From 839fb15f31e70e2bc3f38332a47b8f5e6991e576 Mon Sep 17 00:00:00 2001
From: Sun Jin Kim <sk2521@ic.ac.uk>
Date: Mon, 25 Apr 2022 17:09:24 +0100
Subject: [PATCH] FINISH REFACTORING UCB_LEARNER
---
.../autoaugment_learners/aa_learner.py | 5 +-
.../autoaugment_learners/ucb_learner.py | 101 ++++++++++--------
temp_util/wapp_util.py | 3 +
test/MetaAugment/test_ucb_learner.py | 45 ++++++--
4 files changed, 104 insertions(+), 50 deletions(-)
diff --git a/MetaAugment/autoaugment_learners/aa_learner.py b/MetaAugment/autoaugment_learners/aa_learner.py
index 48d4f051..0eb38d59 100644
--- a/MetaAugment/autoaugment_learners/aa_learner.py
+++ b/MetaAugment/autoaugment_learners/aa_learner.py
@@ -309,7 +309,8 @@ class aa_learner:
child_network_architecture,
train_dataset,
test_dataset,
- logging=False):
+ logging=False,
+ print_every_epoch=True):
"""
Given a policy (using AutoAugment paper terminology), we train a child network
using the policy and return the accuracy (how good the policy is for the dataset and
@@ -384,7 +385,7 @@ class aa_learner:
max_epochs = self.max_epochs,
early_stop_num = self.early_stop_num,
logging = logging,
- print_every_epoch=True)
+ print_every_epoch=print_every_epoch)
# if logging is true, 'accuracy' is actually a tuple: (accuracy, accuracy_log)
return accuracy
diff --git a/MetaAugment/autoaugment_learners/ucb_learner.py b/MetaAugment/autoaugment_learners/ucb_learner.py
index 1a4ddf3a..e22f32ff 100644
--- a/MetaAugment/autoaugment_learners/ucb_learner.py
+++ b/MetaAugment/autoaugment_learners/ucb_learner.py
@@ -1,9 +1,3 @@
-#!/usr/bin/env python
-# coding: utf-8
-
-# In[1]:
-
-
import numpy as np
import torch
import torch.nn as nn
@@ -53,23 +47,24 @@ class ucb_learner(randomsearch_learner):
max_epochs=max_epochs,
early_stop_num=early_stop_num,)
- self.num_policies = num_policies
- # When this learner is initialized we generate `num_policies` number
- # of random policies.
- # generate_new_policy is inherited from the randomsearch_learner class
- self.policies = []
- self.make_more_policies()
+
# attributes used in the UCB1 algorithm
- self.q_values = [0]*self.num_policies
- self.best_q_values = []
+ self.num_policies = num_policies
+
+ self.policies = [self.generate_new_policy() for _ in range(num_policies)]
+
+ self.avg_accs = [None]*self.num_policies
+ self.best_avg_accs = []
+
self.cnts = [0]*self.num_policies
self.q_plus_cnt = [0]*self.num_policies
self.total_count = 0
+
def make_more_policies(self, n):
"""generates n more random policies and adds it to self.policies
@@ -78,50 +73,71 @@ class ucb_learner(randomsearch_learner):
and add to our list of policies
"""
- self.policies.append([self.generate_new_policy() for _ in n])
+ self.policies += [self.generate_new_policy() for _ in range(n)]
+
+ # all the below need to be lengthened to store information for the
+ # new policies
+ self.avg_accs += [None for _ in range(n)]
+ self.cnts += [0 for _ in range(n)]
+ self.q_plus_cnt += [None for _ in range(n)]
+ self.num_policies += n
+
def learn(self,
train_dataset,
test_dataset,
child_network_architecture,
- iterations=15):
+ iterations=15,
+ print_every_epoch=False):
+ """continue the UCB algorithm for `iterations` number of turns
+ """
for this_iter in trange(iterations):
- # get the action to try (either initially in order or using best q_plus_cnt value)
- # TODO: change this if statemetn
- if this_iter >= self.num_policies:
- this_policy_idx = np.argmax(self.q_plus_cnt)
+ # choose which policy we want to test
+ if None in self.avg_accs:
+ # if there is a policy we haven't tested yet, we
+ # test that one
+ this_policy_idx = self.avg_accs.index(None)
this_policy = self.policies[this_policy_idx]
- else:
- this_policy = this_iter
-
-
- best_acc = self.test_autoaugment_policy(
+ acc = self.test_autoaugment_policy(
this_policy,
child_network_architecture,
train_dataset,
test_dataset,
- logging=False
+ logging=False,
+ print_every_epoch=print_every_epoch
)
-
- # update q_values
- # TODO: change this if statemetn
- if this_iter < self.num_policies:
- self.q_values[this_policy_idx] += best_acc
+ # update q_values (average accuracy)
+ self.avg_accs[this_policy_idx] = acc
else:
- self.q_values[this_policy_idx] = (self.q_values[this_policy_idx]*self.cnts[this_policy_idx] + best_acc) / (self.cnts[this_policy_idx] + 1)
-
- best_q_value = max(self.q_values)
- self.best_q_values.append(best_q_value)
-
+ # if we have tested all policies before, we test the
+ # one with the best q_plus_cnt value
+ this_policy_idx = np.argmax(self.q_plus_cnt)
+ this_policy = self.policies[this_policy_idx]
+ acc = self.test_autoaugment_policy(
+ this_policy,
+ child_network_architecture,
+ train_dataset,
+ test_dataset,
+ logging=False,
+ print_every_epoch=print_every_epoch
+ )
+ # update q_values (average accuracy)
+ self.avg_accs[this_policy_idx] = (self.avg_accs[this_policy_idx]*self.cnts[this_policy_idx] + acc) / (self.cnts[this_policy_idx] + 1)
+
+ # logging the best avg acc up to now
+ best_avg_acc = max([x for x in self.avg_accs if x is not None])
+ self.best_avg_accs.append(best_avg_acc)
+
+ # print progress for user
if (this_iter+1) % 5 == 0:
print("Iteration: {},\tQ-Values: {}, Best this_iter: {}".format(
this_iter+1,
- list(np.around(np.array(self.q_values),2)),
- max(list(np.around(np.array(self.q_values),2)))
+ list(np.around(np.array(self.avg_accs),2)),
+ max(list(np.around(np.array(self.avg_accs),2)))
)
)
@@ -130,10 +146,11 @@ class ucb_learner(randomsearch_learner):
self.total_count += 1
# update q_plus_cnt values every turn after the initial sweep through
- # TODO: change this if statemetn
- if this_iter >= self.num_policies - 1:
- for i in range(self.num_policies):
- self.q_plus_cnt[i] = self.q_values[i] + np.sqrt(2*np.log(self.total_count)/self.cnts[i])
+ for i in range(self.num_policies):
+ if self.avg_accs[i] is not None:
+ self.q_plus_cnt[i] = self.avg_accs[i] + np.sqrt(2*np.log(self.total_count)/self.cnts[i])
+
+ print(self.cnts)
diff --git a/temp_util/wapp_util.py b/temp_util/wapp_util.py
index 78be118a..e48d1c31 100644
--- a/temp_util/wapp_util.py
+++ b/temp_util/wapp_util.py
@@ -17,13 +17,16 @@ from MetaAugment.main import create_toy
import pickle
def parse_users_learner_spec(
+ # aalearner type
auto_aug_learner,
+ # search space settings
ds,
ds_name,
exclude_method,
num_funcs,
num_policies,
num_sub_policies,
+ # child network settings
toy_size,
IsLeNet,
batch_size,
diff --git a/test/MetaAugment/test_ucb_learner.py b/test/MetaAugment/test_ucb_learner.py
index 564ac80d..7c6635ff 100644
--- a/test/MetaAugment/test_ucb_learner.py
+++ b/test/MetaAugment/test_ucb_learner.py
@@ -1,7 +1,18 @@
import MetaAugment.autoaugment_learners as aal
-
+import MetaAugment.child_networks as cn
+import torchvision
+import torchvision.datasets as datasets
+from pprint import pprint
def test_ucb_learner():
+ child_network_architecture = cn.SimpleNet
+ train_dataset = datasets.FashionMNIST(root='./datasets/fashionmnist/train',
+ train=True, download=True, transform=None)
+ test_dataset = datasets.FashionMNIST(root='./datasets/fashionmnist/test',
+ train=False, download=True,
+ transform=torchvision.transforms.ToTensor())
+
+
learner = aal.ucb_learner(
# parameters that define the search space
sp_num=5,
@@ -10,15 +21,37 @@ def test_ucb_learner():
discrete_p_m=True,
# hyperparameters for when training the child_network
batch_size=8,
- toy_flag=False,
- toy_size=0.1,
+ toy_flag=True,
+ toy_size=0.001,
learning_rate=1e-1,
max_epochs=float('inf'),
early_stop_num=30,
# ucb_learner specific hyperparameter
- num_policies=100
+ num_policies=3
)
- print(learner.policies)
+ pprint(learner.policies)
+ assert len(learner.policies)==len(learner.avg_accs), \
+ (len(learner.policies), (len(learner.avg_accs)))
+
+ # learn on the 3 policies we generated
+ learner.learn(
+ train_dataset=train_dataset,
+ test_dataset=test_dataset,
+ child_network_architecture=child_network_architecture,
+ iterations=5
+ )
+
+ # let's say we want to explore more policies:
+ # we generate more new policies
+ learner.make_more_policies(n=4)
+
+ # and let's explore how good those are as well
+ learner.learn(
+ train_dataset=train_dataset,
+ test_dataset=test_dataset,
+ child_network_architecture=child_network_architecture,
+ iterations=7
+ )
if __name__=="__main__":
- test_ucb_learner()
\ No newline at end of file
+ test_ucb_learner()
--
GitLab