hsahovic

def tf_ds_cutmix(ds, shuffling=1024):
    ds_shuffled = ds.shuffle(shuffling)
    
    def cutmix(p1, p2):
        img_1, label_1 = p1
        img_2, label_2 = p2
        
        
        lambda_ = tf.random.uniform((1,))
        

hsahovic

import tensorflow as tf
from tensorflow.keras.losses import categorical_crossentropy

@tf.function
def ohem_crossentropy_loss(y_true, y_pred):
    # You can apply OHEM with any loss you want
    # To do so, you just need to change the base loss below
    cross_entropy = categorical_crossentropy(y_true, y_pred)
    
    # You can tune how many examples are rejected by modifying the `80` value below

hsahovic

let img = await webcam.capture();

let rChannel = img.slice([0,0,0],[192,192,1]).reshape([192,192]);
let gChannel = img.slice([0,0,1],[192,192,1]).reshape([192,192]);
let bChannel = img.slice([0,0,2],[192,192,1]).reshape([192,192]);

img = tf.stack([bChannel, gChannel, rChannel], 2);

img = tf.mul(img, 1/255);
img = tf.add(img, -.5);

hsahovic

from tensorflow.keras.models import Sequential
import tensorflow.keras.layers as keras_layers


clone = Sequential()
# Here, layer[1] is a model. 
layers = [model.layers[0]] + model.layers[1].layers + model.layers[2:]

for layer in layers:
    if str(type(layer)).endswith("InputLayer'>"):

hsahovic

import os
from tqdm.notebook import tqdm
from shutil import copyfile

def split_dir(source_dir, target_dir_train, target_dir_test, ratio=.2):
    files = os.listdir(source_dir)
    to_test = int(len(files) * ratio + .5)
    
    assert files[:to_test] + files[to_test:] == files
    for file in files[:to_test]:

hsahovic

from typing import Generator
import tensorflow as tf

def image_generator_to_tf_ds(generator: Generator) -> tf.data.Dataset:
    """Converts a initialized keras Image Data Generator to an equivalent tf Dataset.
    
    Example usage:

        >>> img_generator = ImageDataGenerator()
        >>> img_generator = img_generator.flow_from_directory(