Skip to main content

Declaratively configured pipeline for image classification

Project description

Classification training pipeline

Build status

My puny attempt to build reusable training pipeline for image classification


Idea for this project came from my first attempts to participate in Kaggle competitions. My programmers heart was painfully damaged by looking on my own code as well as on other people kernels. Code was highly repetitive, suffering from numerous reimplementations of same or almost same things through the kernels, model/experiment configuration was often mixed with models code, in other words - from programmer perspective it all looked horrible.

So I decided to extract repetitive things into framework that will work at least for me and will follow these statements:

  • experiment configurations should be cleanly separated from model definitions;
  • experiment configuration files should be easy to compare and should fully describe experiment that is being performed except for the dataset;
  • common blocks like an architecture, callbacks, storing model metrics, visualizing network predictions, should be written once and be a part of common library


pip install classification_pipeline

Note: this package requires python 3.6

Usage guide

Training a model

Let's start from a simple example of classification. Suppose, your data are structured as follows: a .cvs file with images ids and their labels and a folder with all these images. For training a neural network to classify these images all you need are few lines of python code:

import musket_core
from classification_pipeline import classification
class ProteinDataGenerator:

    def __init__(self, paths, labels):
        self.paths, self.labels = paths, labels

    def __len__(self):
        return len(self.paths)

    def __getitem__(self, idx):
        X,y = self.__load_image(self.paths[idx]),self.labels[idx]
        return PredictionItem(self.paths[idx],X, y)

    def __load_image(self, path):
        R = + '_red.png')
        G = + '_green.png')
        B = + '_blue.png')

        im = np.stack((
        ), -1)
        return im
dataset = ProteinDataGenerator(paths,labels)
cfg = classification.parse("config.yaml")

Looks simple, but there is a config.yaml file in the code, and probably it is the place where everything actually happens.

architecture: DenseNet201 #pre-trained model we are going to use
pooling: avg
augmentation: #define some minimal augmentations on images
 Fliplr: 0.5
 Flipud: 0.5
classes: 28 #define the number of classes
activation: sigmoid #as we have multilabel classification, the activation for last layer is sigmoid
weights: imagenet #we would like to start from network pretrained on imagenet dataset
shape: [224, 224, 3] #our desired input image size, everything will be resized to fit
optimizer: Adam #Adam optimizer is a good default choice
batch: 16 #our batch size will be 16
lr: 0.005
copyWeights: true
metrics: #we would like to track some metrics
  - binary_accuracy
  - macro_f1
primary_metric: val_binary_accuracy #the most interesting metric is val_binary_accuracy
primary_metric_mode: max
callbacks: #configure some minimal callbacks
    patience: 3
    monitor: val_macro_f1
    mode: max
    verbose: 1
    patience: 2
    factor: 0.3
    monitor: val_binary_accuracy
    mode: max
    cooldown: 1
    verbose: 1
loss: binary_crossentropy #we use binary_crossentropy loss
  - epochs: 10 #let's go for 100 epochs

So as you see, we have decomposed our task in two parts, code that actually trains the model and experiment configuration, which determines the model and how it should be trained from the set of predefined building blocks.

What does this code actually do behind the scenes?

  • it splits your data into 5 folds, and trains one model per fold;
  • it takes care of model checkpointing, generates example image/label tuples, collects training metrics. All this data will be stored in the folders just near your config.yaml;
  • All your folds are initialized from fixed default seed, so different experiments will use exactly the same train/validation splits.

Image Augmentations

Framework uses awesome imgaug library for augmentation, so you only need to configure your augmentation process in declarative way like in the following example:

  Fliplr: 0.5
  Flipud: 0.5
    scale: [0.8, 1.5] #random scalings
      x: [-0.2,0.2] #random shifts
      y: [-0.2,0.2]
    rotate: [-16, 16] #random rotations on -16,16 degrees
    shear: [-16, 16] #random shears on -16,16 degrees

Freezing/Unfreezing encoder

Freezing encoder is often used with transfer learning. If you want to start with frozen encoder just add

freeze_encoder: true
  - epochs: 10 #Let's go for 10 epochs with frozen encoder

  - epochs: 100 #Now let's go for 100 epochs with trainable encoder
    unfreeze_encoder: true  

in your experiments configuration, then on some stage configuration just add

unfreeze_encoder: true

to stage settings.

Note: This option is not supported for DeeplabV3 architecture.

Custom datasets

You can declare your own dataset class as in this example:

from musket_core.datasets import PredictionItem
import os
import imageio
import pandas as pd
import numpy as np
import cv2

class Classification:

    def __init__(self,imgPath):
        self.species = ['Black-grass', 'Charlock', 'Cleavers', 'Common Chickweed', 'Common wheat',
           'Fat Hen', 'Loose Silky-bent', 'Maize', 'Scentless Mayweed',
           'Shepherds Purse', 'Small-flowered Cranesbill', 'Sugar beet'] = []
        self.targets = []
        self.ids = []
        for s_id, s in enumerate(self.species):
            s_folder = os.path.join(imgPath,s)
            for file in os.listdir(s_folder):
      , file))

    def __len__(self):
        return len(

    def __getitem__(self, item):
        item_file =[item]
        target = self.targets[item]
        t = np.zeros(len(self.species))
        t[target] = 1.0
        image = self.read_image(item_file, (224,224))
        return PredictionItem(self.ids[item], image, t)

    def read_image(self, filepath, target_size=None):
        img = cv2.imread(filepath, cv2.IMREAD_COLOR)
        img = cv2.resize(img.copy(), target_size, interpolation = cv2.INTER_AREA)
        return img

Multi output classification

Sometimes you need to create network that performs several classification tasks at the same moment, in this situation you need to declare classes , activation and loss as the lists of class counts, activation functions and losses like in the following snippet:

classes: [ 4, 4 ] #define the number of classes
activation: [sigmoid,sigmoid]
  - binary_crossentropy
  - binary_crossentropy
primary_metric: val_loss #the most interesting metric is val_binary_accuracy
primary_metric_mode: min  

it is also very likely that you need to change primary metric to val_loss

Balancing your data

One common case is the situation when part of your images does not contain any objects of interest, like in Airbus ship detection challenge. More over your data may be to heavily inbalanced, so you may want to rebalance it. Alternatively you may want to inject some additional images that do not contain objects of interest to decrease amount of false positives that will be produced by the framework.

These scenarios are supported by negatives and validation_negatives settings of training stage configuration, these settings accept following values:

  • none - exclude negative examples from the data
  • real - include all negative examples
  • integer number(1 or 2 or anything), how many negative examples should be included per one positive example

if you are using this setting your dataset class must support isPositive method which returns true for indexes which contain positive examples:

    def isPositive(self, item):
        for mask in pixels:
            if isinstance(mask, str):
                return True;
        return False

Multistage training

Sometimes you need to split your training into several stages. You can easily do it by adding several stage entries in your experiment configuration file like in the following example:

  - epochs: 6 #Train for 6 epochs
    negatives: none #do not include negative examples in your training set 
    validation_negatives: real #validation should contain all negative examples    

  - lr: 0.0001 #let's use different starting learning rate
    epochs: 6
    negatives: real
    validation_negatives: real

  - loss: lovasz_loss #let's override loss function
    lr: 0.00001
    epochs: 6
    initial_weights: ./fpn-resnext2/weights/best-0.1.weights #let's load weights from this file    

Stage entries allow you to configure custom learning rate, balance of negative examples, callbacks, loss function and even initial weights which should be used on a particular stage.

Composite losses

Framework supports composing loss as a weighted sum of predefined loss functions. For example, following construction

loss: binary_crossentropy+0.1*dice_loss

will result in loss function which is composed from binary_crossentropy and dice_loss functions.

Cyclical learning rates


As told in Cyclical learning rates for training neural networks CLR policies can provide quicker converge for some neural network tasks and architectures.


We support them by adopting Brad Kenstler CLR callback for Keras.

If you want to use them, just add CyclicLR in your experiment configuration file as shown below:

    patience: 40
    monitor: val_binary_accuracy
    verbose: 1
     base_lr: 0.0001
     max_lr: 0.01
     mode: triangular2
     step_size: 300

LR Finder

Estimating optimal learning rate for your model is an important thing, we support this by using slightly changed version of Pavel Surmenok - Keras LR Finder

cfg = classification.parse(config.yaml)
ds = SimplePNGMaskDataSet("./train","./train_mask") - ???????????????????
finder.plot_loss(n_skip_beginning=20, n_skip_end=5)
finder.plot_loss_change(sma=20, n_skip_beginning=20, n_skip_end=5, y_lim=(-0.01, 0.01))

will result in this couple of helpful images:



Training on crops

Your images can be too large to train model on them. In this case you probably want to train model on crops. All that you need to do is to specify number of splits per axis. For example, following lines in config

shape: [768, 768, 3]
crops: 3

will lead to splitting each image into 9 cells (3 horizontal splits and 3 vertical splits) and training model on these splits. Augmentations will be run separately on each cell.

During prediction time, your images will be split into these cells, prediction will be executed on each cell, and then results will be assembled in single final mask. Thus the whole process of cropping will be invisible from a consumer perspective.

Using trained model

Okey, our model is trained, now we need to actually do image classification. Let's say, we need to run image classification on images in the directory and store results in csv file:

predictions = []
images = []

#Now let's use best model from fold 0 to do image segmentation on images from images_to_segment
preds = cfg.predict_all_to_array(dataset_test, 0, 0)
for i, item in enumerate(dataset_test):
    p = np.argmax(preds[i])

#Let's store results in csv
df = pd.DataFrame.from_dict({'file': images, 'species': predictions})
df.to_csv('submission.csv', index=False)

Ensembling predictions

And what if you want to ensemble models from several folds? Just pass a list of fold numbers to predict_all_to_array like in the following example:

 cfg.predict_all_to_array(dataset_test, [0,1,2,3,4], 0)

Another supported option is to ensemble results from extra test time augmentation (flips) by adding keyword arg ttflips=True.

Custom evaluation code

Sometimes you need to run custom evaluation code. In such case you may use: evaluateAll method, which provides an iterator on the batches containing original images, training masks and predicted masks

for batch in cfg.evaluateAll(ds,2):
    for i in range(len(batch.predicted_maps_aug)):
        masks = ds.get_masks([i])
        for d in range(1,20):
            cur_seg = binary_opening(batch.predicted_maps_aug[i].arr > d/20, np.expand_dims(disk(2), -1))
            cm = rle.masks_as_images(rle.multi_rle_encode(cur_seg))
            pr = f2(masks, cm);

Accessing model

You may get trained keras model by calling: cfg.load_model(fold, stage).

Analyzing experiments results

Okey, we have done a lot of experiments and now we need to compare the results and understand what works better. This repository contains script which may be used to analyze folder containing sub folders with experiment configurations and results. This script gathers all configurations, diffs them by doing structural diff, then for each configuration it averages metrics for all folds and generates csv file containing metrics and parameters that was actually changed in your experiment like in the following example

This script accepts following arguments:

  • inputFolder - root folder to search for experiments configurations and results
  • output - file to store aggregated metrics
  • onlyMetric - if you specify this option all other metrics will not be written in the report file
  • sortBy - metric that should be used to sort results


python --inputFolder ./experiments --output ./

What is supported?

At this moment classification pipeline supports following pre-trained models:

Each architecture also supports some specific options, list of options is documented in segmentation RAML library.

Supported augmentations are documented in augmentation RAML library.

Callbacks are documented in callbacks RAML library.

Custom architectures, callbacks, metrics

Classification pipeline uses keras custom objects registry to find entities, so if you need to use custom loss function, activation or metric all that you need to do is to register it in Keras as:

keras.utils.get_custom_objects()["my_loss"]= my_loss

If you want to inject new architecture, you should register it in classification.custom_models dictionary.

For example:


where MyUnet is a function that accepts architecture parameters as arguments and returns an instance of keras model.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for classification-pipeline, version 0.432
Filename, size File type Python version Upload date Hashes
Filename, size classification_pipeline-0.432-py2.py3-none-any.whl (11.7 kB) File type Wheel Python version py2.py3 Upload date Hashes View
Filename, size classification_pipeline-0.432.tar.gz (12.2 kB) File type Source Python version None Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page