Low Accuracy in BERT Ensemble Despite Strong Individual Model Performance

Beginners
1

I observed low accuracy in the BERT ensemble, where each individual BERT model—BanglaBERT, BUETBERT, and DistilBERT—achieved around 70–75% accuracy.
Below is the architecture of my model:
For each model applied this below line first
pip install accelerate -U

=======================

after that
Banglabert:

Imports and Installations

=======================

import os
os.environ[“WANDB_DISABLED”] = “true”

First uninstall transformers fully

#!pip uninstall -y transformers

Install correct version manually

!pip install transformers==4.41.1

!pip install --upgrade transformers datasets --quiet
!pip install datasets
!pip install demoji

import pandas as pd
from sklearn.model_selection import train_test_split
from datasets import Dataset
from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer
from sklearn.metrics import accuracy_score, precision_recall_fscore_support
from collections import defaultdict
import torch
import demoji
import re

=======================

Mount Google Drive

=======================

from google.colab import drive
drive.mount(‘/content/Drive’)

Set working directory

os.chdir(‘/content/Drive/My Drive/MSC thesis 25.04.2025’)

=======================

Text Preprocessing Function

=======================

demoji.download_codes()

def preprocess_bangla_text(text):
text = demoji.replace(text, repl=“”)
whitespace = re.compile(r"[\s\u0020\u00a0\u1680\u180e\u202f\u205f\u3000\u2000-\u200a]+“, re.UNICODE)
bangla_fullstop = u”\u0964"
punctSeq = u"['"“”‘’]+|[.?!,…]+|[:;]+"
punc = u"[(),$%^&*+={}:"|'~`<>/,€¦!?½£¶™¼©⅐⅑⅒⅓⅔⅕⅖⅗⅘⅙⅚⅛⅜⅝⅞⅟↉¤¿º;-]+"
bangla_numbers = u"[০১২৩৪৫৬৭৮৯]+"

text = re.sub(punctSeq, " ", text)
text = re.sub(bangla_fullstop, " ", text)
text = re.sub(punc, " ", text)
text = re.sub(bangla_numbers, "", text)
text = whitespace.sub(" ", text).strip()
return text

=======================

Load Dataset

=======================

Load Tier 1 (strict - will be used for evaluation only)

tier1 = pd.read_excel(‘strict_cleaned_file.xlsx’)
tier1 = tier1[[‘Text’, ‘Region’]]

Load Tier 2 and Tier 3 (flexible and all-no for training)

tier2 = pd.read_excel(‘flexible_filtered.xlsx’)
tier2 = tier2[[‘Text’, ‘Region’]]
tier3 = pd.read_excel(‘all_no_cases.xlsx’)
tier3 = tier3[[‘Text’, ‘Region’]]

Combine Tier 2 + Tier 3 for training

train_df = pd.concat([tier2, tier3], ignore_index=True)

=======================

Preprocessing

=======================

tier1[‘Text’] = tier1[‘Text’].apply(preprocess_bangla_text)
train_df[‘Text’] = train_df[‘Text’].apply(preprocess_bangla_text)

=======================

Special Word-Based Oversampling

=======================

special_words_df = pd.read_excel(‘special_words.xlsx’)

special_vocab = defaultdict(set)
for _, row in special_words_df.iterrows():
special_vocab[row[‘region’]].add(str(row[‘word’]))

original_counts = train_df[‘Region’].value_counts().to_dict()
max_count = max(original_counts.values())

augmented_rows =

for region in train_df[‘Region’].unique():
regional_texts = train_df[train_df[‘Region’] == region]
current_count = len(regional_texts)
needed = max_count - current_count
if needed <= 0:
continue

texts_with_keywords = regional_texts[
    regional_texts['Text'].apply(lambda txt: any(word in str(txt) for word in special_vocab[region]))
]

if len(texts_with_keywords) == 0:
    continue

for _ in range(needed):
    sample_row = texts_with_keywords.sample(n=1, replace=True).iloc[0]
    augmented_rows.append(sample_row)

Append oversampled data

train_df = pd.concat([train_df, pd.DataFrame(augmented_rows)], ignore_index=True)

=======================

Encode Labels

=======================

unique_labels = tier1[‘Region’].unique().tolist()
label2id = {label: i for i, label in enumerate(unique_labels)}
id2label = {i: label for label, i in label2id.items()}

train_df[‘label’] = train_df[‘Region’].map(label2id)
tier1[‘label’] = tier1[‘Region’].map(label2id)

=======================

Tokenization

=======================

tokenizer = AutoTokenizer.from_pretrained(“sagorsarker/bangla-bert-base”)

Prepare datasets

dataset_train = Dataset.from_pandas(train_df[[‘Text’, ‘label’]])
dataset_test = Dataset.from_pandas(tier1[[‘Text’, ‘label’]])

def tokenize(batch):
return tokenizer(batch[‘Text’], padding=“max_length”, truncation=True, max_length=128)

dataset_train = dataset_train.map(tokenize, batched=True)
dataset_test = dataset_test.map(tokenize, batched=True)

=======================

Model Loading

=======================

model = AutoModelForSequenceClassification.from_pretrained(
“sagorsarker/bangla-bert-base”,
num_labels=len(unique_labels)
)

=======================

Training Arguments

=======================

training_args = TrainingArguments(

output_dir="./results_banglabert_tier2_tier3",
per_device_train_batch_size=8,
per_device_eval_batch_size=8,
num_train_epochs=3,
logging_dir="./logs_banglabert_tier2_tier3",
logging_steps=10,
eval_strategy="epoch",
save_strategy="epoch",
load_best_model_at_end=True,
metric_for_best_model="accuracy"

)

=======================

Compute Metrics

=======================

def compute_metrics(eval_pred):
logits, labels = eval_pred
predictions = logits.argmax(axis=-1)
precision, recall, f1, _ = precision_recall_fscore_support(labels, predictions, average=‘weighted’)
acc = accuracy_score(labels, predictions)
return {
‘accuracy’: acc,
‘f1’: f1,
‘precision’: precision,
‘recall’: recall
}

=======================

Trainer Setup

=======================

trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset_train,
eval_dataset=dataset_test,
tokenizer=tokenizer,
compute_metrics=compute_metrics
)

=======================

Training and Evaluation

=======================

trainer.train()
trainer.evaluate()

:white_check_mark: Add label mappings before saving

model.config.label2id = label2id
model.config.id2label = id2label

Save model

trainer.save_model(“/content/Drive/MyDrive/MSC thesis 25.04.2025/three_bert_tier2/banglabert_saved_model”)
for buetbert:

=======================

Imports and Installations

=======================

import os
os.environ[“WANDB_DISABLED”] = “true”

First uninstall transformers fully

#!pip uninstall -y transformers

Install correct version manually

!pip install transformers==4.41.1

!pip install --upgrade transformers datasets --quiet
!pip install datasets
!pip install demoji

import pandas as pd
from sklearn.model_selection import train_test_split
from datasets import Dataset
from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer
from sklearn.metrics import accuracy_score, precision_recall_fscore_support
from collections import defaultdict
import torch
import demoji
import re

=======================

Mount Google Drive

=======================

from google.colab import drive
drive.mount(‘/content/Drive’)

Set working directory

os.chdir(‘/content/Drive/My Drive/MSC thesis 25.04.2025’)

=======================

Text Preprocessing Function

=======================

demoji.download_codes()

def preprocess_bangla_text(text):
text = demoji.replace(text, repl=“”)
whitespace = re.compile(r"[\s\u0020\u00a0\u1680\u180e\u202f\u205f\u3000\u2000-\u200a]+“, re.UNICODE)
bangla_fullstop = u”\u0964"
punctSeq = u"['"“”‘’]+|[.?!,…]+|[:;]+"
punc = u"[(),$%^&*+={}:"|'~`<>/,€¦!?½£¶™¼©⅐⅑⅒⅓⅔⅕⅖⅗⅘⅙⅚⅛⅜⅝⅞⅟↉¤¿º;-]+"
bangla_numbers = u"[০১২৩৪৫৬৭৮৯]+"

text = re.sub(punctSeq, " ", text)
text = re.sub(bangla_fullstop, " ", text)
text = re.sub(punc, " ", text)
text = re.sub(bangla_numbers, "", text)
text = whitespace.sub(" ", text).strip()
return text

=======================

Load Dataset

=======================

Load Tier 1 (strict - will be used for evaluation only)

tier1 = pd.read_excel(‘strict_cleaned_file.xlsx’)
tier1 = tier1[[‘Text’, ‘Region’]]

Load Tier 2 and Tier 3 (flexible and all-no for training)

tier2 = pd.read_excel(‘flexible_filtered.xlsx’)
tier2 = tier2[[‘Text’, ‘Region’]]
tier3 = pd.read_excel(‘all_no_cases.xlsx’)
tier3 = tier3[[‘Text’, ‘Region’]]

Combine Tier 2 + Tier 3 for training

train_df = pd.concat([tier2, tier3], ignore_index=True)

=======================

Preprocessing

=======================

tier1[‘Text’] = tier1[‘Text’].apply(preprocess_bangla_text)
train_df[‘Text’] = train_df[‘Text’].apply(preprocess_bangla_text)

=======================

Special Word-Based Oversampling

=======================

special_words_df = pd.read_excel(‘special_words.xlsx’)

special_vocab = defaultdict(set)
for _, row in special_words_df.iterrows():
special_vocab[row[‘region’]].add(str(row[‘word’]))

original_counts = train_df[‘Region’].value_counts().to_dict()
max_count = max(original_counts.values())

augmented_rows =

for region in train_df[‘Region’].unique():
regional_texts = train_df[train_df[‘Region’] == region]
current_count = len(regional_texts)
needed = max_count - current_count
if needed <= 0:
continue

texts_with_keywords = regional_texts[
    regional_texts['Text'].apply(lambda txt: any(word in str(txt) for word in special_vocab[region]))
]

if len(texts_with_keywords) == 0:
    continue

for _ in range(needed):
    sample_row = texts_with_keywords.sample(n=1, replace=True).iloc[0]
    augmented_rows.append(sample_row)

Append oversampled data

train_df = pd.concat([train_df, pd.DataFrame(augmented_rows)], ignore_index=True)

=======================

Encode Labels

=======================

unique_labels = tier1[‘Region’].unique().tolist()
label2id = {label: i for i, label in enumerate(unique_labels)}
id2label = {i: label for label, i in label2id.items()}

train_df[‘label’] = train_df[‘Region’].map(label2id)
tier1[‘label’] = tier1[‘Region’].map(label2id)

=======================

Tokenization

=======================

tokenizer = AutoTokenizer.from_pretrained(“csebuetnlp/banglabert”)

Prepare datasets

dataset_train = Dataset.from_pandas(train_df[[‘Text’, ‘label’]])
dataset_test = Dataset.from_pandas(tier1[[‘Text’, ‘label’]])

def tokenize(batch):
return tokenizer(batch[‘Text’], padding=“max_length”, truncation=True, max_length=128)

dataset_train = dataset_train.map(tokenize, batched=True)
dataset_test = dataset_test.map(tokenize, batched=True)

=======================

Model Loading

=======================

model = AutoModelForSequenceClassification.from_pretrained(
“csebuetnlp/banglabert”,
num_labels=len(unique_labels)
)

=======================

Training Arguments

=======================

training_args = TrainingArguments(

output_dir="./results_buetbert_tier2_tier3",
per_device_train_batch_size=8,
per_device_eval_batch_size=8,
num_train_epochs=3,
logging_dir="./logs_buetbert_tier2_tier3",
logging_steps=10,
eval_strategy="epoch",
save_strategy="epoch",
load_best_model_at_end=True,
metric_for_best_model="accuracy"

)

=======================

Compute Metrics

=======================

def compute_metrics(eval_pred):
logits, labels = eval_pred
predictions = logits.argmax(axis=-1)
precision, recall, f1, _ = precision_recall_fscore_support(labels, predictions, average=‘weighted’)
acc = accuracy_score(labels, predictions)
return {
‘accuracy’: acc,
‘f1’: f1,
‘precision’: precision,
‘recall’: recall
}

=======================

Trainer Setup

=======================

trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset_train,
eval_dataset=dataset_test,
tokenizer=tokenizer,
compute_metrics=compute_metrics
)

=======================

Training and Evaluation

=======================

trainer.train()
trainer.evaluate()

:white_check_mark: Add label mappings before saving

model.config.label2id = label2id
model.config.id2label = id2label

Save model

trainer.save_model(“/content/Drive/MyDrive/MSC thesis 25.04.2025/three_bert_tier2/buetbert_saved_model”)
for DistilBERT

=======================

Imports and Installations

=======================

import os
os.environ[“WANDB_DISABLED”] = “true”

First uninstall transformers fully

#!pip uninstall -y transformers

Install correct version manually

!pip install transformers==4.41.1

!pip install --upgrade transformers datasets --quiet
!pip install datasets
!pip install demoji

import pandas as pd
from sklearn.model_selection import train_test_split
from datasets import Dataset
from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer
from sklearn.metrics import accuracy_score, precision_recall_fscore_support
from collections import defaultdict
import torch
import demoji
import re

=======================

Mount Google Drive

=======================

from google.colab import drive
drive.mount(‘/content/Drive’)

Set working directory

os.chdir(‘/content/Drive/My Drive/MSC thesis 25.04.2025’)

=======================

Text Preprocessing Function

=======================

demoji.download_codes()

def preprocess_bangla_text(text):
text = demoji.replace(text, repl=“”)
whitespace = re.compile(r"[\s\u0020\u00a0\u1680\u180e\u202f\u205f\u3000\u2000-\u200a]+“, re.UNICODE)
bangla_fullstop = u”\u0964"
punctSeq = u"['"“”‘’]+|[.?!,…]+|[:;]+"
punc = u"[(),$%^&*+={}:"|'~`<>/,€¦!?½£¶™¼©⅐⅑⅒⅓⅔⅕⅖⅗⅘⅙⅚⅛⅜⅝⅞⅟↉¤¿º;-]+"
bangla_numbers = u"[০১২৩৪৫৬৭৮৯]+"

text = re.sub(punctSeq, " ", text)
text = re.sub(bangla_fullstop, " ", text)
text = re.sub(punc, " ", text)
text = re.sub(bangla_numbers, "", text)
text = whitespace.sub(" ", text).strip()
return text

=======================

Load Dataset

=======================

Load Tier 1 (strict - will be used for evaluation only)

tier1 = pd.read_excel(‘strict_cleaned_file.xlsx’)
tier1 = tier1[[‘Text’, ‘Region’]]

Load Tier 2 and Tier 3 (flexible and all-no for training)

tier2 = pd.read_excel(‘flexible_filtered.xlsx’)
tier2 = tier2[[‘Text’, ‘Region’]]
tier3 = pd.read_excel(‘all_no_cases.xlsx’)
tier3 = tier3[[‘Text’, ‘Region’]]

Combine Tier 2 + Tier 3 for training

train_df = pd.concat([tier2, tier3], ignore_index=True)

=======================

Preprocessing

=======================

tier1[‘Text’] = tier1[‘Text’].apply(preprocess_bangla_text)
train_df[‘Text’] = train_df[‘Text’].apply(preprocess_bangla_text)

=======================

Special Word-Based Oversampling

=======================

special_words_df = pd.read_excel(‘special_words.xlsx’)

special_vocab = defaultdict(set)
for _, row in special_words_df.iterrows():
special_vocab[row[‘region’]].add(str(row[‘word’]))

original_counts = train_df[‘Region’].value_counts().to_dict()
max_count = max(original_counts.values())

augmented_rows =

for region in train_df[‘Region’].unique():
regional_texts = train_df[train_df[‘Region’] == region]
current_count = len(regional_texts)
needed = max_count - current_count
if needed <= 0:
continue

texts_with_keywords = regional_texts[
    regional_texts['Text'].apply(lambda txt: any(word in str(txt) for word in special_vocab[region]))
]

if len(texts_with_keywords) == 0:
    continue

for _ in range(needed):
    sample_row = texts_with_keywords.sample(n=1, replace=True).iloc[0]
    augmented_rows.append(sample_row)

Append oversampled data

train_df = pd.concat([train_df, pd.DataFrame(augmented_rows)], ignore_index=True)

=======================

Encode Labels

=======================

unique_labels = tier1[‘Region’].unique().tolist()
label2id = {label: i for i, label in enumerate(unique_labels)}
id2label = {i: label for label, i in label2id.items()}

train_df[‘label’] = train_df[‘Region’].map(label2id)
tier1[‘label’] = tier1[‘Region’].map(label2id)

=======================

Tokenization

=======================

tokenizer = AutoTokenizer.from_pretrained(“distilbert-base-multilingual-cased”)

Prepare datasets

dataset_train = Dataset.from_pandas(train_df[[‘Text’, ‘label’]])
dataset_test = Dataset.from_pandas(tier1[[‘Text’, ‘label’]])

def tokenize(batch):
return tokenizer(batch[‘Text’], padding=“max_length”, truncation=True, max_length=128)

dataset_train = dataset_train.map(tokenize, batched=True)
dataset_test = dataset_test.map(tokenize, batched=True)

=======================

Model Loading

=======================

model = AutoModelForSequenceClassification.from_pretrained(
“distilbert-base-multilingual-cased”,
num_labels=len(unique_labels)
)

=======================

Training Arguments

=======================

training_args = TrainingArguments(

output_dir="./results_distilbert_tier2_tier3",
per_device_train_batch_size=8,
per_device_eval_batch_size=8,
num_train_epochs=3,
logging_dir="./logs_distilbert_tier2_tier3",
logging_steps=10,
eval_strategy="epoch",
save_strategy="epoch",
load_best_model_at_end=True,
metric_for_best_model="accuracy"

)

=======================

Compute Metrics

=======================

def compute_metrics(eval_pred):
logits, labels = eval_pred
predictions = logits.argmax(axis=-1)
precision, recall, f1, _ = precision_recall_fscore_support(labels, predictions, average=‘weighted’)
acc = accuracy_score(labels, predictions)
return {
‘accuracy’: acc,
‘f1’: f1,
‘precision’: precision,
‘recall’: recall
}

=======================

Trainer Setup

=======================

trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset_train,
eval_dataset=dataset_test,
tokenizer=tokenizer,
compute_metrics=compute_metrics
)

=======================

Training and Evaluation

=======================

trainer.train()
trainer.evaluate()

:white_check_mark: Add label mappings before saving

model.config.label2id = label2id
model.config.id2label = id2label

Save model

trainer.save_model(“/content/Drive/MyDrive/MSC thesis 25.04.2025/three_bert_tier2/distilbert_saved_model”)

However, after applying the ensemble, the accuracy dropped to 14%. I had saved the individual models to my drive before ensembling. Why did this happen?"
ensemble bert:

:soap: Uninstall broken numpy

!pip uninstall -y numpy

:white_check_mark: Reinstall clean and compatible numpy version

!pip install numpy==1.26.4
!pip install datasets==2.19.1

=== INSTALLATION ===

!pip install -q transformers==4.36.2 datasets openpyxl

=== IMPORTS ===

import pandas as pd
import numpy as np
import re
import torch
import matplotlib.pyplot as plt
from collections import defaultdict
from sklearn.metrics import accuracy_score, precision_recall_fscore_support, confusion_matrix, ConfusionMatrixDisplay
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from datasets import Dataset
from google.colab import drive

=== DRIVE MOUNT ===

drive.mount(‘/content/Drive’)

=== TEXT CLEANING ===

def preprocess_bangla_text(text):
whitespace = re.compile(r"[\s\u0020\u00a0\u1680\u180e\u202f\u205f\u3000\u2000-\u200a]+“, re.UNICODE)
bangla_fullstop = u”\u0964"
punctSeq = u"['"“”‘’]+|[.?!,…]+|[:;]+"
punc = u"[(),$%^&*+={}:"|'~`<>/!?;–—]"
bangla_numbers = u"[০-৯]+"
text = re.sub(punctSeq, " ", text)
text = re.sub(bangla_fullstop, " ", text)
text = re.sub(punc, " “, text)
text = re.sub(bangla_numbers, “”, text)
text = whitespace.sub(” ", text).strip()
return text

=== LOAD DATA ===

df = pd.read_excel(“/content/Drive/MyDrive/MSC thesis 25.04.2025/strict_cleaned_file.xlsx”)[[“Text”, “Region”]]
df[“Text”] = df[“Text”].astype(str).apply(preprocess_bangla_text)

unique_labels = sorted(df[“Region”].unique())
label2id = {label: i for i, label in enumerate(unique_labels)}
id2label = {i: label for label, i in label2id.items()}
df[“label”] = df[“Region”].map(label2id)

=== TOKENIZE DATASET ===

dataset = Dataset.from_pandas(df[[“Text”, “label”]])

=== MODEL INFO ===

models_info = {
“banglabert”: {
“path”: “/content/Drive/MyDrive/MSC thesis 25.04.2025/three_bert_tier2/banglabert_saved_model”,
“tokenizer”: AutoTokenizer.from_pretrained(“sagorsarker/bangla-bert-base”)
},
“buetbert”: {
“path”: “/content/Drive/MyDrive/MSC thesis 25.04.2025/three_bert_tier2/buetbert_saved_model”,
“tokenizer”: AutoTokenizer.from_pretrained(“csebuetnlp/banglabert”)
},
“distilbert”: {
“path”: “/content/Drive/MyDrive/MSC thesis 25.04.2025/three_bert_tier2/distilbert_saved_model”,
“tokenizer”: AutoTokenizer.from_pretrained(“distilbert-base-multilingual-cased”)
}
}

device = torch.device(“cuda” if torch.cuda.is_available() else “cpu”)

all_preds =

for name, model_info in models_info.items():

model = AutoModelForSequenceClassification.from_pretrained(model_info[“path”]).to(device)

tokenizer = model_info[“tokenizer”]

def tokenize_fn(batch):

return tokenizer(batch[“Text”], padding=“max_length”, truncation=True, max_length=128)

tokenized = dataset.map(tokenize_fn, batched=True)

tokenized.set_format(type=‘torch’, columns=[‘input_ids’, ‘attention_mask’, ‘label’]) # :green_circle: FIXED INDENTATION

preds =

with torch.no_grad():

for batch in torch.utils.data.DataLoader(tokenized, batch_size=16):

# :white_check_mark: Remove ‘label’ key before feeding into the model

batch = {k: v for k, v in batch.items() if k != ‘label’}

inputs = {k: v.to(device) for k, v in batch.items()}

outputs = model(**inputs)

logits = outputs.logits.cpu().numpy()

preds.extend(np.argmax(logits, axis=1))

all_preds.append(preds)

# === MAJORITY VOTING ===

# final_preds =

# all_preds_np = np.array(all_preds)

# for i in range(all_preds_np.shape[1]):

# votes = list(all_preds_np[:, i])

# final_preds.append(max(set(votes), key=votes.count))

# === MAJORITY VOTING (Safe Transpose Version) ===

all_preds_np = np.array(all_preds) # Shape: (num_models, num_samples)

print(“Prediction array shape (models, samples):”, all_preds_np.shape)

# Transpose to (num_samples, num_models) for row-wise voting

all_preds_np = all_preds_np.T

# Perform majority voting

final_preds = [np.bincount(row).argmax() for row in all_preds_np]

all_preds = # Start collecting predictions

for name, model_info in models_info.items():
model = AutoModelForSequenceClassification.from_pretrained(model_info[“path”]).to(device)
tokenizer = model_info[“tokenizer”]

def tokenize_fn(batch):
    return tokenizer(batch["Text"], padding="max_length", truncation=True, max_length=128)

tokenized = dataset.map(tokenize_fn, batched=True)
tokenized.set_format(type='torch', columns=['input_ids', 'attention_mask', 'label'])

preds = []
with torch.no_grad():
    for batch in torch.utils.data.DataLoader(tokenized, batch_size=16):
        batch = {k: v for k, v in batch.items() if k != 'label'}  # ✅ REMOVE label
        inputs = {k: v.to(device) for k, v in batch.items()}
        outputs = model(**inputs)

        logits = outputs.logits.cpu().numpy()
        preds.extend(np.argmax(logits, axis=1))

all_preds.append(preds)  # ✅ Still inside the model loop

:green_circle: :white_check_mark: Place this block OUTSIDE the for loop over models:

=== MAJORITY VOTING (Safe Transpose Version) ===

all_preds_np = np.array(all_preds)
print(“Prediction array shape (models, samples):”, all_preds_np.shape)

Transpose to shape (samples, models)

all_preds_np = all_preds_np.T

Majority vote

final_preds = [np.bincount(row).argmax() for row in all_preds_np]

y_true = df[“label”].tolist()

=== METRICS ===

acc = accuracy_score(y_true, final_preds)
precision, recall, f1, _ = precision_recall_fscore_support(y_true, final_preds, average=“weighted”)

print(“:bar_chart: Ensemble Evaluation Metrics:”)
print(f"Accuracy: {acc:.4f}“)
print(f"Precision: {precision:.4f}”)
print(f"Recall: {recall:.4f}“)
print(f"F1 Score: {f1:.4f}”)

=== CONFUSION MATRIX ===

cm = confusion_matrix(y_true, final_preds)
disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=unique_labels)
plt.figure(figsize=(10, 7))
disp.plot(cmap=“Blues”, xticks_rotation=45)
plt.title(“Confusion Matrix - Ensemble Voting”)
plt.show()

1 Like