common_config.py

"""
Provides some common settings for RETURNN configs (training + search).

Use `from ... import *` directly in your config for this module.
"""

import sys
from .multi_gpu_horovod import *  # noqa
from returnn.config import get_global_config

config = get_global_config()

# task
use_tensorflow = True
task = config.value("task", "train")

debug_mode = False
if int(os.environ.get("RETURNN_DEBUG", "0")):
    print("** DEBUG MODE", file=sys.stderr)
    # By itself, this doesn't do anything.
    # In your main config, you might select a smaller batch size, or other things,
    # depending on this flag.
    debug_mode = True
    dry_run = True

# Enforce usage of GPU. (Disable this for testing when you only have a CPU.)
device = os.environ.get("RETURNN_DEVICE", None if debug_mode else "gpu")
# allow_growth should be used when the GPU is shared (e.g. with Xorg).
# However, don't set by default (for performance reasons).
if os.environ.get("RETURNN_TF_SESSION_OPTS"):
    tf_session_opts = eval(os.environ["RETURNN_TF_SESSION_OPTS"])
# tf_session_opts = {"gpu_options": {"allow_growth": True}}

if config.has("beam_size"):
    beam_size = config.int("beam_size", 0)
    print("** beam_size %i" % beam_size, file=sys.stderr)
else:
    beam_size = 12

search_output_layer = "decision"
debug_print_layer_output_template = True  # doesn't cost anything. always recommended

log_batch_size = True  # only relevant with verbosity 5
tf_log_memory_usage = True

# Verbosity 5 will print some stats (loss, batch size, maybe other things) per single minibatch.
# All lower verbosity will not. (If you don't want 5, then recommended would be 4 or 3.)
# (You will get an interactive bar for verbosity<5 for the epoch progress if you use an interactive shell.)
log_verbosity = 5