into 3 modules -- interpret , augment , and inference .

Install

pip install fastai2_extensions

Interpretation

ClassificationInterpretationEx

Extends fastai's ClassificationInterpretation to plot model confidence and per-label accuracy bar graphs. It also adds some convenience to grab filenames based on these confidence levels.

This part of the library is currently suitable for Softmax classifiers only. Multilabel support will be added soon.

from fastai2.vision.all import *
from fastai2_extensions.interpret.all import *

learn = load_learner('/Users/rahulsomani/Desktop/shot-lighting-cast/fastai2-110-epoch-model.pkl')

interp = ClassificationInterpretationEx.from_learner(learn)

plt.style.use('ggplot')
interp.plot_accuracy()

interp.plot_label_confidence()

GradCam

The GradCam object takes in 3 args:

• learn: a fastai Learner
• fname: path to the image file to draw the heatcam over
• labels: list of labels to draw the heatmap for. If None, draws for the highest predicted class

There's quite a few plotting options. For more options, see the docs.

import PIL
fname = '../assets/imgs/alice-in-wonderland.jpg'
PIL.Image.open(fname).resize((550,270))

gcam = GradCam(learn, fname, None)
gcam.plot(full_size=True, plot_original=True, figsize=(12,6))

gcam = GradCam(learn, fname, ['shot_lighting_cast_hard', 'shot_lighting_cast_soft'])
gcam.plot(full_size=False, plot_original=False, figsize=(12,4))

Comparing Multiple Models

compare_venn lets you compares 2 or models trained evaluated on the same dataset to inspect model agreement. If you only input 2 or 3 models, then you can also see Venn Diagrams for the same.

For simplicity, I'm using the same model here with smaller versions of the validation set to display this functionality.

interp1 = ClassificationInterpretationEx.from_learner(learn1)
interp2 = ClassificationInterpretationEx.from_learner(learn2)
interp3 = ClassificationInterpretationEx.from_learner(learn3)
interp1.compute_label_confidence()
interp2.compute_label_confidence()
interp3.compute_label_confidence()

%%capture
fig,common_labels = compare_venn(
    conf_level=(0,99),  interps=[interp1,interp2],
    mode='accurate',
    return_common=True, return_fig=True,
    set_color='tomato'
)

fig

%%capture
fig,common_labels = compare_venn(
    conf_level=(0,99),  interps=[interp1,interp2,interp3],
    mode='accurate',
    return_common=True, return_fig=True,
    set_color='tomato'
)

fig

Augmentation

ApplyPILFilter, not surprisingly, lets you apply one or more PIL.ImageFilters as a data augmentation.

There's also a convenience function read_lut which lets you read in a LUT file (commonly found with .cube extensions), and construct a PIL.ImageFilter.Color3dLUT to apply as a transform.

The idea place for this in a fastai2 pipeline is as an item_tfms as it's a lossless transform and can be done right after reading the image from disk. A full example is shown in the docs.

from fastai2_extensions.augment.pil_filters import *

lut   = read_lut('../assets/luts/2strip.cube')
fname = '../assets/imgs/office-standoff.png'

img_raw  = PILImage.create(fname)
img_filt = ApplyPILFilter(lut,p=1.0)(fname, split_idx=0)

%%capture
fig,ax = plt.subplots(nrows=1, ncols=2, figsize=(16,6))
show_tensor = lambda x,ax: ToTensor()(x).show(ctx=ax)

show_tensor(img_raw,ax[0])
show_tensor(img_filt,ax[1])

ax[0].set_title('Original')
ax[1].set_title('LUT Transformed')

fig

Export

Convenience wrappers to export to ONNX.
Other frameworks will be added soon.

ONNX

#hide_output
from fastai2_extensions.inference.export import *

torch_to_onnx(learn.model,
              activation   = nn.Softmax(-1),
              save_path    = Path.home()/'Desktop',
              model_fname  = 'onnx-model',
              input_shape  = (1,3,224,224),
              input_name   = 'input_image',
              output_names = 'output')

Loading, polishing, and optimising exported model from /Users/rahulsomani/Desktop/onnx-model.onnx
Exported successfully

path_onnx_model = '/Users/rahulsomani/Desktop/onnx-model.onnx'
fname = '../assets/imgs/odyssey-ape.png'

from onnxruntime import InferenceSession

session = InferenceSession(path_onnx_model)
x = {session.get_inputs()[0].name:
     torch_to_numpy(preprocess_one(fname))} # preprocessing - varies based on your training
session.run(None, x)

[array([[0.6942669 , 0.30573303]], dtype=float32)]