This repository contains the code used for the paper:
Deep Neural Networks Regularization for Structured Output Prediction, S.Belharbi, R.Hérault, C.Chatelain, S.Adam. 2017. ArXiv.
Implementation of deep convolutional auto-encoder using transposed convolution "deconvolution".
Please check the code in the file: debug_deep_conv_autoenc.py. This is the architecture we used:
# Deep conv.ae
# configure encoder
encode_cae_l0 = {"type": "conv",
"rng": rnd,
"filter_shape": (16, 3, 3),
"activation": "relu",
"padding": (1, 1),
"W": None,
"b": None,
"b_v": 0.,
"stride": (1, 1)}
encode_cae_l1 = {"type": "downsample",
"poolsize": (2, 2),
"ignore_border": False,
"mode": "max"}
# Dim: 16 * 25x25.
encode_cae_l2 = {"type": "conv",
"rng": rnd,
"filter_shape": (8, 3, 3),
"activation": "relu",
"padding": (1, 1),
"W": None,
"b": None,
"b_v": 0.,
"stride": (1, 1)}
encode_cae_l3 = {"type": "downsample",
"poolsize": (2, 2),
"ignore_border": True,
"mode": "max"}
# Dim: 8 * 12x12.
encoder_config = [encode_cae_l0, encode_cae_l1, encode_cae_l2, encode_cae_l3]
# configure decoder
decode_cae_l0 = {"type": "conv",
"rng": rnd,
"filter_shape": (8, 3, 3),
"activation": "relu",
"padding": (1, 1),
"W": None,
"b": None,
"b_v": 0.,
"stride": (1, 1)}
decode_cae_l1 = {"type": "upsample",
"ratio": 2,
"use_1D_kernel": False}
# Dim: 8 * 24x24
decode_cae_l2 = {"type": "conv",
"rng": rnd,
"filter_shape": (16, 3, 3),
"activation": "relu",
"padding": (1, 1),
"W": None,
"b": None,
"b_v": 0.,
"stride": (1, 1)}
decode_cae_l3 = {"type": "upsample",
"ratio": 2,
"use_1D_kernel": False}
# Dim: 16 * 48x48
decode_cae_l4 = {"type": "conv",
"rng": rnd,
"filter_shape": (1, 3, 3),
"activation": "sigmoid",
"padding": (2, 2),
"W": None,
"b": None,
"b_v": 0.,
"stride": (1, 1)}
# Dim: 1 * 50x50
decoder_config = [decode_cae_l0, decode_cae_l1, decode_cae_l2, decode_cae_l3,
decode_cae_l4]Please cite this paper if you use the code in this repository as part of a published research project.
This code is a Theano based implementation (Python 2.7). To run it, you must install Theano and its dependencies (numpy, matplotlib, ...). Octave and oct2py are needed for the evaluation.
Before running the code, you need to download the data by running the script sop_embed/download_face_data.sh.
To run the experiments mentioned in the paper, see the files in the folder: sop_embed/experiments/*.py.