Symulowanie-wizualne/sw_lab7.ipynb

Zadanie 7 - Alexnet

Aleksandra Jonas, Aleksandra Gronowska, Iwona Christop

Przygotowanie danych

from IPython.display import Image, SVG, display

import sys
import subprocess
import pkg_resources
import numpy as np

required = { 'scikit-image'}
installed = {pkg.key for pkg in pkg_resources.working_set}
missing = required - installed
# Alexnet requires images to be of dim = (227, 227, 3)
newSize = (227,227)

if missing: 
    python = sys.executable
    subprocess.check_call([python, '-m', 'pip', 'install', *missing], stdout=subprocess.DEVNULL)

def load_train_data(input_dir):
    import numpy as np
    import pandas as pd
    import os
    from skimage.io import imread
    import cv2 as cv
    from pathlib import Path
    import random
    from shutil import copyfile, rmtree
    import json

    import seaborn as sns
    import matplotlib.pyplot as plt

    import matplotlib
    
    image_dir = Path(input_dir)
    categories_name = []
    for file in os.listdir(image_dir):
        d = os.path.join(image_dir, file)
        if os.path.isdir(d):
            categories_name.append(file)

    folders = [directory for directory in image_dir.iterdir() if directory.is_dir()]

    train_img = []
    categories_count=[]
    labels=[]
    for i, direc in enumerate(folders):
        count = 0
        for obj in direc.iterdir():
            if os.path.isfile(obj) and os.path.basename(os.path.normpath(obj)) != 'desktop.ini':
                labels.append(os.path.basename(os.path.normpath(direc)))
                count += 1
                img = imread(obj)#zwraca ndarry postaci xSize x ySize x colorDepth
                img = img[:, :, :3]
                img = cv.resize(img, newSize, interpolation=cv.INTER_AREA)# zwraca ndarray
                img = img / 255 #normalizacja
                train_img.append(img)
        categories_count.append(count)
    X={}
    X["values"] = np.array(train_img)
    X["categories_name"] = categories_name
    X["categories_count"] = categories_count
    X["labels"]=labels
    return X

def load_test_data(input_dir):
    import numpy as np
    import pandas as pd
    import os
    from skimage.io import imread
    import cv2 as cv
    from pathlib import Path
    import random
    from shutil import copyfile, rmtree
    import json

    import seaborn as sns
    import matplotlib.pyplot as plt

    import matplotlib

    image_path = Path(input_dir)

    labels_path = image_path.parents[0] / 'test_labels.json'

    jsonString = labels_path.read_text()
    objects = json.loads(jsonString)

    categories_name = []
    categories_count=[]
    count = 0
    c = objects[0]['value']
    for e in  objects:
        if e['value'] != c:
            categories_count.append(count)
            c = e['value']
            count = 1
        else:
            count += 1
        if not e['value'] in categories_name:
            categories_name.append(e['value'])

    categories_count.append(count)
    
    test_img = []

    labels=[]
    for e in objects:
        p = image_path / e['filename']
        img = imread(p)#zwraca ndarry postaci xSize x ySize x colorDepth
        img = img[:, :, :3]
        img = cv.resize(img, newSize, interpolation=cv.INTER_AREA)# zwraca ndarray
        img = img / 255#normalizacja
        test_img.append(img)
        labels.append(e['value'])

    X={}
    X["values"] = np.array(test_img)
    X["categories_name"] = categories_name
    X["categories_count"] = categories_count
    X["labels"]=labels
    return X

# Data load
data_train = load_train_data("./train_test_sw/train_sw")
values_train = data_train['values']
labels_train = data_train['labels']

data_test = load_test_data("./train_test_sw/test_sw")
X_test = data_test['values']
y_test = data_test['labels']

from sklearn.model_selection import train_test_split

X_train, X_validate, y_train, y_validate = train_test_split(values_train, labels_train, test_size=0.2, random_state=42)

from sklearn.preprocessing import LabelEncoder

class_le = LabelEncoder()
y_train_enc = class_le.fit_transform(y_train)
y_validate_enc = class_le.fit_transform(y_validate)
y_test_enc = class_le.fit_transform(y_test)

import tensorflow as tf

train_ds = tf.data.Dataset.from_tensor_slices((X_train, y_train_enc))
validation_ds = tf.data.Dataset.from_tensor_slices((X_validate, y_validate_enc))
test_ds = tf.data.Dataset.from_tensor_slices((X_test, y_test_enc))

train_ds_size = tf.data.experimental.cardinality(train_ds).numpy()
test_ds_size = tf.data.experimental.cardinality(test_ds).numpy()
validation_ds_size = tf.data.experimental.cardinality(validation_ds).numpy()
print("Training data size:", train_ds_size)
print("Test data size:", test_ds_size)
print("Validation data size:", validation_ds_size)

Training data size: 820
Test data size: 259
Validation data size: 206

train_ds = (train_ds
                  .shuffle(buffer_size=train_ds_size)
                  .batch(batch_size=32, drop_remainder=True))
test_ds = (test_ds
                  .shuffle(buffer_size=train_ds_size)
                  .batch(batch_size=32, drop_remainder=True))
validation_ds = (validation_ds
                  .shuffle(buffer_size=train_ds_size)
                  .batch(batch_size=32, drop_remainder=True))

Model 1 - batch size = 32

from tensorflow import keras
import os
import time

model = keras.models.Sequential([
    keras.layers.Conv2D(filters=96, kernel_size=(11,11), strides=(4,4), activation='relu', input_shape=(227,227,3)),
    keras.layers.MaxPool2D(pool_size=(3,3), strides=(2,2)),
    keras.layers.Conv2D(filters=256, kernel_size=(5,5), strides=(1,1), activation='relu', padding="same"),
    keras.layers.MaxPool2D(pool_size=(3,3), strides=(2,2)),
    keras.layers.Conv2D(filters=384, kernel_size=(3,3), strides=(1,1), activation='relu', padding="same"),
    keras.layers.Conv2D(filters=384, kernel_size=(3,3), strides=(1,1), activation='relu', padding="same"),
    keras.layers.Conv2D(filters=256, kernel_size=(3,3), strides=(1,1), activation='relu', padding="same"),
    keras.layers.MaxPool2D(pool_size=(3,3), strides=(2,2)),
    keras.layers.Flatten(),
    keras.layers.Dense(4096, activation='relu'),
    keras.layers.Dense(4096, activation='relu'),
    keras.layers.Dense(10, activation='softmax')
])

root_logdir = os.path.join(os.curdir, "logs\\\\fit\\\\")
def get_run_logdir():
    run_id = time.strftime("run_%Y_%m_%d-%H_%M_%S")
    return os.path.join(root_logdir, run_id)
run_logdir = get_run_logdir()
tensorboard_cb = keras.callbacks.TensorBoard(run_logdir)

model.compile(loss='sparse_categorical_crossentropy', optimizer=tf.optimizers.SGD(lr=.001), metrics=['accuracy'])
model.summary()

WARNING:absl:`lr` is deprecated, please use `learning_rate` instead, or use the legacy optimizer, e.g.,tf.keras.optimizers.legacy.SGD.

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d (Conv2D)             (None, 55, 55, 96)        34944     
                                                                 
 max_pooling2d (MaxPooling2D  (None, 27, 27, 96)       0         
 )                                                               
                                                                 
 conv2d_1 (Conv2D)           (None, 27, 27, 256)       614656    
                                                                 
 max_pooling2d_1 (MaxPooling  (None, 13, 13, 256)      0         
 2D)                                                             
                                                                 
 conv2d_2 (Conv2D)           (None, 13, 13, 384)       885120    
                                                                 
 conv2d_3 (Conv2D)           (None, 13, 13, 384)       1327488   
                                                                 
 conv2d_4 (Conv2D)           (None, 13, 13, 256)       884992    
                                                                 
 max_pooling2d_2 (MaxPooling  (None, 6, 6, 256)        0         
 2D)                                                             
                                                                 
 flatten (Flatten)           (None, 9216)              0         
                                                                 
 dense (Dense)               (None, 4096)              37752832  
                                                                 
 dense_1 (Dense)             (None, 4096)              16781312  
                                                                 
 dense_2 (Dense)             (None, 10)                40970     
                                                                 
=================================================================
Total params: 58,322,314
Trainable params: 58,322,314
Non-trainable params: 0
_________________________________________________________________

model.fit(train_ds,
          epochs=100,
          validation_data=validation_ds,
          validation_freq=1,
          callbacks=[tensorboard_cb])

Epoch 1/100
25/25 [==============================] - 53s 2s/step - loss: 2.2516 - accuracy: 0.2150 - val_loss: 2.1732 - val_accuracy: 0.2656
Epoch 2/100
25/25 [==============================] - 40s 2s/step - loss: 1.9127 - accuracy: 0.2362 - val_loss: 1.6225 - val_accuracy: 0.2188
Epoch 3/100
25/25 [==============================] - 40s 2s/step - loss: 1.6217 - accuracy: 0.2612 - val_loss: 1.5785 - val_accuracy: 0.3438
Epoch 4/100
25/25 [==============================] - 41s 2s/step - loss: 1.5799 - accuracy: 0.3038 - val_loss: 1.5517 - val_accuracy: 0.2917
Epoch 5/100
25/25 [==============================] - 41s 2s/step - loss: 1.5565 - accuracy: 0.2988 - val_loss: 1.5266 - val_accuracy: 0.3750
Epoch 6/100
25/25 [==============================] - 41s 2s/step - loss: 1.4953 - accuracy: 0.3738 - val_loss: 1.6185 - val_accuracy: 0.3750
Epoch 7/100
25/25 [==============================] - 41s 2s/step - loss: 1.4873 - accuracy: 0.3613 - val_loss: 1.4834 - val_accuracy: 0.4219
Epoch 8/100
25/25 [==============================] - 41s 2s/step - loss: 1.4110 - accuracy: 0.4187 - val_loss: 1.3432 - val_accuracy: 0.4740
Epoch 9/100
25/25 [==============================] - 40s 2s/step - loss: 1.3515 - accuracy: 0.4363 - val_loss: 1.3292 - val_accuracy: 0.4688
Epoch 10/100
25/25 [==============================] - 40s 2s/step - loss: 1.3174 - accuracy: 0.4563 - val_loss: 1.3137 - val_accuracy: 0.4375
Epoch 11/100
25/25 [==============================] - 41s 2s/step - loss: 1.2342 - accuracy: 0.4950 - val_loss: 1.3598 - val_accuracy: 0.5260
Epoch 12/100
25/25 [==============================] - 40s 2s/step - loss: 1.2386 - accuracy: 0.4787 - val_loss: 1.1717 - val_accuracy: 0.5052
Epoch 13/100
25/25 [==============================] - 40s 2s/step - loss: 1.1891 - accuracy: 0.5025 - val_loss: 1.1926 - val_accuracy: 0.4896
Epoch 14/100
25/25 [==============================] - 41s 2s/step - loss: 1.1791 - accuracy: 0.4950 - val_loss: 1.1991 - val_accuracy: 0.4271
Epoch 15/100
25/25 [==============================] - 42s 2s/step - loss: 1.1458 - accuracy: 0.5275 - val_loss: 1.1069 - val_accuracy: 0.5104
Epoch 16/100
25/25 [==============================] - 40s 2s/step - loss: 1.0808 - accuracy: 0.5450 - val_loss: 1.0976 - val_accuracy: 0.5521
Epoch 17/100
25/25 [==============================] - 42s 2s/step - loss: 1.0491 - accuracy: 0.5650 - val_loss: 1.2781 - val_accuracy: 0.4740
Epoch 18/100
25/25 [==============================] - 42s 2s/step - loss: 1.0335 - accuracy: 0.5575 - val_loss: 1.1958 - val_accuracy: 0.5417
Epoch 19/100
25/25 [==============================] - 41s 2s/step - loss: 1.0198 - accuracy: 0.5675 - val_loss: 1.0505 - val_accuracy: 0.5729
Epoch 20/100
25/25 [==============================] - 40s 2s/step - loss: 0.9765 - accuracy: 0.5900 - val_loss: 1.0172 - val_accuracy: 0.5938
Epoch 21/100
25/25 [==============================] - 39s 2s/step - loss: 0.9793 - accuracy: 0.5913 - val_loss: 1.0369 - val_accuracy: 0.5365
Epoch 22/100
25/25 [==============================] - 42s 2s/step - loss: 0.9380 - accuracy: 0.6162 - val_loss: 1.0670 - val_accuracy: 0.5104
Epoch 23/100
25/25 [==============================] - 40s 2s/step - loss: 0.9409 - accuracy: 0.5813 - val_loss: 0.9529 - val_accuracy: 0.6094
Epoch 24/100
25/25 [==============================] - 44s 2s/step - loss: 0.8277 - accuracy: 0.6375 - val_loss: 1.2964 - val_accuracy: 0.4635
Epoch 25/100
25/25 [==============================] - 52s 2s/step - loss: 0.9220 - accuracy: 0.5938 - val_loss: 0.9610 - val_accuracy: 0.6042
Epoch 26/100
25/25 [==============================] - 39s 2s/step - loss: 0.8962 - accuracy: 0.6137 - val_loss: 1.0367 - val_accuracy: 0.5365
Epoch 27/100
25/25 [==============================] - 42s 2s/step - loss: 0.8650 - accuracy: 0.6237 - val_loss: 1.0654 - val_accuracy: 0.5156
Epoch 28/100
25/25 [==============================] - 45s 2s/step - loss: 0.8186 - accuracy: 0.6413 - val_loss: 0.9914 - val_accuracy: 0.6094
Epoch 29/100
25/25 [==============================] - 41s 2s/step - loss: 0.8347 - accuracy: 0.6313 - val_loss: 0.9955 - val_accuracy: 0.5990
Epoch 30/100
25/25 [==============================] - 42s 2s/step - loss: 0.7907 - accuracy: 0.6513 - val_loss: 0.9453 - val_accuracy: 0.6146
Epoch 31/100
25/25 [==============================] - 42s 2s/step - loss: 0.7743 - accuracy: 0.6675 - val_loss: 0.9493 - val_accuracy: 0.6042
Epoch 32/100
25/25 [==============================] - 41s 2s/step - loss: 0.7444 - accuracy: 0.6938 - val_loss: 0.9506 - val_accuracy: 0.6146
Epoch 33/100
25/25 [==============================] - 42s 2s/step - loss: 0.7630 - accuracy: 0.6525 - val_loss: 0.8973 - val_accuracy: 0.6354
Epoch 34/100
25/25 [==============================] - 46s 2s/step - loss: 0.7529 - accuracy: 0.6850 - val_loss: 0.9552 - val_accuracy: 0.5833
Epoch 35/100
25/25 [==============================] - 39s 2s/step - loss: 0.6825 - accuracy: 0.7063 - val_loss: 1.0233 - val_accuracy: 0.5729
Epoch 36/100
25/25 [==============================] - 39s 2s/step - loss: 0.6654 - accuracy: 0.7287 - val_loss: 0.9992 - val_accuracy: 0.6250
Epoch 37/100
25/25 [==============================] - 39s 2s/step - loss: 0.7306 - accuracy: 0.7075 - val_loss: 1.1470 - val_accuracy: 0.5833
Epoch 38/100
25/25 [==============================] - 40s 2s/step - loss: 0.6823 - accuracy: 0.7237 - val_loss: 0.9150 - val_accuracy: 0.6406
Epoch 39/100
25/25 [==============================] - 41s 2s/step - loss: 0.7293 - accuracy: 0.6900 - val_loss: 0.9105 - val_accuracy: 0.6719
Epoch 40/100
25/25 [==============================] - 39s 2s/step - loss: 0.6359 - accuracy: 0.7225 - val_loss: 0.8538 - val_accuracy: 0.6823
Epoch 41/100
25/25 [==============================] - 38s 2s/step - loss: 0.6523 - accuracy: 0.7287 - val_loss: 1.5683 - val_accuracy: 0.5417
Epoch 42/100
25/25 [==============================] - 39s 2s/step - loss: 0.6885 - accuracy: 0.7237 - val_loss: 0.9864 - val_accuracy: 0.6458
Epoch 43/100
25/25 [==============================] - 37s 1s/step - loss: 0.5583 - accuracy: 0.7775 - val_loss: 0.9455 - val_accuracy: 0.6198
Epoch 44/100
25/25 [==============================] - 38s 1s/step - loss: 0.5613 - accuracy: 0.7588 - val_loss: 0.8001 - val_accuracy: 0.6771
Epoch 45/100
25/25 [==============================] - 39s 2s/step - loss: 0.5904 - accuracy: 0.7850 - val_loss: 0.8891 - val_accuracy: 0.6719
Epoch 46/100
25/25 [==============================] - 40s 2s/step - loss: 0.5847 - accuracy: 0.7600 - val_loss: 0.7383 - val_accuracy: 0.7135
Epoch 47/100
25/25 [==============================] - 40s 2s/step - loss: 0.5609 - accuracy: 0.7650 - val_loss: 0.9535 - val_accuracy: 0.6354
Epoch 48/100
25/25 [==============================] - 41s 2s/step - loss: 0.5933 - accuracy: 0.7700 - val_loss: 0.8282 - val_accuracy: 0.6823
Epoch 49/100
25/25 [==============================] - 40s 2s/step - loss: 0.5148 - accuracy: 0.7862 - val_loss: 1.0083 - val_accuracy: 0.6094
Epoch 50/100
25/25 [==============================] - 41s 2s/step - loss: 0.5318 - accuracy: 0.7763 - val_loss: 0.7793 - val_accuracy: 0.6927
Epoch 51/100
25/25 [==============================] - 41s 2s/step - loss: 0.4546 - accuracy: 0.8138 - val_loss: 0.7321 - val_accuracy: 0.7396
Epoch 52/100
25/25 [==============================] - 41s 2s/step - loss: 0.4888 - accuracy: 0.7987 - val_loss: 0.7415 - val_accuracy: 0.7292
Epoch 53/100
25/25 [==============================] - 41s 2s/step - loss: 0.4525 - accuracy: 0.8263 - val_loss: 1.1359 - val_accuracy: 0.5938
Epoch 54/100
25/25 [==============================] - 41s 2s/step - loss: 0.4635 - accuracy: 0.8100 - val_loss: 0.8153 - val_accuracy: 0.7083
Epoch 55/100
25/25 [==============================] - 40s 2s/step - loss: 0.3715 - accuracy: 0.8587 - val_loss: 0.8006 - val_accuracy: 0.7083
Epoch 56/100
25/25 [==============================] - 40s 2s/step - loss: 0.7150 - accuracy: 0.7650 - val_loss: 0.6763 - val_accuracy: 0.7604
Epoch 57/100
25/25 [==============================] - 40s 2s/step - loss: 0.4236 - accuracy: 0.8400 - val_loss: 1.2931 - val_accuracy: 0.5625
Epoch 58/100
25/25 [==============================] - 40s 2s/step - loss: 0.4233 - accuracy: 0.8338 - val_loss: 0.7108 - val_accuracy: 0.7188
Epoch 59/100
25/25 [==============================] - 39s 2s/step - loss: 0.4240 - accuracy: 0.8263 - val_loss: 0.8515 - val_accuracy: 0.7656
Epoch 60/100
25/25 [==============================] - 40s 2s/step - loss: 0.2996 - accuracy: 0.8913 - val_loss: 1.1627 - val_accuracy: 0.6719
Epoch 61/100
25/25 [==============================] - 41s 2s/step - loss: 0.4147 - accuracy: 0.8438 - val_loss: 0.8675 - val_accuracy: 0.7656
Epoch 62/100
25/25 [==============================] - 40s 2s/step - loss: 0.3373 - accuracy: 0.8575 - val_loss: 1.8903 - val_accuracy: 0.5260
Epoch 63/100
25/25 [==============================] - 41s 2s/step - loss: 0.3147 - accuracy: 0.8913 - val_loss: 0.9597 - val_accuracy: 0.6927
Epoch 64/100
25/25 [==============================] - 40s 2s/step - loss: 0.4258 - accuracy: 0.8625 - val_loss: 0.7272 - val_accuracy: 0.7448
Epoch 65/100
25/25 [==============================] - 40s 2s/step - loss: 0.2791 - accuracy: 0.8950 - val_loss: 0.7932 - val_accuracy: 0.7396
Epoch 66/100
25/25 [==============================] - 40s 2s/step - loss: 0.2791 - accuracy: 0.8963 - val_loss: 1.1467 - val_accuracy: 0.6823
Epoch 67/100
25/25 [==============================] - 41s 2s/step - loss: 0.2417 - accuracy: 0.9050 - val_loss: 0.8308 - val_accuracy: 0.7344
Epoch 68/100
25/25 [==============================] - 43s 2s/step - loss: 0.4000 - accuracy: 0.8725 - val_loss: 0.8193 - val_accuracy: 0.6875
Epoch 69/100
25/25 [==============================] - 41s 2s/step - loss: 0.2515 - accuracy: 0.9162 - val_loss: 0.8325 - val_accuracy: 0.7396
Epoch 70/100
25/25 [==============================] - 40s 2s/step - loss: 0.2121 - accuracy: 0.9187 - val_loss: 0.9849 - val_accuracy: 0.7240
Epoch 71/100
25/25 [==============================] - 41s 2s/step - loss: 0.1987 - accuracy: 0.9262 - val_loss: 0.8387 - val_accuracy: 0.7760
Epoch 72/100
25/25 [==============================] - 40s 2s/step - loss: 0.2786 - accuracy: 0.8975 - val_loss: 0.7462 - val_accuracy: 0.7917
Epoch 73/100
25/25 [==============================] - 40s 2s/step - loss: 0.1309 - accuracy: 0.9625 - val_loss: 1.0813 - val_accuracy: 0.7448
Epoch 74/100
25/25 [==============================] - 40s 2s/step - loss: 0.3271 - accuracy: 0.9013 - val_loss: 0.7063 - val_accuracy: 0.7604
Epoch 75/100
25/25 [==============================] - 40s 2s/step - loss: 0.2142 - accuracy: 0.9187 - val_loss: 0.9269 - val_accuracy: 0.7708
Epoch 76/100
25/25 [==============================] - 40s 2s/step - loss: 0.1603 - accuracy: 0.9438 - val_loss: 0.8590 - val_accuracy: 0.7448
Epoch 77/100
25/25 [==============================] - 42s 2s/step - loss: 0.1331 - accuracy: 0.9513 - val_loss: 1.0895 - val_accuracy: 0.7083
Epoch 78/100
25/25 [==============================] - 41s 2s/step - loss: 0.1177 - accuracy: 0.9638 - val_loss: 1.0417 - val_accuracy: 0.7500
Epoch 79/100
25/25 [==============================] - 41s 2s/step - loss: 0.4039 - accuracy: 0.8775 - val_loss: 0.8521 - val_accuracy: 0.7240
Epoch 80/100
25/25 [==============================] - 41s 2s/step - loss: 0.1795 - accuracy: 0.9488 - val_loss: 0.9234 - val_accuracy: 0.7344
Epoch 81/100
25/25 [==============================] - 40s 2s/step - loss: 0.0812 - accuracy: 0.9850 - val_loss: 0.9745 - val_accuracy: 0.7656
Epoch 82/100
25/25 [==============================] - 40s 2s/step - loss: 0.0976 - accuracy: 0.9625 - val_loss: 1.0829 - val_accuracy: 0.7500
Epoch 83/100
25/25 [==============================] - 40s 2s/step - loss: 0.1220 - accuracy: 0.9513 - val_loss: 1.8248 - val_accuracy: 0.5885
Epoch 84/100
25/25 [==============================] - 40s 2s/step - loss: 0.7078 - accuracy: 0.8188 - val_loss: 0.7013 - val_accuracy: 0.7552
Epoch 85/100
25/25 [==============================] - 41s 2s/step - loss: 0.1607 - accuracy: 0.9613 - val_loss: 0.8556 - val_accuracy: 0.7656
Epoch 86/100
25/25 [==============================] - 41s 2s/step - loss: 0.0831 - accuracy: 0.9812 - val_loss: 1.0032 - val_accuracy: 0.7552
Epoch 87/100
25/25 [==============================] - 40s 2s/step - loss: 0.1305 - accuracy: 0.9575 - val_loss: 0.8887 - val_accuracy: 0.7604
Epoch 88/100
25/25 [==============================] - 40s 2s/step - loss: 0.0602 - accuracy: 0.9825 - val_loss: 1.1799 - val_accuracy: 0.7135
Epoch 89/100
25/25 [==============================] - 41s 2s/step - loss: 0.0469 - accuracy: 0.9887 - val_loss: 1.1216 - val_accuracy: 0.7396
Epoch 90/100
25/25 [==============================] - 40s 2s/step - loss: 0.1920 - accuracy: 0.9325 - val_loss: 1.2466 - val_accuracy: 0.6927
Epoch 91/100
25/25 [==============================] - 41s 2s/step - loss: 0.2436 - accuracy: 0.9275 - val_loss: 0.8773 - val_accuracy: 0.7656
Epoch 92/100
25/25 [==============================] - 41s 2s/step - loss: 0.0651 - accuracy: 0.9887 - val_loss: 1.0198 - val_accuracy: 0.7448
Epoch 93/100
25/25 [==============================] - 41s 2s/step - loss: 0.0424 - accuracy: 0.9912 - val_loss: 1.1194 - val_accuracy: 0.7448
Epoch 94/100
25/25 [==============================] - 41s 2s/step - loss: 0.0291 - accuracy: 0.9975 - val_loss: 1.2345 - val_accuracy: 0.7292
Epoch 95/100
25/25 [==============================] - 43s 2s/step - loss: 0.0237 - accuracy: 1.0000 - val_loss: 1.2051 - val_accuracy: 0.7292
Epoch 96/100
25/25 [==============================] - 43s 2s/step - loss: 0.9514 - accuracy: 0.8100 - val_loss: 0.9362 - val_accuracy: 0.6562
Epoch 97/100
25/25 [==============================] - 40s 2s/step - loss: 0.3801 - accuracy: 0.8913 - val_loss: 0.8821 - val_accuracy: 0.7396
Epoch 98/100
25/25 [==============================] - 39s 2s/step - loss: 0.1676 - accuracy: 0.9588 - val_loss: 0.8548 - val_accuracy: 0.7292
Epoch 99/100
25/25 [==============================] - 39s 2s/step - loss: 0.0977 - accuracy: 0.9812 - val_loss: 1.1352 - val_accuracy: 0.7135
Epoch 100/100
25/25 [==============================] - 39s 2s/step - loss: 0.1027 - accuracy: 0.9663 - val_loss: 0.8968 - val_accuracy: 0.8021

<keras.callbacks.History at 0x18ee5f3b490>

model.evaluate(test_ds)

8/8 [==============================] - 7s 327ms/step - loss: 1.0574 - accuracy: 0.7070

[1.0574449300765991, 0.70703125]

Wizualizacja filtrów obrazowych na poszczególnych warstwach

layers_names = []
for layer in model.layers:
    # check for convolutional layer
    if 'conv' not in layer.name:
        continue
    layers_names.append(layer.name)
    filters, biases = layer.get_weights()
    print(layer.name, filters.shape)

conv2d (11, 11, 3, 96)
conv2d_1 (5, 5, 96, 256)
conv2d_2 (3, 3, 256, 384)
conv2d_3 (3, 3, 384, 384)
conv2d_4 (3, 3, 384, 256)

import matplotlib.pyplot as plt

filters, biases = model.layers[0].get_weights()
fmin, fmax = filters.min(), filters.max()
filters = (filters - fmin) / (fmax - fmin)

nb_filters, tmp = 3, 1

for i in range(nb_filters):
    f = filters[:, :, :, i]
    for j in range(3):
        ax = plt.subplot(nb_filters, 3, tmp)
        ax.set_xticks([])
        ax.set_yticks([])
        plt.imshow(f[:, :, j], cmap='gray')
        tmp += 1
plt.show()

img_width, img_height = 227, 227

layer = model.get_layer(name=layers_names[0])
feature_extractor = keras.Model(inputs=model.inputs, outputs=layer.output)

def initialize_image():
    # random noisy img
    img = tf.random.uniform((1, img_width, img_height, 3))
    return img

def loss_calc(input_image, filter_index, fex):
    activation = fex(input_image)
    # getting rid of the border pixels so they don't inlfuence our results in any fun way
    filter_activation = activation[:, 2:-2, 2:-2, filter_index]
    return tf.reduce_mean(filter_activation)

def gradient_ascent_step(img, filter_index, learning_rate, fex):
    with tf.GradientTape() as tape:
        tape.watch(img)
        loss = loss_calc(img, filter_index, fex)
    grads = tape.gradient(loss, img)
    grads = tf.math.l2_normalize(grads)
    img += learning_rate * grads
    return loss, img

def deprocess_image(img):
    img -= img.mean()
    img /= img.std() + 1e-5
    img *= 0.15

    img = img[25:-25, 25:-25, :]

    img += 0.5
    img = np.clip(img, 0, 1)

    img *= 255
    img = np.clip(img, 0, 255).astype("uint8")
    return img

def visualize_filter(filter_index, fex):
    # 20 steps grad desc
    iterations = 30
    learning_rate = 10.0
    img = initialize_image()
    for iteration in range(iterations):
        loss, img = gradient_ascent_step(img, filter_index, learning_rate, fex)
    img = deprocess_image(img[0].numpy())
    return loss, img

loss, img = visualize_filter(0, feature_extractor)
keras.preprocessing.image.save_img("lab7props/0.png", img)

display(Image("lab7props/0.png"))

def visualize_layer(layer_name, n):
    from IPython.display import Image, display
    
    layer = model.get_layer(name=layer_name)
    fex = keras.Model(inputs=model.inputs, outputs=layer.output)
    
    print(f"Getting started with layer {layer_name}")
    all_imgs = []
    for filter_index in range(n**2):
        # print("Processing filter %d" % (filter_index,))
        loss, img = visualize_filter(filter_index, fex)
        all_imgs.append(img)

    margin = 5
    cropped_width = img_width - 25 * 2
    cropped_height = img_height - 25 * 2
    width = n * cropped_width + (n - 1) * margin
    height = n * cropped_height + (n - 1) * margin
    stitched_filters = np.zeros((width, height, 3))

    for i in range(n):
        for j in range(n):
            img = all_imgs[i * n + j]
            stitched_filters[
                (cropped_width + margin) * i : (cropped_width + margin) * i + cropped_width,
                (cropped_height + margin) * j : (cropped_height + margin) * j
                + cropped_height,
                :,
            ] = img

    filename = f"lab7props/{layer_name}_stitched_filters.png"
    keras.preprocessing.image.save_img(filename, stitched_filters)
    print(f"{layer_name} done")

layers_names

['conv2d', 'conv2d_1', 'conv2d_2', 'conv2d_3', 'conv2d_4']

for name in layers_names:
    visualize_layer(name, 8)

Getting started with layer conv2d
conv2d done
Getting started with layer conv2d_1
conv2d_1 done
Getting started with layer conv2d_2
conv2d_2 done
Getting started with layer conv2d_3
conv2d_3 done
Getting started with layer conv2d_4
conv2d_4 done

display(Image(f"lab7props/{layers_names[0]}_stitched_filters.png"))