Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

add distillation #225

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

add distillation #225

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
0470648
add distillation
Feb 26, 2024
80bcbe4
swap not
Feb 26, 2024
a6b7447
set teacher to none if not used
masterkni6 Mar 2, 2024
c359435
make teacher outputs none if no teacher
masterkni6 Mar 2, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
100 changes: 60 additions & 40 deletions tf/tfprocess.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -146,7 +146,7 @@ def reset(self):

class TFProcess:

def __init__(self, cfg):
def __init__(self, cfg, is_teacher=False):
self.cfg = cfg
self.net = Net()
self.root_dir = os.path.join(self.cfg['training']['path'],
Expand Down Expand Up @@ -329,21 +329,21 @@ def __init__(self, cfg):
self.renorm_max_d = self.cfg['training'].get('renorm_max_d', 0)
self.renorm_momentum = self.cfg['training'].get(
'renorm_momentum', 0.99)

if self.cfg['gpu'] == 'all':
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
self.strategy = tf.distribute.MirroredStrategy()
tf.distribute.experimental_set_strategy(self.strategy)
else:
gpus = tf.config.experimental.list_physical_devices('GPU')
print(gpus)
tf.config.experimental.set_visible_devices(gpus[self.cfg['gpu']],
'GPU')
tf.config.experimental.set_memory_growth(gpus[self.cfg['gpu']],
True)
self.strategy = None
if not is_teacher:
if self.cfg['gpu'] == 'all':
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
self.strategy = tf.distribute.MirroredStrategy()
tf.distribute.experimental_set_strategy(self.strategy)
else:
gpus = tf.config.experimental.list_physical_devices('GPU')
print(gpus)
tf.config.experimental.set_visible_devices(gpus[self.cfg['gpu']],
'GPU')
tf.config.experimental.set_memory_growth(gpus[self.cfg['gpu']],
True)
self.strategy = None
if self.model_dtype == tf.float16:
tf.keras.mixed_precision.experimental.set_policy('mixed_float16')

Expand Down Expand Up @@ -377,7 +377,10 @@ def init(self, train_dataset, test_dataset, validation_dataset=None):
else:
self.init_net()

def init_net(self):
def set_teacher(self, teacher):
self.teacher = teacher

def init_net(self, is_teacher=False):
self.l2reg = tf.keras.regularizers.l2(l=0.5 * (0.0001))
input_var = tf.keras.Input(shape=(112, 8, 8))
outputs = self.construct_net(input_var)
Expand Down Expand Up @@ -423,21 +426,25 @@ def init_net(self):
import tensorflow_addons as tfa
self.optimizer = tfa.optimizers.Lookahead(self.optimizer)

def correct_policy(target, output):
def correct_policy(target, output, teacher_output=None):
output = tf.cast(output, tf.float32)
# Calculate loss on policy head
if self.cfg['training'].get('mask_legal_moves'):
# extract mask for legal moves from target policy
move_is_legal = tf.greater_equal(target, 0)
# replace logits of illegal moves with large negative value (so that it doesn't affect policy of legal moves) without gradient
if teacher_output is not None:
illegal_filler = tf.zeros_like(teacher_output) - 1.0e10
target = tf.nn.softmax(tf.where(move_is_legal, teacher_output, illegal_filler))
illegal_filler = tf.zeros_like(output) - 1.0e10
output = tf.where(move_is_legal, output, illegal_filler)
# y_ still has -1 on illegal moves, flush them to 0
target = tf.nn.relu(target)
if teacher_output is None:
target = tf.nn.relu(target)
return target, output

def policy_loss(target, output):
target, output = correct_policy(target, output)
def policy_loss(target, output, teacher_output=None):
target, output = correct_policy(target, output, teacher_output)
policy_cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
labels=tf.stop_gradient(target), logits=output)
return tf.reduce_mean(input_tensor=policy_cross_entropy)
Expand Down Expand Up @@ -597,25 +604,26 @@ def accuracy(target, output):
self.cfg['training']['lr_boundaries'].sort()
self.warmup_steps = self.cfg['training'].get('warmup_steps', 0)
self.lr = self.cfg['training']['lr_values'][0]
self.test_writer = tf.summary.create_file_writer(
os.path.join(os.getcwd(),
"leelalogs/{}-test".format(self.cfg['name'])))
self.train_writer = tf.summary.create_file_writer(
os.path.join(os.getcwd(),
"leelalogs/{}-train".format(self.cfg['name'])))
if vars(self).get('validation_dataset', None) is not None:
self.validation_writer = tf.summary.create_file_writer(
os.path.join(
os.getcwd(),
"leelalogs/{}-validation".format(self.cfg['name'])))
if self.swa_enabled:
self.swa_writer = tf.summary.create_file_writer(
if not is_teacher:
self.test_writer = tf.summary.create_file_writer(
os.path.join(os.getcwd(),
"leelalogs/{}-swa-test".format(self.cfg['name'])))
self.swa_validation_writer = tf.summary.create_file_writer(
os.path.join(
os.getcwd(),
"leelalogs/{}-swa-validation".format(self.cfg['name'])))
"leelalogs/{}-test".format(self.cfg['name'])))
self.train_writer = tf.summary.create_file_writer(
os.path.join(os.getcwd(),
"leelalogs/{}-train".format(self.cfg['name'])))
if vars(self).get('validation_dataset', None) is not None:
self.validation_writer = tf.summary.create_file_writer(
os.path.join(
os.getcwd(),
"leelalogs/{}-validation".format(self.cfg['name'])))
if self.swa_enabled:
self.swa_writer = tf.summary.create_file_writer(
os.path.join(os.getcwd(),
"leelalogs/{}-swa-test".format(self.cfg['name'])))
self.swa_validation_writer = tf.summary.create_file_writer(
os.path.join(
os.getcwd(),
"leelalogs/{}-swa-validation".format(self.cfg['name'])))
self.checkpoint = tf.train.Checkpoint(optimizer=self.orig_optimizer,
model=self.model,
global_step=self.global_step,
Expand All @@ -628,6 +636,10 @@ def accuracy(target, output):
keep_checkpoint_every_n_hours=24,
checkpoint_name=self.cfg['name'])

if is_teacher:
assert self.manager.latest_checkpoint is not None, "Teacher has no checkpoint"


def replace_weights(self, proto_filename, ignore_errors=False):
self.net.parse_proto(proto_filename)

Expand Down Expand Up @@ -749,11 +761,19 @@ def read_weights(self):

@tf.function()
def process_inner_loop(self, x, y, z, q, m):
if self.teacher:
teacher_outputs = self.teacher.model(x, training=False)
teacher_policy = teacher_outputs[0]
z = tf.nn.softmax(teacher_outputs[1])
m = teacher_outputs[2]
else:
teacher_policy = None

with tf.GradientTape() as tape:
outputs = self.model(x, training=True)
policy = outputs[0]
value = outputs[1]
policy_loss = self.policy_loss_fn(y, policy)
policy_loss = self.policy_loss_fn(y, policy, teacher_policy)
reg_term = sum(self.model.losses)
if self.wdl:
value_ce_loss = self.value_loss_fn(self.qMix(z, q), value)
Expand Down
15 changes: 15 additions & 0 deletions tf/train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -219,6 +219,17 @@ def main(cmd):

tfprocess.restore()

if cmd.teacher:
teacher_cfg = yaml.safe_load(cmd.teacher.read())
print("Creating Teacher")
teacher_tfprocess = TFProcess(teacher_cfg, is_teacher=True)
teacher_tfprocess.init_net(is_teacher=True)
teacher_tfprocess.restore()
tfprocess.set_teacher(teacher_tfprocess)
print("Teacher Created")
else:
tfprocess.set_teacher(None)

# If number of test positions is not given
# sweeps through all test chunks statistically
# Assumes average of 10 samples per test game.
Expand Down Expand Up @@ -253,6 +264,10 @@ def main(cmd):
type=str,
help='file to store weights in')

argparser.add_argument('--teacher',
type=argparse.FileType('r'),
help='yaml for teacher')

#mp.set_start_method('spawn')
main(argparser.parse_args())
mp.freeze_support()