def initialize_data_from_leavesdb(dataset_name='PNAS', splits={'train':0.7,'validation':0.3}, threshold=50, exclude_classes=[], include_classes=[]): datasets = { 'PNAS': pnas_dataset.PNASDataset(), 'Leaves': leaves_dataset.LeavesDataset(), 'Fossil': fossil_dataset.FossilDataset() } data_files = datasets[dataset_name] data_files.exclude_rare_classes(threshold=threshold) encoder = base_dataset.LabelEncoder(data_files.classes) classes = list((set(encoder.classes)-set(exclude_classes)).union(set(include_classes))) data_files, excluded_data_files = data_files.enforce_class_whitelist(class_names=classes) x = list(data_files.data['path'].values) y = np.array(encoder.encode(data_files.data['family'])) shuffled_data = list(zip(x,y)) random.shuffle(shuffled_data) partitioned_data = partition_data(data=shuffled_data, partitions=OrderedDict(splits)) split_data = {k:v for k,v in partitioned_data.items() if len(v)>0} for subset, subset_data in split_data.items(): split_data[subset] = [list(i) for i in unzip(subset_data)] return split_data, data_files, excluded_data_files
def test_leaves_dataset(): dataset = leaves_dataset.LeavesDataset() assert dataset.name = "Leaves" assert dataset.num_samples == 26953 assert dataset.num_classes == 376 assert dataset._threshold == 0
def test_base_dataset(): # dataset = base_dataset.BaseDataset() fossil_data = fossil_dataset.FossilDataset() leaves_data = leaves_dataset.LeavesDataset() added_data = (leaves_data+fossil_data) for class_name in added_data.classes: include, exclude = added_data.leave_one_class_out(class_name) assert include.shape[0]+exclude.shape[0]==added_data.data.shape[0]
def load_data(dataset_name='PNAS', splits={ 'train': 0.7, 'validation': 0.3 }, threshold=50): datasets = { 'PNAS': pnas_dataset.PNASDataset(), 'Leaves': leaves_dataset.LeavesDataset(), 'Fossil': fossil_dataset.FossilDataset() } data_files = datasets[dataset_name] data_files.exclude_rare_classes(threshold=threshold) encoder = base_dataset.LabelEncoder(data_files.classes) data_files, _ = data_files.enforce_class_whitelist( class_names=encoder.classes) x = list(data_files.data['path'].values) y = np.array(encoder.encode(data_files.data['family'])) shuffled_data = list(zip(x, y)) random.shuffle(shuffled_data) partitioned_data = partition_data(data=shuffled_data, partitions=OrderedDict(splits)) split_data = {k: v for k, v in partitioned_data.items() if len(v) > 0} for subset, subset_data in split_data.items(): split_data[subset] = [list(i) for i in unzip(subset_data)] paths = tf.data.Dataset.from_tensor_slices(split_data['train'][0]) labels = tf.data.Dataset.from_tensor_slices(split_data['train'][1]) train_data = tf.data.Dataset.zip((paths, labels)) train_data = train_data.map( lambda x, y: (tf.image.convert_image_dtype(load_img(x) * 255.0, dtype=tf.uint8), y), num_parallel_calls=-1) paths = tf.data.Dataset.from_tensor_slices(split_data['validation'][0]) labels = tf.data.Dataset.from_tensor_slices(split_data['validation'][1]) validation_data = tf.data.Dataset.zip((paths, labels)) validation_data = validation_data.map( lambda x, y: (tf.image.convert_image_dtype(load_img(x) * 255.0, dtype=tf.uint8), y), num_parallel_calls=-1) return {'train': train_data, 'validation': validation_data}, data_files
def create_experiment__A_train_val_test(src_db=pyleaves.DATABASE_PATH, include_runs=['all']): start = time.time() experiment_view = select_by_col(table=tables['runs'], column='experiment_type', value='A_train_val_test') datasets = { 'PNAS': pnas_dataset.PNASDataset(src_db=src_db), 'Leaves': leaves_dataset.LeavesDataset(src_db=src_db), 'Fossil': fossil_dataset.FossilDataset(src_db=src_db) } logger = stuf({}) for i, run in experiment_view.iterrows(): run_config = stuf(run.to_dict()) if include_runs != ['all'] and (run_config['run_id'] not in include_runs): print(f"SKIPPING run with run_id={run_config['run_id']}") continue print('BEGINNING ', run) config = BaseConfig().parse(namespace=run_config) name = config.dataset_A logger[name] = stuf({}) tfrecords_table = TFRecordsTable( db_path=config.user_config.experiments_db) data_config_A = config.data_config.dataset_A num_shards = data_config_A.num_shards resolution = data_config_A.resolution csv_logger = CSVLogger(config, csv_dir=data_config_A.csv_dir) data = datasets[name] encoder = base_dataset.LabelEncoder(data.data.family) ensure_dir_exists(data_config_A.csv_dir) num_classes = data_config_A.num_classes = encoder.num_classes processed = base_dataset.preprocess_data(data, encoder, data_config_A) for subset in processed.keys(): Item = partial( TFRecordItem, **{ 'run_id': run_config['run_id'], 'experiment_type': run_config['experiment_type'], 'file_group': subset, 'dataset_stage': 'dataset_A', 'dataset_name': name, 'resolution': resolution, 'num_channels': 3, 'num_classes': num_classes, 'num_shards': num_shards }) random.shuffle(processed[subset]) x, y = [list(i) for i in unzip(processed[subset])] save_csv_data(x, y, filepath=data_config_A.csv_data[subset]) encoder.save_config(data_config_A.label_dir) class_counts = base_dataset.calculate_class_counts(y_data=y) class_weights = base_dataset.calculate_class_weights(y_data=y) class_counts_fpath = csv_logger.log_dict( class_counts, filepath=f'{subset}_class_counts.csv') class_weights_fpath = csv_logger.log_dict( class_weights, filepath=f'{subset}_class_weights.csv') print( f'logged class counts and weights to {class_counts_fpath} and {class_weights_fpath}, respectively' ) logger[name][subset] = stuf({'start': time.time()}) file_log = save_tfrecords(data=processed[subset], output_dir=data_config_A.tfrecord_dir, file_prefix=subset, target_size=(resolution, resolution), num_channels=3, num_classes=encoder.num_classes, num_shards=num_shards, TFRecordItem_factory=Item, tfrecords_table=tfrecords_table, verbose=True) logger[name][subset].end = time.time() logger[name][subset].total = logger[name][subset].end - logger[ name][subset].start print( name, subset, f'took {logger[name][subset].total:.2f} sec to collect/convert to TFRecords' ) print(f'Full experiment took a total of {time.time()-start}')
#CALLBACKS experiment_start_time = arrow.utcnow().format('YYYY-MM-DD_HH-mm-ss') log_dir =os.path.join(experiment_dir, experiment_name, 'log_dir',PARAMS['loss'], experiment_start_time) weights_best = os.path.join(log_dir, 'model_ckpt.h5') restore_best_weights=False histogram_freq=0 patience=25 num_epochs = PARAMS['num_epochs'] shuffle=True initial_epoch=0 src_db = pyleaves.DATABASE_PATH datasets = { 'PNAS': pnas_dataset.PNASDataset(src_db=src_db), 'Leaves': leaves_dataset.LeavesDataset(src_db=src_db), 'Fossil': fossil_dataset.FossilDataset(src_db=src_db) } data = datasets[PARAMS['dataset_name']] data_config = stuf(threshold=0, data_splits_meta={ 'train':PARAMS['train_size'], 'val':PARAMS['val_size'], 'test':PARAMS['test_size'] } ) preprocess_input = get_preprocessing_func(PARAMS['model_name']) preprocess_input(tf.zeros([4, 32, 32, 3]), tf.zeros([4, 32])) load_example = partial(_load_example, img_size=PARAMS['image_size'], num_classes=data.num_classes)
def main(**kwargs): import sys for k, v in kwargs.items(): sys.argv += [k, v] from pprint import pprint import argparse import datetime import json import os parser = argparse.ArgumentParser() parser.add_argument('--neptune_project_name', default='jacobarose/sandbox', type=str, help='Neptune.ai project name to log under') parser.add_argument('--experiment_name', default='pnas_minimal_example', type=str, help='Neptune.ai experiment name to log under') parser.add_argument('--config_path', default=r'/home/jacob/projects/pyleaves/pyleaves/configs/example_configs/pnas_resnet_config.json', type=str, help='JSON config file') parser.add_argument('-gpu', '--gpu_id', default='1', type=str, help='integer number of gpu to train on', dest='gpu_id') parser.add_argument('-tags', '--add-tags', default=[], type=str, nargs='*', help='Add arbitrary list of tags to apply to this run in neptune', dest='tags') parser.add_argument('-f', default=None) args = parser.parse_args() with open(args.config_path, 'r') as config_file: PARAMS = json.load(config_file) # print(gpu) # os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu_id) pprint(PARAMS) import tensorflow as tf import neptune # tf.debugging.set_log_device_placement(True) print(tf.__version__) import arrow import numpy as np import pandas as pd import seaborn as sns import matplotlib.pyplot as plt import io from stuf import stuf from more_itertools import unzip from functools import partial # import tensorflow as tf # tf.compat.v1.enable_eager_execution() AUTOTUNE = tf.data.experimental.AUTOTUNE from pyleaves.leavesdb.tf_utils.tf_utils import set_random_seed, reset_keras_session import pyleaves from pyleaves.utils.img_utils import random_pad_image from pyleaves.utils.utils import ensure_dir_exists from pyleaves.datasets import leaves_dataset, fossil_dataset, pnas_dataset, base_dataset from pyleaves.models.vgg16 import VGG16, VGG16GrayScale from pyleaves.models import resnet, vgg16 from tensorflow.compat.v1.keras.callbacks import Callback, ModelCheckpoint, TensorBoard, LearningRateScheduler, EarlyStopping from tensorflow.keras import metrics from tensorflow.keras.preprocessing.image import load_img, img_to_array from tensorflow.keras import layers from tensorflow.keras import backend as K import tensorflow_datasets as tfds import neptune_tensorboard as neptune_tb seed = 346 # set_random_seed(seed) # reset_keras_session() def get_preprocessing_func(model_name): if model_name.startswith('resnet'): from tensorflow.keras.applications.resnet_v2 import preprocess_input elif model_name == 'vgg16': from tensorflow.keras.applications.vgg16 import preprocess_input elif model_name=='shallow': def preprocess_input(x): return x/255.0 # ((x/255.0)-0.5)*2.0 return preprocess_input #lambda x,y: (preprocess_input(x),y) def _load_img(image_path):#, img_size=(224,224)): img = tf.io.read_file(image_path) img = tf.image.decode_jpeg(img, channels=3) img = tf.image.convert_image_dtype(img, tf.float32) return img # return tf.compat.v1.image.resize_image_with_pad(img, *img_size) def _encode_label(label, num_classes=19): label = tf.cast(label, tf.int32) label = tf.one_hot(label, depth=num_classes) return label def _load_example(image_path, label, num_classes=19): img = _load_img(image_path) one_hot_label = _encode_label(label, num_classes=num_classes) return img, one_hot_label def _load_uint8_example(image_path, label, num_classes=19): img = tf.image.convert_image_dtype(_load_img(image_path)*255.0, dtype=tf.uint8) one_hot_label = _encode_label(label, num_classes=num_classes) return img, one_hot_label def rgb2gray_3channel(img, label): ''' Convert rgb image to grayscale, but keep num_channels=3 ''' img = tf.image.rgb_to_grayscale(img) img = tf.image.grayscale_to_rgb(img) return img, label def rgb2gray_1channel(img, label): ''' Convert rgb image to grayscale, num_channels from 3 to 1 ''' img = tf.image.rgb_to_grayscale(img) return img, label def log_data(logs): for k, v in logs.items(): neptune.log_metric(k, v) neptune_logger = tf.keras.callbacks.LambdaCallback(on_epoch_end=lambda epoch, logs: log_data(logs)) def focal_loss(gamma=2.0, alpha=4.0): gamma = float(gamma) alpha = float(alpha) def focal_loss_fixed(y_true, y_pred): """Focal loss for multi-classification FL(p_t)=-alpha(1-p_t)^{gamma}ln(p_t) Notice: y_pred is probability after softmax gradient is d(Fl)/d(p_t) not d(Fl)/d(x) as described in paper d(Fl)/d(p_t) * [p_t(1-p_t)] = d(Fl)/d(x) Focal Loss for Dense Object Detection https://arxiv.org/abs/1708.02002 Arguments: y_true {tensor} -- ground truth labels, shape of [batch_size, num_cls] y_pred {tensor} -- model's output, shape of [batch_size, num_cls] Keyword Arguments: gamma {float} -- (default: {2.0}) alpha {float} -- (default: {4.0}) Returns: [tensor] -- loss. """ epsilon = 1.e-9 y_true = tf.convert_to_tensor(y_true, tf.float32) y_pred = tf.convert_to_tensor(y_pred, tf.float32) model_out = tf.add(y_pred, epsilon) ce = tf.multiply(y_true, -tf.log(model_out)) weight = tf.multiply(y_true, tf.pow(tf.subtract(1., model_out), gamma)) fl = tf.multiply(alpha, tf.multiply(weight, ce)) reduced_fl = tf.reduce_max(fl, axis=1) return tf.reduce_mean(reduced_fl) return focal_loss_fixed def per_class_accuracy(y_true, y_pred): return tf.metrics.mean_per_class_accuracy(y_true, y_pred, num_classes=PARAMS['num_classes']) def build_model(model_params, optimizer, loss, METRICS): if model_params['name']=='vgg16': model_builder = vgg16.VGG16GrayScale(model_params) elif model_params['name'].startswith('resnet'): model_builder = resnet.ResNet(model_params) base = model_builder.build_base() model = model_builder.build_head(base) model.compile(optimizer=optimizer, loss=loss, metrics=METRICS) return model def build_shallow(input_shape=(224,224,3), num_classes=10, optimizer=None, loss=None, METRICS=None): model = tf.keras.models.Sequential() model.add(layers.Conv2D(64, (7, 7), activation='relu', input_shape=input_shape, kernel_initializer=tf.initializers.GlorotNormal())) model.add(layers.MaxPooling2D((2, 2))) model.add(layers.Conv2D(64, (7, 7), activation='relu', kernel_initializer=tf.initializers.GlorotNormal())) model.add(layers.MaxPooling2D((2, 2))) model.add(layers.Conv2D(64, (7, 7), activation='relu', kernel_initializer=tf.initializers.GlorotNormal())) model.add(layers.Flatten()) model.add(layers.Dense(64*2, activation='relu', kernel_initializer=tf.initializers.GlorotNormal())) model.add(layers.Dense(num_classes,activation='softmax', kernel_initializer=tf.initializers.GlorotNormal())) model.compile(optimizer=optimizer, loss=loss, metrics=METRICS) return model class ImageLogger: '''Tensorflow 2.0 version''' def __init__(self, log_dir: str, max_images: int, name: str): self.file_writer = tf.summary.create_file_writer(log_dir) self.log_dir = log_dir self.max_images = max_images self.name = name self._counter = tf.Variable(0, dtype=tf.int64) self.filepaths = [] def add_log(self, img, counter=None, name=None): ''' Intention is to generalize this to an abstract class for logging to any experiment management platform (e.g. neptune, mlflow, etc) Currently takes a filepath pointing to an image file and logs to current neptune experiment. ''' # scaled_images = (img - tf.math.reduce_min(img))/(tf.math.reduce_max(img) - tf.math.reduce_min(img)) # keep = 0 # scaled_images = tf.image.convert_image_dtype(tf.squeeze(scaled_images[keep,:,:,:]), dtype=tf.uint8) # scaled_images = tf.expand_dims(scaled_images, 0) # tf.summary.image(name=self.name, data=scaled_images, step=self._counter, max_outputs=self.max_images) scaled_img = (img - np.min(img))/(np.max(img) - np.min(img)) * 255.0 scaled_img = scaled_img.astype(np.uint32) neptune.log_image(log_name= name or self.name, x=counter, y=scaled_img) return scaled_img def __call__(self, images, labels): with self.file_writer.as_default(): scaled_images = (images - tf.math.reduce_min(images))/(tf.math.reduce_max(images) - tf.math.reduce_min(images)) keep = 0 scaled_images = tf.image.convert_image_dtype(tf.squeeze(scaled_images[keep,:,:,:]), dtype=tf.uint8) scaled_images = tf.expand_dims(scaled_images, 0) labels = tf.argmax(labels[[keep], :],axis=1) tf.summary.image(name=self.name, data=scaled_images, step=self._counter, max_outputs=self.max_images) filepath = os.path.join(self.log_dir,'sample_images',f'{self.name}-{self._counter}.jpg') scaled_images = tf.image.encode_jpeg(tf.squeeze(scaled_images)) tf.io.write_file(filename=tf.constant(filepath), contents=scaled_images) # self.add_log(scaled_images) self._counter.assign_add(1) return images, labels def _cond_apply(x, y, func, prob): """Conditionally apply func to x and y with probability prob. Parameters ---------- x : type Input to conditionally pass through func y : type Label func : type Function to conditionally be applied to x and y prob : type Probability of applying function, within range [0.0,1.0] Returns ------- x, y """ return tf.cond((tf.random.uniform([], 0, 1) >= (1.0 - prob)), lambda: func(x,y), lambda: (x,y)) class ImageAugmentor: """Short summary. Parameters ---------- augmentations : dict Maps a sequence of named augmentations to a scalar probability, according to which they'll be conditionally applied in order. resize_w_pad : tuple, default=None Description of parameter `resize_w_pad`. random_crop : tuple, default=None Description of parameter `random_crop`. random_jitter : dict First applies resize_w_pad, then random_crop. If user desires only 1 of these, set this to None. Should be a dict with 2 keys: 'resize':(height, width) 'crop_size':(crop_height,crop_width, channels) Only 1 of these 3 kwargs should be provided to any given augmentor: {'resize_w_pad', 'random_crop', 'random_jitter'} Example values for each: resize_w_pad=(224,224) random_crop=(224,224,3) random_jitter={'resize':(338,338), 'crop_size':(224,224, 3)} seed : int, default=None Random seed to apply to all augmentations Examples ------- Examples should be written in doctest format, and should illustrate how to use the function/class. >>> Attributes ---------- augmentations """ def __init__(self, name='', augmentations={'rotate':1.0, 'flip':1.0, 'color':1.0, 'rgb2gray_3channel':1.0}, resize_w_pad=None, random_crop=None, random_jitter={'resize':(338,338), 'crop_size':(224,224,3)}, log_dir=None, seed=None): self.name = name self.augmentations = augmentations self.seed = seed if resize_w_pad: self.target_h = resize_w_pad[0] self.target_w = resize_w_pad[1] # self.resize = self.resize_w_pad elif random_crop: self.crop_size = random_crop self.target_h = self.crop_size[0] self.target_w = self.crop_size[1] # self.resize = self.random_crop elif random_jitter: # self.target_h = tf.random.uniform([], random_jitter['crop_size'][0], random_jitter['resize'][0], dtype=tf.int32, seed=self.seed) # self.target_w = tf.random.uniform([], random_jitter['crop_size'][1], random_jitter['resize'][1], dtype=tf.int32, seed=self.seed) self.crop_size = random_jitter['crop_size'] # self.resize = self.random_jitter self.target_h = random_jitter['crop_size'][0] self.target_w = random_jitter['crop_size'][1] self.resize = self.resize_w_pad self.maps = {'rotate':self.rotate, 'flip':self.flip, 'color':self.color, 'rgb2gray_3channel':self.rgb2gray_3channel, 'rgb2gray_1channel':self.rgb2gray_1channel} self.log_dir = log_dir def rotate(self, x: tf.Tensor, label: tf.Tensor) -> tf.Tensor: """Rotation augmentation Args: x, tf.Tensor: Image label, tf.Tensor: arbitrary tensor, passes through unchanged Returns: Augmented image, label """ # Rotate 0, 90, 180, 270 degrees return tf.image.rot90(x, tf.random.uniform(shape=[], minval=0, maxval=4, dtype=tf.int32,seed=self.seed)), label def flip(self, x: tf.Tensor, label: tf.Tensor) -> tf.Tensor: """Flip augmentation Args: x, tf.Tensor: Image to flip label, tf.Tensor: arbitrary tensor, passes through unchanged Returns: Augmented image, label """ x = tf.image.random_flip_left_right(x, seed=self.seed) x = tf.image.random_flip_up_down(x, seed=self.seed) return x, label def color(self, x: tf.Tensor, label: tf.Tensor) -> tf.Tensor: """Color augmentation Args: x, tf.Tensor: Image label, tf.Tensor: arbitrary tensor, passes through unchanged Returns: Augmented image, label """ x = tf.image.random_hue(x, 0.08, seed=self.seed) x = tf.image.random_saturation(x, 0.6, 1.6, seed=self.seed) x = tf.image.random_brightness(x, 0.05, seed=self.seed) x = tf.image.random_contrast(x, 0.7, 1.3, seed=self.seed) return x, label def rgb2gray_3channel(self, x: tf.Tensor, label: tf.Tensor) -> tf.Tensor: """Convert RGB image -> grayscale image, maintain number of channels = 3 Args: x, tf.Tensor: Image label, tf.Tensor: arbitrary tensor, passes through unchanged Returns: Augmented image, label """ x = tf.image.rgb_to_grayscale(x) x = tf.image.grayscale_to_rgb(x) return x, label def rgb2gray_1channel(self, x: tf.Tensor, label: tf.Tensor) -> tf.Tensor: """Convert RGB image -> grayscale image, reduce number of channels from 3 -> 1 Args: x, tf.Tensor: Image label, tf.Tensor: arbitrary tensor, passes through unchanged Returns: Augmented image, label """ x = tf.image.rgb_to_grayscale(x) return x, label def resize_w_pad(self, x: tf.Tensor, label: tf.Tensor) -> tf.Tensor: # TODO Finish this # random_pad_image(x,min_image_size=None,max_image_size=None,pad_color=None,seed=self.seed) return tf.image.resize_with_pad(x, target_height=self.target_h, target_width=self.target_w), label def random_crop(self, x: tf.Tensor, label: tf.Tensor) -> tf.Tensor: return tf.image.random_crop(x, size=self.crop_size), label @tf.function def random_jitter(self, x: tf.Tensor, label: tf.Tensor) -> tf.Tensor: x, label = self.resize_w_pad(x, label) x, label = self.random_crop(x, label) return x, label def apply_augmentations(self, dataset: tf.data.Dataset): """ Call this function to apply all of the augmentation in the order of specification provided to the constructor __init__() of ImageAugmentor. Args: dataset, tf.data.Dataset: must yield individual examples of form (x, y) Returns: Augmented dataset """ dataset = dataset.map(self.resize, num_parallel_calls=AUTOTUNE) for aug_name, aug_p in self.augmentations.items(): aug = self.maps[aug_name] dataset = dataset.map(lambda x,y: _cond_apply(x, y, aug, prob=aug_p), num_parallel_calls=AUTOTUNE) # dataset = dataset.map(lambda x,y: _cond_apply(x, y, func=aug, prob=aug_p), num_parallel_calls=AUTOTUNE) return dataset class ImageLoggerCallback(Callback): '''Tensorflow 2.0 version Callback that keeps track of a tf.data.Dataset and logs the correct batch to neptune based on the current batch. ''' def __init__(self, data :tf.data.Dataset, freq=1, max_images=-1, name='', encoder=None): self.data = data self.freq = freq self.max_images = max_images self.name = name self.encoder=encoder self.init_iterator() def init_iterator(self): self.data_iter = iter(self.data) self._batch = 0 self._count = 0 self.finished = False def yield_batch(self): batch_data = next(self.data_iter) self._batch += 1 self._count += batch_data[0].shape[0] return batch_data def add_log(self, img, counter=None, name=None): ''' Intention is to generalize this to an abstract class for logging to any experiment management platform (e.g. neptune, mlflow, etc) Currently takes a filepath pointing to an image file and logs to current neptune experiment. ''' scaled_img = (img - np.min(img))/(np.max(img) - np.min(img)) * 255.0 scaled_img = scaled_img.astype(np.uint32) neptune.log_image(log_name= name or self.name, x=counter, y=scaled_img) return scaled_img def on_train_batch_begin(self, batch, logs=None): if batch % self.freq or self.finished: return while batch >= self._batch: x, y = self.yield_batch() if self.max_images==-1: self.max_images=x.shape[0] if x.ndim==3: np.newaxis(x, axis=0) if x.shape[0]>self.max_images: x = x[:self.max_images,...] y = y[:self.max_images,...] x = x.numpy() y = np.argmax(y.numpy(),axis=1) if self.encoder: y = self.encoder.decode(y) for i in range(x.shape[0]): # self.add_log(x[i,...], counter=i, name = f'{self.name}-{y[i]}-batch_{str(self._batch).zfill(3)}') self.add_log(x[i,...], counter=self._count+i, name = f'{self.name}-{y[i]}') print(f'Batch {self._batch}: Logged {np.max([x.shape[0],self.max_images])} {self.name} images to neptune') def on_epoch_end(self, epoch, logs={}): self.finished = True class ConfusionMatrixCallback(Callback): '''Tensorflow 2.0 version''' def __init__(self, log_dir, imgs : dict, labels : dict, classes, freq=1, include_train=False, seed=None): self.file_writer = tf.summary.create_file_writer(log_dir) self.log_dir = log_dir self.seed = seed self._counter = 0 assert np.all(np.array(imgs.keys()) == np.array(labels.keys())) self.imgs = imgs for k,v in labels.items(): if v.ndim==2: labels[k] = tf.argmax(v,axis=-1) self.labels = labels self.num_samples = {k:l.numpy().shape[0] for k,l in labels.items()} self.classes = classes self.freq = freq self.include_train = include_train def log_confusion_matrix(self, model, imgs, labels, epoch, name='', norm_cm=False): pred_labels = model.predict_classes(imgs) # pred_labels = tf.argmax(pred_labels,axis=-1) pred_labels = pred_labels[:,None] con_mat = tf.math.confusion_matrix(labels=labels, predictions=pred_labels, num_classes=len(self.classes)).numpy() if norm_cm: con_mat = np.around(con_mat.astype('float') / con_mat.sum(axis=1)[:, np.newaxis], decimals=2) con_mat_df = pd.DataFrame(con_mat, index = self.classes, columns = self.classes) figure = plt.figure(figsize=(12, 12)) sns.heatmap(con_mat_df, annot=True, cmap=plt.cm.Blues) plt.tight_layout() plt.ylabel('True label') plt.xlabel('Predicted label') buf = io.BytesIO() plt.savefig(buf, format='png') buf.seek(0) image = tf.image.decode_png(buf.getvalue(), channels=4) image = tf.expand_dims(image, 0) with self.file_writer.as_default(): tf.summary.image(name=name+'_confusion_matrix', data=image, step=self._counter) neptune.log_image(log_name=name+'_confusion_matrix', x=self._counter, y=figure) plt.close(figure) self._counter += 1 return image def on_epoch_end(self, epoch, logs={}): if (not self.freq) or (epoch%self.freq != 0): return if self.include_train: cm_summary_image = self.log_confusion_matrix(self.model, self.imgs['train'], self.labels['train'], epoch=epoch, name='train') cm_summary_image = self.log_confusion_matrix(self.model, self.imgs['val'], self.labels['val'], epoch=epoch, name='val') #################################################################################### #################################################################################### #################################################################################### neptune.init(project_qualified_name=args.neptune_project_name) # neptune_tb.integrate_with_tensorflow() experiment_dir = '/media/data/jacob/sandbox_logs' experiment_name = args.experiment_name experiment_start_time = arrow.utcnow().format('YYYY-MM-DD_HH-mm-ss') log_dir =os.path.join(experiment_dir, experiment_name, 'log_dir',PARAMS['loss'], experiment_start_time) ensure_dir_exists(log_dir) print('Tensorboard log_dir: ', log_dir) # os.system(f'neptune tensorboard {log_dir} --project {args.neptune_project_name}') weights_best = os.path.join(log_dir, 'model_ckpt.h5') restore_best_weights=False histogram_freq=0 patience=25 num_epochs = PARAMS['num_epochs'] initial_epoch=0 src_db = pyleaves.DATABASE_PATH datasets = { 'PNAS': pnas_dataset.PNASDataset(src_db=src_db), 'Leaves': leaves_dataset.LeavesDataset(src_db=src_db), 'Fossil': fossil_dataset.FossilDataset(src_db=src_db) } # data = datasets[PARAMS['dataset_name']] data_config = stuf(threshold=PARAMS['data_threshold'], num_classes=PARAMS['num_classes'] , data_splits_meta={ 'train':PARAMS['train_size'], 'val':PARAMS['val_size'], 'test':PARAMS['test_size'] } ) preprocess_input = get_preprocessing_func(PARAMS['model_name']) preprocess_input(tf.zeros([4, 224, 224, 3])) load_example = partial(_load_uint8_example, num_classes=data_config.num_classes) # load_example = partial(_load_example, num_classes=data_config.num_classes) if PARAMS['num_channels']==3: color_aug = {'rgb2gray_3channel':1.0} elif PARAMS['num_channels']==1: color_aug = {'rgb2gray_1channel':1.0} resize_w_pad=None random_jitter=None if not PARAMS['random_jitter']['resize']: resize_w_pad = PARAMS['image_size'] else: random_jitter=PARAMS['random_jitter'] TRAIN_image_augmentor = ImageAugmentor(name='train', augmentations={**PARAMS["augmentations"], **color_aug},#'rotate':1.0,'flip':1.0,**color_aug}, resize_w_pad=resize_w_pad, random_crop=None, random_jitter=random_jitter, log_dir=log_dir, seed=None) VAL_image_augmentor = ImageAugmentor(name='val', augmentations={**color_aug}, resize_w_pad=PARAMS['image_size'], random_crop=None, random_jitter=None, log_dir=log_dir, seed=None) TEST_image_augmentor = ImageAugmentor(name='test', augmentations={**color_aug}, resize_w_pad=PARAMS['image_size'], random_crop=None, random_jitter=None, log_dir=log_dir, seed=None) def neptune_log_augmented_images(split_data, num_demo_samples=40, PARAMS=PARAMS): num_demo_samples = 40 cm_data_x = {'train':[],'val':[]} cm_data_y = {'train':[],'val':[]} cm_data_x['train'], cm_data_y['train'] = next(iter(get_data_loader(data=split_data['train'], data_subset_mode='train', batch_size=num_demo_samples, infinite=True, augment=False,seed=2836))) cm_data_x['val'], cm_data_y['val'] = next(iter(get_data_loader(data=split_data['val'], data_subset_mode='val', batch_size=num_demo_samples, infinite=True, augment=False, seed=2836))) for (k_x,v_x), (k_y, v_y) in zip(cm_data_x.items(), cm_data_y.items()): x = tf.data.Dataset.from_tensor_slices(v_x) y = tf.data.Dataset.from_tensor_slices(v_y) xy_data = tf.data.Dataset.zip((x, y)) v = xy_data.map(VAL_image_augmentor.resize, num_parallel_calls=AUTOTUNE) v_aug = TRAIN_image_augmentor.apply_augmentations(xy_data) v_x, v_y = [i.numpy() for i in next(iter(v.batch(10*num_demo_samples)))] v_x_aug, v_y_aug = [i.numpy() for i in next(iter(v_aug.batch(10*num_demo_samples)))] k = k_x for i in range(num_demo_samples): print(f'Neptune: logging {k}_{i}') print(f'{v_x[i].shape}, {v_x_aug[i].shape}') idx = np.random.randint(0,len(v_x)) if True: #'train' in k: TRAIN_image_augmentor.logger.add_log(v_x[idx],counter=i, name=k) TRAIN_image_augmentor.logger.add_log(v_x_aug[idx],counter=i, name=k+'_aug') def get_data_loader(data : tuple, data_subset_mode='train', batch_size=32, num_classes=None, infinite=True, augment=True, seed=2836): num_samples = len(data[0]) x = tf.data.Dataset.from_tensor_slices(data[0]) labels = tf.data.Dataset.from_tensor_slices(data[1]) data = tf.data.Dataset.zip((x, labels)) data = data.cache() if data_subset_mode == 'train': data = data.shuffle(buffer_size=num_samples) # data = data.map(lambda x,y: (tf.image.convert_image_dtype(load_img(x)*255.0,dtype=tf.uint8),y), num_parallel_calls=-1) # data = data.map(load_example, num_parallel_calls=AUTOTUNE) data = data.map(load_example, num_parallel_calls=AUTOTUNE) data = data.map(lambda x,y: (preprocess_input(x), y), num_parallel_calls=AUTOTUNE) if infinite: data = data.repeat() if data_subset_mode == 'train': data = data.shuffle(buffer_size=200, seed=seed) augmentor = TRAIN_image_augmentor elif data_subset_mode == 'val': augmentor = VAL_image_augmentor elif data_subset_mode == 'test': augmentor = TEST_image_augmentor if augment: data = augmentor.apply_augmentations(data) data = data.batch(batch_size, drop_remainder=True) return data.prefetch(AUTOTUNE) def get_tfds_data_loader(data : tf.data.Dataset, data_subset_mode='train', batch_size=32, num_samples=100, num_classes=19, infinite=True, augment=True, seed=2836): def encode_example(x, y): x = tf.image.convert_image_dtype(x, tf.float32) * 255.0 y = _encode_label(y, num_classes=num_classes) return x, y test_d = next(iter(data)) print(test_d[0].numpy().min()) print(test_d[0].numpy().max()) data = data.shuffle(buffer_size=num_samples) \ .cache() \ .map(encode_example, num_parallel_calls=AUTOTUNE) test_d = next(iter(data)) print(test_d[0].numpy().min()) print(test_d[0].numpy().max()) data = data.map(preprocess_input, num_parallel_calls=AUTOTUNE) test_d = next(iter(data)) print(test_d[0].numpy().min()) print(test_d[0].numpy().max()) if data_subset_mode == 'train': data = data.shuffle(buffer_size=100, seed=seed) augmentor = TRAIN_image_augmentor elif data_subset_mode == 'val': augmentor = VAL_image_augmentor elif data_subset_mode == 'test': augmentor = TEST_image_augmentor if augment: data = augmentor.apply_augmentations(data) test_d = next(iter(data)) print(test_d[0].numpy().min()) print(test_d[0].numpy().max()) data = data.batch(batch_size, drop_remainder=True) if infinite: data = data.repeat() return data.prefetch(AUTOTUNE) # y_true = [[0, 1, 0], [0, 0, 1]] # y_pred = [[0.05, 0.95, 0], [0.1, 0.8, 0.1]] def accuracy(y_true, y_pred): y_pred = tf.argmax(y_pred, axis=-1) y_true = tf.argmax(y_true, axis=-1) return tf.reduce_mean(tf.cast(tf.equal(y_true, y_pred), tf.float32)) def true_pos(y_true, y_pred): # y_true = K.ones_like(y_true) return K.sum(K.round(K.clip(y_true * y_pred, 0, 1))) def false_pos(y_true, y_pred): # y_true = K.ones_like(y_true) true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1))) all_positives = K.sum(K.round(K.clip(y_true, 0, 1))) return all_positives - true_positives def true_neg(y_true, y_pred): # y_true = K.ones_like(y_true) return K.sum(1-K.round(K.clip(y_true * y_pred, 0, 1))) def recall(y_true, y_pred): # y_true = K.ones_like(y_true) true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1))) all_positives = K.sum(K.round(K.clip(y_true, 0, 1))) recall = true_positives / (all_positives + K.epsilon()) return recall def precision(y_true, y_pred): y_true = K.ones_like(y_true) true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1))) predicted_positives = K.sum(K.round(K.clip(y_pred, 0, 1))) precision = true_positives / (predicted_positives + K.epsilon()) # tf.print(y_true, y_pred) return precision def f1_score(y_true, y_pred): m_precision = precision(y_true, y_pred) m_recall = recall(y_true, y_pred) # pdb.set_trace() return 2*((m_precision*m_recall)/(m_precision+m_recall+K.epsilon())) # def false_neg(y_true, y_pred): # y_true = K.ones_like(~y_true) # true_neg = K.sum(K.round(K.clip(y_true * y_pred, 0, 1))) # all_negative = K.sum(K.round(K.clip(y_true, 0, 1))) # return all_negatives - true_ # return K.mean(K.argmax(y_true,axis=1)*K.argmax(y_pred,axis=1)) # 'accuracy', # metrics.TrueNegatives(name='tn'), # metrics.FalseNegatives(name='fn'), METRICS = [ f1_score, metrics.TruePositives(name='tp'), metrics.FalsePositives(name='fp'), metrics.CategoricalAccuracy(name='accuracy'), metrics.TopKCategoricalAccuracy(name='top_3_categorical_accuracy', k=3), metrics.TopKCategoricalAccuracy(name='top_5_categorical_accuracy', k=5) ] PARAMS['sys.argv'] = ' '.join(sys.argv) with neptune.create_experiment(name=experiment_name, params=PARAMS, upload_source_files=[__file__]): print('Logging experiment tags:') for tag in args.tags: print(tag) neptune.append_tag(tag) neptune.append_tag(PARAMS['dataset_name']) neptune.append_tag(PARAMS['model_name']) neptune.log_artifact(args.config_path) cm_data_x = {'train':[],'val':[]} cm_data_y = {'train':[],'val':[]} if PARAMS['dataset_name'] in tfds.list_builders(): num_demo_samples=40 tfds_builder = tfds.builder(PARAMS['dataset_name']) tfds_builder.download_and_prepare() num_samples = tfds_builder.info.splits['train'].num_examples num_samples_dict = {'train':int(num_samples*PARAMS['train_size']), 'val':int(num_samples*PARAMS['val_size']), 'test':int(num_samples*PARAMS['test_size'])} classes = tfds_builder.info.features['label'].names num_classes = len(classes) train_slice = [0,int(PARAMS['train_size']*100)] val_slice = [int(PARAMS['train_size']*100), int((PARAMS['train_size']+PARAMS['val_size'])*100)] test_slice = [100 - int(PARAMS['test_size']*100), 100] tfds_train_data = tfds.load(PARAMS['dataset_name'], split=f"train[{train_slice[0]}%:{train_slice[1]}%]", shuffle_files=True, as_supervised=True) tfds_validation_data = tfds.load(PARAMS['dataset_name'], split=f"train[{val_slice[0]}%:{val_slice[1]}%]", shuffle_files=True, as_supervised=True) tfds_test_data = tfds.load(PARAMS['dataset_name'], split=f"train[{test_slice[0]}%:{test_slice[1]}%]", shuffle_files=True, as_supervised=True) # PARAMS['batch_size']=1 train_data = get_tfds_data_loader(data = tfds_train_data, data_subset_mode='train', batch_size=PARAMS['batch_size'], num_samples=num_samples_dict['train'], num_classes=num_classes, infinite=True, augment=True, seed=2836) validation_data = get_tfds_data_loader(data = tfds_validation_data, data_subset_mode='val', batch_size=PARAMS['batch_size'], num_samples=num_samples_dict['val'], num_classes=num_classes, infinite=True, augment=True, seed=2837) test_data = get_tfds_data_loader(data = tfds_test_data, data_subset_mode='test', batch_size=PARAMS['batch_size'], num_samples=num_samples_dict['test'], num_classes=num_classes, infinite=True, augment=True, seed=2838) # tfds_train_data = tfds.load(PARAMS['dataset_name'], split=f"train[{train_slice[0]}%:{train_slice[1]}%]", shuffle_files=True, as_supervised=True) # tfds_validation_data = tfds.load(PARAMS['dataset_name'], split=f"train[{val_slice[0]}%:{val_slice[1]}%]", shuffle_files=True, as_supervised=True) # tfds_test_data = tfds.load(PARAMS['dataset_name'], split=f"train[{test_slice[0]}%:{test_slice[1]}%]", shuffle_files=True, as_supervised=True) split_data = {'train':get_tfds_data_loader(data = tfds_train_data, data_subset_mode='train', batch_size=num_demo_samples, num_samples=num_samples_dict['train'], num_classes=num_classes, infinite=True, augment=True, seed=2836), 'val':get_tfds_data_loader(data = tfds_validation_data, data_subset_mode='val', batch_size=num_demo_samples, num_samples=num_samples_dict['val'], num_classes=num_classes, infinite=True, augment=True, seed=2837), 'test':get_tfds_data_loader(data = tfds_test_data, data_subset_mode='test', batch_size=num_demo_samples, num_samples=num_samples_dict['test'], num_classes=num_classes, infinite=True, augment=True, seed=2838) } steps_per_epoch=num_samples_dict['train']//PARAMS['batch_size'] validation_steps=num_samples_dict['val']//PARAMS['batch_size'] cm_data_x['train'], cm_data_y['train'] = next(iter(split_data['train'])) cm_data_x['val'], cm_data_y['val'] = next(iter(split_data['val'])) else: data = datasets[PARAMS['dataset_name']] neptune.set_property('num_classes',data.num_classes) neptune.set_property('class_distribution',data.metadata.class_distribution) encoder = base_dataset.LabelEncoder(data.data.family) split_data = base_dataset.preprocess_data(data, encoder, data_config) # import pdb;pdb.set_trace() for subset, subset_data in split_data.items(): split_data[subset] = [list(i) for i in unzip(subset_data)] PARAMS['batch_size'] = 32 steps_per_epoch=len(split_data['train'][0])//PARAMS['batch_size']#//10 validation_steps=len(split_data['val'][0])//PARAMS['batch_size']#//10 split_datasets = { k:base_dataset.BaseDataset.from_dataframe( pd.DataFrame({ 'path':v[0], 'family':v[1] })) \ for k,v in split_data.items() } for k,v in split_datasets.items(): print(k, v.num_classes) classes = split_datasets['train'].classes train_data=get_data_loader(data=split_data['train'], data_subset_mode='train', batch_size=PARAMS['batch_size'], infinite=True, augment=True, seed=2836) validation_data=get_data_loader(data=split_data['val'], data_subset_mode='val', batch_size=PARAMS['batch_size'], infinite=True, augment=True, seed=2837) if 'test' in split_data.keys(): test_data=get_data_loader(data=split_data['test'], data_subset_mode='test', batch_size=PARAMS['batch_size'], infinite=True, augment=True, seed=2838) num_demo_samples=150 # neptune_log_augmented_images(split_data, num_demo_samples=num_demo_samples, PARAMS=PARAMS) cm_data_x['train'], cm_data_y['train'] = next(iter(get_data_loader(data=split_data['train'], data_subset_mode='train', batch_size=num_demo_samples, infinite=True, augment=True, seed=2836))) cm_data_x['val'], cm_data_y['val'] = next(iter(get_data_loader(data=split_data['val'], data_subset_mode='val', batch_size=num_demo_samples, infinite=True, augment=True, seed=2836))) ######################################################################################## train_image_logger_cb = ImageLoggerCallback(data=train_data, freq=20, max_images=-1, name='train', encoder=encoder) val_image_logger_cb = ImageLoggerCallback(data=validation_data, freq=20, max_images=-1, name='val', encoder=encoder) ######################################################################################## cm_callback = ConfusionMatrixCallback(log_dir, cm_data_x, cm_data_y, classes=classes, seed=PARAMS['seed'], include_train=True) checkpoint = ModelCheckpoint(weights_best, monitor='val_loss', verbose=0, save_best_only=True, save_weights_only=False, mode='min',restore_best_weights=restore_best_weights) tfboard = TensorBoard(log_dir=log_dir, histogram_freq=histogram_freq, write_images=True) early = EarlyStopping(monitor='val_loss', patience=patience, verbose=1) callbacks = [checkpoint,tfboard,early, cm_callback, neptune_logger, train_image_logger_cb, val_image_logger_cb] ########################## if PARAMS['optimizer'] == 'Adam': optimizer = tf.keras.optimizers.Adam( learning_rate=PARAMS['lr'] ) elif PARAMS['optimizer'] == 'Nadam': optimizer = tf.keras.optimizers.Nadam( learning_rate=PARAMS['lr'] ) elif PARAMS['optimizer'] == 'SGD': optimizer = tf.keras.optimizers.SGD( learning_rate=PARAMS['lr'] ) ########################## if PARAMS['loss']=='focal_loss': loss = focal_loss(gamma=2.0, alpha=4.0) elif PARAMS['loss']=='categorical_crossentropy': loss = 'categorical_crossentropy' ########################## model_params = stuf(name=PARAMS['model_name'], model_dir=os.path.join(experiment_dir, experiment_name, 'models'), num_classes=PARAMS['num_classes'], frozen_layers = PARAMS['frozen_layers'], input_shape = (*PARAMS['image_size'],PARAMS['num_channels']), base_learning_rate = PARAMS['lr'], regularization = PARAMS['regularization']) #### if PARAMS['model_name']=='shallow': model = build_shallow(input_shape=model_params.input_shape, num_classes=PARAMS['num_classes'], optimizer=optimizer, loss=loss, METRICS=METRICS) else: model = build_model(model_params, optimizer, loss, METRICS) print(f"TRAINING {PARAMS['model_name']}") model.summary(print_fn=lambda x: neptune.log_text('model_summary', x)) history = model.fit(train_data, epochs=num_epochs, callbacks=callbacks, validation_data=validation_data, shuffle=True, initial_epoch=initial_epoch, steps_per_epoch=steps_per_epoch, validation_steps=validation_steps) if 'test' in split_data: results = model.evaluate(test_data, steps=len(split_data['test'][0])) else: results = model.evaluate(validation_data, steps=validation_steps)
def load_data(dataset_name='PNAS', splits={ 'train': 0.7, 'validation': 0.3 }, threshold=50, seed=None, shuffle_train=True, exclude_classes=[], include_classes=[], use_tfrecords=False, tfrecord_dir=None, samples_per_shard=800): datasets = { 'PNAS': pnas_dataset.PNASDataset(), 'Leaves': leaves_dataset.LeavesDataset(), 'Fossil': fossil_dataset.FossilDataset() } data_files = datasets[dataset_name] data_files.exclude_rare_classes(threshold=threshold) encoder = base_dataset.LabelEncoder(data_files.classes) classes = list((set(encoder.classes) - set(exclude_classes)).union( set(include_classes))) data_files, excluded_data_files = data_files.enforce_class_whitelist( class_names=classes) x = list(data_files.data['path'].values) y = np.array(encoder.encode(data_files.data['family'])) shuffled_data = list(zip(x, y)) random.shuffle(shuffled_data) partitioned_data = partition_data(data=shuffled_data, partitions=OrderedDict(splits)) split_data = {k: v for k, v in partitioned_data.items() if len(v) > 0} for subset, subset_data in split_data.items(): split_data[subset] = [list(i) for i in unzip(subset_data)] if use_tfrecords: split_datasets = load_data_from_tfrecords( tfrecord_dir=tfrecord_dir, data=split_data, samples_per_shard=samples_per_shard, num_classes=len(classes)) train_data, validation_data = split_datasets['train'], split_datasets[ 'validation'] else: paths = tf.data.Dataset.from_tensor_slices(split_data['train'][0]) labels = tf.data.Dataset.from_tensor_slices(split_data['train'][1]) train_data = tf.data.Dataset.zip((paths, labels)) if shuffle_train: train_data = train_data.shuffle(int(data_files.num_samples * splits['train']), seed=seed, reshuffle_each_iteration=True) train_data = train_data.cache() train_data = train_data.map(lambda x, y: (tf.image.convert_image_dtype( load_img(x) * 255.0, dtype=tf.uint8), y), num_parallel_calls=-1) paths = tf.data.Dataset.from_tensor_slices(split_data['validation'][0]) labels = tf.data.Dataset.from_tensor_slices( split_data['validation'][1]) validation_data = tf.data.Dataset.zip((paths, labels)) validation_data = validation_data.cache() validation_data = validation_data.map( lambda x, y: (tf.image.convert_image_dtype(load_img(x) * 255.0, dtype=tf.uint8), y), num_parallel_calls=-1) return { 'train': train_data, 'validation': validation_data }, data_files, excluded_data_files