:rocket: Code is stable

app.py

@@ -1,4 +1,5 @@
 import logging
+import os
 import random
 from typing import Any
 
@@ -6,13 +7,14 @@ import numpy as np
 import pandas as pd
 from pytorch_lightning import Trainer, LightningModule, LightningDataModule
 from pytorch_lightning.utilities.types import OptimizerLRScheduler, STEP_OUTPUT, EVAL_DATALOADERS, TRAIN_DATALOADERS
+from torch.nn.utils.rnn import pad_sequence
 from torch.utils.data import DataLoader, Dataset
 from torchmetrics.classification import BinaryAccuracy, BinaryAUROC, BinaryF1Score, BinaryPrecision, BinaryRecall
 from transformers import BertModel, BatchEncoding, BertTokenizer, TrainingArguments
 from transformers.modeling_outputs import BaseModelOutputWithPoolingAndCrossAttentions
 import torch
 from torch import nn
-from datasets import load_dataset
+from datasets import load_dataset, IterableDataset
 from huggingface_hub import PyTorchModelHubMixin
 
 timber = logging.getLogger()
@@ -136,19 +138,25 @@ def insert_debug_motif_at_random_position(seq, DEBUG_MOTIF):
   return output
 
 
-class MQTLDataset(Dataset):
-  def __init__(self, dataset, check_if_pipeline_is_ok_by_inserting_debug_motif=False):
-    self.dataset = dataset
+class MQTLDataset(IterableDataset):
+  def __init__(self, m_dataset, seq_len, check_if_pipeline_is_ok_by_inserting_debug_motif=False):
+    self.dataset = m_dataset
     self.check_if_pipeline_is_ok_by_inserting_debug_motif = check_if_pipeline_is_ok_by_inserting_debug_motif
     self.debug_motif = "ATCGCCTA"
+    self.seq_len = seq_len
     pass
 
-  def __len__(self):
-    return len(self.dataset)
-
-  def __getitem__(self, idx):
-    seq = self.dataset[idx]['sequence']  # Fetch the 'sequence' column
-    label = self.dataset[idx]['label']  # Fetch the 'label' column (or whatever target you use)
+  def __iter__(self):
+    for row in self.dataset:
+      processed = self.preprocess(row)
+      if processed is not None:
+        yield processed
+
+  def preprocess(self, row):
+    seq = row['sequence']  # Fetch the 'sequence' column
+    if len(seq) != self.seq_len:
+      return None  # skip problematic row!
+    label = row['label']  # Fetch the 'label' column (or whatever target you use)
     if label == 1 and self.check_if_pipeline_is_ok_by_inserting_debug_motif:
       seq = insert_debug_motif_at_random_position(seq=seq, DEBUG_MOTIF=self.debug_motif)
     seq_rc = reverse_complement_dna_seq(seq)
@@ -162,16 +170,36 @@ class MQTLDataset(Dataset):
     return [ohe_seq, ohe_seq_rc], label_np_array
 
 
+# def collate_fn(batch):
+#   sequences, labels = zip(*batch)
+#   ohe_seq, ohe_seq_rc = sequences[0], sequences[1]
+#   # Pad sequences to the maximum length in this batch
+#   padded_sequences = pad_sequence(ohe_seq, batch_first=True, padding_value=0)
+#   padded_sequences_rc = pad_sequence(ohe_seq_rc, batch_first=True, padding_value=0)
+#   # Convert labels to a tensor
+#   labels = torch.stack(labels)
+#   return [padded_sequences, padded_sequences_rc], labels
+
+
 class MqtlDataModule(LightningDataModule):
   def __init__(self, train_ds: Dataset, val_ds: Dataset, test_ds: Dataset, batch_size=16):
     super().__init__()
     self.batch_size = batch_size
-    self.train_loader = DataLoader(train_ds, batch_size=self.batch_size, shuffle=True, num_workers=15,
-                                   persistent_workers=True)
-    self.validate_loader = DataLoader(val_ds, batch_size=self.batch_size, shuffle=False, num_workers=15,
-                                      persistent_workers=True)
-    self.test_loader = DataLoader(test_ds, batch_size=self.batch_size, shuffle=False, num_workers=15,
-                                  persistent_workers=True)
+    self.train_loader = DataLoader(train_ds, batch_size=self.batch_size, shuffle=False,
+                                   # collate_fn=collate_fn,
+                                   num_workers=15,
+                                   # persistent_workers=True
+                                   )
+    self.validate_loader = DataLoader(val_ds, batch_size=self.batch_size, shuffle=False,
+                                      # collate_fn=collate_fn,
+                                      num_workers=15,
+                                      # persistent_workers=True
+                                      )
+    self.test_loader = DataLoader(test_ds, batch_size=self.batch_size, shuffle=False,
+                                  # collate_fn=collate_fn,
+                                  num_workers=15,
+                                  # persistent_workers=True
+                                  )
     pass
 
   def prepare_data(self):
@@ -376,11 +404,34 @@ def start(classifier_model, model_save_path, is_attention_model=False, m_optimiz
   if is_binned:
     file_suffix = "_binned"
 
-  dataset_map = load_dataset("fahimfarhan/mqtl-classification-datasets")
-
-  train_dataset = MQTLDataset(dataset_map[f"train_binned_{WINDOW}"], check_if_pipeline_is_ok_by_inserting_debug_motif=is_debug)
-  val_dataset = MQTLDataset(dataset_map[f"validate_binned_{WINDOW}"], check_if_pipeline_is_ok_by_inserting_debug_motif=is_debug)
-  test_dataset = MQTLDataset(dataset_map[f"test_binned_{WINDOW}"], check_if_pipeline_is_ok_by_inserting_debug_motif=is_debug)
+  data_files = {
+    # small samples
+    "train_binned_200": "/home/soumic/Codes/mqtl-classification/src/inputdata/dataset_200_train_binned.csv",
+    "validate_binned_200": "/home/soumic/Codes/mqtl-classification/src/inputdata/dataset_200_validate_binned.csv",
+    "test_binned_200": "/home/soumic/Codes/mqtl-classification/src/inputdata/dataset_200_test_binned.csv",
+    # large samples
+    "train_binned_4000": "/home/soumic/Codes/mqtl-classification/src/inputdata/dataset_4000_train_binned.csv",
+    "validate_binned_4000": "/home/soumic/Codes/mqtl-classification/src/inputdata/dataset_4000_validate_binned.csv",
+    "test_binned_4000": "/home/soumic/Codes/mqtl-classification/src/inputdata/dataset_4000_test_binned.csv",
+  }
+
+  dataset_map = None
+  is_my_laptop = os.path.isfile("/home/soumic/Codes/mqtl-classification/src/inputdata/dataset_4000_test_binned.csv")
+  if is_my_laptop:
+    dataset_map = load_dataset("csv", data_files=data_files, streaming=True)
+  else:
+    dataset_map = load_dataset("fahimfarhan/mqtl-classification-datasets", streaming=True)
+
+  train_dataset = MQTLDataset(dataset_map[f"train_binned_{WINDOW}"],
+                              check_if_pipeline_is_ok_by_inserting_debug_motif=is_debug,
+                              seq_len=WINDOW
+                              )
+  val_dataset = MQTLDataset(dataset_map[f"validate_binned_{WINDOW}"],
+                            check_if_pipeline_is_ok_by_inserting_debug_motif=is_debug,
+                            seq_len=WINDOW)
+  test_dataset = MQTLDataset(dataset_map[f"test_binned_{WINDOW}"],
+                             check_if_pipeline_is_ok_by_inserting_debug_motif=is_debug,
+                             seq_len=WINDOW)
 
   data_module = MqtlDataModule(train_ds=train_dataset, val_ds=val_dataset, test_ds=test_dataset)
 
@@ -409,7 +460,8 @@ def start(classifier_model, model_save_path, is_attention_model=False, m_optimiz
   classifier_model.save_pretrained(f"my-awesome-model-{WINDOW}")
 
   # push to the hub
-  classifier_model.push_to_hub(repo_id="fahimfarhan/mqtl-classifier-model", commit_message=":tada: Push model using huggingface_hub")
+  classifier_model.push_to_hub(repo_id="fahimfarhan/mqtl-classifier-model",
+                               commit_message=":tada: Push model using huggingface_hub")
 
   # reload
   model = classifier_model.from_pretrained(f"my-awesome-model-{WINDOW}")
@@ -434,13 +486,16 @@ if __name__ == '__main__':
   simple_cnn.enable_logging = True
 
   start(classifier_model=simple_cnn, model_save_path=simple_cnn.file_name, WINDOW=WINDOW,
-        dataset_folder_prefix="inputdata/", is_debug=True, max_epochs=3)
+        dataset_folder_prefix="inputdata/", is_debug=True, max_epochs=5)
 
   pass
 
-
 """
 lightning_logs/
 *.pth
 my-awesome-model
-"""
+
+INFO:root:validate_acc = 0.5625, validate_auc = 0.5490195751190186, validate_f1_score = 0.30000001192092896, validate_precision = 0.6000000238418579, validate_recall = 0.20000000298023224
+/home/soumic/Codes/mqtl-classification/venv/lib/python3.10/site-packages/pytorch_lightning/trainer/connectors/data_connector.py:424: The 'train_dataloader' does not have many workers which may be a bottleneck. Consider increasing the value of the `num_workers` argument` to `num_workers=15` in the `DataLoader` to improve performance.
+
+"""