2martens/masterthesis
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Added debug prints for ref counts

Browse Source 
Signed-off-by: Jim Martens <github@2martens.de>
This commit is contained in:
Jim Martens
2019-05-01 11:56:12 +02:00
parent 58d867d84f
commit 54cacfd45d
1 changed files with 25 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
src/twomartens/masterthesis/ssd.py
View File

@ -180,14 +180,19 @@ def _predict_one_epoch(dataset: tf.data.Dataset,
# go through the data set
counter = 0
import gc
import sys
from tensorflow.python.eager import context
for inputs in dataset:
decoded_predictions_batch = []
if use_dropout:
for _ in range(forward_passes_per_image):
decoded_predictions_pass = ssd(inputs)
decoded_predictions_batch.append(np.array(decoded_predictions_pass))
result = np.array(ssd(inputs))
decoded_predictions_batch.append(result)
else:
decoded_predictions_batch.append(np.array(ssd(inputs)))
result = np.array(ssd(inputs))
decoded_predictions_batch.append(result)
# save predictions batch-wise to prevent memory problems
if nr_digits is not None:
@ -197,14 +202,28 @@ def _predict_one_epoch(dataset: tf.data.Dataset,
filename = f"{output_file}-{counter:d}.npy"
with open(filename, 'wb') as file:
np.save(file, np.array(decoded_predictions_batch), allow_pickle=False, fix_imports=False)
decoded_predictions_batch_np = np.array(decoded_predictions_batch)
np.save(file, decoded_predictions_batch_np, allow_pickle=False, fix_imports=False)
counter += 1
print((
f"counter: {counter}"
f"list: {sys.getrefcount(decoded_predictions_batch)}"
f"result: {sys.getrefcount(result)}"
f"np list: {sys.getrefcount(decoded_predictions_batch_np)}"
))
tf.set_random_seed(1)
from tensorflow.python.eager import context
context.context()._clear_caches()
import gc
gc.collect()
print((
f"counter: {counter}"
f"list: {sys.getrefcount(decoded_predictions_batch)}"
f"result: {sys.getrefcount(result)}"
f"np list: {sys.getrefcount(decoded_predictions_batch_np)}"
))
epoch_end_time = time.time()
per_epoch_time = epoch_end_time - epoch_start_time