def main(params):
opts = get_params(params)
print(opts, flush=True)
device = opts.device
force_eos = opts.force_eos == 1
if opts.probs == 'uniform':
probs = np.ones(opts.n_features)
elif opts.probs == 'powerlaw':
probs = 1 / np.arange(1, opts.n_features + 1, dtype=np.float32)
else:
probs = np.array([float(x) for x in opts.probs.split(',')],
dtype=np.float32)
probs /= probs.sum()
train_loader = OneHotLoader(n_features=opts.n_features,
batch_size=opts.batch_size,
batches_per_epoch=opts.batches_per_epoch,
probs=probs)
# single batches with 1s on the diag
test_loader = UniformLoader(opts.n_features)
if opts.sender_cell == 'transformer':
sender = Sender(n_features=opts.n_features,
n_hidden=opts.sender_embedding)
sender = core.TransformerSenderReinforce(
agent=sender,
vocab_size=opts.vocab_size,
embed_dim=opts.sender_embedding,
max_len=opts.max_len,
num_layers=opts.sender_num_layers,
num_heads=opts.sender_num_heads,
hidden_size=opts.sender_hidden,
force_eos=opts.force_eos,
generate_style=opts.sender_generate_style,
causal=opts.causal_sender)
else:
sender = Sender(n_features=opts.n_features,
n_hidden=opts.sender_hidden)
sender = core.RnnSenderReinforce(sender,
opts.vocab_size,
opts.sender_embedding,
opts.sender_hidden,
cell=opts.sender_cell,
max_len=opts.max_len,
num_layers=opts.sender_num_layers,
force_eos=force_eos)
if opts.receiver_cell == 'transformer':
receiver = Receiver(n_features=opts.n_features,
n_hidden=opts.receiver_embedding)
receiver = core.TransformerReceiverDeterministic(
receiver,
opts.vocab_size,
opts.max_len,
opts.receiver_embedding,
opts.receiver_num_heads,
opts.receiver_hidden,
opts.receiver_num_layers,
causal=opts.causal_receiver)
else:
receiver = Receiver(n_features=opts.n_features,
n_hidden=opts.receiver_hidden)
if not opts.impatient:
receiver = Receiver(n_features=opts.n_features,
n_hidden=opts.receiver_hidden)
receiver = core.RnnReceiverDeterministic(
receiver,
opts.vocab_size,
opts.receiver_embedding,
opts.receiver_hidden,
cell=opts.receiver_cell,
num_layers=opts.receiver_num_layers)
else:
receiver = Receiver(n_features=opts.receiver_hidden,
n_hidden=opts.vocab_size)
# If impatient 1
receiver = RnnReceiverImpatient(
receiver,
opts.vocab_size,
opts.receiver_embedding,
opts.receiver_hidden,
cell=opts.receiver_cell,
num_layers=opts.receiver_num_layers,
max_len=opts.max_len,
n_features=opts.n_features)
sender.load_state_dict(
torch.load(opts.sender_weights, map_location=torch.device('cpu')))
receiver.load_state_dict(
torch.load(opts.receiver_weights, map_location=torch.device('cpu')))
if not opts.impatient:
game = core.SenderReceiverRnnReinforce(
sender,
receiver,
loss,
sender_entropy_coeff=opts.sender_entropy_coeff,
receiver_entropy_coeff=opts.receiver_entropy_coeff,
length_cost=opts.length_cost,
unigram_penalty=opts.unigram_pen)
else:
game = SenderImpatientReceiverRnnReinforce(
sender,
receiver,
loss,
sender_entropy_coeff=opts.sender_entropy_coeff,
receiver_entropy_coeff=opts.receiver_entropy_coeff,
length_cost=opts.length_cost,
unigram_penalty=opts.unigram_pen)
optimizer = core.build_optimizer(game.parameters())
trainer = core.Trainer(
game=game,
optimizer=optimizer,
train_data=train_loader,
validation_data=test_loader,
callbacks=[EarlyStopperAccuracy(opts.early_stopping_thr)])
# Debut test position
position_sieve = np.zeros((opts.n_features, opts.max_len))
for position in range(opts.max_len):
dataset = [[torch.eye(opts.n_features).to(device), None]]
if opts.impatient:
sender_inputs, messages, receiver_inputs, receiver_outputs, _ = \
dump_test_position_impatient(trainer.game,
dataset,
position=position,
voc_size=opts.vocab_size,
gs=False,
device=device,
variable_length=True)
else:
sender_inputs, messages, receiver_inputs, receiver_outputs, _ = \
dump_test_position(trainer.game,
dataset,
position=position,
voc_size=opts.vocab_size,
gs=False,
device=device,
variable_length=True)
acc_pos = []
for sender_input, message, receiver_output in zip(
sender_inputs, messages, receiver_outputs):
input_symbol = sender_input.argmax()
output_symbol = receiver_output.argmax()
acc = (input_symbol == output_symbol).float().item()
acc_pos.append(acc)
acc_pos = np.array(acc_pos)
position_sieve[:, position] = acc_pos
# Put -1 for position after message_length
_, messages = dump(trainer.game, opts.n_features, device, False)
# Convert messages to numpy array
messages_np = []
for x in messages:
x = x.cpu().numpy()
messages_np.append(x)
for i in range(len(messages_np)):
# Message i
message_i = messages_np[i]
id_0 = np.where(message_i == 0)[0]
if id_0.shape[0] > 0:
for j in range(id_0[0] + 1, opts.max_len):
position_sieve[i, j] = -1
np.save("analysis/position_sieve.npy", position_sieve)
core.close()
def main(params):
opts = get_params(params)
print(opts, flush=True)
device = opts.device
force_eos = opts.force_eos == 1
if opts.probs == 'uniform':
probs = np.ones(opts.n_features)
elif opts.probs == 'powerlaw':
probs = 1 / np.arange(1, opts.n_features+1, dtype=np.float32)
else:
probs = np.array([float(x) for x in opts.probs.split(',')], dtype=np.float32)
probs /= probs.sum()
train_loader = OneHotLoader(n_features=opts.n_features, batch_size=opts.batch_size,
batches_per_epoch=opts.batches_per_epoch, probs=probs)
# single batches with 1s on the diag
test_loader = UniformLoader(opts.n_features)
if opts.sender_cell == 'transformer':
sender = Sender(n_features=opts.n_features, n_hidden=opts.sender_embedding)
sender = core.TransformerSenderReinforce(agent=sender, vocab_size=opts.vocab_size,
embed_dim=opts.sender_embedding, max_len=opts.max_len,
num_layers=opts.sender_num_layers, num_heads=opts.sender_num_heads,
hidden_size=opts.sender_hidden,
force_eos=opts.force_eos,
generate_style=opts.sender_generate_style,
causal=opts.causal_sender)
else:
sender = Sender(n_features=opts.n_features, n_hidden=opts.sender_hidden)
sender = core.RnnSenderReinforce(sender,
opts.vocab_size, opts.sender_embedding, opts.sender_hidden,
cell=opts.sender_cell, max_len=opts.max_len, num_layers=opts.sender_num_layers,
force_eos=force_eos)
if opts.receiver_cell == 'transformer':
receiver = Receiver(n_features=opts.n_features, n_hidden=opts.receiver_embedding)
receiver = core.TransformerReceiverDeterministic(receiver, opts.vocab_size, opts.max_len,
opts.receiver_embedding, opts.receiver_num_heads, opts.receiver_hidden,
opts.receiver_num_layers, causal=opts.causal_receiver)
else:
receiver = Receiver(n_features=opts.n_features, n_hidden=opts.receiver_hidden)
if not opts.impatient:
receiver = Receiver(n_features=opts.n_features, n_hidden=opts.receiver_hidden)
receiver = core.RnnReceiverDeterministic(receiver, opts.vocab_size, opts.receiver_embedding,
opts.receiver_hidden, cell=opts.receiver_cell,
num_layers=opts.receiver_num_layers)
else:
receiver = Receiver(n_features=opts.receiver_hidden, n_hidden=opts.vocab_size)
# If impatient 1
receiver = RnnReceiverImpatient(receiver, opts.vocab_size, opts.receiver_embedding,
opts.receiver_hidden, cell=opts.receiver_cell,
num_layers=opts.receiver_num_layers, max_len=opts.max_len, n_features=opts.n_features)
# If impatient 2
#receiver = RnnReceiverImpatient2(receiver, opts.vocab_size, opts.receiver_embedding,
# opts.receiver_hidden, cell=opts.receiver_cell,
# num_layers=opts.receiver_num_layers, max_len=opts.max_len, n_features=opts.n_features)
sender.load_state_dict(torch.load(opts.sender_weights,map_location=torch.device('cpu')))
receiver.load_state_dict(torch.load(opts.receiver_weights,map_location=torch.device('cpu')))
if not opts.impatient:
game = core.SenderReceiverRnnReinforce(sender, receiver, loss, sender_entropy_coeff=opts.sender_entropy_coeff,
receiver_entropy_coeff=opts.receiver_entropy_coeff,
length_cost=opts.length_cost,unigram_penalty=opts.unigram_pen)
else:
game = SenderImpatientReceiverRnnReinforce(sender, receiver, loss, sender_entropy_coeff=opts.sender_entropy_coeff,
receiver_entropy_coeff=opts.receiver_entropy_coeff,
length_cost=opts.length_cost,unigram_penalty=opts.unigram_pen)
optimizer = core.build_optimizer(game.parameters())
trainer = core.Trainer(game=game, optimizer=optimizer, train_data=train_loader,
validation_data=test_loader, callbacks=[EarlyStopperAccuracy(opts.early_stopping_thr)])
# Test impose message
if not opts.impatient:
acc_vec,messages=dump(trainer.game, opts.n_features, device, False)
else:
acc_vec,messages=dump_impatient(trainer.game, opts.n_features, device, False,save_dir=opts.save_dir)
all_messages=[]
for x in messages:
x = x.cpu().numpy()
all_messages.append(x)
all_messages = np.asarray(all_messages)
messages=-1*np.ones((opts.n_features,opts.max_len))
for i in range(len(all_messages)):
for j in range(all_messages[i].shape[0]):
messages[i,j]=all_messages[i][j]
np.save(opts.save_dir+"messages_analysis.npy",messages)
core.close()
def main(params):
opts = get_params(params)
print(opts, flush=True)
device = opts.device
force_eos = opts.force_eos == 1
if opts.probs == 'uniform':
probs = np.ones(opts.n_features)
elif opts.probs == 'powerlaw':
probs = 1 / np.arange(1, opts.n_features+1, dtype=np.float32)
else:
probs = np.array([float(x) for x in opts.probs.split(',')], dtype=np.float32)
probs /= probs.sum()
train_loader = OneHotLoader(n_features=opts.n_values, batch_size=opts.batch_size*opts.n_attributes,
batches_per_epoch=opts.batches_per_epoch, probs=probs)
# single batches with 1s on the diag
test_loader = UniformLoader(opts.n_values)
### SENDER ###
sender = Sender(n_features=opts.n_attributes*opts.n_values, n_hidden=opts.sender_hidden)
sender = core.RnnSenderReinforce(sender,opts.vocab_size, opts.sender_embedding, opts.sender_hidden,
cell=opts.sender_cell, max_len=opts.max_len, num_layers=opts.sender_num_layers,
force_eos=force_eos)
### RECEIVER ###
receiver = Receiver(n_features=opts.n_values, n_hidden=opts.receiver_hidden)
if not opts.impatient:
receiver = Receiver(n_features=opts.n_features, n_hidden=opts.receiver_hidden)
receiver = RnnReceiverCompositionality(receiver, opts.vocab_size, opts.receiver_embedding,
opts.receiver_hidden, cell=opts.receiver_cell,
num_layers=opts.receiver_num_layers, max_len=opts.max_len, n_attributes=opts.n_attributes, n_values=opts.n_values)
else:
receiver = Receiver(n_features=opts.receiver_hidden, n_hidden=opts.vocab_size)
# If impatient 1
receiver = RnnReceiverImpatientCompositionality(receiver, opts.vocab_size, opts.receiver_embedding,
opts.receiver_hidden, cell=opts.receiver_cell,
num_layers=opts.receiver_num_layers, max_len=opts.max_len, n_attributes=opts.n_attributes, n_values=opts.n_values)
sender.load_state_dict(torch.load(opts.sender_weights,map_location=torch.device('cpu')))
receiver.load_state_dict(torch.load(opts.receiver_weights,map_location=torch.device('cpu')))
if not opts.impatient:
game = CompositionalitySenderReceiverRnnReinforce(sender, receiver, loss_compositionality, sender_entropy_coeff=opts.sender_entropy_coeff,
n_attributes=opts.n_attributes,n_values=opts.n_values,att_weights=[1],receiver_entropy_coeff=opts.receiver_entropy_coeff,
length_cost=opts.length_cost,unigram_penalty=opts.unigram_pen,reg=opts.reg)
else:
game = CompositionalitySenderImpatientReceiverRnnReinforce(sender, receiver, loss_impatient_compositionality, sender_entropy_coeff=opts.sender_entropy_coeff,
n_attributes=opts.n_attributes,n_values=opts.n_values,att_weights=[1],receiver_entropy_coeff=opts.receiver_entropy_coeff,
length_cost=opts.length_cost,unigram_penalty=opts.unigram_pen,reg=opts.reg)
optimizer = core.build_optimizer(game.parameters())
trainer = CompoTrainer(n_attributes=opts.n_attributes,n_values=opts.n_values,game=game, optimizer=optimizer, train_data=train_loader,
validation_data=test_loader, callbacks=[EarlyStopperAccuracy(opts.early_stopping_thr)])
# Debut test position
position_sieve=np.zeros((opts.n_attributes**opts.n_values,opts.max_len,opts.n_attributes))
for position in range(opts.max_len):
one_hots = torch.eye(opts.n_values)
val=np.arange(opts.n_values)
combination=list(itertools.product(val,repeat=opts.n_attributes))
dataset=[]
for i in range(len(combination)):
new_input=torch.zeros(0)
for j in combination[i]:
new_input=torch.cat((new_input,one_hots[j]))
dataset.append(new_input)
dataset=torch.stack(dataset)
dataset=[[dataset,None]]
if opts.impatient:
sender_inputs, messages, receiver_inputs, receiver_outputs, _ = \
dump_test_position_impatient_compositionality(trainer.game,
dataset,
position=position,
voc_size=opts.vocab_size,
gs=False,
device=device,
variable_length=True)
else:
sender_inputs, messages, receiver_inputs, receiver_outputs, _ = \
dump_test_position_compositionality(trainer.game,
dataset,
position=position,
voc_size=opts.vocab_size,
gs=False,
device=device,
variable_length=True)
for i in range(len(receiver_outputs)):
message=messages[i]
correct=True
for j in range(len(list(combination[i]))):
if receiver_outputs[i][j]==list(combination[i])[j]:
position_sieve[i,position,j]=1
# Put -1 for position after message_length
if not opts.impatient:
acc_vec,messages=dump_compositionality(trainer.game, opts.n_attributes, opts.n_values, device, False,0)
else:
acc_vec,messages=dump_impatient_compositionality(trainer.game, opts.n_attributes, opts.n_values, device, False,0)
# Convert messages to numpy array
messages_np=[]
for x in messages:
x = x.cpu().numpy()
messages_np.append(x)
for i in range(len(messages_np)):
# Message i
message_i=messages_np[i]
id_0=np.where(message_i==0)[0]
if id_0.shape[0]>0:
for j in range(id_0[0]+1,opts.max_len):
position_sieve[i,j]=-1
np.save("analysis/position_sieve.npy",position_sieve)
core.close()