[Python] Does the fact that i have just train, test split using keras image_from_dataset...

I've been trying to figure out why my metrics were low after training. I had an F1 score of 75% which is not what i expected at all. I reviewed the code and I suspected the issue was the train, validation split but also the fact that i was evaluating the model on the validation set y_pred = (model.predict(val_ds) > 0.5).astype(int). I read the keras documentation to see if i made mistakes or if there was a way to split my dataset into train, validation, test split but i didn't find anything that could point out if i made a mistake. Here's my code:

import tensorflow as tf
from tensorflow.keras.applications import DenseNet169
from tensorflow.keras.layers import Dense, GlobalAveragePooling2D, RandomFlip, RandomRotation, Dropout
from tensorflow.keras.models import Model, Sequential
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.regularizers import l2
from tensorflow.keras.applications.densenet import preprocess_input
from sklearn.metrics import confusion_matrix, f1_score
import numpy as np

# Define constants
IMG_SIZE = (224, 224) # Match DenseNet169 input size
BATCH_SIZE = 32
EPOCHS = 15
DATA_DIR = "colon_images"
CLASS_NAMES = ["cancerous", "normal"]
SPLIT_RATIO = 0.2

# Load datasets directly at 224x224
def create_dataset(subset):
return tf.keras.utils.image_dataset_from_directory(
DATA_DIR,
validation_split=SPLIT_RATIO,
subset=subset,
seed=42,
image_size=IMG_SIZE, # Directly load at target size
batch_size=BATCH_SIZE,
label_mode='binary'
)

train_ds = create_dataset("training")
val_ds = create_dataset("validation")

# Preprocessing with augmentation (augmentations only active during training)
preprocessing = Sequential([
RandomFlip("horizontal"),
RandomRotation(0.1),
tf.keras.layers.Lambda(preprocess_input) # Correct normalization
])

# Build model
base_model = DenseNet169(
weights='imagenet',
include_top=False,
input_shape=IMG_SIZE + (3,)
)

inputs = tf.keras.Input(shape=IMG_SIZE + (3,))
x = preprocessing(inputs)
x = base_model(x)
x = GlobalAveragePooling2D()(x)
x = Dense(512, activation='relu', kernel_regularizer=l2(0.01))(x)
x = Dropout(0.5)(x) # Regularization
outputs = Dense(1, activation='sigmoid')(x)
model = Model(inputs, outputs)

# Phase 1: Train top layers
base_model.trainable = False
model.compile(
optimizer=Adam(1e-3),
loss='binary_crossentropy',
metrics=['accuracy', tf.keras.metrics.AUC(name='auc'),
tf.keras.metrics.Precision(name='precision'),
tf.keras.metrics.Recall(name='recall')]
)

# Train with callbacks monitoring AUC
early_stop = tf.keras.callbacks.EarlyStopping(
monitor='val_auc', patience=3, mode='max', verbose=1
)
checkpoint = tf.keras.callbacks.ModelCheckpoint(
'best_model.h5', save_best_only=True, monitor='val_auc', mode='max'
)

history = model.fit(
train_ds,
validation_data=val_ds,
epochs=EPOCHS,
callbacks=[early_stop, checkpoint]
)

# Phase 2: Fine-tune entire model
base_model.trainable = True
model.compile(
optimizer=Adam(1e-5), # Very low learning rate
loss='binary_crossentropy',
metrics=['accuracy', tf.keras.metrics.AUC(name='auc')]
)

history_fine = model.fit(
train_ds,
validation_data=val_ds,
epochs=EPOCHS,
initial_epoch=history.epoch[-1],
callbacks=[early_stop, checkpoint]
)

# Evaluation
y_pred = (model.predict(val_ds) > 0.5).astype(int)
y_true = np.concatenate([y for _, y in val_ds], axis=0)

print(f"F1 Score: {f1_score(y_true, y_pred):.3f}")


`

Continue reading...