1641926220

End-to-End Object Detection with Transformers

DE⫶TR: End-to-End Object Detection with Transformers

PyTorch training code and pretrained models for DETR (DEtection TRansformer). We replace the full complex hand-crafted object detection pipeline with a Transformer, and match Faster R-CNN with a ResNet-50, obtaining 42 AP on COCO using half the computation power (FLOPs) and the same number of parameters. Inference in 50 lines of PyTorch.

What it is. Unlike traditional computer vision techniques, DETR approaches object detection as a direct set prediction problem. It consists of a set-based global loss, which forces unique predictions via bipartite matching, and a Transformer encoder-decoder architecture. Given a fixed small set of learned object queries, DETR reasons about the relations of the objects and the global image context to directly output the final set of predictions in parallel. Due to this parallel nature, DETR is very fast and efficient.

About the code. We believe that object detection should not be more difficult than classification, and should not require complex libraries for training and inference. DETR is very simple to implement and experiment with, and we provide a standalone Colab Notebook showing how to do inference with DETR in only a few lines of PyTorch code. Training code follows this idea - it is not a library, but simply a main.py importing model and criterion definitions with standard training loops.

Additionnally, we provide a Detectron2 wrapper in the d2/ folder. See the readme there for more information.

For details see End-to-End Object Detection with Transformers by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko.

Model Zoo

We provide baseline DETR and DETR-DC5 models, and plan to include more in future. AP is computed on COCO 2017 val5k, and inference time is over the first 100 val5k COCO images, with torchscript transformer.

	name	backbone	schedule	inf_time	box AP	url	size
0	DETR	R50	500	0.036	42.0	model \| logs	159Mb
1	DETR-DC5	R50	500	0.083	43.3	model \| logs	159Mb
2	DETR	R101	500	0.050	43.5	model \| logs	232Mb
3	DETR-DC5	R101	500	0.097	44.9	model \| logs	232Mb

COCO val5k evaluation results can be found in this gist.

The models are also available via torch hub, to load DETR R50 with pretrained weights simply do:

model = torch.hub.load('facebookresearch/detr', 'detr_resnet50', pretrained=True)

COCO panoptic val5k models:

	name	backbone	box AP	segm AP	PQ	url	size
0	DETR	R50	38.8	31.1	43.4	download	165Mb
1	DETR-DC5	R50	40.2	31.9	44.6	download	165Mb
2	DETR	R101	40.1	33	45.1	download	237Mb

Checkout our panoptic colab to see how to use and visualize DETR's panoptic segmentation prediction.

Notebooks

We provide a few notebooks in colab to help you get a grasp on DETR:

DETR's hands on Colab Notebook: Shows how to load a model from hub, generate predictions, then visualize the attention of the model (similar to the figures of the paper)
Standalone Colab Notebook: In this notebook, we demonstrate how to implement a simplified version of DETR from the grounds up in 50 lines of Python, then visualize the predictions. It is a good starting point if you want to gain better understanding the architecture and poke around before diving in the codebase.
Panoptic Colab Notebook: Demonstrates how to use DETR for panoptic segmentation and plot the predictions.

Usage - Object detection

There are no extra compiled components in DETR and package dependencies are minimal, so the code is very simple to use. We provide instructions how to install dependencies via conda. First, clone the repository locally:

git clone https://github.com/facebookresearch/detr.git

Then, install PyTorch 1.5+ and torchvision 0.6+:

conda install -c pytorch pytorch torchvision

Install pycocotools (for evaluation on COCO) and scipy (for training):

conda install cython scipy
pip install -U 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

That's it, should be good to train and evaluate detection models.

(optional) to work with panoptic install panopticapi:

pip install git+https://github.com/cocodataset/panopticapi.git

Data preparation

Download and extract COCO 2017 train and val images with annotations from http://cocodataset.org. We expect the directory structure to be the following:

path/to/coco/
  annotations/  # annotation json files
  train2017/    # train images
  val2017/      # val images

Training

To train baseline DETR on a single node with 8 gpus for 300 epochs run:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --coco_path /path/to/coco

A single epoch takes 28 minutes, so 300 epoch training takes around 6 days on a single machine with 8 V100 cards. To ease reproduction of our results we provide results and training logs for 150 epoch schedule (3 days on a single machine), achieving 39.5/60.3 AP/AP50.

We train DETR with AdamW setting learning rate in the transformer to 1e-4 and 1e-5 in the backbone. Horizontal flips, scales and crops are used for augmentation. Images are rescaled to have min size 800 and max size 1333. The transformer is trained with dropout of 0.1, and the whole model is trained with grad clip of 0.1.

Evaluation

To evaluate DETR R50 on COCO val5k with a single GPU run:

python main.py --batch_size 2 --no_aux_loss --eval --resume https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth --coco_path /path/to/coco

We provide results for all DETR detection models in this gist. Note that numbers vary depending on batch size (number of images) per GPU. Non-DC5 models were trained with batch size 2, and DC5 with 1, so DC5 models show a significant drop in AP if evaluated with more than 1 image per GPU.

Multinode training

Distributed training is available via Slurm and submitit:

pip install submitit

Train baseline DETR-6-6 model on 4 nodes for 300 epochs:

python run_with_submitit.py --timeout 3000 --coco_path /path/to/coco

Usage - Segmentation

We show that it is relatively straightforward to extend DETR to predict segmentation masks. We mainly demonstrate strong panoptic segmentation results.

Data preparation

For panoptic segmentation, you need the panoptic annotations additionally to the coco dataset (see above for the coco dataset). You need to download and extract the annotations. We expect the directory structure to be the following:

path/to/coco_panoptic/
  annotations/  # annotation json files
  panoptic_train2017/    # train panoptic annotations
  panoptic_val2017/      # val panoptic annotations

Training

We recommend training segmentation in two stages: first train DETR to detect all the boxes, and then train the segmentation head. For panoptic segmentation, DETR must learn to detect boxes for both stuff and things classes. You can train it on a single node with 8 gpus for 300 epochs with:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --coco_path /path/to/coco  --coco_panoptic_path /path/to/coco_panoptic --dataset_file coco_panoptic --output_dir /output/path/box_model

For instance segmentation, you can simply train a normal box model (or used a pre-trained one we provide).

Once you have a box model checkpoint, you need to freeze it, and train the segmentation head in isolation. For panoptic segmentation you can train on a single node with 8 gpus for 25 epochs:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --masks --epochs 25 --lr_drop 15 --coco_path /path/to/coco  --coco_panoptic_path /path/to/coco_panoptic  --dataset_file coco_panoptic --frozen_weights /output/path/box_model/checkpoint.pth --output_dir /output/path/segm_model

For instance segmentation only, simply remove the dataset_file and coco_panoptic_path arguments from the above command line.

Contributing

We actively welcome your pull requests! Please see CONTRIBUTING.md and CODE_OF_CONDUCT.md for more info.

Author: Facebookresearch
Source Code: https://github.com/facebookresearch/detr
License: Apache-2.0 License

#python #dockerfile #pytorch

What is GEEK

Buddha Community

End-to-End Object Detection with Transformers

Sasha Lee

1650636000

Dl4clj: Clojure Wrapper for Deeplearning4j.

dl4clj

Port of deeplearning4j to clojure

Contact info

If you have any questions,

my email is will@yetanalytics.com
I'm will_hoyt in the clojurians slack
twitter is @FeLungz (don't check very often)

TODO

update examples dir
finish README
- add in examples using Transfer Learning
finish tests
- eval is missing regression tests, roc tests
- nn-test is missing regression tests
- spark tests need to be redone
- need dl4clj.core tests
revist spark for updates
write specs for user facing functions
- this is very important, match isnt strict for maps
- provides 100% certianty of the input -> output flow
- check the args as they come in, dispatch once I know its safe, test the pure output
collapse overlapping api namespaces
add to core use case flows

Features

Stable Features with tests

Neural Networks DSL
Early Stopping Training
Transfer Learning
Evaluation
Data import

Features being worked on for 0.1.0

Clustering (testing in progress)
Spark (currently being refactored)
Front End (maybe current release, maybe future release. Not sure yet)
Version of dl4j is 0.0.8 in this project. Current dl4j version is 0.0.9
Parallelism
Kafka support
Other items mentioned in TODO

Features being worked on for future releases

NLP
Computational Graphs
Reinforement Learning
Arbiter

Artifacts

NOT YET RELEASED TO CLOJARS

fork or clone to try it out

If using Maven add the following repository definition to your pom.xml:

<repository>
  <id>clojars.org</id>
  <url>http://clojars.org/repo</url>
</repository>

Latest release

With Leiningen:

n/a

With Maven:

n/a

<dependency>
  <groupId>_</groupId>
  <artifactId>_</artifactId>
  <version>_</version>
</dependency>

Usage

Things you need to know

All functions for creating dl4j objects return code by default

All of these functions have an option to return the dl4j object
- :as-code? = false
This because all builders require the code representation of dl4j objects
- this requirement is not going to change
INDarray creation fns default to objects, this is for convenience
- :as-code? is still respected

API functions return code when all args are provided as code

API functions return the value of calling the wrapped method when args are provided as a mixture of objects and code or just objects

The tests are there to help clarify behavior, if you are unsure of how to use a fn, search the tests

for questions about spark, refer to the spark section bellow

Example of obj/code duality

(ns my.ns
  (:require [dl4clj.nn.conf.builders.layers :as l]))

;; as code (the default)

(l/dense-layer-builder
 :activation-fn :relu
 :learning-rate 0.006
 :weight-init :xavier
 :layer-name "example layer"
 :n-in 10
 :n-out 1)

;; =>

(doto
 (org.deeplearning4j.nn.conf.layers.DenseLayer$Builder.)
 (.nOut 1)
 (.activation (dl4clj.constants/value-of {:activation-fn :relu}))
 (.weightInit (dl4clj.constants/value-of {:weight-init :xavier}))
 (.nIn 10)
 (.name "example layer")
 (.learningRate 0.006))

;; as an object

(l/dense-layer-builder
 :activation-fn :relu
 :learning-rate 0.006
 :weight-init :xavier
 :layer-name "example layer"
 :n-in 10
 :n-out 1
 :as-code? false)

;; =>

#object[org.deeplearning4j.nn.conf.layers.DenseLayer 0x69d7d160 "DenseLayer(super=FeedForwardLayer(super=Layer(layerName=example layer, activationFn=relu, weightInit=XAVIER, biasInit=NaN, dist=null, learningRate=0.006, biasLearningRate=NaN, learningRateSchedule=null, momentum=NaN, momentumSchedule=null, l1=NaN, l2=NaN, l1Bias=NaN, l2Bias=NaN, dropOut=NaN, updater=null, rho=NaN, epsilon=NaN, rmsDecay=NaN, adamMeanDecay=NaN, adamVarDecay=NaN, gradientNormalization=null, gradientNormalizationThreshold=NaN), nIn=10, nOut=1))"]

General usage examples

Importing data

Loading data from a file (here its a csv)


(ns my.ns
 (:require [dl4clj.datasets.input-splits :as s]
           [dl4clj.datasets.record-readers :as rr]
           [dl4clj.datasets.api.record-readers :refer :all]
           [dl4clj.datasets.iterators :as ds-iter]
           [dl4clj.datasets.api.iterators :refer :all]
           [dl4clj.helpers :refer [data-from-iter]]))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; file splits (convert the data to records)
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def poker-path "resources/poker-hand-training.csv")
;; this is not a complete dataset, it is just here to sever as an example

(def file-split (s/new-filesplit :path poker-path))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; record readers, (read the records created by the file split)
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def csv-rr (initialize-rr! :rr (rr/new-csv-record-reader :skip-n-lines 0 :delimiter ",")
                                 :input-split file-split))

;; lets look at some data
(println (next-record! :rr csv-rr :as-code? false))
;; => #object[java.util.ArrayList 0x2473e02d [1, 10, 1, 11, 1, 13, 1, 12, 1, 1, 9]]
;; this is our first line from the csv


;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; record readers dataset iterators (turn our writables into a dataset)
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def rr-ds-iter (ds-iter/new-record-reader-dataset-iterator
                 :record-reader csv-rr
                 :batch-size 1
                 :label-idx 10
                 :n-possible-labels 10))

;; we use our record reader created above
;; we want to see one example per dataset obj returned (:batch-size = 1)
;; we know our label is at the last index, so :label-idx = 10
;; there are 10 possible types of poker hands so :n-possible-labels = 10
;; you can also set :label-idx to -1 to use the last index no matter the size of the seq

(def other-rr-ds-iter (ds-iter/new-record-reader-dataset-iterator
                       :record-reader csv-rr
                       :batch-size 1
                       :label-idx -1
                       :n-possible-labels 10))

(str (next-example! :iter rr-ds-iter :as-code? false))
;; =>
;;===========INPUT===================
;;[1.00, 10.00, 1.00, 11.00, 1.00, 13.00, 1.00, 12.00, 1.00, 1.00]
;;=================OUTPUT==================
;;[0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 1.00]


;; and to show that :label-idx = -1 gives us the same output

(= (next-example! :iter rr-ds-iter :as-code? false)
   (next-example! :iter other-rr-ds-iter :as-code? false)) ;; => true

INDArrays and Datasets from clojure data structures


(ns my.ns
  (:require [nd4clj.linalg.factory.nd4j :refer [vec->indarray matrix->indarray
                                                indarray-of-zeros indarray-of-ones
                                                indarray-of-rand vec-or-matrix->indarray]]
            [dl4clj.datasets.new-datasets :refer [new-ds]]
            [dl4clj.datasets.api.datasets :refer [as-list]]
            [dl4clj.datasets.iterators :refer [new-existing-dataset-iterator]]
            [dl4clj.datasets.api.iterators :refer :all]
            [dl4clj.datasets.pre-processors :as ds-pp]
            [dl4clj.datasets.api.pre-processors :refer :all]
            [dl4clj.core :as c]))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; INDArray creation
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;;TODO: consider defaulting to code

;; can create from a vector

(vec->indarray [1 2 3 4])
;; => #object[org.nd4j.linalg.cpu.nativecpu.NDArray 0x269df212 [1.00, 2.00, 3.00, 4.00]]

;; or from a matrix

(matrix->indarray [[1 2 3 4] [2 4 6 8]])
;; => #object[org.nd4j.linalg.cpu.nativecpu.NDArray 0x20aa7fe1
;; [[1.00, 2.00, 3.00, 4.00], [2.00, 4.00, 6.00, 8.00]]]


;; will fill in spareness with zeros

(matrix->indarray [[1 2 3 4] [2 4 6 8] [10 12]])
;; => #object[org.nd4j.linalg.cpu.nativecpu.NDArray 0x8b7796c
;;[[1.00, 2.00, 3.00, 4.00],
;; [2.00, 4.00, 6.00, 8.00],
;; [10.00, 12.00, 0.00, 0.00]]]

;; can create an indarray of all zeros with specified shape
;; defaults to :rows = 1 :columns = 1

(indarray-of-zeros :rows 3 :columns 2)
;; => #object[org.nd4j.linalg.cpu.nativecpu.NDArray 0x6f586a7e
;;[[0.00, 0.00],
;; [0.00, 0.00],
;; [0.00, 0.00]]]

(indarray-of-zeros) ;; => #object[org.nd4j.linalg.cpu.nativecpu.NDArray 0xe59ffec 0.00]

;; and if only one is supplied, will get a vector of specified length

(indarray-of-zeros :rows 2)
;; => #object[org.nd4j.linalg.cpu.nativecpu.NDArray 0x2899d974 [0.00, 0.00]]

(indarray-of-zeros :columns 2)
;; => #object[org.nd4j.linalg.cpu.nativecpu.NDArray 0xa5b9782 [0.00, 0.00]]

;; same considerations/defaults for indarray-of-ones and indarray-of-rand

(indarray-of-ones :rows 2 :columns 3)
;; => #object[org.nd4j.linalg.cpu.nativecpu.NDArray 0x54f08662 [[1.00, 1.00, 1.00], [1.00, 1.00, 1.00]]]

(indarray-of-rand :rows 2 :columns 3)
;; all values are greater than 0 but less than 1
;; => #object[org.nd4j.linalg.cpu.nativecpu.NDArray 0x2f20293b [[0.85, 0.86, 0.13], [0.94, 0.04, 0.36]]]



;; vec-or-matrix->indarray is built into all functions which require INDArrays
;; so that you can use clojure data structures
;; but you still have the option of passing existing INDArrays

(def example-array (vec-or-matrix->indarray [1 2 3 4]))
;; => #object[org.nd4j.linalg.cpu.nativecpu.NDArray 0x5c44c71f [1.00, 2.00, 3.00, 4.00]]

(vec-or-matrix->indarray example-array)
;; => #object[org.nd4j.linalg.cpu.nativecpu.NDArray 0x607b03b0 [1.00, 2.00, 3.00, 4.00]]

(vec-or-matrix->indarray (indarray-of-rand :rows 2))
;; => #object[org.nd4j.linalg.cpu.nativecpu.NDArray 0x49143b08 [0.76, 0.92]]

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; data-set creation
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def ds-with-single-example (new-ds :input [1 2 3 4]
                                    :output [0.0 1.0 0.0]))

(as-list :ds ds-with-single-example :as-code? false)
;; =>
;; #object[java.util.ArrayList 0x5d703d12
;;[===========INPUT===================
;;[1.00, 2.00, 3.00, 4.00]
;;=================OUTPUT==================
;;[0.00, 1.00, 0.00]]]

(def ds-with-multiple-examples (new-ds
                                :input [[1 2 3 4] [2 4 6 8]]
                                :output [[0.0 1.0 0.0] [0.0 0.0 1.0]]))

(as-list :ds ds-with-multiple-examples :as-code? false)
;; =>
;;#object[java.util.ArrayList 0x29c7a9e2
;;[===========INPUT===================
;;[1.00, 2.00, 3.00, 4.00]
;;=================OUTPUT==================
;;[0.00, 1.00, 0.00],
;;===========INPUT===================
;;[2.00, 4.00, 6.00, 8.00]
;;=================OUTPUT==================
;;[0.00, 0.00, 1.00]]]

;; we can create a dataset iterator from the code which creates datasets
;; and set the labels for our outputs (optional)

(def ds-with-multiple-examples
  (new-ds
   :input [[1 2 3 4] [2 4 6 8]]
   :output [[0.0 1.0 0.0] [0.0 0.0 1.0]]))

;; iterator
(def training-rr-ds-iter
  (new-existing-dataset-iterator
   :dataset ds-with-multiple-examples
   :labels ["foo" "baz" "foobaz"]))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; data-set normalization
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; this gathers statistics on the dataset and normalizes the data
;; and applies the transformation to all dataset objects in the iterator
(def train-iter-normalized
  (c/normalize-iter! :iter training-rr-ds-iter
                     :normalizer (ds-pp/new-standardize-normalization-ds-preprocessor)
                     :as-code? false))

;; above returns the normalized iterator
;; to get fit normalizer

(def the-normalizer
  (get-pre-processor train-iter-normalized))

Model configuration

Creating a neural network configuration with singe and multiple layers

(ns my.ns
  (:require [dl4clj.nn.conf.builders.layers :as l]
            [dl4clj.nn.conf.builders.nn :as nn]
            [dl4clj.nn.conf.distributions :as dist]
            [dl4clj.nn.conf.input-pre-processor :as pp]
            [dl4clj.nn.conf.step-fns :as s-fn]))

;; nn/builder has 3 types of args
;; 1) args which set network configuration params
;; 2) args which set default values for layers
;; 3) args which set multi layer network configuration params

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; single layer nn configuration
;; here we are setting network configuration
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(nn/builder :optimization-algo :stochastic-gradient-descent
            :seed 123
            :iterations 1
            :minimize? true
            :use-drop-connect? false
            :lr-score-based-decay-rate 0.002
            :regularization? false
            :step-fn :default-step-fn
            :layers {:dense-layer {:activation-fn :relu
                                   :updater :adam
                                   :adam-mean-decay 0.2
                                   :adam-var-decay 0.1
                                   :learning-rate 0.006
                                   :weight-init :xavier
                                   :layer-name "single layer model example"
                                   :n-in 10
                                   :n-out 20}})

;; there are several options within a nn-conf map which can be configuration maps
;; or calls to fns
;; It doesn't matter which option you choose and you don't have to stay consistent
;; the list of params which can be passed as config maps or fn calls will
;; be enumerated at a later date

(nn/builder :optimization-algo :stochastic-gradient-descent
            :seed 123
            :iterations 1
            :minimize? true
            :use-drop-connect? false
            :lr-score-based-decay-rate 0.002
            :regularization? false
            :step-fn (s-fn/new-default-step-fn)
            :build? true
            ;; dont need to specify layer order, theres only one
            :layers (l/dense-layer-builder
                    :activation-fn :relu
                    :updater :adam
                    :adam-mean-decay 0.2
                    :adam-var-decay 0.1
                    :dist (dist/new-normal-distribution :mean 0 :std 1)
                    :learning-rate 0.006
                    :weight-init :xavier
                    :layer-name "single layer model example"
                    :n-in 10
                    :n-out 20))

;; these configurations are the same

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; multi-layer configuration
;; here we are also setting layer defaults
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; defaults will apply to layers which do not specify those value in their config

(nn/builder
 :optimization-algo :stochastic-gradient-descent
 :seed 123
 :iterations 1
 :minimize? true
 :use-drop-connect? false
 :lr-score-based-decay-rate 0.002
 :regularization? false
 :default-activation-fn :sigmoid
 :default-weight-init :uniform

 ;; we need to specify the layer order
 :layers {0 (l/activation-layer-builder
             :activation-fn :relu
             :updater :adam
             :adam-mean-decay 0.2
             :adam-var-decay 0.1
             :learning-rate 0.006
             :weight-init :xavier
             :layer-name "example first layer"
             :n-in 10
             :n-out 20)
          1 {:output-layer {:n-in 20
                            :n-out 2
                            :loss-fn :mse
                            :layer-name "example output layer"}}})

;; specifying multi-layer config params

(nn/builder
 ;; network args
 :optimization-algo :stochastic-gradient-descent
 :seed 123
 :iterations 1
 :minimize? true
 :use-drop-connect? false
 :lr-score-based-decay-rate 0.002
 :regularization? false

 ;; layer defaults
 :default-activation-fn :sigmoid
 :default-weight-init :uniform

 ;; the layers
 :layers {0 (l/activation-layer-builder
             :activation-fn :relu
             :updater :adam
             :adam-mean-decay 0.2
             :adam-var-decay 0.1
             :learning-rate 0.006
             :weight-init :xavier
             :layer-name "example first layer"
             :n-in 10
             :n-out 20)
          1 {:output-layer {:n-in 20
                            :n-out 2
                            :loss-fn :mse
                            :layer-name "example output layer"}}}
 ;; multi layer network args
 :backprop? true
 :backprop-type :standard
 :pretrain? false
 :input-pre-processors {0 (pp/new-zero-mean-pre-pre-processor)
                        1 {:unit-variance-processor {}}})

Configuration to Trained models

Multi Layer models

an implementation of the dl4j mnist classification example

(ns my.ns
  (:require [dl4clj.datasets.iterators :as iter]
            [dl4clj.datasets.input-splits :as split]
            [dl4clj.datasets.record-readers :as rr]
            [dl4clj.optimize.listeners :as listener]
            [dl4clj.nn.conf.builders.nn :as nn]
            [dl4clj.nn.multilayer.multi-layer-network :as mln]
            [dl4clj.nn.api.model :refer [init! set-listeners!]]
            [dl4clj.nn.api.multi-layer-network :refer [evaluate-classification]]
            [dl4clj.datasets.api.record-readers :refer [initialize-rr!]]
            [dl4clj.eval.api.eval :refer [get-stats get-accuracy]]
            [dl4clj.core :as c]))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; nn-conf -> multi-layer-network
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def nn-conf
  (nn/builder
   ;; network args
   :optimization-algo :stochastic-gradient-descent
   :seed 123 :iterations 1 :regularization? true

   ;; setting layer defaults
   :default-activation-fn :relu :default-l2 7.5e-6
   :default-weight-init :xavier :default-learning-rate 0.0015
   :default-updater :nesterovs :default-momentum 0.98

   ;; setting layer configuration
   :layers {0 {:dense-layer
               {:layer-name "example first layer"
                :n-in 784 :n-out 500}}
            1 {:dense-layer
               {:layer-name "example second layer"
                :n-in 500 :n-out 100}}
            2 {:output-layer
               {:n-in 100 :n-out 10
                ;; layer specific params
                :loss-fn :negativeloglikelihood
                :activation-fn :softmax
                :layer-name "example output layer"}}}

   ;; multi layer args
   :backprop? true
   :pretrain? false))

(def multi-layer-network (c/model-from-conf nn-conf))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; local cpu training with dl4j pre-built iterators
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; lets use the pre-built Mnist data set iterator

(def train-mnist-iter
  (iter/new-mnist-data-set-iterator
   :batch-size 64
   :train? true
   :seed 123))

(def test-mnist-iter
  (iter/new-mnist-data-set-iterator
   :batch-size 64
   :train? false
   :seed 123))

;; and lets set a listener so we can know how training is going

(def score-listener (listener/new-score-iteration-listener :print-every-n 5))

;; and attach it to our model

;; TODO: listeners are broken, look into log4j warnning
(def mln-with-listener (set-listeners! :model multi-layer-network
                                       :listeners [score-listener]))

(def trained-mln (mln/train-mln-with-ds-iter! :mln mln-with-listener
                                              :iter train-mnist-iter
                                              :n-epochs 15
                                              :as-code? false))

;; training happens because :as-code? = false
;; if it was true, we would still just have a data structure
;; we now have a trained model that has seen the training dataset 15 times
;; time to evaluate our model

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;Create an evaluation object
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def eval-obj (evaluate-classification :mln trained-mln
                                       :iter test-mnist-iter))

;; always remember that these objects are stateful, dont use the same eval-obj
;; to eval two different networks
;; we trained the model on a training dataset.  We evaluate on a test set

(println (get-stats :evaler eval-obj))
;; this will print the stats to standard out for each feature/label pair

;;Examples labeled as 0 classified by model as 0: 968 times
;;Examples labeled as 0 classified by model as 1: 1 times
;;Examples labeled as 0 classified by model as 2: 1 times
;;Examples labeled as 0 classified by model as 3: 1 times
;;Examples labeled as 0 classified by model as 5: 1 times
;;Examples labeled as 0 classified by model as 6: 3 times
;;Examples labeled as 0 classified by model as 7: 1 times
;;Examples labeled as 0 classified by model as 8: 2 times
;;Examples labeled as 0 classified by model as 9: 2 times
;;Examples labeled as 1 classified by model as 1: 1126 times
;;Examples labeled as 1 classified by model as 2: 2 times
;;Examples labeled as 1 classified by model as 3: 1 times
;;Examples labeled as 1 classified by model as 5: 1 times
;;Examples labeled as 1 classified by model as 6: 2 times
;;Examples labeled as 1 classified by model as 7: 1 times
;;Examples labeled as 1 classified by model as 8: 2 times
;;Examples labeled as 2 classified by model as 0: 3 times
;;Examples labeled as 2 classified by model as 1: 2 times
;;Examples labeled as 2 classified by model as 2: 1006 times
;;Examples labeled as 2 classified by model as 3: 2 times
;;Examples labeled as 2 classified by model as 4: 3 times
;;Examples labeled as 2 classified by model as 6: 3 times
;;Examples labeled as 2 classified by model as 7: 7 times
;;Examples labeled as 2 classified by model as 8: 6 times
;;Examples labeled as 3 classified by model as 2: 4 times
;;Examples labeled as 3 classified by model as 3: 990 times
;;Examples labeled as 3 classified by model as 5: 3 times
;;Examples labeled as 3 classified by model as 7: 3 times
;;Examples labeled as 3 classified by model as 8: 3 times
;;Examples labeled as 3 classified by model as 9: 7 times
;;Examples labeled as 4 classified by model as 2: 2 times
;;Examples labeled as 4 classified by model as 3: 1 times
;;Examples labeled as 4 classified by model as 4: 967 times
;;Examples labeled as 4 classified by model as 6: 4 times
;;Examples labeled as 4 classified by model as 7: 1 times
;;Examples labeled as 4 classified by model as 9: 7 times
;;Examples labeled as 5 classified by model as 0: 2 times
;;Examples labeled as 5 classified by model as 3: 6 times
;;Examples labeled as 5 classified by model as 4: 1 times
;;Examples labeled as 5 classified by model as 5: 874 times
;;Examples labeled as 5 classified by model as 6: 3 times
;;Examples labeled as 5 classified by model as 7: 1 times
;;Examples labeled as 5 classified by model as 8: 3 times
;;Examples labeled as 5 classified by model as 9: 2 times
;;Examples labeled as 6 classified by model as 0: 4 times
;;Examples labeled as 6 classified by model as 1: 3 times
;;Examples labeled as 6 classified by model as 3: 2 times
;;Examples labeled as 6 classified by model as 4: 4 times
;;Examples labeled as 6 classified by model as 5: 4 times
;;Examples labeled as 6 classified by model as 6: 939 times
;;Examples labeled as 6 classified by model as 7: 1 times
;;Examples labeled as 6 classified by model as 8: 1 times
;;Examples labeled as 7 classified by model as 1: 7 times
;;Examples labeled as 7 classified by model as 2: 4 times
;;Examples labeled as 7 classified by model as 3: 3 times
;;Examples labeled as 7 classified by model as 7: 1005 times
;;Examples labeled as 7 classified by model as 8: 2 times
;;Examples labeled as 7 classified by model as 9: 7 times
;;Examples labeled as 8 classified by model as 0: 3 times
;;Examples labeled as 8 classified by model as 2: 3 times
;;Examples labeled as 8 classified by model as 3: 2 times
;;Examples labeled as 8 classified by model as 4: 4 times
;;Examples labeled as 8 classified by model as 5: 3 times
;;Examples labeled as 8 classified by model as 6: 2 times
;;Examples labeled as 8 classified by model as 7: 4 times
;;Examples labeled as 8 classified by model as 8: 947 times
;;Examples labeled as 8 classified by model as 9: 6 times
;;Examples labeled as 9 classified by model as 0: 2 times
;;Examples labeled as 9 classified by model as 1: 2 times
;;Examples labeled as 9 classified by model as 3: 4 times
;;Examples labeled as 9 classified by model as 4: 8 times
;;Examples labeled as 9 classified by model as 6: 1 times
;;Examples labeled as 9 classified by model as 7: 4 times
;;Examples labeled as 9 classified by model as 8: 2 times
;;Examples labeled as 9 classified by model as 9: 986 times

;;==========================Scores========================================
;; Accuracy:        0.9808
;; Precision:       0.9808
;; Recall:          0.9807
;; F1 Score:        0.9807
;;========================================================================

;; can get the stats that are printed via fns in the evaluation namespace
;; after running eval-model-whole-ds

(get-accuracy :evaler evaler-with-stats) ;; => 0.9808

Model Tuning

Early Stopping (controlling training)

it is recommened you start here when designing models

using dl4clj.core


(ns my.ns
  (:require [dl4clj.earlystopping.termination-conditions :refer :all]
            [dl4clj.earlystopping.model-saver :refer [new-in-memory-saver]]
            [dl4clj.nn.api.multi-layer-network :refer [evaluate-classification]]
            [dl4clj.eval.api.eval :refer [get-stats]]
            [dl4clj.nn.conf.builders.nn :as nn]
            [dl4clj.datasets.iterators :as iter]
            [dl4clj.core :as c]))

(def nn-conf
  (nn/builder
   ;; network args
   :optimization-algo :stochastic-gradient-descent
   :seed 123
   :iterations 1
   :regularization? true

   ;; setting layer defaults
   :default-activation-fn :relu
   :default-l2 7.5e-6
   :default-weight-init :xavier
   :default-learning-rate 0.0015
   :default-updater :nesterovs
   :default-momentum 0.98

   ;; setting layer configuration
   :layers {0 {:dense-layer
               {:layer-name "example first layer"
                :n-in 784 :n-out 500}}
            1 {:dense-layer
               {:layer-name "example second layer"
                :n-in 500 :n-out 100}}
            2 {:output-layer
               {:n-in 100 :n-out 10
                ;; layer specific params
                :loss-fn :negativeloglikelihood
                :activation-fn :softmax
                :layer-name "example output layer"}}}

   ;; multi layer args
   :backprop? true
   :pretrain? false))

(def train-iter
  (iter/new-mnist-data-set-iterator
   :batch-size 64
   :train? true
   :seed 123))

(def test-iter
  (iter/new-mnist-data-set-iterator
   :batch-size 64
   :train? false
   :seed 123))

(def invalid-score-condition (new-invalid-score-iteration-termination-condition))

(def max-score-condition (new-max-score-iteration-termination-condition
                          :max-score 20.0))

(def max-time-condition (new-max-time-iteration-termination-condition
                         :max-time-val 10
                         :max-time-unit :minutes))

(def score-doesnt-improve-condition (new-score-improvement-epoch-termination-condition
                                     :max-n-epoch-no-improve 5))

(def target-score-condition (new-best-score-epoch-termination-condition
                             :best-expected-score 0.009))

(def max-number-epochs-condition (new-max-epochs-termination-condition :max-n 20))

(def in-mem-saver (new-in-memory-saver))

(def trained-mln
;; defaults to returning the model
  (c/train-with-early-stopping
   :nn-conf nn-conf
   :training-iter train-mnist-iter
   :testing-iter test-mnist-iter
   :eval-every-n-epochs 1
   :iteration-termination-conditions [invalid-score-condition
                                      max-score-condition
                                      max-time-condition]
   :epoch-termination-conditions [score-doesnt-improve-condition
                                  target-score-condition
                                  max-number-epochs-condition]
   :save-last-model? true
   :model-saver in-mem-saver
   :as-code? false))

(def model-evaler
  (evaluate-classification :mln trained-mln :iter test-mnist-iter))

(println (get-stats :evaler model-evaler))

explicit, step by step way of doing this

(ns my.ns
  (:require [dl4clj.earlystopping.early-stopping-config :refer [new-early-stopping-config]]
            [dl4clj.earlystopping.termination-conditions :refer :all]
            [dl4clj.earlystopping.model-saver :refer [new-in-memory-saver new-local-file-model-saver]]
            [dl4clj.earlystopping.score-calc :refer [new-ds-loss-calculator]]
            [dl4clj.earlystopping.early-stopping-trainer :refer [new-early-stopping-trainer]]
            [dl4clj.earlystopping.api.early-stopping-trainer :refer [fit-trainer!]]
            [dl4clj.nn.conf.builders.nn :as nn]
            [dl4clj.nn.multilayer.multi-layer-network :as mln]
            [dl4clj.utils :refer [load-model!]]
            [dl4clj.datasets.iterators :as iter]
            [dl4clj.core :as c]))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; start with our network config
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def nn-conf
  (nn/builder
   ;; network args
   :optimization-algo :stochastic-gradient-descent
   :seed 123 :iterations 1 :regularization? true
   ;; setting layer defaults
   :default-activation-fn :relu :default-l2 7.5e-6
   :default-weight-init :xavier :default-learning-rate 0.0015
   :default-updater :nesterovs :default-momentum 0.98
   ;; setting layer configuration
   :layers {0 {:dense-layer
               {:layer-name "example first layer"
                :n-in 784 :n-out 500}}
            1 {:dense-layer
               {:layer-name "example second layer"
                :n-in 500 :n-out 100}}
            2 {:output-layer
               {:n-in 100 :n-out 10
                ;; layer specific params
                :loss-fn :negativeloglikelihood
                :activation-fn :softmax
                :layer-name "example output layer"}}}
   ;; multi layer args
   :backprop? true
   :pretrain? false))

(def mln (c/model-from-conf nn-conf))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; the training/testing data
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def train-iter
  (iter/new-mnist-data-set-iterator
   :batch-size 64
   :train? true
   :seed 123))

(def test-iter
  (iter/new-mnist-data-set-iterator
   :batch-size 64
   :train? false
   :seed 123))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; we are going to need termination conditions
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; these allow us to control when we exit training

;; this can be based off of iterations or epochs

;; iteration termination conditions

(def invalid-score-condition (new-invalid-score-iteration-termination-condition))

(def max-score-condition (new-max-score-iteration-termination-condition
                          :max-score 20.0))

(def max-time-condition (new-max-time-iteration-termination-condition
                         :max-time-val 10
                         :max-time-unit :minutes))

;; epoch termination conditions

(def score-doesnt-improve-condition (new-score-improvement-epoch-termination-condition
                                     :max-n-epoch-no-improve 5))

(def target-score-condition (new-best-score-epoch-termination-condition :best-expected-score 0.009))

(def max-number-epochs-condition (new-max-epochs-termination-condition :max-n 20))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; we also need a way to save our model
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; can be in memory or to a local directory

(def in-mem-saver (new-in-memory-saver))

(def local-file-saver (new-local-file-model-saver :directory "resources/tmp/readme/"))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; set up your score calculator
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def score-calcer (new-ds-loss-calculator :iter test-iter
                                          :average? true))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; create an early stopping configuration
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; termination conditions
;; a way to save our model
;; a way to calculate the score of our model on the dataset

(def early-stopping-conf
  (new-early-stopping-config
   :epoch-termination-conditions [score-doesnt-improve-condition
                                  target-score-condition
                                  max-number-epochs-condition]
   :iteration-termination-conditions [invalid-score-condition
                                      max-score-condition
                                      max-time-condition]
   :eval-every-n-epochs 5
   :model-saver local-file-saver
   :save-last-model? true
   :score-calculator score-calcer))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; create an early stopping trainer from our data, model and early stopping conf
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def es-trainer (new-early-stopping-trainer :early-stopping-conf early-stopping-conf
                                            :mln mln
                                            :iter train-iter))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; fit and use our early stopping trainer
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def es-trainer-fitted (fit-trainer! es-trainer :as-code? false))

;; when the trainer terminates, you will see something like this
;;[nREPL-worker-24] BaseEarlyStoppingTrainer INFO  Completed training epoch 14
;;[nREPL-worker-24] BaseEarlyStoppingTrainer INFO  New best model: score = 0.005225599372851298,
;;                                                   epoch = 14 (previous: score = 0.018243224899038346, epoch = 7)
;;[nREPL-worker-24] BaseEarlyStoppingTrainer INFO Hit epoch termination condition at epoch 14.
;;                                           Details: BestScoreEpochTerminationCondition(0.009)

;; and if we look at the es-trainer-fitted object we see

;;#object[org.deeplearning4j.earlystopping.EarlyStoppingResult 0x5ab74f27 EarlyStoppingResult
;;(terminationReason=EpochTerminationCondition,details=BestScoreEpochTerminationCondition(0.009),
;; bestModelEpoch=14,bestModelScore=0.005225599372851298,totalEpochs=15)]

;; and our model has been saved to /resources/tmp/readme/bestModel.bin
;; there we have our model config, model params and our updater state

;; we can then load this model to use it or continue refining it

(def loaded-model (load-model! :path "resources/tmp/readme/bestModel.bin"
                               :load-updater? true))

Transfer Learning (freezing layers)


;; TODO: need to write up examples

Spark Training

dl4j Spark usage

How it is done in dl4clj

Uses dl4clj.core
- This example uses a fn which takes care of most steps for you
  - allows you to pass args as code bc the fn accounts for the multiple spark contexts issue encountered when everything is just a data structure


(ns my.ns
  (:require [dl4clj.nn.conf.builders.layers :as l]
            [dl4clj.nn.conf.builders.nn :as nn]
            [dl4clj.datasets.iterators :refer [new-iris-data-set-iterator]]
            [dl4clj.eval.api.eval :refer [get-stats]]
            [dl4clj.spark.masters.param-avg :as master]
            [dl4clj.spark.data.java-rdd :refer [new-java-spark-context
                                                java-rdd-from-iter]]
            [dl4clj.spark.api.dl4j-multi-layer :refer [eval-classification-spark-mln
                                                       get-spark-context]]
            [dl4clj.core :as c]))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Step 1, create your model config
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def mln-conf
  (nn/builder
   :optimization-algo :stochastic-gradient-descent
   :default-learning-rate 0.006
   :layers {0 (l/dense-layer-builder :n-in 4 :n-out 2 :activation-fn :relu)
            1 {:output-layer
               {:loss-fn :negativeloglikelihood
                :n-in 2 :n-out 3
                :activation-fn :soft-max
                :weight-init :xavier}}}
   :backprop? true
   :backprop-type :standard))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Step 2, training master
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def training-master
  (master/new-parameter-averaging-training-master
   :build? true
   :rdd-n-examples 10
   :n-workers 4
   :averaging-freq 10
   :batch-size-per-worker 2
   :export-dir "resources/spark/master/"
   :rdd-training-approach :direct
   :repartition-data :always
   :repartition-strategy :balanced
   :seed 1234
   :save-updater? true
   :storage-level :none))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Step 3, spark context
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def your-spark-context
  (new-java-spark-context :app-name "example app"))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Step 4, training data
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def iris-iter
  (new-iris-data-set-iterator
   :batch-size 1
   :n-examples 5))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Step 5, spark mln
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def fitted-spark-mln
  (c/train-with-spark :spark-context your-spark-context
                      :mln-conf mln-conf
                      :training-master training-master
                      :iter iris-iter
                      :n-epochs 1
                      :as-code? false))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Step 5, use spark context from spark-mln to create rdd
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; TODO: eliminate this step

(def our-rdd
  (let [sc (get-spark-context fitted-spark-mln :as-code? false)]
    (java-rdd-from-iter :spark-context sc
                        :iter iris-iter)))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Step 6, evaluation model and print stats (poor performance of model expected)
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def eval-obj
  (eval-classification-spark-mln
   :spark-mln fitted-spark-mln
   :rdd our-rdd))

(println (get-stats :evaler eval-obj))

this example demonstrates the dl4j workflow
- NOTE: unlike the previous example, this one requires dl4j objects to be used
  - this is becaues spark only wants you to have one spark context at a time

(ns my.ns
  (:require [dl4clj.nn.conf.builders.layers :as l]
            [dl4clj.nn.conf.builders.nn :as nn]
            [dl4clj.datasets.iterators :refer [new-iris-data-set-iterator]]
            [dl4clj.eval.api.eval :refer [get-stats]]
            [dl4clj.spark.masters.param-avg :as master]
            [dl4clj.spark.data.java-rdd :refer [new-java-spark-context java-rdd-from-iter]]
            [dl4clj.spark.dl4j-multi-layer :as spark-mln]
            [dl4clj.spark.api.dl4j-multi-layer :refer [fit-spark-mln!
                                                       eval-classification-spark-mln]]))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Step 1, create your model
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def mln-conf
  (nn/builder
   :optimization-algo :stochastic-gradient-descent
   :default-learning-rate 0.006
   :layers {0 (l/dense-layer-builder :n-in 4 :n-out 2 :activation-fn :relu)
            1 {:output-layer
               {:loss-fn :negativeloglikelihood
                :n-in 2 :n-out 3
                :activation-fn :soft-max
                :weight-init :xavier}}}
   :backprop? true
   :as-code? false
   :backprop-type :standard))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Step 2, create a training master
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; not all options specified, but most are

(def training-master
  (master/new-parameter-averaging-training-master
   :build? true
   :rdd-n-examples 10
   :n-workers 4
   :averaging-freq 10
   :batch-size-per-worker 2
   :export-dir "resources/spark/master/"
   :rdd-training-approach :direct
   :repartition-data :always
   :repartition-strategy :balanced
   :seed 1234
   :as-code? false
   :save-updater? true
   :storage-level :none))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Step 3, create a Spark Multi Layer Network
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def your-spark-context
  (new-java-spark-context :app-name "example app" :as-code? false))

;; new-java-spark-context will turn an existing spark-configuration into a java spark context
;; or create a new java spark context with master set to "local[*]" and the app name
;; set to :app-name


(def spark-mln
  (spark-mln/new-spark-multi-layer-network
   :spark-context your-spark-context
   :mln mln-conf
   :training-master training-master
   :as-code? false))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Step 4, load your data
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; one way is via a dataset-iterator
;; can make one directly from a dataset (iterator data-set)
;; see: nd4clj.linalg.dataset.api.data-set and nd4clj.linalg.dataset.data-set
;; we are going to use a pre-built one

(def iris-iter
  (new-iris-data-set-iterator
   :batch-size 1
   :n-examples 5
   :as-code? false))

;; now lets convert the data into a javaRDD

(def our-rdd
  (java-rdd-from-iter :spark-context your-spark-context
                      :iter iris-iter))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Step 5, fit and evaluate the model
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(def fitted-spark-mln
  (fit-spark-mln!
   :spark-mln spark-mln
   :rdd our-rdd
   :n-epochs 1))
;; this fn also has the option to supply :path-to-data instead of :rdd
;; that path should point to a directory containing a number of dataset objects

(def eval-obj
  (eval-classification-spark-mln
   :spark-mln fitted-spark-mln
   :rdd our-rdd))
;; we would want to have different testing and training rdd's but here we are using
;; the data we trained on

;; lets get the stats for how our model performed

(println (get-stats :evaler eval-obj))

Terminology

Coming soon

Packages to come back to:

Implement ComputationGraphs and the classes which use them

NLP

Parallelism

TSNE

https://deeplearning4j.org/doc/org/deeplearning4j/plot/package-summary.html

UI

Author: yetanalytics
Source Code: https://github.com/yetanalytics/dl4clj
License: BSD-2-Clause License

#machine-learning #deep-learning

Hermann Frami

1651383480

A Simple Wrapper Around Amplify AppSync Simulator

This serverless plugin is a wrapper for amplify-appsync-simulator made for testing AppSync APIs built with serverless-appsync-plugin.

Install

npm install serverless-appsync-simulator
# or
yarn add serverless-appsync-simulator

Usage

This plugin relies on your serverless yml file and on the serverless-offline plugin.

plugins:
  - serverless-dynamodb-local # only if you need dynamodb resolvers and you don't have an external dynamodb
  - serverless-appsync-simulator
  - serverless-offline

Note: Order is important serverless-appsync-simulator must go before serverless-offline

To start the simulator, run the following command:

sls offline start

You should see in the logs something like:

...
Serverless: AppSync endpoint: http://localhost:20002/graphql
Serverless: GraphiQl: http://localhost:20002
...

Configuration

Put options under custom.appsync-simulator in your serverless.yml file

| option | default | description | | ------------------------ | -------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------- | | apiKey | 0123456789 | When using API_KEY as authentication type, the key to authenticate to the endpoint. | | port | 20002 | AppSync operations port; if using multiple APIs, the value of this option will be used as a starting point, and each other API will have a port of lastPort + 10 (e.g. 20002, 20012, 20022, etc.) | | wsPort | 20003 | AppSync subscriptions port; if using multiple APIs, the value of this option will be used as a starting point, and each other API will have a port of lastPort + 10 (e.g. 20003, 20013, 20023, etc.) | | location | . (base directory) | Location of the lambda functions handlers. | | refMap | {} | A mapping of resource resolutions for the Ref function | | getAttMap | {} | A mapping of resource resolutions for the GetAtt function | | importValueMap | {} | A mapping of resource resolutions for the ImportValue function | | functions | {} | A mapping of external functions for providing invoke url for external fucntions | | dynamoDb.endpoint | http://localhost:8000 | Dynamodb endpoint. Specify it if you're not using serverless-dynamodb-local. Otherwise, port is taken from dynamodb-local conf | | dynamoDb.region | localhost | Dynamodb region. Specify it if you're connecting to a remote Dynamodb intance. | | dynamoDb.accessKeyId | DEFAULT_ACCESS_KEY | AWS Access Key ID to access DynamoDB | | dynamoDb.secretAccessKey | DEFAULT_SECRET | AWS Secret Key to access DynamoDB | | dynamoDb.sessionToken | DEFAULT_ACCESS_TOKEEN | AWS Session Token to access DynamoDB, only if you have temporary security credentials configured on AWS | | dynamoDb.* | | You can add every configuration accepted by DynamoDB SDK | | rds.dbName | | Name of the database | | rds.dbHost | | Database host | | rds.dbDialect | | Database dialect. Possible values (mysql | postgres) | | rds.dbUsername | | Database username | | rds.dbPassword | | Database password | | rds.dbPort | | Database port | | watch | - *.graphql
- *.vtl | Array of glob patterns to watch for hot-reloading. |

Example:

custom:
  appsync-simulator:
    location: '.webpack/service' # use webpack build directory
    dynamoDb:
      endpoint: 'http://my-custom-dynamo:8000'

Hot-reloading

By default, the simulator will hot-relad when changes to *.graphql or *.vtl files are detected. Changes to *.yml files are not supported (yet? - this is a Serverless Framework limitation). You will need to restart the simulator each time you change yml files.

Hot-reloading relies on watchman. Make sure it is installed on your system.

You can change the files being watched with the watch option, which is then passed to watchman as the match expression.

e.g.

custom:
  appsync-simulator:
    watch:
      - ["match", "handlers/**/*.vtl", "wholename"] # => array is interpreted as the literal match expression
      - "*.graphql"                                 # => string like this is equivalent to `["match", "*.graphql"]`

Or you can opt-out by leaving an empty array or set the option to false

Note: Functions should not require hot-reloading, unless you are using a transpiler or a bundler (such as webpack, babel or typescript), un which case you should delegate hot-reloading to that instead.

Resource CloudFormation functions resolution

This plugin supports some resources resolution from the Ref, Fn::GetAtt and Fn::ImportValue functions in your yaml file. It also supports some other Cfn functions such as Fn::Join, Fb::Sub, etc.

Note: Under the hood, this features relies on the cfn-resolver-lib package. For more info on supported cfn functions, refer to the documentation

Basic usage

You can reference resources in your functions' environment variables (that will be accessible from your lambda functions) or datasource definitions. The plugin will automatically resolve them for you.

provider:
  environment:
    BUCKET_NAME:
      Ref: MyBucket # resolves to `my-bucket-name`

resources:
  Resources:
    MyDbTable:
      Type: AWS::DynamoDB::Table
      Properties:
        TableName: myTable
      ...
    MyBucket:
      Type: AWS::S3::Bucket
      Properties:
        BucketName: my-bucket-name
    ...

# in your appsync config
dataSources:
  - type: AMAZON_DYNAMODB
    name: dynamosource
    config:
      tableName:
        Ref: MyDbTable # resolves to `myTable`

Override (or mock) values

Sometimes, some references cannot be resolved, as they come from an Output from Cloudformation; or you might want to use mocked values in your local environment.

In those cases, you can define (or override) those values using the refMap, getAttMap and importValueMap options.

refMap takes a mapping of resource name to value pairs
getAttMap takes a mapping of resource name to attribute/values pairs
importValueMap takes a mapping of import name to values pairs

Example:

custom:
  appsync-simulator:
    refMap:
      # Override `MyDbTable` resolution from the previous example.
      MyDbTable: 'mock-myTable'
    getAttMap:
      # define ElasticSearchInstance DomainName
      ElasticSearchInstance:
        DomainEndpoint: 'localhost:9200'
    importValueMap:
      other-service-api-url: 'https://other.api.url.com/graphql'

# in your appsync config
dataSources:
  - type: AMAZON_ELASTICSEARCH
    name: elasticsource
    config:
      # endpoint resolves as 'http://localhost:9200'
      endpoint:
        Fn::Join:
          - ''
          - - https://
            - Fn::GetAtt:
                - ElasticSearchInstance
                - DomainEndpoint

Key-value mock notation

In some special cases you will need to use key-value mock nottation. Good example can be case when you need to include serverless stage value (${self:provider.stage}) in the import name.

This notation can be used with all mocks - refMap, getAttMap and importValueMap

provider:
  environment:
    FINISH_ACTIVITY_FUNCTION_ARN:
      Fn::ImportValue: other-service-api-${self:provider.stage}-url

custom:
  serverless-appsync-simulator:
    importValueMap:
      - key: other-service-api-${self:provider.stage}-url
        value: 'https://other.api.url.com/graphql'

Limitations

This plugin only tries to resolve the following parts of the yml tree:

provider.environment
functions[*].environment
custom.appSync

If you have the need of resolving others, feel free to open an issue and explain your use case.

For now, the supported resources to be automatically resovled by Ref: are:

DynamoDb tables
S3 Buckets

Feel free to open a PR or an issue to extend them as well.

External functions

When a function is not defined withing the current serverless file you can still call it by providing an invoke url which should point to a REST method. Make sure you specify "get" or "post" for the method. Default is "get", but you probably want "post".

custom:
  appsync-simulator:
    functions:
      addUser:
        url: http://localhost:3016/2015-03-31/functions/addUser/invocations
        method: post
      addPost:
        url: https://jsonplaceholder.typicode.com/posts
        method: post

Supported Resolver types

This plugin supports resolvers implemented by amplify-appsync-simulator, as well as custom resolvers.

From Aws Amplify:

NONE
AWS_LAMBDA
AMAZON_DYNAMODB
PIPELINE

Implemented by this plugin

AMAZON_ELASTIC_SEARCH
HTTP
RELATIONAL_DATABASE

Relational Database

Sample VTL for a create mutation

#set( $cols = [] )
#set( $vals = [] )
#foreach( $entry in $ctx.args.input.keySet() )
  #set( $regex = "([a-z])([A-Z]+)")
  #set( $replacement = "$1_$2")
  #set( $toSnake = $entry.replaceAll($regex, $replacement).toLowerCase() )
  #set( $discard = $cols.add("$toSnake") )
  #if( $util.isBoolean($ctx.args.input[$entry]) )
      #if( $ctx.args.input[$entry] )
        #set( $discard = $vals.add("1") )
      #else
        #set( $discard = $vals.add("0") )
      #end
  #else
      #set( $discard = $vals.add("'$ctx.args.input[$entry]'") )
  #end
#end
#set( $valStr = $vals.toString().replace("[","(").replace("]",")") )
#set( $colStr = $cols.toString().replace("[","(").replace("]",")") )
#if ( $valStr.substring(0, 1) != '(' )
  #set( $valStr = "($valStr)" )
#end
#if ( $colStr.substring(0, 1) != '(' )
  #set( $colStr = "($colStr)" )
#end
{
  "version": "2018-05-29",
  "statements":   ["INSERT INTO <name-of-table> $colStr VALUES $valStr", "SELECT * FROM    <name-of-table> ORDER BY id DESC LIMIT 1"]
}

Sample VTL for an update mutation

#set( $update = "" )
#set( $equals = "=" )
#foreach( $entry in $ctx.args.input.keySet() )
  #set( $cur = $ctx.args.input[$entry] )
  #set( $regex = "([a-z])([A-Z]+)")
  #set( $replacement = "$1_$2")
  #set( $toSnake = $entry.replaceAll($regex, $replacement).toLowerCase() )
  #if( $util.isBoolean($cur) )
      #if( $cur )
        #set ( $cur = "1" )
      #else
        #set ( $cur = "0" )
      #end
  #end
  #if ( $util.isNullOrEmpty($update) )
      #set($update = "$toSnake$equals'$cur'" )
  #else
      #set($update = "$update,$toSnake$equals'$cur'" )
  #end
#end
{
  "version": "2018-05-29",
  "statements":   ["UPDATE <name-of-table> SET $update WHERE id=$ctx.args.input.id", "SELECT * FROM <name-of-table> WHERE id=$ctx.args.input.id"]
}

Sample resolver for delete mutation

{
  "version": "2018-05-29",
  "statements":   ["UPDATE <name-of-table> set deleted_at=NOW() WHERE id=$ctx.args.id", "SELECT * FROM <name-of-table> WHERE id=$ctx.args.id"]
}

Sample mutation response VTL with support for handling AWSDateTime

#set ( $index = -1)
#set ( $result = $util.parseJson($ctx.result) )
#set ( $meta = $result.sqlStatementResults[1].columnMetadata)
#foreach ($column in $meta)
    #set ($index = $index + 1)
    #if ( $column["typeName"] == "timestamptz" )
        #set ($time = $result["sqlStatementResults"][1]["records"][0][$index]["stringValue"] )
        #set ( $nowEpochMillis = $util.time.parseFormattedToEpochMilliSeconds("$time.substring(0,19)+0000", "yyyy-MM-dd HH:mm:ssZ") )
        #set ( $isoDateTime = $util.time.epochMilliSecondsToISO8601($nowEpochMillis) )
        $util.qr( $result["sqlStatementResults"][1]["records"][0][$index].put("stringValue", "$isoDateTime") )
    #end
#end
#set ( $res = $util.parseJson($util.rds.toJsonString($util.toJson($result)))[1][0] )
#set ( $response = {} )
#foreach($mapKey in $res.keySet())
    #set ( $s = $mapKey.split("_") )
    #set ( $camelCase="" )
    #set ( $isFirst=true )
    #foreach($entry in $s)
        #if ( $isFirst )
          #set ( $first = $entry.substring(0,1) )
        #else
          #set ( $first = $entry.substring(0,1).toUpperCase() )
        #end
        #set ( $isFirst=false )
        #set ( $stringLength = $entry.length() )
        #set ( $remaining = $entry.substring(1, $stringLength) )
        #set ( $camelCase = "$camelCase$first$remaining" )
    #end
    $util.qr( $response.put("$camelCase", $res[$mapKey]) )
#end
$utils.toJson($response)

Using Variable Map

Variable map support is limited and does not differentiate numbers and strings data types, please inject them directly if needed.

Will be escaped properly: null, true, and false values.

{
  "version": "2018-05-29",
  "statements":   [
    "UPDATE <name-of-table> set deleted_at=NOW() WHERE id=:ID",
    "SELECT * FROM <name-of-table> WHERE id=:ID and unix_timestamp > $ctx.args.newerThan"
  ],
  variableMap: {
    ":ID": $ctx.args.id,
##    ":TIMESTAMP": $ctx.args.newerThan -- This will be handled as a string!!!
  }
}

Requires

serverless framework
serverless-appsync-plugin
serverless-offline
serverless-dynamodb-local (when using dynamodb resolvers only)
watchman (if Hot-reloading is desactivated it is not required)

Author: Serverless-appsync
Source Code: https://github.com/serverless-appsync/serverless-appsync-simulator
License: MIT License

#serverless #sync #graphql

Arvel Parker

1591611780

How to Find Ulimit For user on Linux

How can I find the correct ulimit values for a user account or process on Linux systems?

For proper operation, we must ensure that the correct ulimit values set after installing various software. The Linux system provides means of restricting the number of resources that can be used. Limits set for each Linux user account. However, system limits are applied separately to each process that is running for that user too. For example, if certain thresholds are too low, the system might not be able to server web pages using Nginx/Apache or PHP/Python app. System resource limits viewed or set with the NA command. Let us see how to use the ulimit that provides control over the resources available to the shell and processes.

#[object object] #[object object] #[object object] #[object object] #[object object] #[object object] #[object object] #[object object] #[object object] #[object object]

German Donnelly

1591993440

MEAN Stack Tutorial MongoDB ExpressJS AngularJS NodeJS

We are going to build a full stack Todo App using the MEAN (MongoDB, ExpressJS, AngularJS and NodeJS). This is the last part of three-post series tutorial.

MEAN Stack tutorial series:

AngularJS tutorial for beginners (Part I)
Creating RESTful APIs with NodeJS and MongoDB Tutorial (Part II)
MEAN Stack Tutorial: MongoDB, ExpressJS, AngularJS and NodeJS (Part III) 👈 you are here
Before completing the app, let’s cover some background about the this stack. If you rather jump to the hands-on part click here to get started.

#[object object] #[object object] #[object object] #[object object] #[object object] #[object object] #[object object] #[object object]

Yoshiko Jones

1598195340

How to configure AWS SES with Postfix MTA

How do I configure Amazon SES With Postfix mail server to send email under a CentOS/RHEL/Fedora/Ubuntu/Debian Linux server?

Amazon Simple Email Service (SES) is a hosted email service for you to send and receive email using your email addresses and domains. Typically SES used for sending bulk email or routing emails without hosting MTA. We can use Perl/Python/PHP APIs to send an email via SES. Another option is to configure Linux or Unix box running Postfix to route all outgoing emails via SES.

» Remove sendmail
» Install postfix
» Configuring postfix for SES
» Test postfix

Procedure to configure AWS SES with Postfix

Before getting started with Amazon SES and Postfix, you need to sign up for AWS, including SES. You need to verify your email address and other settings. Make sure you create a user for SES access and download credentials too.

Step 1 – Uninstall Sendmail if installed

If sendmail installed remove it. Debian/Ubuntu Linux user type the following apt command/apt-get command:

$`` sudo apt --purge remove sendmail

CentOS/RHEL user type the following yum command or dnf command on Fedora/CentOS/RHEL 8.x:

$`` sudo yum remove sendmail

$`` sudo dnf remove sendmail

Sample outputs from CentOS 8 server:

Dependencies resolved.
===============================================================================
 Package           Architecture  Version               Repository         Size
===============================================================================
Removing:
 sendmail          x86_64        8.15.2-32.el8         @AppStream        2.4 M
Removing unused dependencies:
 cyrus-sasl        x86_64        2.1.27-1.el8          @BaseOS           160 k
 procmail          x86_64        3.22-47.el8           @AppStream        369 k

Transaction Summary
===============================================================================
Remove  3 Packages

Freed space: 2.9 M
Is this ok [y/N]: y

#[object object] #[object object] #[object object] #[object object] #[object object] #[object object] #[object object]