Merge pull request #46 from IBM/develop

Release 0.3

.github/PULL_REQUEST_TEMPLATE/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,23 @@
+## Pull Request template
+Please, go through these steps before you submit a PR.
+
+1. Make sure that your PR is not a duplicate.
+2. If not, then make sure that:
+
+     2.1. You have done your changes in a separate branch. Branches should have descriptive names that start with either the `fix/` or `feature/` prefixes. Good examples are: `fix/signin-issue` or `feature/new-model`.
+
+     2.2. You have descriptive commits messages with short titles (first line).
+
+     2.3. You have only one commit (if not, squash them into one commit).
+
+3. **After** these steps, you're ready to open a pull request.
+
+     3.1. Give a descriptive title to your PR.
+
+     3.2. Provide a description of your changes.
+
+     3.3. Put `closes #XXXX` in your comment to auto-close the issue that your PR fixes (if such).
+
+Important: Please review the [CONTRIBUTING.md](../CONTRIBUTING.md) file for detailed contributing guidelines.
+
+*Please remove this template before submitting.*

.gitignore

@@ -7,19 +7,21 @@
 !Readme.md
 !readthedocs.yml
 !setup.py
+!doc_build.sh
 !__init__.py
 !/configs/**
 !/docs/**
 !/miprometheus/**
+!.github/**
 
 # You can be specific with these rules
 __pycache__*
 *.swp
 *.vector_cache
 !.gitignore
 
-# not sure if those are needed
-problems/.DS_Store
-problems/image_text_to_class/.DS_Store
-problems/image_text_to_class/CLEVR_v1.0
-CLEVR_v1.0/
+# Ignore every DS_Store
+**/.DS_Store
+
+# Ignore build directory
+/build/** 

README.md

@@ -51,15 +51,15 @@ The dependencies of MI-prometheus are:
 
    * pytorch (v. 0.4.0)
    * numpy
-   * torchvision (v. 0.2.0)
+   * torchvision
    * torchtext
    * tensorboardx
    * matplotlib 
+   * psutil (enables grid-* to span child processes on MacOS and Ubuntu)
    * PyYAML
    * tqdm
    * nltk
    * h5py
-   * six
    * pyqt5 (v. 5.10.1)
 
 

configs/maes_baselines/default_training.yaml

@@ -5,6 +5,10 @@ training:
         initial_max_sequence_length: 5
     #    must_finish: false
 
+    # Sampler.
+    sampler:
+        name: RandomSampler
+
     # Optimizer parameters:
     optimizer:
         # Exact name of the pytorch optimizer function

configs/vision/alexnet_cifar10.yaml

@@ -5,10 +5,12 @@ training:
     problem:
         name: &name CIFAR10
         batch_size: &b 64
-        index: [0, 40000]
         use_train_data: True
-        padding: &p [0,0,0,0] # ex: (x1, x2, x3, x4) pad last dim by (x1, x2) and 2nd to last by (x3, x4)
-        up_scaling: &scale True # if up_scale true the image is resized to 224 x 224
+        resize: [224, 224]
+    # Use sampler that operates on a subset.
+    sampler:
+        name: SubsetRandomSampler
+        indices: [0, 45000]
     # optimizer parameters:
     optimizer:
         name: Adam
@@ -22,19 +24,20 @@ validation:
     problem:
         name: *name
         batch_size: *b
-        index: [40000, 49999]
         use_train_data: True # True because we are splitting the training set to: validation and training
-        padding: *p
-        up_scaling: *scale
+        resize: [224, 224]
+    # Use sampler that operates on a subset.
+    sampler:
+        name: SubsetRandomSampler
+        indices: [45000, 50000]
 
 # Problem parameters:
 testing:
     problem:
         name: *name
         batch_size: *b
         use_train_data: False
-        padding: *p
-        up_scaling: *scale
+        resize: [224, 224]
 
 # Model parameters:
 model:

configs/vision/alexnet_mnist.yaml

@@ -5,17 +5,19 @@ training:
     problem:
         name: &name MNIST
         batch_size: &b 64
-        index: [0, 54999]
         use_train_data: True
-        padding: &p [0,0,0,0] # ex: (x1, x2, x3, x4) pad last dim by (x1, x2) and 2nd to last by (x3, x4)
-        up_scaling: &scale True # if up_scale true, the image is resized to 224 x 224
+        resize: [224, 224]
+    # Use sampler that operates on a subset.
+    sampler:
+        name: SubsetRandomSampler
+        indices: [0, 55000]
     # optimizer parameters:
     optimizer:
         name: Adam
         lr: 0.01
     # settings parameters
     terminal_conditions:
-        loss_stop: 1.0e-5
+        loss_stop: 1.0e-2
         episode_limit: 50000
         epochs_limit: 10
 
@@ -24,19 +26,16 @@ validation:
     problem:
         name: *name
         batch_size: *b
-        index: [54999, 59999]
         use_train_data: True  # True because we are splitting the training set to: validation and training
-        padding: *p
-        up_scaling: *scale
+        resize: [224, 224]
 
 # Problem parameters:
 testing:
     problem:
         name: *name
         batch_size: *b
         use_train_data: False
-        padding: *p
-        up_scaling: *scale
+        resize: [224, 224]
 
 # Model parameters:
 model:

configs/vision/grid_trainer_mnist.yaml

@@ -0,0 +1,43 @@
+grid_tasks:
+    -
+        default_configs: configs/vision/lenet5_mnist.yaml
+    -
+        default_configs: configs/vision/simplecnn_mnist.yaml
+
+# Set exactly the same experiment conditions for the 2 tasks.
+grid_overwrite:
+    training:
+        problem:
+            batch_size: &b 1000
+        sampler:
+            name: SubsetRandomSampler
+            indices: [0, 55000]
+       # Set the same optimizer parameters.
+        optimizer:
+            name: Adam
+            lr: 0.01
+        # Set the same terminal conditions.
+        terminal_conditions:
+            loss_stop: 4.0e-2
+            episode_limit: 10000
+            epoch_limit: 10
+
+    # Problem parameters:
+    validation:
+        problem:
+            batch_size: *b
+        sampler:
+            name: SubsetRandomSampler
+            indices: [55000, 60000]
+
+    testing:
+        problem:
+            batch_size: *b
+
+grid_settings:
+    # Set number of repetitions of each experiments.
+    experiment_repetitions: 5
+    # Set number of concurrent running experiments.
+    max_concurrent_runs: 4
+    # Set trainer.
+    trainer: mip-online-trainer

configs/vision/lenet5_mnist.yaml

@@ -0,0 +1,48 @@
+# Training parameters:
+training:
+  problem:
+    name: &name MNIST
+    batch_size: &b 64
+    use_train_data: True
+    data_folder: &folder '~/data/mnist'
+    resize: [32, 32]
+  # Use sampler that operates on a subset.
+  sampler:
+    name: SubsetRandomSampler
+    indices: [0, 55000]
+  # optimizer parameters:
+  optimizer:
+    name: Adam
+    lr: 0.01
+  # settings parameters
+  terminal_conditions:
+    loss_stop: 1.0e-2
+    episode_limit: 10000
+    epoch_limit: 10
+
+# Validation parameters:
+validation:
+  #partial_validation_interval: 100
+  problem:
+    name: *name 
+    batch_size: *b
+    use_train_data: True  # True because we are splitting the training set to: validation and training
+    data_folder: *folder
+    resize: [32, 32]
+  # Use sampler that operates on a subset.
+  sampler:
+    name: SubsetRandomSampler
+    indices: [55000, 60000]
+
+# Testing parameters:
+testing:
+  problem:
+    name: *name
+    batch_size: *b
+    use_train_data: False
+    data_folder: *folder
+    resize: [32, 32]
+
+# Model parameters:
+model:
+    name: LeNet5

configs/vision/simplecnn_cifar10.yaml

@@ -5,10 +5,11 @@ training:
     problem:
         name: &name CIFAR10
         batch_size: &b 64
-        index: [0, 40000]
         use_train_data: True
-        padding: &p [0,0,0,0] # ex: (x1, x2, x3, x4) pad last dim by (x1, x2) and 2nd to last by (x3, x4)
-        up_scaling: &scale False # if up_scale true the image is resized to 224 x 224
+    # Use sampler that operates on a subset.
+    sampler:
+        name: SubsetRandomSampler
+        indices: [0, 45000]
     # optimizer parameters:
     optimizer:
         name: Adam
@@ -22,19 +23,18 @@ validation:
     problem:
         name: *name
         batch_size: *b
-        index: [40000, 49999]
         use_train_data: True # True because we are splitting the training set to: validation and training
-        padding: *p
-        up_scaling: *scale
+    # Use sampler that operates on a subset.
+    sampler:
+        name: SubsetRandomSampler
+        indices: [45000, 50000]
 
 # Problem parameters:
 testing:
     problem:
         name: *name
         batch_size: *b
         use_train_data: False
-        padding: *p
-        up_scaling: *scale
 
 # Model parameters:
 model:

configs/vision/simplecnn_mnist.yaml

@@ -5,38 +5,48 @@ training:
     problem:
         name: &name MNIST
         batch_size: &b 64
-        index: [0, 54999]
+        data_folder: &folder '~/data/mnist'
         use_train_data: True
-        padding: &p [0,0,0,0] # ex: (x1, x2, x3, x4) pad last dim by (x1, x2) and 2nd to last by (x3, x4)
-        up_scaling: &scale False # if up_scale true, the image is resized to 224 x 224
+        resize: [32, 32]
+    sampler:
+        name: SubsetRandomSampler
+        indices: [0, 55000]
+        #indices: ~/data/mnist/split_a.txt 
     # optimizer parameters:
     optimizer:
         name: Adam
         lr: 0.01
     # settings parameters
     terminal_conditions:
-        loss_stop: 1.0e-5
-        episode_limit: 50000
-        epochs_limit: 10
+        loss_stop: 1.0e-2
+        episode_limit: 1000
+        epoch_limit: 1
 
 # Problem parameters:
 validation:
     problem:
         name: *name
         batch_size: *b
-        index: [54999, 59999]
+        data_folder: *folder
         use_train_data: True  # True because we are splitting the training set to: validation and training
-        padding: *p
-        up_scaling: *scale
+        resize: [32, 32]
+    sampler:
+        name: SubsetRandomSampler
+        indices: [55000, 60000]
+        #indices: ~/data/mnist/split_b.txt 
+    #dataloader:
+    #    drop_last: True
 
 # Problem parameters:
 testing:
+    #seed_numpy: 4354
+    #seed_torch: 2452
     problem:
         name: *name
         batch_size: *b
+        data_folder: *folder
         use_train_data: False
-        padding: *p
-        up_scaling: *scale
+        resize: [32, 32]
 
 
 # Model parameters:

doc_build.sh

@@ -0,0 +1,14 @@
+#!/usr/bin/env bash
+
+cd docs
+rm -rf build
+
+# create html pages
+sphinx-build -b html source build
+make html
+
+# open web browser(s) to master table of content
+if which firefox
+then
+    firefox build/index.html
+fi