keras_train.py

import numpy as np
import awkward

import tensorflow as tf
from tensorflow import keras
from tf_keras_model import get_particle_net, get_particle_net_lite

import os
import logging

logging.basicConfig(level=logging.INFO, format='[%(asctime)s] %(levelname)s: %(message)s')


def stack_arrays(a, keys, axis=-1):
    flat_arr = np.stack([a[k].flatten() for k in keys], axis=axis)
    return awkward.JaggedArray.fromcounts(a[keys[0]].counts, flat_arr)


def pad_array(a, maxlen, value=0., dtype='float32'):
    x = (np.ones((len(a), maxlen)) * value).astype(dtype)
    for idx, s in enumerate(a):
        if not len(s):
            continue
        trunc = s[:maxlen].astype(dtype)
        x[idx, :len(trunc)] = trunc
    return x


class Dataset(object):

    def __init__(self, filepath, feature_dict = {}, label='label', pad_len=100, data_format='channel_first'):
        self.filepath = filepath
        self.feature_dict = feature_dict
        if len(feature_dict)==0:
            feature_dict['points'] = ['part_etarel', 'part_phirel']
            feature_dict['features'] = ['part_pt_log', 'part_e_log', 'part_etarel', 'part_phirel']
            feature_dict['mask'] = ['part_pt_log']
        self.label = label
        self.pad_len = pad_len
        assert data_format in ('channel_first', 'channel_last')
        self.stack_axis = 1 if data_format=='channel_first' else -1
        self._values = {}
        self._label = None
        self._load()

    def _load(self):
        logging.info('Start loading file %s' % self.filepath)
        counts = None
        with awkward.load(self.filepath) as a:
            self._label = a[self.label]
            for k in self.feature_dict:
                cols = self.feature_dict[k]
                if not isinstance(cols, (list, tuple)):
                    cols = [cols]
                arrs = []
                for col in cols:
                    if counts is None:
                        counts = a[col].counts
                    else:
                        assert np.array_equal(counts, a[col].counts)
                    arrs.append(pad_array(a[col], self.pad_len))
                self._values[k] = np.stack(arrs, axis=self.stack_axis)
        logging.info('Finished loading file %s' % self.filepath)


    def __len__(self):
        return len(self._label)

    def __getitem__(self, key):
        if key==self.label:
            return self._label
        else:
            return self._values[key]
    
    @property
    def X(self):
        return self._values
    
    @property
    def y(self):
        return self._label

    def shuffle(self, seed=None):
        if seed is not None:
            np.random.seed(seed)
        shuffle_indices = np.arange(self.__len__())
        np.random.shuffle(shuffle_indices)
        for k in self._values:
            self._values[k] = self._values[k][shuffle_indices]
        self._label = self._label[shuffle_indices]


def lr_schedule(epoch):
    lr = 1e-3
    if epoch > 10:
        lr *= 0.1
    elif epoch > 20:
        lr *= 0.01
    logging.info('Learning rate: %f'%lr)
    return lr


if __name__ == "__main__":
        
    train_dataset = Dataset('converted/train_file_0.awkd', data_format='channel_last')
    val_dataset = Dataset('converted/val_file_0.awkd', data_format='channel_last')

    model_type = 'particle_net_lite' # choose between 'particle_net' and 'particle_net_lite'
    num_classes = train_dataset.y.shape[1]
    input_shapes = {k:train_dataset[k].shape[1:] for k in train_dataset.X}
    if 'lite' in model_type:
        model = get_particle_net_lite(num_classes, input_shapes)
    else:
        model = get_particle_net(num_classes, input_shapes)
        
    # Training parameters
    batch_size = 1024 if 'lite' in model_type else 384
    epochs = 30

    model.compile(loss='categorical_crossentropy',
                optimizer=keras.optimizers.Adam(learning_rate=lr_schedule(0)),
                metrics=['accuracy'])
    model.summary()


    save_dir = 'model_checkpoints'
    model_name = '%s_model.{epoch:03d}.h5' % model_type
    if not os.path.isdir(save_dir):
        os.makedirs(save_dir)
    filepath = os.path.join(save_dir, model_name)


    # Prepare callbacks for model saving and for learning rate adjustment.
    checkpoint = keras.callbacks.ModelCheckpoint(filepath=filepath,
                                monitor='val_accuracy',
                                verbose=1,
                                save_best_only=False)


    lr_scheduler = keras.callbacks.LearningRateScheduler(lr_schedule)
    progress_bar = keras.callbacks.ProgbarLogger()
    callbacks = [checkpoint, lr_scheduler, progress_bar]


    train_dataset.shuffle()
    model.fit(train_dataset.X, train_dataset.y,
            batch_size=batch_size,
    #           epochs=epochs,
            epochs=1, # --- train only for 1 epoch here for demonstration ---
            validation_data=(val_dataset.X, val_dataset.y),
            shuffle=True,
            callbacks=callbacks)