def main(args):
c = utils.make_config(args)
d = input_pipeline.DataInputPipeline(args.inputs, args.vocab, args.labels,
c)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
# gpu_options = tf.GPUOptions(allow_growth=True)
sess = tf.Session(
) #config=tf.ConfigProto(gpu_options=gpu_options, allow_soft_placement=True))
m = models.Model(c, sess, d)
print 'INFO: loading model from checkpoint...'
m.load(dir=args.checkpoint)
print 'INFO: done!'
sess.run(tf.global_variables_initializer())
print 'INFO: starting inference...'
prog = utils.Progbar(target=d.get_num_batches())
source_out = {'source': [], 'len': [], 'attn': [], 'ids': []}
sales_out = {'label': [], 'pred': []}
price_out = {'label': [], 'pred': []}
shop_out = {'label': [], 'pred': []}
category_out = {'label': [], 'pred': []}
loss_out = []
for i, batch in enumerate(d.batch_iter()):
sales_hat, price_hat, shop_hat, category_hat, loss, attn = \
m.test_on_batch(*batch[:-1])
prog.update(i, [('train loss', loss)])
# record results
source, source_len, log_sales, price, shop, category, ids = batch
source_out['source'] += source
source_out['len'] += source_len
source_out['attn'] += attn.tolist()
source_out['ids'] += ids
sales_out['label'] += log_sales
sales_out['pred'] += sales_hat.tolist()
price_out['label'] += price
price_out['pred'] += price_hat.tolist()
shop_out['label'] += shop
shop_out['pred'] += shop_hat.tolist()
category_out['label'] += category
category_out['pred'] += category_hat.tolist()
loss_out += loss
if args.output is not None:
print '\nINFO: dumping output to ', args.output
write_pickle(source_out, args.output)
print 'INFO: done \(^o^)/'
from utils import make_config
sys.path.append('..')
from seismiqb import SeismicGeometry
# Help message
MSG = """Convert SEG-Y cube to HDF5.
Input SEG-Y file must have correctly filled `INLINE_3D` and `CROSSLINE_3D` headers.
A lot of various statistics about traces are also inferred and stored in the resulting file,
so this script takes some time.
"""
# Argname, description, dtype, default
ARGS = [
('cube-path', 'path to the SEG-Y cube to convert to HDF5', str, None),
]
if __name__ == '__main__':
config = make_config(MSG, ARGS, os.path.basename(__file__).split('.')[0])
geometry = SeismicGeometry(
config['cube-path'],
headers=SeismicGeometry.HEADERS_POST_FULL,
index_headers=SeismicGeometry.INDEX_POST,
collect_stats=True, spatial=True,
)
geometry.make_hdf5()
def main(args):
if not os.path.exists(args.checkpoint):
os.makedirs(args.checkpoint)
c = utils.make_config(args)
d = input_pipeline.DataInputPipeline(args.inputs, args.vocab, args.labels,
c)
os.environ[
'CUDA_VISIBLE_DEVICES'] = args.gpu # Or whichever device you would like to use
# gpu_options = tf.GPUOptions(allow_growth=True)
sess = tf.Session(
) #config=tf.ConfigProto(gpu_options=gpu_options, allow_soft_placement=True))
m = models.Model(c, sess, d)
sess.run(tf.global_variables_initializer())
if args.train:
print 'INFO: starting training...'
prog = utils.Progbar(target=d.get_num_batches())
epoch = 1
for _ in range(c.num_epochs):
epoch_loss = 0
for i, batch in enumerate(d.batch_iter()):
sales_hat, price_hat, shop_hat, category_hat, loss = \
m.train_on_batch(*batch[:-1])
prog.update(i + 1, [('train loss', loss)])
epoch_loss += loss
print '\n INFO: EPOCH ', epoch, ' MEAN LOSS: ', epoch_loss / float(
d.get_num_batches())
print 'INFO: saving checkpoint...'
m.save(args.checkpoint)
print 'INFO: ...done!'
epoch += 1
else:
print 'INFO: loading model from checkpoint...'
m.load(dir=args.checkpoint)
print 'INFO: done!'
if args.output is not None:
print 'INFO: starting inference...'
prog = utils.Progbar(target=d.get_num_batches())
source_out = {'source': [], 'len': [], 'attn': [], 'ids': []}
sales_out = {'label': [], 'pred': []}
price_out = {'label': [], 'pred': []}
shop_out = {'label': [], 'pred': []}
category_out = {'label': [], 'pred': []}
loss_out = []
for i, batch in enumerate(d.batch_iter()):
sales_hat, price_hat, shop_hat, category_hat, loss, attn = \
m.test_on_batch(*batch[:-1])
prog.update(i, [('train loss', loss)])
# record results
source, source_len, log_sales, price, shop, category, ids = batch
source_out['source'] += source
source_out['len'] += source_len
source_out['attn'] += attn.tolist()
source_out['ids'] += ids
sales_out['label'] += log_sales
sales_out['pred'] += sales_hat.tolist()
price_out['label'] += price
price_out['pred'] += price_hat.tolist()
shop_out['label'] += shop
shop_out['pred'] += shop_hat.tolist()
category_out['label'] += category
category_out['pred'] += category_hat.tolist()
loss_out += loss
print '\nINFO: dumping output to ', args.output
utils.write_pickle(source_out, args.output)
print '\n\nINFO: \(^o^)/ done \(^o^)/'