The Submission for SIMMC 2.0 Challenge 2021

• challenge website

Requirements

• python 3.8.8
• pytorch 1.8.1
• transformers 4.8.2
• apex for multi-gpu
• nltk

Preprocessing

1. Download Data

• Download the data provided by the challenge organizer and put it in the data folder.
• Unzip data files

2. Image saving

• Preprocess the image files in advance. The preprocessed result has the image name as the key and visual as the value.

python3 image_preprocessor.py
python3 image_preprocessor_final.py

• The result(.pickle) is saved in res folder.

Step 1 (ITM)

First, the model is post-trained by image-to-text matching. Here, image is each object and text is the visual metadata of the object. Code is provided in the ITM folder.

Step 2 (BTM)

Second, pretraining is performed to use background reprsentation of image in subtasks. Similar to ITM, it is trained to match image and text, and the image is the background of the dialog and the text is the entire context of the dialog. Code is provided in the BTM folder.

Step 3

This is the learning process for each subtask. You can train the model in each folder (sub1, sub2_1, sub2_2, sub2_3, sub2_4, sub4).

Model

All models can be downloaded from the following link

model.pt is a model for evaluating devtest, and the result is saved in the dstc10-simmc-entry folder. model_final.pt is a model for evaluating teststd, and the result is saved in the dstc10-simmc-final-entry folder. However, the training of the model was not completed within the challenge period, so we inferred to model.pt for the teststd data in subtask2.

Evlauation

Using the evaluation script suggested by the challenge organizer

• Evaluation script for subtask1
• Evaluation script for subtask2
• Evaluation script for subtask4-generation

The SIMMC organizers introduce the scripts:

     
      
$ python tools/disambiguator_evaluation.py \
	--pred_file="{PATH_TO_PRED_FILE}" \
	--test_file="{PATH_TO_TEST_FILE}" \


      
       
(line-by-line evaluation)
$ python -m gpt2_dst.scripts.evaluate \
  --input_path_target={PATH_TO_GROUNDTRUTH_TARGET} \
  --input_path_predicted={PATH_TO_MODEL_PREDICTIONS} \
  --output_path_report={PATH_TO_REPORT}

(Or, dialog level evaluation)
$ python -m utils.evaluate_dst \
    --input_path_target={PATH_TO_GROUNDTRUTH_TARGET} \
    --input_path_predicted={PATH_TO_MODEL_PREDICTIONS} \
    --output_path_report={PATH_TO_REPORT}
    

       
        
$ python tools/response_evaluation.py \
    --data_json_path={PATH_TO_GOLD_RESPONSES} \
    --model_response_path={PATH_TO_MODEL_RESPONSES} \
    --single_round_evaluation


        
         
$ python tools/retrieval_evaluation.py \
    --retrieval_json_path={PATH_TO_GROUNDTRUTH_RETRIEVAL} \
    --model_score_path={PATH_TO_MODEL_CANDIDATE_SCORES} \
    --single_round_evaluation    

        
       
      
     

DevTest Results

Subtask #1: Multimodal Disambiguation

Subtask #2: Multimodal Coreference Resolution

Subtask #3: Multimodal Dialog State Tracking

No Training/Testing

Subtask #4: Multimodal Dialog Response Generation

Generation

Retrieval

No Training/Testing