523 lines
23 KiB
Python
523 lines
23 KiB
Python
import importlib
|
|
import pdb
|
|
import os
|
|
import numpy as np
|
|
import concurrent.futures
|
|
import json
|
|
import requests
|
|
import pandas as pd
|
|
import config as cfg
|
|
import shutil
|
|
import re
|
|
|
|
class FindingApi:
|
|
'''
|
|
Methods for accessing eBay's FindingApi services
|
|
'''
|
|
|
|
def __init__(self, service, pageNumber):
|
|
self.service = [
|
|
'findItemsAdvanced', 'findCompletedItems',
|
|
'findItemsByKeywords', 'findItemsIneBayStores', 'findItemsByCategory',
|
|
'findItemsByProduct'
|
|
][service] # Currently using only index 4, i.e., service = 4
|
|
self.pageNumber = list(range(1, pageNumber)) # 77 pgs will give equal weights to cats given call constraints
|
|
|
|
# examples of additional params you may want to add:
|
|
# 'itemFilter(0).value':'Used' consider using this with findCompletedItems call
|
|
# 'itemFilter(1).name':'ListingType'
|
|
# 'itemFilter(1).value':'AuctionWithBIN'
|
|
|
|
def get_data(self, category_id, i):
|
|
|
|
'''
|
|
Gets raw JSON data fom FindingApi service call. Currently being used to
|
|
get itemIDs from categories;
|
|
'''
|
|
|
|
'''
|
|
consider using the sortOrder param to update by the latest listings first.
|
|
Also consider using the exlude duplicates param
|
|
'''
|
|
params = {
|
|
"OPERATION-NAME":self.service,
|
|
"SECURITY-APPNAME":cfg.sec['SECURITY-APPNAME'],
|
|
"SERVICE-VERSION":"1.13.0",
|
|
"RESPONSE-DATA-FORMAT":"JSON",
|
|
"categoryId":category_id,
|
|
"paginationInput.entriesPerPage":"100",
|
|
"paginationInput.PageNumber":i,
|
|
"itemFilter(0).name":"Condition",
|
|
"itemFilter(0).value":"Used"
|
|
}
|
|
|
|
# TODO add try excepts here
|
|
try:
|
|
response = requests.get("https://svcs.ebay.com/services/search/FindingService/v1",
|
|
params=params, timeout=4)
|
|
response.raise_for_status()
|
|
|
|
except requests.exceptions.RequestException:
|
|
print('connection error') #TODO DECIDE HOW TO HANDLE EXCEPTION
|
|
data = response.json()
|
|
return data
|
|
|
|
# TODO add some other options to finding call api such as for possibly filtering for used items only. This might give you a better dataset for training. Or maybe a mixture of new and used. Maybe
|
|
# try and come up with a way to mathematically determine your odds of maximizing the number of pictures in your training set while reducing the number of useless images. Say for example, if you took a
|
|
# random set of 3 of 8 pictures total from each listing you might have a better chance of getting 3 good pictures in addition to increasing your training set. Or maybe you would have better luck with limiting
|
|
# it to the first 5 pictures instead of random.
|
|
|
|
# You may even have more consistency with used shoes since they are "one-off" items without confusing multiple variations and colors. What else you can do is run small training sets on both new and used
|
|
# to see which one is more accurate or if a combo of both is more accurate.
|
|
|
|
def get_ids_from_cats(self):
|
|
'''
|
|
Creates a 20-itemId list to use for the ShoppingApi
|
|
call
|
|
'''
|
|
pages = self.pageNumber
|
|
itemid_results_list = []
|
|
|
|
with open('cat_list.txt') as jf:
|
|
cat_list = json.load(jf)
|
|
|
|
for category_id in cat_list:
|
|
|
|
args = [(category_id, i) for i in pages] # NOTE alternatively you can use args.extend(args) to create master list of tuples with all cats
|
|
# instead of running concurrent.futures.ThreadPoolExecutor in a loop. Might be faster
|
|
|
|
with concurrent.futures.ThreadPoolExecutor() as executor:
|
|
for future in executor.map(lambda p: self.get_data(*p), args):
|
|
data = future
|
|
|
|
'''
|
|
These try excepts may be unnecessary.
|
|
'''
|
|
try: # TODO if conditionals are not working due to each thread checking the same unedited item_id_results list
|
|
training = pd.read_csv('training.csv')
|
|
for item in data['findItemsByCategoryResponse'][0]['searchResult'][0]['item']:
|
|
if (item not in training.values) and (item not in itemid_results_list): # might not be required
|
|
itemid_results_list.append(item['itemId'][0])
|
|
|
|
except (pd.errors.EmptyDataError, FileNotFoundError):
|
|
for item in data['findItemsByCategoryResponse'][0]['searchResult'][0]['item']:
|
|
if item not in itemid_results_list:
|
|
itemid_results_list.append(item['itemId'][0])
|
|
|
|
item_id_results = list(set(itemid_results_list))
|
|
item_id_results = [','.join(itemid_results_list[n:n+20]) for n in list(range(0,
|
|
len(itemid_results_list), 20))] # 20-ItemID list created to maximize dataset/decrease calls given call constraints
|
|
return item_id_results
|
|
|
|
# TODO during your try except conditionals just check the csv files. At the end you can create sets. You can creat another condition that says if the final set is smaller than 100k then you can call finding
|
|
# service on more pages (but only pages you haven't tried) and repeat the search process.
|
|
|
|
# TODO instead of running through multiple try except loops try to implement set methods for efficiency and ease. Remember symmetric_difference, difference, intersection, set()
|
|
# for category_id in cat_list:
|
|
|
|
class ShoppingApi:
|
|
'''
|
|
Creates objects from ShoppingApi service calls that can interact with
|
|
pandas dataframes
|
|
'''
|
|
|
|
def update_cats(self):
|
|
'''
|
|
Updates cat_list.txt
|
|
'''
|
|
|
|
parent_cats = ['3034', '93427'] # Women's and Men's shoe departments
|
|
cat_list = []
|
|
|
|
for department in parent_cats:
|
|
|
|
params = {
|
|
"callname":"GetCategoryInfo",
|
|
"appid":cfg.sec['SECURITY-APPNAME'],
|
|
"version":"671",
|
|
"responseencoding":"JSON",
|
|
"CategoryID":department,
|
|
"IncludeSelector":"ChildCategories",
|
|
}
|
|
|
|
try:
|
|
response = requests.get("https://open.api.ebay.com/shopping?", params=params, timeout=4)
|
|
response.raise_for_status()
|
|
|
|
except requests.exceptions.RequestException:
|
|
print('connection error')
|
|
|
|
response = response.json()
|
|
response = response['CategoryArray']['Category'][1:] # excludes index
|
|
# 0 as this is parent node, i.e., women's or men's dept.
|
|
|
|
temp_cat_list = [cat['CategoryID'] for cat in response]
|
|
cat_list.extend(temp_cat_list)
|
|
|
|
with open('cat_list.txt', 'w') as f:
|
|
json.dump(cat_list, f)
|
|
|
|
# leaf_list = [node['LeafCategory'] for node in response]
|
|
|
|
def get_item_from_findItemsByCategory(self, twenty_id):
|
|
'''
|
|
Gets raw JSON data from multiple live listings given multiple itemIds
|
|
'''
|
|
params = {
|
|
"callname":"GetMultipleItems",
|
|
"appid":cfg.sec['SECURITY-APPNAME'], # TODO check ebay.yaml for deprication. Might be why you're getting connection errors
|
|
"version":"671",
|
|
"responseencoding":"JSON",
|
|
"ItemID":twenty_id,
|
|
"IncludeSelector":"ItemSpecifics",
|
|
}
|
|
|
|
try:
|
|
response = requests.get("https://open.api.ebay.com/shopping?", params=params, timeout=4)
|
|
response.raise_for_status()
|
|
|
|
except requests.exceptions.RequestException: # TODO need better handling
|
|
print('connection error')
|
|
|
|
response = response.json()
|
|
response = response['Item']
|
|
|
|
return response
|
|
|
|
def conky(self):
|
|
'''
|
|
Runs get_item_from_findItemsByCategory in multiple threads to get relevant
|
|
data for creating training sets
|
|
'''
|
|
try:
|
|
with open('raw_data.txt') as f:
|
|
data = json.load(f)
|
|
except (FileNotFoundError, ValueError):
|
|
data = []
|
|
|
|
service_dict = {
|
|
0: 'findItemsAdvanced', 1: 'findCompletedItems',
|
|
2: 'findItemsByKeywords', 3: 'findItemsIneBayStores',
|
|
4: 'findItemsByCategory', 5:'findItemsByProduct'}
|
|
service_dict
|
|
|
|
fnd_srvc = input(str(service_dict) + "choose Finding call: (press 'enter' for default(4))")
|
|
pg_num = int(input('how many pages? (76 max)'))
|
|
|
|
optional_params = {
|
|
"itemFilter(0).name":"Condition",
|
|
"itemFilter(0).value":"Used"
|
|
} # NOTE setting as default in get_data() method
|
|
|
|
if fnd_srvc != '':
|
|
fnd_srvc = int(fnd_srvc)
|
|
finding = FindingApi(fnd_srvc, pg_num)
|
|
else:
|
|
fnd_srvc = 4
|
|
finding = FindingApi(fnd_srvc, pg_num)
|
|
|
|
item_id_results = finding.get_ids_from_cats()
|
|
with concurrent.futures.ThreadPoolExecutor() as executor:
|
|
for future in executor.map(self.get_item_from_findItemsByCategory, item_id_results):
|
|
for item in future:
|
|
data.append(item) # The end result should be a list of dicts where each dict in the list is a listing
|
|
# data.update(future)
|
|
with open('raw_data.txt', 'w') as f:
|
|
json.dump(data, f)
|
|
return data # each future is a list of dictionaries because the output of any multithreader in this method is a list.
|
|
|
|
# data dictionary can't update from list of dicts unless iterated over. Might need a different way to update.
|
|
# TODO It seems like the problem with updating the dictionary/csv file is starting here possibly; I think the item data is getting appended out of order from the item itself.
|
|
|
|
# NOTE:
|
|
|
|
# Limited to 5000 calls to shopping api per day, and getMultpileitems service maxes out at 20 items
|
|
# per call leaving you 100,000 items per day for you pandas dataframe initially. So you'll have
|
|
# to divide these up into the categories. This will leave you with about 6.25K results per cat.
|
|
# More than enough data for your dataset.
|
|
|
|
|
|
class CurateData:
|
|
'''
|
|
Contains methods for curating data for machine learning training sets;
|
|
Takes item in data from ShoppingApi request as argument and extracts/ creates key
|
|
value pairs that gets updated to custom dataframe used in Ml training sets.
|
|
'''
|
|
|
|
def import_raw(self):
|
|
'''
|
|
imports raw response json from local file. This is data from
|
|
GetMultipleItems call in ShoppingApi
|
|
'''
|
|
with open('raw_data.txt') as f:
|
|
raw_data = json.load(f)
|
|
return raw_data
|
|
|
|
def raw_df(self, raw_data): # TODO not dropping dupes, and is appending raw_data for some reason
|
|
'''
|
|
creates pandas df from raw json and saves master raw csv file as raw_df.csv.
|
|
Indended to be used inline with direct
|
|
data stream from ebay's APIs
|
|
'''
|
|
to_json = json.dumps(raw_data)
|
|
raw_df = pd.read_json(to_json)
|
|
raw_df.to_csv('raw_df.csv') # NOTE not append mode because raw_df is made from the master raw_data.txt file
|
|
#raw_df = pd.read_csv('raw_df.csv', index_col=0)
|
|
#raw_df.drop_duplicates(subset=['ItemID']).reset_index(drop=True) # may not need this
|
|
#raw_df.to_csv('raw_df.csv')
|
|
|
|
# TODO still saving "Unnamed:0" column
|
|
|
|
return raw_df
|
|
|
|
def to_training(self, raw_data):
|
|
'''
|
|
creates first pass of potential labels for training set. This is the base
|
|
df used to produce other training sets to use.
|
|
'''
|
|
raw_df = self.raw_df(raw_data)
|
|
interm_df1 = raw_df.loc[:,['ItemID', 'PictureURL', 'PrimaryCategoryID', 'PrimaryCategoryName', 'Title', 'ItemSpecifics']]
|
|
interm_df1[['ItemID', 'PrimaryCAegoryID']] = interm_df1.loc[:, ['ItemID', 'PrimaryCategoryID']].astype(str)
|
|
training = interm_df1.dropna(subset=['ItemSpecifics'])
|
|
return training # TODO RENAME THIS FUNC AND its RETURN VALUE
|
|
|
|
def class_training(self, training):
|
|
'''Training set for multiclass portion of training set. Used to train
|
|
seprately from multilabel portion
|
|
'''
|
|
class_training = training.loc[:, ['PictureURL', 'PrimaryCategoryID']]
|
|
return class_training
|
|
|
|
def nvl_training(self, training):
|
|
'''
|
|
Training set for multilabel portion
|
|
'''
|
|
interm_df1 = pd.Series(training.ItemSpecifics)
|
|
interm_df1 = interm_df1.apply(lambda x: x['NameValueList'])
|
|
|
|
# Necessary for json_normalize():
|
|
|
|
nvl_dict = interm_df1.apply(lambda x: {k:v for (k, v) in zip([n['Name'] for n in x], [v['Value'] for v in x])})
|
|
nvl_df = pd.json_normalize(nvl_dict)
|
|
nvl_training = pd.concat([pd.Series(training.PictureURL), nvl_df], axis=1)
|
|
|
|
return nvl_training
|
|
|
|
def extract_df(self, df):
|
|
'''
|
|
converts single-value lists of strings of any df to string if not null
|
|
'''
|
|
extracted_df = df.applymap(lambda x: ' '.join(x) if isinstance(x, list) else np.nan if pd.isnull(x) else x)
|
|
|
|
return extracted_df
|
|
|
|
def drop_nvl_cols(self, nvl_training): # NOTE this is wonky
|
|
col_drop = [
|
|
'Fabric Type', 'Type of Sport', 'Mid Sole', 'Modified Item',
|
|
'Modification Description', 'Article Type', 'Customized',
|
|
'Character', 'Features', 'Colors', 'Shade', 'Product ID',
|
|
'Personalized', 'Platform Height', 'Year Manufactured',
|
|
'Trim Material', 'Fashion Element', 'Shaft Material',
|
|
'Character Family', 'Heel to Toe Drop', 'Custom Bundle',
|
|
'California Prop 65 Warning', 'Manufacturer Color', 'Main Color',
|
|
'Collection', 'Midsole Type', 'Signed', 'US Shoe Size (Men#!#s)',
|
|
'Calf Circumference', 'Handmade', 'Safety Standards',
|
|
'Customised', 'Cleat Type', 'Cushioning Level', 'AU Shoe Size',
|
|
'Country/Region of Manufacture', 'Type of Sport', 'Main Colour',
|
|
'Look', 'Sole Type', 'Manufacturer Colour', 'Sole Material',
|
|
'Toe Material', 'Feature', 'Length', 'Width', 'Size Chart',
|
|
'Boot Height', 'Water Resistance Level', 'Material Composition',
|
|
'Calf Width', 'Insole Material', 'UPC', 'Size Type'
|
|
]
|
|
col_keep = [
|
|
'PictureURL', 'Style', 'Department', 'Type', 'Gender', 'Closure', 'Performance/Activity',
|
|
'Accents', 'Occasion', 'Toe Shape', 'Pattern', 'Activity',
|
|
'Heel Style', 'Fastening', 'Heel Type', 'Toe Type', 'Departement',
|
|
'Product Type', 'Sub Style', 'Season', 'Theme', 'Upper Material'
|
|
]
|
|
# May be no difference between Product type and sub style; fastening and
|
|
# closure; toe shape and toe type; occasion and performance/activity;
|
|
# see if you can combine these somehow (you may not want this though).
|
|
# Also consider keeping only cols that have plenty of values
|
|
# Run some value_count() analysis to determine frequencies and filter
|
|
# user created item specifics, leaving only predefined ebay item specs
|
|
|
|
user_input = input('drop or keep cols?:')
|
|
|
|
'''
|
|
dropping and or keeping/masking functions to create your filtered df below is
|
|
producing errors due to some column lables in your predefined lists not being present.
|
|
Look at documentation to see if option exists to ignore items not present
|
|
|
|
keep col option is ideal due to users inputting crappy custom fields in
|
|
item specifics. Use this if you can
|
|
'''
|
|
if 'keep' in user_input:
|
|
dropd = nvl_training.reindex([col_keep]) # TODO ERRORS HERE USING LOC OR REINDEX WITH MULTIPLE COL LABELS
|
|
else:
|
|
dropd = nvl_training#.drop(col_drop, errors='ignore', axis=1) # errors='ignore' for non existent labels
|
|
|
|
return dropd
|
|
|
|
# for future reference, to deal with inconsistent values in the nvl (due to sellers inputting custom values in the fields) you can drop either listings or k/v pairs that are unique which
|
|
# can be determined from applying a function to determine frequency of k/v pairs--> list of unique k/v pairs--> function to determine frequency of unique k/v pairs--> drop those that have 1.
|
|
|
|
# Check the above list of cols I want to keep to see if there are duplicates with diff spelling and phrasing (e.g., Departement and Department, or Fastening and Closure Type)
|
|
|
|
def expand_nvlclass(self, class_training, dropd):
|
|
'''
|
|
takes image url list from each cell and expands them into separate/duplicate
|
|
instances. Modifies both class training and dropd dfs. Appends custom
|
|
image url dict {'source':'target'}.
|
|
|
|
* consider applying this function to other cells that have multiple values in their lists
|
|
'''
|
|
expanded_class = class_training.explode('PictureURL').reset_index(drop=True)
|
|
expanded_class = expanded_class.dropna(subset=['PictureURL'])
|
|
expanded_class = expanded_class.drop_duplicates(subset=['PictureURL']).reset_index(drop=True)
|
|
|
|
expanded_dropd = dropd.explode('PictureURL').reset_index(drop=True)
|
|
expanded_dropd = expanded_dropd.dropna(subset=['PictureURL'])
|
|
expanded_dropd = expanded_dropd.drop_duplicates(subset=['PictureURL']).reset_index(drop=True)
|
|
|
|
expanded_dropd = self.extract_df(expanded_dropd) # convert lists to values
|
|
|
|
temp_pics_source_list = list(set(expanded_class.PictureURL.to_list()))
|
|
|
|
|
|
try:
|
|
with open('temp_pics_source_list.txt') as f:
|
|
tpsl = json.load(f)
|
|
tpsl.extend(temp_pics_source_list)
|
|
temp_pics_source_list = list(set(tpsl))
|
|
with open('temp_pics_source_list.txt', 'w') as f:
|
|
json.dump(temp_pics_source_list, f)
|
|
except (ValueError, FileNotFoundError):
|
|
with open('temp_pics_source_list.txt', 'w') as f:
|
|
json.dump(temp_pics_source_list, f)
|
|
|
|
# Append to master training dataframes, drop potential dupes and save
|
|
|
|
expanded_class.to_csv('expanded_class.csv')
|
|
# expanded_class = pd.read_csv('expanded_class.csv', index_col=0)
|
|
# expanded_class.drop_duplicates(subset=['PictureURL']).reset_index(drop=True)
|
|
# expanded_class.to_csv('expanded_class.csv', mode='a', encoding='utf-8') # TODO see line 235 about views and copies
|
|
|
|
expanded_dropd.to_csv('expanded_dropd.csv')
|
|
# expanded_dropd = pd.read_csv('expanded_dropd.csv', index_col=0)
|
|
# expanded_dropd.drop_duplicates(subset=['PictureURL']).reset_index(drop=True)
|
|
# expanded_dropd.to_csv('expanded_dropd.csv', mode='a', encoding='utf-8')
|
|
|
|
return expanded_class, expanded_dropd
|
|
|
|
def dl_pictures(self, *args):
|
|
'''
|
|
Downloads pictures from api to local storage using temp_pics_source_list
|
|
and creates custom {source:target} dictionary as dict_pics
|
|
'''
|
|
|
|
try:
|
|
with open('target_dirs.txt', 'r+') as f: # TODO you can add option to change directory here, too. Look up how to have optional arguments
|
|
target_dir = json.load(f)
|
|
except (ValueError, FileNotFoundError):
|
|
target_dir = input('No target dirctory found. Create One? [y] or [n]:')
|
|
if target_dir == ('y' or 'Y'):
|
|
target_dir = input('Please provide full URL to destination folder:') # TODO need to catch human syntax errors here
|
|
with open('target_dirs.txt','w') as f:
|
|
json.dump(target_dir, f)
|
|
else:
|
|
os.mkdir(os.getcwd()+os.sep+'training_images')
|
|
target_dir = os.getcwd()+os.sep+'training_images'
|
|
with open('target_dirs.txt','w') as f:
|
|
json.dump(target_dir, f)
|
|
print('Creating default folder in current directory @ ' + target_dir)
|
|
|
|
with open('temp_pics_source_list.txt') as f:
|
|
try:
|
|
if args:
|
|
temp_pics_source_list = args[0]
|
|
else:
|
|
temp_pics_source_list = json.load(f)
|
|
except (ValueError, FileNotFoundError):
|
|
if args:
|
|
temp_pics_sources_list = args[0]
|
|
else:
|
|
print('url list not found. download aborted')
|
|
return
|
|
|
|
temp_dict_pics = {k:target_dir + os.sep + re.search(r'[^/]+(?=/\$_|.jpg)', k, re.IGNORECASE).group() + '.jpg' for k in temp_pics_source_list}
|
|
|
|
try:
|
|
with open('dict_pics.txt') as f:
|
|
dict_pics = json.load(f)
|
|
dict_pics.update(temp_dict_pics) # TODO This still creates duplicates
|
|
with open('dict_pics.txt', 'w') as f:
|
|
json.dump(dict_pics, f)
|
|
|
|
except (ValueError, FileNotFoundError):
|
|
with open('dict_pics.txt', 'w') as f:
|
|
json.dump(temp_dict_pics, f)
|
|
dict_pics = temp_dict_pics
|
|
|
|
def dl_pic(dict_pics, pic):
|
|
|
|
if os.path.exists(dict_pics[pic]): # or call temp_dict_pics[pic] can work
|
|
pass # TODO This is not catching duplicates for some reason....possibly not? Upon inspection, files aren't duplicates...but why?
|
|
#TODO it would mean that temp_pics_source_list is changing for some reason?
|
|
|
|
else:
|
|
r = requests.get(pic, stream=True)
|
|
r.raw.decode_content = True
|
|
with open(temp_dict_pics[pic], 'wb') as f: # Or call dict_pics[pic] can work
|
|
shutil.copyfileobj(r.raw, f)
|
|
|
|
bargs = [(dict_pics, pic) for pic in temp_pics_source_list]
|
|
|
|
with concurrent.futures.ThreadPoolExecutor() as executor:
|
|
for future in executor.map(lambda p: dl_pic(*p), bargs):
|
|
future
|
|
|
|
os.remove('temp_pics_source_list.txt') # Deletes file after downloads complete successfully
|
|
|
|
class PreProcessing:
|
|
'''
|
|
Includes methods for pre-processing training set input and labels in the
|
|
training set created from CurateData class. Whereas CurateData training
|
|
sets provided trimmed down data from the raw json response from the
|
|
ShoppingApi call and provided a bare minimum format for the dataframe to be
|
|
used in training, PreProcessing optimizes that dataframe for training and
|
|
includes methods for image manipulation, creating test/train/validation
|
|
splits, etc.
|
|
'''
|
|
|
|
def stt_training(self, dict_pics, expanded_class, expanded_dropd):
|
|
'''
|
|
Source to target training. Replaces source image URL with target URL
|
|
determined by values in dict_pics variable.
|
|
'''
|
|
pass
|
|
|
|
# TODO pipeline gameplan: 5 files: dict_pics.txt,raw_json.txt, raw_json.csv, expanded_class.csv, expanded_dropd.csv
|
|
# cont... open raw_json.txt and append, same with csv --> process new data --> pull out image source+dest and expand new dfs for the additional pictures
|
|
# if not exists and append to master img download dict
|
|
# --> concat m_class_training df and m_nvl_training dfs with new data. Need to add inclusion tests for all files when opened and appended/concatted
|
|
|
|
def main():
|
|
'''
|
|
Main program creates/updates a csv file to use for ML training from live
|
|
ebay listings
|
|
'''
|
|
pass
|
|
# main goes here:
|
|
|
|
if __name__ == "__main__":
|
|
main()
|
|
|
|
'''
|
|
Based on your sample set of 10 images, if you have an average of 5 images per
|
|
listing and you download a hundred listings, you will have about 102 Gb of
|
|
image data. That's just for one day. If you have more than a million listings
|
|
you're looking at a little over 1Tb of image data. You don't even know if this
|
|
is good data yet.
|
|
'''
|