Raise error for train and validation mismatch
#459
by
madhavanvenkatesh
- opened
- geneformer/mtl/data.py +117 -105
geneformer/mtl/data.py
CHANGED
@@ -1,150 +1,162 @@
|
|
1 |
import os
|
2 |
-
|
3 |
from .collators import DataCollatorForMultitaskCellClassification
|
4 |
from .imports import *
|
5 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
def load_and_preprocess_data(dataset_path, config, is_test=False, dataset_type=""):
|
|
|
8 |
try:
|
9 |
dataset = load_from_disk(dataset_path)
|
10 |
|
|
|
11 |
task_names = [f"task{i+1}" for i in range(len(config["task_columns"]))]
|
12 |
task_to_column = dict(zip(task_names, config["task_columns"]))
|
13 |
config["task_names"] = task_names
|
14 |
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
raise KeyError(
|
20 |
-
f"Column {column} not found in the dataset. Available columns: {list(available_columns)}"
|
21 |
-
)
|
22 |
-
|
23 |
-
label_mappings = {}
|
24 |
-
task_label_mappings = {}
|
25 |
-
cell_id_mapping = {}
|
26 |
-
num_labels_list = []
|
27 |
-
|
28 |
-
# Load or create task label mappings
|
29 |
-
if not is_test:
|
30 |
-
for task, column in task_to_column.items():
|
31 |
-
unique_values = sorted(set(dataset[column])) # Ensure consistency
|
32 |
-
label_mappings[column] = {
|
33 |
-
label: idx for idx, label in enumerate(unique_values)
|
34 |
-
}
|
35 |
-
task_label_mappings[task] = label_mappings[column]
|
36 |
-
num_labels_list.append(len(unique_values))
|
37 |
-
|
38 |
-
# Print the mappings for each task with dataset type prefix
|
39 |
-
for task, mapping in task_label_mappings.items():
|
40 |
-
print(
|
41 |
-
f"{dataset_type.capitalize()} mapping for {task}: {mapping}"
|
42 |
-
) # sanity check, for train/validation splits
|
43 |
-
|
44 |
-
# Save the task label mappings as a pickle file
|
45 |
-
with open(f"{config['results_dir']}/task_label_mappings.pkl", "wb") as f:
|
46 |
-
pickle.dump(task_label_mappings, f)
|
47 |
-
else:
|
48 |
-
# Load task label mappings from pickle file for test data
|
49 |
-
with open(f"{config['results_dir']}/task_label_mappings.pkl", "rb") as f:
|
50 |
-
task_label_mappings = pickle.load(f)
|
51 |
-
|
52 |
-
# Infer num_labels_list from task_label_mappings
|
53 |
-
for task, mapping in task_label_mappings.items():
|
54 |
-
num_labels_list.append(len(mapping))
|
55 |
-
|
56 |
-
# Store unique cell IDs in a separate dictionary
|
57 |
-
for idx, record in enumerate(dataset):
|
58 |
-
cell_id = record.get("unique_cell_id", idx)
|
59 |
-
cell_id_mapping[idx] = cell_id
|
60 |
-
|
61 |
-
# Transform records to the desired format
|
62 |
-
transformed_dataset = []
|
63 |
-
for idx, record in enumerate(dataset):
|
64 |
-
transformed_record = {}
|
65 |
-
transformed_record["input_ids"] = torch.tensor(
|
66 |
-
record["input_ids"], dtype=torch.long
|
67 |
-
)
|
68 |
|
69 |
-
|
70 |
-
|
71 |
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
transformed_record["label"] = label_dict
|
80 |
-
else:
|
81 |
-
# Create dummy labels for test data
|
82 |
-
label_dict = {task: -1 for task in config["task_names"]}
|
83 |
-
transformed_record["label"] = label_dict
|
84 |
|
85 |
-
|
|
|
|
|
|
|
86 |
|
87 |
return transformed_dataset, cell_id_mapping, num_labels_list
|
|
|
88 |
except KeyError as e:
|
89 |
-
|
90 |
except Exception as e:
|
91 |
-
|
92 |
-
return None, None, None
|
93 |
|
94 |
|
95 |
def preload_and_process_data(config):
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
)
|
100 |
-
val_dataset, val_cell_id_mapping, _ = load_and_preprocess_data(
|
101 |
-
config["val_path"], config, dataset_type="validation"
|
102 |
-
)
|
103 |
-
return (
|
104 |
-
train_dataset,
|
105 |
-
train_cell_id_mapping,
|
106 |
-
val_dataset,
|
107 |
-
val_cell_id_mapping,
|
108 |
-
num_labels_list,
|
109 |
-
)
|
110 |
|
|
|
|
|
111 |
|
112 |
-
|
113 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
114 |
|
115 |
-
data_collator = DataCollatorForMultitaskCellClassification()
|
116 |
|
117 |
-
|
|
|
|
|
118 |
preprocessed_dataset,
|
119 |
batch_size=batch_size,
|
120 |
shuffle=True,
|
121 |
-
collate_fn=
|
122 |
-
num_workers=
|
123 |
pin_memory=True,
|
124 |
)
|
125 |
-
return loader
|
126 |
|
127 |
|
128 |
def preload_data(config):
|
129 |
-
|
130 |
-
train_loader = get_data_loader(
|
131 |
-
val_loader = get_data_loader(
|
132 |
return train_loader, val_loader
|
133 |
|
134 |
|
135 |
def load_and_preprocess_test_data(config):
|
136 |
-
"""
|
137 |
-
Load and preprocess test data, treating it as unlabeled.
|
138 |
-
"""
|
139 |
return load_and_preprocess_data(config["test_path"], config, is_test=True)
|
140 |
|
141 |
|
142 |
def prepare_test_loader(config):
|
143 |
-
"""
|
144 |
-
|
145 |
-
"""
|
146 |
-
test_dataset, cell_id_mapping, num_labels_list = load_and_preprocess_test_data(
|
147 |
-
config
|
148 |
-
)
|
149 |
test_loader = get_data_loader(test_dataset, config["batch_size"])
|
150 |
return test_loader, cell_id_mapping, num_labels_list
|
|
|
1 |
import os
|
|
|
2 |
from .collators import DataCollatorForMultitaskCellClassification
|
3 |
from .imports import *
|
4 |
|
5 |
+
def validate_columns(dataset, required_columns, dataset_type):
|
6 |
+
"""Ensures required columns are present in the dataset."""
|
7 |
+
missing_columns = [col for col in required_columns if col not in dataset.column_names]
|
8 |
+
if missing_columns:
|
9 |
+
raise KeyError(
|
10 |
+
f"Missing columns in {dataset_type} dataset: {missing_columns}. "
|
11 |
+
f"Available columns: {dataset.column_names}"
|
12 |
+
)
|
13 |
+
|
14 |
+
|
15 |
+
def create_label_mappings(dataset, task_to_column):
|
16 |
+
"""Creates label mappings for the dataset."""
|
17 |
+
task_label_mappings = {}
|
18 |
+
num_labels_list = []
|
19 |
+
for task, column in task_to_column.items():
|
20 |
+
unique_values = sorted(set(dataset[column]))
|
21 |
+
mapping = {label: idx for idx, label in enumerate(unique_values)}
|
22 |
+
task_label_mappings[task] = mapping
|
23 |
+
num_labels_list.append(len(unique_values))
|
24 |
+
return task_label_mappings, num_labels_list
|
25 |
+
|
26 |
+
|
27 |
+
def save_label_mappings(mappings, path):
|
28 |
+
"""Saves label mappings to a pickle file."""
|
29 |
+
with open(path, "wb") as f:
|
30 |
+
pickle.dump(mappings, f)
|
31 |
+
|
32 |
+
|
33 |
+
def load_label_mappings(path):
|
34 |
+
"""Loads label mappings from a pickle file."""
|
35 |
+
with open(path, "rb") as f:
|
36 |
+
return pickle.load(f)
|
37 |
+
|
38 |
+
|
39 |
+
def transform_dataset(dataset, task_to_column, task_label_mappings, config, is_test):
|
40 |
+
"""Transforms the dataset to the required format."""
|
41 |
+
transformed_dataset = []
|
42 |
+
cell_id_mapping = {}
|
43 |
+
|
44 |
+
for idx, record in enumerate(dataset):
|
45 |
+
transformed_record = {
|
46 |
+
"input_ids": torch.tensor(record["input_ids"], dtype=torch.long),
|
47 |
+
"cell_id": idx, # Index-based cell ID
|
48 |
+
}
|
49 |
+
|
50 |
+
if not is_test:
|
51 |
+
label_dict = {
|
52 |
+
task: task_label_mappings[task][record[column]]
|
53 |
+
for task, column in task_to_column.items()
|
54 |
+
}
|
55 |
+
else:
|
56 |
+
label_dict = {task: -1 for task in config["task_names"]}
|
57 |
+
|
58 |
+
transformed_record["label"] = label_dict
|
59 |
+
transformed_dataset.append(transformed_record)
|
60 |
+
cell_id_mapping[idx] = record.get("unique_cell_id", idx)
|
61 |
+
|
62 |
+
return transformed_dataset, cell_id_mapping
|
63 |
+
|
64 |
|
65 |
def load_and_preprocess_data(dataset_path, config, is_test=False, dataset_type=""):
|
66 |
+
"""Main function to load and preprocess data."""
|
67 |
try:
|
68 |
dataset = load_from_disk(dataset_path)
|
69 |
|
70 |
+
# Setup task and column mappings
|
71 |
task_names = [f"task{i+1}" for i in range(len(config["task_columns"]))]
|
72 |
task_to_column = dict(zip(task_names, config["task_columns"]))
|
73 |
config["task_names"] = task_names
|
74 |
|
75 |
+
label_mappings_path = os.path.join(
|
76 |
+
config["results_dir"],
|
77 |
+
f"task_label_mappings{'_val' if dataset_type == 'validation' else ''}.pkl"
|
78 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
|
80 |
+
if not is_test:
|
81 |
+
validate_columns(dataset, task_to_column.values(), dataset_type)
|
82 |
|
83 |
+
# Create and save label mappings
|
84 |
+
task_label_mappings, num_labels_list = create_label_mappings(dataset, task_to_column)
|
85 |
+
save_label_mappings(task_label_mappings, label_mappings_path)
|
86 |
+
else:
|
87 |
+
# Load existing mappings for test data
|
88 |
+
task_label_mappings = load_label_mappings(label_mappings_path)
|
89 |
+
num_labels_list = [len(mapping) for mapping in task_label_mappings.values()]
|
|
|
|
|
|
|
|
|
|
|
90 |
|
91 |
+
# Transform dataset
|
92 |
+
transformed_dataset, cell_id_mapping = transform_dataset(
|
93 |
+
dataset, task_to_column, task_label_mappings, config, is_test
|
94 |
+
)
|
95 |
|
96 |
return transformed_dataset, cell_id_mapping, num_labels_list
|
97 |
+
|
98 |
except KeyError as e:
|
99 |
+
raise ValueError(f"Configuration error or dataset key missing: {e}")
|
100 |
except Exception as e:
|
101 |
+
raise RuntimeError(f"Error during data loading or preprocessing: {e}")
|
|
|
102 |
|
103 |
|
104 |
def preload_and_process_data(config):
|
105 |
+
"""Preloads and preprocesses train and validation datasets."""
|
106 |
+
# Process train data and save mappings
|
107 |
+
train_data = load_and_preprocess_data(config["train_path"], config, dataset_type="train")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
108 |
|
109 |
+
# Process validation data and save mappings
|
110 |
+
val_data = load_and_preprocess_data(config["val_path"], config, dataset_type="validation")
|
111 |
|
112 |
+
# Validate that the mappings match
|
113 |
+
validate_label_mappings(config)
|
114 |
+
|
115 |
+
return (*train_data, *val_data[:2]) # Return train and val data along with mappings
|
116 |
+
|
117 |
+
|
118 |
+
def validate_label_mappings(config):
|
119 |
+
"""Ensures train and validation label mappings are consistent."""
|
120 |
+
train_mappings_path = os.path.join(config["results_dir"], "task_label_mappings.pkl")
|
121 |
+
val_mappings_path = os.path.join(config["results_dir"], "task_label_mappings_val.pkl")
|
122 |
+
train_mappings = load_label_mappings(train_mappings_path)
|
123 |
+
val_mappings = load_label_mappings(val_mappings_path)
|
124 |
+
|
125 |
+
for task_name in config["task_names"]:
|
126 |
+
if train_mappings[task_name] != val_mappings[task_name]:
|
127 |
+
raise ValueError(
|
128 |
+
f"Mismatch in label mappings for task '{task_name}'.\n"
|
129 |
+
f"Train Mapping: {train_mappings[task_name]}\n"
|
130 |
+
f"Validation Mapping: {val_mappings[task_name]}"
|
131 |
+
)
|
132 |
|
|
|
133 |
|
134 |
+
def get_data_loader(preprocessed_dataset, batch_size):
|
135 |
+
"""Creates a DataLoader with optimal settings."""
|
136 |
+
return DataLoader(
|
137 |
preprocessed_dataset,
|
138 |
batch_size=batch_size,
|
139 |
shuffle=True,
|
140 |
+
collate_fn=DataCollatorForMultitaskCellClassification(),
|
141 |
+
num_workers=os.cpu_count(),
|
142 |
pin_memory=True,
|
143 |
)
|
|
|
144 |
|
145 |
|
146 |
def preload_data(config):
|
147 |
+
"""Preprocesses train and validation data for trials."""
|
148 |
+
train_loader = get_data_loader(*preload_and_process_data(config)[:2], config["batch_size"])
|
149 |
+
val_loader = get_data_loader(*preload_and_process_data(config)[2:4], config["batch_size"])
|
150 |
return train_loader, val_loader
|
151 |
|
152 |
|
153 |
def load_and_preprocess_test_data(config):
|
154 |
+
"""Loads and preprocesses test data."""
|
|
|
|
|
155 |
return load_and_preprocess_data(config["test_path"], config, is_test=True)
|
156 |
|
157 |
|
158 |
def prepare_test_loader(config):
|
159 |
+
"""Prepares DataLoader for test data."""
|
160 |
+
test_dataset, cell_id_mapping, num_labels_list = load_and_preprocess_test_data(config)
|
|
|
|
|
|
|
|
|
161 |
test_loader = get_data_loader(test_dataset, config["batch_size"])
|
162 |
return test_loader, cell_id_mapping, num_labels_list
|