final_commission

callous-youth · Sep 27, 2020 · 903769e · 903769e
1 parent 0375e6c
commit 903769e
Show file tree

Hide file tree

Showing 25 changed files with 476 additions and 145 deletions.
diff --git a/.travis.yml b/.travis.yml
@@ -15,13 +15,13 @@ script:
     - coverage run  -p test_meta_init.py --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=200 --T=1 --meta-lr=0.001 --lr=0.1 --method=MetaInit --inner_method=Simple --outer_method=Simple --learn_lr=True --logdir=../tmp/ --clip_value=10.0 --first_order=True --print-interval=100 --save_interval=100 --use_t=True --Notes=toolkit_TNet
     - coverage run  -p test_meta_init.py --mode=test --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=200 --T=1 --meta-lr=0.001 --lr=0.1 --method=MetaInit --inner_method=Simple --outer_method=Simple --logdir=../tmp/ --expdir=../tmp/  --clip_value=10.0 --iterations_to_test=100 --print-interval=100 --save_interval=100 --use_t=True --Notes=toolkit_TNet
     - coverage run  -p test_meta_init.py --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=10 --T=2 --meta-lr=0.001 --lr=0.1 --method=MetaInit --inner_method=Simple --outer_method=Simple --logdir=../tmp/ --expdir=../tmp/  --print-interval=100 --first_order=True --save_interval=100 --use_warp=True --Notes=toolkit_warp
-    - coverage run -p test_meta_repr.py  --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=100 --T=2 --meta-lr=0.001 --lr=0.1 --method=MetaRepr --inner_method=Trad --outer_method=Reverse --inner_opt=Adam --outer_opt=Adam  --clip_value=10.0 --logdir=../tmp/ --print-interval=100 --truncate_iter=1 --save_interval=100 --Notes=toolkit_test_TRHG
-    - coverage run -p test_meta_repr.py  --mode=test --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=100 --T=2 --meta-lr=0.001 --lr=0.1 --method=MetaRepr --inner_method=Trad --outer_method=Reverse --inner_opt=Adam --outer_opt=Adam  --clip_value=10.0 --logdir=../tmp/ --expdir=../tmp/ --print-interval=100 --truncate_iter=1 --save_interval=100 --Notes=toolkit_test_TRHG
-    - coverage run -p test_meta_repr.py  --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=10 --T=1 --meta-lr=0.001 --lr=0.1 --method=MetaRepr --inner_method=Aggr --learn_alpha_itr=True --learn_st=True --outer_method=Reverse --inner_opt=Momentum --outer_opt=Momentum --logdir=../tmp/ --alpha=0.7 --print-interval=100 --save_interval=100 --Notes=toolkit_test_BDA
-    - coverage run -p test_meta_repr.py  --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=10 --T=1 --meta-lr=0.001 --lr=0.1 --method=MetaRepr --inner_method=Aggr --learn_alpha=True --learn_st=True --outer_method=Reverse --inner_opt=Momentum --outer_opt=Momentum --logdir=../tmp/ --alpha=0.3 --print-interval=100 --save_interval=100 --Notes=toolkit_test_BDA_alpha_iter
-    - coverage run -p test_meta_repr.py  --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=2 --meta_train_iterations=10 --T=1 --meta-lr=0.001 --lr=0.1 --method=MetaRepr --inner_method=Aggr --outer_method=Darts --alpha=0.7 --logdir=../tmp/ --print-interval=100 --save_interval=100 --Notes=toolkit_test_DARTS
-    - coverage run -p test_meta_repr.py  --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=2 --meta_train_iterations=10 --T=1 --meta-lr=0.001 --lr=0.1 --method=MetaRepr --inner_method=Aggr --outer_method=Darts --alpha=0.7 --learn_alpha_itr=True --logdir=../tmp/ --print-interval=100 --save_interval=100 --Notes=toolkit_test_DARTS_alpha_iter
-    - coverage run -p test_meta_repr.py  --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=10 --T=1 --meta-lr=0.001 --lr=0.1 --method=MetaRepr --inner_method=Trad --outer_method=Implicit --inner_opt=SGD --outer_opt=Adam --logdir=../tmp/ --print-interval=100 --save_interval=100 --Notes=toolkit_test_Implicit
+    - coverage run -p test_meta_feat.py  --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=100 --T=2 --meta-lr=0.001 --lr=0.1 --method=MetaFeat --inner_method=Trad --outer_method=Reverse --inner_opt=Adam --outer_opt=Adam  --clip_value=10.0 --logdir=../tmp/ --print-interval=100 --truncate_iter=1 --save_interval=100 --Notes=toolkit_test_TRHG
+    - coverage run -p test_meta_feat.py  --mode=test --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=100 --T=2 --meta-lr=0.001 --lr=0.1 --method=MetaFeat --inner_method=Trad --outer_method=Reverse --inner_opt=Adam --outer_opt=Adam  --clip_value=10.0 --logdir=../tmp/ --expdir=../tmp/ --print-interval=100 --truncate_iter=1 --save_interval=100 --Notes=toolkit_test_TRHG
+    - coverage run -p test_meta_feat.py  --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=10 --T=1 --meta-lr=0.001 --lr=0.1 --method=MetaFeat --inner_method=Aggr --learn_alpha_itr=True --learn_st=True --outer_method=Reverse --inner_opt=Momentum --outer_opt=Momentum --logdir=../tmp/ --alpha=0.7 --print-interval=100 --save_interval=100 --Notes=toolkit_test_BDA
+    - coverage run -p test_meta_feat.py  --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=10 --T=1 --meta-lr=0.001 --lr=0.1 --method=MetaFeat --inner_method=Aggr --learn_alpha=True --learn_st=True --outer_method=Reverse --inner_opt=Momentum --outer_opt=Momentum --logdir=../tmp/ --alpha=0.3 --print-interval=100 --save_interval=100 --Notes=toolkit_test_BDA_alpha_iter
+    - coverage run -p test_meta_feat.py  --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=2 --meta_train_iterations=10 --T=1 --meta-lr=0.001 --lr=0.1 --method=MetaFeat --inner_method=Aggr --outer_method=Darts --alpha=0.7 --logdir=../tmp/ --print-interval=100 --save_interval=100 --Notes=toolkit_test_DARTS
+    - coverage run -p test_meta_feat.py  --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=2 --meta_train_iterations=10 --T=1 --meta-lr=0.001 --lr=0.1 --method=MetaFeat --inner_method=Aggr --outer_method=Darts --alpha=0.7 --learn_alpha_itr=True --logdir=../tmp/ --print-interval=100 --save_interval=100 --Notes=toolkit_test_DARTS_alpha_iter
+    - coverage run -p test_meta_feat.py  --mode=train --dataset=omniglot --classes=5 --examples_train=1 --examples_test=1 --meta_batch_size=1 --meta_train_iterations=10 --T=1 --meta-lr=0.001 --lr=0.1 --method=MetaFeat --inner_method=Trad --outer_method=Implicit --inner_opt=SGD --outer_opt=Adam --logdir=../tmp/ --print-interval=100 --save_interval=100 --Notes=toolkit_test_Implicit
     - coverage combine
 
 after_success:

diff --git a/README.md b/README.md
@@ -11,7 +11,7 @@
 
 BOML is a modularized optimization library that unifies several ML algorithms into a common bilevel optimization framework. It provides interfaces to implement popular bilevel optimization algorithms, so that you could quickly build your own meta learning neural network and test its performance.
 
-ReadMe.md file contains brief introduction to implement meta-initialization-based and meta-feature-based methods in few-shot learning field. Except for algorithms which have been proposed, various combinations of lower level and upper level strategies are available. 
+ReadMe.md contains brief introduction to implement meta-initialization-based and meta-feature-based methods in few-shot classification field. Except for algorithms which have been proposed, various combinations of lower level and upper level strategies are available. 
 
 ## Meta Learning 
 
@@ -54,7 +54,7 @@ ex = boml.BOMLExperiment(dataset)
 
 ```
 
-### Build network structure and define hyperparameters
+### Build network structure and define parameters for meta-learner and base-learner
 ```python
 boml_ho = boml.BOMLOptimizer(
     method="MetaInit", inner_method="Simple", outer_method="Simple"
@@ -76,7 +76,7 @@ inner_grad = boml_ho.ll_problem(
 ```
 ### Define UL objectives and UL calculation process
 ```python
-loss_outer = utils.cross_entropy(pred=ex.model.re_forward(ex.x_).out, label=ex.y_)
+loss_outer = utils.cross_entropy(pred=ex.model.re_forward(ex.x_).out, label=ex.y_) # loss function
 boml_ho.ul_problem(
     outer_objective=loss_outer,
     meta_learning_rate=args.meta_lr,
@@ -108,14 +108,14 @@ with tf.Session() as sess:
 ## Related Methods 
  - [Hyperparameter optimization with approximate gradient(HOAG)](https://arxiv.org/abs/1602.02355)
  - [Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks(MAML)](https://arxiv.org/abs/1703.03400)
- - [On First-Order Meta-Learning Algorithms(FOMAML)](https://arxiv.org/abs/1803.02999)
+ - [On First-Order Meta-Learning Algorithms(FMAML)](https://arxiv.org/abs/1703.03400)
  - [Meta-SGD: Learning to Learn Quickly for Few-Shot Learning(Meta-SGD)](https://arxiv.org/pdf/1707.09835.pdf)
  - [Bilevel Programming for Hyperparameter Optimization and Meta-Learning(RHG)](http://export.arxiv.org/pdf/1806.04910)
  - [Truncated Back-propagation for Bilevel Optimization(TG)](https://arxiv.org/pdf/1810.10667.pdf)
  - [Gradient-Based Meta-Learning with Learned Layerwise Metric and Subspace(MT-net)](http://proceedings.mlr.press/v80/lee18a/lee18a.pdf)
  - [Meta-Learning with warped gradient Descent(WarpGrad))](https://arxiv.org/abs/1909.00025)
  - [DARTS: Differentiable Architecture Search(DARTS)](https://arxiv.org/pdf/1806.09055.pdf)
- - [A Generic First-Order Algorithmic Framework for Bi-Level Programming Beyond Lower-Level Singleton(BA)](https://arxiv.org/pdf/2006.04045.pdf)
+ - [A Generic First-Order Algorithmic Framework for Bi-Level Programming Beyond Lower-Level Singleton(BDA)](https://arxiv.org/pdf/2006.04045.pdf)
 
 
 

diff --git a/boml/__init__.py b/boml/__init__.py
@@ -1,3 +1,8 @@
+try:
+    import tensorflow as tf
+except ImportError:
+    print("No compatible version of TensorFlow has been installed!")
+
 from boml.load_data.experiment import BOMLExperiment
 from boml.load_data import (
     meta_mini_imagenet,
@@ -12,17 +17,17 @@
     BOMLNetMiniMetaInitV1,
 )
 from boml.setup_model.meta_repr_v1 import (
-    BOMLNetMetaReprV1,
-    BOMLNetMiniMetaReprV1,
-    BOMLNetOmniglotMetaReprV1,
+    BOMLNetMetaFeatV1,
+    BOMLNetMiniMetaFeatV1,
+    BOMLNetOmniglotMetaFeatV1,
 )
 from boml.setup_model.meta_init_v2 import (
     BOMLNetOmniglotMetaInitV2,
     BOMLNetMiniMetaInitV2,
 )
 from boml.setup_model.meta_repr_v2 import (
-    BOMLNetOmniglotMetaReprV2,
-    BOMLNetMiniMetaReprV2,
+    BOMLNetOmniglotMetaFeatV2,
+    BOMLNetMiniMetaFeatV2,
 )
 from boml.setup_model.network import BOMLNet
 from boml.setup_model.feedforward import BOMLNetFeedForward
@@ -31,3 +36,5 @@
 from boml.extension import get_outerparameter
 from boml import utils
 from boml import extension
+
+
diff --git a/boml/boml_optimizer/optimizer.py b/boml/boml_optimizer/optimizer.py
@@ -20,7 +20,7 @@
 
 class BOMLOptimizer(object):
     """
-    Wrapper for performing gradient-based metaparameter optimization
+    Wrapper for performing gradient-based parameter optimization
     """
 
     def __init__(
@@ -32,26 +32,27 @@ def __init__(
         experiments=[],
     ):
         """
-        BMLHOptimizer is the main class in `pybml`, which takes responsibility for
+        BMLHOptimizer is the main class in `boml`, which takes responsibility for
         the whole process of model construnction and back propagation.
 
-        :param Method: define basic method for following training process, it should be included in ['MetaInit', 'MetaRepr'],
-         'MetaInit' type includes meta-initialization-based methods like 'MAML, FOMAML, MT-net, WarpGrad';
-         'MetaRepr' type includes meta-feature-based methods like
-         'BDA, RHG, Truncated, Implicit HG, DARTS';
-        :param inner_method: method chosen for solving LLproblem, including ['Trad' ,'Simple', 'Aggr'], 'MetaRepr' type choose
-        either 'Trad' for traditional optimization strategies or 'Aggr' for Gradient Aggragation optimization 'MetaInit' type
-        should choose 'Simple', and set specific parameters for detailed method choices like FOMAML or TNet.
-        :param outer_method: method chosen for solving LLproblem, including ['Reverse' ,'Simple', 'Forward', 'Implcit'],
-        'MetaInit' type should choose Simple, and set specific parameters for detailed method choices like FOMAML
+        :param method: define basic method for following training process, it should be included in
+        ['MetaInit', 'MetaFeat'], 'MetaInit' type includes meta-initialization-based methods like
+        'MAML, FOMAML, MT-net, WarpGrad'; 'MetaFeat' type includes meta-feature-based methods like
+         'BDA, RHG, TRHG, Implicit HG, DARTS';
+        :param inner_method: method chosen for solving LLproblem, including ['Trad' ,'Simple', 'Aggr'],
+        'MetaFeat' type choose either 'Trad' for traditional optimization strategies or 'Aggr' for
+        Gradient Aggregation optimization 'MetaInit' type
+        should choose 'Simple', and set specific parameters for detailed method choices like FMAML or TNet.
+        :param outer_method: method chosen for solving LLproblem, including ['Reverse' ,'Simple', 'Forward', 'Implicit'],
+        'MetaInit' type should choose Simple, and set specific parameters for detailed method choices like FMAML
         :param truncate_iter: specific parameter for Truncated Reverse method, defining number of iterations to truncate
          in the Back propagation process
         :param experiments: list of experiment objects that has already been initialized
         :return :an initialized instance of BMLHOptimizer
         """
-        assert method in ("MetaRepr", "MetaInit"), (
-            "initialize method arguement, should be in list \an [MetaRepr,MetaInitl] "
-            "MetaRepr based methods include [BDA,FHG,RHG,TRHG],"
+        assert method in ("MetaFeat", "MetaInit"), (
+            "initialize method arguement, should be in list \an [MetaFeat,MetaInitl] "
+            "MetaFeat based methods include [BDA,FHG,RHG,TRHG],"
             "HperOptim based methods include [MAML,FOMAML,MSGD]"
         )
         self._method = method
@@ -172,7 +173,7 @@ def base_learner(
         :param _input: orginal input for neural network construction of task-specific module
         :param meta_learner: returned value of Meta_model function, which is a instance of BMLNet or its child classes
         :param name: name for Base model modules used for BMLNet initialization
-        :param weights_initializer: initializer function for task_specific network, called by 'MetaRepr' method
+        :param weights_initializer: initializer function for task_specific network, called by 'MetaFeat' method
         :return: task-specific model part
         """
         if self.method == "MetaInit":
@@ -194,7 +195,7 @@ def base_learner(
                 outer_method=self.outer_method,
                 use_warp=meta_learner.use_warp,
             )
-        elif self.method == "MetaRepr":
+        elif self.method == "MetaFeat":
             base_learner = getattr(boml_networks, "BOMLNetFeedForward")(
                 _input=_input,
                 dims=self.data_set.train.dim_target,
@@ -203,7 +204,7 @@ def base_learner(
             )
         else:
             print(
-                "initialize method arguement, should be in list \an [MetaRepr,MetaInitl]"
+                "initialize method arguement, should be in list \an [MetaFeat,MetaInitl]"
             )
             raise AssertionError
         return base_learner
@@ -214,6 +215,7 @@ def ll_problem(
         learning_rate,
         T,
         var_list=None,
+        experiment=None,
         inner_objective_optimizer="SGD",
         outer_objective=None,
         learn_lr=False,
@@ -226,7 +228,6 @@ def ll_problem(
         first_order=False,
         loss_func=utils.cross_entropy,
         momentum=0.5,
-        experiment=None,
         **inner_kargs
     ):
         """
@@ -288,10 +289,10 @@ def ll_problem(
             "found {} instead".format(type(self.io_opt))
         )
         assert self._method in (
-            "MetaRepr",
+            "MetaFeat",
             "MetaInit",
-        ), "illegal initialization value for argument:method, should be in [MetaRepr, MetaInit]"
-        if self.method == "MetaRepr":
+        ), "illegal initialization value for argument:method, should be in [MetaFeat, MetaInit]"
+        if self.method == "MetaFeat":
             if self.inner_method == "Aggr":
                 assert (
                     outer_objective is not None

diff --git a/boml/extension.py b/boml/extension.py
@@ -8,7 +8,7 @@
 
 from boml import utils
 
-Meta_Repr_Method = ["MAML", "FOMAML", "Meta-SGD", "MT-net", "WarpGrad"]
+Meta_Repr_Method = ["MAML", "FMAML", "Meta-SGD", "MT-net", "WarpGrad"]
 Meta_Init_Method = ["RHG", "TRHG", "DARTS", "BDA", "Implicit"]
 METHOD_COLLECTIONS = [Meta_Repr_Method, Meta_Init_Method]
 

diff --git a/boml/load_data/datasets/dl_utils.py b/boml/load_data/datasets/dl_utils.py
@@ -1,4 +1,3 @@
-import os
 from functools import reduce
 
 import numpy as np
@@ -8,6 +7,12 @@
 
 
 def get_indices_balanced_classes(n_examples, labels, forbidden_indices):
+    """
+    :param n_examples: number of examples
+    :param labels: paired labels
+    :param forbidden_indices:
+    :return:
+    """
     N = len(labels)
     n_classes = len(labels[0])
 
@@ -28,6 +33,10 @@ def get_indices_balanced_classes(n_examples, labels, forbidden_indices):
 
 
 def test_if_balanced(dataset):
+    """
+    :param dataset:
+    :return: just for testing
+    """
     labels = dataset.target
     n_classes = len(labels[0])
     class_counter = [0] * n_classes
@@ -37,10 +46,19 @@ def test_if_balanced(dataset):
 
 
 def maybe_cast_to_scalar(what):
+    """
+    :param what: input
+    :return: return the scalar of input if the length of input equals 1.
+    """
     return what[0] if len(what) == 1 else what
 
 
 def pad(_example, _size):
+    """
+    :param _example:
+    :param _size:
+    :return: performs concatenation
+    """
     return np.concatenate([_example] * _size)
 
 
@@ -68,6 +86,10 @@ def vstack(lst):
 
 
 def convert_sparse_matrix_to_sparse_tensor(X):
+    """
+    :param X: sparse matrix
+    :return: sparse tensor of X
+    """
     if isinstance(X, sc_sp.csr.csr_matrix):
         coo = X.tocoo()
         indices = np.mat([coo.row, coo.col]).transpose()
@@ -78,6 +100,10 @@ def convert_sparse_matrix_to_sparse_tensor(X):
 
 
 def get_data(d_set):
+    """
+    :param d_set: instance of dataset
+    :return: inputs of datasets
+    """
     if hasattr(d_set, "images"):
         data = d_set.images
     elif hasattr(d_set, "data"):
@@ -88,6 +114,10 @@ def get_data(d_set):
 
 
 def get_targets(d_set):
+    """
+    :param d_set: instance of dataset
+    :return: labels of datasets
+    """
     if hasattr(d_set, "labels"):
         return d_set.labels
     elif hasattr(d_set, "target"):
@@ -147,18 +177,14 @@ def as_tuple_or_list(obj):
 
 
 def to_one_hot_enc(seq, dimension=None):
+    """
+
+    :param seq: sequence
+    :param dimension: the dimension of output
+    :return: output of one-hot encoding
+    """
     da_max = dimension or int(np.max(seq)) + 1
     _tmp = np.zeros((len(seq), da_max))
     _tmp[range(len(_tmp)), np.array(seq, dtype=int)] = 1
     return _tmp
-    #
-    # def create_and_set(_p):
-    #     _tmp = np.zeros(da_max)
-    #     _tmp[int(_p)] = 1
-    #     return _tmp
-    #
-    # return np.array([create_and_set(_v) for _v in seq])
-
 
-from_env = os.getenv("DATASETS_FOLDER")
-DATA_FOLDER = from_env
diff --git a/boml/load_data/datasets/load_full_dataset.py b/boml/load_data/datasets/load_full_dataset.py
@@ -75,14 +75,8 @@
 
 print("Data folder is", DATA_FOLDER)
 
-# kind of private
-TIMIT_DIR = os.path.join(DATA_FOLDER, "timit4python")
-XRMB_DIR = os.path.join(DATA_FOLDER, "XRMB")
-IROS15_BASE_FOLDER = os.path.join(
-    DATA_FOLDER, os.path.join("dls_collaboration", "Learning")
-)
 
-# easy to find!
+# easy to find
 IRIS_TRAINING = os.path.join(DATA_FOLDER, "iris", "training.csv")
 IRIS_TEST = os.path.join(DATA_FOLDER, "iris", "test.csv")
 MNIST_DIR = os.path.join(DATA_FOLDER, "mnist_data")