cvmodelz

cvmodelz

"Why is it written with 'z'? Because 'cvmodels' already exists ¯\_(ツ)_/¯"

This is a package for fast and easy loading of chainer models. It also supports chainercv and chainercv2 models.

Installation

pip install cvmodelz

Built-in scripts

python -m cvmodelz.model_info cvmodelz.InceptionV3
python -m cvmodelz.model_info cvmodelz.InceptionV3 -size 427 --device 0
python -m cvmodelz.model_info cvmodelz.ResNet50 -size 448 --device 0

Basic Usage

from cvmodelz.models import ModelFactory

model = ModelFactory.new("cvmodelz.ResNet50", input_size=224)
print(model.meta)
# ModelInfo:
#   name: ResNet50
#   input_size: 224
#   feature_size: 2048
#   n_conv_maps: 2048
#   conv_map_layer: res5
#   feature_layer: pool5
#   classifier_layers:
#     - fc6
#   prepare_func: '<bound method BaseResNet.prepare of <cvmodelz.models.pretrained.resnet.ResNet50
#     object at 0x7fa037f13f70>>'
print(model.clf_layer)
# Linear(in_size=2048, out_size=1000, nobias=False)

You can also easily use the class methods ModelFactory.get_all_models and ModelFactory.get_models to restrict the options in argparse:

from cvmodelz.models import ModelFactory

print("\n".join(ModelFactory.get_all_models()))
# chainer.ResNet50Layers
# chainer.ResNet101Layers
# chainer.ResNet152Layers
# chainercv.SSD300
# chainercv.FasterRCNNVGG16
# chainercv2.resnet50
# chainercv2.resnet18
# chainercv2.inceptionv3
# chainercv2.inceptionresnetv1
# chainercv2.resnext50_32x4d
# cvmodelz.VGG16
# cvmodelz.VGG19
# cvmodelz.ResNet35
# cvmodelz.ResNet35HD
# cvmodelz.ResNet50
# cvmodelz.ResNet50HD
# cvmodelz.ResNet101
# cvmodelz.ResNet152
# cvmodelz.InceptionV3
# cvmodelz.InceptionV3HD


import argparse
parser = argparse.ArgumentParser()

parser.add_argument("--model_type", choices=ModelFactory.get_models())
# or restrict it to particular modules:
parser.add_argument("--model_type", choices=ModelFactory.get_models(["chainercv2", "cvmodelz"]))

args = parser.parse_args()

model = ModelFactory.new(args.model_type)

Models from the cvmodelz package have a method to preprocess an input so it fits the pixel values:

import numpy as np

from cvmodelz.models import ModelFactory

incv3 = ModelFactory.new("cvmodelz.InceptionV3")

array = np.random.randint(0, 255, size=(1080, 1920, 3)).astype(np.uint8)

print(incv3.prepare(array).shape)
# (3, 299, 299)
print(incv3.prepare(array, keep_ratio=True).shape)
# (3, 299, 531)
print(incv3.prepare(array, size=427, keep_ratio=True).shape)
# (3, 427, 759)

x = incv3.prepare(array, keep_ratio=True)
print(x.min(), x.max(), x.dtype)
# 0.00030062173 0.9944885 float32
print(array.min(), array.max(), array.dtype)
# 0 225 uint8

Classifiers

The model objects returned by the ModelFactory are "just" backbones that can be used in a classifier object:

from cvmodelz.models import ModelFactory
from cvmodelz.classifiers import Classifier

model = ModelFactory.new("cvmodelz.InceptionV3")

clf = Classifier(model=model)

The main ideas behind this separation are the following:

• In the forward pass, the classifier object is responsible for (1) calling the model, (2) computing the evaluation metrics, and (3) computing the loss and returning it, so it can be minimized by an updater.
• cvmodelz.classifiers.SeparateClassifier can be used instead of the default classifier class, which then creates a duplicate of the given model. By overriding the forward method, one can for example easily implement a global- and part-based classification.
• The loading of the weights is abstracted away in the different implementations but also gives the possibility to override this behavior (see cvmodelz.classifiers.base.Classifier.load method for further information).