Skip to content
Snippets Groups Projects
Commit 8bdeaf68 authored by Sun Jin Kim's avatar Sun Jin Kim
Browse files

edit aa_learner

parent 5fda1aca
No related branches found
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
MetaAugment/CP2_Max.py
+ 3
− 1
View file @ 8bdeaf68
......@@ -37,8 +37,10 @@ class Learner(nn.Module):
self.relu3 = nn.ReLU()
self.fc2 = nn.Linear(120, 84)
self.relu4 = nn.ReLU()
self.fc3 = nn.Linear(84, 13)
self.fc3 = nn.Linear(84, num_transforms + 21)
# self.sig = nn.Sigmoid()
# Currently using discrete outputs for the probabilities
def forward(self, x):
y = self.conv1(x)
......
MetaAugment/__pycache__/main.cpython-38.pyc
+ 0
− 0
View file @ 8bdeaf68
No preview for this file type
MetaAugment/autoaugment_learners/aa_learner.py
+ 92
− 87
View file @ 8bdeaf68
......@@ -53,109 +53,114 @@ class aa_learner:
# TODO: We should probably use a different way to store results than self.history
self.history = []
def generate_new_policy(self):
def generate_new_discrete_operation(self, fun_num=14, p_bins=10, m_bins=10):
'''
Generate a new random policy in the form of
[
(("Invert", 0.8, None), ("Contrast", 0.2, 6)),
(("Rotate", 0.7, 2), ("Invert", 0.8, None)),
(("Sharpness", 0.8, 1), ("Sharpness", 0.9, 3)),
(("ShearY", 0.5, 8), ("Invert", 0.7, None)),
]
generate a new random operation in the form of a tensor of dimension:
(fun_num + 11 + 10)
The first fun_num dimensions is a 1-hot encoding to specify which function to use.
The next 11 dimensions specify which 'probability' to choose.
(0.0, 0.1, ..., 1.0)
The next 10 dimensions specify which 'magnitude' to choose.
(0, 1, ..., 9)
'''
fun = np.random.randint(0, fun_num)
prob = np.random.randint(p_bins+1, fun_num)
mag = np.random.randint(m_bins, fun_num)
fun_t= torch.zeros(fun_num)
fun_t[fun] = 1
prob_t = torch.zeros(p_bins+1)
prob_t[prob] = 1
mag_t = torch.zeros(m_bins)
mag_t[mag] = 1
def generate_new_discrete_operation(fun_num=14, p_bins=10, m_bins=10):
'''
generate a new random operation in the form of a tensor of dimension:
(fun_num + 11 + 10)
The first fun_num dimensions is a 1-hot encoding to specify which function to use.
The next 11 dimensions specify which 'probability' to choose.
(0.0, 0.1, ..., 1.0)
The next 10 dimensions specify which 'magnitude' to choose.
(0, 1, ..., 9)
'''
fun = np.random.randint(0, fun_num)
prob = np.random.randint(p_bins+1, fun_num)
mag = np.random.randint(m_bins, fun_num)
fun_t= torch.zeros(fun_num)
fun_t[fun] = 1
prob_t = torch.zeros(p_bins+1)
prob_t[prob] = 1
mag_t = torch.zeros(m_bins)
mag_t[mag] = 1
return torch.cat([fun_t, prob_t, mag_t])
return torch.cat([fun_t, prob_t, mag_t])
def generate_new_continuous_operation(self, fun_num=14, p_bins=10, m_bins=10):
'''
Returns operation_tensor, which is a tensor representation of a random operation with
dimension:
(fun_num + 1 + 1)
The first fun_num dimensions is a 1-hot encoding to specify which function to use.
The next 1 dimensions specify which 'probability' to choose.
0 < x < 1
The next 1 dimensions specify which 'magnitude' to choose.
0 < x < 9
'''
fun = np.random.randint(0, fun_num)
fun_p_m = torch.zeros(fun_num + 2)
fun_p_m[fun] = 1
fun_p_m[-2] = np.random.uniform() # 0<prob<1
fun_p_m[-1] = np.random.uniform() * (m_bins-1) # 0<mag<9
return fun_p_m
def generate_new_continuous_operation(fun_num=14, p_bins=10, m_bins=10):
'''
Returns operation_tensor, which is a tensor representation of a random operation with
dimension:
(fun_num + 1 + 1)
The first fun_num dimensions is a 1-hot encoding to specify which function to use.
The next 1 dimensions specify which 'probability' to choose.
0 < x < 1
The next 1 dimensions specify which 'magnitude' to choose.
0 < x < 9
'''
fun = np.random.randint(0, fun_num)
fun_p_m = torch.zeros(fun_num + 2)
fun_p_m[fun] = 1
fun_p_m[-2] = np.random.uniform() # 0<prob<1
fun_p_m[-1] = np.random.uniform() * (m_bins-1) # 0<mag<9
return fun_p_m
def translate_operation_tensor(operation_tensor, fun_num=14,
def translate_operation_tensor(self, operation_tensor, fun_num=14,
p_bins=10, m_bins=10,
discrete_p_m=False):
'''
takes in a tensor representing a operation and returns an actual operation which
is in the form of:
("Invert", 0.8, None)
or
("Contrast", 0.2, 6)
Args:
operation_tensor
continuous_p_m (boolean): whether the operation_tensor has continuous representations
of probability and magnitude
'''
# if input operation_tensor is discrete
if discrete_p_m:
fun_t = operation_tensor[:fun_num]
prob_t = operation_tensor[fun_num:fun_num+p_bins+1]
mag_t = operation_tensor[-m_bins:]
fun = torch.argmax(fun_t)
prob = torch.argmax(prob_t)
mag = torch.argmax(mag_t)
raise NotImplementedError("U can implement this if u want")
# process continuous operation_tensor
'''
takes in a tensor representing a operation and returns an actual operation which
is in the form of:
("Invert", 0.8, None)
or
("Contrast", 0.2, 6)
Args:
operation_tensor
continuous_p_m (boolean): whether the operation_tensor has continuous representations
of probability and magnitude
'''
# if input operation_tensor is discrete
if discrete_p_m:
fun_t = operation_tensor[:fun_num]
p = operation_tensor[-2].item() # 0 < p < 1
m = operation_tensor[-1].item() # 0 < m < 9
prob_t = operation_tensor[fun_num:fun_num+p_bins+1]
mag_t = operation_tensor[-m_bins:]
fun = torch.argmax(fun_t)
prob = torch.argmax(prob_t) # 0 <= p <= 10
mag = torch.argmax(mag_t) # 0 <= m <= 9
fun = augmentation_space[fun][0]
prob = prob/10
fun_num = torch.argmax(fun_t)
function = augmentation_space[fun_num][0]
p = round(p, 1) # round to nearest first decimal digit
m = round(m) # round to nearest integer
return (function, p, m)
return (fun, prob, mag)
# process continuous operation_tensor
fun_t = operation_tensor[:fun_num]
p = operation_tensor[-2].item() # 0 < p < 1
m = operation_tensor[-1].item() # 0 < m < 9
fun_num = torch.argmax(fun_t)
function = augmentation_space[fun_num][0]
p = round(p, 1) # round to nearest first decimal digit
m = round(m) # round to nearest integer
return (function, p, m)
def generate_new_policy(self):
'''
Generate a new random policy in the form of
[
(("Invert", 0.8, None), ("Contrast", 0.2, 6)),
(("Rotate", 0.7, 2), ("Invert", 0.8, None)),
(("Sharpness", 0.8, 1), ("Sharpness", 0.9, 3)),
(("ShearY", 0.5, 8), ("Invert", 0.7, None)),
]
'''
new_policy = []
for _ in range(self.sp_num):
# generate 2 operations for each subpolicy
ops = []
for i in range(2):
new_op = generate_new_continuous_operation(self.fun_num)
new_op = translate_operation_tensor(new_op)
new_op = self.generate_new_continuous_operation(self.fun_num)
new_op = self.translate_operation_tensor(new_op)
ops.append(new_op)
new_subpolicy = tuple(ops)
......
MetaAugment/autoaugment_learners/randomsearch_learner.py
+ 91
− 87
View file @ 8bdeaf68
......@@ -57,109 +57,113 @@ class randomsearch_learner:
# TODO: We should probably use a different way to store results than self.history
self.history = []
def generate_new_policy(self):
def generate_new_discrete_operation(self, fun_num=14, p_bins=10, m_bins=10):
'''
Generate a new random policy in the form of
[
(("Invert", 0.8, None), ("Contrast", 0.2, 6)),
(("Rotate", 0.7, 2), ("Invert", 0.8, None)),
(("Sharpness", 0.8, 1), ("Sharpness", 0.9, 3)),
(("ShearY", 0.5, 8), ("Invert", 0.7, None)),
]
generate a new random operation in the form of a tensor of dimension:
(fun_num + 11 + 10)
The first fun_num dimensions is a 1-hot encoding to specify which function to use.
The next 11 dimensions specify which 'probability' to choose.
(0.0, 0.1, ..., 1.0)
The next 10 dimensions specify which 'magnitude' to choose.
(0, 1, ..., 9)
'''
fun = np.random.randint(0, fun_num)
prob = np.random.randint(p_bins+1, fun_num)
mag = np.random.randint(m_bins, fun_num)
fun_t= torch.zeros(fun_num)
fun_t[fun] = 1
prob_t = torch.zeros(p_bins+1)
prob_t[prob] = 1
mag_t = torch.zeros(m_bins)
mag_t[mag] = 1
def generate_new_discrete_operation(fun_num=14, p_bins=10, m_bins=10):
'''
generate a new random operation in the form of a tensor of dimension:
(fun_num + 11 + 10)
The first fun_num dimensions is a 1-hot encoding to specify which function to use.
The next 11 dimensions specify which 'probability' to choose.
(0.0, 0.1, ..., 1.0)
The next 10 dimensions specify which 'magnitude' to choose.
(0, 1, ..., 9)
'''
fun = np.random.randint(0, fun_num)
prob = np.random.randint(p_bins+1, fun_num)
mag = np.random.randint(m_bins, fun_num)
fun_t= torch.zeros(fun_num)
fun_t[fun] = 1
prob_t = torch.zeros(p_bins+1)
prob_t[prob] = 1
mag_t = torch.zeros(m_bins)
mag_t[mag] = 1
return torch.cat([fun_t, prob_t, mag_t])
return torch.cat([fun_t, prob_t, mag_t])
def generate_new_continuous_operation(self, fun_num=14, p_bins=10, m_bins=10):
'''
Returns operation_tensor, which is a tensor representation of a random operation with
dimension:
(fun_num + 1 + 1)
The first fun_num dimensions is a 1-hot encoding to specify which function to use.
The next 1 dimensions specify which 'probability' to choose.
0 < x < 1
The next 1 dimensions specify which 'magnitude' to choose.
0 < x < 9
'''
fun = np.random.randint(0, fun_num)
fun_p_m = torch.zeros(fun_num + 2)
fun_p_m[fun] = 1
fun_p_m[-2] = np.random.uniform() # 0<prob<1
fun_p_m[-1] = np.random.uniform() * (m_bins-1) # 0<mag<9
return fun_p_m
def generate_new_continuous_operation(fun_num=14, p_bins=10, m_bins=10):
'''
Returns operation_tensor, which is a tensor representation of a random operation with
dimension:
(fun_num + 1 + 1)
The first fun_num dimensions is a 1-hot encoding to specify which function to use.
The next 1 dimensions specify which 'probability' to choose.
0 < x < 1
The next 1 dimensions specify which 'magnitude' to choose.
0 < x < 9
'''
fun = np.random.randint(0, fun_num)
fun_p_m = torch.zeros(fun_num + 2)
fun_p_m[fun] = 1
fun_p_m[-2] = np.random.uniform() # 0<prob<1
fun_p_m[-1] = np.random.uniform() * (m_bins-1) # 0<mag<9
return fun_p_m
def translate_operation_tensor(operation_tensor, fun_num=14,
def translate_operation_tensor(self, operation_tensor, fun_num=14,
p_bins=10, m_bins=10,
discrete_p_m=False):
'''
takes in a tensor representing a operation and returns an actual operation which
is in the form of:
("Invert", 0.8, None)
or
("Contrast", 0.2, 6)
Args:
operation_tensor
continuous_p_m (boolean): whether the operation_tensor has continuous representations
of probability and magnitude
'''
# if input operation_tensor is discrete
if discrete_p_m:
fun_t = operation_tensor[:fun_num]
prob_t = operation_tensor[fun_num:fun_num+p_bins+1]
mag_t = operation_tensor[-m_bins:]
fun = torch.argmax(fun_t)
prob = torch.argmax(prob_t)
mag = torch.argmax(mag_t)
raise NotImplementedError("U can implement this if u want")
# process continuous operation_tensor
'''
takes in a tensor representing a operation and returns an actual operation which
is in the form of:
("Invert", 0.8, None)
or
("Contrast", 0.2, 6)
Args:
operation_tensor
continuous_p_m (boolean): whether the operation_tensor has continuous representations
of probability and magnitude
'''
# if input operation_tensor is discrete
if discrete_p_m:
fun_t = operation_tensor[:fun_num]
p = operation_tensor[-2].item() # 0 < p < 1
m = operation_tensor[-1].item() # 0 < m < 9
prob_t = operation_tensor[fun_num:fun_num+p_bins+1]
mag_t = operation_tensor[-m_bins:]
fun = torch.argmax(fun_t)
prob = torch.argmax(prob_t) # 0 <= p <= 10
mag = torch.argmax(mag_t) # 0 <= m <= 9
fun_num = torch.argmax(fun_t)
function = augmentation_space[fun_num][0]
p = round(p, 1) # round to nearest first decimal digit
m = round(m) # round to nearest integer
return (function, p, m)
fun = augmentation_space[fun][0]
prob = prob/10
return (fun, prob, mag)
# process continuous operation_tensor
fun_t = operation_tensor[:fun_num]
p = operation_tensor[-2].item() # 0 < p < 1
m = operation_tensor[-1].item() # 0 < m < 9
fun_num = torch.argmax(fun_t)
function = augmentation_space[fun_num][0]
p = round(p, 1) # round to nearest first decimal digit
m = round(m) # round to nearest integer
return (function, p, m)
def generate_new_policy(self):
'''
Generate a new random policy in the form of
[
(("Invert", 0.8, None), ("Contrast", 0.2, 6)),
(("Rotate", 0.7, 2), ("Invert", 0.8, None)),
(("Sharpness", 0.8, 1), ("Sharpness", 0.9, 3)),
(("ShearY", 0.5, 8), ("Invert", 0.7, None)),
]
'''
new_policy = []
for _ in range(self.sp_num):
# generate 2 operations for each subpolicy
ops = []
for i in range(2):
new_op = generate_new_continuous_operation(self.fun_num)
new_op = translate_operation_tensor(new_op)
new_op = self.generate_new_continuous_operation(self.fun_num)
new_op = self.translate_operation_tensor(new_op)
ops.append(new_op)
new_subpolicy = tuple(ops)
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment