cSAXS/cdtools
0
0
Fork 0
mirror of https://github.com/cdtools-developers/cdtools.git synced 2026-09-09 21:12:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #1 from cdtools-developers/refactor/reconstructors

Browse Source 
Various suggested changes for the reconstructors branch
This commit is contained in:
Dayne Yoshiki Sasaki
2025-10-17 15:58:47 -07:00
committed by GitHub
parent 82ce845385 7d39015719
commit 6443b16f5c
14 changed files with 293 additions and 176 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/source/examples.rst
+8 -2
View File
@@ -31,9 +31,9 @@ When reading this script, note the basic workflow. After the data is loaded, a m
Next, the model is moved to the GPU using the :code:`model.to` function. Any device understood by :code:`torch.Tensor.to` can be specified here. The next line is a bit more subtle - the dataset is told to move patterns to the GPU before passing them to the model using the :code:`dataset.get_as` function. This function does not move the stored patterns to the GPU. If there is sufficient GPU memory, the patterns can also be pre-moved to the GPU using :code:`dataset.to`, but the speedup is empirically quite small.
Once the device is selected, a reconstruction is run using :code:`model.Adam_optimize`. This is a generator function which will yield at every epoch, to allow some monitoring code to be run.
Once the device is selected, a reconstruction is run using :code:`model.Adam_optimize`. This is a generator function which will yield at the end of every epoch, to allow some monitoring code to be run.
Finally, the results can be studied using :code:`model.inspect(dataet)`, which creates or updates a set of plots showing the current state of the model parameters. :code:`model.compare(dataset)` is also called, which shows how the simulated diffraction patterns compare to the measured diffraction patterns in the dataset.
Finally, the results can be studied using :code:`model.inspect(dataset)`, which creates or updates a set of plots showing the current state of the model parameters. :code:`model.compare(dataset)` is also called, which shows how the simulated diffraction patterns compare to the measured diffraction patterns in the dataset.
Fancy Ptycho
@@ -63,6 +63,12 @@ By default, FancyPtycho will also optimize over the following model parameters,
These corrections can be turned off (on) by calling :code:`model.<parameter>.requires_grad = False #(True)`.
Note as well two other changes that are made in this script, when compared to `simple_ptycho.py`. First, a `Reconstructor` object is explicitly created, in this case an `AdamReconstructor`. This object stores a model, dataset, and pytorch optimizer. It is then used to orchestrate the later reconstruction using a call to `Reconstructor.optimize()`.
We use this pattern, instead of the simpler call to `model.Adam_optimize()`, because having the reconstructor store the optimizer as well as the model and dataset allows the moment estimates to persist between multiple rounds of optimization. This leads to the second change: In this script, we run two optimization loops. The first loop aggressively refines the probe, with a low minibatch size and a high learning rate. The second loop has a smaller learning rate and a larger batch size, which allow for a more precise final estimation of the object.
In this case, we used one reconstructor, but it is possible to create additional reconstructors to zero out all the persistant information in the optimizer, if desired, or even to instantiate multiple reconstructors on the same model with different optimization algorithms (e.g. `model.LBFGS_optimize()`).
Gold Ball Ptycho
----------------
docs/source/index.rst
+1
View File
@@ -9,6 +9,7 @@
general
datasets
models
reconstructors
tools/index
indices_tables
docs/source/reconstructors.rst
+5
View File
@@ -0,0 +1,5 @@
Reconstructors
==============
.. automodule:: cdtools.reconstructors
:members:
examples/fancy_ptycho.py
+12 -4
View File
@@ -19,19 +19,27 @@ device = 'cuda'
model.to(device=device)
dataset.get_as(device=device)
# For this script, we use a slightly different pattern where we explicitly
# create a `Reconstructor` class to orchestrate the reconstruction. The
# reconstructor will store the model and dataset and create an appropriate
# optimizer. This allows the optimizer to persist between loops, along with
# e.g. estimates of the moments of individual parameters
recon = cdtools.reconstructors.AdamReconstructor(model, dataset)
# The learning rate parameter sets the alpha for Adam.
# The beta parameters are (0.9, 0.999) by default
# The batch size sets the minibatch size
for loss in model.Adam_optimize(50, dataset, lr=0.02, batch_size=10):
for loss in recon.optimize(50, lr=0.02, batch_size=10):
print(model.report())
# Plotting is expensive, so we only do it every tenth epoch
if model.epoch % 10 == 0:
model.inspect(dataset)
# It's common to chain several different reconstruction loops. Here, we
# started with an aggressive refinement to find the probe, and now we
# polish the reconstruction with a lower learning rate and larger minibatch
for loss in model.Adam_optimize(50, dataset, lr=0.005, batch_size=50):
# started with an aggressive refinement to find the probe in the previous
# loop, and now we polish the reconstruction with a lower learning rate
# and larger minibatch
for loss in recon.optimize(50, lr=0.005, batch_size=50):
print(model.report())
if model.epoch % 10 == 0:
model.inspect(dataset)
examples/gold_ball_ptycho.py
+7 -4
View File
@@ -29,6 +29,7 @@ model = cdtools.models.FancyPtycho.from_dataset(
probe_fourier_crop=pad
)
# This is a trick that my grandmother taught me, to combat the raster grid
# pathology: we randomze the our initial guess of the probe positions.
# The units here are pixels in the object array.
@@ -42,17 +43,20 @@ device = 'cuda'
model.to(device=device)
dataset.get_as(device=device)
# Create the reconstructor
recon = cdtools.reconstructors.AdamReconstructor(model, dataset)
# This will save out the intermediate results if an exception is thrown
# during the reconstruction
with model.save_on_exception(
'example_reconstructions/gold_balls_earlyexit.h5', dataset):
for loss in model.Adam_optimize(20, dataset, lr=0.005, batch_size=50):
for loss in recon.optimize(20, lr=0.005, batch_size=50):
print(model.report())
if model.epoch % 10 == 0:
model.inspect(dataset)
for loss in model.Adam_optimize(50, dataset, lr=0.002, batch_size=100):
for loss in recon.optimize(50, lr=0.002, batch_size=100):
print(model.report())
if model.epoch % 10 == 0:
model.inspect(dataset)
@@ -64,8 +68,7 @@ with model.save_on_exception(
# Setting schedule=True automatically lowers the learning rate if
# the loss fails to improve after 10 epochs
for loss in model.Adam_optimize(100, dataset, lr=0.001, batch_size=100,
schedule=True):
for loss in recon.optimize(100, lr=0.001, batch_size=100, schedule=True):
print(model.report())
if model.epoch % 10 == 0:
model.inspect(dataset)
examples/gold_ball_split.py
+6 -3
View File
@@ -36,15 +36,18 @@ for label, dataset in zip(labels, datasets):
model.to(device=device)
dataset.get_as(device=device)
# Create the reconstructor
recon = cdtools.reconstructors.AdamReconstructor(model, dataset)
# For batched reconstructions like this, there's no need to live-plot
# the progress
for loss in model.Adam_optimize(20, dataset, lr=0.005, batch_size=50):
for loss in recon.optimize(20, lr=0.005, batch_size=50):
print(model.report())
for loss in model.Adam_optimize(50, dataset, lr=0.002, batch_size=100):
for loss in recon.optimize(50, lr=0.002, batch_size=100):
print(model.report())
for loss in model.Adam_optimize(100, dataset, lr=0.001, batch_size=100,
for loss in recon.optimize(100, lr=0.001, batch_size=100,
schedule=True):
print(model.report())
examples/simple_ptycho.py
+10 -1
View File
@@ -1,3 +1,12 @@
"""
Runs a very simple reconstruction using the SimplePtycho model, which was
designed to be an easy introduction to show how the models are made and used.
For a more realistic example of how to use cdtools for real-world data,
look at fancy_ptycho.py and gold_ball_ptycho.py, both of which use the
more powerful FancyPtycho model and include more information on how to
correct for common sources of error.
"""
import cdtools
from matplotlib import pyplot as plt
@@ -13,7 +22,7 @@ device = 'cuda'
model.to(device=device)
dataset.get_as(device=device)
# We run the actual reconstruction
# We run the reconstruction
for loss in model.Adam_optimize(100, dataset, batch_size=10):
# We print a quick report of the optimization status
print(model.report())
src/cdtools/models/base.py
+49 -67
View File
@@ -40,6 +40,7 @@ import time
from scipy import io
from contextlib import contextmanager
from cdtools.tools.data import nested_dict_to_h5, h5_to_nested_dict, nested_dict_to_numpy, nested_dict_to_torch
from cdtools.reconstructors import AdamReconstructor, LBFGSReconstructor, SGDReconstructor
from cdtools.datasets import CDataset
from typing import List, Union, Tuple
import os
@@ -374,31 +375,24 @@ class CDIModel(t.nn.Module):
only the calculation speed.
"""
# We want to have model.Adam_optimize call AND store
# cdtools.reconstructors.AdamReconstructor to perform reconstructions
# without creating a new reconstructor each time we update the
# hyperparameters.
#
# The only way to do this is to make the Adam reconstructor an
# attribute of the model. But since the Adam reconstructor also
# depends on CDIModel, this seems to give rise to a circular import
# error unless we import cdtools.reconstructors within this method:
if not hasattr(self, 'reconstructor'):
from cdtools.reconstructors import AdamReconstructor
self.reconstructor = AdamReconstructor(model=self,
dataset=dataset,
subset=subset)
reconstructor = AdamReconstructor(
model=self,
dataset=dataset,
subset=subset,
)
# Run some reconstructions
return self.reconstructor.optimize(iterations=iterations,
batch_size=batch_size,
lr=lr,
betas=betas,
schedule=schedule,
amsgrad=amsgrad,
regularization_factor=regularization_factor, # noqa
thread=thread,
calculation_width=calculation_width)
return reconstructor.optimize(
iterations=iterations,
batch_size=batch_size,
lr=lr,
betas=betas,
schedule=schedule,
amsgrad=amsgrad,
regularization_factor=regularization_factor, # noqa
thread=thread,
calculation_width=calculation_width,
)
def LBFGS_optimize(self,
iterations: int,
@@ -445,29 +439,23 @@ class CDIModel(t.nn.Module):
round of gradient accumulation. Does not affect the result, only
the calculation speed.
"""
# We want to have model.LBFGS_optimize store
# cdtools.reconstructors.LBFGSReconstructor as an attribute to run
# reconstructions without generating new reconstructors each time
# CDIModel.LBFGS_optimize is called.
#
# Since the LBFGS reconstructor also depends on CDIModel, a circular
# import error arises unless we import cdtools.reconstructors within
# this method:
if not hasattr(self, 'reconstructor'):
from cdtools.reconstructors import LBFGSReconstructor
self.reconstructor = LBFGSReconstructor(model=self,
dataset=dataset,
subset=subset)
reconstructor = LBFGSReconstructor(
model=self,
dataset=dataset,
subset=subset,
)
# Run some reconstructions
return self.reconstructor.optimize(iterations=iterations,
lr=lr,
history_size=history_size,
regularization_factor=regularization_factor, # noqa
thread=thread,
calculation_width=calculation_width,
line_search_fn=line_search_fn)
return reconstructor.optimize(
iterations=iterations,
lr=lr,
history_size=history_size,
regularization_factor=regularization_factor, # noqa
thread=thread,
calculation_width=calculation_width,
line_search_fn=line_search_fn,
)
def SGD_optimize(self,
iterations: int,
dataset: CDataset,
@@ -520,31 +508,25 @@ class CDIModel(t.nn.Module):
round of gradient accumulation.
"""
# We want to have model.SGD_optimize store
# cdtools.reconstructors.SGDReconstructor as an attribute to run
# reconstructions without generating new reconstructors each time
# CDIModel.SGD_optimize is called.
#
# Since the SGD reconstructor also depends on CDIModel, a circular
# import error arises unless we import cdtools.reconstructors within
# this method:
if not hasattr(self, 'reconstructor'):
from cdtools.reconstructors import SGDReconstructor
self.reconstructor = SGDReconstructor(model=self,
dataset=dataset,
subset=subset)
reconstructor = SGDReconstructor(
model=self,
dataset=dataset,
subset=subset,
)
# Run some reconstructions
return self.reconstructor.optimize(iterations=iterations,
batch_size=batch_size,
lr=lr,
momentum=momentum,
dampening=dampening,
weight_decay=weight_decay,
nesterov=nesterov,
regularization_factor=regularization_factor, # noqa
thread=thread,
calculation_width=calculation_width)
return reconstructor.optimize(
iterations=iterations,
batch_size=batch_size,
lr=lr,
momentum=momentum,
dampening=dampening,
weight_decay=weight_decay,
nesterov=nesterov,
regularization_factor=regularization_factor, # noqa
thread=thread,
calculation_width=calculation_width,
)
def report(self):
src/cdtools/reconstructors/adam.py
+36 -18
View File
@@ -6,12 +6,17 @@ The Reconstructor class is designed to resemble so-called
'Trainer' classes that (in the language of the AI/ML folks) handles
the 'training' of a model given some dataset and optimizer.
"""
from __future__ import annotations
from typing import TYPE_CHECKING
import torch as t
from cdtools.datasets.ptycho_2d_dataset import Ptycho2DDataset
from cdtools.models import CDIModel
from typing import Tuple, List, Union
from cdtools.reconstructors import Reconstructor
if TYPE_CHECKING:
from cdtools.models import CDIModel
from cdtools.datasets.ptycho_2d_dataset import Ptycho2DDataset
__all__ = ['AdamReconstructor']
@@ -46,10 +51,12 @@ class AdamReconstructor(Reconstructor):
dataset: Ptycho2DDataset,
subset: List[int] = None):
super().__init__(model, dataset, subset)
# Define the optimizer for use in this subclass
self.optimizer = t.optim.Adam(self.model.parameters())
optimizer = t.optim.Adam(model.parameters())
super().__init__(model, dataset, optimizer, subset=subset)
def adjust_optimizer(self,
lr: int = 0.005,
@@ -74,11 +81,13 @@ class AdamReconstructor(Reconstructor):
param_group['betas'] = betas
param_group['amsgrad'] = amsgrad
def optimize(self,
iterations: int,
batch_size: int = 15,
lr: float = 0.005,
betas: Tuple[float] = (0.9, 0.999),
custom_data_loader: t.utils.data.DataLoader = None,
schedule: bool = False,
amsgrad: bool = False,
regularization_factor: Union[float, List[float]] = None,
@@ -94,6 +103,12 @@ class AdamReconstructor(Reconstructor):
(formerly `CDIModel.AD_optimize`) to run a round of reconstruction
once the dataloader and optimizer hyperparameters have been
set up.
The `batch_size` parameter sets the batch size for the default
dataloader. If a custom data loader is desired, it can be passed
in to the `custom_data_loader` argument, which will override the
`batch_size` and `shuffle` parameters
Parameters
----------
@@ -110,6 +125,9 @@ class AdamReconstructor(Reconstructor):
schedule : bool
Optional, create a learning rate scheduler
(torch.optim.lr_scheduler._LRScheduler).
custom_data_loader : t.utils.data.DataLoader
Optional, a custom DataLoader to use. If set, will override
batch_size.
amsgrad : bool
Optional, whether to use the AMSGrad variant of this algorithm.
regularization_factor : float or list(float)
@@ -133,18 +151,13 @@ class AdamReconstructor(Reconstructor):
f'{regularization_factor}, and schedule = {schedule}.\n'
)
# 1) The subset statement is contained in Reconstructor.__init__
# 2) Set up / re-initialize the data laoder
self.setup_dataloader(batch_size=batch_size, shuffle=shuffle)
# 3) The optimizer is created in self.__init__, but the
# hyperparameters need to be set up with self.adjust_optimizer
# The optimizer is created in self.__init__, but the
# hyperparameters need to be set up with self.adjust_optimizer
self.adjust_optimizer(lr=lr,
betas=betas,
amsgrad=amsgrad)
# 4) Set up the scheduler
# Set up the scheduler
if schedule:
self.scheduler = \
t.optim.lr_scheduler.ReduceLROnPlateau(self.optimizer,
@@ -153,8 +166,13 @@ class AdamReconstructor(Reconstructor):
else:
self.scheduler = None
# 5) This is analagous to making a call to CDIModel.AD_optimize
return super(AdamReconstructor, self).optimize(iterations,
regularization_factor,
thread,
calculation_width)
# Now, we run the optimize routine defined in the base class
return super(AdamReconstructor, self).optimize(
iterations,
batch_size=batch_size,
custom_data_loader=custom_data_loader,
regularization_factor=regularization_factor,
thread=thread,
calculation_width=calculation_width,
shuffle=shuffle,
)
src/cdtools/reconstructors/base.py
+83 -27
View File
@@ -8,16 +8,21 @@ the 'training' of a model given some dataset and optimizer.
The subclasses of Reconstructor are required to implement
their own data loaders and optimizer adjusters
"""
from __future__ import annotations
from typing import TYPE_CHECKING
import torch as t
from torch.utils import data as td
import threading
import queue
import time
from cdtools.datasets import CDataset
from cdtools.models import CDIModel
from typing import List, Union
if TYPE_CHECKING:
from cdtools.models import CDIModel
from cdtools.datasets import CDataset
__all__ = ['Reconstructor']
@@ -35,46 +40,51 @@ class Reconstructor:
Model for CDI/ptychography reconstruction
dataset: CDataset
The dataset to reconstruct against
optimizer: torch.optim.Optimizer
The optimizer to use for the reconstruction
subset : list(int) or int
Optional, a pattern index or list of pattern indices to use
Important attributes:
- **model** -- Always points to the core model used.
- **optimizer** -- A `torch.optim.Optimizer` that must be defined when
initializing the Reconstructor subclass.
- **scheduler** -- A `torch.optim.lr_scheduler` that may be defined during
the `optimize` method.
- **data_loader** -- A torch.utils.data.DataLoader that is defined by
calling the `setup_dataloader` method.
Attributes
----------
model : CDIModel
Points to the core model used.
optimizer : torch.optim.Optimizer
Must be defined when initializing the Reconstructor subclass.
scheduler : torch.optim.lr_scheduler, optional
May be defined during the ``optimize`` method.
data_loader : torch.utils.data.DataLoader
Defined by calling the ``setup_dataloader`` method.
"""
def __init__(self,
model: CDIModel,
dataset: CDataset,
optimizer: t.optim.Optimizer,
subset: Union[int, List[int]] = None):
# Store parameters as attributes of Reconstructor
self.subset = subset
# Initialize attributes that must be defined by the subclasses
self.optimizer = None
self.scheduler = None
self.data_loader = None
# Store the original model
self.model = model
self.optimizer = optimizer
# Store the dataset
# Store the dataset, clipping it to a subset if needed
if subset is not None:
# if subset is just one pattern, turn into a list for convenience
if isinstance(subset, int):
subset = [subset]
dataset = td.Subset(dataset, subset)
self.dataset = dataset
# Initialize attributes that must be defined by the subclasses
self.scheduler = None
self.data_loader = None
def setup_dataloader(self,
batch_size: int = None,
shuffle: bool = True):
"""
Sets up / re-initializes the dataloader.
Sets up or re-initializes the dataloader.
Parameters
----------
@@ -91,6 +101,7 @@ class Reconstructor:
else:
self.data_loader = td.Dataloader(self.dataset)
def adjust_optimizer(self, **kwargs):
"""
Change hyperparameters for the utilized optimizer.
@@ -100,7 +111,8 @@ class Reconstructor:
"""
raise NotImplementedError()
def _run_epoch(self,
def run_epoch(self,
stop_event: threading.Event = None,
regularization_factor: Union[float, List[float]] = None,
calculation_width: int = 10):
@@ -128,6 +140,18 @@ class Reconstructor:
diffraction pattern intensity
"""
# Setting this as an explicit catch makes me feel more comfortable
# exposing it as a public method. This way a user won't be confused
# if they try to use this directly
if self.data_loader is None:
raise RuntimeError(
'No data loader was defined. Please run '
'Reconstructor.setup_dataloader() before running '
'Reconstructor.run_epoch(), or use Reconstructor.optimize(), '
'which does it automatically.'
)
# Initialize some tracking variables
normalization = 0
loss = 0
@@ -207,9 +231,12 @@ class Reconstructor:
def optimize(self,
iterations: int,
batch_size: int = 1,
custom_data_loader: torch.utils.data.DataLoader = None,
regularization_factor: Union[float, List[float]] = None,
thread: bool = True,
calculation_width: int = 10):
calculation_width: int = 10,
shuffle=True):
"""
Runs a round of reconstruction using the provided optimizer
@@ -228,10 +255,25 @@ class Reconstructor:
the plots. This behavior can be turned off by setting the keyword
argument 'thread' to False.
The `batch_size` parameter sets the batch size for the default
dataloader. If a custom data loader is desired, it can be passed
in to the `custom_data_loader` argument, which will override the
`batch_size` and `shuffle` parameters
Please see `AdamReconstructor.optimize()` for an example of how to
override this function when designing a subclass
Parameters
----------
iterations : int
How many epochs of the algorithm to run.
batch_size : int
Optional, the batch size to use. Default is 1. This is typically
overridden by subclasses with an appropriate default for the
specific optimizer.
custom_data_loader : torch.utils.data.DataLoader
Optional, a custom DataLoader to use. Will override batch_size
if set.
regularization_factor : float or list(float)
Optional, if the model has a regularizer defined, the set of
parameters to pass the regularizer method.
@@ -242,6 +284,10 @@ class Reconstructor:
Default 10, how many translations to pass through at once for each
round of gradient accumulation. This does not affect the result,
but may affect the calculation speed.
shuffle : bool
Optional, enable/disable shuffling of the dataset. This option
is intended for diagnostic purposes and should be left as True.
Yields
------
@@ -250,6 +296,11 @@ class Reconstructor:
diffraction pattern intensity.
"""
if custom_data_loader is None:
self.setup_dataloader(batch_size=batch_size, shuffle=shuffle)
else:
self.data_loader = custom_data_loader
# We store the current optimizer as a model parameter so that
# it can be saved and loaded for checkpointing
self.current_optimizer = self.optimizer
@@ -265,8 +316,10 @@ class Reconstructor:
yield float('nan')
continue
yield self._run_epoch(regularization_factor=regularization_factor, # noqa
calculation_width=calculation_width)
yield self.run_epoch(
regularization_factor=regularization_factor, # noqa
calculation_width=calculation_width,
)
# But if we do want to thread, it's annoying:
else:
@@ -277,9 +330,12 @@ class Reconstructor:
def target():
try:
result_queue.put(
self._run_epoch(stop_event=stop_event,
regularization_factor=regularization_factor, # noqa
calculation_width=calculation_width))
self.run_epoch(
stop_event=stop_event,
regularization_factor=regularization_factor, # noqa
calculation_width=calculation_width,
)
)
except Exception as e:
# If something bad happens, put the exception into the
# result queue
src/cdtools/reconstructors/lbfgs.py
+26 -17
View File
@@ -6,12 +6,18 @@ The Reconstructor class is designed to resemble so-called
'Trainer' classes that (in the language of the AI/ML folks) handles
the 'training' of a model given some dataset and optimizer.
"""
from __future__ import annotations
from typing import TYPE_CHECKING
import torch as t
from cdtools.datasets.ptycho_2d_dataset import Ptycho2DDataset
from cdtools.models import CDIModel
from typing import List, Union
from cdtools.reconstructors import Reconstructor
if TYPE_CHECKING:
from cdtools.models import CDIModel
from cdtools.datasets.ptycho_2d_dataset import Ptycho2DDataset
__all__ = ['LBFGSReconstructor']
@@ -47,10 +53,16 @@ class LBFGSReconstructor(Reconstructor):
dataset: Ptycho2DDataset,
subset: List[int] = None):
super().__init__(model, dataset, subset)
# Define the optimizer for use in this subclass
self.optimizer = t.optim.LBFGS(self.model.parameters())
optimizer = t.optim.LBFGS(model.parameters())
super().__init__(
model,
dataset,
optimizer,
subset=subset,
)
def adjust_optimizer(self,
lr: int = 0.005,
@@ -115,20 +127,17 @@ class LBFGSReconstructor(Reconstructor):
round of gradient accumulation. Does not affect the result, only
the calculation speed.
"""
# 1) The subset statement is contained in Reconstructor.__init__
# 2) Set up / re-initialize the data loader. For LBFGS, we load
# all the data at once.
self.setup_dataloader(batch_size=len(self.dataset))
# 3) The optimizer is created in self.__init__, but the
# hyperparameters need to be set up with self.adjust_optimizer
# The optimizer is created in self.__init__, but the
# hyperparameters need to be set up with self.adjust_optimizer
self.adjust_optimizer(lr=lr,
history_size=history_size,
line_search_fn=line_search_fn)
# 4) This is analagous to making a call to CDIModel.AD_optimize
return super(LBFGSReconstructor, self).optimize(iterations,
regularization_factor,
thread,
calculation_width)
# Now, we run the optimize routine defined in the base class
return super(LBFGSReconstructor, self).optimize(
iterations,
batch_size=len(self.dataset),
regularization_factor=regularization_factor,
thread=thread,
calculation_width=calculation_width)
src/cdtools/reconstructors/sgd.py
+28 -21
View File
@@ -6,12 +6,18 @@ The Reconstructor class is designed to resemble so-called
'Trainer' classes that (in the language of the AI/ML folks) handles
the 'training' of a model given some dataset and optimizer.
"""
from __future__ import annotations
from typing import TYPE_CHECKING
import torch as t
from cdtools.datasets.ptycho_2d_dataset import Ptycho2DDataset
from cdtools.models import CDIModel
from typing import List, Union
from cdtools.reconstructors import Reconstructor
if TYPE_CHECKING:
from cdtools.models import CDIModel
from cdtools.datasets.ptycho_2d_dataset import Ptycho2DDataset
__all__ = ['SGDReconstructor']
@@ -43,11 +49,17 @@ class SGDReconstructor(Reconstructor):
dataset: Ptycho2DDataset,
subset: List[int] = None):
super().__init__(model, dataset, subset)
# Define the optimizer for use in this subclass
self.optimizer = t.optim.SGD(self.model.parameters())
optimizer = t.optim.SGD(model.parameters())
super().__init__(
model,
dataset,
optimizer,
subset=subset,
)
def adjust_optimizer(self,
lr: int = 0.005,
momentum: float = 0,
@@ -82,7 +94,7 @@ class SGDReconstructor(Reconstructor):
def optimize(self,
iterations: int,
batch_size: int = None,
batch_size: int = 15,
lr: float = 2e-7,
momentum: float = 0,
dampening: float = 0,
@@ -133,25 +145,20 @@ class SGDReconstructor(Reconstructor):
Optional, enable/disable shuffling of the dataset. This option
is intended for diagnostic purposes and should be left as True.
"""
# 1) The subset statement is contained in Reconstructor.__init__
# 2) Set up / re-initialize the data laoder
if batch_size is not None:
self.setup_dataloader(batch_size=batch_size, shuffle=shuffle)
else:
# Use default torch dataloader parameters
self.setup_dataloader(batch_size=1, shuffle=False)
# 3) The optimizer is created in self.__init__, but the
# hyperparameters need to be set up with self.adjust_optimizer
# The optimizer is created in self.__init__, but the
# hyperparameters need to be set up with self.adjust_optimizer
self.adjust_optimizer(lr=lr,
momentum=momentum,
dampening=dampening,
weight_decay=weight_decay,
nesterov=nesterov)
# 4) This is analagous to making a call to CDIModel.AD_optimize
return super(SGDReconstructor, self).optimize(iterations,
regularization_factor,
thread,
calculation_width)
# Now, we run the optimize routine defined in the base class
return super(SGDReconstructor, self).optimize(
iterations,
batch_size=batch_size,
regularization_factor=regularization_factor,
thread=thread,
calculation_width=calculation_width,
)
tests/models/test_fancy_ptycho.py
+7 -2
View File
@@ -56,8 +56,8 @@ def test_lab_ptycho(lab_ptycho_cxi, reconstruction_device, show_plot):
dataset,
n_modes=3,
oversampling=2,
exponentiate_obj=True,
dm_rank=2,
exponentiate_obj=True,
probe_support_radius=120,
propagation_distance=5e-3,
units='mm',
@@ -70,7 +70,7 @@ def test_lab_ptycho(lab_ptycho_cxi, reconstruction_device, show_plot):
model.to(device=reconstruction_device)
dataset.get_as(device=reconstruction_device)
for loss in model.Adam_optimize(70, dataset, lr=0.02, batch_size=10):
for loss in model.Adam_optimize(50, dataset, lr=0.02, batch_size=10):
print(model.report())
if show_plot and model.epoch % 10 == 0:
model.inspect(dataset)
@@ -79,6 +79,11 @@ def test_lab_ptycho(lab_ptycho_cxi, reconstruction_device, show_plot):
print(model.report())
if show_plot and model.epoch % 10 == 0:
model.inspect(dataset)
for loss in model.Adam_optimize(25, dataset, lr=0.001, batch_size=50):
print(model.report())
if show_plot and model.epoch % 10 == 0:
model.inspect(dataset)
model.tidy_probes()
tests/test_reconstructors.py
+15 -10
View File
@@ -16,7 +16,8 @@ def test_Adam_gold_balls(gold_ball_cxi, reconstruction_device, show_plot):
2) We are only using the single-GPU dataloading method
3) Ensure `recon.model` points to the original `model`
4) Reconstructions performed by `Adam.optimize` and
`model.Adam_optimize` calls produce identical results.
`model.Adam_optimize` calls produce identical results when
run over one round of optimization.
5) The quality of the reconstruction remains below a specified
threshold.
5) Ensure that the FancyPtycho model works fine and dandy with the
@@ -91,7 +92,8 @@ def test_Adam_gold_balls(gold_ball_cxi, reconstruction_device, show_plot):
' reconstruction_device,', reconstruction_device)
t.manual_seed(0)
for i, iterations in enumerate(epoch_tup):
# We only need to test the first loop to ensure it's identical
for i, iterations in enumerate(epoch_tup[:1]):
for loss in model.Adam_optimize(iterations,
dataset,
lr=lr_tup[i],
@@ -106,14 +108,15 @@ def test_Adam_gold_balls(gold_ball_cxi, reconstruction_device, show_plot):
model.inspect(dataset)
model.compare(dataset)
# Ensure equivalency between the model reconstructions
assert np.allclose(model_recon.loss_history[-1], model.loss_history[-1])
# Ensure equivalency between the model reconstructions during the first
# pass, where they should be identical
assert np.allclose(model_recon.loss_history[:epoch_tup[0]], model.loss_history[:epoch_tup[0]])
# Ensure reconstructions have reached a certain loss tolerance. This just
# comes from running a reconstruction when it was working well and
# choosing a rough value. If it triggers this assertion error, something
# changed to make the final quality worse!
assert model.loss_history[-1] < 0.0001
assert model_recon.loss_history[-1] < 0.0001
@pytest.mark.slow
@@ -127,7 +130,8 @@ def test_LBFGS_RPI(optical_data_ss_cxi,
hyperparameters
2) Ensure `recon.model` points to the original `model`
3) Reconstructions performed by `LBFGS.optimize` and
`model.LBFGS_optimize` calls produce identical results.
`model.LBFGS_optimize` calls produce identical results when
run over one round of reconstruction.
4) The quality of the reconstruction remains below a specified
threshold.
5) Ensure that the RPI model works fine and dandy with the
@@ -184,7 +188,7 @@ def test_LBFGS_RPI(optical_data_ss_cxi,
print('Running reconstruction using CDIModel.LBFGS_optimize.' +
'optimize on provided reconstruction_device,', reconstruction_device)
t.manual_seed(0)
for i, iterations in enumerate(epoch_tup):
for i, iterations in enumerate(epoch_tup[:1]):
for loss in model.LBFGS_optimize(iterations,
dataset,
lr=0.4,
@@ -198,12 +202,12 @@ def test_LBFGS_RPI(optical_data_ss_cxi,
model.compare(dataset)
# Check loss equivalency between the two reconstructions
assert np.allclose(model.loss_history[-1], model_recon.loss_history[-1])
assert np.allclose(model.loss_history[:epoch_tup[0]], model_recon.loss_history[:epoch_tup[0]])
# The final loss when testing this was 2.28607e-3. Based on this, we set
# a threshold of 2.3e-3 for the tested loss. If this value has been
# exceeded, the reconstructions have gotten worse.
assert model.loss_history[-1] < 0.0023
assert model_recon.loss_history[-1] < 0.0023
@pytest.mark.slow
@@ -214,7 +218,8 @@ def test_SGD_gold_balls(gold_ball_cxi, reconstruction_device, show_plot):
hyperparameters
3) Ensure `recon.model` points to the original `model`
4) Reconstructions performed by `SGD.optimize` and
`model.SGD_optimize` calls produce identical results.
`model.SGD_optimize` calls produce identical results
when run over one round of reconstruction.
5) The quality of the reconstruction remains below a specified
threshold.
5) Ensure that the FancyPtycho model works fine and dandy with the