Cospel

# https://github.com/neel-dey/tf2-keras-gcnn
# https://github.com/neel-dey/tf2-GrouPy
import numpy as np
import tensorflow as tf

from tensorflow.keras import Model
from tensorflow.keras.layers import Input, Activation, MaxPooling2D

from keras_gcnn.layers import GConv2D, GBatchNorm, GroupPool

Cospel

import pprint

from ximilar.client import DetectionClient


client = DetectionClient("__API_AUTH_TOKEN__")
results = client.detect_on_task([{"_file": "test/IMG_20181228_102636.jpg"}], "__TASK_ID__")

pprint.pprint(results)

Cospel

import os
import csv

from ximilar.client import DetectionClient


client = DetectionClient("__YOUR_API_TOKEN__")

# mapping of your ids
labels = {

Cospel

class BatchRenormCallback(tf.keras.callbacks.Callback):
    def __init__(self, log_dir, bvalues):
        super().__init__()
        self.writer = tf.summary.create_file_writer(log_dir)
        self.bvalues = bvalues

    def on_epoch_begin(self, epoch, logs=None):
        if epoch in self.bvalues:
            self.change_renorm(epoch)

Cospel

# cuda10, cuddn7.5, bazel 2.0, gcc7 through devtoolset-7, our custom tf2 build

0. # scl enable devtoolset-7 -- bash # for gcc on centos

1.
edit files build_deps/toolchains/* change 10.1 to 10.0

edit build_deps/toolchains/gcc7_manylinux2010-nvcc-cuda10.1/BUILD
-        "/usr/local/cuda-10.1/targets/x86_64-linux/include",
-        "/usr/local/cuda-10.1/include",

Cospel

import tensorflow as tf

base_model = tf.keras.applications.MobileNetV2(
    weights="imagenet", input_shape=self.shape, include_top=False
)

# Freeze the base_model
base_model.trainable = False

# Set the base model training=False so batch statistics is not updated

Cospel

@tf.function
def cutout_channel(image, probability=0.5, iterations=3):
    image = tf.cast(image, tf.float32)
    r, g, b = tf.split(image, 3, axis=2)

    # for cc in tf.range(iterations):
    r = random_function(r, cutout_base, probability, None, **{"sl": 0.02, "sh": 0.08, "r1": 0.3})
    g = random_function(g, cutout_base, probability, None, **{"sl": 0.02, "sh": 0.08, "r1": 0.3})
    b = random_function(b, cutout_base, probability, None, **{"sl": 0.02, "sh": 0.08, "r1": 0.3})

Cospel

@tf.function
def preprocess_caffe(x):
    """
    Preprocessing for Keras (VGG, ResnetV1).
    ! This works only for channels_last
    :param x: np.asarray([image, image, ...], dtype="float32") in RGB
    :return: normalized image vgg style (BGR)
    """
    batch, height, width, channels = x.shape
    x = tf.cast(x, tf.float32)

Cospel

tf.keras.models.save_model(model, "model")
model_sm = tf.keras.models.load_model("model")

tf.keras.models.save_model(model, "model.h5")
model_h5 = tf.keras.models.load_model("model.h5", custom_objects={"PreprocessTFLayer": PreprocessTFLayer})

Cospel

import tensorflow as tf

class BatchNormalization(tf.keras.layers.BatchNormalization):
    """
    Replace BatchNormalization layers with this new layer.
    This layer has fixed momentum 0.9.
    """
    def __init__(self, momentum=0.9, name=None, **kwargs):
        super(BatchNormalization, self).__init__(momentum=0.9, name=name, **kwargs)