def prune():
with session() as sess:
"""prune machines which haven't removed themselves properly"""
sess.execute(f"""
delete from machines where updated_at < now() - interval '5 minutes'
""")
sess.commit()
def await_job(jid):
with session() as sess:
params = {'jid': jid}
i = 0
while True:
time.sleep(1)
job = sess.execute(
text("""
select state, method from jobs where id=:jid
"""), params).fetchone()
# TODO notify them of error?
# 10 minutes, give up
if job.state == 'error' or i > 60 * 10:
return Box(method=job.method, data_out=False)
if job.state == 'done':
## don't delete actually, let Job.prune handle that. Need last_job (created_at)
#delete from jobs where id=:jid returning method, data_out
job = sess.execute(
text("""
select method, data_out from jobs where id=:jid
"""), params).fetchone()
sess.commit()
return job
i += 1
def ga(uid: Union[str, UUID4], category: str, action: str):
"""
I'm only tracking interesting server-side events right now (no cookies), I want to see what features
are being used and important bits like sign-ups & book-thumbs. user-id is obfuscated
https://developers.google.com/analytics/devguides/collection/protocol/v1/devguide#event
"""
# actually don't care about uid, just need a unique identifier. Note this is
# a 1-way hash (right?), so I can't even decrypt - just want unique per-feature track
uid_ = str(uid).encode() # to bytes
uid_ = hashlib.sha256(uid_).hexdigest()
url = "https://ssl.google-analytics.com/"
url += "debug/collect" if DEBUG else "collect"
res = requests.post(url,
params=dict(v=1,
tid=vars.GA,
cid=uid_,
t='event',
ec=category,
ea=action))
# if DEBUG: print(res.json())
if action in ('register', 'like', 'dislike', 'therapist', 'notes'):
with session() as sess:
if sess.execute(
text("""
select is_superuser su from users where id=:uid
"""), dict(uid=uid)).fetchone().su:
# don't notify of my own or Lisa's entries
return
send_mail('*****@*****.**', 'action',
dict(category=category, action=action))
def nlp_on_rows(method, id, job_id):
for_entries = method == 'entries'
with session() as sess:
uids = nlp_on_rows_(method, id, job_id, sess, set())
# profiles doesn't use uids, but create 1-el array to iterate anyway
if not for_entries: uids = [None]
for i, uid in enumerate(uids):
uid = str(uid)
gen_keywords(for_entries, uid, sess)
if for_entries:
# Then add a book job for every user who was affected. Delay by x minutes per
# job (jobs are pruned by updated_at < ?, so posting to the future)
# 9131155e: only update every x entries
future = datetime.utcnow() + timedelta(minutes=i*3)
sess.add(M.Job(
method='books',
data_in={'args': [uid]},
created_at=future,
updated_at=future,
))
sess.commit()
else:
match_profiles(sess)
return {}
def themes(eids, algo='agglomorative'):
logger.info("Themes")
with session() as sess:
# use Model to decrypt fields
E, CE = M.Entry, M.CacheEntry
res = sess.query(CE) \
.with_entities(CE.paras, CE.clean, CE.vectors) \
.join(E, E.id == CE.entry_id) \
.filter(E.id.in_(eids), func.array_length(CE.vectors,1)>0) \
.order_by(E.created_at.desc()) \
.all()
# assert len(eids) == len(res)
entries = pd.Series([e for r in res for e in r.paras])
stripped = pd.Series([c for r in res for c in r.clean])
vecs = np.vstack([r.vectors for r in res]).astype(np.float32)
chain = Similars(vecs)
if False and os.path.exists(vars.AE_PATH):
chain = chain.autoencode(filename=vars.AE_PATH).cluster(algo=algo)
else:
chain = chain.normalize().cluster(algo=algo)
clusters = chain.value()
labels = chain.data.labels
topics = []
for l, center in enumerate(clusters):
mask = labels == l
n_entries = mask.sum().item()
print('n_entries', n_entries)
if n_entries < 2:
print('skipping')
continue
vecs_, stripped_, entries_ = vecs[mask], stripped[mask], entries[mask]
dists = Similars(center,
vecs_).normalize().cosine(abs=True).value().squeeze()
entries_ = entries_.iloc[dists.argsort()].tolist()[:5]
terms = top_terms(stripped_.tolist())
topics.append({
'n_entries': n_entries,
'terms': terms,
'summary': entries_, # add full thing, will batch-compute next
'sentiment': None,
})
groups = [t['summary'] for t in topics]
batch_summaries = nlp_.summarization(groups, min_length=50, max_length=300)
for i, res in enumerate(batch_summaries):
print(res)
topics[i]['summary'] = res['summary']
topics[i]['sentiment'] = res['sentiment']
topics = {'terms': top_terms(stripped, 10), 'themes': topics}
return topics
def prune():
with session() as sess:
"""prune completed or stuck jobs. Completed jobs aren't too useful for admins; error is."""
sess.execute(f"""
delete from jobs
where updated_at < now() - interval '10 minutes'
and state in ('working', 'done')
""")
sess.commit()
def db(client):
"""await client to init_db"""
with D.session() as sess:
# wait for GPU to restart from no-db crash
while True:
sql = "select 1 from machines where status='on'"
if M.await_row(sess, sql): break
time.sleep(.5)
yield sess
def themes(eids, algo='agglomorative'):
with session() as sess:
# use Model to decrypt fields
res = sess.query(M.CacheEntry)\
.with_entities(M.CacheEntry.paras, M.CacheEntry.clean, M.CacheEntry.vectors)\
.join(M.Entry, M.Entry.id == M.CacheEntry.entry_id)\
.filter(M.Entry.id.in_(eids))\
.order_by(M.Entry.created_at.desc())\
.all()
# assert len(eids) == len(res)
entries = pd.Series([e for r in res for e in r.paras])
stripped = pd.Series([c for r in res for c in r.clean])
vecs = []
for r in res:
if r.vectors: vecs += r.vectors
# if not vecs: return False # TODO somethign else to return?
vecs = np.vstack(vecs).astype(np.float32)
clusters = Similars(vecs).normalize().cluster(algo=algo).value()
topics = []
for l in range(clusters.max()):
in_clust = clusters == l
n_entries = in_clust.sum().item()
print('n_entries', n_entries)
if n_entries < 2:
print('skipping')
continue
vecs_, stripped_, entries_ = vecs[in_clust],\
stripped.iloc[in_clust], entries.iloc[in_clust]
center = vecs_.mean(axis=0)[np.newaxis, :]
dists = Similars(center, vecs_).normalize().cosine().value().squeeze()
entries_ = entries_.iloc[dists.argsort()].tolist()[:5]
terms = top_terms(stripped_.tolist())
topics.append({
'n_entries': n_entries,
'terms': terms,
'summary': entries_, # add full thing, will batch-compute next
'sentiment': None,
})
groups = [t['summary'] for t in topics]
batch_summaries = nlp_.summarization(groups, min_length=50, max_length=300)
for i, res in enumerate(batch_summaries):
print(res)
topics[i]['summary'] = res['summary']
topics[i]['sentiment'] = res['sentiment']
topics = {'terms': top_terms(stripped, 10), 'themes': topics}
return topics
def influencers():
with session() as sess:
users = sess.execute(
text(f"""
select id::text from users
where
-- has logged in recently
updated_at > now() - interval '2 days' and
-- has been 1d since last-run (or never run)
(extract(day from now() - last_influencers) >= 1 or last_influencers is null)
""")).fetchall()
for u in users:
uid_ = dict(uid=u.id)
sess.execute(
text(f"""
update users set last_influencers=now() where id=:uid
"""), uid_)
sess.commit()
res = influencers_(sess, u.id)
if not res: continue
# A field can get deleted while running XGB, causing a fkey constraint error.
# https://docs.sqlalchemy.org/en/13/dialects/postgresql.html
# Can't do on_conflict for FK constraints, get fresh ids and filter out missing ones.
fids = [
x.id for x in sess.execute(
text("""
select id::text from fields where user_id=:uid
"""), uid_).fetchall()
]
next_preds, importances, all_imps = res
for fid, others in importances.items():
if fid not in fids: continue
inf_score, next_pred = all_imps[fid], next_preds[fid]
insert = postgresql.insert(M.Influencer.__table__).values([
dict(field_id=fid, influencer_id=inf_id, score=score)
for inf_id, score in others.items() if inf_id in fids
])
sess.execute(
insert.on_conflict_do_update(
constraint=M.Influencer.__table__.primary_key,
set_=dict(score=insert.excluded.score)))
sess.execute(
text("""
update fields set influencer_score=:score, next_pred=:pred
where id=:fid;
"""), dict(score=inf_score, pred=next_pred, fid=fid))
sess.commit()
return {}
def cloud_up_maybe():
if is_dev(): return
with session() as sess:
if M.User.last_checkin(sess) > 15: return
if M.Machine.gpu_status(sess) in ("on", "pending"): return
logger.warning("Initing AWS Batch")
M.Machine.notify_online(sess, 'batch', 'pending')
boto3.client('batch').submit_job(
jobName=str(uuid4()),
jobQueue='gnothi-jq',
jobDefinition='gnothi-jd',
)
def predict_books(user_id, vecs_user, n_recs=30, centroids=False):
with session() as sess:
# TODO should I move this down further, to get more lines to test?
fixt = fixtures.load_books(user_id)
if fixt is not None: return fixt
vecs_books, books = load_books(sess)
sql = "select book_id as id, user_id, shelf from bookshelf where user_id=%(uid)s"
shelf = pd.read_sql(sql, sess.bind, params={'uid': user_id}).set_index('id', drop=False)
shelf_idx = books.id.isin(shelf.id)
# normalize for cosine, and downstream DNN
chain = Similars(vecs_user, vecs_books).normalize()
vecs_user, vecs_books = chain.value()
logger.info("Finding cosine similarities")
if centroids:
labels = chain.agglomorative().value()
lhs = np.vstack([
vecs_user[labels == l].mean(0)
for l in range(labels.max())
])
chain = Similars(lhs, vecs_user)
# Take best cluster-score for every book
dist = chain.cosine().value().min(axis=0)
# 0f29e591: minmax_scale(dist). norm_out=True works better
# then map back onto books, so they're back in order (pandas index-matching)
books['dist'] = dist
if shelf_idx.sum() > 0:
like, dislike = dist.min() - dist.std(), dist.max() + dist.std()
shelf_map = dict(like=like, already_read=like, recommend=like, dislike=dislike, remove=None, ai=None)
shelf['dist'] = shelf.shelf.apply(lambda k: shelf_map[k])
shelf.dist.fillna(books.dist, inplace=True) # fill in "remove"
books.loc[shelf.index, 'dist'] = shelf.dist # indexes(id) match, so assigns correctly
assert not books.dist.isna().any(), "Messed up merging shelf/books.dist by index"
# e2eaea3f: save/load dnn
dnn = train_books_predictor(books, vecs_books, shelf_idx)
books['dist'] = dnn.predict(vecs_books)
# dupes by title in libgen
# r = books.sort_values('dist')\
df = books[~shelf_idx].sort_values('dist')\
.drop_duplicates('title', keep='first')\
.iloc[:n_recs]
fixtures.save_books(user_id, df)
return df
def multiple_book_jobs(uids):
with session() as sess:
sess.execute(
satext("""
update users set last_books=null where id in :uids;
"""), dict(uids=tuple(uids)))
sess.commit()
# TODO handle this in run.py when it's consuming jobs
def delay_books(uid, i):
time.sleep(i * 60 * 5) # 5m
Job.create_job('books', data_in=dict(args=[str(uid)]))
for i, uid in enumerate(uids):
threading.Thread(target=delay_books, args=(uid, i)).start()
def wrap_job(jid, method, fn):
logger.info(f"Run job {method}")
try:
start = time.time()
res = fn()
sql = "update jobs set state='done', data_out=:data where id=:jid"
logger.info(f"Job {method} complete {time.time() - start}")
except Exception as err:
err = str(traceback.format_exc()) # str(err)
res = dict(error=err)
sql = "update jobs set state='error', data_out=:data where id=:jid"
logger.error(f"Job {method} error {time.time() - start} {err}")
with session() as sess:
sess.execute(satext(sql), dict(data=jsonb(res), jid=str(jid)))
sess.commit()
def match_profiles():
with session() as sess:
df = pd.read_sql(
"""
select e.user_id, c.vectors from cache_entries c
inner join entries e on e.id=c.entry_id
where c.vectors is not null
""", sess.bind)
if not df.shape[0]: return
# flatten multi-paragraph entries
df['vectors'] = df.vectors.apply(mean_)
# then mean the semantic of all entries for this user.
# TODO cluster or something, just mean-ing all their entries is stupid
df = df.groupby(['user_id']).vectors.agg(mean_)
uids = df.index.tolist()
vecs_entries = np.vstack(df.values)
# TODO add community (M.User.public == True)
df = pd.read_sql(
"""
select c.user_id, c.vectors from cache_users c
inner join users u on c.user_id=u.id
where u.therapist=true and c.vectors is not null
""", sess.bind)
if not df.shape[0]: return
match_ids = df.user_id.tolist()
# This on the other hand is OK to mean, it's just their profile
vecs_profiles = np.vstack(df.vectors.apply(mean_).values)
dists = Similars(vecs_entries,
vecs_profiles).normalize().cosine().value()
sess.execute(
text("""
delete from profile_matches where user_id in :uids
"""), dict(uids=tuple(uids)))
sess.commit()
# everything is in same order at this point
sess.bulk_save_objects([
M.ProfileMatch(user_id=uid, match_id=mid, score=dists[i, j])
for i, uid in enumerate(uids) for j, mid in enumerate(match_ids)
])
sess.commit()
def clear_fixtures(self):
all_ = FRESH == 'all'
if 'books' in FRESH or all_:
self.rm(f"{BASE}/books.pkl")
if 'entries' in FRESH or all_:
self.rm(f"{BASE}/entries.pkl")
self.rm(f"{BASE}/nlp_entries.pkl")
if 'profiles' in FRESH or all_:
self.rm(f"{BASE}/nlp_profiles.pkl")
if 'wiki' in FRESH or all_:
self.rm(f"{BASE}/wiki", isdir=True)
if 'influencers' in FRESH or all_:
self.rm(f"{BASE}/xgb_hypers.pkl")
if 'liben' in FRESH or all_:
self.rm(f"/storage/libgen_testing.npy")
with session() as sess:
sess.execute("delete from books")
sess.commit()
def run_job(job):
jid_, k = str(job.id), job.method
jid = {'jid': jid_}
with session() as sess:
data = sess.execute("select data_in from jobs where id=:jid", jid).fetchone().data_in
args = data.get('args', [])
kwargs = data.get('kwargs', {})
if k in ('entries', 'profiles'):
kwargs['job_id'] = jid_
if k == 'books':
nlp_.clear()
os.system(f"python app/books.py --jid={jid_} --uid={args[0]}")
return
def fn(): return m[k](*args, **kwargs)
M.Job.wrap_job(jid_, k, fn)
def cloud_up_maybe():
if is_dev(): return
with session() as sess:
if M.User.last_checkin(sess) > 10: return
if M.Machine.gpu_status(sess) != "off": return
logger.warning("Initing Paperspace")
M.Machine.notify_online(sess, 'paperspace', 'pending')
jobs = job_client.list()
if any([j.state in up_states for j in jobs]):
return
vars_ = {**dict(vars), **{'MACHINE': 'paperspace'}}
return job_client.create(machine_type='K80',
container='lefnire/gnothi:gpu-0.0.13',
project_id=vars.PAPERSPACE_PROJECT_ID,
is_preemptible=True,
command='python app/run.py',
env_vars=vars_)
def run_books(user_id):
with session() as sess:
user_id = str(user_id)
uid = {'uid': user_id}
# don't run if ran recently (notice the inverse if & comparator, simpler)
if sess.execute(text(f"""
select 1 from users
where id=:uid and last_books > {utcnow} - interval '10 minutes'
"""), uid).fetchone():
return
sess.execute(text(f"""
update users set last_books={utcnow} where id=:uid
"""), uid)
sess.commit()
entries = sess.execute(text("""
select c.vectors from cache_entries c
inner join entries e on e.id=c.entry_id and e.user_id=:uid
order by e.created_at desc;
"""), uid).fetchall()
profile = sess.execute(text("""
select vectors from cache_users where user_id=:uid
"""), uid).fetchone()
vecs = []
if profile and profile.vectors:
vecs += profile.vectors
for e in entries:
if e.vectors: vecs += e.vectors
vecs = np.vstack(vecs).astype(np.float32)
res = predict_books(user_id, vecs)
sess.execute(text("""
delete from bookshelf where user_id=:uid and shelf='ai'
"""), uid)
sess.commit()
res = res.rename(columns=dict(id='book_id', dist='score'))[['book_id', 'score']]
res['user_id'] = user_id
res['shelf'] = 'ai'
res['created_at'] = res['updated_at'] = datetime.datetime.utcnow()
res.to_sql('bookshelf', sess.bind, if_exists='append', index=False)
def create_job(method, data_in={}, **kwargs):
"""
Ensures certain jobs only created once at a time. Never manually add Job() call this instead
"""
with session() as sess:
arg0 = data_in.get('args', [None])[0]
if type(arg0) != str: arg0 = None
# For entries, profiles: set ai_ran=False to queue them into the next batch,
# then arg0 isn't used downstream (was previously).
if method in ('entries', 'profiles') and arg0:
table = dict(entries='entries', profiles='users')[method]
sess.execute(
satext(f"""
update {table} set ai_ran=False where id=:id;
"""), dict(id=arg0))
sess.commit()
exists = sess.execute(
satext("""
select 1 from jobs
-- maybe if we're mid-job, things have changed; so don't incl. working? rethink
--where method=:method and state in ('new', 'working') and
where method=:method and state='new' and
case
when method='influencers' then true
when method='books' and data_in->'args'->>0=:arg0 then true
when method='entries' then true
when method='profiles' then true
when method='habitica' then true
else false
end
"""), dict(method=method, arg0=arg0)).fetchone()
if exists: return False
j = Job(method=method, data_in=data_in, **kwargs)
sess.add(j)
sess.commit()
sess.refresh(j)
return str(j.id)
def predict_books(user_id, vecs_user, n_recs=30):
with session() as sess:
# TODO should I move this down further, to get more lines to test?
fixt = fixtures.load_books(user_id)
if fixt is not None: return fixt
vecs_books, books = load_books(sess, user_id)
# normalize for cosine, and downstream DNN
chain = Similars(vecs_user, vecs_books).normalize()
vecs_user, vecs_books = chain.value()
logger.info("Finding cosine similarities")
# Take best cluster-score for every book
dist = chain.cosine(abs=True).value().min(axis=0)
# 0f29e591: minmax_scale(dist). norm_out=True works better
# then map back onto books, so they're back in order (pandas index-matching)
# Push highly-rated books up, low-rated books down. Do that even stronger for user's own ratings.
# Using negative-score because cosine DISTANCE (less is better)
books['dist'] = dist
books['dist'] = books.dist \
+ (books.dist.std() * -books.global_score / 2.) \
+ (books.dist.std() * -books.user_score)
assert not books.dist.isna().any(), "Messed up merging shelf/books.dist by index"
# e2eaea3f: save/load dnn
dnn = train_books_predictor(books, vecs_books)
books['dist'] = dnn.predict(vecs_books)
# dupes by title in libgen
# r = books.sort_values('dist')\
df = books[~books.user_rated].sort_values('dist')\
.drop_duplicates('title', keep='first')\
.iloc[:n_recs]
fixtures.save_books(user_id, df)
return df
def await_job(jid):
with session() as sess:
params = {'jid': jid}
i = 0
while True:
time.sleep(1)
job = sess.execute(
text("""
select state, method from jobs where id=:jid
"""), params).fetchone()
# TODO notify them of error?
# 10 minutes, give up
if job.state == 'error' or i > 60 * 10:
return Box(method=job.method, data_out=False)
if job.state == 'done':
job = sess.execute(
text("""
select method, data_out from jobs where id=:jid
"""), params).fetchone()
sess.commit()
return job
i += 1
return m[k](*args, **kwargs)
M.Job.wrap_job(jid_, k, fn)
# 3eb71b3: unloading models. multiprocessing handles better
if __name__ == '__main__':
logger.info(f"torch.cuda.current_device() {torch.cuda.current_device()}")
logger.info(f"torch.cuda.device(0) {torch.cuda.device(0)}")
logger.info(f"torch.cuda.device_count() {torch.cuda.device_count()}")
logger.info(
f"torch.cuda.get_device_name(0) {torch.cuda.get_device_name(0)}")
logger.info(f"torch.cuda.is_available() {torch.cuda.is_available()}")
logger.info("\n\n")
with session() as sess:
while True:
M.Machine.notify_online(sess, vars.MACHINE)
cloud_down_maybe(sess)
# only allow 2 jobs at a time.
if M.Machine.job_ct_on_machine(sess, vars.MACHINE) >= 2:
time.sleep(1)
continue
# Find jobs
job = M.Job.take_job(sess, "run_on='gpu'")
if job:
# aaf1ec95: multiprocessing.Process for problem models
threading.Thread(target=run_job, args=(job, )).start()
# run_job(job.id)
def eid_to_title(self, eid):
with session() as sess:
return sess.execute(
text("""
select title from entries where id=:eid
"""), dict(eid=eid)).fetchone().title
def uid_to_email(self, uid):
with session() as sess:
return sess.execute(
text("""
select email from users where id=:uid
"""), dict(uid=uid)).fetchone().email
def load_books_df(sess):
# sort asc since that's how we mapped to vecs in first place (order_values)
df = pd.read_sql("select * from books order by id asc", sess.bind)
if df.shape[0]:
return df.set_index('id', drop=False)
logger.info("Load books MySQL")
FIND_PROBLEMS = False
ALL_BOOKS = False
# for-sure psych. See tmp/topics.txt, or libgen.sql topics(lang='en')
psych_topics = 'psychology|self-help|therapy'
# good other psych topics, either mis-categorized or other
psych_topics += '|anthropology|social|religion'
psych_topics += '|^history|^education'
sql = Box(
select="select u.ID, u.Title, u.Author, d.descr, t.topic_descr",
body="""
from updated u
inner join description d on d.md5=u.MD5
inner join topics t on u.Topic=t.topic_id
and t.lang='en'
where u.Language = 'English'
and title not regexp 'sams|teach yourself' -- remove junk
and (length(d.descr) + length(u.Title)) > 200
and u.ID not in ('62056','72779','111551','165602','165606','239835','240399','272945','310202','339718','390651','530739','570667','581466','862274','862275','879029','935149','1157279','1204687','1210652','1307307','1410416','1517634','1568907','1592543','2103755','2128089','2130515','2187329','2270690','2270720','2275684','2275804','2277017','2284616','2285559','2314405','2325313','2329959','2340421','2347272','2374055','2397307','2412259','2420958','2421152','2421413','2423975')
""",
# handle u.Topic='' (1326526 rows)
just_psych=f"and t.topic_descr regexp '{psych_topics}'",
# find_problems
just_ids="select distinct u.ID",
where_id="and u.ID=:id"
)
if FIND_PROBLEMS:
# # Those MD5s: UnicodeDecodeError: 'charmap' codec can't decode byte 0x9d in position 636: character maps to <undefined>
# TODO try instead create_engine(convert_unicode=True)
with session('books') as sessb:
ids = ' '.join([sql.just_ids, sql.body])
ids = [x.ID for x in sessb.execute(ids).fetchall()]
problem_ids = []
for i, id in enumerate(tqdm(ids)):
if i % 10000 == 0:
problems = len(problem_ids) / len(ids) * 100
logger.info(f"{problems}% problems")
try:
row = ' '.join([sql.select, sql.body, sql.where_id])
sessb.execute(text(row), {'id': id})
except:
problem_ids.append(id)
problem_ids = ','.join([f"'{id}'" for id in problem_ids])
logger.info(f"and u.ID not in ({problem_ids})")
exit(0)
sql_ = [sql.select, sql.body]
if not ALL_BOOKS: sql_ += [sql.just_psych]
sql_ = ' '.join(sql_)
with session('books') as sessb:
df = pd.read_sql(sql_, sessb.bind)
df = df.drop_duplicates(['Title', 'Author'])
logger.info(f"n_books before cleanup {df.shape[0]}")
logger.info("Removing HTML")
broken = '(\?\?\?|\#\#\#)' # russian / other FIXME better way to handle
df = df[~(df.Title + df.descr).str.contains(broken)] \
.drop_duplicates(['Title', 'Author']) # TODO reconsider
df['descr'] = cleantext.multiple(df.descr.tolist(), [
cleantext.strip_html,
cleantext.fix_punct,
cleantext.only_ascii,
cleantext.multiple_whitespace,
cleantext.unmark
])
# books = books[books.clean.apply(lambda x: detect(x) == 'en')]
logger.info(f"n_books after cleanup {df.shape[0]}")
df = df.rename(columns=dict(
ID='id',
descr='text',
Title='title',
Author='author',
topic_descr='topic',
))
# drop dupes, keep longest desc
df['txt_len'] = df.text.str.len()
df = df.sort_values('txt_len', ascending=False)\
.drop_duplicates('id')\
.drop(columns=['txt_len'])\
.sort_values('id')
df['thumbs'] = 0
logger.info(f"Saving books to DB")
df.to_sql('books', sess.bind, index=False, chunksize=500, if_exists='append', method='multi')
return df.set_index('id', drop=False)
create_database(vars.DB_FULL)
import pytest, time
from box import Box
from fastapi.testclient import TestClient
from lorem_text import lorem
import common.database as D
import common.models as M
from common.fixtures import fixtures
from app.main import app
import logging
logger = logging.getLogger(__name__)
with D.session() as sess:
for t in """
bookshelf
cache_entries
cache_users
entries
entries_tags
field_entries
fields
jobs
machines
notes
people
shares
shares_tags
tags
def db_books():
with D.session('books') as sess:
yield sess
def load_df(self):
if exists(paths.df):
logger.info("Load books.df")
self.df = pd.read_feather(paths.df)\
.drop(columns=['index'])\
.set_index('id', drop=False)
return
# invalidate embeddings, they're out of sync
try:
os.remove(paths.vecs)
except:
pass
logger.info("Load books MySQL")
# 58fbd36a: limit to psychology topics
sql = f"""
select u.ID, u.Title, u.Author, d.descr, t.topic_descr
from updated u
inner join description d on d.md5=u.MD5
inner join topics t on u.Topic=t.topic_id
-- later more languages; but I think it's only Russian in Libgen?
and t.lang='en'
where u.Language = 'English'
-- Make sure there's some content to work with
and length(d.descr) > 200 and length(u.Title) > 1
"""
with session('books') as sessb:
df = pd.read_sql(sql, sessb.bind)
df = df.rename(columns=dict(
ID='id',
descr='text',
Title='title',
Author='author',
topic_descr='topic',
))
logger.info(f"n_books before cleanup {df.shape[0]}")
logger.info("Remove HTML")
# some books are literally just ########
df = df[~(df.title + df.text).str.contains('(\?\?\?|\#\#\#)')]
df['text'] = CleanText(df.text.tolist())\
.strip_html()\
.only_ascii()\
.multiple_whitespace()\
.value()
df['txt_len'] = df.text.str.len()
# Ensure has content. Drop dupes, keeping those w longest description
df = df[df.txt_len > 150]\
.sort_values('txt_len', ascending=False)\
.drop_duplicates('id')\
.drop_duplicates(['title', 'author'])\
.drop(columns=['txt_len'])
# books = books[books.clean.apply(lambda x: detect(x) == 'en')]
logger.info(f"n_books after cleanup {df.shape[0]}")
logger.info(f"Save books.df")
# Error: feather does not support serializing a non-default index for the index; you can .reset_index() to make the index into column(s)
# I get ^ even though no index has yet been set. Have to manually reset_index() anyway
df = df.reset_index()
df.to_feather(paths.df)
# call self, which returns newly-saved df (ensures consistent order, etc)
self.load_df()
def influencers_(user_id):
logging.info("Influencers")
with session() as sess:
fes = pd.read_sql("""
-- remove duplicates, use average. FIXME find the dupes bug
with fe_clean as (
select field_id, created_at::date, avg(value) as value
from field_entries
group by field_id, created_at::date
),
-- ensure enough data
fe_ct as (
select field_id from fe_clean
group by field_id having count(value) > 5
)
select
fe.created_at, -- index
fe.field_id::text, -- column, uuid->string
fe.value -- value
from fe_clean fe
inner join fe_ct on fe_ct.field_id=fe.field_id -- just removes rows
inner join fields f on f.id=fe.field_id
where f.user_id=%(uid)s
and f.excluded_at is null
order by fe.created_at asc
""", sess.bind, params={'uid': user_id})
if not fes.size: return None # not enough entries
params = dict(
uid=user_id,
fids=tuple(fes.field_id.unique().tolist())
)
fs = pd.read_sql("""
select id::text, default_value, default_value_value
from fields
where user_id=%(uid)s
and id in %(fids)s
and excluded_at is null
""", sess.bind, params=params)
fs = {r.id: r for i, r in fs.iterrows()}
# Easier pivot debugging
# fields['field_id'] = fields.field_id.apply(lambda x: x[0:4])
fes = fes.pivot(index='created_at', columns='field_id', values='value')
# fes = fes.resample('D')
cols = fes.columns.tolist()
hypers = hyperopt(fes, fs, user_id)
xgb_args = {} # {'tree_method': 'gpu_hist', 'gpu_id': 0}
next_preds = {}
importances = {}
all_imps = []
for t in cols:
# remove up until they start tracking; we'll impute from there on up
fvi = fes[t].first_valid_index()
fes_ = impute_and_roll(fes[fvi:].copy(), fs)
### Next Preds
### ----------
# For next-pred, we keep target column. Yes, likely most predictive; but a rolling
# trend is important info
X = fes_
y = X[t]
model = XGBRegressor(**xgb_args, **hypers)
model.fit(X, y)
preds = model.predict(X.iloc[-1:])
next_preds[t] = float(preds[0])
# model.fit(X, y) # what's this? was I fitting twice?
### Importances
### -----------
X = fes_.drop(columns=[t])
y = fes_[t]
model = XGBRegressor(**xgb_args, **hypers)
model.fit(X, y)
imps = [float(x) for x in model.feature_importances_]
# FIXME
# /xgboost/sklearn.py:695: RuntimeWarning: invalid value encountered in true_divide return all_features / all_features.sum()
# I think this is due to target having no different value, in which case
# just leave like this.
imps = [0. if np.isnan(imp) else imp for imp in imps]
# put target col back in
imps.insert(cols.index(t), 0.0)
dict_ = dict(zip(cols, imps))
all_imps.append(dict_)
importances[t] = dict_
all_imps = dict(pd.DataFrame(all_imps).mean())
return next_preds, importances, all_imps
def nlp_on_rows(method='entries'):
for_entries = method == 'entries' # else is_profile
with session() as sess:
if for_entries:
rows = sess.query(M.Entry) \
.filter(
func.length(M.Entry.text) > 64,
M.Entry.no_ai.isnot(True),
M.Entry.ai_ran.isnot(True)
)
else:
rows = sess.query(M.User) \
.filter(
func.length(M.User.bio) > 32,
M.User.ai_ran.isnot(True)
)
rows = rows.all()
if not rows: return {}
paras_grouped = []
uids = set()
for r in rows:
txt = r.text if for_entries \
else r.bio # r.profile_to_text() # TODO profile_to_text adds people
paras_grouped.append(cleantext.markdown_split_paragraphs([txt]))
if for_entries:
uids.add(r.user_id)
paras_flat = [p for paras in paras_grouped for p in paras]
fkeys = [r.title for r in rows] \
if for_entries else [r.email for r in rows]
fixt = fixtures.load_nlp_rows(fkeys, method=method)
if fixt:
if for_entries:
clean_txt, embeds, titles, texts = fixt
else:
clean_txt, embeds = fixt
else:
clean_txt = cleantext.keywords(
paras_flat, postags=['NOUN', 'ADJ', 'VERB', 'PROPN'])
embeds = nlp_.sentence_encode(paras_flat).tolist()
if for_entries:
titles = nlp_.summarization(paras_grouped,
min_length=5,
max_length=20,
with_sentiment=False)
texts = nlp_.summarization(paras_grouped,
min_length=30,
max_length=250)
for i, r in enumerate(rows):
CM = M.CacheEntry if for_entries else M.CacheUser
c = sess.query(CM).get(r.id)
if not c:
c = CM(entry_id=r.id) if for_entries else CM(user_id=r.id)
sess.add(c)
# Save the cache_entry (paras,clean,vectors)
paras = paras_grouped[i]
c.paras = paras
ct = len(paras)
c.clean = [' '.join(e) for e in clean_txt[:ct]]
c.vectors = embeds[:ct]
sess.commit()
# Save the fixture for later
fixt = (clean_txt[:ct], embeds[:ct], titles[i], texts[i]) \
if for_entries else (clean_txt[:ct], embeds[:ct])
fixt_k = r.title if for_entries else r.email
fixtures.save_nlp_row(fixt_k, fixt, method=method)
clean_txt, embeds = clean_txt[ct:], embeds[ct:]
if for_entries:
r.title_summary = titles[i]["summary"]
r.text_summary = texts[i]["summary"]
r.sentiment = texts[i]["sentiment"]
r.ai_ran = True
sess.commit()
if for_entries:
# 9131155e: only update every x entries
M.Job.multiple_book_jobs(list(uids))
return {}
