def test_snapshoting(): CHECKPOINT_PATH = Path("./test_checkpoints") core.init() sender = core.GumbelSoftmaxWrapper(ToyAgent(), temperature=1) receiver = Receiver() loss = lambda sender_input, message, receiver_input, receiver_output, labels, aux_input: ( F.cross_entropy(receiver_output, labels), {}, ) game = core.SymbolGameGS(sender, receiver, loss) optimizer = torch.optim.Adam(game.parameters()) data = Dataset() trainer = core.Trainer( game, optimizer, train_data=data, validation_data=None, callbacks=[core.CheckpointSaver(checkpoint_path=CHECKPOINT_PATH)], ) trainer.train(2) assert (CHECKPOINT_PATH / Path("1.tar")).exists() assert (CHECKPOINT_PATH / Path("2.tar")).exists() assert (CHECKPOINT_PATH / Path("final.tar")).exists() shutil.rmtree(CHECKPOINT_PATH) # Clean-up """
def test_snapshoting(): CHECKPOINT_PATH = Path('./test_checkpoints') core.init() sender = core.GumbelSoftmaxWrapper(ToyAgent(), temperature=1) receiver = Receiver() loss = lambda sender_input, message, receiver_input, receiver_output, labels: \ (F.cross_entropy(receiver_output, labels), {}) game = core.SymbolGameGS(sender, receiver, loss) optimizer = torch.optim.Adam(game.parameters()) data = Dataset() trainer = core.Trainer( game, optimizer, train_data=data, validation_data=None, callbacks=[core.CheckpointSaver(checkpoint_path=CHECKPOINT_PATH)]) trainer.train(2) assert (CHECKPOINT_PATH / Path('1.tar')).exists() assert (CHECKPOINT_PATH / Path('2.tar')).exists() assert (CHECKPOINT_PATH / Path('final.tar')).exists() del trainer trainer = core.Trainer(game, optimizer, train_data=data) # Re-instantiate trainer trainer.load_from_latest(CHECKPOINT_PATH) assert trainer.start_epoch == 2 trainer.train(3) shutil.rmtree(CHECKPOINT_PATH) # Clean-up
def test_max_snapshoting(): CHECKPOINT_PATH = Path("./test_checkpoints") core.init() sender = core.GumbelSoftmaxWrapper(ToyAgent(), temperature=1) receiver = Receiver() loss = lambda sender_input, message, receiver_input, receiver_output, labels: ( F.cross_entropy(receiver_output, labels), {}, ) game = core.SymbolGameGS(sender, receiver, loss) optimizer = torch.optim.Adam(game.parameters()) data = Dataset() trainer = core.Trainer( game, optimizer, train_data=data, validation_data=None, callbacks=[ core.CheckpointSaver(checkpoint_path=CHECKPOINT_PATH, max_checkpoints=2) ], ) trainer.train(n_epochs=6) assert (CHECKPOINT_PATH / Path("5.tar")).exists() assert (CHECKPOINT_PATH / Path("6.tar")).exists() assert (CHECKPOINT_PATH / Path("final.tar")).exists() assert len([x for x in CHECKPOINT_PATH.glob("**/*") if x.is_file()]) == 3 del trainer trainer = core.Trainer(game, optimizer, train_data=data) # Re-instantiate trainer trainer.load_from_latest(CHECKPOINT_PATH) assert trainer.start_epoch == 6 trainer.train(3) shutil.rmtree(CHECKPOINT_PATH) # Clean-up
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() print('the probs are: ', probs, flush=True) 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) ################################# # define sender (speaker) agent # ################################# sender = Sender(n_features=opts.n_features, n_hidden=opts.sender_hidden) sender = 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, noise_loc=opts.sender_noise_loc, noise_scale=opts.sender_noise_scale) #################################### # define receiver (listener) agent # #################################### receiver = Receiver(n_features=opts.n_features, n_hidden=opts.receiver_hidden) receiver = RnnReceiverDeterministic(receiver, opts.vocab_size, opts.receiver_embedding, opts.receiver_hidden, cell=opts.receiver_cell, num_layers=opts.receiver_num_layers, noise_loc=opts.receiver_noise_loc, noise_scale=opts.receiver_noise_scale) ################### # define channel # ################### channel = Channel(vocab_size=opts.vocab_size, p=opts.channel_repl_prob) game = SenderReceiverRnnReinforce( sender, receiver, loss, sender_entropy_coeff=opts.sender_entropy_coeff, receiver_entropy_coeff=opts.receiver_entropy_coeff, length_cost=opts.length_cost, effective_max_len=opts.effective_max_len, channel=channel, sender_entropy_common_ratio=opts.sender_entropy_common_ratio) optimizer = core.build_optimizer(game.parameters()) callbacks = [ EarlyStopperAccuracy(opts.early_stopping_thr), core.ConsoleLogger(as_json=True, print_train_loss=True) ] if opts.checkpoint_dir: ''' info in checkpoint_name: - n_features as f - vocab_size as vocab - random_seed as rs - lr as lr - sender_hidden as shid - receiver_hidden as rhid - sender_entropy_coeff as sentr - length_cost as reg - max_len as max_len - sender_noise_scale as sscl - receiver_noise_scale as rscl - channel_repl_prob as crp - sender_entropy_common_ratio as scr ''' checkpoint_name = ( f'{opts.name}' + '_aer' + ('_uniform' if opts.probs == 'uniform' else '') + f'_f{opts.n_features}' + f'_vocab{opts.vocab_size}' + f'_rs{opts.random_seed}' + f'_lr{opts.lr}' + f'_shid{opts.sender_hidden}' + f'_rhid{opts.receiver_hidden}' + f'_sentr{opts.sender_entropy_coeff}' + f'_reg{opts.length_cost}' + f'_max_len{opts.max_len}' + f'_sscl{opts.sender_noise_scale}' + f'_rscl{opts.receiver_noise_scale}' + f'_crp{opts.channel_repl_prob}' + f'_scr{opts.sender_entropy_common_ratio}') callbacks.append( core.CheckpointSaver(checkpoint_path=opts.checkpoint_dir, checkpoint_freq=opts.checkpoint_freq, prefix=checkpoint_name)) trainer = core.Trainer(game=game, optimizer=optimizer, train_data=train_loader, validation_data=test_loader, callbacks=callbacks) trainer.train(n_epochs=opts.n_epochs) print('<div id="prefix test without eos">') prefix_test(trainer.game, opts.n_features, device, add_eos=False) print('</div>') print('<div id="prefix test with eos">') prefix_test(trainer.game, opts.n_features, device, add_eos=True) print('<div id="suffix test">') suffix_test(trainer.game, opts.n_features, device) print('</div>') print('<div id="replacement test">') replacement_test(trainer.game, opts.n_features, opts.vocab_size, device) print('</div>') print('<div id="dump">') dump(trainer.game, opts.n_features, device, False) print('</div>') 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() print('the probs are: ', probs, flush=True) 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) receiver = core.RnnReceiverDeterministic(receiver, opts.vocab_size, opts.receiver_embedding, opts.receiver_hidden, cell=opts.receiver_cell, num_layers=opts.receiver_num_layers) empty_logger = LoggingStrategy.minimal() game = core.SenderReceiverRnnReinforce(sender, receiver, loss, sender_entropy_coeff=opts.sender_entropy_coeff, receiver_entropy_coeff=opts.receiver_entropy_coeff, train_logging_strategy=empty_logger, length_cost=opts.length_cost) optimizer = core.build_optimizer(game.parameters()) callbacks = [EarlyStopperAccuracy(opts.early_stopping_thr), core.ConsoleLogger(as_json=True, print_train_loss=True)] if opts.checkpoint_dir: checkpoint_name = f'{opts.name}_vocab{opts.vocab_size}_rs{opts.random_seed}_lr{opts.lr}_shid{opts.sender_hidden}_rhid{opts.receiver_hidden}_sentr{opts.sender_entropy_coeff}_reg{opts.length_cost}_max_len{opts.max_len}' callbacks.append(core.CheckpointSaver(checkpoint_path=opts.checkpoint_dir, prefix=checkpoint_name)) trainer = core.Trainer(game=game, optimizer=optimizer, train_data=train_loader, validation_data=test_loader, callbacks=callbacks) trainer.train(n_epochs=opts.n_epochs) game.logging_strategy = LoggingStrategy.maximal() # now log everything dump(trainer.game, opts.n_features, device, False) core.close()
def main(params): opts = get_params(params) print(opts, flush=True) # For compatibility, after https://github.com/facebookresearch/EGG/pull/130 # the meaning of `length` changed a bit. Before it included the EOS symbol; now # it doesn't. To ensure that hyperparameters/CL arguments do not change, # we subtract it here. opts.max_len -= 1 device = opts.device 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() print("the probs are: ", probs, flush=True) 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, 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, ) 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) receiver = core.RnnReceiverDeterministic( receiver, opts.vocab_size, opts.receiver_embedding, opts.receiver_hidden, cell=opts.receiver_cell, num_layers=opts.receiver_num_layers, ) empty_logger = LoggingStrategy.minimal() game = core.SenderReceiverRnnReinforce( sender, receiver, loss, sender_entropy_coeff=opts.sender_entropy_coeff, receiver_entropy_coeff=opts.receiver_entropy_coeff, train_logging_strategy=empty_logger, length_cost=opts.length_cost, ) optimizer = core.build_optimizer(game.parameters()) callbacks = [ EarlyStopperAccuracy(opts.early_stopping_thr), core.ConsoleLogger(as_json=True, print_train_loss=True), ] if opts.checkpoint_dir: checkpoint_name = f"{opts.name}_vocab{opts.vocab_size}_rs{opts.random_seed}_lr{opts.lr}_shid{opts.sender_hidden}_rhid{opts.receiver_hidden}_sentr{opts.sender_entropy_coeff}_reg{opts.length_cost}_max_len{opts.max_len}" callbacks.append( core.CheckpointSaver( checkpoint_path=opts.checkpoint_dir, prefix=checkpoint_name ) ) trainer = core.Trainer( game=game, optimizer=optimizer, train_data=train_loader, validation_data=test_loader, callbacks=callbacks, ) trainer.train(n_epochs=opts.n_epochs) game.logging_strategy = LoggingStrategy.maximal() # now log everything dump(trainer.game, opts.n_features, device, False) core.close()