reflectorch 1.0.0__py3-none-any.whl

reflectorch/models/networks/mlp_networks.py

@@ -0,0 +1,256 @@
+# -*- coding: utf-8 -*-
+
+import math
+from typing import Optional
+import torch
+from torch import nn, cat, split, Tensor
+
+from reflectorch.models.networks.residual_net import ResidualMLP
+from reflectorch.models.encoders.conv_encoder import ConvEncoder
+from reflectorch.models.encoders.fno import FnoEncoder
+from reflectorch.models.activations import activation_by_name
+
+class NetworkWithPriorsConvEmb(nn.Module):
+    """MLP network with 1D CNN embedding network
+
+    .. image:: ../documentation/FigureReflectometryNetwork.png
+        :width: 800px
+        :align: center
+
+    Args:
+        in_channels (int, optional): the number of input channels of the 1D CNN. Defaults to 1.
+        hidden_channels (tuple, optional): list with the number of channels for each layer of the 1D CNN. Defaults to (32, 64, 128, 256, 512).
+        dim_embedding (int, optional): the dimension of the embedding produced by the 1D CNN. Defaults to 128.
+        dim_avpool (int, optional): the type of activation function in the 1D CNN. Defaults to 1.
+        embedding_net_activation (str, optional): the type of activation function in the 1D CNN. Defaults to 'gelu'.
+        use_batch_norm (bool, optional): whether to use batch normalization (in both the 1D CNN and the MLP). Defaults to False.
+        dim_out (int, optional): the dimension of the output produced by the MLP. Defaults to 8.
+        layer_width (int, optional): the width of a linear layer in the MLP. Defaults to 512.
+        num_blocks (int, optional): the number of residual blocks in the MLP. Defaults to 4.
+        repeats_per_block (int, optional): the number of normalization/activation/linear repeats in a block. Defaults to 2.
+        mlp_activation (str, optional):  the type of activation function in the MLP. Defaults to 'gelu'.
+        dropout_rate (float, optional): dropout rate for each block. Defaults to 0.0.
+        use_selu_init (bool, optional): whether to use the special weights initialization for the 'selu' activation function. Defaults to False.
+        pretrained_embedding_net (str, optional): the path to the weights of a pretrained embedding network. Defaults to None.
+        residual (bool, optional): whether the blocks have a residual skip connection. Defaults to True.
+        adaptive_activation (bool, optional): must be set to ``True`` if the activation function is adaptive. Defaults to False.
+        conditioning (str, optional): the manner in which the prior bounds are provided as input to the network. Defaults to 'concat'.
+    """
+    def __init__(self,
+                 in_channels: int = 1,
+                 hidden_channels: tuple = (32, 64, 128, 256, 512),
+                 dim_embedding: int = 128,
+                 dim_avpool: int = 1,
+                 embedding_net_activation: str = 'gelu',
+                 use_batch_norm: bool = False,
+                 dim_out: int = 8,
+                 layer_width: int = 512,
+                 num_blocks: int = 4,
+                 repeats_per_block: int = 2,
+                 mlp_activation: str = 'gelu',
+                 dropout_rate: float = 0.0,
+                 use_selu_init: bool = False,
+                 pretrained_embedding_net: str = None,
+                 residual: bool = True,
+                 adaptive_activation: bool = False,
+                 conditioning: str = 'concat',
+                 ):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.conditioning = conditioning
+
+        self.embedding_net = ConvEncoder(
+            in_channels=in_channels, 
+            hidden_channels=hidden_channels, 
+            dim_latent=dim_embedding, 
+            dim_avpool=dim_avpool,
+            use_batch_norm=use_batch_norm, 
+            activation=embedding_net_activation
+        )
+        
+        self.dim_prior_bounds = 2 * dim_out
+
+        if conditioning == 'concat':
+            dim_mlp_in = dim_embedding + self.dim_prior_bounds
+            dim_condition = 0
+        elif conditioning == 'glu' or conditioning == 'film':
+            dim_mlp_in = dim_embedding
+            dim_condition = self.dim_prior_bounds
+        else:
+            raise NotImplementedError
+
+        self.mlp = ResidualMLP(
+            dim_in=dim_mlp_in,
+            dim_out=dim_out,
+            dim_condition=dim_condition,
+            layer_width=layer_width,
+            num_blocks=num_blocks,
+            repeats_per_block=repeats_per_block,
+            activation=mlp_activation,
+            use_batch_norm=use_batch_norm,
+            dropout_rate=dropout_rate,
+            residual=residual,
+            adaptive_activation=adaptive_activation,
+            conditioning=conditioning,
+        )
+
+        if use_selu_init and embedding_net_activation == 'selu':
+            self.embedding_net.apply(selu_init)
+
+        if use_selu_init and mlp_activation == 'selu':
+            self.mlp.apply(selu_init)
+
+        if pretrained_embedding_net:
+            self.embedding_net.load_weights(pretrained_embedding_net)
+
+
+    def forward(self, curves: Tensor, bounds: Tensor, q_values: Optional[Tensor] = None):
+        """
+        Args:
+            curves (Tensor): reflectivity curves
+            bounds (Tensor): prior bounds
+            q_values (Tensor, optional): q values. Defaults to None.
+
+        Returns:
+            Tensor: prediction
+        """
+        if q_values is not None:
+            curves = torch.cat([curves[:, None, :], q_values[:, None, :]], dim=1)
+
+        if self.conditioning == 'concat': 
+            x = torch.cat([self.embedding_net(curves), bounds], dim=-1)
+            x = self.mlp(x)
+
+        elif self.conditioning == 'glu' or self.conditioning == 'film':
+            x = self.mlp(self.embedding_net(curves), condition=bounds)
+
+        return x
+    
+    
+class NetworkWithPriorsFnoEmb(nn.Module):
+    """MLP network with FNO embedding network
+
+    Args:
+        in_channels (int, optional): the number of input channels to the FNO-based embedding network. Defaults to 2.
+        dim_embedding (int, optional): the dimension of the embedding produced by the FNO. Defaults to 128.
+        modes (int, optional):  the number of Fourier modes that are utilized. Defaults to 16.
+        width_fno (int, optional): the number of channels in the FNO blocks. Defaults to 64.
+        embedding_net_activation (str, optional):  the type of activation function in the embedding network. Defaults to 'gelu'.
+        n_fno_blocks (int, optional): the number of FNO blocks. Defaults to 6.
+        fusion_self_attention (bool, optional): if ``True`` a fusion layer is used after the FNO blocks to produce the final output. Defaults to False.
+        dim_out (int, optional): the dimension of the output produced by the MLP. Defaults to 8.
+        layer_width (int, optional): the width of a linear layer in the MLP. Defaults to 512.
+        num_blocks (int, optional): the number of residual blocks in the MLP. Defaults to 4.
+        repeats_per_block (int, optional): the number of normalization/activation/linear repeats in a block. Defaults to 2.
+        use_batch_norm (bool, optional): whether to use batch normalization (only in the MLP). Defaults to False.
+        mlp_activation (str, optional):  the type of activation function in the MLP. Defaults to 'gelu'.
+        dropout_rate (float, optional): dropout rate for each block. Defaults to 0.0.
+        use_selu_init (bool, optional): whether to use the special weights initialization for the 'selu' activation function. Defaults to False.
+        residual (bool, optional): whether the blocks have a residual skip connection. Defaults to True.
+        adaptive_activation (bool, optional): must be set to ``True`` if the activation function is adaptive. Defaults to False.
+        conditioning (str, optional): the manner in which the prior bounds are provided as input to the network. Defaults to 'concat'.
+    """
+    def __init__(self,
+                 in_channels: int = 2,
+                 dim_embedding: int = 128,
+                 modes: int = 16,
+                 width_fno: int = 64,
+                 embedding_net_activation: str = 'gelu',
+                 n_fno_blocks : int = 6,
+                 fusion_self_attention: bool = False,
+                 dim_out: int = 8,
+                 layer_width: int = 512,
+                 num_blocks: int = 4,
+                 repeats_per_block: int = 2,
+                 use_batch_norm: bool = False,
+                 mlp_activation: str = 'gelu',
+                 dropout_rate: float = 0.0,
+                 use_selu_init: bool = False,
+                 residual: bool = True,
+                 adaptive_activation: bool = False,
+                 conditioning: str = 'concat',
+                 ):
+        super().__init__()
+
+        self.conditioning = conditioning
+
+        self.embedding_net = FnoEncoder(
+            ch_in=in_channels, 
+            dim_embedding=dim_embedding, 
+            modes=modes, 
+            width_fno=width_fno, 
+            n_fno_blocks=n_fno_blocks, 
+            activation=embedding_net_activation,
+            fusion_self_attention=fusion_self_attention
+        )
+
+        self.dim_prior_bounds = 2 * dim_out
+
+        if conditioning == 'concat':
+            dim_mlp_in = dim_embedding + self.dim_prior_bounds
+            dim_condition = 0
+        elif conditioning == 'glu' or conditioning == 'film':
+            dim_mlp_in = dim_embedding
+            dim_condition = self.dim_prior_bounds
+        else:
+            raise NotImplementedError
+
+        self.mlp = ResidualMLP(
+            dim_in=dim_mlp_in,
+            dim_out=dim_out,
+            dim_condition=dim_condition,
+            layer_width=layer_width,
+            num_blocks=num_blocks,
+            repeats_per_block=repeats_per_block,
+            activation=mlp_activation,
+            use_batch_norm=use_batch_norm,
+            dropout_rate=dropout_rate,
+            residual=residual,
+            adaptive_activation=adaptive_activation,
+            conditioning=conditioning,
+        )
+
+        if use_selu_init and embedding_net_activation == 'selu':
+            self.FnoEncoder.apply(selu_init)
+
+        if use_selu_init and mlp_activation == 'selu':
+            self.mlp.apply(selu_init)
+
+            
+    def forward(self, curves: Tensor, bounds: Tensor, q_values: Optional[Tensor] =None):
+        """
+        Args:
+            curves (Tensor): reflectivity curves
+            bounds (Tensor): prior bounds
+            q_values (Tensor, optional): q values. Defaults to None.
+
+        Returns:
+            Tensor: prediction
+        """
+        if curves.dim() < 3:
+            curves = curves[:, None, :]
+        if q_values is not None:
+            curves = torch.cat([curves, q_values[:, None, :]], dim=1)
+
+        if self.conditioning == 'concat': 
+            x = torch.cat([self.embedding_net(curves), bounds], dim=-1)
+            x = self.mlp(x)
+
+        elif self.conditioning == 'glu' or self.conditioning == 'film':
+            x = self.mlp(self.embedding_net(curves), condition=bounds)
+
+        return x
+
+
+
+def selu_init(m):
+    if isinstance(m, (nn.Conv1d, nn.Linear)):
+        m.weight.data.normal_(0.0, 0.5 / math.sqrt(m.weight.numel()))
+        nn.init.constant_(m.bias, 0)
+    elif isinstance(m, nn.BatchNorm1d):
+        size = m.weight.size()
+        fan_in = size[0]
+
+        m.weight.data.normal_(0.0, 1.0 / math.sqrt(fan_in))
+        m.bias.data.fill_(0)

reflectorch/models/networks/residual_net.py

@@ -0,0 +1,131 @@
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torch.nn import init
+
+from reflectorch.models.activations import activation_by_name
+
+class ResidualMLP(nn.Module):
+    """Multilayer perceptron with residual blocks (BN-Act-Linear-BN-Act-Linear)"""
+
+    def __init__(
+            self,
+            dim_in: int,
+            dim_out: int,
+            dim_condition: int = 0,
+            layer_width: int = 512,
+            num_blocks: int = 4,
+            repeats_per_block: int = 2,
+            activation: str = 'relu',
+            use_batch_norm: bool = True,
+            dropout_rate: float = 0.0,
+            residual: bool = True,
+            adaptive_activation: bool = False,
+            conditioning: str = 'glu',
+    ):
+        super().__init__()
+
+        dim_first_layer = dim_in + dim_condition
+        self.first_layer = nn.Linear(dim_first_layer, layer_width)
+        self.blocks = nn.ModuleList(
+            [
+                ResidualBlock(
+                    layer_width=layer_width,
+                    dim_condition=dim_condition,
+                    repeats_per_block=repeats_per_block,
+                    activation=activation,
+                    use_batch_norm=use_batch_norm,
+                    dropout_rate=dropout_rate,
+                    residual=residual,
+                    adaptive_activation=adaptive_activation,
+                    conditioning = conditioning,
+                )
+                for _ in range(num_blocks)
+            ]
+        )
+        self.last_layer = nn.Linear(layer_width, dim_out)
+
+    def forward(self, x, condition=None):
+        if condition is None:
+            x = self.first_layer(x)
+        else:
+            x = self.first_layer(torch.cat([x, condition], dim=-1))
+
+        for block in self.blocks:
+            x = block(x, condition=condition)
+        x = self.last_layer(x)
+
+        return x
+
+
+class ResidualBlock(nn.Module):
+    """Residual block (BN-Act-Linear-BN-Act-Linear)"""
+
+    def __init__(
+            self,
+            layer_width: int,
+            dim_condition: int = 0,
+            repeats_per_block: int = 2,
+            activation: str = 'relu',
+            use_batch_norm: bool = False,
+            dropout_rate: float = 0.0,
+            residual: bool = True,
+            adaptive_activation: bool = False,
+            conditioning: str = 'glu',
+    ):
+        super().__init__()
+         
+        self.residual = residual
+        self.repeats_per_block = repeats_per_block
+        self.use_batch_norm = use_batch_norm
+        self.dropout_rate = dropout_rate
+        self.adaptive_activation = adaptive_activation
+        self.conditioning = conditioning
+
+        if not adaptive_activation:
+            self.activation = activation_by_name(activation)()
+        else:
+            self.activation_layers = nn.ModuleList(
+                [activation_by_name(activation)() for _ in range(repeats_per_block)]
+            )
+
+        if use_batch_norm:
+            self.batch_norm_layers = nn.ModuleList(
+                [nn.BatchNorm1d(layer_width, eps=1e-3) for _ in range(repeats_per_block)]
+            )
+
+        if dim_condition:
+            if conditioning == 'glu':
+                self.condition_layer = nn.Linear(dim_condition, layer_width)
+            elif conditioning == 'film':
+                self.condition_layer = nn.Linear(dim_condition, 2*layer_width)
+
+        self.linear_layers = nn.ModuleList(
+            [nn.Linear(layer_width, layer_width) for _ in range(repeats_per_block)]
+        )
+
+        if self.dropout_rate > 0:
+            self.dropout = nn.Dropout(p=dropout_rate)
+
+    def forward(self, x, condition=None):
+        x0 = x
+         
+        for i in range(self.repeats_per_block):
+            if self.use_batch_norm:
+                x = self.batch_norm_layers[i](x)
+            if not self.adaptive_activation:
+                x = self.activation(x)
+            else:
+                x = self.activation_layers[i](x)
+            if self.dropout_rate > 0 and i == self.repeats_per_block - 1:
+                x = self.dropout(x)
+            x = self.linear_layers[i](x)
+        
+        if condition is not None:
+            if self.conditioning == 'glu':
+                x = F.glu(torch.cat((x, self.condition_layer(condition)), dim=-1), dim=-1)
+            elif self.conditioning == 'film':
+                gamma, beta = torch.chunk(self.condition_layer(condition), chunks=2, dim=-1)
+                x = x * gamma + beta
+
+        return x0 + x if self.residual else x

reflectorch/paths.py

@@ -0,0 +1,33 @@
+# -*- coding: utf-8 -*-
+#
+#
+# This source code is licensed under the GPL license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Union
+from pathlib import Path
+
+__all__ = [
+    'ROOT_DIR',
+    'SAVED_MODELS_DIR',
+    'SAVED_LOSSES_DIR',
+    'RUN_SCRIPTS_DIR',
+    'CONFIG_DIR',
+    'TESTS_PATH',
+    'TEST_DATA_PATH',
+    'listdir',
+]
+
+ROOT_DIR: Path = Path(__file__).parents[1]
+SAVED_MODELS_DIR: Path = ROOT_DIR / 'saved_models'
+SAVED_LOSSES_DIR: Path = ROOT_DIR / 'saved_losses'
+RUN_SCRIPTS_DIR: Path = ROOT_DIR / 'runs'
+CONFIG_DIR: Path = ROOT_DIR / 'configs'
+TESTS_PATH: Path = ROOT_DIR / 'tests'
+TEST_DATA_PATH = TESTS_PATH / 'data'
+
+
+def listdir(path: Union[Path, str], pattern: str = '*', recursive: bool = False, *, sort_key=None, reverse=False):
+    path = Path(path)
+    func = path.rglob if recursive else path.glob
+    return sorted(list(func(pattern)), key=sort_key, reverse=reverse)

reflectorch/runs/__init__.py

@@ -0,0 +1,35 @@
+# -*- coding: utf-8 -*-
+#
+#
+# This source code is licensed under the GPL license found in the
+# LICENSE file in the root directory of this source tree.
+
+from reflectorch.runs.train import (
+    run_train,
+    run_train_on_cluster,
+    run_test_config,
+)
+
+from reflectorch.runs.utils import (
+    train_from_config,
+    get_trainer_from_config,
+    get_paths_from_config,
+    get_callbacks_from_config,
+    get_trainer_by_name,
+    get_callbacks_by_name,
+)
+
+from reflectorch.runs.config import load_config
+
+__all__ = [
+    'run_train',
+    'run_train_on_cluster',
+    'train_from_config',
+    'run_test_config',
+    'get_trainer_from_config',
+    'get_paths_from_config',
+    'get_callbacks_from_config',
+    'get_trainer_by_name',
+    'get_callbacks_by_name',
+    'load_config',
+]

reflectorch/runs/config.py

@@ -0,0 +1,31 @@
+# -*- coding: utf-8 -*-
+#
+#
+# This source code is licensed under the GPL license found in the
+# LICENSE file in the root directory of this source tree.
+
+import yaml
+
+from pathlib import Path
+from reflectorch.paths import CONFIG_DIR
+
+
+def load_config(config_name: str, config_dir: str = None) -> dict:
+    """Loads a configuration dictionary from a YAML configuration file located in the configuration directory
+
+    Args:
+        config_name (str): name of the YAML configuration file
+        config_dir (str): path of the configuration directory
+
+    Returns:
+        dict: the configuration dictionary
+    """
+    if not config_name.endswith('.yaml'):
+        config_name = f'{config_name}.yaml'
+    config_dir = Path(config_dir) if config_dir else CONFIG_DIR
+    path = config_dir / config_name
+    with open(path, 'r') as f:
+        config = yaml.safe_load(f)
+    config['config_path'] = str(path.absolute())
+
+    return config

reflectorch/runs/slurm_utils.py

@@ -0,0 +1,99 @@
+# -*- coding: utf-8 -*-
+#
+#
+# This source code is licensed under the GPL license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Tuple, Union
+from pathlib import Path
+import subprocess
+
+from reflectorch.paths import RUN_SCRIPTS_DIR
+
+
+def save_sbatch_and_run(
+        name: str,
+        args: str,
+        time: str,
+        partition: str = None,
+        reservation: bool = False,
+        chdir: str = '~/maxwell_output',
+        run_dir: Path = None,
+        confirm: bool = False,
+) -> Union[Tuple[str, str], None]:
+    run_dir = Path(run_dir) if run_dir else RUN_SCRIPTS_DIR
+    sbatch_path = run_dir / f'{name}.sh'
+
+    if sbatch_path.is_file():
+        import warnings
+        warnings.warn(f'Sbatch file {str(sbatch_path)} already exists!')
+        if confirm and not confirm_input('Continue?'):
+            return
+
+    file_content = _generate_sbatch_str(
+        name,
+        args,
+        time=time,
+        reservation=reservation,
+        partition=partition,
+        chdir=chdir,
+    )
+
+    if confirm:
+        print(f'Generated file content: \n{file_content}\n')
+        if not confirm_input(f'Save to {str(sbatch_path)} and run?'):
+            return
+
+    with open(str(sbatch_path), 'w') as f:
+        f.write(file_content)
+
+    res = submit_job(str(sbatch_path))
+    return res
+
+
+def _generate_sbatch_str(name: str,
+                         args: str,
+                         time: str,
+                         partition: str = None,
+                         reservation: bool = False,
+                         chdir: str = '~/maxwell_output',
+                         entry_point: str = 'python -m reflectorch.train',
+                         ) -> str:
+    chdir = str(Path(chdir).expanduser().absolute())
+    partition_keyword = 'reservation' if reservation else 'partition'
+
+    return f'''#!/bin/bash
+#SBATCH --chdir {chdir}
+#SBATCH --{partition_keyword}={partition}
+#SBATCH --constraint=P100
+#SBATCH --nodes=1
+#SBATCH --job-name {name}
+#SBATCH --time={time}
+#SBATCH --output {name}.out
+#SBATCH --error {name}.err
+
+{entry_point} {args}
+'''
+
+
+def confirm_input(message: str) -> bool:
+    yes = ('y', 'yes')
+    no = ('n', 'no')
+    res = ''
+    valid_results = list(yes) + list(no)
+    message = f'{message} Y/n: '
+
+    while res not in valid_results:
+        res = input(message).lower()
+    return res in yes
+
+
+def submit_job(sbatch_path: str) -> Tuple[str, str]:
+    process = subprocess.Popen(
+        ['sbatch', str(sbatch_path)],
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+    )
+
+    stdout, stderr = process.communicate()
+    return stdout.decode(), stderr.decode()

reflectorch/runs/train.py

@@ -0,0 +1,85 @@
+# -*- coding: utf-8 -*-
+#
+#
+# This source code is licensed under the GPL license found in the
+# LICENSE file in the root directory of this source tree.
+
+import click
+
+from reflectorch.runs.slurm_utils import save_sbatch_and_run
+from reflectorch.runs.utils import train_from_config
+from reflectorch.runs.config import load_config
+
+__all__ = [
+    'run_train',
+    'run_train_on_cluster',
+    'run_test_config',
+]
+
+
+@click.command()
+@click.argument('config_name', type=str)
+def run_train(config_name: str):
+    """Runs the training from the command line interface
+       Example: python -m reflectorch.train 'conf_name'
+
+    Args:
+        config_name (str): name of the YAML configuration file
+    """
+    config = load_config(config_name)
+    train_from_config(config)
+
+
+@click.command()
+@click.argument('config_name', type=str)
+@click.argument('batch_size', type=int, default=512)
+@click.argument('num_iterations', type=int, default=10)
+def run_test_config(config_name: str, batch_size: int, num_iterations: int):
+    """Run for the purpose of testing the configuration file.
+       Example: python -m reflectorch.test_config 'conf_name.yaml' 512 10
+
+    Args:
+        config_name (str): name of the YAML configuration file
+        batch_size (int): overwrites the batch size in the configuration file
+        num_iterations (int): overwrites the number of iterations in the configuration file
+    """
+    config = load_config(config_name)
+    config = _change_to_test_config(config, batch_size=batch_size, num_iterations=num_iterations)
+    train_from_config(config)
+
+
+@click.command()
+@click.argument('config_name')
+def run_train_on_cluster(config_name: str):
+    config = load_config(config_name)
+    name = config['general']['name']
+    slurm_conf = config['slurm']
+
+    res = save_sbatch_and_run(
+        name,
+        config_name,
+        time=slurm_conf['time'],
+        partition=slurm_conf['partition'],
+        reservation=slurm_conf.get('reservation', False),
+        chdir=slurm_conf.get('chdir', '~/maxwell_output'),
+        run_dir=slurm_conf.get('run_dir', None),
+        confirm=slurm_conf.get('confirm', True),
+    )
+    if not res:
+        print('Aborted.')
+        return
+    out, err = res
+
+    if err:
+        print('Error occurred: ', err)
+    else:
+        print('Success!', out)
+
+
+def _change_to_test_config(config, batch_size: int, num_iterations: int):
+    config = dict(config)
+    config['training']['logger']['use_neptune'] = False
+    config['training']['num_iterations'] = num_iterations
+    config['training']['batch_size'] = batch_size
+    config['training']['update_tqdm_freq'] = 1
+    return config