view multimodal_learner.xml @ 0:375c36923da1 draft default tip

planemo upload for repository https://github.com/goeckslab/gleam.git commit 1c6c1ad7a1b2bd3645aa0eafa2167784820b52e0

<tool id="multimodal_learner" name="Multimodal Learner" version="0.1.0" profile="22.01">
  <description>Train and evaluate an AutoGluon Multimodal model (tabular + image + text)</description>

  <requirements>
    <container type='docker'>quay.io/goeckslab/multimodal-learner:1.4.0</container>
  </requirements>

  <required_files>
    <include path="multimodal_learner.py"/>
    <include path="utils.py"/>
    <include path="split_logic.py"/>
    <include path="training_pipeline.py"/>
    <include path="test_pipeline.py"/>
    <include path="metrics_logic.py"/>
    <include path="plot_logic.py"/>
    <include path="report_utils.py"/>
    <include path="feature_help_modal.py"/>
  </required_files>

  <stdio>
    <exit_code range="137" level="fatal_oom" description="Out of Memory"/>
    <exit_code range="1:" level="fatal" description="Tool failed — see Tool Standard Error"/>
  </stdio>

  <command detect_errors="exit_code"><![CDATA[
#import re

#set $image_zip_paths = []
#if $use_images_conditional.use_images == "yes"
  #for $zip_file in $use_images_conditional.images_zip_repeat
    #set $image_zip_paths = $image_zip_paths + [$zip_file.images_zip]
  #end for
#end if
#if len($image_zip_paths) > 0
  #set $images_zip_cli = " ".join(["'%s'" % z for z in $image_zip_paths])
#else
  #set $images_zip_cli = None
#end if

set -e;
ln -sf '$input_csv' 'train_input.csv';
#if $test_dataset_conditional.has_test_dataset == "yes"
ln -sf '$test_dataset_conditional.input_test' 'test_input.csv';
#end if

python '$__tool_directory__/multimodal_learner.py'
  --input_csv_train 'train_input.csv'
  #if $test_dataset_conditional.has_test_dataset == "yes"
  --input_csv_test 'test_input.csv'
  #end if
  --target_column '$target_column'

  #if $use_images_conditional.use_images == "yes"
    #if $images_zip_cli
    --images_zip $images_zip_cli
    #end if
    --missing_image_strategy '$use_images_conditional.missing_image_strategy'
    #if $use_images_conditional.backbone_image
    --backbone_image '$use_images_conditional.backbone_image'
    #end if
  #end if

  #if $backbone_text not in ("", None)
  --backbone_text '$backbone_text'
  #end if

  --preset '$preset'
  --eval_metric '$eval_metric'

  --random_seed '$random_seed'
  #if $time_limit
  --time_limit $time_limit
  #end if
  #if $deterministic == "true"
  --deterministic
  #end if

  #if $customize_defaults_conditional.customize_defaults == "yes"
    #if $customize_defaults_conditional.validation_size not in ("", None)
      --validation_size $customize_defaults_conditional.validation_size
    #end if
    #if $customize_defaults_conditional.split_probabilities and str($customize_defaults_conditional.split_probabilities).strip()
      --split_probabilities #echo " ".join([str(float(x)) for x in str($customize_defaults_conditional.split_probabilities).replace(",", " ").split() if x.strip()]) #
    #end if
    #if $customize_defaults_conditional.cross_validation == "true"
      --cross_validation true
      --num_folds $customize_defaults_conditional.num_folds
    #end if
    #if $customize_defaults_conditional.epochs
      --epochs $customize_defaults_conditional.epochs
    #end if
    #if $customize_defaults_conditional.learning_rate
      --learning_rate $customize_defaults_conditional.learning_rate
    #end if
    #if $customize_defaults_conditional.batch_size
      --batch_size $customize_defaults_conditional.batch_size
    #end if
    #if $customize_defaults_conditional.threshold
      --threshold $customize_defaults_conditional.threshold
    #end if
    #if $customize_defaults_conditional.hyperparameters
      --hyperparameters '$customize_defaults_conditional.hyperparameters'
    #end if
  #end if

  --output_json '$output_json'
  --output_html '$output_html'
  --output_config '$output_config'
]]></command>

  <inputs>
    <param name="input_csv" type="data" format="csv,tsv" label="Training dataset (CSV/TSV)" help="Must contain the target column and optional image paths"/>
    <param name="target_column" type="data_column" data_ref="input_csv" numerical="false" use_header_names="true" label="Target / Label column"/>

    <conditional name="test_dataset_conditional">
      <param name="has_test_dataset" type="boolean" truevalue="yes" falsevalue="no" checked="false" label="Provide separate test dataset?"/>
      <when value="yes">
        <param name="input_test" type="data" format="csv,tsv" optional="true" label="Test dataset (CSV/TSV)"/>
      </when>
      <when value="no"/>
    </conditional>

    <param name="backbone_text" type="select" label="Text backbone" optional="true">
      <option value="microsoft/deberta-v3-base" selected="true">microsoft/deberta-v3-base</option>
      <option value="microsoft/deberta-v3-small">microsoft/deberta-v3-small</option>
      <option value="google/electra-base-discriminator">google/electra-base-discriminator</option>
      <option value="google/electra-small-discriminator">google/electra-small-discriminator</option>
      <option value="roberta-base">roberta-base</option>
      <option value="bert-base-uncased">bert-base-uncased</option>
      <option value="distilroberta-base">distilroberta-base</option>
      <option value="albert-base-v2">albert-base-v2</option>
    </param>

    <conditional name="use_images_conditional">
      <param name="use_images" type="boolean" truevalue="yes" falsevalue="no" checked="false" label="Use image modality?"/>
      <when value="yes">
        <repeat name="images_zip_repeat" title="Image archive(s)" min="1">
          <param name="images_zip" type="data" format="zip" label="ZIP file containing images"/>
        </repeat>
        <param name="backbone_image" type="select" label="Image backbone" optional="true">
          <option value='swin_base_patch4_window7_224' selected='true'>swin_base_patch4_window7_224</option>
          <option value='swin_large_patch4_window12_384.in22k_ft_in1k'>swin_large_patch4_window12_384.in22k_ft_in1k</option>
          <option value='swin_small_patch4_window7_224'>swin_small_patch4_window7_224</option>
          <option value='swin_tiny_patch4_window7_224'>swin_tiny_patch4_window7_224</option>
          <option value='caformer_b36.in21k_ft_in1k'>caformer_b36.in21k_ft_in1k</option>
          <option value='caformer_m36.in21k_ft_in1k'>caformer_m36.in21k_ft_in1k</option>
          <option value='caformer_s36.in21k_ft_in1k'>caformer_s36.in21k_ft_in1k</option>
          <option value='caformer_s18.in1k'>caformer_s18.in1k</option>
          <option value='caformer_b36.sail_in22k_ft_in1k'>caformer_b36.sail_in22k_ft_in1k</option>
          <option value='caformer_m36.sail_in22k_ft_in1k'>caformer_m36.sail_in22k_ft_in1k</option>
          <option value='caformer_s36.sail_in22k_ft_in1k'>caformer_s36.sail_in22k_ft_in1k</option>
          <option value='vit_base_patch16_224'>vit_base_patch16_224</option>
          <option value='vit_large_patch14_224'>vit_large_patch14_224</option>
          <option value='convnext_base'>convnext_base</option>
          <option value='eva02_base_patch14_448.mim_in22k_ft_in22k_in1k'>eva02_base_patch14_448.mim_in22k_ft_in22k_in1k</option>
          <option value='resnet50'>resnet50</option>
        </param>
        <param name="missing_image_strategy" type="boolean" truevalue="true" falsevalue="false" checked="false"
               label="Drop rows with missing images?" help="True = drop, False = replace with placeholder (default)"/>
      </when>
      <when value="no"/>
    </conditional>

    <param name="preset" type="select" label="Quality preset">
      <option value="medium_quality" selected="true">Medium quality (fast)</option>
      <option value="high_quality">High quality</option>
      <option value="best_quality">Best quality (slowest)</option>
    </param>

    <param name="eval_metric" type="select" label="Primary evaluation metric">
      <option value="auto" selected="true">Auto (let AutoGluon choose)</option>
      <option value="roc_auc">ROC AUC</option>
      <option value="accuracy">Accuracy</option>
      <option value="balanced_accuracy">Balanced Accuracy</option>
      <option value="f1">F1</option>
      <option value="f1_macro">F1 Macro</option>
      <option value="f1_micro">F1 Micro</option>
      <option value="f1_weighted">F1 Weighted</option>
      <option value="precision">Precision</option>
      <option value="precision_macro">Precision Macro</option>
      <option value="precision_micro">Precision Micro</option>
      <option value="precision_weighted">Precision Weighted</option>
      <option value="recall">Recall</option>
      <option value="recall_macro">Recall Macro</option>
      <option value="recall_micro">Recall Micro</option>
      <option value="recall_weighted">Recall Weighted</option>
      <option value="average_precision">Average Precision</option>
      <option value="roc_auc_ovo_macro">ROC AUC OVO Macro</option>
      <option value="roc_auc_ovo_weighted">ROC AUC OVO Weighted</option>
      <option value="roc_auc_ovr_macro">ROC AUC OVR Macro</option>
      <option value="roc_auc_ovr_weighted">ROC AUC OVR Weighted</option>
      <option value="log_loss">Log Loss</option>
      <option value="mse">MSE</option>
      <option value="rmse">RMSE</option>
      <option value="mae">MAE</option>
      <option value="msle">MSLE</option>
      <option value="r2">R2</option>
    </param>

    <param name="random_seed" type="integer" value="42" label="Random seed"/>

    <param name="time_limit" type="integer" optional="true" min="60" label="Time limit (seconds)" help="Total training time budget. Recommended: 3600+ for real runs"/>
    <param name="deterministic" type="boolean" truevalue="true" falsevalue="false" checked="false"
           label="Enable deterministic mode" help="Use deterministic algorithms and CuDNN settings to reduce run-to-run variance (may slow training)"/>

    <conditional name="customize_defaults_conditional">
      <param name="customize_defaults" type="boolean" truevalue="yes" falsevalue="no" checked="false" label="Advanced: customize training settings?"/>
      <when value="yes">
        <param name="validation_size" type="float" value="0.2" label="Validation fraction (when test set provided)"/>
        <param name="split_probabilities" type="text" value="0.7 0.1 0.2" label="Train / Val / Test split (space-separated) when no test set"/>
        <param name="cross_validation" type="boolean" truevalue="true" falsevalue="false" checked="false" label="Enable k-fold cross-validation"/>
        <param name="num_folds" type="integer" value="5" label="Number of CV folds"/>
        <param name="epochs" type="integer" optional="true" label="Max epochs"/>
        <param name="learning_rate" type="float" optional="true" label="Learning rate"/>
        <param name="batch_size" type="integer" optional="true" label="Batch size"/>
        <param name="threshold" type="float" optional="true" min="0" max="1" label="Binary classification threshold"/>
        <param name="hyperparameters" type="text" optional="true" label="Extra AutoGluon hyperparameters (JSON or YAML string)"/>
      </when>
      <when value="no"/>
    </conditional>
  </inputs>

  <outputs>
    <data name="output_html" format="html" label="Multimodal Learner analysis report on data ${input_csv.name}"/>
    <data name="output_config" format="yaml" label="Multimodal Learner training config on data ${input_csv.name}"/>
    <data name="output_json" format="json" label="Multimodal Learner metric results on data ${input_csv.name}"/>
  </outputs>

  <tests>
    <!-- Basic run with images + external test set -->
    <test expect_num_outputs="3">
      <param name="input_csv" value="train.csv"/>
      <param name="target_column" value="7"/>
      <param name="test_dataset_conditional|has_test_dataset" value="yes"/>
      <param name="test_dataset_conditional|input_test" value="test.csv"/>
      <param name="use_images_conditional|use_images" value="yes"/>
      <param name="use_images_conditional|images_zip_repeat_0|images_zip" value="images.zip"/>
      <param name="use_images_conditional|backbone_image" value="resnet50"/>
      <param name="backbone_text" value="google/electra-base-discriminator"/>
      <output name="output_html">
        <assert_contents>
          <has_text text="Model Performance Summary"/>
        </assert_contents>
      </output>
    </test>

    <!-- Custom threshold -->
    <test expect_num_outputs="3">
      <param name="input_csv" value="train.csv"/>
      <param name="target_column" value="7"/>
      <param name="test_dataset_conditional|has_test_dataset" value="yes"/>
      <param name="test_dataset_conditional|input_test" value="test.csv"/>
      <param name="use_images_conditional|use_images" value="yes"/>
      <param name="use_images_conditional|images_zip_repeat_0|images_zip" value="images.zip"/>
      <param name="customize_defaults_conditional|customize_defaults" value="yes"/>
      <param name="customize_defaults_conditional|threshold" value="0.4"/>
      <output name="output_json">
        <assert_contents>
          <has_text text="&quot;threshold&quot;: 0.4"/>
        </assert_contents>
      </output>
    </test>

    <!-- No external test set; internal split -->
    <test expect_num_outputs="3">
      <param name="input_csv" value="train.csv"/>
      <param name="target_column" value="7"/>
      <param name="test_dataset_conditional|has_test_dataset" value="no"/>
      <param name="use_images_conditional|use_images" value="yes"/>
      <param name="use_images_conditional|images_zip_repeat_0|images_zip" value="images.zip"/>
      <output name="output_json">
        <assert_contents>
          <has_text text="&quot;val&quot;"/>
        </assert_contents>
      </output>
    </test>

    <!-- Text/tabular only (ignore images) -->
    <test expect_num_outputs="3">
      <param name="input_csv" value="train.csv"/>
      <param name="target_column" value="7"/>
      <param name="test_dataset_conditional|has_test_dataset" value="yes"/>
      <param name="test_dataset_conditional|input_test" value="test.csv"/>
      <param name="use_images_conditional|use_images" value="no"/>
      <output name="output_html">
        <assert_contents>
          <has_text text="Train and Validation Performance Summary"/>
        </assert_contents>
      </output>
    </test>
  </tests>

  <help><![CDATA[
**AutoGluon Multimodal Learner**

Trains a powerful multimodal model combining tabular features, images, and text using AutoGluon-Multimodal.

- Handles missing images intelligently
- Supports cross-validation
- Produces detailed HTML reports and transparent metrics
- Fully reproducible

Ideal for medical imaging + clinical data, product images + descriptions, etc.
]]></help>

  <citations>
    <citation type="bibtex">
@article{AutoGluon2023,
  author = {Erickson, Nick and Mueller, Jonas and Wang, Yizhou and others},
  title = {AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data},
  journal = {arXiv preprint arXiv:2003.06505},
  year = {2023}
}
    </citation>
  </citations>
</tool>