cut-mix.ipynb

{"metadata":{"kernelspec":{"display_name":"Python 3","language":"python","name":"python3"},"language_info":{"codemirror_mode":{"name":"ipython","version":3},"file_extension":".py","mimetype":"text/x-python","name":"python","nbconvert_exporter":"python","pygments_lexer":"ipython3","version":"3.7.12"},"papermill":{"default_parameters":{},"duration":2956.009394,"end_time":"2022-11-22T16:04:31.886387","environment_variables":{},"exception":null,"input_path":"__notebook__.ipynb","output_path":"__notebook__.ipynb","parameters":{},"start_time":"2022-11-22T15:15:15.876993","version":"2.3.4"}},"nbformat_minor":5,"nbformat":4,"cells":[{"cell_type":"markdown","source":"**Initial setup**","metadata":{"papermill":{"duration":0.010884,"end_time":"2022-11-22T15:15:23.782371","exception":false,"start_time":"2022-11-22T15:15:23.771487","status":"completed"},"tags":[]}},{"cell_type":"markdown","source":"**To use keras_vc we have to do these initial operations**","metadata":{}},{"cell_type":"code","source":"! apt -y install --allow-change-held-packages libcudnn8=8.6.0.163-1+cuda11.8","metadata":{"papermill":{"duration":123.591281,"end_time":"2022-11-22T15:17:27.383538","exception":false,"start_time":"2022-11-22T15:15:23.792257","status":"completed"},"tags":[],"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"! pip uninstall -y tensorflow tensorflow-gpu \n! pip uninstall -y tensorflow-io\n! pip uninstall -y tensorflow-transform","metadata":{"papermill":{"duration":22.803617,"end_time":"2022-11-22T15:17:50.246085","exception":false,"start_time":"2022-11-22T15:17:27.442468","status":"completed"},"tags":[],"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"! pip install tensorflow-transform\n! pip install tensorflow==2.10.1 tensorflow-gpu==2.10.1\n! pip install keras-cv --upgrade","metadata":{"papermill":{"duration":167.134905,"end_time":"2022-11-22T15:20:37.395916","exception":false,"start_time":"2022-11-22T15:17:50.261011","status":"completed"},"tags":[],"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"**Libraries**","metadata":{"papermill":{"duration":0.075928,"end_time":"2022-11-22T15:20:37.548374","exception":false,"start_time":"2022-11-22T15:20:37.472446","status":"completed"},"tags":[]}},{"cell_type":"code","source":"import tensorflow as tf\nimport numpy as np\nfrom numpy import asarray\nimport os\nimport random\nimport pandas as pd\nimport seaborn as sns\nimport matplotlib as mpl\nimport matplotlib.pyplot as plt\n\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score\nfrom sklearn.metrics import confusion_matrix\n\nfrom PIL import Image\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator\nfrom sklearn.utils import class_weight\nimport datetime \nimport pathlib\n\nimport keras_cv\n\nprint('Keras_CV : ', keras_cv.__version__)\ntfk = tf.keras\ntfkl = tf.keras.layers\nprint('Tensorflow : ', tf.__version__)","metadata":{"id":"f_sOaV1Y8NsL","outputId":"8d72ef20-426f-4bd6-961a-17c6521288a2","papermill":{"duration":5.341585,"end_time":"2022-11-22T15:20:42.965152","exception":false,"start_time":"2022-11-22T15:20:37.623567","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:29:08.584621Z","iopub.execute_input":"2022-11-22T22:29:08.585068Z","iopub.status.idle":"2022-11-22T22:29:11.441742Z","shell.execute_reply.started":"2022-11-22T22:29:08.584978Z","shell.execute_reply":"2022-11-22T22:29:11.440603Z"},"trusted":true},"execution_count":1,"outputs":[{"name":"stderr","text":"2022-11-22 22:29:08.597940: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 AVX512F FMA\nTo enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.\n2022-11-22 22:29:08.771863: E tensorflow/stream_executor/cuda/cuda_blas.cc:2981] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered\n2022-11-22 22:29:09.643269: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda/lib64:/usr/local/cuda/lib:/usr/local/lib/x86_64-linux-gnu:/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/usr/local/cuda/lib64:/usr/local/cuda/lib:/usr/local/lib/x86_64-linux-gnu:/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/usr/local/nvidia/lib:/usr/local/nvidia/lib64\n2022-11-22 22:29:09.643408: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /usr/local/cuda/lib64:/usr/local/cuda/lib:/usr/local/lib/x86_64-linux-gnu:/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/usr/local/cuda/lib64:/usr/local/cuda/lib:/usr/local/lib/x86_64-linux-gnu:/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/usr/local/nvidia/lib:/usr/local/nvidia/lib64\n2022-11-22 22:29:09.643424: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.\n","output_type":"stream"},{"name":"stdout","text":"Keras_CV :  0.3.4\nTensorflow :  2.10.1\n","output_type":"stream"}]},{"cell_type":"code","source":"seed = 42127\nrandom.seed(seed)\nos.environ['PYTHONHASHSEED'] = str(seed)\nnp.random.seed(seed)\ntf.random.set_seed(seed)\ntf.compat.v1.set_random_seed(seed)","metadata":{"id":"C7HYua8HUHIj","papermill":{"duration":0.087517,"end_time":"2022-11-22T15:20:43.129806","exception":false,"start_time":"2022-11-22T15:20:43.042289","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:29:13.122066Z","iopub.execute_input":"2022-11-22T22:29:13.122986Z","iopub.status.idle":"2022-11-22T22:29:13.130383Z","shell.execute_reply.started":"2022-11-22T22:29:13.122945Z","shell.execute_reply":"2022-11-22T22:29:13.129352Z"},"trusted":true},"execution_count":2,"outputs":[]},{"cell_type":"markdown","source":"### Suppress warnings","metadata":{"id":"BDM82PpE3VSg","papermill":{"duration":0.076309,"end_time":"2022-11-22T15:20:43.282432","exception":false,"start_time":"2022-11-22T15:20:43.206123","status":"completed"},"tags":[]}},{"cell_type":"code","source":"import warnings\nimport logging\n\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'\nwarnings.simplefilter(action='ignore', category=FutureWarning)\nwarnings.simplefilter(action='ignore', category=Warning)\ntf.get_logger().setLevel('INFO')\ntf.autograph.set_verbosity(0)\n\ntf.get_logger().setLevel(logging.ERROR)\ntf.get_logger().setLevel('ERROR')\ntf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)","metadata":{"id":"5fXtacjAqOIq","papermill":{"duration":0.087235,"end_time":"2022-11-22T15:20:43.445170","exception":false,"start_time":"2022-11-22T15:20:43.357935","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:29:17.372926Z","iopub.execute_input":"2022-11-22T22:29:17.373400Z","iopub.status.idle":"2022-11-22T22:29:17.381035Z","shell.execute_reply.started":"2022-11-22T22:29:17.373359Z","shell.execute_reply":"2022-11-22T22:29:17.379886Z"},"trusted":true},"execution_count":3,"outputs":[]},{"cell_type":"markdown","source":"# **Dataset preparation**","metadata":{"id":"pkxOZykBOLpX","papermill":{"duration":0.076624,"end_time":"2022-11-22T15:20:43.598826","exception":false,"start_time":"2022-11-22T15:20:43.522202","status":"completed"},"tags":[]}},{"cell_type":"code","source":"imbalClasses = tf.keras.utils.image_dataset_from_directory( \n    '/kaggle/input/dataset-1and6/only1and6',\n    labels='inferred',\n    label_mode='categorical',\n    class_names=None,\n    color_mode='rgb',\n    batch_size=32,\n    image_size=(96,96),\n    subset = 'training',\n    validation_split=0.15,\n    seed=seed,\n)","metadata":{"execution":{"iopub.status.busy":"2022-11-22T22:29:20.999129Z","iopub.execute_input":"2022-11-22T22:29:20.999513Z","iopub.status.idle":"2022-11-22T22:29:22.495048Z","shell.execute_reply.started":"2022-11-22T22:29:20.999482Z","shell.execute_reply":"2022-11-22T22:29:22.494081Z"},"trusted":true},"execution_count":4,"outputs":[{"name":"stdout","text":"Found 414 files belonging to 8 classes.\nUsing 352 files for training.\n","output_type":"stream"},{"name":"stderr","text":"2022-11-22 22:29:21.258515: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2022-11-22 22:29:21.272404: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2022-11-22 22:29:21.273178: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2022-11-22 22:29:21.274303: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 AVX512F FMA\nTo enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.\n2022-11-22 22:29:21.274642: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2022-11-22 22:29:21.275342: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2022-11-22 22:29:21.276024: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2022-11-22 22:29:22.098792: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2022-11-22 22:29:22.099714: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2022-11-22 22:29:22.100462: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2022-11-22 22:29:22.101072: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1616] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 15361 MB memory:  -> device: 0, name: Tesla P100-PCIE-16GB, pci bus id: 0000:00:04.0, compute capability: 6.0\n","output_type":"stream"}]},{"cell_type":"code","source":"AUTOTUNE = tf.data.AUTOTUNE\ndatapath = '/kaggle/input/dataset/training_data_final'\ntraining_set = tf.keras.utils.image_dataset_from_directory(\n    datapath,\n    labels='inferred',\n    label_mode='categorical',\n    color_mode='rgb',\n    image_size=(96,96),\n    shuffle=True,\n    seed=seed,\n    validation_split=0.15,\n    subset='training'\n)\n\nvalidation_set = tf.keras.utils.image_dataset_from_directory(\n    datapath,\n    labels='inferred',\n    label_mode='categorical',\n    color_mode='rgb',\n    image_size=(96,96),\n    shuffle=True,\n    seed=seed,\n    validation_split=0.15,\n    subset='validation'\n)\n","metadata":{"id":"9uJm0KSXDQcS","outputId":"c5705de6-b239-432e-9a85-8610f09208dc","papermill":{"duration":2.903079,"end_time":"2022-11-22T15:20:46.899959","exception":false,"start_time":"2022-11-22T15:20:43.996880","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:29:36.220449Z","iopub.execute_input":"2022-11-22T22:29:36.221672Z","iopub.status.idle":"2022-11-22T22:29:37.606122Z","shell.execute_reply.started":"2022-11-22T22:29:36.221598Z","shell.execute_reply":"2022-11-22T22:29:37.605091Z"},"trusted":true},"execution_count":5,"outputs":[{"name":"stdout","text":"Found 3542 files belonging to 8 classes.\nUsing 3011 files for training.\nFound 3542 files belonging to 8 classes.\nUsing 531 files for validation.\n","output_type":"stream"}]},{"cell_type":"markdown","source":"# **Dataset augmentation**","metadata":{"id":"bU42tKrxOLpc","papermill":{"duration":0.085232,"end_time":"2022-11-22T15:20:50.017429","exception":false,"start_time":"2022-11-22T15:20:49.932197","status":"completed"},"tags":[]}},{"cell_type":"markdown","source":"**Augmenter** \n\n**We build this Augmenter, made of three layers: a generic augmentation layer (implemented with keras_cv layer RandAugment), a CutMix layer and a MixUp layer.**","metadata":{"papermill":{"duration":0.113318,"end_time":"2022-11-22T15:21:04.786512","exception":false,"start_time":"2022-11-22T15:21:04.673194","status":"completed"},"tags":[]}},{"cell_type":"code","source":"\naugmenter = keras_cv.layers.Augmenter( \n  layers=[\n      keras_cv.layers.RandAugment( \n        value_range=(0, 255),\n        augmentations_per_image=3,\n        magnitude=0.3,\n        magnitude_stddev=0.2,\n        rate=0.5,),\n      keras_cv.layers.CutMix(),\n      keras_cv.layers.MixUp()\n    ]\n)\n\ndef augment_data(images, labels):\n    inputs = {\"images\": images, \"labels\": labels}\n    outputs = augmenter(inputs)\n    return outputs['images'], outputs['labels']\n","metadata":{"papermill":{"duration":3.148457,"end_time":"2022-11-22T15:21:08.043902","exception":false,"start_time":"2022-11-22T15:21:04.895445","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:29:58.008554Z","iopub.execute_input":"2022-11-22T22:29:58.008955Z","iopub.status.idle":"2022-11-22T22:29:58.027648Z","shell.execute_reply.started":"2022-11-22T22:29:58.008924Z","shell.execute_reply":"2022-11-22T22:29:58.026637Z"},"trusted":true},"execution_count":9,"outputs":[]},{"cell_type":"markdown","source":"**Now we do the actual augmentation**\n\n**As usual, we first do the selective augmentation on classes 1 and 6 and then the total augmentation on the training set**","metadata":{}},{"cell_type":"code","source":"dataAugmImbal = imbalClasses.map(augment_data, num_parallel_calls=tf.data.AUTOTUNE) \ndataAugmImbal = dataAugmImbal.concatenate(imbalClasses.map(augment_data, num_parallel_calls=tf.data.AUTOTUNE))","metadata":{"execution":{"iopub.status.busy":"2022-11-22T22:30:01.131972Z","iopub.execute_input":"2022-11-22T22:30:01.132352Z","iopub.status.idle":"2022-11-22T22:30:08.910592Z","shell.execute_reply.started":"2022-11-22T22:30:01.132319Z","shell.execute_reply":"2022-11-22T22:30:08.909562Z"},"trusted":true},"execution_count":10,"outputs":[]},{"cell_type":"code","source":"training_set = training_set.concatenate(training_set.map(augment_data, num_parallel_calls=tf.data.AUTOTUNE) ) #total augm\ntraining_set = training_set.concatenate(dataAugmImbal)\ntraining_set = training_set.concatenate(imbalClasses) #OVERSAMPLING classes 1 and 6 ","metadata":{"execution":{"iopub.status.busy":"2022-11-22T22:30:08.912429Z","iopub.execute_input":"2022-11-22T22:30:08.913346Z","iopub.status.idle":"2022-11-22T22:30:10.722912Z","shell.execute_reply.started":"2022-11-22T22:30:08.913307Z","shell.execute_reply":"2022-11-22T22:30:10.721788Z"},"trusted":true},"execution_count":11,"outputs":[]},{"cell_type":"code","source":"training_set = training_set.cache().shuffle(buffer_size=1000).prefetch(buffer_size=AUTOTUNE)\nvalidation_set = validation_set.cache().prefetch(buffer_size=AUTOTUNE)","metadata":{"id":"0E0fe9CVDQcU","papermill":{"duration":0.133408,"end_time":"2022-11-22T15:21:14.146612","exception":false,"start_time":"2022-11-22T15:21:14.013204","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:30:10.969173Z","iopub.execute_input":"2022-11-22T22:30:10.969841Z","iopub.status.idle":"2022-11-22T22:30:10.985255Z","shell.execute_reply.started":"2022-11-22T22:30:10.969795Z","shell.execute_reply":"2022-11-22T22:30:10.984344Z"},"trusted":true},"execution_count":12,"outputs":[]},{"cell_type":"markdown","source":"**Preprocessing**","metadata":{"id":"a7QaFDJgOLpd","papermill":{"duration":0.160179,"end_time":"2022-11-22T15:21:14.421966","exception":false,"start_time":"2022-11-22T15:21:14.261787","status":"completed"},"tags":[]}},{"cell_type":"code","source":"training_set = training_set.map(lambda x,y: (tf.keras.applications.inception_v3.preprocess_input(x), y))\nvalidation_set = validation_set.map(lambda x,y: (tf.keras.applications.inception_v3.preprocess_input(x),y))","metadata":{"id":"BiniALJ1DQcW","papermill":{"duration":0.177269,"end_time":"2022-11-22T15:21:14.718382","exception":false,"start_time":"2022-11-22T15:21:14.541113","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:30:14.649227Z","iopub.execute_input":"2022-11-22T22:30:14.649650Z","iopub.status.idle":"2022-11-22T22:30:14.706560Z","shell.execute_reply.started":"2022-11-22T22:30:14.649613Z","shell.execute_reply":"2022-11-22T22:30:14.705700Z"},"trusted":true},"execution_count":13,"outputs":[]},{"cell_type":"markdown","source":"# **Network**","metadata":{"id":"ELg2uP8hNXCv","papermill":{"duration":0.116096,"end_time":"2022-11-22T15:21:14.949514","exception":false,"start_time":"2022-11-22T15:21:14.833418","status":"completed"},"tags":[]}},{"cell_type":"code","source":"myInception = tfk.applications.inception_v3.InceptionV3(include_top = False, weights='imagenet')","metadata":{"papermill":{"duration":6.472166,"end_time":"2022-11-22T15:21:22.004880","exception":false,"start_time":"2022-11-22T15:21:15.532714","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:30:17.853521Z","iopub.execute_input":"2022-11-22T22:30:17.854249Z","iopub.status.idle":"2022-11-22T22:30:20.711681Z","shell.execute_reply.started":"2022-11-22T22:30:17.854210Z","shell.execute_reply":"2022-11-22T22:30:20.710645Z"},"trusted":true},"execution_count":14,"outputs":[]},{"cell_type":"code","source":"myInception.trainable = True\n","metadata":{"id":"RnrDwtZDOLpe","papermill":{"duration":0.180344,"end_time":"2022-11-22T15:21:22.302779","exception":false,"start_time":"2022-11-22T15:21:22.122435","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:30:20.713422Z","iopub.execute_input":"2022-11-22T22:30:20.713798Z","iopub.status.idle":"2022-11-22T22:30:20.729324Z","shell.execute_reply.started":"2022-11-22T22:30:20.713767Z","shell.execute_reply":"2022-11-22T22:30:20.728121Z"},"trusted":true},"execution_count":15,"outputs":[]},{"cell_type":"code","source":"input_shape = (96, 96, 3)\nepochs = 100\n\nl1 = 0.001\nl2 = 0.001\ndropout_rate = 0.35\n\nlearning_rate = 1e-4 \n","metadata":{"id":"VwX67L6CNqt2","papermill":{"duration":0.210408,"end_time":"2022-11-22T15:21:22.867141","exception":false,"start_time":"2022-11-22T15:21:22.656733","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:30:29.089954Z","iopub.execute_input":"2022-11-22T22:30:29.090345Z","iopub.status.idle":"2022-11-22T22:30:29.096236Z","shell.execute_reply.started":"2022-11-22T22:30:29.090311Z","shell.execute_reply":"2022-11-22T22:30:29.095136Z"},"trusted":true},"execution_count":16,"outputs":[]},{"cell_type":"markdown","source":"**Classifier definition**","metadata":{"id":"u88qenMWOLpf","papermill":{"duration":0.17984,"end_time":"2022-11-22T15:21:23.229509","exception":false,"start_time":"2022-11-22T15:21:23.049669","status":"completed"},"tags":[]}},{"cell_type":"code","source":"custom_classifier = tfk.Sequential([ \n        tfkl.Input(shape = (2048,)), \n        tfkl.experimental.RandomFourierFeatures(output_dim=1024, trainable = True , kernel_initializer=\"gaussian\"),  \n        tfkl.Dropout(rate = dropout_rate), \n        tfkl.Dense(units = 80,\n                   activation = \"relu\", \n                   kernel_regularizer = tfk.regularizers.L1L2(l2 = l2),\n                   kernel_initializer=tfk.initializers.HeUniform(seed*2),\n                   name = 'Dense'),\n        tfkl.Dense(units=8, \n                   activation='softmax', \n                   kernel_initializer=tfk.initializers.GlorotUniform(seed*2), \n                   name='out', \n                   kernel_regularizer = tfk.regularizers.L2(l2 = l2)                   \n                   )\n    ],\n    name='classifier'\n)","metadata":{"id":"xaRzXuN3rrqb","papermill":{"duration":0.275782,"end_time":"2022-11-22T15:21:23.693007","exception":false,"start_time":"2022-11-22T15:21:23.417225","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:30:32.784944Z","iopub.execute_input":"2022-11-22T22:30:32.785614Z","iopub.status.idle":"2022-11-22T22:30:32.854691Z","shell.execute_reply.started":"2022-11-22T22:30:32.785556Z","shell.execute_reply":"2022-11-22T22:30:32.853748Z"},"trusted":true},"execution_count":17,"outputs":[]},{"cell_type":"code","source":"custom_classifier.summary() ","metadata":{"papermill":{"duration":0.175794,"end_time":"2022-11-22T15:21:24.049709","exception":false,"start_time":"2022-11-22T15:21:23.873915","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:30:36.453495Z","iopub.execute_input":"2022-11-22T22:30:36.454444Z","iopub.status.idle":"2022-11-22T22:30:36.477502Z","shell.execute_reply.started":"2022-11-22T22:30:36.454395Z","shell.execute_reply":"2022-11-22T22:30:36.476701Z"},"trusted":true},"execution_count":18,"outputs":[{"name":"stdout","text":"Model: \"classifier\"\n_________________________________________________________________\n Layer (type)                Output Shape              Param #   \n=================================================================\n random_fourier_features (Ra  (None, 1024)             2098177   \n ndomFourierFeatures)                                            \n                                                                 \n dropout (Dropout)           (None, 1024)              0         \n                                                                 \n Dense (Dense)               (None, 80)                82000     \n                                                                 \n out (Dense)                 (None, 8)                 648       \n                                                                 \n=================================================================\nTotal params: 2,180,825\nTrainable params: 82,649\nNon-trainable params: 2,098,176\n_________________________________________________________________\n","output_type":"stream"}]},{"cell_type":"markdown","source":"**Model definition**","metadata":{"id":"MQ0CVKoMOLpf","papermill":{"duration":0.126469,"end_time":"2022-11-22T15:21:24.294615","exception":false,"start_time":"2022-11-22T15:21:24.168146","status":"completed"},"tags":[]}},{"cell_type":"code","source":"def build_model(input_shape): \n    input_layer = tfkl.Input(input_shape)\n    x = tfkl.Resizing(height = 299, width = 299, interpolation = 'bicubic' )(input_layer)\n    x = myInception(x)\n    x = tfkl.Dropout(rate=dropout_rate)(x)\n    x = tfkl.GlobalAveragePooling2D()(x) \n    output_layer = custom_classifier(x)\n    model = tfk.Model(inputs=input_layer, outputs=output_layer, name='model')\n    model.compile(loss=tfk.losses.CategoricalCrossentropy(label_smoothing=0.1), optimizer=tf.keras.optimizers.Adam(learning_rate), metrics='accuracy')\n    return model \n\nmodel = build_model(input_shape)","metadata":{"id":"qXlq70zBNMCb","papermill":{"duration":0.890165,"end_time":"2022-11-22T15:21:25.306212","exception":false,"start_time":"2022-11-22T15:21:24.416047","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:30:42.157131Z","iopub.execute_input":"2022-11-22T22:30:42.157543Z","iopub.status.idle":"2022-11-22T22:30:42.871918Z","shell.execute_reply.started":"2022-11-22T22:30:42.157510Z","shell.execute_reply":"2022-11-22T22:30:42.870890Z"},"trusted":true},"execution_count":19,"outputs":[]},{"cell_type":"markdown","source":"**Callbacks definition**","metadata":{"id":"RKV6Ps1pOLpg","papermill":{"duration":0.12096,"end_time":"2022-11-22T15:21:25.800970","exception":false,"start_time":"2022-11-22T15:21:25.680010","status":"completed"},"tags":[]}},{"cell_type":"code","source":"from datetime import datetime\ndef create_folders_and_callbacks(model_name):\n    exps_dir = os.path.join('experiments')\n    if not os.path.exists(exps_dir):\n        os.makedirs(exps_dir)\n    now = datetime.now().strftime('%b%d_%H-%M-%S')\n    exp_dir = os.path.join(exps_dir, model_name + '_' + str(now))\n    if not os.path.exists(exp_dir):\n        os.makedirs(exp_dir)\n    callbacks = []\n    \n    # Model checkpoint\n      # ----------------\n    ckpt_dir = os.path.join(exp_dir, 'ckpts')\n    if not os.path.exists(ckpt_dir):\n        os.makedirs(ckpt_dir)\n          \n    ckpt_callback = tf.keras.callbacks.ModelCheckpoint(filepath=os.path.join(ckpt_dir, 'cp.ckpt'), save_weights_only=True, save_best_only=True) #play here \n      \n    callbacks.append(ckpt_callback)\n    \n    # Early Stopping\n    # --------------\n    es_callback = tf.keras.callbacks.EarlyStopping(monitor='val_accuracy', patience=16, restore_best_weights=True) #tuning point\n    callbacks.append(es_callback)\n    return callbacks","metadata":{"id":"6qbixQiprQl_","papermill":{"duration":0.135685,"end_time":"2022-11-22T15:21:26.059431","exception":false,"start_time":"2022-11-22T15:21:25.923746","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:30:46.753473Z","iopub.execute_input":"2022-11-22T22:30:46.754168Z","iopub.status.idle":"2022-11-22T22:30:46.762327Z","shell.execute_reply.started":"2022-11-22T22:30:46.754133Z","shell.execute_reply":"2022-11-22T22:30:46.761152Z"},"trusted":true},"execution_count":20,"outputs":[]},{"cell_type":"code","source":"model.summary()","metadata":{"id":"hlVvuyP-DQcY","outputId":"3a9ba965-7d76-408d-a7f1-e6c7abd55b22","papermill":{"duration":0.166004,"end_time":"2022-11-22T15:21:26.345451","exception":false,"start_time":"2022-11-22T15:21:26.179447","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:30:50.570075Z","iopub.execute_input":"2022-11-22T22:30:50.570467Z","iopub.status.idle":"2022-11-22T22:30:50.617786Z","shell.execute_reply.started":"2022-11-22T22:30:50.570434Z","shell.execute_reply":"2022-11-22T22:30:50.616793Z"},"trusted":true},"execution_count":21,"outputs":[{"name":"stdout","text":"Model: \"model\"\n_________________________________________________________________\n Layer (type)                Output Shape              Param #   \n=================================================================\n input_3 (InputLayer)        [(None, 96, 96, 3)]       0         \n                                                                 \n resizing (Resizing)         (None, 299, 299, 3)       0         \n                                                                 \n inception_v3 (Functional)   (None, None, None, 2048)  21802784  \n                                                                 \n dropout_1 (Dropout)         (None, 8, 8, 2048)        0         \n                                                                 \n global_average_pooling2d (G  (None, 2048)             0         \n lobalAveragePooling2D)                                          \n                                                                 \n classifier (Sequential)     (None, 8)                 2180825   \n                                                                 \n=================================================================\nTotal params: 23,983,609\nTrainable params: 21,851,001\nNon-trainable params: 2,132,608\n_________________________________________________________________\n","output_type":"stream"}]},{"cell_type":"code","source":"\nfor i, layer in enumerate(model.get_layer('inception_v3').layers[:-1]):\n    layer.trainable=False\n\nfor i, layer in enumerate(model.get_layer('inception_v3').layers[:30]):\n    layer.trainable=True\n    \nfor i, layer in enumerate(model.get_layer('inception_v3').layers[-40:-1]):\n    layer.trainable = True\n\nmodel.summary()","metadata":{"execution":{"iopub.status.busy":"2022-11-22T22:30:55.212822Z","iopub.execute_input":"2022-11-22T22:30:55.213450Z","iopub.status.idle":"2022-11-22T22:30:55.272048Z","shell.execute_reply.started":"2022-11-22T22:30:55.213412Z","shell.execute_reply":"2022-11-22T22:30:55.271030Z"},"trusted":true},"execution_count":22,"outputs":[{"name":"stdout","text":"Model: \"model\"\n_________________________________________________________________\n Layer (type)                Output Shape              Param #   \n=================================================================\n input_3 (InputLayer)        [(None, 96, 96, 3)]       0         \n                                                                 \n resizing (Resizing)         (None, 299, 299, 3)       0         \n                                                                 \n inception_v3 (Functional)   (None, None, None, 2048)  21802784  \n                                                                 \n dropout_1 (Dropout)         (None, 8, 8, 2048)        0         \n                                                                 \n global_average_pooling2d (G  (None, 2048)             0         \n lobalAveragePooling2D)                                          \n                                                                 \n classifier (Sequential)     (None, 8)                 2180825   \n                                                                 \n=================================================================\nTotal params: 23,983,609\nTrainable params: 6,494,745\nNon-trainable params: 17,488,864\n_________________________________________________________________\n","output_type":"stream"}]},{"cell_type":"code","source":"MODEL_NAME = 'cut_mix' ","metadata":{"id":"29DMTAkdaISz","papermill":{"duration":0.133415,"end_time":"2022-11-22T15:21:26.597775","exception":false,"start_time":"2022-11-22T15:21:26.464360","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:30:59.900910Z","iopub.execute_input":"2022-11-22T22:30:59.901314Z","iopub.status.idle":"2022-11-22T22:30:59.906696Z","shell.execute_reply.started":"2022-11-22T22:30:59.901280Z","shell.execute_reply":"2022-11-22T22:30:59.905501Z"},"trusted":true},"execution_count":23,"outputs":[]},{"cell_type":"markdown","source":"# **Training**","metadata":{"id":"v6HIUYTFOLph","papermill":{"duration":0.136377,"end_time":"2022-11-22T15:21:26.854546","exception":false,"start_time":"2022-11-22T15:21:26.718169","status":"completed"},"tags":[]}},{"cell_type":"code","source":"callbacks = create_folders_and_callbacks(model_name = MODEL_NAME)\n\nhistory = model.fit(\n    x = training_set,\n    epochs = epochs,\n    validation_data = validation_set, \n    callbacks = callbacks,\n).history","metadata":{"id":"pXAHAuoSz-_S","outputId":"27d152ec-c0a0-4ba4-a3d9-e0f9004c9b11","papermill":{"duration":1492.81767,"end_time":"2022-11-22T15:46:19.826482","exception":false,"start_time":"2022-11-22T15:21:27.008812","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T22:31:02.424257Z","iopub.execute_input":"2022-11-22T22:31:02.424661Z","iopub.status.idle":"2022-11-22T23:01:54.010834Z","shell.execute_reply.started":"2022-11-22T22:31:02.424627Z","shell.execute_reply":"2022-11-22T23:01:54.009838Z"},"trusted":true},"execution_count":24,"outputs":[{"name":"stdout","text":"Epoch 1/100\n","output_type":"stream"},{"name":"stderr","text":"2022-11-22 22:31:20.873329: I tensorflow/core/kernels/data/shuffle_dataset_op.cc:390] Filling up shuffle buffer (this may take a while): 149 of 1000\n2022-11-22 22:31:31.026145: I tensorflow/core/kernels/data/shuffle_dataset_op.cc:390] Filling up shuffle buffer (this may take a while): 206 of 1000\n2022-11-22 22:31:32.067092: I tensorflow/core/kernels/data/shuffle_dataset_op.cc:415] Shuffle buffer filled.\n2022-11-22 22:31:32.591367: I tensorflow/stream_executor/cuda/cuda_dnn.cc:384] Loaded cuDNN version 8600\n2022-11-22 22:31:33.205644: W tensorflow/stream_executor/gpu/asm_compiler.cc:111] *** WARNING *** You are using ptxas 11.0.221, which is older than 11.1. ptxas before 11.1 is known to miscompile XLA code, leading to incorrect results or invalid-address errors.\n\nYou may not need to update to CUDA 11.1; cherry-picking the ptxas binary is often sufficient.\n","output_type":"stream"},{"name":"stdout","text":"223/223 [==============================] - 96s 289ms/step - loss: 1.9739 - accuracy: 0.4087 - val_loss: 1.4188 - val_accuracy: 0.6347\nEpoch 2/100\n223/223 [==============================] - 61s 274ms/step - loss: 1.5321 - accuracy: 0.6769 - val_loss: 1.1761 - val_accuracy: 0.7684\nEpoch 3/100\n223/223 [==============================] - 60s 271ms/step - loss: 1.3001 - accuracy: 0.8186 - val_loss: 1.1650 - val_accuracy: 0.7552\nEpoch 4/100\n223/223 [==============================] - 61s 272ms/step - loss: 1.2062 - accuracy: 0.8593 - val_loss: 1.0289 - val_accuracy: 0.8324\nEpoch 5/100\n223/223 [==============================] - 59s 264ms/step - loss: 1.1545 - accuracy: 0.8823 - val_loss: 1.0450 - val_accuracy: 0.8060\nEpoch 6/100\n223/223 [==============================] - 61s 272ms/step - loss: 1.1149 - accuracy: 0.8901 - val_loss: 1.0053 - val_accuracy: 0.8343\nEpoch 7/100\n223/223 [==============================] - 61s 271ms/step - loss: 1.0841 - accuracy: 0.8993 - val_loss: 0.9498 - val_accuracy: 0.8531\nEpoch 8/100\n223/223 [==============================] - 61s 272ms/step - loss: 1.0574 - accuracy: 0.8952 - val_loss: 0.9430 - val_accuracy: 0.8606\nEpoch 9/100\n223/223 [==============================] - 59s 263ms/step - loss: 1.0419 - accuracy: 0.9001 - val_loss: 0.9797 - val_accuracy: 0.8211\nEpoch 10/100\n223/223 [==============================] - 60s 271ms/step - loss: 1.0226 - accuracy: 0.9045 - val_loss: 0.8876 - val_accuracy: 0.8964\nEpoch 11/100\n223/223 [==============================] - 59s 264ms/step - loss: 1.0076 - accuracy: 0.8983 - val_loss: 0.9719 - val_accuracy: 0.8418\nEpoch 12/100\n223/223 [==============================] - 59s 263ms/step - loss: 0.9974 - accuracy: 0.9008 - val_loss: 0.9509 - val_accuracy: 0.8550\nEpoch 13/100\n223/223 [==============================] - 61s 272ms/step - loss: 0.9832 - accuracy: 0.9051 - val_loss: 0.8867 - val_accuracy: 0.8814\nEpoch 14/100\n223/223 [==============================] - 59s 264ms/step - loss: 0.9726 - accuracy: 0.9065 - val_loss: 0.8885 - val_accuracy: 0.8795\nEpoch 15/100\n223/223 [==============================] - 59s 263ms/step - loss: 0.9601 - accuracy: 0.9070 - val_loss: 0.9353 - val_accuracy: 0.8493\nEpoch 16/100\n223/223 [==============================] - 59s 264ms/step - loss: 0.9507 - accuracy: 0.9127 - val_loss: 0.9540 - val_accuracy: 0.8343\nEpoch 17/100\n223/223 [==============================] - 59s 263ms/step - loss: 0.9411 - accuracy: 0.9048 - val_loss: 0.9232 - val_accuracy: 0.8493\nEpoch 18/100\n223/223 [==============================] - 60s 271ms/step - loss: 0.9325 - accuracy: 0.9141 - val_loss: 0.8835 - val_accuracy: 0.8701\nEpoch 19/100\n223/223 [==============================] - 60s 271ms/step - loss: 0.9230 - accuracy: 0.9076 - val_loss: 0.8766 - val_accuracy: 0.8588\nEpoch 20/100\n223/223 [==============================] - 59s 263ms/step - loss: 0.9181 - accuracy: 0.9164 - val_loss: 0.9479 - val_accuracy: 0.8437\nEpoch 21/100\n223/223 [==============================] - 60s 271ms/step - loss: 0.9116 - accuracy: 0.9117 - val_loss: 0.8581 - val_accuracy: 0.8719\nEpoch 22/100\n223/223 [==============================] - 59s 263ms/step - loss: 0.9048 - accuracy: 0.9154 - val_loss: 0.8893 - val_accuracy: 0.8531\nEpoch 23/100\n223/223 [==============================] - 59s 264ms/step - loss: 0.8981 - accuracy: 0.9186 - val_loss: 0.8962 - val_accuracy: 0.8663\nEpoch 24/100\n223/223 [==============================] - 60s 271ms/step - loss: 0.8896 - accuracy: 0.9171 - val_loss: 0.8181 - val_accuracy: 0.8908\nEpoch 25/100\n223/223 [==============================] - 59s 263ms/step - loss: 0.8847 - accuracy: 0.9212 - val_loss: 0.8368 - val_accuracy: 0.8738\nEpoch 26/100\n223/223 [==============================] - 59s 264ms/step - loss: 0.8797 - accuracy: 0.9200 - val_loss: 0.9619 - val_accuracy: 0.8324\n","output_type":"stream"}]},{"cell_type":"code","source":"model.save(\"models/model_cutmix\")","metadata":{"execution":{"iopub.status.busy":"2022-11-22T23:14:07.154950Z","iopub.execute_input":"2022-11-22T23:14:07.155379Z","iopub.status.idle":"2022-11-22T23:14:45.424873Z","shell.execute_reply.started":"2022-11-22T23:14:07.155344Z","shell.execute_reply":"2022-11-22T23:14:45.423819Z"},"trusted":true},"execution_count":26,"outputs":[]},{"cell_type":"code","source":"import shutil\nshutil.make_archive(\"model_cutmix\", 'zip', \"./models/model_cutmix\")","metadata":{"execution":{"iopub.status.busy":"2022-11-22T23:15:11.290055Z","iopub.execute_input":"2022-11-22T23:15:11.290475Z","iopub.status.idle":"2022-11-22T23:15:26.604164Z","shell.execute_reply.started":"2022-11-22T23:15:11.290440Z","shell.execute_reply":"2022-11-22T23:15:26.603074Z"},"trusted":true},"execution_count":27,"outputs":[{"execution_count":27,"output_type":"execute_result","data":{"text/plain":"'/kaggle/working/model_cutmix.zip'"},"metadata":{}}]},{"cell_type":"markdown","source":"# **Model assessment**","metadata":{"id":"TbdS6hN0OLph","papermill":{"duration":0.763892,"end_time":"2022-11-22T15:47:11.458066","exception":false,"start_time":"2022-11-22T15:47:10.694174","status":"completed"},"tags":[]}},{"cell_type":"code","source":"metrics = model.evaluate(validation_set, return_dict=True)\nprint(metrics)\n\nDATASET_EVAL = validation_set\nlabels = np.concatenate([y for x, y in DATASET_EVAL], axis=0) \npredictions = model.predict(validation_set) \n\nassert len(labels) == len(predictions)\n\n# Compute the confusion matrix\ncm = confusion_matrix(np.argmax(labels, axis=-1), np.argmax(predictions, axis=-1))\n\n# Compute the classification metrics\naccuracy = accuracy_score(np.argmax(labels, axis=-1), np.argmax(predictions, axis=-1))\nprecision = precision_score(np.argmax(labels, axis=-1), np.argmax(predictions, axis=-1), average='macro')\nrecall = recall_score(np.argmax(labels, axis=-1), np.argmax(predictions, axis=-1), average='macro')\nf1 = f1_score(np.argmax(labels, axis=-1), np.argmax(predictions, axis=-1), average='macro')\nf1_classes = f1_score(np.argmax(labels, axis=-1), np.argmax(predictions, axis=-1), average=None)\n\nprint('Accuracy:',accuracy.round(4))\nprint('Precision:',precision.round(4))\nprint('Recall:',recall.round(4))\nprint('F1:',f1.round(4))\nprint('F1 (per class):')\nprint(f1_classes)\n\nprint('WORST CLASS: Species:', np.argmin(f1_classes)+1)\n\n# Plot the confusion matrix\nig, ax = plt.subplots(figsize=(7.5, 7.5))\nax.matshow(cm, cmap=plt.cm.Blues, alpha=0.3)\nfor i in range(cm.shape[0]):\n    for j in range(cm.shape[1]):\n        ax.text(x=j, y=i,s=cm[i, j], va='center', ha='center', size='xx-large')\n\n        \nplt.xlabel('Predictions', fontsize=18)\nplt.ylabel('Actuals', fontsize=18)\nplt.title('Confusion Matrix', fontsize=18)\nplt.show()","metadata":{"id":"dDH2Jm_yDQcZ","papermill":{"duration":6.830463,"end_time":"2022-11-22T15:47:18.995696","exception":false,"start_time":"2022-11-22T15:47:12.165233","status":"completed"},"tags":[],"execution":{"iopub.status.busy":"2022-11-22T23:02:33.529594Z","iopub.execute_input":"2022-11-22T23:02:33.530511Z","iopub.status.idle":"2022-11-22T23:03:23.277603Z","shell.execute_reply.started":"2022-11-22T23:02:33.530474Z","shell.execute_reply":"2022-11-22T23:03:23.276523Z"},"trusted":true},"execution_count":25,"outputs":[{"name":"stdout","text":"17/17 [==============================] - 1s 85ms/step - loss: 0.8876 - accuracy: 0.8964\n{'loss': 0.887619137763977, 'accuracy': 0.8964218497276306}\n17/17 [==============================] - 3s 80ms/step\nAccuracy: 0.8964\nPrecision: 0.8951\nRecall: 0.8919\nF1: 0.8923\nF1 (per class):\n[0.80851064 0.85314685 0.93877551 0.90909091 0.89041096 0.95652174\n 0.93793103 0.84393064]\nWORST CLASS: Species: 1\n","output_type":"stream"},{"output_type":"display_data","data":{"text/plain":"<Figure size 540x540 with 1 Axes>","image/png":"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\n"},"metadata":{"needs_background":"light"}}]}]}