def main(params):
# initialize the egg lib
opts = core.init(params=params)
# get pre-defined common line arguments (batch/vocab size, etc).
# See egg/core/util.py for a list
# prepare the dataset
kwargs = {"num_workers": 1, "pin_memory": True} if opts.cuda else {}
transform = transforms.ToTensor()
train_loader = torch.utils.data.DataLoader(datasets.MNIST(
"./data", train=True, download=True, transform=transform),
batch_size=opts.batch_size,
shuffle=True,
**kwargs)
test_loader = torch.utils.data.DataLoader(datasets.MNIST(
"./data", train=False, transform=transform),
batch_size=opts.batch_size,
shuffle=True,
**kwargs)
# initialize the agents and the game
sender = Sender(opts.vocab_size) # the "data" transform part of an agent
sender = core.GumbelSoftmaxWrapper(
sender, temperature=1.0) # wrapping into a GS interface
receiver = Receiver()
receiver = core.SymbolReceiverWrapper(receiver,
vocab_size=opts.vocab_size,
agent_input_size=400)
# setting up as a standard Sender/Receiver game with 1 symbol communication
game = core.SymbolGameGS(sender, receiver, loss)
# This callback would be called at the end of each epoch by the Trainer; it reduces the sampling
# temperature used by the GS
temperature_updater = core.TemperatureUpdater(agent=sender,
decay=0.75,
minimum=0.01)
# get an optimizer that is set up by common command line parameters,
# defaults to Adam
optimizer = core.build_optimizer(game.parameters())
# initialize and launch the trainer
trainer = core.Trainer(
game=game,
optimizer=optimizer,
train_data=train_loader,
validation_data=test_loader,
callbacks=[
temperature_updater,
core.ConsoleLogger(as_json=True, print_train_loss=True),
],
)
trainer.train(n_epochs=opts.n_epochs)
core.close()
def test_symbol_wrapper():
core.init()
receiver = core.SymbolReceiverWrapper(Receiver(), vocab_size=15, agent_input_size=5)
# when trained with REINFORCE, the message would be encoded as long ids
message_rf = torch.randint(high=15, size=(16,)).long()
output_rf = receiver(message_rf)
assert output_rf.size() == torch.Size((16, 5))
# when trained with Gumbel-Softmax, the message would be encoded as one-hots
message_gs = torch.zeros((16, 15))
message_gs.scatter_(1, message_rf.unsqueeze(1), 1.0) # same message, one-hot-encoded
output_gs = receiver(message_gs)
assert output_rf.eq(output_gs).all().item() == 1
def main(params):
opts = get_params(params)
print(opts)
device = opts.device
train_loader = OneHotLoader(n_bits=opts.n_bits,
bits_s=opts.bits_s,
bits_r=opts.bits_r,
batch_size=opts.batch_size,
batches_per_epoch=opts.n_examples_per_epoch /
opts.batch_size)
test_loader = UniformLoader(n_bits=opts.n_bits,
bits_s=opts.bits_s,
bits_r=opts.bits_r)
test_loader.batch = [x.to(device) for x in test_loader.batch]
if not opts.variable_length:
sender = Sender(n_bits=opts.n_bits,
n_hidden=opts.sender_hidden,
vocab_size=opts.vocab_size)
if opts.mode == 'gs':
sender = core.GumbelSoftmaxWrapper(agent=sender,
temperature=opts.temperature)
receiver = Receiver(n_bits=opts.n_bits,
n_hidden=opts.receiver_hidden)
receiver = core.SymbolReceiverWrapper(
receiver,
vocab_size=opts.vocab_size,
agent_input_size=opts.receiver_hidden)
game = core.SymbolGameGS(sender, receiver, diff_loss)
elif opts.mode == 'rf':
sender = core.ReinforceWrapper(agent=sender)
receiver = Receiver(n_bits=opts.n_bits,
n_hidden=opts.receiver_hidden)
receiver = core.SymbolReceiverWrapper(
receiver,
vocab_size=opts.vocab_size,
agent_input_size=opts.receiver_hidden)
receiver = core.ReinforceDeterministicWrapper(agent=receiver)
game = core.SymbolGameReinforce(
sender,
receiver,
diff_loss,
sender_entropy_coeff=opts.sender_entropy_coeff)
elif opts.mode == 'non_diff':
sender = core.ReinforceWrapper(agent=sender)
receiver = ReinforcedReceiver(n_bits=opts.n_bits,
n_hidden=opts.receiver_hidden)
receiver = core.SymbolReceiverWrapper(
receiver,
vocab_size=opts.vocab_size,
agent_input_size=opts.receiver_hidden)
game = core.SymbolGameReinforce(
sender,
receiver,
non_diff_loss,
sender_entropy_coeff=opts.sender_entropy_coeff,
receiver_entropy_coeff=opts.receiver_entropy_coeff)
else:
if opts.mode != 'rf':
print('Only mode=rf is supported atm')
opts.mode = 'rf'
if opts.sender_cell == 'transformer':
receiver = Receiver(n_bits=opts.n_bits,
n_hidden=opts.receiver_hidden)
sender = Sender(
n_bits=opts.n_bits,
n_hidden=opts.sender_hidden,
vocab_size=opts.sender_hidden) # TODO: not really vocab
sender = core.TransformerSenderReinforce(
agent=sender,
vocab_size=opts.vocab_size,
embed_dim=opts.sender_emb,
max_len=opts.max_len,
num_layers=1,
num_heads=1,
hidden_size=opts.sender_hidden)
else:
receiver = Receiver(n_bits=opts.n_bits,
n_hidden=opts.receiver_hidden)
sender = Sender(
n_bits=opts.n_bits,
n_hidden=opts.sender_hidden,
vocab_size=opts.sender_hidden) # TODO: not really vocab
sender = core.RnnSenderReinforce(agent=sender,
vocab_size=opts.vocab_size,
embed_dim=opts.sender_emb,
hidden_size=opts.sender_hidden,
max_len=opts.max_len,
cell=opts.sender_cell)
if opts.receiver_cell == 'transformer':
receiver = Receiver(n_bits=opts.n_bits, n_hidden=opts.receiver_emb)
receiver = core.TransformerReceiverDeterministic(
receiver,
opts.vocab_size,
opts.max_len,
opts.receiver_emb,
num_heads=1,
hidden_size=opts.receiver_hidden,
num_layers=1)
else:
receiver = Receiver(n_bits=opts.n_bits,
n_hidden=opts.receiver_hidden)
receiver = core.RnnReceiverDeterministic(receiver,
opts.vocab_size,
opts.receiver_emb,
opts.receiver_hidden,
cell=opts.receiver_cell)
game = core.SenderReceiverRnnGS(sender, receiver, diff_loss)
game = core.SenderReceiverRnnReinforce(
sender,
receiver,
diff_loss,
sender_entropy_coeff=opts.sender_entropy_coeff,
receiver_entropy_coeff=opts.receiver_entropy_coeff)
optimizer = torch.optim.Adam([
dict(params=sender.parameters(), lr=opts.sender_lr),
dict(params=receiver.parameters(), lr=opts.receiver_lr)
])
loss = game.loss
intervention = CallbackEvaluator(test_loader,
device=device,
is_gs=opts.mode == 'gs',
loss=loss,
var_length=opts.variable_length,
input_intervention=True)
trainer = core.Trainer(game=game,
optimizer=optimizer,
train_data=train_loader,
validation_data=test_loader,
callbacks=[
core.ConsoleLogger(as_json=True),
EarlyStopperAccuracy(opts.early_stopping_thr),
intervention
])
trainer.train(n_epochs=opts.n_epochs)
core.close()