Fashion-Gen: The Generative Fashion Dataset and Challenge 论文解读＆数据集介绍

论文解读

论文地址：https://arxiv.org/abs/1806.08317

数据集划分

数目	train	val	test
293, 008	260, 480	32, 528	32, 528

类别介绍

数据集中有48个主类，121个子类。
如下是训练集、测试集中类别占比

图片统计

如下是训练集中主类、子类的图片数目统计

文本描述

如下是文本描述长度的统计

如下是从文本中提取的颜色分布

chanllenge

• Generating high-resolution images using P-GANs
• Text-to-Image synthesis

评估方法

• Inception Score
• Human Evaluation（因为Inception Score没考虑文本图片之间的相关性）

数据集下载

看到FashionBERT论文里的数据集FashionGEN，想了解一下，但是官网上填了个表单就没信了，地址为：https://fashion-gen.com/于是又在网上找了相关内容，找到一个网址https://github.com/menardai/FashionGenAttnGAN

上面有3个文件（注：没有提供测试集，论文中说不会提供测试集，被集成在了论文的docker中）

• fashiongen_256_256_train.h5
• fashiongen_256_256_validation.h5
• fashiongen_consume_data_example.pdf

分析代码

参考https://docs.h5py.org/en/stable/quick.html用以下代码进行分析

import h5py
import numpy as np
BATCH_SIZE = 32

def get_batch(file_h5, features, batch_number, batch_size=32):
    """Get a batch of the dataset Args: file_h5(str): path of the dataset features(list(str)): list of names of features present in the dataset that should be returned. batch_number(int): the id of the batch to be returned. batch_size(int): the mini-batch size Returns: A list of numpy arrays of the requested features"""
    list_of_arrays = []
    lb, ub = batch_number * batch_size, (batch_number + 1) * batch_size
    for feature in features:
        list_of_arrays.append(file_h5[feature][lb: ub])
    return list_of_arrays

# open the file
# file_h5 = h5py.File('fashiongen_256_256_train.h5', mode='r')
file_h5 = h5py.File('fashiongen_256_256_validation.h5', mode='r')
# define the features to be retrieved
list_of_features = ['input_image', 'input_description']
dataset_len = len(file_h5['input_image'])  
nb_batches = int(dataset_len / BATCH_SIZE)
batch_nb = np.random.randint(0, nb_batches)
# get the first batch of the data
list_of_arrays = get_batch(file_h5, list_of_features, batch_nb, BATCH_SIZE)
# close the file
file_h5.close()

得到训练集数目260490、验证集数目32528
数据集是个类似dict的结构，keys分别为
[‘index’, ‘index_2’, ‘input_brand’, ‘input_category’, ‘input_composition’, ‘input_concat_description’, ‘input_department’, ‘input_description’, ‘input_gender’, ‘input_image’, ‘input_msrpUSD’, ‘input_name’, ‘input_pose’, ‘input_productID’, ‘input_season’, ‘input_subcategory’]

图片的维度为：

(256, 256, 3)

内容分析

以验证集为例，接下来一个一个分析内容

1. index

file_h5['index'].shape
# (32528, 1)
file_h5['index'][0:][0:]
# 输出以下
[[    24]
 [    25]
 [    26]
 ...
 [342153]
 [342154]
 [342155]]

2. index_2

file_h5['index_2'].shape
# (32528,)
file_h5['index_2'][0:]
# 输出以下
[    0     1     2 ... 32525 32526 32527]

3. input_brand

file_h5['input_brand'].shape
# (32528, 1)
file_h5['input_brand'][0:][0:]
# 输出以下
array([[b'Diesel'],
       [b'Diesel'],
       [b'Diesel'],
       ...,
       [b'Calvin Klein 205W39NYC'],
       [b'Calvin Klein 205W39NYC'],
       [b'Calvin Klein 205W39NYC']], dtype='|S100')

4. input_category

file_h5['input_category'].shape
# (32528, 1)
file_h5['input_category'][0:][0:]
# 输出以下
array([[b'JACKETS & COATS'],
       [b'JACKETS & COATS'],
       [b'JACKETS & COATS'],
       ...,
       [b'SHIRTS'],
       [b'SHIRTS'],
       [b'SHIRTS']], dtype='|S100')

5. input_composition

file_h5['input_composition'].shape
# (32528, 1)
file_h5['input_composition'][0:][0:]
# 输出以下
array([[b'90% cotton, 8% polyester, 2% elastane.'],
       [b'90% cotton, 8% polyester, 2% elastane.'],
       [b'90% cotton, 8% polyester, 2% elastane.'],
       ...,
       [b'100% cotton.'],
       [b'100% cotton.'],
       [b'100% cotton.']], dtype='|S200')

6. input_concat_description

file_h5['input_concat_description'].shape
# (32528, 1)
file_h5['input_concat_description'][0:][0:]
# 输出以下
array([[b'Denim-like jogg jacket in blue. Fading and whiskering throughout. Spread collar. Copper tone button closures at front. Flap pockets at chest with metallic logo plaque. Seam pockets at sides. Cinch tabs at back waistband. Single button sleeve cuffs. Tone on tone stitching. Semi-sheer t-shirt in heather white. Crewneck collar. Patch pocket at breast. Tonal stitching.'],
       [b'Denim-like jogg jacket in blue. Fading and whiskering throughout. Spread collar. Copper tone button closures at front. Flap pockets at chest with metallic logo plaque. Seam pockets at sides. Cinch tabs at back waistband. Single button sleeve cuffs. Tone on tone stitching. Semi-sheer t-shirt in heather white. Crewneck collar. Patch pocket at breast. Tonal stitching.'],
       [b'Denim-like jogg jacket in blue. Fading and whiskering throughout. Spread collar. Copper tone button closures at front. Flap pockets at chest with metallic logo plaque. Seam pockets at sides. Cinch tabs at back waistband. Single button sleeve cuffs. Tone on tone stitching. Semi-sheer t-shirt in heather white. Crewneck collar. Patch pocket at breast. Tonal stitching.'],
       ...,
       [b"Long sleeve cotton poplin shirt in 'optic' white. Spread collar featuring carved silver-tone hardware. Button closure at front. Single-button barrel cuffs. Tonal stitching. Slim-fit 'uniform' twill trousers in black. Mid-rise. Four-pocket styling. Central pleat at front and back legs. Grosgrain tape striped in blue and purple at outseams. Zip-fly. Partially lined. Tonal stitching. Long sleeve coated cotton-blend trench coat in beige. Notched lapel collar. Concealed button closure at front. Detachable pin-buckle belt and welt pockets at waist. Buttoned tab at central back vent and cuffs. Epaulets. Storm flap. Unlined. Tonal stitching."],
       [b"Long sleeve cotton poplin shirt in 'optic' white. Spread collar featuring carved silver-tone hardware. Button closure at front. Single-button barrel cuffs. Tonal stitching. Slim-fit 'uniform' twill trousers in black. Mid-rise. Four-pocket styling. Central pleat at front and back legs. Grosgrain tape striped in blue and purple at outseams. Zip-fly. Partially lined. Tonal stitching. Long sleeve coated cotton-blend trench coat in beige. Notched lapel collar. Concealed button closure at front. Detachable pin-buckle belt and welt pockets at waist. Buttoned tab at central back vent and cuffs. Epaulets. Storm flap. Unlined. Tonal stitching."],
       [b"Long sleeve cotton poplin shirt in 'optic' white. Spread collar featuring carved silver-tone hardware. Button closure at front. Single-button barrel cuffs. Tonal stitching. Slim-fit 'uniform' twill trousers in black. Mid-rise. Four-pocket styling. Central pleat at front and back legs. Grosgrain tape striped in blue and purple at outseams. Zip-fly. Partially lined. Tonal stitching. Long sleeve coated cotton-blend trench coat in beige. Notched lapel collar. Concealed button closure at front. Detachable pin-buckle belt and welt pockets at waist. Buttoned tab at central back vent and cuffs. Epaulets. Storm flap. Unlined. Tonal stitching."]],
      dtype='|S800')

7. input_department

file_h5['input_department'].shape
# (32528, 1)
file_h5['input_department'][0:][0:]
# 输出以下
array([[b'CLOTHING'],
       [b'CLOTHING'],
       [b'CLOTHING'],
       ...,
       [b'CLOTHING'],
       [b'CLOTHING'],
       [b'CLOTHING']], dtype='|S100')

8. input_description

file_h5['input_description'].shape
# (32528, 1)
file_h5['input_description'][0:][0:]
# 输出以下
array([[b'Denim-like jogg jacket in blue. Fading and whiskering throughout. Spread collar. Copper tone button closures at front. Flap pockets at chest with metallic logo plaque. Seam pockets at sides. Cinch tabs at back waistband. Single button sleeve cuffs. Tone on tone stitching.'],
       [b'Denim-like jogg jacket in blue. Fading and whiskering throughout. Spread collar. Copper tone button closures at front. Flap pockets at chest with metallic logo plaque. Seam pockets at sides. Cinch tabs at back waistband. Single button sleeve cuffs. Tone on tone stitching.'],
       [b'Denim-like jogg jacket in blue. Fading and whiskering throughout. Spread collar. Copper tone button closures at front. Flap pockets at chest with metallic logo plaque. Seam pockets at sides. Cinch tabs at back waistband. Single button sleeve cuffs. Tone on tone stitching.'],
       ...,
       [b"Long sleeve cotton poplin shirt in 'optic' white. Spread collar featuring carved silver-tone hardware. Button closure at front. Single-button barrel cuffs. Tonal stitching."],
       [b"Long sleeve cotton poplin shirt in 'optic' white. Spread collar featuring carved silver-tone hardware. Button closure at front. Single-button barrel cuffs. Tonal stitching."],
       [b"Long sleeve cotton poplin shirt in 'optic' white. Spread collar featuring carved silver-tone hardware. Button closure at front. Single-button barrel cuffs. Tonal stitching."]],
      dtype='|S400')

9. input_gender

file_h5['input_gender'].shape
# (32528, 1)
file_h5['input_gender'][0:][0:]
# 输出以下
array([[b'Men'],
       [b'Men'],
       [b'Men'],
       ...,
       [b'Men'],
       [b'Men'],
       [b'Men']], dtype='|S30')

10. input_image

file_h5['input_image'].shape
# (32528, 256, 256, 3)
file_h5['input_image'][0].shape
# (256, 256, 3)

11. input_msrpUSD

file_h5['input_msrpUSD'].shape
(32528, 1)
file_h5['input_msrpUSD'][0:][0:]
# 输出以下
array([[335.],
       [335.],
       [335.],
       ...,
       [990.],
       [990.],
       [990.]], dtype=float32)

12. input_name

file_h5['input_name'].shape
# (32528, 1)
file_h5['input_name'][0:][0:]
# 输出以下
array([[b'Blue Faded Elshar Jogg Jacket'],
       [b'Blue Faded Elshar Jogg Jacket'],
       [b'Blue Faded Elshar Jogg Jacket'],
       ...,
       [b'White Pointed Collar Shirt'],
       [b'White Pointed Collar Shirt'],
       [b'White Pointed Collar Shirt']], dtype='|S100')

13. ‘input_pose

file_h5['input_pose'].shape
# (32528, 1)
file_h5['input_pose'][0:][0:]
# 输出以下
array([[b'id_gridfs_1'],
       [b'id_gridfs_2'],
       [b'id_gridfs_3'],
       ...,
       [b'id_gridfs_3'],
       [b'id_gridfs_4'],
       [b'id_gridfs_5']], dtype='|S40')

14. input_productID

file_h5['input_productID'].shape
# (32528, 1)
file_h5['input_productID'][0:][0:]
# 输出以下
array([[  86605],
       [  86605],
       [  86605],
       ...,
       [2938688],
       [2938688],
       [2938688]], dtype=int32)

15. input_season

file_h5['input_season'].shape
(32528, 1)
file_h5['input_season'][0:][0:]
# 输出以下
array([[b'SS2014'],
       [b'SS2014'],
       [b'SS2014'],
       ...,
       [b'SS2018'],
       [b'SS2018'],
       [b'SS2018']], dtype='|S10')

16. input_subcategory

file_h5['input_subcategory'].shape
# (32528, 1)
file_h5['input_subcategory'][0:][0:]
# 输出以下
array([[b'DENIM JACKETS'],
       [b'DENIM JACKETS'],
       [b'DENIM JACKETS'],
       ...,
       [b'SHIRTS'],
       [b'SHIRTS'],
       [b'SHIRTS']], dtype='|S100')

可视化数据

keys	1	2	3
index	24，25，26，27	73，74，75，76	93，94，95，96
index_2	0, 1, 2, 3	4, 5, 6, 7	8, 9, 10, 11
input_brand	b’Diesel’	b’Dsquared2’	b’Diesel Black Gold’
input_category	b’JACKETS & COATS’	b’JEANS’	b’JACKETS & COATS’
input_composition	b’90% cotton, 8% polyester, 2% elastane.’	b’98% cotton, 2% elastane.’	b’Body: 100% lambskin. Contrast: 100% goatskin. Lining: 51% cotton, 49% rayon. Sleeve lining: 54% acetate, 46% polyester.’
input_concat_description	b’Denim-like jogg jacket in blue. Fading and whiskering throughout. Spread collar. Copper tone button closures at front. Flap pockets at chest with metallic logo plaque. Seam pockets at sides. Cinch tabs at back waistband. Single button sleeve cuffs. Tone on tone stitching. Semi-sheer t-shirt in heather white. Crewneck collar. Patch pocket at breast. Tonal stitching.’	b’Slim-fit jeans in light grey. Distressing and fading throughout. Seven-pocket styling. Textured black leather logo patch at back waist. Tonal stitching. Red logo tab at button-fly. Short sleeve t-shirt in deep slate blue. Crewneck collar. Tonal stitching.’	b’Long sleeve suede jacket in black. Tonal grained leather paneling throughout. Stand collar. Zip closure and zippered welt pockets at front. Zippered vents at back hem. Welt pockets at interior. Fully lined. Tonal stitching. Zippered expansion panels at sleeve cuffs. Long sleeve coated denim shirt in indigo blue. Irregular overdye effect throughout in black. Spread collar. Flap pockets at breast, one with metallic logo piece. Press-stud closure at front. Tonal stitching. Three press-studs at barrel cuffs. Slim-fit cropped pleated wool trousers in black. Four-pocket styling. Tonal stitching. Button-fly.’
input_department	b’CLOTHING’	b’CLOTHING’	b’CLOTHING’
input_description	b’Denim-like jogg jacket in blue. Fading and whiskering throughout. Spread collar. Copper tone button closures at front. Flap pockets at chest with metallic logo plaque. Seam pockets at sides. Cinch tabs at back waistband. Single button sleeve cuffs. Tone on tone stitching.’	b’Slim-fit jeans in light grey. Distressing and fading throughout. Seven-pocket styling. Textured black leather logo patch at back waist. Tonal stitching. Red logo tab at button-fly.’	b’Long sleeve suede jacket in black. Tonal grained leather paneling throughout. Stand collar. Zip closure and zippered welt pockets at front. Zippered vents at back hem. Welt pockets at interior. Fully lined. Tonal stitching. Zippered expansion panels at sleeve cuffs.’
input_gender	b’Men’	b’Men’	b’Men’
input_image1
input_image2
input_image3
input_image4
input_msrpUSD	335.	630.	1215.
input_pose	b’id_gridfs_1’，b’id_gridfs_2’，b’id_gridfs_3’，b’id_gridfs_4’	b’id_gridfs_1’，b’id_gridfs_2’，b’id_gridfs_3’，b’id_gridfs_4’	b’id_gridfs_1’，b’id_gridfs_2’，b’id_gridfs_3’，b’id_gridfs_4’
input_productID	86605	86773	86711
input_season	b’SS2014’	b’SS2014’	b’SS2014’
input_subcategory	b’DENIM JACKETS’	b’JEANS’	b’LEATHER JACKETS’

后续数据集内容待更新…如有问题敬请指教哦～