def test_spec_set(tmpdir, client):
gdf_test = cudf.DataFrame({
"ad_id": [1, 2, 2, 6, 6, 8, 3, 3],
"source_id": [2, 4, 4, 7, 5, 2, 5, 2],
"platform": [1, 2, np.nan, 2, 1, 3, 3, 1],
"cont": [1, 2, np.nan, 2, 1, 3, 3, 1],
"clicked": [1, 0, 1, 0, 0, 1, 1, 0],
})
p = nvt.Workflow(
cat_names=["ad_id", "source_id", "platform"],
cont_names=["cont"],
label_name=["clicked"],
client=client,
)
p.add_feature(ops.FillMissing())
p.add_feature(ops.Normalize())
p.add_feature(ops.Categorify())
p.add_feature(
ops.TargetEncoding(
cat_groups=["ad_id", "source_id", "platform"],
cont_target="clicked",
kfold=5,
fold_seed=42,
p_smooth=20,
))
p.apply(nvt.Dataset(gdf_test), record_stats=True)
assert p.stats
def test_normalize(tmpdir, df, dataset, gpu_memory_frac, engine, op_columns):
cont_features = op_columns >> ops.Normalize()
processor = nvtabular.Workflow(cont_features)
processor.fit(dataset)
new_gdf = processor.transform(dataset).to_ddf().compute()
new_gdf.index = df.index # Make sure index is aligned for checks
for col in op_columns:
assert math.isclose(df[col].mean(),
processor.column_group.op.means[col],
rel_tol=1e-4)
assert math.isclose(df[col].std(),
processor.column_group.op.stds[col],
rel_tol=1e-4)
df[col] = (df[col] - processor.column_group.op.means[col]
) / processor.column_group.op.stds[col]
assert np.all((df[col] - new_gdf[col]).abs().values <= 1e-2)
# our normalize op also works on dicts of cupy/numpy tensors. make sure this works like we'd
# expect
df = dataset.compute()
cupy_inputs = {col: df[col].values for col in op_columns}
cupy_outputs = cont_features.op.transform(op_columns, cupy_inputs)
for col in op_columns:
assert np.allclose(cupy_outputs[col], new_gdf[col].values)
def test_normalize(tmpdir, df, dataset, gpu_memory_frac, engine, op_columns):
cat_names = ["name-cat", "name-string"
] if engine == "parquet" else ["name-string"]
cont_names = ["x", "y"]
label_name = ["label"]
config = nvt.workflow.get_new_config()
config["PP"]["continuous"] = [ops.Moments(columns=op_columns)]
processor = nvtabular.Workflow(cat_names=cat_names,
cont_names=cont_names,
label_name=label_name,
config=config)
processor.update_stats(dataset)
op = ops.Normalize()
columns_ctx = {}
columns_ctx["continuous"] = {}
columns_ctx["continuous"]["base"] = op_columns or cont_names
new_gdf = op.apply_op(df,
columns_ctx,
"continuous",
stats_context=processor.stats)
df["x"] = (df["x"] -
processor.stats["means"]["x"]) / processor.stats["stds"]["x"]
assert new_gdf["x"].equals(df["x"])
def test_dask_normalize(client, tmpdir, datasets, engine):
paths = glob.glob(str(datasets[engine]) + "/*." + engine.split("-")[0])
df1 = cudf.read_parquet(paths[0])[mycols_pq]
df2 = cudf.read_parquet(paths[1])[mycols_pq]
df0 = cudf.concat([df1, df2], axis=0)
cat_names = ["name-cat", "name-string"]
cont_names = ["x", "y", "id"]
label_name = ["label"]
normalize = ops.Normalize()
conts = cont_names >> ops.FillMissing() >> normalize
workflow = Workflow(conts + cat_names + label_name, client=client)
dataset = Dataset(paths, engine)
result = workflow.fit_transform(dataset).to_ddf().compute()
# Make sure we collected accurate statistics
means = df0[cont_names].mean()
stds = df0[cont_names].std()
for name in cont_names:
assert math.isclose(means[name], normalize.means[name], rel_tol=1e-3)
assert math.isclose(stds[name], normalize.stds[name], rel_tol=1e-3)
# New (normalized) means should all be close to zero
new_means = result[cont_names].mean()
for name in cont_names:
assert new_means[name] < 1e-3
def test_s3_dataset(s3, paths, engine, df):
# create a mocked out bucket here
bucket = "testbucket"
s3.create_bucket(Bucket=bucket)
s3_paths = []
for path in paths:
s3_path = f"s3://{bucket}/{path}"
with fsspec.open(s3_path, "wb") as f:
f.write(open(path, "rb").read())
s3_paths.append(s3_path)
# create a basic s3 dataset
dataset = nvt.Dataset(s3_paths)
# make sure the iteration API works
columns = mycols_pq if engine == "parquet" else mycols_csv
gdf = cudf.concat(list(dataset.to_iter()))[columns]
assert_eq(gdf.reset_index(drop=True), df.reset_index(drop=True))
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.FillMissing(), ops.Clip(min_value=0), ops.LogOp()])
processor.add_preprocess(ops.Normalize())
processor.add_preprocess(ops.Categorify(cat_cache="host"))
processor.finalize()
processor.update_stats(dataset)
def test_parquet_output(client, use_client, tmpdir, shuffle):
out_files_per_proc = 2
n_workers = len(client.cluster.workers) if use_client else 1
out_path = str(tmpdir.mkdir("processed"))
path = str(tmpdir.join("simple.parquet"))
size = 25
row_group_size = 5
df = pd.DataFrame({"a": np.arange(size)})
df.to_parquet(path, row_group_size=row_group_size, engine="pyarrow")
columns = ["a"]
dataset = nvt.Dataset(path, engine="parquet", row_groups_per_part=1)
workflow = nvt.Workflow(columns >> ops.Normalize(),
client=client if use_client else None)
workflow.fit_transform(dataset).to_parquet(
output_path=out_path,
shuffle=shuffle,
out_files_per_proc=out_files_per_proc)
# Check that the number of output files is correct
result = glob.glob(os.path.join(out_path, "*.parquet"))
assert len(result) == out_files_per_proc * n_workers
# Make sure _metadata exists
meta_path = os.path.join(out_path, "_metadata")
assert os.path.exists(meta_path)
# Make sure _metadata makes sense
_metadata = cudf.io.read_parquet_metadata(meta_path)
assert _metadata[0] == size
assert _metadata[2] == columns
def test_empty_cols(tmpdir, df, dataset, engine, cat_names, cont_names, label_name):
# test out https://github.com/NVIDIA/NVTabular/issues/149 making sure we can iterate over
# empty cats/conts
# first with no continuous columns
processor = nvt.Workflow(cat_names=cat_names, cont_names=cont_names, label_name=label_name)
processor.add_feature([ops.FillMedian()])
processor.add_feature(ops.Normalize())
processor.add_feature(ops.Categorify())
output_train = os.path.join(tmpdir, "train/")
os.mkdir(output_train)
processor.apply(
dataset,
apply_offline=True,
record_stats=True,
shuffle=nvt.io.Shuffle.PER_PARTITION,
output_format=None,
)
df_out = processor.get_ddf().compute(scheduler="synchronous")
data_itr = torch_dataloader.TorchAsyncItr(
nvt.Dataset(df_out), cats=cat_names, conts=cont_names, labels=label_name, batch_size=1
)
for nvt_batch in data_itr:
cats, conts, labels = nvt_batch
if cat_names:
assert cats.shape[-1] == len(cat_names)
if cont_names:
assert conts.shape[-1] == len(cont_names)
if label_name:
assert labels.shape[-1] == len(label_name)
def test_normalize_upcastfloat64(tmpdir, dataset, gpu_memory_frac, engine,
op_columns):
df = cudf.DataFrame(
{
"x": [1.9e10, 2.3e16, 3.4e18, 1.6e19],
"label": [1, 0, 1, 0]
},
dtype="float32")
cont_features = op_columns >> ops.Normalize()
processor = nvtabular.Workflow(cont_features)
dataset = nvt.Dataset(df)
processor.fit(dataset)
new_gdf = processor.transform(dataset).to_ddf().compute()
for col in op_columns:
assert math.isclose(df[col].mean(),
processor.column_group.op.means[col],
rel_tol=1e-4)
assert math.isclose(df[col].std(),
processor.column_group.op.stds[col],
rel_tol=1e-4)
df[col] = (df[col] - processor.column_group.op.means[col]
) / processor.column_group.op.stds[col]
assert np.all((df[col] - new_gdf[col]).abs().values <= 1e-2)
def test_error_handling(tmpdir):
df = _make_df({"x": np.arange(10), "y": np.arange(10)})
def custom_transform(col):
if len(col) == 2:
raise ValueError("Lets cause some problems")
return col
features = ["x", "y"
] >> ops.FillMissing() >> ops.Normalize() >> custom_transform
workflow = nvt.Workflow(features)
workflow.fit(nvt.Dataset(df))
model_name = "test_error_handling"
triton.generate_nvtabular_model(workflow,
model_name,
tmpdir + f"/{model_name}",
backend=BACKEND)
with run_triton_server(tmpdir) as client:
inputs = triton.convert_df_to_triton_input(["x", "y"], df[:2])
with pytest.raises(
tritonclient.utils.InferenceServerException) as exception_info:
client.infer(model_name, inputs)
assert "ValueError: Lets cause some problems" in str(
exception_info.value)
def test_dask_normalize(client, tmpdir, datasets, engine):
paths = glob.glob(str(datasets[engine]) + "/*." + engine.split("-")[0])
df1 = cudf.read_parquet(paths[0])[mycols_pq]
df2 = cudf.read_parquet(paths[1])[mycols_pq]
df0 = cudf.concat([df1, df2], axis=0)
cat_names = ["name-cat", "name-string"]
cont_names = ["x", "y", "id"]
label_name = ["label"]
processor = Workflow(
client=client, cat_names=cat_names, cont_names=cont_names, label_name=label_name
)
processor.add_preprocess(ops.Normalize())
processor.finalize()
dataset = Dataset(paths, engine)
processor.apply(dataset)
result = processor.get_ddf().compute()
# Make sure we collected accurate statistics
means = df0[cont_names].mean()
stds = df0[cont_names].std()
counts = df0[cont_names].count()
for name in cont_names:
assert math.isclose(means[name], processor.stats["means"][name], rel_tol=1e-3)
assert math.isclose(stds[name], processor.stats["stds"][name], rel_tol=1e-3)
assert math.isclose(counts[name], processor.stats["counts"][name], rel_tol=1e-3)
# New (normalized) means should all be close to zero
new_means = result[cont_names].mean()
for name in cont_names:
assert new_means[name] < 1e-3
def test_schema_write_read_dataset(tmpdir, dataset, engine):
cat_names = ["name-cat", "name-string"
] if engine == "parquet" else ["name-string"]
cont_names = ["x", "y", "id"]
label_name = ["label"]
norms = ops.Normalize()
cat_features = cat_names >> ops.Categorify(cat_cache="host")
cont_features = cont_names >> ops.FillMissing() >> ops.Clip(
min_value=0) >> ops.LogOp >> norms
workflow = Workflow(cat_features + cont_features + label_name)
workflow.fit(dataset)
workflow.transform(dataset).to_parquet(
tmpdir,
out_files_per_proc=10,
)
schema_path = Path(tmpdir)
proto_schema = PbTxt_SchemaWriter._read(schema_path / "schema.pbtxt")
new_dataset = Dataset(glob.glob(str(tmpdir) + "/*.parquet"))
assert """name: "name-cat"\n min: 0\n max: 27\n""" in str(
proto_schema)
assert new_dataset.schema == workflow.output_schema
def test_empty_cols(tmpdir, df, dataset, engine, cat_names, cont_names, label_name):
features = []
if cont_names:
features.append(cont_names >> ops.FillMedian() >> ops.Normalize())
if cat_names:
features.append(cat_names >> ops.Categorify())
# test out https://github.com/NVIDIA/NVTabular/issues/149 making sure we can iterate over
# empty cats/conts
graph = sum(features, nvt.ColumnGroup(label_name))
if not graph.columns:
# if we don't have conts/cats/labels we're done
return
processor = nvt.Workflow(sum(features, nvt.ColumnGroup(label_name)))
output_train = os.path.join(tmpdir, "train/")
os.mkdir(output_train)
df_out = processor.fit_transform(dataset).to_ddf().compute(scheduler="synchronous")
data_itr = torch_dataloader.TorchAsyncItr(
nvt.Dataset(df_out), cats=cat_names, conts=cont_names, labels=label_name, batch_size=1
)
for nvt_batch in data_itr:
cats, conts, labels = nvt_batch
if cat_names:
assert cats.shape[-1] == len(cat_names)
if cont_names:
assert conts.shape[-1] == len(cont_names)
if label_name:
assert labels.shape[-1] == len(label_name)
def test_normalize_upcastfloat64(tmpdir, dataset, gpu_memory_frac, engine, op_columns):
df = cudf.DataFrame(
{"x": [1.9e10, 2.3e16, 3.4e18, 1.6e19], "label": [1, 0, 1, 0]}, dtype="float32"
)
cat_names = []
cont_names = ["x"]
label_name = ["label"]
config = nvt.workflow.get_new_config()
config["PP"]["continuous"] = [ops.Moments(columns=op_columns)]
processor = nvtabular.Workflow(
cat_names=cat_names, cont_names=cont_names, label_name=label_name, config=config
)
processor.update_stats(dataset)
op = ops.Normalize()
columns_ctx = {}
columns_ctx["continuous"] = {}
columns_ctx["continuous"]["base"] = op_columns or cont_names
new_gdf = op.apply_op(df, columns_ctx, "continuous", stats_context=processor.stats)
df["x"] = (df["x"] - processor.stats["means"]["x"]) / processor.stats["stds"]["x"]
assert new_gdf["x"].equals(df["x"])
def test_concatenate_dataframe(tmpdir, output_model):
# we were seeing an issue in the rossmann workflow where we dropped certain columns,
# https://github.com/NVIDIA/NVTabular/issues/961
df = _make_df({
"cat": ["aaaa", "bbbb", "cccc", "aaaa", "bbbb", "aaaa"],
"cont": [0.0, 1.0, 2.0, 3.0, 4.0, 5],
})
# this bug only happened with a dataframe representation: force this by using a lambda
cats = ["cat"] >> ops.LambdaOp(lambda col: _hash_series(col) % 1000)
conts = ["cont"] >> ops.Normalize() >> ops.FillMissing() >> ops.LogOp()
dataset = Dataset(df)
workflow = nvt.Workflow(cats + conts).fit_schema(dataset.infer_schema())
if output_model == "pytorch":
model_info = {
"cat": {
"columns": ["cat"],
"dtype": "int32"
},
"cont": {
"columns": ["cont"],
"dtype": "float32"
},
}
else:
model_info = None
_verify_workflow_on_tritonserver(tmpdir, workflow, df,
"test_concatenate_dataframe",
output_model, model_info)
def test_normalize_std_zero(cpu):
df = pd.DataFrame({"a": 7 * [10]})
dataset = nvt.Dataset(df, cpu=cpu)
processor = nvtabular.Workflow(["a"] >> ops.Normalize())
processor.fit(dataset)
result = processor.transform(dataset).compute()["a"]
assert (result == 0).all()
def test_kill_dl(tmpdir, df, dataset, part_mem_fraction, engine):
cat_names = ["name-cat", "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()])
processor.add_preprocess(ops.Normalize())
processor.add_preprocess(ops.Categorify())
output_train = os.path.join(tmpdir, "train/")
os.mkdir(output_train)
processor.apply(
dataset,
apply_offline=True,
record_stats=True,
shuffle=nvt.io.Shuffle.PER_PARTITION,
output_path=output_train,
)
tar_paths = [
os.path.join(output_train, x) for x in os.listdir(output_train) if x.endswith("parquet")
]
nvt_data = nvt.Dataset(tar_paths[0], engine="parquet", part_mem_fraction=part_mem_fraction)
data_itr = torch_dataloader.TorchAsyncItr(
nvt_data, cats=cat_names, conts=cont_names, labels=["label"]
)
results = {}
for batch_size in [2 ** i for i in range(9, 25, 1)]:
print("Checking batch size: ", batch_size)
num_iter = max(10 * 1000 * 1000 // batch_size, 100) # load 10e7 samples
# import pdb; pdb.set_trace()
data_itr.batch_size = batch_size
start = time.time()
for i, data in enumerate(data_itr):
if i >= num_iter:
break
del data
stop = time.time()
throughput = i * batch_size / (stop - start)
results[batch_size] = throughput
print(
"batch size: ",
batch_size,
", throughput: ",
throughput,
"items",
i * batch_size,
"time",
stop - start,
)
def test_gpu_dl_break(tmpdir, df, dataset, batch_size, part_mem_fraction,
engine, device):
cat_names = ["name-cat", "name-string"]
cont_names = ["x", "y", "id"]
label_name = ["label"]
conts = cont_names >> ops.FillMedian() >> ops.Normalize()
cats = cat_names >> ops.Categorify()
processor = nvt.Workflow(conts + cats + label_name)
output_train = os.path.join(tmpdir, "train/")
os.mkdir(output_train)
processor.fit_transform(dataset).to_parquet(
shuffle=nvt.io.Shuffle.PER_PARTITION,
output_path=output_train,
out_files_per_proc=2,
)
tar_paths = [
os.path.join(output_train, x) for x in os.listdir(output_train)
if x.endswith("parquet")
]
nvt_data = nvt.Dataset(tar_paths[0],
engine="parquet",
part_mem_fraction=part_mem_fraction)
data_itr = torch_dataloader.TorchAsyncItr(
nvt_data,
batch_size=batch_size,
cats=cat_names,
conts=cont_names,
labels=["label"],
device=device,
)
len_dl = len(data_itr) - 1
first_chunk = 0
idx = 0
for idx, chunk in enumerate(data_itr):
if idx == 0:
first_chunk = len(chunk[0])
last_chk = len(chunk[0])
print(last_chk)
if idx == 1:
break
del chunk
assert idx < len_dl
first_chunk_2 = 0
for idx, chunk in enumerate(data_itr):
if idx == 0:
first_chunk_2 = len(chunk[0])
del chunk
assert idx == len_dl
assert first_chunk == first_chunk_2
def test_gpu_workflow(tmpdir, 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"]
norms = ops.Normalize()
conts = cont_names >> ops.FillMissing() >> ops.Clip(min_value=0) >> norms
cats = cat_names >> ops.Categorify()
workflow = nvt.Workflow(conts + cats + label_name)
workflow.fit(dataset)
if dump:
workflow_dir = os.path.join(tmpdir, "workflow")
workflow.save(workflow_dir)
workflow = None
workflow = Workflow.load(workflow_dir)
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(), norms.means["x"], rel_tol=1e-4)
assert math.isclose(get_norms(df.y).mean(), norms.means["y"], rel_tol=1e-4)
assert math.isclose(get_norms(df.x).std(), norms.stds["x"], rel_tol=1e-3)
assert math.isclose(get_norms(df.y).std(), norms.stds["y"], rel_tol=1e-3)
# Check that categories match
if engine == "parquet":
cats_expected0 = df["name-cat"].unique().values_host
cats0 = get_cats(workflow, "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(workflow, "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
workflow.transform(dataset).to_parquet(
output_path=tmpdir,
out_files_per_proc=10,
shuffle=nvt.io.Shuffle.PER_PARTITION)
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_fit_schema_works_when_subtracting_column_names():
schema = Schema(["x", "y", "id"])
cont_features = (ColumnSelector(
["x", "y"]) >> ops.FillMissing() >> ops.Clip(min_value=0) >> ops.LogOp
>> ops.Normalize() >> ops.Rename(postfix="_renamed"))
workflow1 = Workflow(cont_features - "y_renamed")
workflow1.fit_schema(schema)
assert workflow1.output_schema.column_names == ["x_renamed"]
def test_fit_schema():
schema = Schema(["x", "y", "id"])
cont_features = (ColumnSelector(schema.column_names) >> ops.FillMissing()
>> ops.Clip(min_value=0) >> ops.LogOp >> ops.Normalize()
>> ops.Rename(postfix="_renamed"))
workflow = Workflow(cont_features)
workflow.fit_schema(schema)
assert workflow.output_schema.column_names == [
"x_renamed", "y_renamed", "id_renamed"
]
def test_gpu_dl(tmpdir, df, dataset, batch_size, gpu_memory_frac, engine):
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()])
processor.add_preprocess(ops.Normalize())
processor.add_preprocess(ops.Categorify())
output_train = os.path.join(tmpdir, "train/")
os.mkdir(output_train)
processor.apply(
dataset,
apply_offline=True,
record_stats=True,
shuffle=True,
output_path=output_train,
num_out_files=2,
)
tar_paths = [
os.path.join(output_train, x) for x in os.listdir(output_train) if x.endswith("parquet")
]
data_itr = nvt.torch_dataloader.TorchTensorBatchDatasetItr(
tar_paths[0],
engine="parquet",
sub_batch_size=batch_size,
gpu_memory_frac=gpu_memory_frac,
cats=cat_names,
conts=cont_names,
labels=["label"],
names=mycols_csv,
sep="\t",
)
num_rows, num_row_groups, col_names = cudf.io.read_parquet_metadata(tar_paths[0])
rows = 0
for idx, chunk in enumerate(data_itr):
rows += len(chunk)
del chunk
# accounts for incomplete batches at the end of chunks
# that dont necesssarily have the full batch_size
assert (idx + 1) * batch_size >= rows
assert rows == num_rows
if os.path.exists(output_train):
shutil.rmtree(output_train)
def test_generate_triton_model(tmpdir, engine, output_model, df):
tmpdir = "./tmp"
conts = ["x", "y", "id"] >> ops.FillMissing() >> ops.Normalize()
cats = ["name-cat", "name-string"] >> ops.Categorify(cat_cache="host")
workflow = nvt.Workflow(conts + cats)
workflow.fit(nvt.Dataset(df))
expected = workflow.transform(nvt.Dataset(df)).to_ddf().compute()
# save workflow to triton / verify we see some expected output
if output_model == "pytorch":
model_info = {
"name-cat": {
"columns": ["name-cat"],
"dtype": "int64"
},
"name-string": {
"columns": ["name-string"],
"dtype": "int64"
},
"id": {
"columns": ["id"],
"dtype": "float32"
},
"x": {
"columns": ["x"],
"dtype": "float32"
},
"y": {
"columns": ["y"],
"dtype": "float32"
},
}
else:
model_info = None
repo = os.path.join(tmpdir, "models")
triton.generate_nvtabular_model(
workflow=workflow,
name="model",
output_path=repo,
version=1,
output_model=output_model,
output_info=model_info,
)
workflow = None
assert os.path.exists(os.path.join(repo, "config.pbtxt"))
workflow = nvt.Workflow.load(os.path.join(repo, "1", "workflow"))
transformed = workflow.transform(nvt.Dataset(df)).to_ddf().compute()
assert_eq(expected, transformed)
def test_remove_columns():
# _remove_columns was failing to export the criteo example, because
# the label column was getting inserted into the subgroups of the output node
# https://github.com/NVIDIA-Merlin/NVTabular/issues/1198
label_columns = ["label"]
cats = ["a"] >> ops.Categorify()
conts = ["b"] >> ops.Normalize()
workflow = nvt.Workflow(cats + conts + label_columns)
df = pd.DataFrame({"a": ["a", "b"], "b": [1.0, 2.0], "label": [0, 1]})
workflow.fit(nvt.Dataset(df))
removed = ensemble._remove_columns(workflow, label_columns)
assert set(removed.output_dtypes.keys()) == {"a", "b"}
def test_normalize(tmpdir, df, dataset, gpu_memory_frac, engine, op_columns):
cont_features = op_columns >> ops.Normalize()
processor = nvtabular.Workflow(cont_features)
processor.fit(dataset)
new_gdf = processor.transform(dataset).to_ddf().compute()
new_gdf.index = df.index # Make sure index is aligned for checks
for col in op_columns:
assert math.isclose(df[col].mean(), processor.column_group.op.means[col], rel_tol=1e-4)
assert math.isclose(df[col].std(), processor.column_group.op.stds[col], rel_tol=1e-4)
df[col] = (df[col] - processor.column_group.op.means[col]) / processor.column_group.op.stds[
col
]
assert np.all((df[col] - new_gdf[col]).abs().values <= 1e-2)
def test_hugectr(tmpdir, df, dataset, output_format, engine, op_columns):
cat_names = ["name-cat", "name-string"] if engine == "parquet" else ["name-string"]
cont_names = ["x", "y"]
label_names = ["label"]
# set variables
nfiles = 10
ext = ""
outdir = tmpdir + "/hugectr"
os.mkdir(outdir)
# process data
processor = nvt.Workflow(cat_names=cat_names, cont_names=cont_names, label_name=label_names)
processor.add_feature([ops.ZeroFill(columns=op_columns), ops.LogOp()])
processor.add_preprocess(ops.Normalize())
processor.add_preprocess(ops.Categorify())
processor.finalize()
# Need to collect statistics first (for now)
processor.update_stats(dataset)
# Second "online" pass to write HugeCTR output
processor.apply(
dataset,
apply_offline=False,
record_stats=False,
output_path=outdir,
out_files_per_proc=nfiles,
output_format=output_format,
shuffle=False,
)
# Check files
ext = ""
if output_format == "parquet":
ext = "parquet"
assert os.path.isfile(outdir + "/metadata.json")
elif output_format == "hugectr":
ext = "data"
assert os.path.isfile(outdir + "/file_list.txt")
for n in range(nfiles):
assert os.path.isfile(os.path.join(outdir, str(n) + "." + ext))
def test_generate_triton_model(tmpdir, engine, df):
tmpdir = "./tmp"
conts = ["x", "y", "id"] >> ops.FillMissing() >> ops.Normalize()
cats = ["name-cat", "name-string"] >> ops.Categorify(cat_cache="host")
workflow = nvt.Workflow(conts + cats)
workflow.fit(nvt.Dataset(df))
expected = workflow.transform(nvt.Dataset(df)).to_ddf().compute()
# save workflow to triton / verify we see some expected output
repo = os.path.join(tmpdir, "models")
triton.generate_nvtabular_model(workflow, "model", repo)
workflow = None
assert os.path.exists(os.path.join(repo, "config.pbtxt"))
workflow = nvt.Workflow.load(os.path.join(repo, "1", "workflow"))
transformed = workflow.transform(nvt.Dataset(df)).to_ddf().compute()
assert_eq(expected, transformed)
def test_workflow_transform_ddf_dtypes():
# Initial Dataset
df = cudf.datasets.timeseries().reset_index()
ddf = dask_cudf.from_cudf(df, npartitions=2)
dataset = Dataset(ddf)
# Create and Execute Workflow
cols = ["name", "x", "y", "timestamp"]
cat_cols = ["id"] >> ops.Normalize()
workflow = Workflow(cols + cat_cols)
workflow.fit(dataset)
transformed_ddf = workflow.transform(dataset).to_ddf()
# no transforms on the pass through cols, should have original dtypes
for col in cols:
assert_eq(ddf.dtypes[col], transformed_ddf.dtypes[col])
# Followup dask-cudf sorting used to throw an exception because of dtype issues,
# check that it works now
transformed_ddf.sort_values(["id", "timestamp"]).compute()
def test_dask_preproc_cpu(client, tmpdir, datasets, engine, shuffle, cpu):
paths = glob.glob(str(datasets[engine]) + "/*." + engine.split("-")[0])
if engine == "parquet":
df1 = cudf.read_parquet(paths[0])[mycols_pq]
df2 = cudf.read_parquet(paths[1])[mycols_pq]
elif engine == "csv":
df1 = cudf.read_csv(paths[0], header=0)[mycols_csv]
df2 = cudf.read_csv(paths[1], header=0)[mycols_csv]
else:
df1 = cudf.read_csv(paths[0], names=allcols_csv)[mycols_csv]
df2 = cudf.read_csv(paths[1], names=allcols_csv)[mycols_csv]
df0 = cudf.concat([df1, df2], axis=0)
if engine in ("parquet", "csv"):
dataset = Dataset(paths, part_size="1MB", cpu=cpu)
else:
dataset = Dataset(paths, names=allcols_csv, part_size="1MB", cpu=cpu)
# Simple transform (normalize)
cat_names = ["name-string"]
cont_names = ["x", "y", "id"]
label_name = ["label"]
conts = cont_names >> ops.FillMissing() >> ops.Normalize()
workflow = Workflow(conts + cat_names + label_name, client=client)
transformed = workflow.fit_transform(dataset)
# Write out dataset
output_path = os.path.join(tmpdir, "processed")
transformed.to_parquet(output_path=output_path,
shuffle=shuffle,
out_files_per_proc=4)
# Check the final result
df_disk = dd_read_parquet(output_path, engine="pyarrow").compute()
assert_eq(
df0.sort_values(["id", "x"])[["name-string", "label"]],
df_disk.sort_values(["id", "x"])[["name-string", "label"]],
check_index=False,
)
def test_dataloader_schema(tmpdir, df, dataset, batch_size, engine, device):
cat_names = ["name-cat", "name-string"]
cont_names = ["x", "y", "id"]
label_name = ["label"]
conts = cont_names >> ops.FillMedian() >> ops.Normalize()
cats = cat_names >> ops.Categorify()
processor = nvt.Workflow(conts + cats + label_name)
output_train = os.path.join(tmpdir, "train/")
os.mkdir(output_train)
processor.fit_transform(dataset).to_parquet(
shuffle=nvt.io.Shuffle.PER_PARTITION,
output_path=output_train,
out_files_per_proc=2,
)
tar_paths = [
os.path.join(output_train, x) for x in os.listdir(output_train)
if x.endswith("parquet")
]
nvt_data = nvt.Dataset(tar_paths, engine="parquet")
data_loader = torch_dataloader.TorchAsyncItr(
nvt_data,
batch_size=batch_size,
shuffle=False,
labels=label_name,
)
batch = next(iter(data_loader))
assert all(name in batch[0] for name in cat_names)
assert all(name in batch[0] for name in cont_names)
num_label_cols = batch[1].shape[1] if len(batch[1].shape) > 1 else 1
assert num_label_cols == len(label_name)
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"]
norms = ops.Normalize()
cat_features = cat_names >> ops.Categorify()
if replace:
cont_features = cont_names >> ops.FillMissing() >> ops.LogOp >> norms
else:
fillmissing_logop = (cont_names >> ops.FillMissing() >> ops.LogOp >>
ops.Rename(postfix="_FillMissing_1_LogOp_1"))
cont_features = cont_names + fillmissing_logop >> norms
workflow = Workflow(cat_features + cont_features + label_name,
client=client)
workflow.fit(dataset)
if dump:
workflow_dir = os.path.join(tmpdir, "workflow")
workflow.save(workflow_dir)
workflow = None
workflow = Workflow.load(workflow_dir, client=client)
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; Clip, Log
concat_ops = "_FillMissing_1_LogOp_1"
if replace:
concat_ops = ""
assert math.isclose(get_norms(df.x).mean(),
norms.means["x" + concat_ops],
rel_tol=1e-1)
assert math.isclose(get_norms(df.y).mean(),
norms.means["y" + concat_ops],
rel_tol=1e-1)
assert math.isclose(get_norms(df.x).std(),
norms.stds["x" + concat_ops],
rel_tol=1e-1)
assert math.isclose(get_norms(df.y).std(),
norms.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(workflow, "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(workflow, "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
workflow.transform(dataset).to_parquet(
tmpdir,
out_files_per_proc=10,
shuffle=nvt.io.Shuffle.PER_PARTITION,
)
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)
