Diabetic_Retinopathy_Detection_Using_CNN.ipynb

Kernel: Python 3

Vision Transformer Code: https://colab.research.google.com/drive/1UD_bybmnndzhi-tvuaqvHNHP5RWIq5zY?usp=sharing

from tensorflow import lite
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import numpy as np
import pandas as pd
import random, os
import shutil
import matplotlib.pyplot as plt
from matplotlib.image import imread
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.metrics import categorical_accuracy
from sklearn.model_selection import train_test_split

import kagglehub

# Download latest version
path = kagglehub.dataset_download("sovitrath/diabetic-retinopathy-224x224-2019-data")

print("Path to dataset files:", path)

Downloading from https://www.kaggle.com/api/v1/datasets/download/sovitrath/diabetic-retinopathy-224x224-2019-data?dataset_version_number=4...

100%|██████████| 238M/238M [00:02<00:00, 95.5MB/s]

Extracting files...

Path to dataset files: /root/.cache/kagglehub/datasets/sovitrath/diabetic-retinopathy-224x224-2019-data/versions/4

# Add an additional column, mapping to the type
df = pd.read_csv(r'../root/.cache/kagglehub/datasets/sovitrath/diabetic-retinopathy-224x224-2019-data/versions/4/train.csv')

diagnosis_dict_binary = {
    0: 'No_DR',
    1: 'DR',
    2: 'DR',
    3: 'DR',
    4: 'DR'
}

diagnosis_dict = {
    0: 'No_DR',
    1: 'Mild',
    2: 'Moderate',
    3: 'Severe',
    4: 'Proliferate_DR',
}


df['binary_type'] =  df['diagnosis'].map(diagnosis_dict_binary.get)
df['type'] = df['diagnosis'].map(diagnosis_dict.get)
df.head()

train_intermediate, val = train_test_split(df, test_size = 0.15, stratify = df['type'])
train, test = train_test_split(train_intermediate, test_size = 0.15 / (1 - 0.15), stratify = train_intermediate['type'])

#print number
print(train['type'].value_counts(), '\n')

type
No_DR             1263
Moderate           699
Mild               258
Proliferate_DR     207
Severe             135
Name: count, dtype: int64 

#create directories for train, val and test. update it everytime the code runs with shutil.rmtree
base_dir = '.\dataset'


train_dir = os.path.join(base_dir, 'train')
val_dir = os.path.join(base_dir, 'val')
test_dir = os.path.join(base_dir, 'test')

if os.path.exists(base_dir):
    shutil.rmtree(base_dir)

#make directories for train, val and test
os.makedirs(train_dir, exist_ok=True)
os.makedirs(val_dir, exist_ok=True)
os.makedirs(test_dir, exist_ok=True)
#ensure the directories have the same columns
valid_types = ['No_DR', 'Mild', 'Moderate', 'Severe', 'Proliferate_DR']
df = df[df['type'].isin(valid_types)]
assert all(train['type'].isin(valid_types))
assert all(val['type'].isin(valid_types))
assert all(test['type'].isin(valid_types))

#copy the files from the sourcefile to train, validation and test directory respectively
src_dir = r'../root/.cache/kagglehub/datasets/sovitrath/diabetic-retinopathy-224x224-2019-data/versions/4/colored_images'
for index, row in train.iterrows():
    diagnosis = row['type']
    binary_diagnosis = row['binary_type']
    id_code = row['id_code'] + ".png"
    srcfile = os.path.join(src_dir, diagnosis, id_code)
    dstfile = os.path.join(train_dir, diagnosis)
    os.makedirs(dstfile, exist_ok = True)
    if os.path.exists(srcfile):
      shutil.copy(srcfile, dstfile)
for index, row in val.iterrows():
    diagnosis = row['type']
    binary_diagnosis = row['binary_type']
    id_code = row['id_code'] + ".png"
    srcfile = os.path.join(src_dir, diagnosis, id_code)
    dstfile = os.path.join(val_dir, diagnosis)
    os.makedirs(dstfile, exist_ok=True)
    if os.path.exists(srcfile):
        shutil.copy(srcfile, dstfile)

for index, row in test.iterrows():
    diagnosis = row['type']
    binary_diagnosis = row['binary_type']
    id_code = row['id_code'] + ".png"
    srcfile = os.path.join(src_dir, diagnosis, id_code)
    dstfile = os.path.join(test_dir, diagnosis)
    os.makedirs(dstfile, exist_ok=True)
    if os.path.exists(srcfile):
        shutil.copy(srcfile, dstfile)





for subdir in [train_dir, val_dir, test_dir]:
    print(f"\nContents of {subdir}:")
    for root, dirs, files in os.walk(subdir):
        print(f"{root}: {len(files)} files")

Contents of .\dataset/train:
.\dataset/train: 0 files
.\dataset/train/Moderate: 699 files
.\dataset/train/Mild: 258 files
.\dataset/train/Severe: 135 files
.\dataset/train/Proliferate_DR: 207 files
.\dataset/train/No_DR: 1263 files

Contents of .\dataset/val:
.\dataset/val: 0 files
.\dataset/val/Moderate: 150 files
.\dataset/val/Mild: 56 files
.\dataset/val/Severe: 29 files
.\dataset/val/Proliferate_DR: 44 files
.\dataset/val/No_DR: 271 files

Contents of .\dataset/test:
.\dataset/test: 0 files
.\dataset/test/Moderate: 150 files
.\dataset/test/Mild: 56 files
.\dataset/test/Severe: 29 files
.\dataset/test/Proliferate_DR: 44 files
.\dataset/test/No_DR: 271 files

train_path = train_dir
val_path = val_dir
test_path = test_dir

train_batches = ImageDataGenerator(rescale = 1./255).flow_from_directory(train_path, target_size=(224,224), shuffle = True)
val_batches = ImageDataGenerator(rescale = 1./255).flow_from_directory(val_path, target_size=(224,224), shuffle = True)
test_batches = ImageDataGenerator(rescale = 1./255).flow_from_directory(test_path, target_size=(224,224), shuffle = False)

Found 2562 images belonging to 5 classes.
Found 550 images belonging to 5 classes.
Found 550 images belonging to 5 classes.

#cnn model
model = tf.keras.Sequential([
    layers.Conv2D(8, (3,3), padding="valid", input_shape=(224,224,3), activation = 'relu'),
    layers.MaxPooling2D(pool_size=(2,2)),
    layers.BatchNormalization(),

    layers.Conv2D(16, (3,3), padding="valid", activation = 'relu'),
    layers.MaxPooling2D(pool_size=(2,2)),
    layers.BatchNormalization(),

    layers.Conv2D(32, (4,4), padding="valid", activation = 'relu'),
    layers.MaxPooling2D(pool_size=(2,2)),
    layers.BatchNormalization(),
    layers.Flatten(),
    layers.Dense(32, activation = 'relu'),
    layers.Dropout(0.15),
    layers.Dense(5, activation = 'softmax')
])

model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=1e-5),
              loss='categorical_crossentropy',
              metrics=['accuracy'])


history = model.fit(train_batches,
                    epochs=50,
                    validation_data=val_batches)

/usr/local/lib/python3.10/dist-packages/keras/src/layers/convolutional/base_conv.py:107: UserWarning: Do not pass an `input_shape`/`input_dim` argument to a layer. When using Sequential models, prefer using an `Input(shape)` object as the first layer in the model instead.
  super().__init__(activity_regularizer=activity_regularizer, **kwargs)

Epoch 1/10

/usr/local/lib/python3.10/dist-packages/keras/src/trainers/data_adapters/py_dataset_adapter.py:122: UserWarning: Your `PyDataset` class should call `super().__init__(**kwargs)` in its constructor. `**kwargs` can include `workers`, `use_multiprocessing`, `max_queue_size`. Do not pass these arguments to `fit()`, as they will be ignored.
  self._warn_if_super_not_called()

81/81 ━━━━━━━━━━━━━━━━━━━━ 101s 1s/step - accuracy: 0.4029 - loss: 1.5315 - val_accuracy: 0.5673 - val_loss: 1.5060
Epoch 2/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 145s 1s/step - accuracy: 0.6735 - loss: 0.9081 - val_accuracy: 0.6727 - val_loss: 1.3945
Epoch 3/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 139s 1s/step - accuracy: 0.7026 - loss: 0.8166 - val_accuracy: 0.6945 - val_loss: 1.2591
Epoch 4/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 142s 1s/step - accuracy: 0.7062 - loss: 0.8137 - val_accuracy: 0.6855 - val_loss: 1.1358
Epoch 5/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 144s 1s/step - accuracy: 0.7130 - loss: 0.7965 - val_accuracy: 0.7127 - val_loss: 0.9882
Epoch 6/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 139s 1s/step - accuracy: 0.7532 - loss: 0.7223 - val_accuracy: 0.7327 - val_loss: 0.8859
Epoch 7/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 144s 1s/step - accuracy: 0.7584 - loss: 0.7100 - val_accuracy: 0.7382 - val_loss: 0.8226
Epoch 8/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 94s 1s/step - accuracy: 0.7351 - loss: 0.7226 - val_accuracy: 0.7327 - val_loss: 0.7908
Epoch 9/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 143s 1s/step - accuracy: 0.7378 - loss: 0.7089 - val_accuracy: 0.7545 - val_loss: 0.7670
Epoch 10/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 140s 1s/step - accuracy: 0.7330 - loss: 0.7138 - val_accuracy: 0.7545 - val_loss: 0.7665

import matplotlib.pyplot as plt
import numpy as np
from sklearn.metrics import confusion_matrix, f1_score, precision_score, recall_score

# Custom Callback to compute metrics after each epoch
class MetricsCallback(tf.keras.callbacks.Callback):
    def __init__(self, validation_batches):
        self.val_batches = validation_batches
        self.sensitivity = []
        self.specificity = []
        self.f1_scores = []

    def on_epoch_end(self, epoch, logs=None):
        # Get predictions and ground truths
        y_pred_probs = self.model.predict(self.val_batches)  # Probabilities
        y_pred = np.argmax(y_pred_probs, axis=1)  # Predicted classes
        y_true = self.val_batches.labels  # True labels

        # Compute confusion matrix
        cm = confusion_matrix(y_true, y_pred)
        tp = np.diag(cm)  # True positives for each class
        fp = np.sum(cm, axis=0) - tp  # False positives for each class
        fn = np.sum(cm, axis=1) - tp  # False negatives for each class
        tn = np.sum(cm) - (tp + fp + fn)  # True negatives for each class

        # Avoid division by zero
        recall = tp / (tp + fn + np.finfo(float).eps)  # Sensitivity (Recall)
        specificity = tn / (tn + fp + np.finfo(float).eps)
        precision = tp / (tp + fp + np.finfo(float).eps)
        f1 = 2 * (precision * recall) / (precision + recall + np.finfo(float).eps)

        # Store averages across classes
        self.sensitivity.append(np.mean(recall))
        self.specificity.append(np.mean(specificity))
        self.f1_scores.append(np.mean(f1))

# Initialize callback
metrics_callback = MetricsCallback(validation_batches=val_batches)

# Train the model with the callback
history = model.fit(train_batches,
                    epochs=50,
                    validation_data=val_batches,
                    callbacks=[metrics_callback])

# Plot Accuracy
plt.plot(history.history['accuracy'], label='Train Accuracy')
plt.plot(history.history['val_accuracy'], label='Validation Accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.title('Accuracy vs Epochs')
plt.legend()
plt.savefig('cnn acc vs epoch.png')
plt.show()

# Plot Sensitivity, Specificity, and F1-score
epochs = range(1, len(metrics_callback.sensitivity) + 1)
plt.plot(epochs, metrics_callback.sensitivity, label='Sensitivity')
plt.plot(epochs, metrics_callback.specificity, label='Specificity')
plt.plot(epochs, metrics_callback.f1_scores, label='F1 Score')
plt.xlabel('Epochs')
plt.ylabel('Metrics')
plt.title('Sensitivity, Specificity, and F1-Score vs Epochs')
plt.legend()
plt.savefig('cnn spec f1 vs epoch.png')
plt.show()

Epoch 1/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 440ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 107s 1s/step - accuracy: 0.8073 - loss: 0.5413 - val_accuracy: 0.7327 - val_loss: 0.7587
Epoch 2/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 418ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 138s 1s/step - accuracy: 0.8078 - loss: 0.5459 - val_accuracy: 0.7400 - val_loss: 0.7507
Epoch 3/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 327ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 141s 1s/step - accuracy: 0.8205 - loss: 0.5216 - val_accuracy: 0.7327 - val_loss: 0.7526
Epoch 4/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 320ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 104s 1s/step - accuracy: 0.8220 - loss: 0.5207 - val_accuracy: 0.7327 - val_loss: 0.7552
Epoch 5/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 332ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 102s 1s/step - accuracy: 0.8213 - loss: 0.5000 - val_accuracy: 0.7364 - val_loss: 0.7481
Epoch 6/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 316ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 107s 1s/step - accuracy: 0.8227 - loss: 0.5007 - val_accuracy: 0.7400 - val_loss: 0.7574
Epoch 7/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 429ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 101s 1s/step - accuracy: 0.8283 - loss: 0.4953 - val_accuracy: 0.7364 - val_loss: 0.7528
Epoch 8/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 324ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 144s 1s/step - accuracy: 0.8353 - loss: 0.4914 - val_accuracy: 0.7418 - val_loss: 0.7406
Epoch 9/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 326ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 104s 1s/step - accuracy: 0.8286 - loss: 0.4643 - val_accuracy: 0.7382 - val_loss: 0.7535
Epoch 10/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 322ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 107s 1s/step - accuracy: 0.8268 - loss: 0.4771 - val_accuracy: 0.7436 - val_loss: 0.7535
Epoch 11/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 432ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 101s 1s/step - accuracy: 0.8484 - loss: 0.4625 - val_accuracy: 0.7382 - val_loss: 0.7537
Epoch 12/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 331ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 144s 1s/step - accuracy: 0.8515 - loss: 0.4411 - val_accuracy: 0.7345 - val_loss: 0.7652
Epoch 13/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 322ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 106s 1s/step - accuracy: 0.8459 - loss: 0.4426 - val_accuracy: 0.7364 - val_loss: 0.7536
Epoch 14/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 326ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 111s 1s/step - accuracy: 0.8416 - loss: 0.4430 - val_accuracy: 0.7345 - val_loss: 0.7559
Epoch 15/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 324ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 138s 1s/step - accuracy: 0.8490 - loss: 0.4353 - val_accuracy: 0.7455 - val_loss: 0.7509
Epoch 16/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 436ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 103s 1s/step - accuracy: 0.8572 - loss: 0.4196 - val_accuracy: 0.7400 - val_loss: 0.7536
Epoch 17/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 337ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 145s 1s/step - accuracy: 0.8584 - loss: 0.4279 - val_accuracy: 0.7509 - val_loss: 0.7486
Epoch 18/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 7s 380ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 109s 1s/step - accuracy: 0.8700 - loss: 0.4032 - val_accuracy: 0.7491 - val_loss: 0.7484
Epoch 19/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 317ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 105s 1s/step - accuracy: 0.8678 - loss: 0.4065 - val_accuracy: 0.7436 - val_loss: 0.7501
Epoch 20/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 329ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 141s 1s/step - accuracy: 0.8708 - loss: 0.3996 - val_accuracy: 0.7473 - val_loss: 0.7510
Epoch 21/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 433ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 103s 1s/step - accuracy: 0.8808 - loss: 0.3935 - val_accuracy: 0.7382 - val_loss: 0.7557
Epoch 22/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 434ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 105s 1s/step - accuracy: 0.8671 - loss: 0.4118 - val_accuracy: 0.7400 - val_loss: 0.7632
Epoch 23/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 429ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 102s 1s/step - accuracy: 0.8693 - loss: 0.3885 - val_accuracy: 0.7527 - val_loss: 0.7595
Epoch 24/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 314ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 141s 1s/step - accuracy: 0.8671 - loss: 0.3895 - val_accuracy: 0.7418 - val_loss: 0.7526
Epoch 25/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 433ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 147s 1s/step - accuracy: 0.8796 - loss: 0.3842 - val_accuracy: 0.7364 - val_loss: 0.7678
Epoch 26/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 429ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 139s 1s/step - accuracy: 0.8792 - loss: 0.3600 - val_accuracy: 0.7491 - val_loss: 0.7655
Epoch 27/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 431ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 101s 1s/step - accuracy: 0.8941 - loss: 0.3391 - val_accuracy: 0.7400 - val_loss: 0.7691
Epoch 28/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 321ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 145s 1s/step - accuracy: 0.8877 - loss: 0.3532 - val_accuracy: 0.7418 - val_loss: 0.7658
Epoch 29/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 325ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 102s 1s/step - accuracy: 0.8948 - loss: 0.3400 - val_accuracy: 0.7436 - val_loss: 0.7761
Epoch 30/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 324ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 147s 1s/step - accuracy: 0.8936 - loss: 0.3344 - val_accuracy: 0.7527 - val_loss: 0.7652
Epoch 31/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 320ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 137s 1s/step - accuracy: 0.8954 - loss: 0.3314 - val_accuracy: 0.7327 - val_loss: 0.7747
Epoch 32/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 7s 370ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 108s 1s/step - accuracy: 0.9006 - loss: 0.3306 - val_accuracy: 0.7436 - val_loss: 0.7662
Epoch 33/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 433ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 137s 1s/step - accuracy: 0.8943 - loss: 0.3268 - val_accuracy: 0.7436 - val_loss: 0.7721
Epoch 34/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 314ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 142s 1s/step - accuracy: 0.9086 - loss: 0.3117 - val_accuracy: 0.7436 - val_loss: 0.7863
Epoch 35/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 315ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 100s 1s/step - accuracy: 0.9065 - loss: 0.2979 - val_accuracy: 0.7418 - val_loss: 0.7767
Epoch 36/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 334ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 106s 1s/step - accuracy: 0.9149 - loss: 0.3005 - val_accuracy: 0.7455 - val_loss: 0.7790
Epoch 37/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 431ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 136s 1s/step - accuracy: 0.9074 - loss: 0.2953 - val_accuracy: 0.7364 - val_loss: 0.7779
Epoch 38/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 320ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 143s 1s/step - accuracy: 0.9103 - loss: 0.3055 - val_accuracy: 0.7509 - val_loss: 0.7785
Epoch 39/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 9s 491ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 105s 1s/step - accuracy: 0.9185 - loss: 0.2735 - val_accuracy: 0.7418 - val_loss: 0.7780
Epoch 40/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 422ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 137s 1s/step - accuracy: 0.9141 - loss: 0.2989 - val_accuracy: 0.7364 - val_loss: 0.7859
Epoch 41/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 9s 506ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 102s 1s/step - accuracy: 0.9229 - loss: 0.2660 - val_accuracy: 0.7382 - val_loss: 0.7951
Epoch 42/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 429ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 100s 1s/step - accuracy: 0.9228 - loss: 0.2747 - val_accuracy: 0.7491 - val_loss: 0.7982
Epoch 43/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 422ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 100s 1s/step - accuracy: 0.9129 - loss: 0.2866 - val_accuracy: 0.7345 - val_loss: 0.7970
Epoch 44/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 321ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 145s 1s/step - accuracy: 0.9183 - loss: 0.2825 - val_accuracy: 0.7382 - val_loss: 0.7967
Epoch 45/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 321ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 100s 1s/step - accuracy: 0.9215 - loss: 0.2690 - val_accuracy: 0.7436 - val_loss: 0.7990
Epoch 46/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 317ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 102s 1s/step - accuracy: 0.9264 - loss: 0.2446 - val_accuracy: 0.7364 - val_loss: 0.7988
Epoch 47/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 7s 395ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 143s 1s/step - accuracy: 0.9116 - loss: 0.2727 - val_accuracy: 0.7436 - val_loss: 0.7939
Epoch 48/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 317ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 140s 1s/step - accuracy: 0.9241 - loss: 0.2640 - val_accuracy: 0.7509 - val_loss: 0.7995
Epoch 49/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 10s 537ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 143s 1s/step - accuracy: 0.9366 - loss: 0.2348 - val_accuracy: 0.7545 - val_loss: 0.7973
Epoch 50/50
18/18 ━━━━━━━━━━━━━━━━━━━━ 7s 396ms/step
81/81 ━━━━━━━━━━━━━━━━━━━━ 140s 1s/step - accuracy: 0.9305 - loss: 0.2520 - val_accuracy: 0.7455 - val_loss: 0.7954

files.download('cnn acc vs epoch.png')
files.download('cnn spec f1 vs epoch.png')

<IPython.core.display.Javascript object>

<IPython.core.display.Javascript object>

<IPython.core.display.Javascript object>

<IPython.core.display.Javascript object>

model.save('64x3-CNN.keras')

test_loss, test_accuracy = model.evaluate(test_batches)
print(f"Test Accuracy: {test_accuracy * 100:.2f}%")

18/18 ━━━━━━━━━━━━━━━━━━━━ 6s 329ms/step - accuracy: 0.7009 - loss: 0.8493
Test Accuracy: 73.82%

from sklearn.metrics import roc_curve, auc
from sklearn.preprocessing import label_binarize
import matplotlib.pyplot as plt
import numpy as np
y_test = test_batches.labels
# Binarize the output (e.g., for 5 classes)
y_test_bin = label_binarize(y_test, classes=[0, 1, 2, 3, 4])
n_classes = y_test_bin.shape[1]  # Number of classes

# Ensure predictions are also in the right shape
y_pred_probs = model.predict(test_batches)  # Probabilities for all classes

# Compute ROC curve and AUC for each class
fpr = {}
tpr = {}
roc_auc = {}

for i in range(n_classes):
    fpr[i], tpr[i], _ = roc_curve(y_test_bin[:, i], y_pred_probs[:, i])
    roc_auc[i] = auc(fpr[i], tpr[i])

# Plot all ROC curves
plt.figure()
colors = ['blue', 'green', 'red', 'cyan', 'magenta']
for i in range(n_classes):
    plt.plot(fpr[i], tpr[i], color=colors[i], lw=2,
             label=f'Class {i} (AUC = {roc_auc[i]:.2f})')

# Add random guess line
plt.plot([0, 1], [0, 1], color='gray', linestyle='--')
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver Operating Characteristic for Multi-Class')
plt.legend(loc='lower right')
plt.savefig('cnn roc curve.png')
plt.show()

/usr/local/lib/python3.10/dist-packages/keras/src/trainers/data_adapters/py_dataset_adapter.py:122: UserWarning: Your `PyDataset` class should call `super().__init__(**kwargs)` in its constructor. `**kwargs` can include `workers`, `use_multiprocessing`, `max_queue_size`. Do not pass these arguments to `fit()`, as they will be ignored.
  self._warn_if_super_not_called()

18/18 ━━━━━━━━━━━━━━━━━━━━ 8s 435ms/step

from google.colab import files
files.download('cnn roc curve.png')

<IPython.core.display.Javascript object>

<IPython.core.display.Javascript object>

#generate confusion matrix
from sklearn.metrics import classification_report
y_pred = np.argmax(y_pred_probs, axis=1)  # Predicted class labels
print(classification_report(y_test, y_pred, target_names=test_batches.class_indices.keys()))
from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
cm = confusion_matrix(y_test, y_pred)
disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=test_batches.class_indices.keys())
disp.plot(cmap="Blues")
plt.savefig('cnn confusion matrix.png')
plt.show()

                precision    recall  f1-score   support

          Mild       0.47      0.50      0.49        56
      Moderate       0.60      0.73      0.66       150
         No_DR       0.89      0.97      0.93       271
Proliferate_DR       0.33      0.02      0.04        44
        Severe       0.60      0.21      0.31        29

      accuracy                           0.74       550
     macro avg       0.58      0.49      0.49       550
  weighted avg       0.71      0.74      0.71       550

from google.colab import files
files.download('cnn confusion matrix.png')

<IPython.core.display.Javascript object>

<IPython.core.display.Javascript object>

#evaluate training and validation accuracy on a iteratively halving dataset
from math import ceil, pow
src_dir = r'../root/.cache/kagglehub/datasets/sovitrath/diabetic-retinopathy-224x224-2019-data/versions/4/colored_images'
for i in range(5):
  total_files = len(train)
  denom = pow(2, i)
  print(denom)
  files_to_copy = ceil(total_files / denom)  # Half the files, rounded up if odd
  print(files_to_copy)

  if os.path.exists(train_dir):
    shutil.rmtree(train_dir)
  os.makedirs(train_dir, exist_ok=True)


  # Counter to track the number of files copied
  copied_files = 0

  for index, row in train.iterrows():
    if copied_files >= files_to_copy:
        break  # Stop when half of the files are copied

    diagnosis = row['type']
    binary_diagnosis = row['binary_type']
    id_code = row['id_code'] + ".png"
    srcfile = os.path.join(src_dir, diagnosis, id_code)
    dstfile = os.path.join(train_dir, binary_diagnosis)
    os.makedirs(dstfile, exist_ok=True)

    if os.path.exists(srcfile):
        shutil.copy(srcfile, dstfile)
        copied_files += 1  # Increment the counter

  for index, row in val.iterrows():
    diagnosis = row['type']
    binary_diagnosis = row['binary_type']
    id_code = row['id_code'] + ".png"
    srcfile = os.path.join(src_dir, diagnosis, id_code)
    dstfile = os.path.join(val_dir, binary_diagnosis)
    os.makedirs(dstfile, exist_ok=True)
    if os.path.exists(srcfile):
        shutil.copy(srcfile, dstfile)

  for index, row in test.iterrows():
    diagnosis = row['type']
    binary_diagnosis = row['binary_type']
    id_code = row['id_code'] + ".png"
    srcfile = os.path.join(src_dir, diagnosis, id_code)
    dstfile = os.path.join(test_dir, binary_diagnosis)
    os.makedirs(dstfile, exist_ok=True)
    if os.path.exists(srcfile):
        shutil.copy(srcfile, dstfile)





  for subdir in [train_dir, val_dir, test_dir]:
    print(f"\nContents of {subdir}:")
    for root, dirs, files in os.walk(subdir):
        print(f"{root}: {len(files)} files")

  train_path = train_dir
  val_path = val_dir
  test_path = test_dir

  train_batches = ImageDataGenerator(rescale = 1./255).flow_from_directory(train_path, target_size=(224,224), shuffle = True)
  val_batches = ImageDataGenerator(rescale = 1./255).flow_from_directory(val_path, target_size=(224,224), shuffle = True)
  test_batches = ImageDataGenerator(rescale = 1./255).flow_from_directory(test_path, target_size=(224,224), shuffle = False)

  #cnn model
  model = tf.keras.Sequential([
      layers.Conv2D(8, (3,3), padding="valid", input_shape=(224,224,3), activation = 'relu'),
      layers.MaxPooling2D(pool_size=(2,2)),
      layers.BatchNormalization(),

      layers.Conv2D(16, (3,3), padding="valid", activation = 'relu'),
      layers.MaxPooling2D(pool_size=(2,2)),
      layers.BatchNormalization(),

      layers.Conv2D(32, (4,4), padding="valid", activation = 'relu'),
      layers.MaxPooling2D(pool_size=(2,2)),
      layers.BatchNormalization(),
      layers.Flatten(),
      layers.Dense(32, activation = 'relu'),
      layers.Dropout(0.15),
      layers.Dense(2, activation = 'softmax')
  ])

  model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=1e-5),
                loss='categorical_crossentropy',
                metrics=['accuracy'])


  history = model.fit(train_batches,
                      epochs=10,
                      validation_data=val_batches)

1.0
2562

Contents of .\dataset/train:
.\dataset/train: 0 files
.\dataset/train/No_DR: 1263 files
.\dataset/train/DR: 1299 files

Contents of .\dataset/val:
.\dataset/val: 0 files
.\dataset/val/No_DR: 271 files
.\dataset/val/DR: 279 files

Contents of .\dataset/test:
.\dataset/test: 0 files
.\dataset/test/No_DR: 271 files
.\dataset/test/DR: 279 files
Found 2562 images belonging to 2 classes.
Found 550 images belonging to 2 classes.
Found 550 images belonging to 2 classes.

/usr/local/lib/python3.10/dist-packages/keras/src/layers/convolutional/base_conv.py:107: UserWarning: Do not pass an `input_shape`/`input_dim` argument to a layer. When using Sequential models, prefer using an `Input(shape)` object as the first layer in the model instead.
  super().__init__(activity_regularizer=activity_regularizer, **kwargs)

Epoch 1/10

/usr/local/lib/python3.10/dist-packages/keras/src/trainers/data_adapters/py_dataset_adapter.py:122: UserWarning: Your `PyDataset` class should call `super().__init__(**kwargs)` in its constructor. `**kwargs` can include `workers`, `use_multiprocessing`, `max_queue_size`. Do not pass these arguments to `fit()`, as they will be ignored.
  self._warn_if_super_not_called()

81/81 ━━━━━━━━━━━━━━━━━━━━ 95s 1s/step - accuracy: 0.7844 - loss: 0.4764 - val_accuracy: 0.5073 - val_loss: 0.8788
Epoch 2/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 138s 1s/step - accuracy: 0.9100 - loss: 0.2328 - val_accuracy: 0.5073 - val_loss: 1.0356
Epoch 3/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 87s 1s/step - accuracy: 0.9173 - loss: 0.2120 - val_accuracy: 0.5073 - val_loss: 1.1552
Epoch 4/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 144s 1s/step - accuracy: 0.9285 - loss: 0.1944 - val_accuracy: 0.5109 - val_loss: 0.9722
Epoch 5/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 87s 1s/step - accuracy: 0.9311 - loss: 0.1797 - val_accuracy: 0.5491 - val_loss: 0.6548
Epoch 6/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 143s 1s/step - accuracy: 0.9346 - loss: 0.1593 - val_accuracy: 0.7582 - val_loss: 0.4198
Epoch 7/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 86s 1s/step - accuracy: 0.9482 - loss: 0.1526 - val_accuracy: 0.9164 - val_loss: 0.2454
Epoch 8/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 89s 1s/step - accuracy: 0.9395 - loss: 0.1521 - val_accuracy: 0.9236 - val_loss: 0.1954
Epoch 9/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 87s 1s/step - accuracy: 0.9404 - loss: 0.1574 - val_accuracy: 0.9309 - val_loss: 0.1698
Epoch 10/10
81/81 ━━━━━━━━━━━━━━━━━━━━ 143s 1s/step - accuracy: 0.9452 - loss: 0.1475 - val_accuracy: 0.9327 - val_loss: 0.1669
2.0
1281

Contents of .\dataset/train:
.\dataset/train: 0 files
.\dataset/train/No_DR: 664 files
.\dataset/train/DR: 617 files

Contents of .\dataset/val:
.\dataset/val: 0 files
.\dataset/val/No_DR: 271 files
.\dataset/val/DR: 279 files

Contents of .\dataset/test:
.\dataset/test: 0 files
.\dataset/test/No_DR: 271 files
.\dataset/test/DR: 279 files
Found 1281 images belonging to 2 classes.
Found 550 images belonging to 2 classes.
Found 550 images belonging to 2 classes.
Epoch 1/10
41/41 ━━━━━━━━━━━━━━━━━━━━ 52s 1s/step - accuracy: 0.7730 - loss: 0.4892 - val_accuracy: 0.5073 - val_loss: 0.6800
Epoch 2/10
41/41 ━━━━━━━━━━━━━━━━━━━━ 47s 1s/step - accuracy: 0.8863 - loss: 0.2912 - val_accuracy: 0.5073 - val_loss: 0.6994
Epoch 3/10
41/41 ━━━━━━━━━━━━━━━━━━━━ 51s 1s/step - accuracy: 0.8942 - loss: 0.2338 - val_accuracy: 0.5073 - val_loss: 0.7084
Epoch 4/10
41/41 ━━━━━━━━━━━━━━━━━━━━ 79s 1s/step - accuracy: 0.9099 - loss: 0.2347 - val_accuracy: 0.5073 - val_loss: 0.7110
Epoch 5/10
41/41 ━━━━━━━━━━━━━━━━━━━━ 47s 1s/step - accuracy: 0.9302 - loss: 0.2254 - val_accuracy: 0.5127 - val_loss: 0.6887
Epoch 6/10
41/41 ━━━━━━━━━━━━━━━━━━━━ 52s 1s/step - accuracy: 0.9352 - loss: 0.1886 - val_accuracy: 0.5218 - val_loss: 0.6459
Epoch 7/10
41/41 ━━━━━━━━━━━━━━━━━━━━ 79s 1s/step - accuracy: 0.9431 - loss: 0.1787 - val_accuracy: 0.5327 - val_loss: 0.6414
Epoch 8/10
41/41 ━━━━━━━━━━━━━━━━━━━━ 80s 1s/step - accuracy: 0.9385 - loss: 0.1534 - val_accuracy: 0.5982 - val_loss: 0.5673
Epoch 9/10
41/41 ━━━━━━━━━━━━━━━━━━━━ 82s 1s/step - accuracy: 0.9442 - loss: 0.1668 - val_accuracy: 0.7182 - val_loss: 0.4592
Epoch 10/10
41/41 ━━━━━━━━━━━━━━━━━━━━ 49s 1s/step - accuracy: 0.9510 - loss: 0.1555 - val_accuracy: 0.7745 - val_loss: 0.4030
4.0
641

Contents of .\dataset/train:
.\dataset/train: 0 files
.\dataset/train/No_DR: 335 files
.\dataset/train/DR: 306 files

Contents of .\dataset/val:
.\dataset/val: 0 files
.\dataset/val/No_DR: 271 files
.\dataset/val/DR: 279 files

Contents of .\dataset/test:
.\dataset/test: 0 files
.\dataset/test/No_DR: 271 files
.\dataset/test/DR: 279 files
Found 641 images belonging to 2 classes.
Found 550 images belonging to 2 classes.
Found 550 images belonging to 2 classes.
Epoch 1/10
21/21 ━━━━━━━━━━━━━━━━━━━━ 30s 1s/step - accuracy: 0.6456 - loss: 0.6968 - val_accuracy: 0.6873 - val_loss: 0.6828
Epoch 2/10
21/21 ━━━━━━━━━━━━━━━━━━━━ 27s 1s/step - accuracy: 0.8668 - loss: 0.3267 - val_accuracy: 0.7309 - val_loss: 0.6757
Epoch 3/10
21/21 ━━━━━━━━━━━━━━━━━━━━ 27s 1s/step - accuracy: 0.8835 - loss: 0.2830 - val_accuracy: 0.5982 - val_loss: 0.6689
Epoch 4/10
21/21 ━━━━━━━━━━━━━━━━━━━━ 41s 1s/step - accuracy: 0.8973 - loss: 0.2757 - val_accuracy: 0.5145 - val_loss: 0.6677
Epoch 5/10
21/21 ━━━━━━━━━━━━━━━━━━━━ 41s 1s/step - accuracy: 0.9076 - loss: 0.2304 - val_accuracy: 0.5145 - val_loss: 0.6660
Epoch 6/10
21/21 ━━━━━━━━━━━━━━━━━━━━ 27s 1s/step - accuracy: 0.9293 - loss: 0.1915 - val_accuracy: 0.5145 - val_loss: 0.6673
Epoch 7/10
21/21 ━━━━━━━━━━━━━━━━━━━━ 27s 1s/step - accuracy: 0.9245 - loss: 0.1830 - val_accuracy: 0.5127 - val_loss: 0.6822
Epoch 8/10
21/21 ━━━━━━━━━━━━━━━━━━━━ 27s 1s/step - accuracy: 0.9402 - loss: 0.1654 - val_accuracy: 0.5436 - val_loss: 0.6704
Epoch 9/10
21/21 ━━━━━━━━━━━━━━━━━━━━ 27s 1s/step - accuracy: 0.9411 - loss: 0.1661 - val_accuracy: 0.5455 - val_loss: 0.6825
Epoch 10/10
21/21 ━━━━━━━━━━━━━━━━━━━━ 30s 1s/step - accuracy: 0.9319 - loss: 0.1554 - val_accuracy: 0.5618 - val_loss: 0.6682
8.0
321

Contents of .\dataset/train:
.\dataset/train: 0 files
.\dataset/train/No_DR: 168 files
.\dataset/train/DR: 153 files

Contents of .\dataset/val:
.\dataset/val: 0 files
.\dataset/val/No_DR: 271 files
.\dataset/val/DR: 279 files

Contents of .\dataset/test:
.\dataset/test: 0 files
.\dataset/test/No_DR: 271 files
.\dataset/test/DR: 279 files
Found 321 images belonging to 2 classes.
Found 550 images belonging to 2 classes.
Found 550 images belonging to 2 classes.
Epoch 1/10
11/11 ━━━━━━━━━━━━━━━━━━━━ 19s 1s/step - accuracy: 0.6672 - loss: 0.6967 - val_accuracy: 0.5073 - val_loss: 0.6918
Epoch 2/10
11/11 ━━━━━━━━━━━━━━━━━━━━ 20s 2s/step - accuracy: 0.8493 - loss: 0.3656 - val_accuracy: 0.5073 - val_loss: 0.6956
Epoch 3/10
11/11 ━━━━━━━━━━━━━━━━━━━━ 16s 1s/step - accuracy: 0.8947 - loss: 0.3088 - val_accuracy: 0.5073 - val_loss: 0.7002
Epoch 4/10
11/11 ━━━━━━━━━━━━━━━━━━━━ 16s 1s/step - accuracy: 0.8497 - loss: 0.3160 - val_accuracy: 0.5073 - val_loss: 0.7053
Epoch 5/10
11/11 ━━━━━━━━━━━━━━━━━━━━ 21s 1s/step - accuracy: 0.8942 - loss: 0.2618 - val_accuracy: 0.5073 - val_loss: 0.7103
Epoch 6/10
11/11 ━━━━━━━━━━━━━━━━━━━━ 16s 1s/step - accuracy: 0.9385 - loss: 0.2016 - val_accuracy: 0.5073 - val_loss: 0.7169
Epoch 7/10
11/11 ━━━━━━━━━━━━━━━━━━━━ 21s 2s/step - accuracy: 0.9371 - loss: 0.1914 - val_accuracy: 0.5073 - val_loss: 0.7235
Epoch 8/10
11/11 ━━━━━━━━━━━━━━━━━━━━ 17s 1s/step - accuracy: 0.9385 - loss: 0.1765 - val_accuracy: 0.5073 - val_loss: 0.7315
Epoch 9/10
11/11 ━━━━━━━━━━━━━━━━━━━━ 16s 1s/step - accuracy: 0.9635 - loss: 0.1654 - val_accuracy: 0.5073 - val_loss: 0.7362
Epoch 10/10
11/11 ━━━━━━━━━━━━━━━━━━━━ 16s 1s/step - accuracy: 0.9161 - loss: 0.1862 - val_accuracy: 0.5073 - val_loss: 0.7401
16.0
161

Contents of .\dataset/train:
.\dataset/train: 0 files
.\dataset/train/No_DR: 89 files
.\dataset/train/DR: 72 files

Contents of .\dataset/val:
.\dataset/val: 0 files
.\dataset/val/No_DR: 271 files
.\dataset/val/DR: 279 files

Contents of .\dataset/test:
.\dataset/test: 0 files
.\dataset/test/No_DR: 271 files
.\dataset/test/DR: 279 files
Found 161 images belonging to 2 classes.
Found 550 images belonging to 2 classes.
Found 550 images belonging to 2 classes.
Epoch 1/10
6/6 ━━━━━━━━━━━━━━━━━━━━ 15s 2s/step - accuracy: 0.6488 - loss: 0.7023 - val_accuracy: 0.5691 - val_loss: 0.6915
Epoch 2/10
6/6 ━━━━━━━━━━━━━━━━━━━━ 20s 2s/step - accuracy: 0.7992 - loss: 0.4822 - val_accuracy: 0.5364 - val_loss: 0.6905
Epoch 3/10
6/6 ━━━━━━━━━━━━━━━━━━━━ 20s 2s/step - accuracy: 0.8545 - loss: 0.3240 - val_accuracy: 0.5073 - val_loss: 0.6910
Epoch 4/10
6/6 ━━━━━━━━━━━━━━━━━━━━ 12s 2s/step - accuracy: 0.8747 - loss: 0.3692 - val_accuracy: 0.5073 - val_loss: 0.6940
Epoch 5/10
6/6 ━━━━━━━━━━━━━━━━━━━━ 20s 2s/step - accuracy: 0.9120 - loss: 0.2714 - val_accuracy: 0.5073 - val_loss: 0.6983
Epoch 6/10
6/6 ━━━━━━━━━━━━━━━━━━━━ 21s 2s/step - accuracy: 0.9072 - loss: 0.2942 - val_accuracy: 0.5073 - val_loss: 0.7036
Epoch 7/10
6/6 ━━━━━━━━━━━━━━━━━━━━ 12s 2s/step - accuracy: 0.8756 - loss: 0.2688 - val_accuracy: 0.5073 - val_loss: 0.7078
Epoch 8/10
6/6 ━━━━━━━━━━━━━━━━━━━━ 12s 2s/step - accuracy: 0.9226 - loss: 0.2576 - val_accuracy: 0.5073 - val_loss: 0.7089
Epoch 9/10
6/6 ━━━━━━━━━━━━━━━━━━━━ 20s 2s/step - accuracy: 0.9151 - loss: 0.2522 - val_accuracy: 0.5073 - val_loss: 0.7091
Epoch 10/10
6/6 ━━━━━━━━━━━━━━━━━━━━ 12s 2s/step - accuracy: 0.9447 - loss: 0.2030 - val_accuracy: 0.5073 - val_loss: 0.7103

import matplotlib.pyplot as plt

# Data
database_size = [2562, 1281, 641, 321, 161]
cnn_train_acc = [94.52, 95.10, 93.19, 91.61, 94.47]
cnn_val_acc = [93.27, 77.45, 56.18, 50.73, 50.73]
vit_train_acc = [91.67, 88.21, 70.20, 54.34, 62.67]
vit_val_acc = [92.73, 94, 69.27, 63.09, 63.09]

# Plot
figure, axis = plt.subplots(1, 2, figsize=(10, 5))
# Define the y-ticks with 5 increments
y_ticks = np.arange(50, 101, 5)

# CNN plot
axis[0].plot(database_size, cnn_train_acc, label='CNN Train Accuracy')
axis[0].plot(database_size, cnn_val_acc, label='CNN Validation Accuracy')
axis[0].set_xlabel('Database Size')
axis[0].set_ylabel('Accuracy')
axis[0].set_title('CNN Accuracy vs Database Size')
#axis[0].set_xticks(database_size)  # Set the x-ticks explicitly
axis[0].set_yticks(y_ticks)  # Set consistent y-ticks
axis[0].invert_xaxis()  # Reverse the x-axis
axis[0].legend()

# ViT plot
axis[1].plot(database_size, vit_train_acc, label='ViT Train Accuracy')
axis[1].plot(database_size, vit_val_acc, label='ViT Validation Accuracy')
axis[1].set_xlabel('Database Size')
axis[1].set_ylabel('Accuracy')
axis[1].set_title('ViT Accuracy vs Database Size')
#axis[1].set_xticks(database_size)  # Set the x-ticks explicitly
axis[1].set_yticks(y_ticks)  # Set consistent y-ticks
axis[1].invert_xaxis()  # Reverse the x-axis
axis[1].legend()
plt.savefig('cnn vs vit.png')
plt.tight_layout()
plt.show()

from google.colab import files
files.download('cnn vs vit.png')

<IPython.core.display.Javascript object>

<IPython.core.display.Javascript object>