import io
import os
import statistics
from pathlib import Path
import pandas as pd
from immuneML.environment.Constants import Constants
from immuneML.environment.EnvironmentSettings import EnvironmentSettings
from immuneML.hyperparameter_optimization.config.SplitType import SplitType
from immuneML.hyperparameter_optimization.states.HPAssessmentState import HPAssessmentState
from immuneML.hyperparameter_optimization.states.HPItem import HPItem
from immuneML.hyperparameter_optimization.states.HPLabelState import HPLabelState
from immuneML.hyperparameter_optimization.states.HPSelectionState import HPSelectionState
from immuneML.hyperparameter_optimization.states.TrainMLModelState import TrainMLModelState
from immuneML.ml_methods.util.Util import Util as MLUtil
from immuneML.ml_metrics.ClassificationMetric import ClassificationMetric
from immuneML.presentation.TemplateParser import TemplateParser
from immuneML.presentation.html.Util import Util
from immuneML.reports.ReportResult import ReportResult
from immuneML.util.PathBuilder import PathBuilder
from immuneML.util.StringHelper import StringHelper
[docs]
class HPHTMLBuilder:
"""
A class that will make HTML file(s) out of TrainMLModelState object to show what analysis took place in the TrainMLModel.
"""
CSS_PATH = EnvironmentSettings.html_templates_path / "css/custom.css"
NUM_DIGITS = 3
[docs]
@staticmethod
def build(state: TrainMLModelState = None) -> Path:
"""
Function that builds the HTML files based on the HPOptimization state.
Arguments:
state: HPOptimizationState object with all details on the optimization
Returns:
path to the main HTML file (index.html which is located under state.result_path)
"""
base_path = PathBuilder.build(state.path / "../HTML_output/")
state = HPHTMLBuilder._move_reports_recursive(state, base_path)
html_map = HPHTMLBuilder._make_main_html_map(state, base_path)
result_file = base_path / f"TrainMLModelReport_{state.name}.html"
TemplateParser.parse(template_path=EnvironmentSettings.html_templates_path / "HPOptimization.html",
template_map=html_map, result_path=result_file)
for label_name in state.label_configuration.get_labels_by_name():
for index, item in enumerate(HPHTMLBuilder._make_assessment_pages(state, base_path, label_name)):
TemplateParser.parse(template_path=EnvironmentSettings.html_templates_path / "AssessmentSplitDetails.html",
template_map=item,
result_path=base_path / HPHTMLBuilder._make_assessment_split_path(index, state.name, label_name))
for label_name in state.label_configuration.get_labels_by_name():
for assessment_index in range(state.assessment.split_count):
TemplateParser.parse(template_path=EnvironmentSettings.html_templates_path / "SelectionDetails.html",
template_map=HPHTMLBuilder._make_selection(state, assessment_index, label_name, base_path),
result_path=base_path / HPHTMLBuilder._make_selection_split_path(assessment_index, label_name, state.name))
return result_file
@staticmethod
def _make_assessment_split_path(split_index: int, state_name: str, label_name: str) -> Path:
return Path(f"{state_name}_CV_assessment_split_{split_index + 1}_{label_name}.html")
@staticmethod
def _make_selection_split_path(assessment_index: int, label_name: str, state_name: str) -> Path:
return Path(f"{state_name}_selection_details_{label_name}_split_{assessment_index + 1}.html")
@staticmethod
def _make_selection(state: TrainMLModelState, assessment_index: int, label_name: str, base_path):
selection_state = state.assessment_states[assessment_index].label_states[label_name].selection_state
hp_settings = []
optimal = selection_state.optimal_hp_setting.get_key()
for hp_setting, hp_items in selection_state.hp_items.items():
hp_splits = []
for hp_item in hp_items:
hp_splits.append(HPHTMLBuilder._print_metric(hp_item.performance, state.optimization_metric))
hp_settings.append({
"hp_setting": hp_setting,
"hp_splits": hp_splits,
"optimal": hp_setting == optimal
})
if len(hp_splits) > 1:
hp_settings[-1]["average"] = round(statistics.mean(perf for perf in hp_splits if [isinstance(perf, float)]), HPHTMLBuilder.NUM_DIGITS)
hp_settings[-1]["show_average"] = True
else:
hp_settings[-1]["average"] = None
hp_settings[-1]["show_average"] = False
hp_settings[-1]["hp_splits"] = [{"optimization_metric_val": val} for val in hp_settings[-1]["hp_splits"]]
has_other_metrics = len([metric for metric in state.metrics if metric != state.optimization_metric]) > 0 and \
not (state.selection.split_strategy == SplitType.RANDOM and state.selection.training_percentage == 1)
return {
"css_style": Util.get_css_content(HPHTMLBuilder.CSS_PATH),
"label": label_name,
"assessment_split": assessment_index + 1,
"splits": [{"split_index": i} for i in range(1, state.selection.split_count + 1)],
"split_count": state.selection.split_count,
"optimization_metric": state.optimization_metric.name.lower(),
"has_other_metrics": has_other_metrics,
"metrics": [{"performance": HPHTMLBuilder._extract_selection_performance_per_metric(selection_state, metric, state.selection.split_count),
"metric": HPHTMLBuilder._get_heading_metric_name(metric.name.lower())}
for metric in state.metrics if metric != state.optimization_metric] if has_other_metrics else None,
"hp_settings": hp_settings,
"show_average": any(hps["show_average"] for hps in hp_settings),
"data_split_reports": [
{'split_index': index + 1,
'train': Util.to_dict_recursive(selection_state.train_data_reports[index], base_path)
if len(selection_state.train_data_reports) == state.selection.split_count else None,
'test': Util.to_dict_recursive(selection_state.val_data_reports[index], base_path)
if len(selection_state.train_data_reports) == state.selection.split_count else None}
for index in range(state.selection.split_count)] if len(state.selection.reports.data_split_reports) > 0 else None,
"has_data_split_reports": len(state.selection.reports.data_split_reports) > 0,
"has_reports_per_setting": len(state.selection.reports.encoding_reports) + len(state.selection.reports.model_reports) > 0,
"reports_per_setting": [{
"hp_setting": hp_setting,
"reports": HPHTMLBuilder._make_selection_reports_for_item_list(hp_items, base_path)
} for hp_setting, hp_items in selection_state.hp_items.items()]
}
@staticmethod
def _make_selection_reports_for_item_list(hp_items: list, base_path) -> list:
result = []
for split_index, hp_item in enumerate(hp_items):
result.append({
"split_index": split_index + 1,
"has_encoding_train_reports": len(hp_item.encoding_train_results) > 0,
"has_encoding_test_reports": len(hp_item.encoding_test_results) > 0,
"has_ml_reports": len(hp_item.model_report_results) > 0,
"encoding_train_reports": Util.to_dict_recursive(hp_item.encoding_train_results, base_path) if len(
hp_item.encoding_train_results) > 0 else None,
"encoding_test_reports": Util.to_dict_recursive(hp_item.encoding_test_results, base_path) if len(
hp_item.encoding_test_results) > 0 else None,
"ml_reports": Util.to_dict_recursive(hp_item.model_report_results, base_path) if len(hp_item.model_report_results) > 0 else None,
})
return result if len(result) > 0 else None
@staticmethod
def _make_assessment_pages(state: TrainMLModelState, base_path: Path, label_name: str):
assessment_list = []
for i, assessment_state in enumerate(state.assessment_states):
assessment_item = {"css_style": Util.get_css_content(HPHTMLBuilder.CSS_PATH),
"optimization_metric": state.optimization_metric.name.lower(),
"split_index": assessment_state.split_index + 1,
"hp_settings": [],
"has_reports": len(state.assessment.reports.model_reports) + len(state.assessment.reports.encoding_reports) > 0,
"train_data_reports": Util.to_dict_recursive(assessment_state.train_val_data_reports, base_path),
"test_data_reports": Util.to_dict_recursive(assessment_state.test_data_reports, base_path),
"show_data_reports": len(assessment_state.train_val_data_reports) > 0 or len(assessment_state.test_data_reports) > 0}
if hasattr(assessment_state.train_val_dataset, "metadata_file") and assessment_state.train_val_dataset.metadata_file is not None:
assessment_item["train_metadata_path"] = os.path.relpath(str(assessment_state.train_val_dataset.metadata_file), str(base_path))
assessment_item["train_metadata"] = Util.get_table_string_from_csv(assessment_state.train_val_dataset.metadata_file)
else:
assessment_item["train_metadata_path"] = None
if hasattr(assessment_state.test_dataset, "metadata_file") and assessment_state.test_dataset.metadata_file is not None:
assessment_item['test_metadata_path'] = os.path.relpath(assessment_state.test_dataset.metadata_file, base_path)
assessment_item["test_metadata"] = Util.get_table_string_from_csv(assessment_state.test_dataset.metadata_file)
else:
assessment_item["test_metadata_path"] = None
assessment_item["label"] = label_name
for hp_setting, item in assessment_state.label_states[label_name].assessment_items.items():
optimal = str(assessment_state.label_states[label_name].optimal_hp_setting.get_key())
reports_path = HPHTMLBuilder._make_assessment_reports(state, i, hp_setting, assessment_state, label_name, base_path)
assessment_item["hp_settings"].append({
"optimal": str(hp_setting) == optimal,
"hp_setting": str(hp_setting),
"optimization_metric_val": HPHTMLBuilder._print_metric(item.performance, state.optimization_metric),
"reports_path": reports_path
})
assessment_item["show_non_optimal"] = len(assessment_item["hp_settings"]) > 1
assessment_item["selection_path"] = HPHTMLBuilder._make_selection_split_path(i, label_name, state.name)
assessment_item['performances_per_metric'] = HPHTMLBuilder._extract_assessment_performances_per_metric(state, assessment_state, label_name)
assessment_list.append(assessment_item)
return assessment_list
@staticmethod
def _extract_assessment_performances_per_metric(state: TrainMLModelState, assessment_state: HPAssessmentState, label_name: str) -> str:
performance_metric = {"setting": [], **{metric.name.lower(): [] for metric in state.metrics}}
for hp_setting, hp_item in assessment_state.label_states[label_name].assessment_items.items():
performance_metric['setting'].append(str(hp_setting))
for metric in sorted(state.metrics, key=lambda metric: metric.name.lower()):
performance_metric[metric.name.lower()].append(HPHTMLBuilder._print_metric(hp_item.performance, metric))
s = io.StringIO()
pd.DataFrame(performance_metric).rename(columns={"setting": 'Hyperparameter settings (preprocessing, encoding, ML method)'})\
.to_csv(s, sep="\t", index=False)
return Util.get_table_string_from_csv_string(s.getvalue(), separator="\t")
@staticmethod
def _make_assessment_reports(state, i, hp_setting_key, assessment_state, label_name: str, base_path: Path):
path = base_path / f"{state.name}_{label_name}_{hp_setting_key}_assessment_reports_split_{i + 1}.html"
hp_item = assessment_state.label_states[label_name].assessment_items[hp_setting_key]
data = {
"split_index": i + 1,
"hp_setting": hp_setting_key,
"label": label_name,
"css_style": Util.get_css_content(HPHTMLBuilder.CSS_PATH),
"has_encoding_reports": len(hp_item.encoding_train_results) > 0 or len(hp_item.encoding_test_results) > 0,
"has_ml_reports": len(hp_item.model_report_results) > 0,
"encoding_train_reports": Util.to_dict_recursive(hp_item.encoding_train_results, base_path) if len(
hp_item.encoding_train_results) > 0 else None,
"encoding_test_reports": Util.to_dict_recursive(hp_item.encoding_test_results, base_path) if len(
hp_item.encoding_test_results) > 0 else None,
"ml_reports": Util.to_dict_recursive(hp_item.model_report_results, base_path) if len(
hp_item.model_report_results) > 0 else None
}
if data["has_ml_reports"] or data["has_encoding_reports"]:
TemplateParser.parse(template_path=EnvironmentSettings.html_templates_path / "Reports.html", template_map=data, result_path=path)
return path.name
else:
return None
@staticmethod
def _make_hp_per_label(state: TrainMLModelState):
mapping = []
for label_name in state.label_configuration.get_labels_by_name():
results = []
for i, assessment_state in enumerate(state.assessment_states):
results.append({
"index": assessment_state.split_index + 1,
"hp_setting": assessment_state.label_states[label_name].optimal_assessment_item.hp_setting,
"optimization_metric_val": HPHTMLBuilder._print_metric(assessment_state.label_states[label_name].optimal_assessment_item.performance,
state.optimization_metric),
"split_details_path": HPHTMLBuilder._make_assessment_split_path(assessment_state.split_index, state.name, label_name)
})
mapping.append({"label": label_name, "assessment_results": results})
return mapping
@staticmethod
def _print_metric(performance: dict, metric: ClassificationMetric):
if performance is not None and metric.name.lower() in performance:
if isinstance(performance[metric.name.lower()], float):
return round(performance[metric.name.lower()], HPHTMLBuilder.NUM_DIGITS)
else:
return performance[metric.name.lower()]
else:
return Constants.NOT_COMPUTED
@staticmethod
def _make_model_per_label(state: TrainMLModelState, base_path: Path) -> list:
mapping = []
for label_name in state.label_configuration.get_labels_by_name():
mapping.append({
"label": label_name,
"model_path": Path(os.path.relpath(path=str(state.optimal_hp_item_paths[label_name]), start=str(base_path)))
})
return mapping
@staticmethod
def _make_main_html_map(state: TrainMLModelState, base_path: Path) -> dict:
html_map = {**Util.make_dataset_html_map(state.dataset), **{
"css_style": Util.get_css_content(HPHTMLBuilder.CSS_PATH),
"full_specs": Util.get_full_specs_path(base_path),
"logfile": Util.get_logfile_path(base_path),
"labels": [{"name": label.name, "values": str(label.values)[1:-1]} for label in state.label_configuration.get_label_objects()],
"optimization_metric": state.optimization_metric.name.lower(),
"other_metrics": str([metric.name.lower() for metric in state.metrics])[1:-1].replace("'", ""),
"metrics": [{"name": metric.name.lower()} for metric in state.metrics],
"assessment_desc": state.assessment,
"selection_desc": state.selection,
"show_hp_reports": bool(state.report_results),
'hp_reports': Util.to_dict_recursive(state.report_results, base_path) if state.report_results else None,
"hp_per_label": HPHTMLBuilder._make_hp_per_label(state),
'models_per_label': HPHTMLBuilder._make_model_per_label(state, base_path),
'immuneML_version': MLUtil.get_immuneML_version()
}}
return html_map
@staticmethod
def _move_reports_recursive(obj, path: Path):
for attribute in (vars(obj) if not isinstance(obj, dict) else obj):
attribute_value = getattr(obj, attribute) if not isinstance(obj, dict) else obj[attribute]
if isinstance(attribute_value, list) and all(isinstance(item, ReportResult) for item in attribute_value):
new_attribute_values = []
for report_result in attribute_value:
new_attribute_values.append(Util.update_report_paths(report_result, path))
setattr(obj, attribute, new_attribute_values)
elif isinstance(attribute_value, list) and all(isinstance(item, HPAssessmentState) for item in attribute_value):
obj = HPHTMLBuilder._process_list_recursively(obj, attribute, attribute_value, path)
elif isinstance(attribute_value, dict) and all(
isinstance(item, HPLabelState) or isinstance(item, HPItem) for item in attribute_value.values()):
obj = HPHTMLBuilder._process_dict_recursive(obj, attribute, attribute_value, path)
elif isinstance(attribute_value, dict) and all(isinstance(item, list) for item in attribute_value.values()) and all(
all(isinstance(item, HPItem) for item in item_list) for item_list in attribute_value.values()):
obj = HPHTMLBuilder._process_hp_items(obj, attribute, attribute_value, path)
elif isinstance(attribute_value, HPSelectionState):
setattr(obj, attribute, HPHTMLBuilder._move_reports_recursive(attribute_value, path))
return obj
@staticmethod
def _process_hp_items(obj, attribute, attribute_value, path: Path):
new_attribute_value = {}
for hp_setting, hp_item_list in attribute_value.items():
new_hp_item_list = []
for hp_item in hp_item_list:
new_hp_item_list.append(HPHTMLBuilder._move_reports_recursive(hp_item, path))
new_attribute_value[hp_setting] = new_hp_item_list
setattr(obj, attribute, new_attribute_value)
return obj
@staticmethod
def _process_dict_recursive(obj, attribute, attribute_value, path: Path):
for key, value in attribute_value.items():
attribute_value[key] = HPHTMLBuilder._move_reports_recursive(value, path)
setattr(obj, attribute, attribute_value)
return obj
@staticmethod
def _process_list_recursively(obj, attribute, attribute_value, path: Path):
new_attribute_values = []
for item in attribute_value:
new_attribute_values.append(HPHTMLBuilder._move_reports_recursive(item, path))
setattr(obj, attribute, new_attribute_values)
return obj
@staticmethod
def _extract_selection_performance_per_metric(selection_state: HPSelectionState, metric: ClassificationMetric, split_count):
performance = {"setting": [], **{f"split {i + 1}": [] for i in range(split_count)}}
for hp_setting, hp_item_list in selection_state.hp_items.items():
performance['setting'].append(str(hp_setting))
for index, hp_item in enumerate(hp_item_list):
performance[f'split {index + 1}'].append(HPHTMLBuilder._print_metric(hp_item.performance, metric))
s = io.StringIO()
pd.DataFrame(performance).rename(columns={"setting": 'Hyperparameter settings (preprocessing, encoding, ML method)'}).to_csv(s, sep="\t",
index=False)
return Util.get_table_string_from_csv_string(s.getvalue(), separator="\t")
@staticmethod
def _get_heading_metric_name(metric: str):
if metric != "auc":
return " ".join(metric.split("_")).title()
else:
return metric.upper()