From 19f870527919157f47e18f894c095b566cc16246 Mon Sep 17 00:00:00 2001
From: "michael.divia" <michael.divia@etu.hesge.ch>
Date: Tue, 1 Apr 2025 18:01:46 +0200
Subject: [PATCH] 2 training models
---
pokedex.py | 113 -----------------------------------
pokedex_EfficientNetV2M.py | 118 +++++++++++++++++++++++++++++++++++++
pokedex_ResNet50.py | 118 +++++++++++++++++++++++++++++++++++++
test.py | 66 ---------------------
4 files changed, 236 insertions(+), 179 deletions(-)
delete mode 100644 pokedex.py
create mode 100644 pokedex_EfficientNetV2M.py
create mode 100644 pokedex_ResNet50.py
delete mode 100644 test.py
diff --git a/pokedex.py b/pokedex.py
deleted file mode 100644
index d60dcaf..0000000
--- a/pokedex.py
+++ /dev/null
@@ -1,113 +0,0 @@
-import os
-import numpy as np
-import keras
-from keras import layers
-import matplotlib.pyplot as plt
-from tensorflow import data as tf_data
-import random
-
-#os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
-
-# Load training and validation dataset
-train_ds, val_ds = keras.utils.image_dataset_from_directory(
- "Combined_Dataset",
- labels="inferred",
- label_mode="int",
- image_size=(256, 256),
- batch_size=20,
- shuffle=True,
- validation_split=0.2,
- subset="both",
- seed=random.randint(0,8000)
-)
-
-# Get class (Pokémon) names
-class_names = train_ds.class_names
-
-# Introduce artificial sample diversity
-data_augmentation_layers = [
- layers.RandomFlip("horizontal"),
- layers.RandomRotation(0.1),
-]
-
-def data_augmentation(images):
- for layer in data_augmentation_layers:
- images = layer(images)
- return images
-
-# Apply `data_augmentation` to the training images.
-train_ds = train_ds.map(
- lambda img, label: (data_augmentation(img), label),
- num_parallel_calls=tf_data.AUTOTUNE,
-)
-
-# Prefetching samples in GPU memory helps maximize GPU utilization.
-train_ds = train_ds.prefetch(tf_data.AUTOTUNE)
-val_ds = val_ds.prefetch(tf_data.AUTOTUNE)
-
-
-# MODEL
-def simple_xception_netowkr(input_shape, num_classes):
- inputs = keras.Input(shape=input_shape)
-
- # Entry block
- x = layers.Rescaling(1.0 / 255)(inputs)
- x = layers.Conv2D(128, 3, strides=2, padding="same")(x)
- x = layers.BatchNormalization()(x)
- x = layers.Activation("relu")(x)
-
- previous_block_activation = x # Set aside residual
-
- for size in [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:
- units = 1
- else:
- units = num_classes
-
- x = layers.Dropout(0.25)(x)
- # We specify activation=None so as to return logits
- outputs = layers.Dense(units, activation=None)(x)
- return keras.Model(inputs, outputs)
-
-
-model = simple_xception_netowkr(input_shape=(256, 256) + (3,), num_classes=152)
-
-# Train
-epochs = 25
-
-callbacks = [
- keras.callbacks.ModelCheckpoint("save_at_{epoch}.keras"),
-]
-model.compile(
- optimizer=keras.optimizers.Adam(3e-4),
- loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
- metrics=[keras.metrics.SparseCategoricalAccuracy(name="acc")],
-)
-model.fit(
- train_ds,
- epochs=epochs,
- callbacks=callbacks,
- validation_data=val_ds,
-)
\ No newline at end of file
diff --git a/pokedex_EfficientNetV2M.py b/pokedex_EfficientNetV2M.py
new file mode 100644
index 0000000..b9fd017
--- /dev/null
+++ b/pokedex_EfficientNetV2M.py
@@ -0,0 +1,118 @@
+import os
+import gc
+import keras
+import tensorflow as tf
+from keras import layers
+from tensorflow import data as tf_data
+
+# --- GPU Strategy ---
+strategy = tf.distribute.MirroredStrategy()
+print("Number of GPUs:", strategy.num_replicas_in_sync)
+
+# --- Parameters ---
+data_dir = "/home/users/d/divia/scratch/Combined_Dataset"
+image_size = (240, 240)
+num_classes = 151
+base_batch_size = 32
+base_lr = 1e-3
+
+# --- Auto-Tune Batch Size ---
+def find_max_batch_size(data_dir, image_size, candidate_sizes=[256, 128, 64, 32, 16]):
+ print("Tuning batch size...")
+ for bs in candidate_sizes:
+ try:
+ print(f"Trying global batch size {bs}...")
+ ds = keras.utils.image_dataset_from_directory(
+ data_dir,
+ labels="inferred",
+ label_mode="int",
+ image_size=image_size,
+ batch_size=bs,
+ shuffle=True
+ )
+ for batch in ds.take(1): # Try to load one batch
+ tf.print("✓ Batch size", bs, "works")
+ del ds
+ gc.collect()
+ tf.keras.backend.clear_session()
+ return bs
+ except tf.errors.ResourceExhaustedError:
+ print(f"✗ Batch size {bs} too large.")
+ raise RuntimeError("No suitable batch size found.")
+
+global_batch_size = find_max_batch_size(data_dir, image_size)
+scaled_lr = base_lr * (global_batch_size / base_batch_size)
+
+# --- Load Dataset ---
+full_ds = keras.utils.image_dataset_from_directory(
+ data_dir,
+ labels="inferred",
+ label_mode="int",
+ image_size=image_size,
+ batch_size=global_batch_size,
+ shuffle=True,
+ seed=1234
+)
+
+# --- Train/Val Split ---
+total_batches = tf.data.experimental.cardinality(full_ds).numpy()
+train_size = int(0.8 * total_batches)
+train_ds = full_ds.take(train_size)
+val_ds = full_ds.skip(train_size)
+
+# --- Data Augmentation ---
+data_augmentation_layers = keras.Sequential([
+ layers.RandomFlip("horizontal"),
+ layers.RandomRotation(0.1),
+])
+
+def preprocess_train(img, label):
+ img = data_augmentation_layers(img)
+ label = tf.one_hot(label, num_classes)
+ return img, label
+
+def preprocess_val(img, label):
+ label = tf.one_hot(label, num_classes)
+ return img, label
+
+train_ds = train_ds.map(preprocess_train, num_parallel_calls=tf_data.AUTOTUNE)
+val_ds = val_ds.map(preprocess_val, num_parallel_calls=tf_data.AUTOTUNE)
+
+train_ds = train_ds.prefetch(buffer_size=tf_data.AUTOTUNE)
+val_ds = val_ds.prefetch(buffer_size=tf_data.AUTOTUNE)
+
+# --- Build & Compile Model ---
+with strategy.scope():
+ base_model = tf.keras.applications.EfficientNetV2M(
+ include_top=False,
+ weights='imagenet',
+ input_shape=(240, 240, 3)
+ )
+
+ model = keras.Sequential([
+ base_model,
+ layers.GlobalAveragePooling2D(),
+ layers.Dense(256, activation='relu'),
+ layers.Dropout(0.5),
+ layers.Dense(num_classes, activation='softmax')
+ ])
+
+ optimizer = tf.keras.optimizers.Adam(learning_rate=scaled_lr)
+
+ model.compile(
+ optimizer=optimizer,
+ loss='categorical_crossentropy',
+ metrics=['accuracy']
+ )
+
+# --- Train ---
+callbacks = [
+ keras.callbacks.ModelCheckpoint("EfficientNetV2M/save_at_{epoch}.keras")
+]
+
+model.fit(
+ train_ds,
+ validation_data=val_ds,
+ epochs=10,
+ callbacks=callbacks
+)
\ No newline at end of file
diff --git a/pokedex_ResNet50.py b/pokedex_ResNet50.py
new file mode 100644
index 0000000..9c16244
--- /dev/null
+++ b/pokedex_ResNet50.py
@@ -0,0 +1,118 @@
+import os
+import gc
+import keras
+import tensorflow as tf
+from keras import layers
+from tensorflow import data as tf_data
+
+# --- GPU Strategy ---
+strategy = tf.distribute.MirroredStrategy()
+print("Number of GPUs:", strategy.num_replicas_in_sync)
+
+# --- Parameters ---
+data_dir = "/home/users/d/divia/scratch/Combined_Dataset"
+image_size = (224, 224)
+num_classes = 151
+base_batch_size = 32
+base_lr = 1e-3
+
+# --- Auto-Tune Batch Size ---
+def find_max_batch_size(data_dir, image_size, candidate_sizes=[256, 128, 64, 32, 16]):
+ print("Tuning batch size...")
+ for bs in candidate_sizes:
+ try:
+ print(f"Trying global batch size {bs}...")
+ ds = keras.utils.image_dataset_from_directory(
+ data_dir,
+ labels="inferred",
+ label_mode="int",
+ image_size=image_size,
+ batch_size=bs,
+ shuffle=True
+ )
+ for batch in ds.take(1): # Try to load one batch
+ tf.print("✓ Batch size", bs, "works")
+ del ds
+ gc.collect()
+ tf.keras.backend.clear_session()
+ return bs
+ except tf.errors.ResourceExhaustedError:
+ print(f"✗ Batch size {bs} too large.")
+ raise RuntimeError("No suitable batch size found.")
+
+global_batch_size = find_max_batch_size(data_dir, image_size)
+scaled_lr = base_lr * (global_batch_size / base_batch_size)
+
+# --- Load Dataset ---
+full_ds = keras.utils.image_dataset_from_directory(
+ data_dir,
+ labels="inferred",
+ label_mode="int",
+ image_size=image_size,
+ batch_size=global_batch_size,
+ shuffle=True,
+ seed=1234
+)
+
+# --- Train/Val Split ---
+total_batches = tf.data.experimental.cardinality(full_ds).numpy()
+train_size = int(0.8 * total_batches)
+train_ds = full_ds.take(train_size)
+val_ds = full_ds.skip(train_size)
+
+# --- Data Augmentation ---
+data_augmentation_layers = keras.Sequential([
+ layers.RandomFlip("horizontal"),
+ layers.RandomRotation(0.1),
+])
+
+def preprocess_train(img, label):
+ img = data_augmentation_layers(img)
+ label = tf.one_hot(label, num_classes)
+ return img, label
+
+def preprocess_val(img, label):
+ label = tf.one_hot(label, num_classes)
+ return img, label
+
+train_ds = train_ds.map(preprocess_train, num_parallel_calls=tf_data.AUTOTUNE)
+val_ds = val_ds.map(preprocess_val, num_parallel_calls=tf_data.AUTOTUNE)
+
+train_ds = train_ds.prefetch(buffer_size=tf_data.AUTOTUNE)
+val_ds = val_ds.prefetch(buffer_size=tf_data.AUTOTUNE)
+
+# --- Build & Compile Model ---
+with strategy.scope():
+ base_model = tf.keras.applications.ResNet50(
+ include_top=False,
+ weights='imagenet',
+ input_shape=(224, 224, 3)
+ )
+
+ model = keras.Sequential([
+ base_model,
+ layers.GlobalAveragePooling2D(),
+ layers.Dense(256, activation='relu'),
+ layers.Dropout(0.5),
+ layers.Dense(num_classes, activation='softmax')
+ ])
+
+ optimizer = tf.keras.optimizers.Adam(learning_rate=scaled_lr)
+
+ model.compile(
+ optimizer=optimizer,
+ loss='categorical_crossentropy',
+ metrics=['accuracy']
+ )
+
+# --- Train ---
+callbacks = [
+ keras.callbacks.ModelCheckpoint("ResNet50/save_at_{epoch}.keras")
+]
+
+model.fit(
+ train_ds,
+ validation_data=val_ds,
+ epochs=10,
+ callbacks=callbacks
+)
\ No newline at end of file
diff --git a/test.py b/test.py
deleted file mode 100644
index 1579836..0000000
--- a/test.py
+++ /dev/null
@@ -1,66 +0,0 @@
-import keras
-import matplotlib.pyplot as plt
-import numpy as np
-import os
-import random
-
-#os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
-
-# Get class names from directory structure
-temp_ds = keras.utils.image_dataset_from_directory(
- "Combined_Dataset",
- labels="inferred",
- label_mode="int",
- image_size=(256, 256),
- batch_size=1,
- shuffle=False
-)
-class_names = temp_ds.class_names
-
-# Load model
-model = keras.models.load_model("simple_xception.keras")
-
-# Base path
-base_path = "Combined_Dataset"
-
-# Prepare 2x2 plot
-plt.figure(figsize=(10, 10))
-
-for i in range(4):
- # Pick random class and image
- random_class = random.choice(class_names)
- class_folder = os.path.join(base_path, random_class)
- random_image = random.choice([
- f for f in os.listdir(class_folder)
- if f.lower().endswith(('.png', '.jpg', '.jpeg'))
- ])
- img_path = os.path.join(class_folder, random_image)
-
- # Load and preprocess
- img = keras.utils.load_img(img_path, target_size=(256, 256))
- img_array = keras.utils.img_to_array(img)
- img_array = keras.ops.expand_dims(img_array, 0)
-
- # Predict
- predictions = model.predict(img_array, verbose=0)
- probabilities = keras.ops.softmax(predictions[0])
- predicted_class_index = np.argmax(probabilities)
- predicted_label = class_names[predicted_class_index]
- confidence = 100 * probabilities[predicted_class_index]
-
- # Compare with actual
- is_correct = predicted_label == random_class
-
- # Plot
- ax = plt.subplot(2, 2, i + 1)
- plt.imshow(img)
- plt.axis("off")
- plt.title(
- f"Pred: {predicted_label}\n"
- f"True: {random_class}\n"
- f"{'Yes' if is_correct else 'No'} | {confidence:.1f}%",
- fontsize=10
- )
-
-plt.tight_layout()
-plt.show()
--
GitLab