Guendel19_Chexpert_CWBCE_E20_B32 - TobiasSchmidtDE/DeepL-MedicalImaging GitHub Wiki
Version: 1
Trained Guendel19 architecture using the 'Chexpert CWBCE E20 B32' benchmark. The benchmark was initialized for the chexpert_full dataset with batch size of 32, shuffel set to True and images rescaled to dimension (256, 256). The training was done for 10 epochs using the Adam optimizer and weighted_binary_crossentropy loss. A total of 14 labels/pathologies were included in the training and encoded using the 'uzeroes' method. The traing set included 143398 number of sample, the validation set 35646, and the test set 44370.
from pathlib import Path
import pandas as pd
import matplotlib.pyplot as plt
from matplotlib.pyplot import figure
import json
import os
import re
import pprintdata = json.loads(os.environ['EXP_DATA'])
history = data['history']
for s in data["description"].split(".")[:-1]:
print(s + ".\n")Trained Guendel19 architecture using the 'Chexpert CWBCE E20 B32' benchmark.
The benchmark was initialized for the chexpert_full dataset with batch size of 32, shuffel set to True and images rescaled to dimension (256, 256).
The training was done for 10 epochs using the Adam optimizer and weighted_binary_crossentropy loss.
A total of 14 labels/pathologies were included in the training and encoded using the 'uzeroes' method.
The traing set included 143398 number of sample, the validation set 35646, and the test set 44370.
# if there are any metrics that were renamed, add this new name here as ("default_name":"new_name")
metric_custom_names={"auc":"AUC_ROC"}
metric_names = [re.sub("([a-z0-9])([A-Z])","\g<1> \g<2>",name) for name in data["benchmark"]["metrics"]]
metric_keys = [re.sub("([a-z0-9])([A-Z])","\g<1>_\g<2>",name).lower() for name in data["benchmark"]["metrics"]]
for default_name, custom_name in metric_custom_names.items():
if not default_name in history.keys() and default_name in metric_keys:
#replace default name with custom name
metric_keys[metric_keys.index(default_name)]=custom_namedef print_or_plot_metric(metric_key, metric_name, figure_name):
if len(history[metric_key]) == 1:
print("Data for {m_name} only available for a single epoch. \nSkipping plot and printing data...".format(m_name=metric_name))
print('Train {}: '.format(metric_name), history[metric_key])
print('Validation {}: '.format(metric_name), history['val_'+metric_key])
print()
else:
plot_epoch_metric(metric_key, metric_name, figure_name)
def plot_epoch_metric(metric_key, metric_name, figure_name):
figure(num=None, figsize=(10, 6))
plt.plot(history[metric_key])
if 'val_'+metric_key in history.keys():
plt.plot(history['val_'+metric_key])
plt.title(figure_name)
plt.ylabel(metric_name)
plt.xlabel('Epoch')
if 'val_'+metric_key in history.keys():
plt.legend(['Train', 'Validation'], loc='upper left')
plt.show()
for i, metric_key in enumerate(metric_keys):
print_or_plot_metric(metric_key, metric_names[i], "Model "+metric_names[i])
print_or_plot_metric("loss", "Loss", "Model loss")
if "lr" in history.keys():
plot_epoch_metric("lr", "Learning Rate", "Learning Rate")
if 'classification_report' in data.keys() and data['classification_report']:
print(data['classification_report']) precision recall f1-score support
No Finding 0.10 0.38 0.15 4383
Enlarged Cardiomediastinum 0.05 0.46 0.09 2137
Cardiomegaly 0.12 0.33 0.18 5454
Lung Opacity 0.47 0.40 0.43 21069
Lung Lesion 0.04 0.22 0.07 1931
Edema 0.24 0.24 0.24 10439
Consolidation 0.07 0.25 0.11 3030
Pneumonia 0.03 0.11 0.05 1256
Atelectasis 0.14 0.37 0.21 6421
Pneumothorax 0.09 0.20 0.12 3816
Pleural Effusion 0.37 0.40 0.39 16713
Pleural Other 0.02 0.17 0.03 745
Fracture 0.04 0.17 0.06 1726
Support Devices 0.52 0.44 0.48 23022
micro avg 0.20 0.36 0.26 102142
macro avg 0.16 0.30 0.19 102142
weighted avg 0.33 0.36 0.33 102142
samples avg 0.19 0.34 0.22 102142
if 'test' in data.keys() and data['test']:
for score_name, score in data["test"].items():
print('Test {}: '.format(score_name), score)
Test loss: 1.3039888143539429
Test AUC: 0.7536254525184631
Test Precision: 0.3530292510986328
Test Recall: 0.6324920058250427
Test F2Score: 0.5460412502288818
Test BinaryAccuracy: 0.7489665150642395
pp = pprint.PrettyPrinter(indent=4)
if "benchmark" in data.keys():
pp.pprint(data["benchmark"]){ 'batch_size': 32,
'benchmark_name': 'Chexpert CWBCE E20 B32',
'dataset_folder': 'data/chexpert/full',
'dataset_name': 'chexpert_full',
'dim': [256, 256],
'drop_last': True,
'epochs': 10,
'label_columns': [ 'No Finding',
'Enlarged Cardiomediastinum',
'Cardiomegaly',
'Lung Opacity',
'Lung Lesion',
'Edema',
'Consolidation',
'Pneumonia',
'Atelectasis',
'Pneumothorax',
'Pleural Effusion',
'Pleural Other',
'Fracture',
'Support Devices'],
'loss': 'weighted_binary_crossentropy',
'metrics': ['AUC', 'Precision', 'Recall', 'F2Score', 'BinaryAccuracy'],
'models_dir': 'models',
'n_channels': 3,
'nan_replacement': 0,
'optimizer': 'Adam',
'path_column': 'Path',
'path_column_prefix': '',
'shuffle': True,
'test_num_samples': 44370,
'train_num_samples': 143398,
'u_enc': 'uzeroes',
'unc_value': -1,
'valid_num_samples': 35646}
Version: 1
Trained Guendel19 architecture using the 'Chexpert CWBCE E20 B32' benchmark. The benchmark was initialized for the chexpert_full dataset with batch size of 32, shuffel set to True and images rescaled to dimension (256, 256). The training was done for 10 epochs using the Adam optimizer and weighted_binary_crossentropy loss. A total of 14 labels/pathologies were included in the training and encoded using the 'uzeroes' method. The traing set included 143398 number of sample, the validation set 35646, and the test set 44370.
from pathlib import Path
from dotenv import load_dotenv, find_dotenv
import pandas as pd
import matplotlib.pyplot as plt
from matplotlib.pyplot import figure
import json
import os
import re
import pprint
basepath = Path(os.getcwd())
if basepath.name != "idp-radio-1":
os.chdir(basepath.parent.parent)
print(os.getcwd())
load_dotenv(find_dotenv())
from src.preprocessing.split.train_test_split import train_test_split/srv/idp-radio-1
data = json.loads(os.environ['EXP_DATA'])
history = data['history']for s in data["description"].split(".")[:-1]:
print(s + ".\n")Trained Guendel19 architecture using the 'Chexpert CWBCE E20 B32' benchmark.
The benchmark was initialized for the chexpert_full dataset with batch size of 32, shuffel set to True and images rescaled to dimension (256, 256).
The training was done for 10 epochs using the Adam optimizer and weighted_binary_crossentropy loss.
A total of 14 labels/pathologies were included in the training and encoded using the 'uzeroes' method.
The traing set included 143398 number of sample, the validation set 35646, and the test set 44370.
# if there are any metrics that were renamed, add this new name here as ("default_name":"new_name")
metric_custom_names={"auc":"AUC_ROC"}
metric_names = [re.sub("([a-z0-9])([A-Z])","\g<1> \g<2>",name) for name in data["benchmark"]["metrics"]]
metric_keys = [re.sub("([a-z0-9])([A-Z])","\g<1>_\g<2>",name).lower() for name in data["benchmark"]["metrics"]]
for default_name, custom_name in metric_custom_names.items():
if not default_name in history.keys() and default_name in metric_keys:
#replace default name with custom name
metric_keys[metric_keys.index(default_name)]=custom_namedef print_or_plot_metric(metric_key, metric_name, figure_name):
if len(history[metric_key]) == 1:
print("Data for {m_name} only available for a single epoch. \nSkipping plot and printing data...".format(m_name=metric_name))
print('Train {}: '.format(metric_name), history[metric_key])
print('Validation {}: '.format(metric_name), history['val_'+metric_key])
print()
else:
plot_epoch_metric(metric_key, metric_name, figure_name)
def plot_epoch_metric(metric_key, metric_name, figure_name):
figure(num=None, figsize=(10, 6))
plt.plot(history[metric_key])
if 'val_'+metric_key in history.keys():
plt.plot(history['val_'+metric_key])
plt.title(figure_name)
plt.ylabel(metric_name)
plt.xlabel('Epoch')
if 'val_'+metric_key in history.keys():
plt.legend(['Train', 'Validation'], loc='upper left')
plt.show()
for i, metric_key in enumerate(metric_keys):
print_or_plot_metric(metric_key, metric_names[i], "Model "+metric_names[i])
print_or_plot_metric("loss", "Loss", "Model loss")
if "lr" in history.keys():
plot_epoch_metric("lr", "Learning Rate", "Learning Rate")
if 'classification_report' in data.keys() and data['classification_report']:
print(data['classification_report']) precision recall f1-score support
No Finding 0.10 0.38 0.15 4383
Enlarged Cardiomediastinum 0.05 0.46 0.09 2137
Cardiomegaly 0.12 0.33 0.18 5454
Lung Opacity 0.47 0.40 0.43 21069
Lung Lesion 0.04 0.22 0.07 1931
Edema 0.24 0.24 0.24 10439
Consolidation 0.07 0.25 0.11 3030
Pneumonia 0.03 0.11 0.05 1256
Atelectasis 0.14 0.37 0.21 6421
Pneumothorax 0.09 0.20 0.12 3816
Pleural Effusion 0.37 0.40 0.39 16713
Pleural Other 0.02 0.17 0.03 745
Fracture 0.04 0.17 0.06 1726
Support Devices 0.52 0.44 0.48 23022
micro avg 0.20 0.36 0.26 102142
macro avg 0.16 0.30 0.19 102142
weighted avg 0.33 0.36 0.33 102142
samples avg 0.19 0.34 0.22 102142
if 'test' in data.keys() and data['test']:
for score_name, score in data["test"].items():
print('Test {}: '.format(score_name), score)
Test loss: 1.3039888143539429
Test AUC: 0.7536254525184631
Test Precision: 0.3530292510986328
Test Recall: 0.6324920058250427
Test F2Score: 0.5460412502288818
Test BinaryAccuracy: 0.7489665150642395
pp = pprint.PrettyPrinter(indent=4)
if "benchmark" in data.keys():
pp.pprint(data["benchmark"]){ 'batch_size': 32,
'benchmark_name': 'Chexpert CWBCE E20 B32',
'dataset_folder': 'data/chexpert/full',
'dataset_name': 'chexpert_full',
'dim': [256, 256],
'drop_last': True,
'epochs': 10,
'label_columns': [ 'No Finding',
'Enlarged Cardiomediastinum',
'Cardiomegaly',
'Lung Opacity',
'Lung Lesion',
'Edema',
'Consolidation',
'Pneumonia',
'Atelectasis',
'Pneumothorax',
'Pleural Effusion',
'Pleural Other',
'Fracture',
'Support Devices'],
'loss': 'weighted_binary_crossentropy',
'metrics': ['AUC', 'Precision', 'Recall', 'F2Score', 'BinaryAccuracy'],
'models_dir': 'models',
'n_channels': 3,
'nan_replacement': 0,
'optimizer': 'Adam',
'path_column': 'Path',
'path_column_prefix': '',
'shuffle': True,
'test_num_samples': 44370,
'train_num_samples': 143398,
'u_enc': 'uzeroes',
'unc_value': -1,
'valid_num_samples': 35646}
if 'benchmark' in data.keys() and 'split_seed' in data['benchmark']:
benchmark = data['benchmark']
dataset_path = Path(benchmark['dataset_folder'])
train_labels = benchmark['train_labels'] if 'train_labels' in benchmark.keys() else 'train.csv'
split_test_size = benchmark['split_test_size'] if 'split_test_size' in benchmark.keys() else 0.2
split_valid_size = benchmark['split_valid_size'] if 'split_valid_size' in benchmark.keys() else 0.2
split_group = benchmark['split_group'] if 'split_group' in benchmark.keys() else 'patient_id'
split_seed = benchmark['split_seed']
all_labels = pd.read_csv(dataset_path / train_labels)
train_labels, test_labels = train_test_split(all_labels, test_size=split_test_size, group=split_group, seed=split_seed)
train_labels, validation_labels = train_test_split(train_labels, test_size=split_valid_size, group=split_group, seed=split_seed)from src.datasets.u_encoding import uencode
def get_distribution(labels):
if 'nan_replacement' in benchmark.keys():
labels = labels.fillna(benchmark['nan_replacement'])
data = labels.to_numpy()
data = uencode(benchmark['u_enc'], data, unc_value=benchmark['unc_value'])
data = pd.DataFrame(data, columns=labels.columns)
labels = data[benchmark['label_columns']]
d = {'Pathology': [], 'Positive': [], 'Positive %': [], 'Negative': [], 'Negative %': [],}
for label in labels.columns:
values = labels.groupby(label)
d['Pathology'].append(label)
positive = values.size()[1.0] if 1.0 in values.size() else 0
positive_percent = positive / labels.shape[0] * 100
d['Positive'].append(positive)
d['Positive %'].append(round(positive_percent))
negative = values.size()[-0.0] if -0.0 in values.size() else 0
negative_percent = negative / labels.shape[0] * 100
d['Negative'].append(negative)
d['Negative %'].append(round(negative_percent))
df = pd.DataFrame(d)
df = df.set_index('Pathology')
return dfif 'benchmark' in data.keys() and 'split_seed' in data['benchmark']:
train = get_distribution(train_labels)
val = get_distribution(validation_labels)
test = get_distribution(test_labels)
positives = train[['Positive %']].merge(val[['Positive %']], left_index=True, right_index=True).merge(test[['Positive %']], left_index=True, right_index=True).rename(columns={"Positive %_x": "Positives Train", "Positive %_y": "Positives Validation", "Positive %": "Positives Test", })
positives.copy().plot(kind='bar', figsize=(10,7), title="Positive Labels Distribution")
negatives = train[['Negative %']].merge(val[['Negative %']], left_index=True, right_index=True).merge(test[['Negative %']], left_index=True, right_index=True).rename(columns={"Negative %_x": "Negative Train", "Negative %_y": "Negative Validation", "Negative %": "Negative Test", })
negatives.copy().plot(kind='bar', figsize=(10,7), title="Negative Labels Distribution")
train[['Positive %', 'Negative %']].copy().plot(kind='bar', figsize=(10,7), title="Training set")
val[['Positive %', 'Negative %']].copy().plot(kind='bar', figsize=(10,7), title="Validation set")
test[['Positive %', 'Negative %']].copy().plot(kind='bar', figsize=(10,7), title="Test set")