First code commit

README.md

@@ -1 +1,37 @@
-# training-engine
+# ALOHA Training Engine
+
+This is stand-alone Training Engine module  ALOHA.eu project. http://aloha-h2020.eu
+
+## Setup
+Training Engine was developed using Python 3.6, TensorFlow 1.12, PyTorch 1.0 and was tested under Windows 10
+
+``pip install -r requirements.txt``
+
+## Data Standardization
+Data standardization description described in [data_standardization](data_standardization)
+
+## Architecture deployment
+to be added
+
+## Training Configuration
+to be added
+
+
+## Supported Features
+- [x] training/inference TensorFlow support
+- [x] classification task support
+- [x] segmentation task support
+
+## In development:
+- [ ] training/inference PyTorch support
+- [ ] detection task support
+- [ ] tracking task support
+- [ ] GAN architecture support
+- [ ] hyperparameter optimization
+- [ ] fast/full training mode
+- [ ] accuracy prediction metric
+- [ ] transfer leraning support
+
+
+# Acknowledgements
+This work was carried out by the European Union\`s Horizon H2020 Research and Innovation programme under grant agreement No. 780788

configs/config.py

@@ -0,0 +1,102 @@
+#!/usr/bin/env python
+# -----------------------------------------------------------------------------
+# Copyright (C) Software Competence Center Hagenberg GmbH (SCCH)
+# All rights reserved.
+# -----------------------------------------------------------------------------
+# This document contains proprietary information belonging to SCCH.
+# Passing on and copying of this document, use and communication of its
+# contents is not permitted without prior written authorization.
+# -----------------------------------------------------------------------------
+# Created on : 11/16/2017 11:07 $
+# by : shepeleva $
+# SVN : $
+#
+
+# --- imports -----------------------------------------------------------------
+
+import argparse
+
+
+class ConfigFlags:
+    def __init__(self):
+        """
+
+        """
+        parser = argparse.ArgumentParser(description='Tensorflow DL training pipeline')
+        # mandatory variables
+        parser.add_argument('--net', help='Network',
+                            type=str, default='ConvNet')
+        parser.add_argument('--mode', help='training / test',
+                            type=str, default='training')
+        parser.add_argument('--data_set', help='Name/ID of the used data set',
+                            type=str, default='MNIST')
+        parser.add_argument('--data_dir', help='Learning data directory',
+                            type=str, default='')
+        parser.add_argument('--data_file', help='Data file',
+                            type=str)
+        parser.add_argument('--checkpoint_dir', help='Checkpoint directory',
+                            type=str, default='ckpnt_dir')
+        parser.add_argument('--trainlog_dir', help='Train records directory',
+                            type=str, default='tr_log')
+        parser.add_argument('--lr', help='Learning rate',
+                            type=float, default=0.001)
+        parser.add_argument('--lr_decay', help='Learning rate decay rate',
+                            type=float, default=0.0005)
+        parser.add_argument('--ref_steps', help='Refinement Steps',
+                            type=int, default=5)
+        parser.add_argument('--ref_patience', help='Refinement Patience',
+                            type=int, default=200)
+        parser.add_argument('--batch_size', help='Batch size',
+                            type=int, default=32)
+        parser.add_argument('--num_epochs', help='Number of epochs',
+                            type=int, default=100)
+        parser.add_argument('--loss', help='Loss function',
+                            type=str, default='mse')
+        parser.add_argument('--optimizer', help='Optimizers: \n\t adam - adam optimizer '
+                                                '\n\t gradient - gradient descent '
+                                                '\n\t proximalgrad - proximal gradient descent ',
+                            default='adam')
+        parser.add_argument('--dropout', help='Dropout Rate to use',
+                            type=float, default=0.25)
+        parser.add_argument('--tb_record', help='Tensorboard records on/off',
+                            type=bool, default=True)
+        parser.add_argument('--darkon_record', help='Darkon tool',
+                            type=bool, default=False)
+        parser.add_argument('--gradcam_record', help='GradCam tool',
+                            type=bool, default=False)
+        parser.add_argument('--gradcam_layers', help='Number of inspected convolution layers',
+                            type=int, default=1)
+        parser.add_argument('--gradcam_layers_max', help='Number of convolution layers',
+                            type=int, default=3)
+        parser.add_argument('--mi_record', help='MI tool',
+                            type=bool, default=False)
+        parser.add_argument('--gpu_load', help='GPU load percentage [0.1 : 1]',
+                            type=int, default=0.8)
+        # optional variables
+        parser.add_argument('--image_size', help='Image size',
+                            type=list, default=[128, 128, 1])
+        parser.add_argument('--num_classes', help='Number of labels',
+                            type=int, default=2)
+        parser.add_argument('--num_filters', help='Number of filters',
+                            type=int)
+        # todo actually implement following parameters in NetRunner
+        parser.add_argument('--filter_size', help='Filter size',
+                            type=int)
+        parser.add_argument('--pool_size', help='Pool size',
+                            type=int)
+        parser.add_argument('--stride_size', help='stride size',
+                            type=int)
+        parser.add_argument('--nonlin', help='Nonlinearity',
+                            type=str)
+        parser.add_argument('--upconv', help='Up convolution type: upconv or upsampling',
+                            type=str)
+        parser.add_argument('--multi_task', help="multiple task, i.e. different loss functions",
+                            type=bool)
+        self.args = parser.parse_args()
+
+    def return_flags(self):
+        """
+
+        :return:
+        """
+        return self.args

configs/config_ConvNet.py

@@ -0,0 +1,63 @@
+#!/usr/bin/env python
+# -----------------------------------------------------------------------------
+# Copyright (C) Software Competence Center Hagenberg GmbH (SCCH)
+# All rights reserved.
+# -----------------------------------------------------------------------------
+# This document contains proprietary information belonging to SCCH.
+# Passing on and copying of this document, use and communication of its
+# contents is not permitted without prior written authorization.
+# -----------------------------------------------------------------------------
+# Created on : 01/08/2018 16:29 $
+# by : shepeleva $
+# SVN : $
+#
+
+# --- imports -----------------------------------------------------------------
+
+from configs.config import ConfigFlags
+
+
+def load_config():
+    config = ConfigFlags().return_flags()
+
+    config.net = 'ConvNet'
+    config.training_mode = False
+    config.data_set = 'MNIST'
+    config.image_size = [28, 28, 1]
+
+    config.lr = 0.001
+    config.lr_decay = 0.1
+    config.ref_steps = 10
+    config.ref_patience = 1
+    config.batch_size = 32
+    config.num_epochs = 1000
+    config.loss = 'mse'
+    config.optimizer = 'adam'
+    config.gradcam_record = True
+    config.gradcam_layers = 6
+    config.gradcam_layers_max = 6
+    config.mi_record = False
+    config.gpu_load = 0.8
+    config.num_classes = 10
+    config.class_labels = [i for i in range(9)]
+    config.num_filters = 16
+    config.upconv = 'upconv'
+    config.nonlin = 'relu'
+    config.task_type = 'classification'
+    config.accuracy = 'mse'
+    config.augmentation = {'flip_hor': False,
+                           'flip_vert': False}
+    config.data_split = 0.7
+    config.long_summary = True
+    config.trainable_layers = 'all'
+    config.normalize = True
+    config.zero_center = True
+    config.dropout = 0.4
+    config.chpnt2load = ''
+    config.multi_task = False
+    config.cross_val = 1
+    config.framework = 'tensorflow'
+    config.experiment_path = None
+
+
+    return config

data_standardization/README.md

@@ -0,0 +1,20 @@
+# data_standardization
+
+Description of custom dataset representation accepted by ALOHA tool
+
+Current support of data input into ALOHA tool is following:
+- by picking `dataset`
+    >allowed entries: MNIST, CIFAR10, CIFAR100, <custom>
+
+    if `<custom>` is picked, user have two options:
+    - from `data_file`
+
+        > supported file formats: JSON, TXT
+
+    - from `data_folder`
+
+        > multiple file loading for classification and segmentation case
+
+For more detailed description check corresponding folder
+
+##### Important. Description provided in standardization files are not yet final. Changes my apply during development phase.

data_standardization/json/json_standardization.md

@@ -0,0 +1,113 @@
+# JSON format standardization
+
+## General information
+Rules on dataset description on JSON format.
+This representation supports 3 task scenarios:
+- Classification
+- Segmentation
+- Detection
+- *In development: Traking, Multi-task*
+
+This standardization describes following input cases:
+- IC1: single folder with images
+- IC2: multiple folders with images
+- IC3: single video
+- IC4: multiple videos
+
+## Data loading to ALOHA toolflow
+By using JSON format standardization, please, use parameter `data_file` in the configuration and pick JSON file, which describes your dataset.
+
+Please, be aware that all source locations in the JSON format standardization should contain ABSOLUTE path.
+
+
+## Detailed standardization description
+The standardization should be described as a dictionary of `<key>: <value>` pairs.
+
+Allowed list of `key` values:
+
+- `name` - dataset name
+    >str value, describes dataset name, Ex.: 'random_set'
+- `width` - width of the frame/image
+    >set value only if all frames/images of the same size, otherwise put 0
+- `height` - height of the frame/image
+    >set value only if all frames/images of the same size, otherwise put 0
+- `source` - source of the information (`video` or `image`)
+    >describes which source will be used, avaliable values:  `video` or `image`
+- `source_path` - absolute path to the source
+    >absolute or relative to the .json file path to the data source
+- `mask_path` - absolute path to masks if presented
+    >absolute or relative to the .json file path to the masks (only for Segmentation task)
+- `data_size` - number of frames per video / number of images in dataset
+    >depending on IC representaiton this value coud difer use following example: <br>
+    for IC1 or IC3: 2340 - means that dataset represented by one folder with 2340 images in total OR by one video with 2340 frames in total<br>
+    for IC2 or IC4: [345, 654, 123] - means that dataset represented by 3 folders with 345, 654 and 123 images accordingly OR by 3 videos with 345, 654 and 123 frames accordingly
+- `meta`- detailed information for specific task
+    >in this section every frame/image should get detailed description accordingly to the task
+
+###  `meta` for Classification
+List of following dictionaries:
+- `frame` - frame/image id
+    >for IC1: image_name<br>
+    for IC2: [image_folder, image]<br>
+    for IC3: frame_id<br>
+    for IC4: [video_file_name, frame_id]
+- `width`
+    >width of particular frame, if global width is not setted up
+- `height`
+    >height of particular frame, if global height is not setted up
+- `mask` - mask presence
+    >for this task should be equal to null
+- `objects` - list of the descriptions of objects which should be classified in current frame
+    - `object_class` - object class indentification
+      >integer or one-hot labeling
+    - `bb` - bounding box
+      >for this task should be equal to null
+    - `polygon` - polygon area
+      >for this task should be equal to null
+
+###  `meta` for Segmentation
+List of following dictionaries:
+- `frame` - frame/image id
+    >for IC1: image_name<br>
+    for IC2: [image_folder, image]<br>
+    for IC3: frame_id<br>
+    for IC4: [video_file_name, frame_id]
+- `width`
+    >width of particular frame, if global width is not setted up
+- `height`
+    >height of particular frame, if global height is not setted up
+- `mask`- mask presence, video/image representation might be independent from frame representaiton
+    >for IC1: image_name<br>
+    for IC2: [image_folder, image]<br>
+    for IC3: frame_id<br>
+    for IC4: [video_file_name, frame_id]
+- `objects` - list of the descriptions of objects which should be classified in current frame
+    - `object_class` - object class indentification
+      >integer or one-hot labeling
+    - `bb` - bounding box
+      >for this task should be equal to null
+    - `polygon` - polygon area
+      >list of points describing polygon
+
+### `meta` for Detection
+List of following dictionaries:
+- `frame` - frame/image id
+    >for IC1: image_name<br>
+    for IC2: [image_folder, image]<br>
+    for IC3: frame_id<br>
+    for IC4: [video_file_name, frame_id]
+- `width`
+    >width of particular frame, if global width is not setted up
+- `height`
+    >height of particular frame, if global height is not setted up
+- `mask` - mask presence
+    >for this task should be equal to null
+- `objects` - list of the descriptions of objects which should be classified in current frame
+    - `object_class` - object class indentification
+      >integer or one-hot labeling
+    - `bb` - object boundaries
+      >should be described as [Xmin, Ymin, Xmax, Ymax], where<br>
+      (Xmin, Ymin) pair specifying the lower-left corner<br>
+      (Xmax, Ymax) pair specifying the upper-right corner
+    - `polygon` - polygon area
+      >list of points describing polygon

data_standardization/txt/txt_standardization.md

@@ -0,0 +1,41 @@
+# TXT format standardization
+
+## General information
+Rules on dataset description on TXT format.
+This representation supports 2 task scenarios:
+- Classification
+- Segmentation
+- Detection
+- *In development: Traking, Multi-task*
+
+This standardization describes following input cases:
+- IC1: single TXT file
+
+## Data loading to ALOHA toolflow
+- IC1 case: please, use parameter `data_file` in the configuration and pick TXT file, which describes your dataset.
+
+Please, be aware that all source locations in the TXT format standardization should contain ABSOLUTE path.
+
+## Detailed standardization description
+The standardization should be described as a list of paticular fileds separated by space depending on the task.
+
+### Classification
+File structure: `<image_file> <label>`
+- `image_file` - absolute path to the image
+- `label` - label description, integer or one-hot encoded
+  >for integer description the count start from 0
+
+### Segmentaiton
+File structure: `<image_file> <mask_file>`
+- `image_file` - absolute path to the image
+- `mask_file` -  absolute path to the corresponding mask
+  >segmentation areas should be setted to the integer values starting from 1 (0 - backgound) for a required number of classes in dataset
+
+### Detection
+File structure: `<image_file> <bounding_box>`
+- `image_file` - absolute path to the image
+- `bounding_box` - object boundaries
+  >should be described as [Xmin, Ymin, Xmax, Ymax], where
+  (Xmin, Ymin) pair specifying the lower-left corner<br>
+  (Xmax, Ymax) pair specifying the upper-right corner<br>
+  for more than one object should be represented as list if lists