FiLM框架运行CLEVR数据集

1.搭建虚拟环境

virtualenv -p python3 .env       # Create virtual environment
source .env/bin/activate         # Activate virtual environment
pip install -r requirements.txt  # Install dependencies
echo $PWD > .env/lib/python3.5/site-packages/iep.pth # Add this package to virtual environment
# Work for a while ...
deactivate # Exit virtual environment

---

2.数据集的预处理

CLEVR数据集

（1）下载解压
（2）提取图像特征
提取ResNet-101特征 【train\val\test类似】

python scripts/extract_features.py \
  --input_image_dir data/CLEVR_v1.0/images/train \
  --output_h5_file data/train_features.h5

h5是什么格式的文件？
（3）预处理question【train\val\test类似】

python scripts/preprocess_questions.py \
  --input_questions_json data/CLEVR_v1.0/questions/CLEVR_train_questions.json \
  --output_h5_file data/train_questions.h5 \
  --output_vocab_json data/vocab.json

vocab.json这个词汇表stores the mapping between tokens and indices for questions and programs。训练的时候生成vocab.json，可以重复利用在val\test那里用作输入

python scripts/preprocess_questions.py \
  --input_questions_json data/CLEVR_v1.0/questions/CLEVR_val_questions.json \
  --output_h5_file data/val_questions.h5 \
  --input_vocab_json data/vocab.json

---

3.训练
训练细节和“ Program Generator + Execution Engine model”高度一致，虽然我们是单步训练，他们是三步训练。
三步训练：包括PG、EE、PG+EE
（1）训练Program Generator

python scripts/train_model.py \
  --model_type PG \
  --num_train_samples 18000 \   # a small number of ground-truth programs
  --num_iterations 20000 \
  --checkpoint_every 1000 \
  --checkpoint_path data/program_generator.pt

（2）训练 Execution Engine

python scripts/train_model.py \
  --model_type EE \
  --program_generator_start_from data/program_generator.py \   #用的是PG预测生成的输出
  --num_iterations 100000 \
  --checkpoint_path data/execution_engine.pt

（3）联合训练PG+EE【无gt联合微调】

python scripts/train_model.py \
  --model_type PG+EE \
  --program_generator_start_from data/program_generator.pt \
  --execution_engine_start_from data/execution_engine.pt \
  --checkpoint_path data/joint_pg_ee.pt

---

FiLM的单步训练，直接运行下面的语句就开始训练了。

sh scripts/train/film.sh

生成film.pt文件（模型），film.log文件，film.pt.json文件

---

4.测试

python scripts/run_model.py \
  --program_generator data/program_generator.pt \
  --execution_engine data/execution_engine.pt \
  --input_question_h5 data/val_questions.h5 \
  --input_features_h5 data/val_features.h5

联合微调版

python scripts/run_model.py \
  --program_generator data/joint_pg_ee.pt \
  --execution_engine data/joint_pg_ee.pt \
  --input_question_h5 data/val_questions.h5 \
  --input_features_h5 data/val_features.h5