I copied some sample code straight from Keras official website and edited it to make a machine learning model. I am using Google Colab for my code.
Link: https://keras.io/examples/vision/image_classification_from_scratch/
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
from google.colab import drive
drive.mount("/content/gdrive")
image_size = (50, 50)
batch_size = 400
import random
num = random.randint(1, 400)
#random seed
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
"/content/gdrive/My Drive/pest/train",
seed=num,
image_size=image_size,
batch_size=batch_size,
)
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
"/content/gdrive/My Drive/pest/test",
seed=num,
image_size=image_size,
batch_size=batch_size,
)
#tried data augmentation
data_augmentation = keras.Sequential(
[
layers.RandomFlip("horizontal"),
layers.RandomRotation(0.1),
]
)
def make_model(input_shape, num_classes):
inputs = keras.Input(shape=input_shape)
# Image augmentation block
x = data_augmentation(inputs)
# Entry block
x = layers.Rescaling(1.0 / 255)(x)
x = layers.Conv2D(32, 3, strides=2, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
x = layers.Conv2D(64, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
previous_block_activation = x # Set aside residual
for size in [128, 256, 512, 728]:
x = layers.Activation("relu")(x)
x = layers.SeparableConv2D(size, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
x = layers.SeparableConv2D(size, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.MaxPooling2D(3, strides=2, padding="same")(x)
# Project residual
residual = layers.Conv2D(size, 1, strides=2, padding="same")(
previous_block_activation
)
x = layers.add([x, residual]) # Add back residual
previous_block_activation = x # Set aside next residual
x = layers.SeparableConv2D(1024, 3, padding="same")(x)
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
x = layers.GlobalAveragePooling2D()(x)
if num_classes == 2:
activation = "sigmoid"
units = 1
else:
activation = "softmax"
units = num_classes
x = layers.Dropout(0.5)(x)
outputs = layers.Dense(units, activation=activation)(x)
return keras.Model(inputs, outputs)
model = make_model(input_shape=image_size (3,), num_classes=2)
keras.utils.plot_model(model, show_shapes=True)
epochs = 50
callbacks = [
keras.callbacks.ModelCheckpoint("save_at_{epoch}.h5"),
]
model.compile(
optimizer=keras.optimizers.Adam(1e-3),
loss="binary_crossentropy",
metrics=["accuracy"],
)
model.fit(
train_ds, epochs=epochs, callbacks=callbacks, validation_data=val_ds,
#should have automatic shuffling I think
)
However, when I run it, the result is
Epoch 1/50
2/2 [==============================] - 71s 14s/step - loss: 0.6260 - accuracy: 0.6050 - val_loss: 0.6931 - val_accuracy: 0.5000
Epoch 2/50
2/2 [==============================] - 2s 507ms/step - loss: 0.2689 - accuracy: 0.8867 - val_loss: 0.6932 - val_accuracy: 0.5000
Epoch 3/50
2/2 [==============================] - 2s 536ms/step - loss: 0.1241 - accuracy: 0.9483 - val_loss: 0.6932 - val_accuracy: 0.5000
Epoch 4/50
2/2 [==============================] - 2s 506ms/step - loss: 0.0697 - accuracy: 0.9750 - val_loss: 0.6934 - val_accuracy: 0.5000
Epoch 5/50
2/2 [==============================] - 2s 525ms/step - loss: 0.0479 - accuracy: 0.9867 - val_loss: 0.6936 - val_accuracy: 0.5000
Epoch 6/50
2/2 [==============================] - 2s 534ms/step - loss: 0.0359 - accuracy: 0.9867 - val_loss: 0.6940 - val_accuracy: 0.5000
Epoch 7/50
2/2 [==============================] - 2s 509ms/step - loss: 0.0145 - accuracy: 0.9983 - val_loss: 0.6946 - val_accuracy: 0.5000
Epoch 8/50
2/2 [==============================] - 2s 545ms/step - loss: 0.0124 - accuracy: 0.9967 - val_loss: 0.6954 - val_accuracy: 0.5000
Epoch 9/50
2/2 [==============================] - 2s 544ms/step - loss: 0.0092 - accuracy: 0.9967 - val_loss: 0.6964 - val_accuracy: 0.5000
Epoch 10/50
2/2 [==============================] - 2s 512ms/step - loss: 0.0060 - accuracy: 0.9967 - val_loss: 0.6980 - val_accuracy: 0.5000
Epoch 11/50
2/2 [==============================] - 2s 535ms/step - loss: 0.0036 - accuracy: 0.9983 - val_loss: 0.6998 - val_accuracy: 0.5000
Epoch 12/50
2/2 [==============================] - 2s 503ms/step - loss: 0.0085 - accuracy: 0.9983 - val_loss: 0.7020 - val_accuracy: 0.5000
Epoch 13/50
2/2 [==============================] - 2s 665ms/step - loss: 0.0040 - accuracy: 1.0000 - val_loss: 0.7046 - val_accuracy: 0.5000
Epoch 14/50
2/2 [==============================] - 2s 516ms/step - loss: 0.0017 - accuracy: 1.0000 - val_loss: 0.7078 - val_accuracy: 0.5000
Epoch 15/50
2/2 [==============================] - 2s 520ms/step - loss: 0.0023 - accuracy: 0.9983 - val_loss: 0.7115 - val_accuracy: 0.5000
Epoch 16/50
2/2 [==============================] - 2s 500ms/step - loss: 8.5606e-04 - accuracy: 1.0000 - val_loss: 0.7157 - val_accuracy: 0.5000
Epoch 17/50
2/2 [==============================] - 2s 524ms/step - loss: 0.0018 - accuracy: 1.0000 - val_loss: 0.7205 - val_accuracy: 0.5000
Epoch 18/50
2/2 [==============================] - 2s 499ms/step - loss: 9.0626e-04 - accuracy: 1.0000 - val_loss: 0.7258 - val_accuracy: 0.5000
Epoch 19/50
2/2 [==============================] - 2s 510ms/step - loss: 0.0014 - accuracy: 1.0000 - val_loss: 0.7313 - val_accuracy: 0.5000
Epoch 20/50
2/2 [==============================] - 2s 711ms/step - loss: 0.0013 - accuracy: 1.0000 - val_loss: 0.7371 - val_accuracy: 0.5000
Epoch 21/50
2/2 [==============================] - 2s 511ms/step - loss: 9.9904e-04 - accuracy: 1.0000 - val_loss: 0.7431 - val_accuracy: 0.5000
Epoch 22/50
2/2 [==============================] - 2s 540ms/step - loss: 0.0019 - accuracy: 1.0000 - val_loss: 0.7489 - val_accuracy: 0.5000
Epoch 23/50
2/2 [==============================] - 2s 513ms/step - loss: 4.9861e-04 - accuracy: 1.0000 - val_loss: 0.7553 - val_accuracy: 0.5000
Epoch 24/50
2/2 [==============================] - 2s 542ms/step - loss: 6.6248e-04 - accuracy: 1.0000 - val_loss: 0.7622 - val_accuracy: 0.5000
Epoch 25/50
2/2 [==============================] - 2s 510ms/step - loss: 7.7911e-04 - accuracy: 1.0000 - val_loss: 0.7699 - val_accuracy: 0.5000
Epoch 26/50
2/2 [==============================] - 2s 502ms/step - loss: 3.3703e-04 - accuracy: 1.0000 - val_loss: 0.7781 - val_accuracy: 0.5000
Epoch 27/50
2/2 [==============================] - 2s 539ms/step - loss: 3.7860e-04 - accuracy: 1.0000 - val_loss: 0.7870 - val_accuracy: 0.5000
Epoch 28/50
2/2 [==============================] - 2s 507ms/step - loss: 2.4852e-04 - accuracy: 1.0000 - val_loss: 0.7962 - val_accuracy: 0.5000
Epoch 29/50
2/2 [==============================] - 2s 512ms/step - loss: 1.7709e-04 - accuracy: 1.0000 - val_loss: 0.8058 - val_accuracy: 0.5000
Epoch 30/50
2/2 [==============================] - 2s 538ms/step - loss: 1.6884e-04 - accuracy: 1.0000 - val_loss: 0.8161 - val_accuracy: 0.5000
Epoch 31/50
2/2 [==============================] - 2s 521ms/step - loss: 2.0884e-04 - accuracy: 1.0000 - val_loss: 0.8266 - val_accuracy: 0.5000
Epoch 32/50
2/2 [==============================] - 2s 543ms/step - loss: 1.8691e-04 - accuracy: 1.0000 - val_loss: 0.8375 - val_accuracy: 0.5000
Epoch 33/50
2/2 [==============================] - 2s 520ms/step - loss: 1.7296e-04 - accuracy: 1.0000 - val_loss: 0.8487 - val_accuracy: 0.5000
Epoch 34/50
2/2 [==============================] - 2s 516ms/step - loss: 4.5739e-04 - accuracy: 1.0000 - val_loss: 0.8601 - val_accuracy: 0.5000
Epoch 35/50
2/2 [==============================] - 2s 530ms/step - loss: 9.6831e-05 - accuracy: 1.0000 - val_loss: 0.8720 - val_accuracy: 0.5000
Epoch 36/50
2/2 [==============================] - 2s 553ms/step - loss: 1.2694e-04 - accuracy: 1.0000 - val_loss: 0.8847 - val_accuracy: 0.5000
Epoch 37/50
2/2 [==============================] - 2s 514ms/step - loss: 8.6252e-05 - accuracy: 1.0000 - val_loss: 0.8977 - val_accuracy: 0.5000
Epoch 38/50
2/2 [==============================] - 2s 520ms/step - loss: 2.6762e-04 - accuracy: 1.0000 - val_loss: 0.9115 - val_accuracy: 0.5000
Epoch 39/50
2/2 [==============================] - 2s 542ms/step - loss: 8.1350e-05 - accuracy: 1.0000 - val_loss: 0.9258 - val_accuracy: 0.5000
Epoch 40/50
2/2 [==============================] - 2s 506ms/step - loss: 8.0961e-05 - accuracy: 1.0000 - val_loss: 0.9405 - val_accuracy: 0.5000
Epoch 41/50
2/2 [==============================] - 2s 526ms/step - loss: 6.6102e-05 - accuracy: 1.0000 - val_loss: 0.9555 - val_accuracy: 0.5000
Epoch 42/50
2/2 [==============================] - 2s 549ms/step - loss: 1.1529e-04 - accuracy: 1.0000 - val_loss: 0.9707 - val_accuracy: 0.5000
Epoch 43/50
2/2 [==============================] - 2s 528ms/step - loss: 6.1373e-05 - accuracy: 1.0000 - val_loss: 0.9864 - val_accuracy: 0.5000
Epoch 44/50
2/2 [==============================] - 2s 516ms/step - loss: 7.2809e-05 - accuracy: 1.0000 - val_loss: 1.0025 - val_accuracy: 0.5000
Epoch 45/50
2/2 [==============================] - 2s 513ms/step - loss: 5.9504e-05 - accuracy: 1.0000 - val_loss: 1.0191 - val_accuracy: 0.5000
Epoch 46/50
2/2 [==============================] - 2s 515ms/step - loss: 6.1622e-05 - accuracy: 1.0000 - val_loss: 1.0361 - val_accuracy: 0.5000
Epoch 47/50
2/2 [==============================] - 2s 525ms/step - loss: 7.7296e-05 - accuracy: 1.0000 - val_loss: 1.0534 - val_accuracy: 0.5000
Epoch 48/50
2/2 [==============================] - 2s 512ms/step - loss: 4.5088e-05 - accuracy: 1.0000 - val_loss: 1.0711 - val_accuracy: 0.5000
Epoch 49/50
2/2 [==============================] - 2s 532ms/step - loss: 1.1449e-04 - accuracy: 1.0000 - val_loss: 1.0887 - val_accuracy: 0.5000
Epoch 50/50
2/2 [==============================] - 2s 516ms/step - loss: 6.0932e-05 - accuracy: 1.0000 - val_loss: 1.1071 - val_accuracy: 0.5000
<keras.callbacks.History at 0x7fb4205a20d0>
Since I have 2 classes, my teacher said that a validation accuracy of 0.5 means that it is completely random.
My images are in the format of 50x50 .jpg images in Google Drive. Could that be the problem as my current image size is 50x50? But when I run
import matplotlib.pyplot as plt
plt.figure(figsize=(10, 10))
for images, labels in train_ds.take(1):
for i in range(9):
ax = plt.subplot(3, 3, i 1)
plt.imshow(images[i].numpy().astype("uint8"))
plt.title(int(labels[i]))
plt.axis("off")
The images are correct, as in the entire image is shown and is clear. I tried changing the seed to a random number. The code comes with data augmentation and the model.fit() should automatically shuffle the images (if I understood the online sites correctly).
My teacher does not know what is wrong either. Any solutions?
Edit: this is the dataset https://www.kaggle.com/datasets/simranvolunesia/pest-dataset
Edit2: Sorry for the confusion but I only used two datasets, aphids and bollworm.
CodePudding user response：
the answer provided by Djinn is correct. Also your are including augmentation WITHIN your model. So your model will augment not only the training images but the validation and test images as well. Test and Validation images should not be augmented. If you want augmentation then use the ImageDataGenerator.flow_from_directory. Documentation for that is here
CodePudding user response：
Edit: You are also using binary_crossentropy for a multi-class classification problem, yet you're forcing it to only have two classes when your passed dataset contains nine.
model = make_model(input_shape=image_size (3,), num_classes=2)
According to your dataset, the classes are:
Pests: aphids, armyworm, beetle, bollworm, grasshopper, mites, mosquito, sawfly, stem borer
I don't see where you're only working with two classes, unless there's some code missing somewhere that removes the other seven. This site (https://keras.io/examples/vision/image_classification_from_scratch/) is classifying into two classes: cat or dog. That's probably where you got two classes from.
So that line needs to be changed to:
model = make_model(input_shape=image_size (3,), num_classes=9)
Change this:
model.compile(
optimizer=keras.optimizers.Adam(1e-3),
loss="binary_crossentropy",
metrics=["accuracy"],
)
To:
model.compile(
optimizer=keras.optimizers.Adam(1e-3),
loss="sparse_categorical_crossentropy",
metrics=["accuracy"],
)
You might also need to change that metric from accuracy to binary_accuracy. Try with just accuracy first, then with binary_accuracy.
model.compile(
optimizer=keras.optimizers.Adam(1e-3),
loss="sparse_categorical_crossentropy",
metrics=["binary_accuracy"],
According to the documentation, you are not splitting your validation data correctly and probably dealing with the default shuffling too.
Define your datasets like this (assuming a 20% validation split):
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
"/content/gdrive/My Drive/pest/train",
validation_split=0.2,
subset="training"
seed=num,
image_size=image_size,
batch_size=batch_size,
)
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
"/content/gdrive/My Drive/pest/train",
validation_split=0.2,
subset="validation"
seed=num,
image_size=image_size,
batch_size=batch_size,
)
# with test folder for test set
test_ds = tf.keras.preprocessing.image_dataset_from_directory(
"/content/gdrive/My Drive/pest/test",
image_size=image_size,
batch_size=batch_size,
shuffle=False
)