def test_inspect_datagen(tmpdir, datasets, engine, dist):
# Dataset
paths = glob.glob(str(datasets[engine]) + "/*." + engine.split("-")[0])
# Dataset columns type config
columns_dict = {}
columns_dict["cats"] = ["name-cat", "name-string"
] if engine == "parquet" else ["name-string"]
columns_dict["conts"] = ["x", "y"]
columns_dict["labels"] = ["label"]
# Create inspector and inspect
output_inspect1 = tmpdir + "/dataset_info1.json"
dataset = Dataset(paths, engine=engine)
a = datains.DatasetInspector()
a.inspect(dataset, columns_dict, output_inspect1)
assert os.path.isfile(output_inspect1)
# Generate dataset using data_gen tool
output_datagen = tmpdir + "/datagen"
os.mkdir(output_datagen)
with fsspec.open(output_inspect1) as f:
output1 = json.load(f)
cols = datagen._get_cols_from_schema(output1)
if dist == "uniform":
df_gen = datagen.DatasetGen(datagen.UniformDistro(), gpu_frac=0.00001)
else:
df_gen = datagen.DatasetGen(datagen.PowerLawDistro(0.1),
gpu_frac=0.00001)
output_datagen_files = df_gen.full_df_create(output1["num_rows"],
cols,
entries=True,
output=output_datagen)
# Inspect again and check output are the same
output_inspect2 = tmpdir + "/dataset_info2.json"
dataset = Dataset(output_datagen_files, engine=engine)
a.inspect(dataset, columns_dict, output_inspect2)
assert os.path.isfile(output_inspect2)
# Compare json outputs
with fsspec.open(output_inspect2) as f:
output2 = json.load(f)
for k1 in output1.keys():
if k1 == "num_rows":
assert output1[k1] == output2[k1]
else:
for k2 in output1[k1].keys():
for k3 in output1[k1][k2].keys():
if k3 == "dtype":
if output1[k1][k2][k3] == "object":
assert (output1[k1][k2][k3] == output2[k1][k2][k3]
or output2[k1][k2][k3] == "int64")
else:
assert output1[k1][k2][k3] == output2[k1][k2][k3]
else:
assert output1[k1][k2][k3] == pytest.approx(
output2[k1][k2][k3], rel=1e-0, abs=1e-0)
class FileItrDataset(torch.utils.data.IterableDataset):
gpu_itr = None
def __init__(self, file, **kwargs):
columns = kwargs.pop("columns", None)
self.gpu_itr = Dataset(file, **kwargs).to_iter(columns=columns)
def __iter__(self):
return self.gpu_itr.__iter__()
def __len__(self):
return len(self.gpu_itr)
def __init__(self,
path,
sub_batch_size=1,
cats=None,
conts=None,
labels=None,
pin_memory=False,
**kwargs):
self.apply_ops = kwargs.get("apply_ops", False)
self.cat_cols = cats
self.cont_cols = conts
self.label_cols = labels
self.itr = Dataset(path,
**kwargs).to_iter(columns=cats + conts + labels)
self.batch_size = sub_batch_size
self.num_chunks = len(self.itr)
def main(args):
# Get device configuration
device_size = device_mem_size(kind="total")
device_limit = int(args.device_limit_frac * device_size)
device_pool_size = int(args.device_pool_frac * device_size)
part_size = int(args.part_mem_frac * device_size)
# Get dataset columns
with fsspec.open(args.config_file) as f:
config = json.load(f)
# Create Dataset
dataset = Dataset(args.data_path, engine=args.format, part_size=part_size)
# Call Inspector
with managed_client(args.devices, device_limit, args.protocol) as client:
setup_rmm_pool(client, device_pool_size)
a = datains.DatasetInspector(client)
a.inspect(dataset, config, args.output_file)
def test_full_df(num_rows, tmpdir, distro):
json_sample["num_rows"] = num_rows
cats = list(json_sample["cats"].keys())
cols = datagen._get_cols_from_schema(json_sample, distros=distro)
df_gen = datagen.DatasetGen(datagen.UniformDistro(), gpu_frac=0.00001)
df_files = df_gen.full_df_create(num_rows, cols, entries=True, output=tmpdir)
test_size = 0
full_df = _make_df()
for fi in df_files:
df = Dataset(fi).to_ddf().compute()
test_size = test_size + df.shape[0]
full_df = _concat([full_df, df])
assert test_size == num_rows
conts_rep = cols["conts"]
cats_rep = cols["cats"]
labels_rep = cols["labels"]
assert df.shape[1] == len(conts_rep) + len(cats_rep) + len(labels_rep)
for idx, cat in enumerate(cats[1:]):
dist = cats_rep[idx + 1].distro or df_gen.dist
if HAS_GPU:
if not _is_string_dtype(full_df[cat]._column):
sts, ps = dist.verify(full_df[cat].to_pandas())
assert all(s > 0.9 for s in sts)
else:
if not _is_string_dtype(full_df[cat]):
sts, ps = dist.verify(full_df[cat])
assert all(s > 0.9 for s in sts)
# these are not mh series
assert full_df[cat].nunique() == cats_rep[0].cardinality
assert full_df[cat].str.len().min() == cats_rep[0].min_entry_size
assert full_df[cat].str.len().max() == cats_rep[0].max_entry_size
# check the mh list here cat 0 only
if HAS_GPU:
check_ser = _make_df(list(full_df[cats[0]]._column.elements.values_host))[0]
else:
check_ser = _pull_apart_list(full_df[cats[0]])[0]
assert check_ser.nunique() == cats_rep[0].cardinality
assert check_ser.str.len().min() == cats_rep[0].min_entry_size
assert check_ser.str.len().max() == cats_rep[0].max_entry_size
proc.add_feature([ZeroFill(replace=True), LogOp(replace=True)])
proc.add_preprocess(Normalize(replace=True))
if int(args.freq_thresh) == 0:
proc.add_preprocess(Categorify(replace=True, out_path=args.out_dir))
else:
proc.add_preprocess(
Categorify(
replace=True,
use_frequency=True,
freq_threshold=int(args.freq_thresh),
out_path=args.out_dir,
))
print("Creating Dataset Iterator")
dataset_args = {"sep": "\t"} if args.in_file_type == "csv" else {}
trains_ds = Dataset(train_set,
engine=args.in_file_type,
part_mem_fraction=float(args.gpu_mem_frac),
**dataset_args)
valids_ds = Dataset(valid_set,
engine=args.in_file_type,
part_mem_fraction=float(args.gpu_mem_frac),
**dataset_args)
print("Running apply")
out_train = os.path.join(args.out_dir, "train")
out_valid = os.path.join(args.out_dir, "valid")
start = time()
proc.apply(
trains_ds,
apply_offline=True,
record_stats=True,
def test_gpu_preproc(tmpdir, df, dataset, dump, gpu_memory_frac, engine,
preprocessing):
cat_names = ["name-cat", "name-string"
] if engine == "parquet" else ["name-string"]
cont_names = ["x", "y", "id"]
label_name = ["label"]
processor = nvt.Workflow(cat_names=cat_names,
cont_names=cont_names,
label_name=label_name)
processor.add_feature(
[ops.FillMedian(),
ops.LogOp(preprocessing=preprocessing)])
processor.add_preprocess(ops.Normalize())
processor.add_preprocess(ops.Categorify())
processor.finalize()
processor.update_stats(dataset)
if dump:
config_file = tmpdir + "/temp.yaml"
processor.save_stats(config_file)
processor.clear_stats()
processor.load_stats(config_file)
def get_norms(tar: cudf.Series):
ser_median = tar.dropna().quantile(0.5, interpolation="linear")
gdf = tar.fillna(ser_median)
gdf = np.log(gdf + 1)
return gdf
# Check mean and std - No good right now we have to add all other changes; Zerofill, Log
x_col = "x" if preprocessing else "x_LogOp"
y_col = "y" if preprocessing else "y_LogOp"
assert math.isclose(get_norms(df.x).mean(),
processor.stats["means"][x_col],
rel_tol=1e-2)
assert math.isclose(get_norms(df.y).mean(),
processor.stats["means"][y_col],
rel_tol=1e-2)
assert math.isclose(get_norms(df.x).std(),
processor.stats["stds"][x_col],
rel_tol=1e-2)
assert math.isclose(get_norms(df.y).std(),
processor.stats["stds"][y_col],
rel_tol=1e-2)
# Check median (TODO: Improve the accuracy)
x_median = df.x.dropna().quantile(0.5, interpolation="linear")
y_median = df.y.dropna().quantile(0.5, interpolation="linear")
id_median = df.id.dropna().quantile(0.5, interpolation="linear")
assert math.isclose(x_median, processor.stats["medians"]["x"], rel_tol=1e1)
assert math.isclose(y_median, processor.stats["medians"]["y"], rel_tol=1e1)
assert math.isclose(id_median,
processor.stats["medians"]["id"],
rel_tol=1e1)
# Check that categories match
if engine == "parquet":
cats_expected0 = df["name-cat"].unique().values_host
cats0 = get_cats(processor, "name-cat")
assert cats0.tolist() == [None] + cats_expected0.tolist()
cats_expected1 = df["name-string"].unique().values_host
cats1 = get_cats(processor, "name-string")
assert cats1.tolist() == [None] + cats_expected1.tolist()
# Write to new "shuffled" and "processed" dataset
processor.write_to_dataset(tmpdir,
dataset,
nfiles=10,
shuffle=True,
apply_ops=True)
processor.create_final_cols()
# if preprocessing
if not preprocessing:
for col in cont_names:
assert f"{col}_LogOp" in processor.columns_ctx["final"]["cols"][
"continuous"]
dlc = torch_dataloader.DLCollator(preproc=processor, apply_ops=False)
data_files = [
torch_dataloader.FileItrDataset(x,
use_row_groups=True,
gpu_memory_frac=gpu_memory_frac,
names=allcols_csv)
for x in glob.glob(str(tmpdir) + "/*.parquet")
]
data_itr = torch.utils.data.ChainDataset(data_files)
dl = torch_dataloader.DLDataLoader(data_itr,
collate_fn=dlc.gdf_col,
pin_memory=False,
num_workers=0)
len_df_pp = 0
for chunk in dl:
len_df_pp += len(chunk[0][0])
dataset = Dataset(glob.glob(str(tmpdir) + "/*.parquet"),
part_mem_fraction=gpu_memory_frac)
x = processor.ds_to_tensors(dataset.to_iter(), apply_ops=False)
num_rows, num_row_groups, col_names = cudf.io.read_parquet_metadata(
str(tmpdir) + "/_metadata")
assert len(x[0]) == len_df_pp
itr_ds = torch_dataloader.TensorItrDataset([x[0], x[1], x[2]],
batch_size=512000)
count_tens_itr = 0
for data_gd in itr_ds:
count_tens_itr += len(data_gd[1])
assert data_gd[0].shape[1] > 0
assert data_gd[1].shape[1] > 0
assert len_df_pp == count_tens_itr
def test_gpu_workflow_api(tmpdir, client, df, dataset, gpu_memory_frac, engine,
dump, op_columns, use_client):
cat_names = ["name-cat", "name-string"
] if engine == "parquet" else ["name-string"]
cont_names = ["x", "y", "id"]
label_name = ["label"]
processor = nvt.Workflow(
cat_names=cat_names,
cont_names=cont_names,
label_name=label_name,
client=client if use_client else None,
)
processor.add_feature([ops.ZeroFill(columns=op_columns), ops.LogOp()])
processor.add_preprocess(ops.Normalize())
processor.add_preprocess(ops.Categorify(cat_cache="host"))
processor.finalize()
processor.update_stats(dataset)
if dump:
config_file = tmpdir + "/temp.yaml"
processor.save_stats(config_file)
processor.clear_stats()
processor.load_stats(config_file)
def get_norms(tar: cudf.Series):
gdf = tar.fillna(0)
gdf = gdf * (gdf >= 0).astype("int")
gdf = np.log(gdf + 1)
return gdf
# Check mean and std - No good right now we have to add all other changes; Zerofill, Log
if not op_columns:
assert math.isclose(get_norms(df.y).mean(),
processor.stats["means"]["y"],
rel_tol=1e-1)
assert math.isclose(get_norms(df.y).std(),
processor.stats["stds"]["y"],
rel_tol=1e-1)
assert math.isclose(get_norms(df.x).mean(),
processor.stats["means"]["x"],
rel_tol=1e-1)
assert math.isclose(get_norms(df.x).std(),
processor.stats["stds"]["x"],
rel_tol=1e-1)
# Check that categories match
if engine == "parquet":
cats_expected0 = df["name-cat"].unique().values_host
cats0 = get_cats(processor, "name-cat")
# adding the None entry as a string because of move from gpu
assert cats0.tolist() == [None] + cats_expected0.tolist()
cats_expected1 = df["name-string"].unique().values_host
cats1 = get_cats(processor, "name-string")
# adding the None entry as a string because of move from gpu
assert cats1.tolist() == [None] + cats_expected1.tolist()
# Write to new "shuffled" and "processed" dataset
processor.write_to_dataset(tmpdir,
dataset,
out_files_per_proc=10,
shuffle="partial",
apply_ops=True)
dataset_2 = Dataset(glob.glob(str(tmpdir) + "/*.parquet"),
part_mem_fraction=gpu_memory_frac)
df_pp = cudf.concat(list(dataset_2.to_iter()), axis=0)
if engine == "parquet":
assert is_integer_dtype(df_pp["name-cat"].dtype)
assert is_integer_dtype(df_pp["name-string"].dtype)
num_rows, num_row_groups, col_names = cudf.io.read_parquet_metadata(
str(tmpdir) + "/_metadata")
assert num_rows == len(df_pp)
def test_gpu_workflow_config(tmpdir, client, df, dataset, gpu_memory_frac,
engine, dump, replace):
cat_names = ["name-cat", "name-string"
] if engine == "parquet" else ["name-string"]
cont_names = ["x", "y", "id"]
label_name = ["label"]
config = nvt.workflow.get_new_config()
# add operators with dependencies
config["FE"]["continuous"] = [[
ops.FillMissing(replace=replace),
ops.LogOp(replace=replace)
]]
config["PP"]["continuous"] = [[
ops.LogOp(replace=replace),
ops.Normalize()
]]
config["PP"]["categorical"] = [ops.Categorify()]
processor = nvt.Workflow(
cat_names=cat_names,
cont_names=cont_names,
label_name=label_name,
config=config,
client=client,
)
processor.update_stats(dataset)
if dump:
config_file = tmpdir + "/temp.yaml"
processor.save_stats(config_file)
processor.clear_stats()
processor.load_stats(config_file)
def get_norms(tar: cudf.Series):
ser_median = tar.dropna().quantile(0.5, interpolation="linear")
gdf = tar.fillna(ser_median)
gdf = np.log(gdf + 1)
return gdf
# Check mean and std - No good right now we have to add all other changes; Zerofill, Log
concat_ops = "_FillMissing_LogOp"
if replace:
concat_ops = ""
assert math.isclose(get_norms(df.x).mean(),
processor.stats["means"]["x" + concat_ops],
rel_tol=1e-1)
assert math.isclose(get_norms(df.y).mean(),
processor.stats["means"]["y" + concat_ops],
rel_tol=1e-1)
assert math.isclose(get_norms(df.x).std(),
processor.stats["stds"]["x" + concat_ops],
rel_tol=1e-1)
assert math.isclose(get_norms(df.y).std(),
processor.stats["stds"]["y" + concat_ops],
rel_tol=1e-1)
# Check that categories match
if engine == "parquet":
cats_expected0 = df["name-cat"].unique().values_host
cats0 = get_cats(processor, "name-cat")
# adding the None entry as a string because of move from gpu
assert cats0.tolist() == [None] + cats_expected0.tolist()
cats_expected1 = df["name-string"].unique().values_host
cats1 = get_cats(processor, "name-string")
# adding the None entry as a string because of move from gpu
assert cats1.tolist() == [None] + cats_expected1.tolist()
# Write to new "shuffled" and "processed" dataset
processor.write_to_dataset(tmpdir,
dataset,
out_files_per_proc=10,
shuffle="partial",
apply_ops=True)
dataset_2 = Dataset(glob.glob(str(tmpdir) + "/*.parquet"),
part_mem_fraction=gpu_memory_frac)
df_pp = cudf.concat(list(dataset_2.to_iter()), axis=0)
if engine == "parquet":
assert is_integer_dtype(df_pp["name-cat"].dtype)
assert is_integer_dtype(df_pp["name-string"].dtype)
num_rows, num_row_groups, col_names = cudf.io.read_parquet_metadata(
str(tmpdir) + "/_metadata")
assert num_rows == len(df_pp)
def test_gpu_workflow(tmpdir, client, df, dataset, gpu_memory_frac, engine,
dump):
cat_names = ["name-cat", "name-string"
] if engine == "parquet" else ["name-string"]
cont_names = ["x", "y", "id"]
label_name = ["label"]
config = nvt.workflow.get_new_config()
config["FE"]["continuous"] = [ops.ZeroFill()]
config["PP"]["continuous"] = [[ops.ZeroFill(), ops.Normalize()]]
config["PP"]["categorical"] = [ops.Categorify()]
processor = nvt.Workflow(
cat_names=cat_names,
cont_names=cont_names,
label_name=label_name,
config=config,
client=client,
)
processor.update_stats(dataset)
if dump:
config_file = tmpdir + "/temp.yaml"
processor.save_stats(config_file)
processor.clear_stats()
processor.load_stats(config_file)
def get_norms(tar: cudf.Series):
gdf = tar.fillna(0)
gdf = gdf * (gdf >= 0).astype("int")
return gdf
assert math.isclose(get_norms(df.x).mean(),
processor.stats["means"]["x"],
rel_tol=1e-4)
assert math.isclose(get_norms(df.y).mean(),
processor.stats["means"]["y"],
rel_tol=1e-4)
# assert math.isclose(get_norms(df.id).mean(),
# processor.stats["means"]["id_ZeroFill_LogOp"], rel_tol=1e-4)
assert math.isclose(get_norms(df.x).std(),
processor.stats["stds"]["x"],
rel_tol=1e-3)
assert math.isclose(get_norms(df.y).std(),
processor.stats["stds"]["y"],
rel_tol=1e-3)
# assert math.isclose(get_norms(df.id).std(),
# processor.stats["stds"]["id_ZeroFill_LogOp"], rel_tol=1e-3)
# Check that categories match
if engine == "parquet":
cats_expected0 = df["name-cat"].unique().values_host
cats0 = get_cats(processor, "name-cat")
# adding the None entry as a string because of move from gpu
assert cats0.tolist() == [None] + cats_expected0.tolist()
cats_expected1 = df["name-string"].unique().values_host
cats1 = get_cats(processor, "name-string")
# adding the None entry as a string because of move from gpu
assert cats1.tolist() == [None] + cats_expected1.tolist()
# Write to new "shuffled" and "processed" dataset
processor.write_to_dataset(tmpdir,
dataset,
out_files_per_proc=10,
shuffle="partial",
apply_ops=True)
dataset_2 = Dataset(glob.glob(str(tmpdir) + "/*.parquet"),
part_mem_fraction=gpu_memory_frac)
df_pp = cudf.concat(list(dataset_2.to_iter()), axis=0)
if engine == "parquet":
assert is_integer_dtype(df_pp["name-cat"].dtype)
assert is_integer_dtype(df_pp["name-string"].dtype)
num_rows, num_row_groups, col_names = cudf.io.read_parquet_metadata(
str(tmpdir) + "/_metadata")
assert num_rows == len(df_pp)
def __init__(self, file, **kwargs):
columns = kwargs.pop("columns", None)
self.gpu_itr = Dataset(file, **kwargs).to_iter(columns=columns)
def preprocess_criteo_parquet(
input_path: str,
output_path: str,
client,
frequency_threshold: int,
):
train_days = [str(x) for x in CRITEO_TRAIN_DAYS]
train_files = [
os.path.join(input_path, x) for x in os.listdir(input_path)
if x.startswith("day") and x.split(".")[0].split("_")[-1] in train_days
]
valid_file = os.path.join(input_path, "day_23.part2.parquet")
test_file = os.path.join(input_path, "day_23.part1.parquet")
all_set = train_files + [valid_file] + [test_file]
print(all_set, train_files, valid_file, test_file)
print("Creating Workflow Object")
workflow = Workflow(cat_names=CRITEO_CATEGORICAL_COLUMNS,
cont_names=CRITEO_CONTINUOUS_COLUMNS,
label_name=CRITEO_CLICK_COLUMNS)
# We want to assign 0 to all missing values, and calculate log(x+3) for present values
# so if we set missing values to -2, then the result of log(1+2+(-2)) would be 0
workflow.add_cont_feature([
FillMissing(fill_val=-2.0),
LambdaOp(op_name='Add3ButMinusOneCauseLogAddsOne',
f=lambda col, _: col.add(2.0)),
LogOp(), # Log(1+x)
])
workflow.add_cat_preprocess(
Categorify(freq_threshold=frequency_threshold, out_path=output_path))
workflow.finalize()
print("Creating Dataset Iterator")
all_ds = Dataset(all_set,
engine="parquet",
part_mem_fraction=ALL_DS_MEM_FRAC)
trains_ds = Dataset(train_files,
engine="parquet",
part_mem_fraction=TRAIN_DS_MEM_FRAC)
valid_ds = Dataset(valid_file,
engine="parquet",
part_mem_fraction=TEST_DS_MEM_FRAC)
test_ds = Dataset(test_file,
engine="parquet",
part_mem_fraction=VALID_DS_MEM_FRAC)
print("Running apply")
out_train = os.path.join(output_path, "train")
out_valid = os.path.join(output_path, "validation")
out_test = os.path.join(output_path, "test")
start = time()
workflow.update_stats(all_ds)
print(f"Gathering statistics time: {time() - start}")
start = time()
workflow.apply(trains_ds, record_stats=False, output_path=out_train)
print(f"train preprocess time: {time() - start}")
start = time()
workflow.apply(valid_ds, record_stats=False, output_path=out_valid)
print(f"valid preprocess time: {time() - start}")
start = time()
workflow.apply(test_ds, record_stats=False, output_path=out_test)
print(f"test preprocess time: {time() - start}")
save_model_size_config(workflow, output_path)
