def load_hparam_search(name):
""" Reads json-formatted hparam search file spec to pandas DF,
and loads additional metrics into the dataframe.
"""
hpath = get_git_root() + "experiments/searches/"
df = pd.read_json(hpath + name,
orient='index').rename_axis('id').reset_index()
# JSON convert the tuples in hparam search to list when it's interpreted.
# Convert the values to str to make it workable
df['kernel_size'] = [str(x) for x in df['kernel_size'].values]
# Add additional metrics to df
accs = []
f1 = []
mcc = []
auc = []
for e_id in df['id']:
e = load_experiment(e_id)
accs.append(e['metrics']['accuracy_score'])
f1.append(e['metrics']['f1_score'])
mcc.append(e['metrics']['matthews_corrcoef'])
auc.append(e['metrics']['roc_auc_score'])
df['accuracy_score'] = accs
df['f1_score'] = f1
df['matthews_corrcoef'] = mcc
df['roc_auc_score'] = auc
return df
def set_config(self, config):
""" Set default config for model.fit and additional kfold cross-
validation arguments. If a config is provided to the function,
replace given keys with the ones provided, and add new keys present.
"""
# Get github repo root path
rpath = get_git_root()
self.config = {
'fit_args': {
'batch_size': None,
'epochs': 1,
'verbose': 2,
},
'kfold_args': {
'n_splits': 5,
'shuffle': True,
},
'path_args': {
'repo_root': rpath,
'models': rpath + 'models/',
'figures': rpath + 'figures/',
'experiments': rpath + 'experiments/',
'results': rpath + 'results',
'model_config': rpath + 'experiments/model_config/',
},
'random_seed': None,
}
if config is not None:
for major_key in config.keys():
# Treat the value as dict, save directly if not dict
try:
for k, v in config[major_key].items():
self.config[major_key][k] = v
except AttributeError:
self.config[major_key] = config[major_key]
except KeyError:
self.config[major_key] = config[major_key]
from tensorflow.keras.layers import Conv2D, Dense, Flatten, MaxPooling2D
import tensorflow as tf
import numpy as np
import json
import warnings
warnings.filterwarnings('ignore', category=FutureWarning)
# ================== Config =======================
with open("results_experiment_config.json", 'r') as fp:
config = json.load(fp)
# ================== Callbacks ====================
# ================== Import Data ==================
DATA_PATH = get_git_root() + "data/simulated/"
images = np.load(DATA_PATH + config['data']['images'])
images = images.reshape(images.shape[0], 256)
energies = np.load(DATA_PATH + config['data']['energies'])
positions = np.load(DATA_PATH + config['data']['positions'])
single_indices, double_indices, close_indices = event_indices(positions)
# log-scale the images if desireable
config['scaling'] = "minmax"
if "np.log" in config['scaling']:
images = np.log1p(images)
# set tf random seed
tf.random.set_seed(config['random_seed'])
# ================== Import Data ==================
with tf.device(get_tf_device(20)):
sys.path.append('../../master_scripts')
from master_scripts.noise_addition import *
import numpy as np
from master_scripts.data_functions import get_git_root
import json
import time
#config = {
# 'distfile': "ratiodist.txt",
# 'imagefile': ["images_training_18000000_202012080020.npy","images_test_2000000_202012080020.npy"]
#}
with open(sys.argv[1], 'r') as fp:
config = json.load(fp)
DATA_PATH = get_git_root() + "data/simulated/"
DIST_PATH = get_git_root() + "data/real/"
for data in config['imagefile']:
print(data)
print(type(data))
print("Noising " , data, " this may take some time.")
start = time.time()
images = np.load(DATA_PATH + data).copy()
dist = gen_dist(DIST_PATH + config['distfile'])
for i in range(0, len(images)):
#print("START LOOP")
#print(image[0][0:5])
images[i] = images[i].reshape(16,16) * rnoise_gen(dist).reshape(16,16)
#print('***************************')
#print(image[0][0:5])
def load_experiment(e_id):
repo_root = get_git_root()
e_path = repo_root + "experiments/"
with open(e_path + e_id + ".json", "r") as fp:
e = json.load(fp)
return e
def anodedata_classification_table(experiment_id,
data_name,
return_events=False):
"""Outputs the event dict post-classification as a table
:param experiment_id: unique id of experiment
:param data_name: filename of datafile without type suffix
ex. "anodedata_500k"
:param return_events: bool, return the events dict if True
"""
# Load the event classification results
repo_root = get_git_root()
fname = repo_root + "results/events_classified_" + data_name + "_"
fname += experiment_id + ".json"
with open(fname, "r") as fp:
events = json.load(fp)
# Generate list of unique event descriptors present in the events
descriptors = list(
set([event['event_descriptor'] for event in events.values()]))
# Frequency of each type of descriptor for each event type
desc_class = {
'single': [],
'double': [],
}
for event in events.values():
desc_class[event['event_class']].append(event['event_descriptor'])
# Translation dict for event descriptor
# Note that not all of these correspond to something that may exists.
translate_descriptor = {
1: "Implant",
2: "Decay",
3: "implant + Decay",
4: "Light ion",
5: "Implant + Light Ion",
6: "Decay + Light Ion",
7: "Implant + Decay + Light Ion",
8: "Double (time)",
9: "Implant + Double (time)",
10: "Decay + Double (time)",
11: "Implant + Decay + Double (time)",
12: "Light ion + Double (time)",
13: "Implant + Light Ion + Double (time)",
14: "Decay + Light ion + Double (time)",
15: "Implant + Decay + Light Ion + Double (time)",
16: "Double (space)",
17: "Implant + Double (space)",
18: "Decay + Double (space)"
}
# Print a table-like structure for viewing
print("Classification results for {}:".format(experiment_id))
print(
"|Event descriptor | Event type | singles | doubles |"
)
print(
"| :--- | :---: | :---: | :---: |"
)
for d in descriptors:
print("|{:^17d}|{:^30s}|{:^9d}|{:^9d}|".format(
d, translate_descriptor[d], desc_class['single'].count(d),
desc_class['double'].count(d)))
if return_events:
return events
import numpy as np
import json
import warnings
warnings.filterwarnings('ignore', category=FutureWarning)
# ================== Config =======================
config = {
'fit_args': {
'epochs': 20,
'batch_size': 32,
},
'random_seed': 120,
}
# ================== Import Data ==================
DATA_PATH = get_git_root() + "data/simulated/"
images = np.load(DATA_PATH + "images_200k.npy")
images = images.reshape(images.shape[0], 16, 16, 1)
labels = np.load(DATA_PATH + "labels_200k.npy")
x_idx = np.arange(images.shape[0])
train_idx, val_idx, u1, u2 = train_test_split(
x_idx, x_idx, random_state=config['random_seed'])
# ================== Search params ================
batch_sizes = [32, 64, 128, 256]
# set tf random seed
tf.random.set_seed(config['random_seed'])
id_param = {}
search_name = "batch_size_deeper"
# ============================================================================ # Modifies pixels in the simulated detector images in order to make it more # similar to the real data. # In anodedata_500k.txt pixel 3, 13 is set to 0 due to being unreliably noisy # The same is the case for the bordering pixels y=0 and y=15 # ============================================================================ import numpy as np from master_scripts.data_functions import get_git_root repo_root = get_git_root() images = np.load(repo_root + "data/simulated/images_full.npy") images[:, 3, 13] = 0 images[:, 0, :] = 0 images[:, 15, :] = 0 np.save(repo_root + "data/simulated/images_full_pixelmod.npy", images)
