diff --git a/.gitignore b/.gitignore
index 15f1423..2c32756 100644
--- a/.gitignore
+++ b/.gitignore
@@ -130,6 +130,7 @@ dmypy.json
# Results/Logs
results/*
+results*/*
logs/*.log
# toml files
diff --git a/anemic/__init__.py b/anemic/__init__.py
new file mode 100644
index 0000000..cac0ba3
--- /dev/null
+++ b/anemic/__init__.py
@@ -0,0 +1,7 @@
+from anemic import (
+ datasets,
+ models,
+ modules,
+ trainers,
+ utils,
+)
diff --git a/anemic/datasets/__init__.py b/anemic/datasets/__init__.py
new file mode 100644
index 0000000..047b697
--- /dev/null
+++ b/anemic/datasets/__init__.py
@@ -0,0 +1 @@
+from anemic.datasets import base_dataset
diff --git a/src/datasets/base_dataset.py b/anemic/datasets/base_dataset.py
similarity index 95%
rename from src/datasets/base_dataset.py
rename to anemic/datasets/base_dataset.py
index 912828c..43eda14 100644
--- a/src/datasets/base_dataset.py
+++ b/anemic/datasets/base_dataset.py
@@ -5,9 +5,9 @@
import torch
from torch.utils.data import Dataset
-from src.utils.file_loaders import load_csv_as_df, load_json
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
+from anemic.utils.file_loaders import load_csv_as_df, load_json
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
diff --git a/anemic/models/__init__.py b/anemic/models/__init__.py
new file mode 100644
index 0000000..2457c48
--- /dev/null
+++ b/anemic/models/__init__.py
@@ -0,0 +1,8 @@
+from anemic.models import (
+ auto_models,
+ caml,
+ dcan,
+ fusion,
+ multirescnn,
+ transicd,
+)
diff --git a/anemic/models/auto_models.py b/anemic/models/auto_models.py
new file mode 100644
index 0000000..ec38066
--- /dev/null
+++ b/anemic/models/auto_models.py
@@ -0,0 +1 @@
+from anemic.utils.mapper import ConfigMapper
diff --git a/src/models/caml.py b/anemic/models/caml.py
similarity index 98%
rename from src/models/caml.py
rename to anemic/models/caml.py
index 8828292..73c79c6 100644
--- a/src/models/caml.py
+++ b/anemic/models/caml.py
@@ -12,9 +12,9 @@
from torch.autograd import Variable
from torch.nn.init import xavier_uniform
-from src.utils.model_utils import load_lookups, pad_desc_vecs
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.model_utils import load_lookups, pad_desc_vecs
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
diff --git a/anemic/models/dcan.py b/anemic/models/dcan.py
new file mode 100644
index 0000000..f46c6ba
--- /dev/null
+++ b/anemic/models/dcan.py
@@ -0,0 +1,507 @@
+import copy
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.init import xavier_uniform_
+from torch.nn.utils import weight_norm as weight_norm_
+
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.text_loggers import get_logger
+
+logger = get_logger(__name__)
+
+
+@ConfigMapper.map("models", "dcan")
+class DCAN(nn.Module):
+ """
+ This class is used to create the DCAN model.
+ References:
+ Paper: https://aclanthology.org/2020.clinicalnlp-1.8/
+ GitHub Repository: https://github.com/shaoxiongji/DCAN
+ For the parameters related to convolutional layers, please see this:
+ https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html.
+ Args:
+ num_classes (int): Number of classes (ICD codes).
+ conv_channel_sizes (list): List of lists of integers. Each list
+ represents the channel sizes of convolutional
+ layers in a `TemporalBlock`. So, for example,
+ if the list is [[100, 600, 600],
+ [600, 600, 600]].
+ the `TemporalConvNet` layer will have 2
+ `TemporalBlock`s, each temporal block have
+ 2 convolutional layers:
+ Conv(100, 600), Conv(600, 600) for the first
+ one, and Conv(600, 600), Conv(600, 600). If
+ the `add_emb_size_to_channel_sizes`, we don't
+ have to pass the input channel size. So, in
+ the above case, we can just pass
+ [[600, 600], [600, 600, 600]].
+ add_emb_size_to_channel_sizes (bool): If True, you need not specify
+ the input channel size. Please
+ see the description of
+ `conv_channel_sizes`.
+ kernel_sizes (list): List of list of integers (same format as
+ `conv_channel_sizes`). Each integer represents the
+ kernel size/filter size of the respective
+ convolutional layer in `TemporalBlock` layer.
+ strides (list): List of list of integers (same format as
+ `conv_channel_sizes`). Each integer represents the
+ stride of the respective convolutional layer in
+ `TemporalBlock` layer.
+ paddings (list): List of list of integers (same format as
+ `conv_channel_sizes`). Each integer represents the
+ padding of the respective convolutional layer in
+ `TemporalBlock` layer. in DCAN, this value is set to
+ "(kernel_size - 1) * dilation_size".
+ dilations (list): List of list of integers (same format as
+ `conv_channel_sizes`). Each integer represents the
+ dilation size of the respective convolutional layer
+ `TemporalBlock` layer.` In DCAN, this value is
+ "2^(temporal_block_level)".
+ dropouts (list): List of list of floats (same format as
+ `conv_channel_sizes`). Each float represents the
+ dropout probability of the respective convolutional
+ `TemporalBlock` layer.
+ weight_norm (bool): If True, apply weight normalization to the
+ convolutional layers.
+ activation (str): Activation function to use. Should be one of "relu",
+ "elu", "leaky_relu".
+ """
+
+ def __init__(self, config):
+ super(DCAN, self).__init__()
+ logger.info(f"Initialising {self.__class__.__name__}")
+ logger.debug(
+ f"Initialising {self.__class__.__name__} with " f"config: {config}"
+ )
+
+ self.config = config
+
+ self.word_embedding_layer = WordEmbeddingLayer(
+ **config.word_representation_layer.params.init_params.as_dict()
+ )
+ if config.word_representation_layer.params.freeze_layer:
+ self.freeze_layer(self.word_embedding_layer.embed)
+
+ num_levels = len(config.kernel_sizes)
+ num_inner_conv_levels = len(config.kernel_sizes[0])
+
+ conv_channel_sizes = copy.deepcopy(config.conv_channel_sizes)
+ if config.add_emb_size_to_channel_sizes:
+ conv_channel_sizes[0] = [
+ self.word_embedding_layer.embedding_size
+ ] + conv_channel_sizes[0]
+ dropouts = [
+ [config.dropout for _ in range(num_inner_conv_levels)]
+ for _ in range(num_levels)
+ ]
+
+ self.temporal_conv_net = TemporalConvNet(
+ conv_channel_sizes_=conv_channel_sizes,
+ kernel_sizes_=config.kernel_sizes,
+ strides_=config.strides,
+ paddings_=config.paddings,
+ dilations_=config.dilations,
+ dropouts_=dropouts,
+ weight_norm=config.weight_norm,
+ activation=config.activation,
+ )
+
+ self.linear_layer = nn.Linear(
+ conv_channel_sizes[-1][-1], config.projection_size
+ )
+ self.activation = ConfigMapper.get_object(
+ "activations", config.activation
+ )()
+
+ self.output_layer = OutputLayer(
+ config.projection_size, config.num_classes
+ )
+
+ xavier_uniform_(self.linear_layer.weight)
+
+ def forward(self, data):
+ x = self.word_embedding_layer(data)
+ hid_seq = self.temporal_conv_net(x.transpose(1, 2)).transpose(1, 2)
+ hid_seq = self.activation(self.linear_layer(hid_seq))
+ logits = self.output_layer(hid_seq)
+ return logits
+
+ def freeze_layer(self, layer):
+ for param in layer.parameters():
+ param.requires_grad = False
+
+ def get_input_attention(self):
+ # Use the attention score computed in the forward pass
+ return self.output_layer.label_wise_attn.alpha.cpu().detach().numpy()
+
+
+class OutputLayer(nn.Module):
+ def __init__(self, input_size, num_classes):
+ super(OutputLayer, self).__init__()
+ self.label_wise_attn = LabelWiseAttn(input_size, num_classes)
+
+ self.final = nn.Linear(input_size, num_classes)
+ xavier_uniform_(self.final.weight)
+
+ def forward(self, x):
+ m = self.label_wise_attn(x)
+ logits = self.final.weight.mul(m).sum(dim=2).add(self.final.bias)
+ return logits
+
+
+class WordEmbeddingLayer(nn.Module):
+ """
+ A Word Embedding Layer. This layer loads a pre-trained word embedding matrix
+ , and copies its weights to an nn.Embedding layer.
+
+ Args:
+ embed_dir (str): A directory containing the pre-trained word embedding
+ matrix, among other things. Please see
+ https://github.com/dalgu90/icd-coding-benchmark/blob/main/anemic/modules/embeddings.py#L17
+ for more details.
+ dropout (float): The dropout probability.
+ """
+
+ def __init__(self, embed_dir, dropout):
+ super(WordEmbeddingLayer, self).__init__()
+ logger.debug(
+ f"Initialising {self.__class__.__name__} with "
+ f"embed_dir = {embed_dir}, dropout = {dropout}"
+ )
+
+ # Note: This should be changed, since we won't always use Word2Vec.
+ embedding_cls = ConfigMapper.get_object("embeddings", "word2vec")
+
+ W = torch.Tensor(embedding_cls.load_emb_matrix(embed_dir))
+ self.embed = nn.Embedding(W.size()[0], W.size()[1], padding_idx=0)
+ self.embed.weight.data = W.clone()
+
+ self.embedding_size = self.embed.embedding_dim
+
+ self.dropout = nn.Dropout(dropout)
+
+ def forward(self, x):
+ embedding = self.embed(x)
+ x = self.dropout(embedding)
+ return x
+
+
+class Chomp1d(nn.Module):
+ def __init__(self, chomp_size):
+ super(Chomp1d, self).__init__()
+ self.chomp_size = chomp_size
+
+ def forward(self, x):
+ return x[:, :, : -self.chomp_size].contiguous()
+
+
+class ConvTemporalSubBlock(nn.Module):
+ """
+ A simple temporal convolutional block. Adapted from
+ https://github.com/shaoxiongji/DCAN/blob/master/models.py#L84-L88. This
+ layer has a dilated convolutional layer, a `chomp1d` layer, followed by
+ activation and dropout. For the parameters related to convolutional layers,
+ please see this:
+ https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html.
+
+ Args:
+ in_channels (int): The number of input channels in the convolutional
+ layer.
+ out_channels (int): The number of output channels in the convolutional
+ layer.
+ kernel_size (int): The size of the kernel in the convolutional layer.
+ stride (int): The stride of the convolutional layer.
+ padding (int): The padding of the convolutional layer.
+ dilation (int): The dilation size of the convolutional layer.
+ dropout (float): The dropout probability.
+ weight_norm (bool): Whether to apply weight normalization to the
+ convolutional layer.
+ activation (str): The activation function to use. DCAN uses "relu".
+ For all available activations, see
+ https://github.com/dalgu90/icd-coding-benchmark/blob/main/anemic/modules/activations.py.
+ """
+
+ def __init__(
+ self,
+ in_channels,
+ out_channels,
+ kernel_size,
+ stride,
+ padding,
+ dilation,
+ dropout=0.2,
+ weight_norm=True,
+ activation="relu",
+ ):
+ super(ConvTemporalSubBlock, self).__init__()
+ logger.debug(
+ f"Initialising {self.__class__.__name__} with "
+ f"in_channels = {in_channels}, out_channels = "
+ f"{out_channels}, kernel_size = {kernel_size}, "
+ f"stride = {stride}, padding = {padding}, "
+ f"dilation = {dilation}, dropout = {dropout}, "
+ f"weight_norm = {weight_norm}, activation = {activation}"
+ )
+
+ self.conv_layer = nn.Conv1d(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ kernel_size=kernel_size,
+ stride=stride,
+ padding=padding,
+ dilation=dilation,
+ )
+ if weight_norm:
+ self.conv_layer = weight_norm_(self.conv_layer)
+ self.chomp1d = Chomp1d(padding)
+ self.activation = ConfigMapper.get_object("activations", activation)()
+ self.dropout = nn.Dropout(dropout)
+
+ self.__init_weights__()
+
+ def __init_weights__(self):
+ xavier_uniform_(self.conv_layer.weight)
+
+ def forward(self, x):
+ x = self.conv_layer(x)
+ x = self.chomp1d(x)
+ x = self.activation(x)
+ x = self.dropout(x)
+ return x
+
+
+class TemporalBlock(nn.Module):
+ """
+ A Temporal Block containing stacks of `ConvTemporalSubBlocks`, followed
+ by activation.
+ References:
+ Paper: https://arxiv.org/abs/2009.14578
+ Repository: https://github.com/shaoxiongji/DCAN/blob/master/models.py#L81
+
+ Args:
+ conv_channel_sizes (list): List of integers, with channel sizes of
+ convolutional layers. For example, if the
+ list is [100, 200, 300], there will be two
+ convolutional layers: Conv1d(100, 200) and
+ Conv1d(200, 300).
+ kernel_sizes (list): List of integers, with kernel sizes of every
+ `ConvTemporalSubBlock`.
+ strides (list): List of integers, with strides of convolutional layers.
+ paddings (list): List of integers, with paddings of every
+ `ConvTemporalSubBlock`.
+ dilations (list): List of integers, with dilation sizes of every
+ `ConvTemporalSubBlock`.
+ dropouts (list): List of floats, with dropout probabilities of every
+ `ConvTemporalSubBlock`.
+ weight_norm (bool): Whether to apply weight normalization to every
+ convolutional layer. DCAN uses weight norm.
+ activation (str): The activation function to use. DCAN uses "relu".
+ """
+
+ def __init__(
+ self,
+ conv_channel_sizes,
+ kernel_sizes,
+ strides,
+ paddings,
+ dilations,
+ dropouts,
+ weight_norm=True,
+ activation="relu",
+ ):
+ super(TemporalBlock, self).__init__()
+ conv_channel_size_pairs = list(
+ zip(conv_channel_sizes[:-1], conv_channel_sizes[1:])
+ )
+
+ self.conv_temporal_sub_blocks = nn.ModuleList(
+ [
+ ConvTemporalSubBlock(
+ in_channels=conv_channel_size_pair[0],
+ out_channels=conv_channel_size_pair[1],
+ kernel_size=kernel_size,
+ stride=stride,
+ padding=padding,
+ dilation=dilation,
+ dropout=dropout,
+ weight_norm=weight_norm,
+ activation=activation,
+ )
+ for (
+ conv_channel_size_pair,
+ kernel_size,
+ stride,
+ padding,
+ dilation,
+ dropout,
+ ) in zip(
+ conv_channel_size_pairs,
+ kernel_sizes,
+ strides,
+ paddings,
+ dilations,
+ dropouts,
+ )
+ ]
+ )
+
+ self.downsample = (
+ nn.Conv1d(conv_channel_sizes[0], conv_channel_sizes[-1], 1)
+ if conv_channel_sizes[0] != conv_channel_sizes[-1]
+ else None
+ )
+ self.output_activation = ConfigMapper.get_object(
+ "activations", activation
+ )()
+
+ self.init_weights()
+
+ def init_weights(self):
+ if self.downsample is not None:
+ xavier_uniform_(self.downsample.weight)
+
+ def forward(self, x):
+ conv_layer_output = x
+ for conv_temporal_sub_block in self.conv_temporal_sub_blocks:
+ conv_layer_output = conv_temporal_sub_block(conv_layer_output)
+ res = x if self.downsample is None else self.downsample(x)
+ return self.output_activation(conv_layer_output + res)
+
+
+class TemporalConvNet(nn.Module):
+ """
+ Stack of `TemporalBlock`s. Used in the DCAN model.
+ References:
+ Paper: https://arxiv.org/abs/2009.14578
+ Repository: https://github.com/shaoxiongji/DCAN/blob/master/models.py#L114
+
+ Args:
+ conv_channel_sizes_ (list): List of lists of integers. Each list
+ represents the channel sizes of convolutional
+ layers in a `TemporalBlock`. So, for
+ example, if the list is [[100, 600, 600],
+ [600, 600, 600]].
+ the `TemporalConvNet` layer will have 2
+ `TemporalBlock`s, each temporal block have
+ 2 convolutional layers:
+ Conv(100, 600), Conv(600, 600) for the first
+ one, and Conv(600, 600), Conv(600, 600). If
+ the `add_emb_size_to_channel_sizes`, we
+ don't have to pass the input channel size.
+ So, in the above case, we can just pass
+ [[600, 600], [600, 600, 600]].
+ kernel_sizes_ (list): List of list of integers (same format as
+ `conv_channel_sizes`). Each integer represents the
+ kernel size/filter size of the respective
+ convolutional layer in `TemporalBlock` layer.
+ strides_ (list): List of list of integers (same format as
+ `conv_channel_sizes`). Each integer represents the
+ stride of the respective convolutional layer in
+ `TemporalBlock` layer.
+ paddings_ (list): List of list of integers (same format as
+ `conv_channel_sizes`). Each integer represents the
+ padding of the respective convolutional layer in
+ `TemporalBlock` layer. in DCAN, this value is set to
+ "(kernel_size - 1) * dilation_size".
+ dilations_ (list): List of list of integers (same format as
+ `conv_channel_sizes`). Each integer represents the
+ dilation size of the respective convolutional layer
+ `TemporalBlock` layer.` In DCAN, this value is
+ "2^(temporal_block_level)".
+ dropouts_ (list): List of list of floats (same format as
+ `conv_channel_sizes`). Each float represents the
+ dropout probability of the respective convolutional
+ `TemporalBlock` layer.
+ weight_norm (bool): If True, apply weight normalization to the
+ convolutional layers.
+ activation (str): Activation function to use. DCAN uses "relu".
+ """
+
+ def __init__(
+ self,
+ conv_channel_sizes_,
+ kernel_sizes_,
+ strides_,
+ paddings_,
+ dilations_,
+ dropouts_,
+ weight_norm=True,
+ activation="relu",
+ ):
+ super(TemporalConvNet, self).__init__()
+ logger.debug(
+ f"Initialising {self.__class__.__name__} with "
+ f"conv_channel_sizes_ = {conv_channel_sizes_}, "
+ f"kernel_sizes_ = {kernel_sizes_}, "
+ f"strides_ = {strides_}, paddings_ = {paddings_}, "
+ f"dilations_ = {dilations_}, dropouts_ = {dropouts_}, "
+ f"weight_norm = {weight_norm}, activation = {activation}"
+ )
+
+ self.temporal_blocks = nn.ModuleList(
+ [
+ TemporalBlock(
+ conv_channel_sizes=conv_channel_sizes,
+ kernel_sizes=kernel_sizes,
+ strides=strides,
+ paddings=paddings,
+ dilations=dilations,
+ dropouts=dropouts,
+ weight_norm=weight_norm,
+ activation=activation,
+ )
+ for (
+ conv_channel_sizes,
+ kernel_sizes,
+ strides,
+ paddings,
+ dilations,
+ dropouts,
+ ) in zip(
+ conv_channel_sizes_,
+ kernel_sizes_,
+ strides_,
+ paddings_,
+ dilations_,
+ dropouts_,
+ )
+ ]
+ )
+
+ def forward(self, x):
+ for temporal_block in self.temporal_blocks:
+ x = temporal_block(x)
+ return x
+
+
+class LabelWiseAttn(nn.Module):
+ """
+ A Label-wise Attention layer (as implemented in CAML, DCAN, etc.).
+ References:
+ Papers: https://arxiv.org/abs/1802.05695 (Section 2.2)
+ Repository: https://github.com/jamesmullenbach/caml-mimic/blob/master/learn/models.py#L184
+
+ Args:
+ input_size (int): The size of the input, i.e., the number of channels
+ if the output is from a convolutional layer/embedding
+ size if the output is from a fully connected layer.
+ num_classes (int): The number of classes.
+ """
+
+ def __init__(self, input_size, num_classes):
+ super(LabelWiseAttn, self).__init__()
+ logger.debug(
+ f"Initialising {self.__class__.__name__} with "
+ f"input size = {input_size}, num_classes = {num_classes}"
+ )
+
+ self.U = nn.Linear(input_size, num_classes)
+ xavier_uniform_(self.U.weight)
+
+ def forward(self, x):
+ att = self.U.weight.matmul(x.transpose(1, 2)) # [bs, Y, seq_len]
+ self.alpha = F.softmax(att, dim=2)
+ m = self.alpha.matmul(x) # [bs, Y, dim]
+ return m
diff --git a/src/models/fusion.py b/anemic/models/fusion.py
similarity index 98%
rename from src/models/fusion.py
rename to anemic/models/fusion.py
index e2a5d82..4c154d4 100755
--- a/src/models/fusion.py
+++ b/anemic/models/fusion.py
@@ -11,9 +11,9 @@
import torch.nn.functional as F
from torch.nn.init import xavier_uniform_ as xavier_uniform
-from src.utils.model_utils import load_lookups
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.model_utils import load_lookups
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
@@ -417,7 +417,7 @@ class Fusion(nn.Module):
def __init__(self, config):
super(Fusion, self).__init__()
- logger.info(f"Initialising %s", self.__class__.__name__)
+ logger.info("Initialising %s", self.__class__.__name__)
logger.debug(
"Initialising %s with config: %s", self.__class__.__name__, config
)
diff --git a/src/models/multirescnn.py b/anemic/models/multirescnn.py
similarity index 98%
rename from src/models/multirescnn.py
rename to anemic/models/multirescnn.py
index 95b1241..1b6b147 100644
--- a/src/models/multirescnn.py
+++ b/anemic/models/multirescnn.py
@@ -1,5 +1,6 @@
"""
-ICD Coding from Clinical Text Using Multi-Filter Residual Convolutional Neural Network, 2020
+ICD Coding from Clinical Text Using Multi-Filter Residual Convolutional Neural
+Network, 2020
https://github.com/foxlf823/Multi-Filter-Residual-Convolutional-Neural-Network
"""
@@ -10,9 +11,9 @@
import torch.nn.functional as F
from torch.nn.init import xavier_uniform_ as xavier_uniform
-from src.utils.mapper import ConfigMapper
-from src.utils.model_utils import load_lookups
-from src.utils.text_loggers import get_logger
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.model_utils import load_lookups
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
diff --git a/src/models/transicd/model.py b/anemic/models/transicd.py
similarity index 96%
rename from src/models/transicd/model.py
rename to anemic/models/transicd.py
index a337a07..5c9b62f 100644
--- a/src/models/transicd/model.py
+++ b/anemic/models/transicd.py
@@ -4,8 +4,8 @@
import torch.nn as nn
from torch.autograd import Variable
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
@@ -114,7 +114,7 @@ def forward(self, inputs):
# outputs = torch.zeros(batch_size, self.num_classes).to(inputs.device)
# for code, ff_layer in enumerate(self.ff_layers):
- # outputs[:, code : code + 1] = ff_layer(weighted_outputs[:, code, :])
+ # outputs[:, code : code + 1] = ff_layer(weighted_outputs[:, code])
# Trick: Use one linear layer as per-code linear layers
outputs = (weighted_outputs * self.ff_layer.weight).sum(axis=2)
@@ -135,7 +135,7 @@ class WordEmbeddingLayer(nn.Module):
Args:
embed_dir (str): A directory containing the pre-trained word embedding
matrix, among other things. Please see
- https://github.com/dalgu90/icd-coding-benchmark/blob/main/src/modules/embeddings.py#L17
+ https://github.com/dalgu90/icd-coding-benchmark/blob/main/anemic/modules/embeddings.py#L17
for more details.
"""
@@ -223,9 +223,7 @@ class LabelAttentionLayer(nn.Module):
Defaults to 2.
"""
- def __init__(
- self, embed_size=128, num_classes=50, attn_expansion=2
- ):
+ def __init__(self, embed_size=128, num_classes=50, attn_expansion=2):
super(LabelAttentionLayer, self).__init__()
logger.debug(
f"Initialising {self.__class__.__name__} with "
diff --git a/anemic/modules/__init__.py b/anemic/modules/__init__.py
new file mode 100755
index 0000000..e16fe6a
--- /dev/null
+++ b/anemic/modules/__init__.py
@@ -0,0 +1,12 @@
+from anemic.modules import (
+ activations,
+ dataset_splitters,
+ embeddings,
+ losses,
+ metrics,
+ optimizers,
+ preprocessing_pipelines,
+ preprocessors,
+ schedulers,
+ tokenizers,
+)
diff --git a/src/modules/activations.py b/anemic/modules/activations.py
similarity index 80%
rename from src/modules/activations.py
rename to anemic/modules/activations.py
index 8433949..fb03ee4 100755
--- a/src/modules/activations.py
+++ b/anemic/modules/activations.py
@@ -1,6 +1,6 @@
import torch.nn as nn
-from src.utils.mapper import ConfigMapper
+from anemic.utils.mapper import ConfigMapper
ConfigMapper.map("activations", "relu")(nn.ReLU)
ConfigMapper.map("activations", "logsoftmax")(nn.LogSoftmax)
diff --git a/src/modules/dataset_splitters.py b/anemic/modules/dataset_splitters.py
similarity index 87%
rename from src/modules/dataset_splitters.py
rename to anemic/modules/dataset_splitters.py
index 54a9a6c..2ec195f 100644
--- a/src/modules/dataset_splitters.py
+++ b/anemic/modules/dataset_splitters.py
@@ -1,9 +1,9 @@
import logging
import os
-from src.utils.file_loaders import load_json
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
+from anemic.utils.file_loaders import load_json
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
diff --git a/src/modules/embeddings.py b/anemic/modules/embeddings.py
similarity index 93%
rename from src/modules/embeddings.py
rename to anemic/modules/embeddings.py
index 3b31c1e..ce83881 100755
--- a/src/modules/embeddings.py
+++ b/anemic/modules/embeddings.py
@@ -6,9 +6,9 @@
import gensim
import numpy as np
-from src.utils.file_loaders import load_json, save_json
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
+from anemic.utils.file_loaders import load_json, save_json
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
diff --git a/src/modules/losses.py b/anemic/modules/losses.py
similarity index 96%
rename from src/modules/losses.py
rename to anemic/modules/losses.py
index 8afceaa..0323c72 100755
--- a/src/modules/losses.py
+++ b/anemic/modules/losses.py
@@ -7,8 +7,8 @@
from torch.autograd import Variable
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
-from src.utils.file_loaders import load_json
-from src.utils.mapper import ConfigMapper
+from anemic.utils.file_loaders import load_json
+from anemic.utils.mapper import ConfigMapper
ConfigMapper.map("losses", "mse")(MSELoss)
ConfigMapper.map("losses", "CrossEntropyLoss")(CrossEntropyLoss)
diff --git a/src/modules/metrics.py b/anemic/modules/metrics.py
similarity index 97%
rename from src/modules/metrics.py
rename to anemic/modules/metrics.py
index cdf1a41..f20908a 100755
--- a/src/modules/metrics.py
+++ b/anemic/modules/metrics.py
@@ -10,9 +10,9 @@
roc_auc_score,
)
-from src.utils.configuration import Config
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
+from anemic.utils.configuration import Config
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
@@ -23,6 +23,7 @@ def to_np_array(array):
array = np.array(array)
return array
+
def _auc_job(x):
return roc_auc_score(x[0], x[1])
diff --git a/src/modules/optimizers.py b/anemic/modules/optimizers.py
similarity index 80%
rename from src/modules/optimizers.py
rename to anemic/modules/optimizers.py
index 55edccf..be8106f 100755
--- a/src/modules/optimizers.py
+++ b/anemic/modules/optimizers.py
@@ -1,7 +1,7 @@
"""Method containing activation functions"""
from torch.optim import SGD, Adam, AdamW
-from src.utils.mapper import ConfigMapper
+from anemic.utils.mapper import ConfigMapper
ConfigMapper.map("optimizers", "adam")(Adam)
ConfigMapper.map("optimizers", "adam_w")(AdamW)
diff --git a/src/modules/preprocessing_pipelines.py b/anemic/modules/preprocessing_pipelines.py
similarity index 96%
rename from src/modules/preprocessing_pipelines.py
rename to anemic/modules/preprocessing_pipelines.py
index d7f2bd4..9bbcd00 100644
--- a/src/modules/preprocessing_pipelines.py
+++ b/anemic/modules/preprocessing_pipelines.py
@@ -5,14 +5,11 @@
import pandas as pd
from tqdm.auto import tqdm
-from src.modules.dataset_splitters import *
-from src.modules.embeddings import *
-from src.modules.preprocessors import ClinicalNotePreprocessor, CodeProcessor
-from src.modules.tokenizers import *
-from src.utils.code_based_filtering import TopKCodes
-from src.utils.file_loaders import load_csv_as_df, save_df, save_json
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
+from anemic.modules.preprocessors import ClinicalNotePreprocessor, CodeProcessor
+from anemic.utils.code_based_filtering import TopKCodes
+from anemic.utils.file_loaders import load_csv_as_df, save_df, save_json
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
tqdm.pandas()
diff --git a/src/modules/preprocessors.py b/anemic/modules/preprocessors.py
similarity index 98%
rename from src/modules/preprocessors.py
rename to anemic/modules/preprocessors.py
index 1dc8782..ae9dba0 100755
--- a/src/modules/preprocessors.py
+++ b/anemic/modules/preprocessors.py
@@ -13,8 +13,8 @@
)
from nltk.tokenize import RegexpTokenizer
-from src.utils.file_loaders import load_json
-from src.utils.text_loggers import get_logger
+from anemic.utils.file_loaders import load_json
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
diff --git a/src/modules/schedulers.py b/anemic/modules/schedulers.py
similarity index 90%
rename from src/modules/schedulers.py
rename to anemic/modules/schedulers.py
index b8a07f8..ac6e6d9 100755
--- a/src/modules/schedulers.py
+++ b/anemic/modules/schedulers.py
@@ -8,7 +8,7 @@
)
from transformers import get_linear_schedule_with_warmup
-from src.utils.mapper import ConfigMapper
+from anemic.utils.mapper import ConfigMapper
ConfigMapper.map("schedulers", "step")(StepLR)
ConfigMapper.map("schedulers", "cosineanneal")(CosineAnnealingLR)
diff --git a/src/modules/tokenizers.py b/anemic/modules/tokenizers.py
similarity index 89%
rename from src/modules/tokenizers.py
rename to anemic/modules/tokenizers.py
index 673093a..50d8146 100755
--- a/src/modules/tokenizers.py
+++ b/anemic/modules/tokenizers.py
@@ -2,8 +2,8 @@
import logging
import sys
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
diff --git a/anemic/trainers/__init__.py b/anemic/trainers/__init__.py
new file mode 100644
index 0000000..2364313
--- /dev/null
+++ b/anemic/trainers/__init__.py
@@ -0,0 +1 @@
+from anemic.trainers import base_trainer
diff --git a/src/trainers/base_trainer.py b/anemic/trainers/base_trainer.py
similarity index 95%
rename from src/trainers/base_trainer.py
rename to anemic/trainers/base_trainer.py
index 748121b..ca81606 100755
--- a/src/trainers/base_trainer.py
+++ b/anemic/trainers/base_trainer.py
@@ -7,19 +7,11 @@
from torch.utils.data import DataLoader
from tqdm import tqdm
-from src.modules.embeddings import *
-from src.modules.losses import *
-from src.modules.metrics import *
-from src.modules.optimizers import *
-from src.modules.schedulers import *
-from src.modules.tokenizers import *
-from src.utils.checkpoint_savers import *
-from src.utils.configuration import Config
-from src.utils.file_loaders import save_json
-from src.utils.graph_writers import *
-from src.utils.mapper import ConfigMapper
-from src.utils.misc import *
-from src.utils.text_loggers import get_logger
+from anemic.modules.metrics import load_metric
+from anemic.utils.configuration import Config
+from anemic.utils.file_loaders import save_json
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
@@ -85,17 +77,20 @@ def train(self, model, train_dataset, val_dataset=None):
)
scheduler = None
if self.config.lr_scheduler is not None:
- if 'warmup' in self.config.lr_scheduler.name:
- warm_up_steps = self.config.lr_scheduler.params.warm_up_proportion*(len(train_dataset) // batch_size)
- num_training_steps = (len(train_dataset) // batch_size)
+ if "warmup" in self.config.lr_scheduler.name:
+ warm_up_steps = (
+ self.config.lr_scheduler.params.warm_up_proportion
+ * (len(train_dataset) // batch_size)
+ )
+ num_training_steps = len(train_dataset) // batch_size
scheduler = ConfigMapper.get_object(
"schedulers", self.config.lr_scheduler.name
- )(optimizer,warm_up_steps,num_training_steps)
+ )(optimizer, warm_up_steps, num_training_steps)
logger.debug(
f"Created scheduler {scheduler.__class__.__name__} with "
f"config: {self.config.lr_scheduler.params}"
)
- else :
+ else:
scheduler = ConfigMapper.get_object(
"schedulers", self.config.lr_scheduler.name
)(optimizer, **self.config.lr_scheduler.params.as_dict())
diff --git a/anemic/utils/__init__.py b/anemic/utils/__init__.py
new file mode 100755
index 0000000..e796567
--- /dev/null
+++ b/anemic/utils/__init__.py
@@ -0,0 +1,12 @@
+from anemic.utils import (
+ checkpoint_savers,
+ code_based_filtering,
+ configuration,
+ file_loaders,
+ graph_writers,
+ import_related_ops,
+ mapper,
+ misc,
+ model_utils,
+ text_loggers,
+)
diff --git a/src/utils/checkpoint_savers.py b/anemic/utils/checkpoint_savers.py
similarity index 96%
rename from src/utils/checkpoint_savers.py
rename to anemic/utils/checkpoint_savers.py
index deaabfa..f17d96d 100644
--- a/src/utils/checkpoint_savers.py
+++ b/anemic/utils/checkpoint_savers.py
@@ -6,10 +6,10 @@
import torch
-from src.modules.metrics import load_metric
-from src.utils.file_loaders import load_json, save_json
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
+from anemic.modules.metrics import load_metric
+from anemic.utils.file_loaders import load_json, save_json
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
@@ -186,7 +186,7 @@ def save_ckpt(
def load_ckpt(self, model, ckpt_fname, optimizer=None):
ckpt_fpath = os.path.join(self.config.checkpoint_dir, ckpt_fname)
logger.debug(f"Loading ckpt from {ckpt_fpath}")
- checkpoint = torch.load(ckpt_fpath, map_location='cpu')
+ checkpoint = torch.load(ckpt_fpath, map_location="cpu")
model.load_state_dict(checkpoint["model"])
if optimizer:
optimizer.load_state_dict(checkpoint["optimizer"])
diff --git a/src/utils/code_based_filtering.py b/anemic/utils/code_based_filtering.py
similarity index 95%
rename from src/utils/code_based_filtering.py
rename to anemic/utils/code_based_filtering.py
index 1d845c4..eb2bff8 100644
--- a/src/utils/code_based_filtering.py
+++ b/anemic/utils/code_based_filtering.py
@@ -5,8 +5,8 @@
import pandas as pd
from sklearn.preprocessing import MultiLabelBinarizer
-from src.utils.file_loaders import save_json
-from src.utils.text_loggers import get_logger
+from anemic.utils.file_loaders import save_json
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
diff --git a/src/utils/configuration.py b/anemic/utils/configuration.py
similarity index 95%
rename from src/utils/configuration.py
rename to anemic/utils/configuration.py
index 11e6ccc..eccee07 100755
--- a/src/utils/configuration.py
+++ b/anemic/utils/configuration.py
@@ -2,7 +2,7 @@
import yaml
-from src.utils.mapper import ConfigMapper
+from anemic.utils.mapper import ConfigMapper
def load_yaml(path):
@@ -122,8 +122,10 @@ def __getattr__(self, attr):
if isinstance(self._config[attr], dict):
return Config(dic=self._config[attr])
elif isinstance(self._config[attr], list):
- return [Config(dic=e) if isinstance(e, dict) else e
- for e in self._config[attr]]
+ return [
+ Config(dic=e) if isinstance(e, dict) else e
+ for e in self._config[attr]
+ ]
else:
return self._config[attr]
else:
diff --git a/src/utils/file_loaders.py b/anemic/utils/file_loaders.py
similarity index 95%
rename from src/utils/file_loaders.py
rename to anemic/utils/file_loaders.py
index 7f48210..38a1f6b 100644
--- a/src/utils/file_loaders.py
+++ b/anemic/utils/file_loaders.py
@@ -4,7 +4,7 @@
import pandas as pd
-from src.utils.text_loggers import get_logger
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
diff --git a/src/utils/graph_writers.py b/anemic/utils/graph_writers.py
similarity index 89%
rename from src/utils/graph_writers.py
rename to anemic/utils/graph_writers.py
index 840ad44..06d08e2 100755
--- a/src/utils/graph_writers.py
+++ b/anemic/utils/graph_writers.py
@@ -4,8 +4,8 @@
import torch
from torch.utils.tensorboard import SummaryWriter
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.text_loggers import get_logger
logger = get_logger(__name__)
diff --git a/src/utils/import_related_ops.py b/anemic/utils/import_related_ops.py
similarity index 100%
rename from src/utils/import_related_ops.py
rename to anemic/utils/import_related_ops.py
diff --git a/src/utils/mapper.py b/anemic/utils/mapper.py
similarity index 100%
rename from src/utils/mapper.py
rename to anemic/utils/mapper.py
diff --git a/src/utils/misc.py b/anemic/utils/misc.py
similarity index 96%
rename from src/utils/misc.py
rename to anemic/utils/misc.py
index 03f76d4..ea182c6 100755
--- a/src/utils/misc.py
+++ b/anemic/utils/misc.py
@@ -210,4 +210,3 @@ def html_word_importance(words, importances):
tags.append("")
return "".join(tags)
-
diff --git a/src/utils/model_utils.py b/anemic/utils/model_utils.py
similarity index 97%
rename from src/utils/model_utils.py
rename to anemic/utils/model_utils.py
index ef672b1..74f86ab 100644
--- a/src/utils/model_utils.py
+++ b/anemic/utils/model_utils.py
@@ -9,9 +9,9 @@
import numpy as np
-from src.modules.preprocessors import CodeProcessor
-from src.utils.file_loaders import load_csv_as_df, load_json
-from src.utils.mapper import ConfigMapper
+from anemic.modules.preprocessors import CodeProcessor
+from anemic.utils.file_loaders import load_csv_as_df, load_json
+from anemic.utils.mapper import ConfigMapper
def load_lookups(
diff --git a/src/utils/text_loggers.py b/anemic/utils/text_loggers.py
similarity index 100%
rename from src/utils/text_loggers.py
rename to anemic/utils/text_loggers.py
diff --git a/app.py b/app.py
index 5687010..db1d2a5 100755
--- a/app.py
+++ b/app.py
@@ -6,21 +6,17 @@
import copy
import csv
-from captum.attr import LayerIntegratedGradients
import numpy as np
import pandas as pd
import seaborn as sns
import streamlit as st
import torch
+from captum.attr import LayerIntegratedGradients
-from src.datasets import *
-from src.models import *
-from src.modules.embeddings import *
-from src.modules.preprocessors import ClinicalNotePreprocessor
-from src.utils.checkpoint_savers import *
-from src.utils.configuration import Config
-from src.utils.mapper import ConfigMapper
-from src.utils.misc import html_word_importance
+from anemic.modules.preprocessors import ClinicalNotePreprocessor
+from anemic.utils.configuration import Config
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.misc import html_word_importance
hash_funcs = {
Config: lambda x: hash(str(x)),
@@ -91,7 +87,7 @@ def load_modules(config):
# 3. Load model
model_dict = {}
lig_dict = {}
- if hasattr(config, 'models'):
+ if hasattr(config, "models"):
model_configs = config.models
else:
model_configs = [config]
@@ -121,8 +117,10 @@ def load_modules(config):
)
embed_layer = getattr(model, embed_layer_name)
except:
- raise ValueError(f"Config for {model_config.model.name} does not"
- "specify name of the embedding layer.")
+ raise ValueError(
+ f"Config for {model_config.model.name} does not"
+ "specify name of the embedding layer."
+ )
lig = LayerIntegratedGradients(model, embed_layer)
model_dict[model_config.model.name] = model
@@ -138,6 +136,7 @@ def load_icd_desc(config):
icd_desc = {r[0]: r[1] for r in icd_desc}
return icd_desc
+
# Page setup
st.set_page_config(
page_title="ICD Coding Interactive Demo",
@@ -145,7 +144,8 @@ def load_icd_desc(config):
layout="wide",
)
-st.markdown("""
+st.markdown(
+ """
""", unsafe_allow_html=True)
+""",
+ unsafe_allow_html=True,
+)
# Title & status line
st.title("🩺 ICD Coding Interactive Demo")
@@ -223,15 +225,17 @@ def set_status(text):
"NO",
"Integrated Gradients",
]
- if any(hasattr(model, "get_input_attention") for model in
- model_dict.values()):
+ if any(
+ hasattr(model, "get_input_attention")
+ for model in model_dict.values()
+ ):
vis_score_options.append("Attention score")
vis_score = st.radio(
"Visualize attribution score",
vis_score_options,
help="""Interpretability visualization methods. Attention score is
- available only for attention-based models."""
+ available only for attention-based models.""",
)
vis_code_options = ["Choose ICD code"]
@@ -245,8 +249,10 @@ def set_status(text):
)
# Preprocessing option selection (truncation is not controlled)
- st.markdown(""" Preprocessing
""",
- unsafe_allow_html=True)
+ st.markdown(
+ """ Preprocessing
""",
+ unsafe_allow_html=True,
+ )
pp_config = config.clinical_note_preprocessing
pp_lower_case = st.checkbox(
"Lowercase",
@@ -273,8 +279,10 @@ def set_status(text):
with col2:
# Input text
css_str = "line-height:1; margin-top:1rem; margin-bottom:-2rem;"
- st.markdown(f"""Discharge summary note
""",
- unsafe_allow_html=True)
+ st.markdown(
+ f"""Discharge summary note
""",
+ unsafe_allow_html=True,
+ )
input_text = st.text_area(label="", height=200)
# input_text = st.text_area(label="Discharge summary note", height=200)
input_text = input_text.strip()
@@ -337,7 +345,7 @@ def set_status(text):
}
)
output_df.index += 1
- cmap = sns.light_palette('#AC304B', as_cmap=True)
+ cmap = sns.light_palette("#AC304B", as_cmap=True)
output_df = output_df.style.background_gradient(
cmap=cmap, subset=["Probability"], vmin=0.0, vmax=1.0
).format({"Probability": "{:.4f}"})
@@ -350,8 +358,9 @@ def set_status(text):
st.markdown("**[No attribution method selected]**")
elif target_label == -1:
st.markdown("**[No ICD code selected]**")
- elif (vis_score == "Attention score" and
- not hasattr(model, "get_input_attention")):
+ elif vis_score == "Attention score" and not hasattr(
+ model, "get_input_attention"
+ ):
st.markdown("**[Model does not support attention score]**")
else:
if vis_score == "Integrated Gradients":
@@ -359,10 +368,12 @@ def set_status(text):
attrs, approx_error = lig.attribute(
batch_input,
target=target_label,
- return_convergence_delta=True
+ return_convergence_delta=True,
)
attrs = attrs.sum(dim=2).squeeze(0)
- attrs = (attrs/torch.norm(attrs)).cpu().detach().numpy()
+ attrs = (
+ (attrs / torch.norm(attrs)).cpu().detach().numpy()
+ )
elif vis_score == "Attention score":
attrs = model.get_input_attention()
attrs = attrs[:, target_label].squeeze(0)
@@ -372,8 +383,10 @@ def set_status(text):
assert len(attrs) >= len(tokens)
html_string = html_word_importance(tokens, attrs)
- st.markdown(f"**{vis_score}** for **{vis_code}** "
- f"({icd_desc[vis_code]})")
+ st.markdown(
+ f"**{vis_score}** for **{vis_code}** "
+ f"({icd_desc[vis_code]})"
+ )
st.markdown(html_string, unsafe_allow_html=True)
st.markdown("")
elif token_idxs and len(token_idxs) < config.demo.min_input_len:
diff --git a/run.py b/run.py
index 09a318e..4747d16 100644
--- a/run.py
+++ b/run.py
@@ -6,13 +6,10 @@
import torch
from torchsummaryX import summary
-from src.datasets import *
-from src.models import *
-from src.trainers import *
-from src.utils.configuration import Config
-from src.utils.import_related_ops import pandas_related_ops
-from src.utils.mapper import ConfigMapper
-from src.utils.misc import seed
+from anemic.utils.configuration import Config
+from anemic.utils.import_related_ops import pandas_related_ops
+from anemic.utils.mapper import ConfigMapper
+from anemic.utils.misc import seed
pandas_related_ops()
diff --git a/run_preprocessing.py b/run_preprocessing.py
index d6cdf43..5cbef72 100644
--- a/run_preprocessing.py
+++ b/run_preprocessing.py
@@ -1,8 +1,7 @@
import argparse
-from src.modules.preprocessing_pipelines import *
-from src.utils.configuration import Config
-from src.utils.mapper import ConfigMapper
+from anemic.utils.configuration import Config
+from anemic.utils.mapper import ConfigMapper
# Command line arguments
parser = argparse.ArgumentParser(description="Preprocessing datasets")
diff --git a/src/__init__.py b/src/__init__.py
deleted file mode 100644
index e69de29..0000000
diff --git a/src/datasets/__init__.py b/src/datasets/__init__.py
deleted file mode 100644
index 227088f..0000000
--- a/src/datasets/__init__.py
+++ /dev/null
@@ -1 +0,0 @@
-from src.datasets.base_dataset import BaseDataset
diff --git a/src/models/__init__.py b/src/models/__init__.py
deleted file mode 100644
index ec43bca..0000000
--- a/src/models/__init__.py
+++ /dev/null
@@ -1,6 +0,0 @@
-from src.models.caml import ConvAttnPool as CAML
-from src.models.caml import VanillaConv as CNN
-from src.models.multirescnn import MultiResCNN
-from src.models.dcan import DCAN
-from src.models.fusion import Fusion
-from src.models.transicd.model import TransICD
diff --git a/src/models/auto_models.py b/src/models/auto_models.py
deleted file mode 100644
index 6998953..0000000
--- a/src/models/auto_models.py
+++ /dev/null
@@ -1 +0,0 @@
-from src.utils.mapper import ConfigMapper
diff --git a/src/models/dcan.py b/src/models/dcan.py
deleted file mode 100644
index c0ca9b3..0000000
--- a/src/models/dcan.py
+++ /dev/null
@@ -1,154 +0,0 @@
-import copy
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from torch.nn.init import xavier_uniform_
-
-from src.modules.activations import *
-from src.modules.layers.label_wise_attn import LabelWiseAttn
-from src.modules.layers.temporal_conv_net import TemporalConvNet
-from src.modules.layers.word_embedding_layer import WordEmbeddingLayer
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
-
-logger = get_logger(__name__)
-
-
-@ConfigMapper.map("models", "dcan")
-class DCAN(nn.Module):
- """
- This class is used to create the DCAN model.
- References:
- Paper: https://aclanthology.org/2020.clinicalnlp-1.8/
- GitHub Repository: https://github.com/shaoxiongji/DCAN
- For the parameters related to convolutional layers, please see this:
- https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html.
- Args:
- num_classes (int): Number of classes (ICD codes).
- conv_channel_sizes (list): List of lists of integers. Each list
- represents the channel sizes of convolutional
- layers in a `TemporalBlock`. So, for example,
- if the list is [[100, 600, 600],
- [600, 600, 600]].
- the `TemporalConvNet` layer will have 2
- `TemporalBlock`s, each temporal block have
- 2 convolutional layers:
- Conv(100, 600), Conv(600, 600) for the first
- one, and Conv(600, 600), Conv(600, 600). If
- the `add_emb_size_to_channel_sizes`, we don't
- have to pass the input channel size. So, in
- the above case, we can just pass
- [[600, 600], [600, 600, 600]].
- add_emb_size_to_channel_sizes (bool): If True, you need not specify
- the input channel size. Please
- see the description of
- `conv_channel_sizes`.
- kernel_sizes (list): List of list of integers (same format as
- `conv_channel_sizes`). Each integer represents the
- kernel size/filter size of the respective
- convolutional layer in `TemporalBlock` layer.
- strides (list): List of list of integers (same format as
- `conv_channel_sizes`). Each integer represents the
- stride of the respective convolutional layer in
- `TemporalBlock` layer.
- paddings (list): List of list of integers (same format as
- `conv_channel_sizes`). Each integer represents the
- padding of the respective convolutional layer in
- `TemporalBlock` layer. in DCAN, this value is set to
- "(kernel_size - 1) * dilation_size".
- dilations (list): List of list of integers (same format as
- `conv_channel_sizes`). Each integer represents the
- dilation size of the respective convolutional layer
- `TemporalBlock` layer.` In DCAN, this value is
- "2^(temporal_block_level)".
- dropouts (list): List of list of floats (same format as
- `conv_channel_sizes`). Each float represents the
- dropout probability of the respective convolutional
- `TemporalBlock` layer.
- weight_norm (bool): If True, apply weight normalization to the
- convolutional layers.
- activation (str): Activation function to use. Should be one of "relu",
- "elu", "leaky_relu".
- """
-
- def __init__(self, config):
- super(DCAN, self).__init__()
- logger.info(f"Initialising {self.__class__.__name__}")
- logger.debug(
- f"Initialising {self.__class__.__name__} with " f"config: {config}"
- )
-
- self.config = config
-
- self.word_embedding_layer = WordEmbeddingLayer(
- **config.word_representation_layer.params.init_params.as_dict()
- )
- if config.word_representation_layer.params.freeze_layer:
- self.freeze_layer(self.word_embedding_layer.embed)
-
- num_levels = len(config.kernel_sizes)
- num_inner_conv_levels = len(config.kernel_sizes[0])
-
- conv_channel_sizes = copy.deepcopy(config.conv_channel_sizes)
- if config.add_emb_size_to_channel_sizes:
- conv_channel_sizes[0] = [
- self.word_embedding_layer.embedding_size
- ] + conv_channel_sizes[0]
- dropouts = [
- [config.dropout for _ in range(num_inner_conv_levels)]
- for _ in range(num_levels)
- ]
-
- self.temporal_conv_net = TemporalConvNet(
- conv_channel_sizes_=conv_channel_sizes,
- kernel_sizes_=config.kernel_sizes,
- strides_=config.strides,
- paddings_=config.paddings,
- dilations_=config.dilations,
- dropouts_=dropouts,
- weight_norm=config.weight_norm,
- activation=config.activation,
- )
-
- self.linear_layer = nn.Linear(
- conv_channel_sizes[-1][-1], config.projection_size
- )
- self.activation = ConfigMapper.get_object(
- "activations", config.activation
- )()
-
- self.output_layer = OutputLayer(
- config.projection_size, config.num_classes
- )
-
- xavier_uniform_(self.linear_layer.weight)
-
- def forward(self, data):
- x = self.word_embedding_layer(data)
- hid_seq = self.temporal_conv_net(x.transpose(1, 2)).transpose(1, 2)
- hid_seq = self.activation(self.linear_layer(hid_seq))
- logits = self.output_layer(hid_seq)
- return logits
-
- def freeze_layer(self, layer):
- for param in layer.parameters():
- param.requires_grad = False
-
- def get_input_attention(self):
- # Use the attention score computed in the forward pass
- return self.output_layer.label_wise_attn.alpha.cpu().detach().numpy()
-
-
-class OutputLayer(nn.Module):
- def __init__(self, input_size, num_classes):
- super(OutputLayer, self).__init__()
- self.label_wise_attn = LabelWiseAttn(input_size, num_classes)
-
- self.final = nn.Linear(input_size, num_classes)
- xavier_uniform_(self.final.weight)
-
- def forward(self, x):
- m = self.label_wise_attn(x)
- logits = self.final.weight.mul(m).sum(dim=2).add(self.final.bias)
- return logits
diff --git a/src/models/transicd/__init__.py b/src/models/transicd/__init__.py
deleted file mode 100644
index e69de29..0000000
diff --git a/src/modules/__init__.py b/src/modules/__init__.py
deleted file mode 100755
index e69de29..0000000
diff --git a/src/modules/layers/__init__.py b/src/modules/layers/__init__.py
deleted file mode 100644
index e69de29..0000000
diff --git a/src/modules/layers/label_wise_attn.py b/src/modules/layers/label_wise_attn.py
deleted file mode 100644
index ed98e89..0000000
--- a/src/modules/layers/label_wise_attn.py
+++ /dev/null
@@ -1,40 +0,0 @@
-# flake8: noqa
-
-import torch.nn as nn
-import torch.nn.functional as F
-from torch.nn.init import xavier_uniform_
-
-from src.utils.text_loggers import get_logger
-
-logger = get_logger(__name__)
-
-
-class LabelWiseAttn(nn.Module):
- """
- A Label-wise Attention layer (as implemented in CAML, DCAN, etc.).
- References:
- Papers: https://arxiv.org/abs/1802.05695 (Section 2.2)
- Repository: https://github.com/jamesmullenbach/caml-mimic/blob/master/learn/models.py#L184
-
- Args:
- input_size (int): The size of the input, i.e., the number of channels
- if the output is from a convolutional layer/embedding
- size if the output is from a fully connected layer.
- num_classes (int): The number of classes.
- """
-
- def __init__(self, input_size, num_classes):
- super(LabelWiseAttn, self).__init__()
- logger.debug(
- f"Initialising {self.__class__.__name__} with "
- f"input size = {input_size}, num_classes = {num_classes}"
- )
-
- self.U = nn.Linear(input_size, num_classes)
- xavier_uniform_(self.U.weight)
-
- def forward(self, x):
- att = self.U.weight.matmul(x.transpose(1, 2)) # [bs, Y, seq_len]
- self.alpha = F.softmax(att, dim=2)
- m = self.alpha.matmul(x) # [bs, Y, dim]
- return m
diff --git a/src/modules/layers/temporal_block.py b/src/modules/layers/temporal_block.py
deleted file mode 100644
index cb4c598..0000000
--- a/src/modules/layers/temporal_block.py
+++ /dev/null
@@ -1,193 +0,0 @@
-# flake8: noqa
-
-import torch.nn as nn
-from torch.nn.init import xavier_uniform_
-from torch.nn.utils import weight_norm as weight_norm_
-
-from src.modules.activations import *
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
-
-logger = get_logger(__name__)
-
-
-class Chomp1d(nn.Module):
- def __init__(self, chomp_size):
- super(Chomp1d, self).__init__()
- self.chomp_size = chomp_size
-
- def forward(self, x):
- return x[:, :, : -self.chomp_size].contiguous()
-
-
-class ConvTemporalSubBlock(nn.Module):
- """
- A simple temporal convolutional block. Adapted from
- https://github.com/shaoxiongji/DCAN/blob/master/models.py#L84-L88. This
- layer has a dilated convolutional layer, a `chomp1d` layer, followed by
- activation and dropout. For the parameters related to convolutional layers,
- please see this:
- https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html.
-
- Args:
- in_channels (int): The number of input channels in the convolutional
- layer.
- out_channels (int): The number of output channels in the convolutional
- layer.
- kernel_size (int): The size of the kernel in the convolutional layer.
- stride (int): The stride of the convolutional layer.
- padding (int): The padding of the convolutional layer.
- dilation (int): The dilation size of the convolutional layer.
- dropout (float): The dropout probability.
- weight_norm (bool): Whether to apply weight normalization to the
- convolutional layer.
- activation (str): The activation function to use. DCAN uses "relu".
- For all available activations, see
- https://github.com/dalgu90/icd-coding-benchmark/blob/main/src/modules/activations.py.
- """
-
- def __init__(
- self,
- in_channels,
- out_channels,
- kernel_size,
- stride,
- padding,
- dilation,
- dropout=0.2,
- weight_norm=True,
- activation="relu",
- ):
- super(ConvTemporalSubBlock, self).__init__()
- logger.debug(
- f"Initialising {self.__class__.__name__} with "
- f"in_channels = {in_channels}, out_channels = "
- f"{out_channels}, kernel_size = {kernel_size}, "
- f"stride = {stride}, padding = {padding}, "
- f"dilation = {dilation}, dropout = {dropout}, "
- f"weight_norm = {weight_norm}, activation = {activation}"
- )
-
- self.conv_layer = nn.Conv1d(
- in_channels=in_channels,
- out_channels=out_channels,
- kernel_size=kernel_size,
- stride=stride,
- padding=padding,
- dilation=dilation,
- )
- if weight_norm:
- self.conv_layer = weight_norm_(self.conv_layer)
- self.chomp1d = Chomp1d(padding)
- self.activation = ConfigMapper.get_object("activations", activation)()
- self.dropout = nn.Dropout(dropout)
-
- self.__init_weights__()
-
- def __init_weights__(self):
- xavier_uniform_(self.conv_layer.weight)
-
- def forward(self, x):
- x = self.conv_layer(x)
- x = self.chomp1d(x)
- x = self.activation(x)
- x = self.dropout(x)
- return x
-
-
-class TemporalBlock(nn.Module):
- """
- A Temporal Block containing stacks of `ConvTemporalSubBlocks`, followed
- by activation.
- References:
- Paper: https://arxiv.org/abs/2009.14578
- Repository: https://github.com/shaoxiongji/DCAN/blob/master/models.py#L81
-
- Args:
- conv_channel_sizes (list): List of integers, with channel sizes of
- convolutional layers. For example, if the
- list is [100, 200, 300], there will be two
- convolutional layers: Conv1d(100, 200) and
- Conv1d(200, 300).
- kernel_sizes (list): List of integers, with kernel sizes of every
- `ConvTemporalSubBlock`.
- strides (list): List of integers, with strides of convolutional layers.
- paddings (list): List of integers, with paddings of every
- `ConvTemporalSubBlock`.
- dilations (list): List of integers, with dilation sizes of every
- `ConvTemporalSubBlock`.
- dropouts (list): List of floats, with dropout probabilities of every
- `ConvTemporalSubBlock`.
- weight_norm (bool): Whether to apply weight normalization to every
- convolutional layer. DCAN uses weight norm.
- activation (str): The activation function to use. DCAN uses "relu".
- """
-
- def __init__(
- self,
- conv_channel_sizes,
- kernel_sizes,
- strides,
- paddings,
- dilations,
- dropouts,
- weight_norm=True,
- activation="relu",
- ):
- super(TemporalBlock, self).__init__()
- conv_channel_size_pairs = list(
- zip(conv_channel_sizes[:-1], conv_channel_sizes[1:])
- )
-
- self.conv_temporal_sub_blocks = nn.ModuleList(
- [
- ConvTemporalSubBlock(
- in_channels=conv_channel_size_pair[0],
- out_channels=conv_channel_size_pair[1],
- kernel_size=kernel_size,
- stride=stride,
- padding=padding,
- dilation=dilation,
- dropout=dropout,
- weight_norm=weight_norm,
- activation=activation,
- )
- for (
- conv_channel_size_pair,
- kernel_size,
- stride,
- padding,
- dilation,
- dropout,
- ) in zip(
- conv_channel_size_pairs,
- kernel_sizes,
- strides,
- paddings,
- dilations,
- dropouts,
- )
- ]
- )
-
- self.downsample = (
- nn.Conv1d(conv_channel_sizes[0], conv_channel_sizes[-1], 1)
- if conv_channel_sizes[0] != conv_channel_sizes[-1]
- else None
- )
- self.output_activation = ConfigMapper.get_object(
- "activations", activation
- )()
-
- self.init_weights()
-
- def init_weights(self):
- if self.downsample is not None:
- xavier_uniform_(self.downsample.weight)
-
- def forward(self, x):
- conv_layer_output = x
- for conv_temporal_sub_block in self.conv_temporal_sub_blocks:
- conv_layer_output = conv_temporal_sub_block(conv_layer_output)
- res = x if self.downsample is None else self.downsample(x)
- return self.output_activation(conv_layer_output + res)
diff --git a/src/modules/layers/temporal_conv_net.py b/src/modules/layers/temporal_conv_net.py
deleted file mode 100644
index 3222a95..0000000
--- a/src/modules/layers/temporal_conv_net.py
+++ /dev/null
@@ -1,115 +0,0 @@
-# flake8: noqa
-
-import torch.nn as nn
-
-from src.modules.activations import *
-from src.modules.layers.temporal_block import TemporalBlock
-from src.utils.text_loggers import get_logger
-
-logger = get_logger(__name__)
-
-
-class TemporalConvNet(nn.Module):
- """
- Stack of `TemporalBlock`s. Used in the DCAN model.
- References:
- Paper: https://arxiv.org/abs/2009.14578
- Repository: https://github.com/shaoxiongji/DCAN/blob/master/models.py#L114
-
- Args:
- conv_channel_sizes_ (list): List of lists of integers. Each list
- represents the channel sizes of convolutional
- layers in a `TemporalBlock`. So, for
- example, if the list is [[100, 600, 600],
- [600, 600, 600]].
- the `TemporalConvNet` layer will have 2
- `TemporalBlock`s, each temporal block have
- 2 convolutional layers:
- Conv(100, 600), Conv(600, 600) for the first
- one, and Conv(600, 600), Conv(600, 600). If
- the `add_emb_size_to_channel_sizes`, we
- don't have to pass the input channel size.
- So, in the above case, we can just pass
- [[600, 600], [600, 600, 600]].
- kernel_sizes_ (list): List of list of integers (same format as
- `conv_channel_sizes`). Each integer represents the
- kernel size/filter size of the respective
- convolutional layer in `TemporalBlock` layer.
- strides_ (list): List of list of integers (same format as
- `conv_channel_sizes`). Each integer represents the
- stride of the respective convolutional layer in
- `TemporalBlock` layer.
- paddings_ (list): List of list of integers (same format as
- `conv_channel_sizes`). Each integer represents the
- padding of the respective convolutional layer in
- `TemporalBlock` layer. in DCAN, this value is set to
- "(kernel_size - 1) * dilation_size".
- dilations_ (list): List of list of integers (same format as
- `conv_channel_sizes`). Each integer represents the
- dilation size of the respective convolutional layer
- `TemporalBlock` layer.` In DCAN, this value is
- "2^(temporal_block_level)".
- dropouts_ (list): List of list of floats (same format as
- `conv_channel_sizes`). Each float represents the
- dropout probability of the respective convolutional
- `TemporalBlock` layer.
- weight_norm (bool): If True, apply weight normalization to the
- convolutional layers.
- activation (str): Activation function to use. DCAN uses "relu".
- """
-
- def __init__(
- self,
- conv_channel_sizes_,
- kernel_sizes_,
- strides_,
- paddings_,
- dilations_,
- dropouts_,
- weight_norm=True,
- activation="relu",
- ):
- super(TemporalConvNet, self).__init__()
- logger.debug(
- f"Initialising {self.__class__.__name__} with "
- f"conv_channel_sizes_ = {conv_channel_sizes_}, "
- f"kernel_sizes_ = {kernel_sizes_}, "
- f"strides_ = {strides_}, paddings_ = {paddings_}, "
- f"dilations_ = {dilations_}, dropouts_ = {dropouts_}, "
- f"weight_norm = {weight_norm}, activation = {activation}"
- )
-
- self.temporal_blocks = nn.ModuleList(
- [
- TemporalBlock(
- conv_channel_sizes=conv_channel_sizes,
- kernel_sizes=kernel_sizes,
- strides=strides,
- paddings=paddings,
- dilations=dilations,
- dropouts=dropouts,
- weight_norm=weight_norm,
- activation=activation,
- )
- for (
- conv_channel_sizes,
- kernel_sizes,
- strides,
- paddings,
- dilations,
- dropouts,
- ) in zip(
- conv_channel_sizes_,
- kernel_sizes_,
- strides_,
- paddings_,
- dilations_,
- dropouts_,
- )
- ]
- )
-
- def forward(self, x):
- for temporal_block in self.temporal_blocks:
- x = temporal_block(x)
- return x
diff --git a/src/modules/layers/word_embedding_layer.py b/src/modules/layers/word_embedding_layer.py
deleted file mode 100644
index ecdfd43..0000000
--- a/src/modules/layers/word_embedding_layer.py
+++ /dev/null
@@ -1,44 +0,0 @@
-import torch
-import torch.nn as nn
-
-from src.utils.mapper import ConfigMapper
-from src.utils.text_loggers import get_logger
-
-logger = get_logger(__name__)
-
-
-class WordEmbeddingLayer(nn.Module):
- """
- A Word Embedding Layer. This layer loads a pre-trained word embedding matrix
- , and copies its weights to an nn.Embedding layer.
-
- Args:
- embed_dir (str): A directory containing the pre-trained word embedding
- matrix, among other things. Please see
- https://github.com/dalgu90/icd-coding-benchmark/blob/main/src/modules/embeddings.py#L17
- for more details.
- dropout (float): The dropout probability.
- """
-
- def __init__(self, embed_dir, dropout):
- super(WordEmbeddingLayer, self).__init__()
- logger.debug(
- f"Initialising {self.__class__.__name__} with "
- f"embed_dir = {embed_dir}, dropout = {dropout}"
- )
-
- # Note: This should be changed, since we won't always use Word2Vec.
- embedding_cls = ConfigMapper.get_object("embeddings", "word2vec")
-
- W = torch.Tensor(embedding_cls.load_emb_matrix(embed_dir))
- self.embed = nn.Embedding(W.size()[0], W.size()[1], padding_idx=0)
- self.embed.weight.data = W.clone()
-
- self.embedding_size = self.embed.embedding_dim
-
- self.dropout = nn.Dropout(dropout)
-
- def forward(self, x):
- embedding = self.embed(x)
- x = self.dropout(embedding)
- return x
diff --git a/src/trainers/__init__.py b/src/trainers/__init__.py
deleted file mode 100644
index c6668cb..0000000
--- a/src/trainers/__init__.py
+++ /dev/null
@@ -1 +0,0 @@
-from src.trainers.base_trainer import BaseTrainer
diff --git a/src/utils/__init__.py b/src/utils/__init__.py
deleted file mode 100755
index e69de29..0000000