Im using a code from GitHub. It is used to detect gluacoma from oct images. Ive been working on fundus images but this is a process i was interested in. However this interest is turning into a nightmare. I cant add a link to the github below if anyone wants context. Ive attached it below.
`"""
Network training
"""
from __future__ import print_function
import sys
import numpy as np
import sklearn.metrics as sm
from common import *
from nutsflow.common import StableRandom
from keras import backend as K
from network import create_network
from keras.metrics import categorical_accuracy
from datetime import datetime
def compute_aucs(network, samples, cache=None):
y_pred = (samples >> read_cube >> flatten_cube >> resize_cube >>
make_enface >> If(cache, cache) >> build_pred_batch >>
network.predict() >> Collect())
aucs = {}
idxs = range(len(LABELS))
ijs = [(i, j) for i in idxs for j in idxs if i > j]
for i, j in ijs:
inclass = Filter(lambda s: s[0][1] in {i, j})
fsamples, fpreds = zip(samples, y_pred) >> inclass >> Unzip()
y_true = [1 if s[1] == i else 0 for s in fsamples]
y_score = [p[i] for p in fpreds]
name = LABELS[j] + '-' + LABELS[i]
aucs[name] = sm.roc_auc_score(y_true, y_score)
return aucs
def prepare_cache():
if CLEARCACHE:
cachetrain.clear()
cacheval.clear()
cachetest.clear()
else:
print('CACHE NOT CLEARED!')
def train(fold):
has_params = len(sys.argv) > 1
print('SYS.ARGV', sys.argv)
if has_params:
params = [int(a) for a in sys.argv[1:]]
print('PARAMS', params)
hp.BN, hp.ORDER, hp.ALGO, hp.LR, hp.FLIPEYE, hp.MIXUP = params[:6]
hp.P_AUG = params[6:10]
n, layer_params = len(params[10:]) // 3, params[10:]
hp.N_FILTER, hp.N_CONV, hp.N_STRIDE = (layer_params[0:n],
layer_params[n:2 * n],
layer_params[2 * n:3 * n])
network = create_network('best_weights.h5')
network.print_layers()
if USE_PRETRAINED:
print('LOADING PRETRAINED NET!')
network.load_weights()
do_augment = sum(hp.P_AUG) > 0
rand = StableRandom(fold)
same_sid = lambda s: s[2].split('-')[0]
split_data = SplitRandom((0.8, 0.1, 0.1), constraint=same_sid, rand=rand)
train_samples, val_samples, test_samples = \
read_samples() >> split_data >> Collect()
train_samples = train_samples >> Head(N_TRAIN)
traindist = train_samples >> CountValues(1)
print('#samples', len(train_samples), len(val_samples), len(test_samples))
print('train-dist', traindist)
print('val-dist', val_samples >> CountValues(1))
print('test-dist', test_samples >> CountValues(1))
best_val_auc = 0
for epoch in range(N_EPOCHS):
print('EPOCH:', epoch)
# build_batch returns [epoch0, epoch1...]
# where epochN is [[numpy],[numpy]]
# network train expects [numpy, numpy]
# this is a hack to remove the extrta []
def epch_hack(ep):
return (ep[0][0], ep[1][0])
t_loss, t_acc = (train_samples >> PrintProgress(train_samples) >>
read_cube >> flatten_cube >> resize_cube >>
make_enface >> Stratify(1, traindist) >>
build_batch >> Map(epoch_hack) >> network.train() >> Unzip())
t_loss_mean = t_loss >> Mean()
t_acc_mean = t_acc >> Mean()
train_time = datetime.now() - start_time
print("train loss : {:.5f}".format(t_loss_mean))
print("train acc : {:.5f}".format(t_acc_mean))
print("train time : {:.1f} min".format(train_time.seconds / 60.0))
aucs = compute_aucs(network, val_samples, cacheval)
v_auc = aucs.values() >> Mean()
print("val auc : {:.5f}".format(v_auc))
for name, auc in aucs.items():
print(' val', name, round(auc, 3))
plot_epoch((t_acc_mean, v_auc))
plot_aucs([v_auc] + list(aucs.values()))
v_auc_mean = aucs.values() >> Mean()
best_val_auc = max(best_val_auc, v_auc_mean)
if SAVE_BEST:
network.save_best(v_auc_mean, isloss=False)
print('test -----------------------------------------')
network.load_weights()
aucs = compute_aucs(network, test_samples, cachetest)
for name, auc in aucs.items():
print(' test', name, round(auc, 3))
test_auc = round(aucs.values() >> Mean(), 3)
best_val_auc = round(best_val_auc, 3)
print('test auc', test_auc)
print('best val auc', best_val_auc)
if has_params:
f = open('hyperparams6.csv', 'a')
pad = [0] * (6 - len(hp.N_FILTER))
n_filter = hp.N_FILTER + pad
n_conv = hp.N_CONV + pad
n_stride = hp.N_STRIDE + pad
results = ', '.join(map(str, [test_auc, best_val_auc] +
[hp.BN, hp.ORDER, hp.ALGO, hp.LR, hp.FLIPEYE, hp.MIXUP] +
hp.P_AUG + n_filter + n_conv + n_stride))
print('results', results)
f.write(results + '\n')
f.close()
return best_val_auc, test_auc
if __name__ == "__main__":
val_aucs, test_aucs = [], []
for fold in range(N_FOLDS):
print('FOLD', fold+1, 'of', N_FOLDS)
val_auc, test_auc = train(fold)
val_aucs.append(val_auc)
test_aucs.append(test_auc)
print('***', N_FOLDS, 'fold cross-validation')
print('*** mean val AUC %.3f %.3f' % (val_aucs >> MeanStd()))
print('*** mean test AUC %.3f %.3f' % (test_aucs >> MeanStd()))`
This is the error Im getting. I know that a cant change into integer but i dont know why this is happening or how to fix it Any help will be appreciated!.
FOLD 1 of 1
SYS.ARGV ['/usr/local/lib/python3.10/dist-packages/ipykernel_launcher.py', '-f', '/root/.local/share/jupyter/runtime/kernel-4e1cc6fa-7d56-4bc7-8934-503c656f7615.json']
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-31-2eebb00bf84b> in <cell line: 139>()
141 for fold in range(N_FOLDS):
142 print('FOLD', fold+1, 'of', N_FOLDS)
--> 143 val_auc, test_auc = train(fold)
144 val_aucs.append(val_auc)
145 test_aucs.append(test_auc)
1 frames
<ipython-input-31-2eebb00bf84b> in <listcomp>(.0)
46 print('SYS.ARGV', sys.argv)
47 if has_params:
---> 48 params = [int(a) for a in sys.argv[1:]]
49 print('PARAMS', params)
50 hp.BN, hp.ORDER, hp.ALGO, hp.LR, hp.FLIPEYE, hp.MIXUP = params[:6]
ValueError: invalid literal for int() with base 10: '-f'
Ive been trying and failing to run it.
