predictions_chunk = predictions_chunk[:
chunk_length] # cut off zeros / padding
print(" compute average over transforms")
predictions_chunk_avg = predictions_chunk.reshape(
-1, len(augmentation_transforms), 37).mean(1)
predictions_list.append(predictions_chunk_avg)
time_since_start = time.time() - start_time
print(" %s since start" % load_data.hms(time_since_start))
all_predictions = np.vstack(predictions_list)
print("Write predictions to %s" % target_path_valid)
load_data.save_gz(target_path_valid, all_predictions)
print("Evaluate")
rmse_valid = analysis['losses_valid'][-1]
rmse_augmented = np.sqrt(np.mean((y_valid - all_predictions)**2))
print(" MSE (last iteration):\t%.6f" % rmse_valid)
print(" MSE (augmented):\t%.6f" % rmse_augmented)
if DO_TEST:
print()
print("TEST SET")
print("Compute predictions")
predictions_list = []
start_time = time.time()
for e, (chunk_data, chunk_length) in enumerate(test_gen):
else:
try:
print ''
print 'Re-evalulating and predicting'
if DO_VALID:
evalHist = winsol.evaluate([xs_valid[0], xs_valid[1]],
y_valid=y_valid)
winsol.save_loss(modelname='model_norm_metrics')
evalHist = winsol.load_loss(modelname='model_norm_metrics')
print ''
predictions = winsol.predict([xs_valid[0], xs_valid[1]])
print "Write predictions to %s" % target_path_valid
load_data.save_gz(target_path_valid, predictions)
except KeyboardInterrupt:
print "\ngot keyboard interuption"
save_exit()
except ValueError:
print "\ngot value error, could be the end of the generator in the fit"
save_exit()
evalHist = winsol.load_loss(modelname='model_norm_metrics')
if np.shape(predictions) != np.shape(y_valid):
raise ValueError(
'prediction and validation set have different shapes, %s to %s ' %
(np.shape(predictions), np.shape(y_valid)))
predictions_chunk = np.vstack(predictions_chunk_list)
predictions_chunk = predictions_chunk[:chunk_length] # cut off zeros / padding
print " compute average over transforms"
predictions_chunk_avg = predictions_chunk.reshape(-1, len(augmentation_transforms), 37).mean(1)
predictions_list.append(predictions_chunk_avg)
time_since_start = time.time() - start_time
print " %s since start" % load_data.hms(time_since_start)
all_predictions = np.vstack(predictions_list)
print "Write predictions to %s" % target_path_valid
load_data.save_gz(target_path_valid, all_predictions)
print "Evaluate"
rmse_valid = analysis['losses_valid'][-1]
rmse_augmented = np.sqrt(np.mean((y_valid - all_predictions)**2))
print " MSE (last iteration):\t%.6f" % rmse_valid
print " MSE (augmented):\t%.6f" % rmse_augmented
if DO_TEST:
print
print "TEST SET"
print "Compute predictions"
predictions_list = []
start_time = time.time()
# print " %s" % os.path.basename(path)
predictions = load_data.load_gz(path)
predictions_uniform = predictions * (1.0 / len(predictions_test_paths))
predictions_separate = predictions * weights_separate[None, :]
predictions *= weight # inplace scaling
if blended_predictions is None:
blended_predictions = predictions
blended_predictions_separate = predictions_separate
blended_predictions_uniform = predictions_uniform
else:
blended_predictions += predictions
blended_predictions_separate += predictions_separate
blended_predictions_uniform += predictions_uniform
print
print "Storing blended predictions (shared) in %s" % TARGET_PATH
load_data.save_gz(TARGET_PATH, blended_predictions)
print
print "Storing blended predictions (separate) in %s" % TARGET_PATH_SEPARATE
load_data.save_gz(TARGET_PATH_SEPARATE, blended_predictions_separate)
print
print "Storing uniformly blended predictions in %s" % TARGET_PATH_UNIFORM
load_data.save_gz(TARGET_PATH_UNIFORM, blended_predictions_uniform)
print
print "Done!"
for path, weight, weights_separate in zip(predictions_test_paths, out_s,
out_s2):
# print " %s" % os.path.basename(path)
predictions = load_data.load_gz(path)
predictions_uniform = predictions * (1.0 / len(predictions_test_paths))
predictions_separate = predictions * weights_separate[None, :]
predictions *= weight # inplace scaling
if blended_predictions is None:
blended_predictions = predictions
blended_predictions_separate = predictions_separate
blended_predictions_uniform = predictions_uniform
else:
blended_predictions += predictions
blended_predictions_separate += predictions_separate
blended_predictions_uniform += predictions_uniform
print()
print("Storing blended predictions (shared) in %s" % TARGET_PATH)
load_data.save_gz(TARGET_PATH, blended_predictions)
print()
print("Storing blended predictions (separate) in %s" % TARGET_PATH_SEPARATE)
load_data.save_gz(TARGET_PATH_SEPARATE, blended_predictions_separate)
print()
print("Storing uniformly blended predictions in %s" % TARGET_PATH_UNIFORM)
load_data.save_gz(TARGET_PATH_UNIFORM, blended_predictions_uniform)
print()
print("Done!")
print "SUBSET: %s" % subset
print
if subset == 'train':
num_images = load_data.num_train
ids = load_data.train_ids
elif subset == 'test':
num_images = load_data.num_test
ids = load_data.test_ids
def process(k):
print "image %d/%d (%s)" % (k + 1, num_images, subset)
img_id = ids[k]
img = load_data.load_image(img_id, from_ram=True, subset=subset)
return estimate_params(img)
pool = mp.Pool(NUM_PROCESSES)
estimated_params = pool.map(process, xrange(num_images), chunksize=100)
pool.close()
pool.join()
# estimated_params = map(process, xrange(num_images)) # no mp for debugging
params_array = np.array(estimated_params)
target_path = TARGET_PATTERN % subset
print "Saving to %s..." % target_path
load_data.save_gz(target_path, params_array)
