Pytorch implementation of FiLM: Visual Reasoning with a General Conditioning Layer

Requirements

• Python3
• Pytorch 1.0.0
• TensorBoardX

Differences from the original implementation

The original implementation used some of the pretrained layers in resnet, or four convolution layers with kernel size = 4 and stride = 2 when starting from scratch.

However, in this implementation, considering that the target is a Sort-of-CLEVR, I reduced the number of layers with stride = 2 to three and added two layers with stride = 1 to increase the size of the feature map.

Initial convolution layer configuration for this implementation is:

(Kernel size = 5, stride = 2, padding = 2)
(Kernel size = 3, stride = 2, padding = 1)
(Kernel size = 3, stride = 2, padding = 1)
(Kernel size = 3, stride = 1, padding = 1)
(Kernel size = 3, stride = 1, padding = 1)

Usage

generate sort-of-clevr dataset

python soc_generator.py

train

python train.py 
    --batch_size={64}
    --n_epoch={120}
    --lr={1e-4}
    --weight_decay={1e-4}
    --save_dir={model}
    --dataset={data/sort-of-clevr.pickle}
    --init={kaiming}
    --n_res={6}
    --seed={12345}
    --n_cpu={4}
    [--resume={}]

test

python test.py
    --n_res
    --dataset
    --model

visualize

python visualize.py
    --n_res
    --dataset
    --model
    --save_dir [features]

Example of a visualized feature map image

Result

Sort-of-CLEVR	n_res = 6
Accuracy	98%