3.1. Gradient Operator
The grad_operator
module contains the main operators of ConFIG algorithm. You can use these operators to perform the ConFIG update step for your optimization problem.
Operation Functions¤
conflictfree.grad_operator.ConFIG_update
¤
ConFIG_update(
grads: Union[torch.Tensor, Sequence[torch.Tensor]],
weight_model: WeightModel = EqualWeight(),
length_model: LengthModel = ProjectionLength(),
use_least_square: bool = True,
losses: Optional[Sequence] = None,
) -> torch.Tensor
Performs the standard ConFIG update step.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
grads |
Union[Tensor, Sequence[Tensor]]
|
The gradients to update. It can be a stack of gradient vectors (at dim 0) or a sequence of gradient vectors. |
required |
weight_model |
WeightModel
|
The weight model for calculating the direction weights. Defaults to EqualWeight(), which will make the final update gradient not biased towards any gradient. |
EqualWeight()
|
length_model |
LengthModel
|
The length model for rescaling the length of the final gradient. Defaults to ProjectionLength(), which will project each gradient vector onto the final gradient vector to get the final length. |
ProjectionLength()
|
use_least_square |
bool
|
Whether to use the least square method for calculating the best direction.
If set to False, we will directly calculate the pseudo-inverse of the gradient matrix. See |
True
|
losses |
Optional[Sequence]
|
The losses associated with the gradients. The losses will be passed to the weight and length model. If your weight/length model doesn't require loss information, you can set this value as None. Defaults to None. |
None
|
Returns:
Type | Description |
---|---|
Tensor
|
torch.Tensor: The final update gradient. |
Examples:
from conflictfree.grad_operator import ConFIG_update
from conflictfree.utils import get_gradient_vector,apply_gradient_vector
optimizer=torch.Adam(network.parameters(),lr=1e-3)
for input_i in dataset:
grads=[] # we record gradients rather than losses
for loss_fn in loss_fns:
optimizer.zero_grad()
loss_i=loss_fn(input_i)
loss_i.backward()
grads.append(get_gradient_vector(network)) #get loss-specfic gradient
g_config=ConFIG_update(grads) # calculate the conflict-free direction
apply_gradient_vector(network) # set the condlict-free direction to the network
optimizer.step()
Source code in conflictfree/grad_operator.py
68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 |
|
conflictfree.grad_operator.ConFIG_update_double
¤
ConFIG_update_double(
grad_1: torch.Tensor,
grad_2: torch.Tensor,
weight_model: WeightModel = EqualWeight(),
length_model: LengthModel = ProjectionLength(),
losses: Optional[Sequence] = None,
) -> torch.Tensor
ConFIG update for two gradients where no inverse calculation is needed.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
grad_1 |
Tensor
|
The first gradient. |
required |
grad_2 |
Tensor
|
The second gradient. |
required |
weight_model |
WeightModel
|
The weight model for calculating the direction weights. Defaults to EqualWeight(), which will make the final update gradient not biased towards any gradient. |
EqualWeight()
|
length_model |
LengthModel
|
The length model for rescaling the length of the final gradient. Defaults to ProjectionLength(), which will project each gradient vector onto the final gradient vector to get the final length. |
ProjectionLength()
|
losses |
Optional[Sequence]
|
The losses associated with the gradients. The losses will be passed to the weight and length model. If your weight/length model doesn't require loss information, you can set this value as None. Defaults to None. |
None
|
Returns:
Type | Description |
---|---|
Tensor
|
torch.Tensor: The final update gradient. |
Examples:
from conflictfree.grad_operator import ConFIG_update_double
from conflictfree.utils import get_gradient_vector,apply_gradient_vector
optimizer=torch.Adam(network.parameters(),lr=1e-3)
for input_i in dataset:
grads=[] # we record gradients rather than losses
for loss_fn in [loss_fn1, loss_fn2]:
optimizer.zero_grad()
loss_i=loss_fn(input_i)
loss_i.backward()
grads.append(get_gradient_vector(network)) #get loss-specfic gradient
g_config=ConFIG_update_double(grads) # calculate the conflict-free direction
apply_gradient_vector(network) # set the condlict-free direction to the network
optimizer.step()
Source code in conflictfree/grad_operator.py
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 |
|
Operator Classes¤
conflictfree.grad_operator.ConFIGOperator
¤
Bases: GradientOperator
Operator for the ConFIG algorithm.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
weight_model |
WeightModel
|
The weight model for calculating the direction weights. Defaults to EqualWeight(), which will make the final update gradient not biased towards any gradient. |
EqualWeight()
|
length_model |
LengthModel
|
The length model for rescaling the length of the final gradient. Defaults to ProjectionLength(), which will project each gradient vector onto the final gradient vector to get the final length. |
ProjectionLength()
|
allow_simplified_model |
bool
|
Whether to allow simplified model for calculating the gradient. If set to True, will use simplified form of ConFIG method when there are only two losses (ConFIG_update_double). Defaults to True. |
True
|
use_least_square |
bool
|
Whether to use the least square method for calculating the best direction.
If set to False, we will directly calculate the pseudo-inverse of the gradient matrix. See |
True
|
Examples:
from conflictfree.grad_operator import ConFIGOperator
from conflictfree.utils import get_gradient_vector,apply_gradient_vector
optimizer=torch.Adam(network.parameters(),lr=1e-3)
operator=ConFIGOperator() # initialize operator
for input_i in dataset:
grads=[]
for loss_fn in loss_fns:
optimizer.zero_grad()
loss_i=loss_fn(input_i)
loss_i.backward()
grads.append(get_gradient_vector(network))
g_config=operator.calculate_gradient(grads) # calculate the conflict-free direction
apply_gradient_vector(network) # or simply use `operator.update_gradient(network,grads)` to calculate and set the condlict-free direction to the network
optimizer.step()
Source code in conflictfree/grad_operator.py
177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 |
|
update_gradient
¤
update_gradient(
network: torch.nn.Module,
grads: Union[torch.Tensor, Sequence[torch.Tensor]],
losses: Optional[Sequence] = None,
) -> None
Calculate the gradient and apply the gradient to the network.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
network |
Module
|
The target network. |
required |
grads |
Union[Tensor, Sequence[Tensor]]
|
The gradients to update. It can be a stack of gradient vectors (at dim 0) or a sequence of gradient vectors. |
required |
losses |
Optional[Sequence]
|
The losses associated with the gradients. The losses will be passed to the weight and length model. If your weight/length model doesn't require loss information, you can set this value as None. Defaults to None. |
None
|
Returns:
Type | Description |
---|---|
None
|
None |
Source code in conflictfree/grad_operator.py
158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 |
|
__init__
¤
__init__(
weight_model: WeightModel = EqualWeight(),
length_model: LengthModel = ProjectionLength(),
allow_simplified_model: bool = True,
use_least_square: bool = True,
)
Source code in conflictfree/grad_operator.py
212 213 214 215 216 217 218 219 220 221 |
|
calculate_gradient
¤
calculate_gradient(
grads: Union[torch.Tensor, Sequence[torch.Tensor]],
losses: Optional[Sequence] = None,
) -> torch.Tensor
Calculates the gradient using the ConFIG algorithm.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
grads |
Union[Tensor, Sequence[Tensor]]
|
The gradients to update. It can be a stack of gradient vectors (at dim 0) or a sequence of gradient vectors. |
required |
losses |
Optional[Sequence]
|
The losses associated with the gradients. The losses will be passed to the weight and length model. If your weight/length model doesn't require loss information, you can set this value as None. Defaults to None. |
None
|
Returns:
Type | Description |
---|---|
Tensor
|
torch.Tensor: The calculated gradient. |
Source code in conflictfree/grad_operator.py
223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 |
|
conflictfree.grad_operator.PCGradOperator
¤
Bases: GradientOperator
PCGradOperator class represents a gradient operator for PCGrad algorithm.
@inproceedings{yu2020gradient, title={Gradient surgery for multi-task learning}, author={Yu, Tianhe and Kumar, Saurabh and Gupta, Abhishek and Levine, Sergey and Hausman, Karol and Finn, Chelsea}, booktitle={34th International Conference on Neural Information Processing Systems}, year={2020}, url={https://dl.acm.org/doi/abs/10.5555/3495724.3496213} }
Source code in conflictfree/grad_operator.py
251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 |
|
update_gradient
¤
update_gradient(
network: torch.nn.Module,
grads: Union[torch.Tensor, Sequence[torch.Tensor]],
losses: Optional[Sequence] = None,
) -> None
Calculate the gradient and apply the gradient to the network.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
network |
Module
|
The target network. |
required |
grads |
Union[Tensor, Sequence[Tensor]]
|
The gradients to update. It can be a stack of gradient vectors (at dim 0) or a sequence of gradient vectors. |
required |
losses |
Optional[Sequence]
|
The losses associated with the gradients. The losses will be passed to the weight and length model. If your weight/length model doesn't require loss information, you can set this value as None. Defaults to None. |
None
|
Returns:
Type | Description |
---|---|
None
|
None |
Source code in conflictfree/grad_operator.py
158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 |
|
calculate_gradient
¤
calculate_gradient(
grads: Union[torch.Tensor, Sequence[torch.Tensor]],
losses: Optional[Sequence] = None,
) -> torch.Tensor
Calculates the gradient using the PCGrad algorithm.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
grads |
Union[Tensor, Sequence[Tensor]]
|
The gradients to update. It can be a stack of gradient vectors (at dim 0) or a sequence of gradient vectors. |
required |
losses |
Optional[Sequence]
|
This parameter should not be set for current operator. Defaults to None. |
None
|
Returns:
Type | Description |
---|---|
Tensor
|
torch.Tensor: The calculated gradient using PCGrad method. |
Source code in conflictfree/grad_operator.py
266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 |
|
conflictfree.grad_operator.IMTLGOperator
¤
Bases: GradientOperator
PCGradOperator class represents a gradient operator for IMTL-G algorithm.
@inproceedings{ liu2021towards, title={Towards Impartial Multi-task Learning}, author={Liyang Liu and Yi Li and Zhanghui Kuang and Jing-Hao Xue and Yimin Chen and Wenming Yang and Qingmin Liao and Wayne Zhang}, booktitle={International Conference on Learning Representations}, year={2021}, url={https://openreview.net/forum?id=IMPnRXEWpvr} }
Source code in conflictfree/grad_operator.py
291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 |
|
update_gradient
¤
update_gradient(
network: torch.nn.Module,
grads: Union[torch.Tensor, Sequence[torch.Tensor]],
losses: Optional[Sequence] = None,
) -> None
Calculate the gradient and apply the gradient to the network.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
network |
Module
|
The target network. |
required |
grads |
Union[Tensor, Sequence[Tensor]]
|
The gradients to update. It can be a stack of gradient vectors (at dim 0) or a sequence of gradient vectors. |
required |
losses |
Optional[Sequence]
|
The losses associated with the gradients. The losses will be passed to the weight and length model. If your weight/length model doesn't require loss information, you can set this value as None. Defaults to None. |
None
|
Returns:
Type | Description |
---|---|
None
|
None |
Source code in conflictfree/grad_operator.py
158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 |
|
calculate_gradient
¤
calculate_gradient(
grads: Union[torch.Tensor, Sequence[torch.Tensor]],
losses: Optional[Sequence] = None,
) -> torch.Tensor
Calculates the gradient using the IMTL-G algorithm.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
grads |
Union[Tensor, Sequence[Tensor]]
|
The gradients to update. It can be a stack of gradient vectors (at dim 0) or a sequence of gradient vectors. |
required |
losses |
Optional[Sequence]
|
This parameter should not be set for current operator. Defaults to None. |
None
|
Returns:
Type | Description |
---|---|
Tensor
|
torch.Tensor: The calculated gradient using IMTL-G method. |
Source code in conflictfree/grad_operator.py
307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 |
|
Base Class of Operators¤
conflictfree.grad_operator.GradientOperator
¤
A base class that represents a gradient operator.
Methods:
Name | Description |
---|---|
calculate_gradient |
Calculates the gradient based on the given gradients and losses. |
update_gradient |
Updates the gradient of the network based on the calculated gradient. |
Source code in conflictfree/grad_operator.py
125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 |
|
calculate_gradient
¤
calculate_gradient(
grads: Union[torch.Tensor, Sequence[torch.Tensor]],
losses: Optional[Sequence] = None,
) -> torch.Tensor
Calculates the gradient based on the given gradients and losses.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
grads |
Union[Tensor, Sequence[Tensor]]
|
The gradients to update. It can be a stack of gradient vectors (at dim 0) or a sequence of gradient vectors. |
required |
losses |
Optional[Sequence]
|
The losses associated with the gradients. The losses will be passed to the weight and length model. If your weight/length model doesn't require loss information, you can set this value as None. Defaults to None. |
None
|
Returns:
Type | Description |
---|---|
Tensor
|
torch.Tensor: The calculated gradient. |
Raises:
Type | Description |
---|---|
NotImplementedError
|
If the method is not implemented. |
Source code in conflictfree/grad_operator.py
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 |
|
update_gradient
¤
update_gradient(
network: torch.nn.Module,
grads: Union[torch.Tensor, Sequence[torch.Tensor]],
losses: Optional[Sequence] = None,
) -> None
Calculate the gradient and apply the gradient to the network.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
network |
Module
|
The target network. |
required |
grads |
Union[Tensor, Sequence[Tensor]]
|
The gradients to update. It can be a stack of gradient vectors (at dim 0) or a sequence of gradient vectors. |
required |
losses |
Optional[Sequence]
|
The losses associated with the gradients. The losses will be passed to the weight and length model. If your weight/length model doesn't require loss information, you can set this value as None. Defaults to None. |
None
|
Returns:
Type | Description |
---|---|
None
|
None |
Source code in conflictfree/grad_operator.py
158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 |
|