def _transform(self, dfs):
icustay_index = dfs['icustay_detail'].set_index('icustay_id')
for df_name in ['chartevents', 'labevents', 'ioevents']:
df = dfs[df_name]
df['icustay_intime'] = icustay_index.loc[df['icustay_id'],
'icustay_intime'].values
df = df[df['charttime'] > df['icustay_intime']]
dfs[df_name] = df
demo_history = DemographicTransform().transform(dfs['icustay_detail'])
chart_history = ChartTransform().transform(dfs['chartevents'])
lab_history = LabTransform().transform(dfs['labevents'])
urine_history = UrineTransform().transform(dfs['ioevents'])
all_history = from_chunks(
[chart_history, lab_history, demo_history, urine_history])
all_history['charttime'] = big_dt_to_num(all_history['charttime'])
all_history = all_history.sort_index(
by=['subject_id', 'charttime']).reset_index(drop=True)
all_history['charttime'] = num_to_big_dt(all_history['charttime'])
all_history.drop_duplicates(inplace=True)
all_history = all_history[[
'subject_id', 'charttime', 'category', 'valuenum'
]]
return all_history
def _transform(self, df):
df['days'] = df['end_date'] - df['start_date'] + pd.Timedelta(days=1)
invalid_range_mask = df[
'days'] < 1 #remove rows with invalid (negative) deltas
df = df[~invalid_range_mask]
max_days = ceil(df['days'].max().days)
df['date'] = df['start_date']
del df['start_date'], df['end_date']
expanded = [df]
expand_per_row = False
for i in range(2, max_days):
prev = expanded[-1]
rows_as_long_as = prev[prev['days'] >= pd.Timedelta(days=i)]
rows_as_long_as.loc[:, 'date'] += pd.Timedelta(days=1)
# if there are only a relative few rows remaining with long date ranges,
# it will be more efficient to expand them per-row rather than per-date-range
if len(rows_as_long_as) < max_days / len(
rows_as_long_as.columns)**2:
expand_per_row = True
break
expanded.append(rows_as_long_as)
#expand remaining rows per-row rather than per-date-range
if expand_per_row:
rala = rows_as_long_as
# represent any categorical columns as ints for fast expansion
rala_cat = rala.select_dtypes(include=['category'])
rala_categories = {}
for col in rala_cat.columns:
rala_categories[col] = rala[col].cat.categories
rala[col] = rala[col].cat.codes
# for each row, generate expanded rows as remaining days in the date range
for j, row in rala.iterrows():
date = row['date']
n_days = row['days'].days - i
expanded_row = pd.DataFrame(index=np.arange(n_days + 1),
columns=row.index)
for col in expanded_row.columns:
expanded_row[col] = row[col]
expanded_row['date'] = pd.date_range(
date, date + pd.Timedelta(days=n_days))
#convert numeric back to categorical
for col, categories in rala_categories.items():
expanded_row[col] = pd.Categorical.from_codes(
codes=expanded_row[col],
categories=categories,
name=col)
expanded.append(expanded_row)
# merge all extra rows
df = from_chunks(expanded)
return df
def _transform(self, df):
df['days'] = df['end_date'] - df['start_date'] + pd.Timedelta(days=1)
invalid_range_mask = df['days'] < 1 #remove rows with invalid (negative) deltas
df = df[~invalid_range_mask]
max_days = ceil(df['days'].max().days)
df['date'] = df['start_date']
del df['start_date'], df['end_date']
expanded = [df]
expand_per_row = False
for i in range(2, max_days):
prev = expanded[-1]
rows_as_long_as = prev[prev['days'] >= pd.Timedelta(days=i)]
rows_as_long_as.loc[:, 'date'] += pd.Timedelta(days=1)
# if there are only a relative few rows remaining with long date ranges,
# it will be more efficient to expand them per-row rather than per-date-range
if len(rows_as_long_as) < max_days / len(rows_as_long_as.columns)**2:
expand_per_row = True
break
expanded.append(rows_as_long_as)
#expand remaining rows per-row rather than per-date-range
if expand_per_row:
rala = rows_as_long_as
# represent any categorical columns as ints for fast expansion
rala_cat = rala.select_dtypes(include=['category'])
rala_categories = {}
for col in rala_cat.columns:
rala_categories[col] = rala[col].cat.categories
rala[col] = rala[col].cat.codes
# for each row, generate expanded rows as remaining days in the date range
for j, row in rala.iterrows():
date = row['date']
n_days = row['days'].days - i
expanded_row = pd.DataFrame(index=np.arange(n_days + 1), columns=row.index)
for col in expanded_row.columns:
expanded_row[col] = row[col]
expanded_row['date'] = pd.date_range(date, date+pd.Timedelta(days=n_days))
#convert numeric back to categorical
for col, categories in rala_categories.items():
expanded_row[col] = pd.Categorical.from_codes(
codes=expanded_row[col],
categories=categories,
name=col
)
expanded.append(expanded_row)
# merge all extra rows
df = from_chunks(expanded)
return df
def _transform(self, data):
start = time.time()
print('transforming data of size',
data.memory_usage(index=True).sum(), 'bytes')
store_chunks_jobs = []
transform_jobs = []
hdf_stores = []
try:
print('splitting data into large groups')
group_iter = self._group_iter(
data,
len(data) // self.n_jobs or self.chunksize)
for group_data in group_iter:
if group_data.empty:
continue
f = temp_file.make_temporary_file()
hdf_store = HdfDataStore(self.input_schema(), f)
hdf_stores.append(hdf_store)
hdf_store.store(group_data)
store_chunks_jobs.append(
joblib.delayed(self.store_chunks_job)(hdf_store))
#transform_jobs.append(joblib.delayed(self.transform_job)(hdf_store))
print('breaking data into chunks in parallel')
joblib.Parallel(n_jobs=self.n_jobs)(store_chunks_jobs)
chunk_stores = chain.from_iterable(store.chunk_stores()
for store in hdf_stores)
transform_jobs = [
joblib.delayed(self.transform_job)(chunk_store)
for chunk_store in chunk_stores
]
print('running transforms in', len(transform_jobs),
'parallel jobs')
result_hdf_stores = joblib.Parallel(
n_jobs=self.n_jobs)(transform_jobs)
print('loading and merging the results')
results = from_chunks(r_hdf_store.load()
for r_hdf_store in result_hdf_stores)
print('finished merge')
finally:
for hdf_store in hdf_stores:
hdf_store.delete_chunks()
hdf_store.delete()
end = time.time()
print('took', end - start, 'seconds to transform all data in parallel')
return results
def _transform(self, tables):
# filtering carevue ioevents
io_carevue = VasoIOEventsCarevue().transform(tables)
# filtering metavision ioevents
io_metavision = VasoIOEventsMetavision().transform(tables)
# concatenating ioevents
ioevents = from_chunks([io_carevue, io_metavision])
# getting vaso day counts
vaso_days = VasoDayCounts().transform(ioevents)
return vaso_days
def _transform(self, tables):
# filtering carevue ioevents
io_carevue = VasoIOEventsCarevue().transform(tables)
# filtering metavision ioevents
io_metavision = VasoIOEventsMetavision().transform(tables)
# concatenating ioevents
ioevents = from_chunks([io_carevue, io_metavision])
# getting vaso day counts
vaso_days = VasoDayCounts().transform(ioevents)
return vaso_days
def load(self):
seq = odo.odo(self.odo_target, odo.chunks(pandas.DataFrame),
chunksize=CHUNK_SIZE,
dshape=schema_to_dshape(self.schema))
def conv_chunks(chunks):
for chunk in chunks:
print('typechecking a chunk')
self.schema.conform_df(chunk, skip_sort=True)
yield chunk
print('concatenating df chunks')
df = from_chunks(conv_chunks(seq))
return df
def load(self):
seq = odo.odo(self.odo_target,
odo.chunks(pandas.DataFrame),
chunksize=CHUNK_SIZE,
dshape=schema_to_dshape(self.schema))
def conv_chunks(chunks):
for chunk in chunks:
print('typechecking a chunk')
self.schema.conform_df(chunk, skip_sort=True)
yield chunk
print('concatenating df chunks')
df = from_chunks(conv_chunks(seq))
return df
def _transform(self, data):
start = time.time()
print("transforming data of size", data.memory_usage(index=True).sum(), "bytes")
store_chunks_jobs = []
transform_jobs = []
hdf_stores = []
try:
print("splitting data into large groups")
group_iter = self._group_iter(data, len(data) // self.n_jobs or self.chunksize)
for group_data in group_iter:
if group_data.empty:
continue
f = temp_file.make_temporary_file()
hdf_store = HdfDataStore(self.input_schema(), f)
hdf_stores.append(hdf_store)
hdf_store.store(group_data)
store_chunks_jobs.append(joblib.delayed(self.store_chunks_job)(hdf_store))
# transform_jobs.append(joblib.delayed(self.transform_job)(hdf_store))
print("breaking data into chunks in parallel")
joblib.Parallel(n_jobs=self.n_jobs)(store_chunks_jobs)
chunk_stores = chain.from_iterable(store.chunk_stores() for store in hdf_stores)
transform_jobs = [joblib.delayed(self.transform_job)(chunk_store) for chunk_store in chunk_stores]
print("running transforms in", len(transform_jobs), "parallel jobs")
result_hdf_stores = joblib.Parallel(n_jobs=self.n_jobs)(transform_jobs)
print("loading and merging the results")
results = from_chunks(r_hdf_store.load() for r_hdf_store in result_hdf_stores)
print("finished merge")
finally:
for hdf_store in hdf_stores:
hdf_store.delete_chunks()
hdf_store.delete()
end = time.time()
print("took", end - start, "seconds to transform all data in parallel")
return results
def _transform(self, dfs):
icustay_index = dfs['icustay_detail'].set_index('icustay_id')
for df_name in ['chartevents', 'labevents', 'ioevents']:
df = dfs[df_name]
df['icustay_intime'] = icustay_index.loc[df['icustay_id'], 'icustay_intime'].values
df = df[df['charttime'] > df['icustay_intime']]
dfs[df_name] = df
demo_history = DemographicTransform().transform(dfs['icustay_detail'])
chart_history = ChartTransform().transform(dfs['chartevents'])
lab_history = LabTransform().transform(dfs['labevents'])
urine_history = UrineTransform().transform(dfs['ioevents'])
all_history = from_chunks([chart_history, lab_history, demo_history, urine_history])
all_history['charttime'] = big_dt_to_num(all_history['charttime'])
all_history = all_history.sort_index(by=['subject_id', 'charttime']).reset_index(drop=True)
all_history['charttime'] = num_to_big_dt(all_history['charttime'])
all_history.drop_duplicates(inplace=True)
all_history = all_history[['subject_id', 'charttime', 'category', 'valuenum']]
return all_history
def _transform(self, dfs):
icustay_index = dfs["icustay_detail"].set_index("icustay_id")
for df_name in ["chartevents", "labevents", "ioevents"]:
df = dfs[df_name]
df["icustay_intime"] = icustay_index.loc[df["icustay_id"], "icustay_intime"].values
df = df[df["charttime"] > df["icustay_intime"]]
dfs[df_name] = df
demo_history = DemographicTransform().transform(dfs["icustay_detail"])
chart_history = ChartTransform().transform(dfs["chartevents"])
lab_history = LabTransform().transform(dfs["labevents"])
urine_history = UrineTransform().transform(dfs["ioevents"])
all_history = from_chunks([chart_history, lab_history, demo_history, urine_history])
all_history["charttime"] = big_dt_to_num(all_history["charttime"])
all_history = all_history.sort_index(by=["subject_id", "charttime"]).reset_index(drop=True)
all_history["charttime"] = num_to_big_dt(all_history["charttime"])
all_history.drop_duplicates(inplace=True)
all_history = all_history[["subject_id", "charttime", "category", "valuenum"]]
return all_history
from chatto_transform.config import config
from chatto_transform.schema.ss.ss_sql_raw_schema import appointments
from chatto_transform.datastores.hdf_datastore import HdfDataStore
from chatto_transform.lib.chunks import from_chunks
import time
ds = HdfDataStore(appointments, config.data_dir + 'test.hdf')
chunks = ds.load_chunks()
start = time.time()
df = from_chunks(chunks)
end = time.time()
print('took', end - start, 'seconds to load and concatenate all data')
from chatto_transform.config import config
from chatto_transform.schema.ss.ss_sql_raw_schema import appointments
from chatto_transform.datastores.hdf_datastore import HdfDataStore
from chatto_transform.lib.chunks import from_chunks
import time
ds = HdfDataStore(appointments, config.data_dir+'test.hdf')
chunks = ds.load_chunks()
start = time.time()
df = from_chunks(chunks)
end = time.time()
print('took', end - start, 'seconds to load and concatenate all data')
