if __name__ == '__main__':
trial_name = 'whale_segmentation'
snippets = [
["data", "dataset.py"],
]
with dslib.trial.run_trial(trial_name=trial_name,
trials_dir=trials_dir,
description=[],
snippets=snippets) as trial:
train_ds, valid_ds = dataset.get_train_test_gens()
with train_ds as train_gen, valid_ds as valid_gen:
valid_chunk = valid_gen.next()
import theano
import treeano
import canopy
import canopy.sandbox.monitor_ui
print "Making new model."
model = make_model(params)
print(model)
network = model.network()
import dataset
import tflib
'''
CONSTANTS
'''
BATCH_SIZE = 4
IMWIDTH = 200
IMHEIGHT = 280
TESTSPLIT = 0.20
TRIAL_NAME = 'whale1'
tr, te = dataset.get_train_test_gens(
anno_type='Head', rel_img_path='../imgs/',
desired_output_size=(IMWIDTH, IMHEIGHT),
test_split_percentage=TESTSPLIT,
annotations_dir='../code/right_whale_hunt/annotations/',
chunk_size=BATCH_SIZE)
# input / output placeholders
x_in = tf.placeholder("float", [BATCH_SIZE, IMWIDTH, IMHEIGHT, 3])
y_in = tf.placeholder("float", [BATCH_SIZE, IMWIDTH, IMHEIGHT, 1])
# keep probability for dropout
keep_prob = tf.placeholder("float")
conv1a = tflib.conv_bn_relu(x_in,
kernel_size=[3, 3],
out_filters=32,
scope="conv1a",
summarize=True)
conv1a_do = tflib.dropout(conv1a, keep_prob)