def test_demo_returnn_as_framework(self):
print("Prepare.")
import subprocess
import shutil
from glob import glob
from Util import get_login_username
# echo via subprocess, because this stdout as well as the other will always be visible.
subprocess.check_call(
["echo", "travis_fold:start:test_demo_returnn_as_framework"])
assert os.path.exists("setup.py")
if glob("dist/*.tar.gz"):
# we want it unique below
for fn in glob("dist/*.tar.gz"):
os.remove(fn)
if os.path.exists("MANIFEST"):
os.remove("MANIFEST") # auto-generated. will be recreated
if os.path.exists("docs/crnn"):
os.remove("docs/crnn"
) # this is auto-generated, and confuses setup.py sdist
tmp_model_dir = "/tmp/%s/returnn-demo-as-framework" % get_login_username(
)
if os.path.exists(tmp_model_dir):
shutil.rmtree(tmp_model_dir, ignore_errors=True)
print("setup.py sdist, to create package.")
subprocess.check_call([py, "setup.py", "sdist"])
dist_fns = glob("dist/*.tar.gz")
assert len(dist_fns) == 1
dist_fn = os.path.abspath(dist_fns[0])
pip_path = which_pip()
print("Pip install Returnn.")
in_virtual_env = hasattr(
sys, 'real_prefix') # https://stackoverflow.com/questions/1871549/
cmd = [py, pip_path, "install"]
if not in_virtual_env:
cmd += ["--user"]
cmd += ["-v", dist_fn]
print("$ %s" % " ".join(cmd))
subprocess.check_call(cmd, cwd="/")
print("Running demo now.")
subprocess.check_call([py, "demo-returnn-as-framework.py"],
cwd="demos")
print("Success.")
subprocess.check_call(
["echo", "travis_fold:end:test_demo_returnn_as_framework"])
optimize_move_layers_out = True
debug_print_layer_output_template = True
# trainer
batching = "random"
batch_size = 1000
max_seqs = 5 # TODO: 100
chunking = "0"
truncation = -1
#gradient_clip = 10
#gradient_nan_inf_filter = True
adam = True
gradient_noise = 0.3
learning_rate = 0.0005
learning_rate_control = "newbob"
learning_rate_control_relative_error_relative_lr = True
#model = "./test/transformer-hmm"
model = "/tmp/%s/crnn/%s/model" % (
get_login_username(), demo_name
) # https://github.com/tensorflow/tensorflow/issues/6537
num_epochs = 100
save_interval = 1
debug_mode = True
debug_add_check_numerics_ops = True
# log
log_verbosity = 5
batch_size=5000,
max_seqs=10,
chunking="0",
network={
"fw0": {"class": "rec", "unit": "NativeLstm2", "dropout": 0.1, "n_out": 10},
"output": {"class": "softmax", "loss": "ce", "from": ["fw0"]}
},
# training
nadam=True,
learning_rate=0.01,
num_epochs=100,
debug_add_check_numerics_ops=True,
model="/tmp/%s/returnn-demo-as-framework/model" % get_login_username(),
cleanup_old_models=True,
learning_rate_control="newbob_multi_epoch",
learning_rate_control_relative_error_relative_lr=True,
newbob_multi_num_epochs=3, newbob_multi_update_interval=1, newbob_learning_rate_decay=0.9,
learning_rate_file="/tmp/%s/returnn-demo-as-framework/newbob.data" % get_login_username(),
# log
log_verbosity=3
))
engine = Engine(config)
train_data = init_dataset({"class": "Task12AXDataset", "num_seqs": 1000, "name": "train"})
dev_data = init_dataset({"class": "Task12AXDataset", "num_seqs": 100, "name": "dev", "fixed_random_seed": 1})
network = TransformerNetwork(model_size=128, normalized_loss = True).build() search_output_layer = "decision" optimize_move_layers_out = True debug_print_layer_output_template=True # trainer batching = "random" batch_size = 1000 max_seqs = 5 # TODO: 100 chunking = "0" truncation = -1 #gradient_clip = 10 #gradient_nan_inf_filter = True adam = True gradient_noise = 0.3 learning_rate = 0.0005 learning_rate_control = "newbob" learning_rate_control_relative_error_relative_lr = True #model = "./test/transformer-hmm" model = "/tmp/%s/crnn/%s/model" % (get_login_username(), demo_name) # https://github.com/tensorflow/tensorflow/issues/6537 num_epochs = 100 save_interval = 1 debug_mode = True debug_add_check_numerics_ops = True # log log_verbosity = 5