Share your V2 projects here

Part 1 (2020)
376

I wrote a transform block for an object detection task that has only one class. Since there’s only one class, I don’t need any classification or labeling on the bounding box. I didn’t get this working out of the box with the fastai applications level API, but once I more deeply learned the transform system, I was able to figure it out. Sharing it here in case it’s useful for someone else:

from fastcore.transform import *
from fastai.data.all import *
from fastai.vision.all import *
from fastai.test_utils import synth_learner
import pandas as pd
from pathlib import Path
from nbdev import show_doc
import os
from chessocr import *

#export
class NoLabelBBoxLabeler(Transform):
    """ Bounding box labeler with no label """
    def setups(self, x): noop
    def decode (self, x, **kwargs):
        self.bbox,self.lbls = None,None
        return self._call('decodes', x, **kwargs)

    def decodes(self, x:TensorBBox):
        self.bbox = x
        return self.bbox if self.lbls is None else LabeledBBox(self.bbox, self.lbls)

#export
class BBoxTruth:
    """ get bounding box location from DataFrame """
    def __init__(self, df): self.df=df
        
    def __call__(self, o):
        size,x,y,_,_,_,_=self.df.iloc[int(o.stem)-1]
        return [[x,y, x+size, y+size]]

#export
def iou(pred, target):
    """ Vectorized Intersection Over Union calculation """
    target = Tensor.cpu(target).squeeze(1)
    pred = Tensor.cpu(pred)
    ab = np.stack([pred, target])
    intersect_area = np.maximum(ab[:, :, [2, 3]].min(axis=0) - ab[:, :, [0, 1]].max(axis=0), 0).prod(axis=1)
    union_area = ((ab[:, :, 2] - ab[:, :, 0]) * (ab[:, :, 3] - ab[:, :, 1])).sum(axis=0) - intersect_area
    return (intersect_area / union_area).mean()

#export
NoLabelBBoxBlock = TransformBlock(type_tfms=TensorBBox.create, 
                             item_tfms=[PointScaler, NoLabelBBoxLabeler])

data_url = Path.home()/".fastai/data/chess"
df = pd.read_csv(data_url/'annotations.csv', index_col=0)
block = DataBlock(
    blocks=(ImageBlock, NoLabelBBoxBlock), 
    get_items=get_image_files,
    get_y=[BBoxTruth(df)],
    n_inp=1,
    item_tfms=[Resize(224)])

dls=block.dataloaders(data_url, batch_size=64)

dls.show_batch(max_n=9, figsize=(8, 8))
9 Likes