def multitrainer(trainers, keep_on_gpu=True, parallel=False, cuda_pool=(0, )):
"""Train multiple models in parallel.
Parameters
----------
trainers : sequence[ModelTrainer]
keep_on_gpu : bool, default=True
Keep all models on GPU (risk of out-of-memory)
parallel : bool or int, default=False
Train model in parallel.
If an int, only this number of models will be trained in parallel.
cuda_pool : sequence[int], default=[0]
IDs of GPUs that can be used to dispatch the models.
"""
n = len(trainers)
initial_epoch = max(min(trainer.initial_epoch for trainer in trainers), 1)
nb_epoch = max(trainer.nb_epoch for trainer in trainers)
trainers = tuple(trainers)
if parallel:
parallel = len(trainers) if parallel is True else parallel
chunksize = max(len(trainers) // (2 * parallel), 1)
pool = Pool(parallel)
cuda_pool = py.make_list(pool, parallel)
if not torch.cuda.is_available():
cuda_pool = []
if not keep_on_gpu:
for trainer in trainers:
trainer.to(device='cpu')
# --- init ---
if parallel:
args = zip(trainers, [keep_on_gpu] * n, [cuda_pool] * n)
trainers = list(pool.map(_init1, args, chunksize=chunksize))
else:
args = zip(trainers, [keep_on_gpu] * n, [cuda_pool] * n)
trainers = list(map(_init1, args))
# --- train ---
for epoch in range(initial_epoch + 1, nb_epoch + 1):
if parallel:
args = zip(trainers, [epoch] * n, [keep_on_gpu] * n,
[cuda_pool] * n)
trainers = list(pool.map(_train1, args, chunksize=chunksize))
else:
args = zip(trainers, [epoch] * n, [keep_on_gpu] * n,
[cuda_pool] * n)
trainers = list(map(_train1, args))
def runInParallel(args):
cur_args = []
for i in range(args.r):
cur_args.append({'seed': base_seed + i})
print('total to run', cur_args, 'nProc', args.nproc)
if args.nproc > 1:
pool = Pool(processes=int(args.nproc))
pool.map(single_run_algorithm, cur_args)
pool.close()
pool.join()
else:
for cur_arg in cur_args:
single_run_algorithm(cur_arg)
def dsrg_layer(targets, labels, probs_ori, num_classes, thre_fg, thre_bg,
NUM_WORKERS):
targets = torch.reshape(targets, (-1, 1, 1, num_classes))
labels = torch.transpose(labels, 1, 3) # bx41x41x21
probs = torch.transpose(probs_ori, 1, 3) # bx41x41x21
targets[:, :, :, 0] = 1
# probs = probs.clone().detach()
batch_size = targets.shape[0]
complete_para_list = []
for i in range(batch_size):
params_list = []
params_list.append([targets[i], labels[i], probs[i]])
params_list.append([thre_fg, thre_bg])
complete_para_list.append(params_list)
ret = pool.map(single_generate_seed_step, complete_para_list)
ret = []
# for i in range(batch_size):
# ret.append(single_generate_seed_step(complete_para_list[i]))
new_labels = ret[0]
for i in range(1, batch_size):
new_labels = torch.cat([new_labels, ret[i]], axis=0)
new_labels = torch.transpose(new_labels, 1, 3) # bx21x41x41
return new_labels
def step(self, items_seq, items_ori, items_batch=None, boxes_batch=None):
if (items_batch != None) & (boxes_batch != None):
self.items_batch = items_batch
self.boxes_batch = boxes_batch
self.batch_indx = list(range(self.BATCH_SIZE))
self.expected_items_n = [self.ITEMS_SEQ_LN] * BATCH_SIZE
self.all_outs = {i: [] for i in range(self.BATCH_SIZE)}
self.current_level = 0
self.items_batch_alligned = None
self.boxes_batch_alligned = None
items_seq_ = torch.LongTensor(items_seq).transpose(1, 0).expand(
self.INPUT_SIZE, self.ITEMS_SEQ_LN,
self.BATCH_SIZE).transpose(2, 0)
items_ori_ = items_ori[torch.arange(self.BATCH_SIZE).expand(
self.ITEMS_SEQ_LN, self.BATCH_SIZE).transpose(1, 0),
torch.LongTensor(items_seq).
expand(self.BATCH_SIZE, self.ITEMS_SEQ_LN)]
self.items_batch_alligned = self.items_batch[[
self.base_indx_items, items_seq_, items_ori_
]]
lookup_sm = self.boxes_batch.expand(
self.ITEMS_SEQ_LN,
self.BATCH_SIZE, self.ITEMS_SEQ_LN, self.INPUT_SIZE).transpose(
1, 0) - self.items_batch_alligned.unsqueeze(2)
validities = (lookup_sm >= 0).all(3).any(2).tolist()
new_seq = []
for i, j in zip(items_seq, validities):
new_seq.append([i[k] for k in range(len(i)) if j[k] == True])
self.batch_indx = [i for i in self.batch_indx if len(new_seq[i]) > 0]
items_seq = [i for i in new_seq if len(i) > 0]
zp = list(
zip(self.batch_indx, self.items_batch[self.batch_indx],
self.boxes_batch[self.batch_indx], items_seq,
items_ori[self.batch_indx]))
p = Pool(10)
out = p.map(self.target_func, zp)
p.close()
p.join()
out = [pickle.loads(i) for i in out]
out_series = pd.Series(out)
_ = out_series.apply(lambda x: self.dict_update(x))
# out = [i for i in out if i[1] < i[2]]
self.batch_indx = [i[0] for i in out]
self.current_level += 1
items_seq = [i[5] for i in out]
all_rewards = [i[-1] * i[-2] for i in out]
# filled_items_indx = {i:[i[2] for i in j] for i,j in self.all_outs.items() if len(j) > 0}
# filled_items_HUs = {i:[i[7] for i in j if len(i[7]) > 0] for i,j in self.all_outs.items()}
# all_rewards = [self.calc_reward(self.items_batch[i],i,filled_items_indx,filled_items_HUs) for i in range(self.BATCH_SIZE)]
return all_rewards
def run_main():
deck_lists = list(map(int,args.decklists.split(","))) if args.decklists is not None else None
if deck_lists is None:
deck_lists = list(deck_id_2_name.keys())
else:
assert all(key in deck_id_2_name for key in deck_lists)
if deck_lists == [0, 1, 4, 5, 10, 12]:
deck_lists = [0, 1, 4, 12, 5, 10]
mylogger.info("deck_lists:{}".format(deck_lists))
D = [Deck() for i in range(len(deck_lists))]
deck_index = 0
# sorted_keys = sorted(list(deck_id_2_name.keys()))
# for i in sorted_keys:
# if i not in deck_lists:
# continue
for i in deck_lists:
mylogger.info("{}(deck_id:{}):{}".format(deck_index, i, key_2_tsv_name[i]))
D[deck_index] = tsv_to_deck(key_2_tsv_name[i][0])
D[deck_index].set_leader_class(key_2_tsv_name[i][1])
deck_index += 1
Results = {}
list_range = range(len(deck_lists))
#print(list(itertools.product(list_range,list_range)))
Player1 = Player(9, True, policy=New_Dual_NN_Non_Rollout_OM_ISMCTSPolicy(model_name=args.model_name),
mulligan=Min_cost_mulligan_policy())
if args.opponent is not None:
if args.model_name is not None:
if args.opponent == "Greedy":
Player2 = Player(9, True, policy=NN_GreedyPolicy(model_name=args.model_name),
mulligan=Min_cost_mulligan_policy())
elif args.opponent == "MCTS":
Player2 = Player(9, True, policy=New_Dual_NN_Non_Rollout_OM_ISMCTSPolicy(model_name=args.model_name),
mulligan=Min_cost_mulligan_policy())
else:
Player2 = Player(9, True, policy=Opponent_Modeling_MCTSPolicy(),
mulligan=Min_cost_mulligan_policy())
else:
Player2 = Player(9, True, policy=AggroPolicy(),
mulligan=Min_cost_mulligan_policy())
Player1.name = "Alice"
Player2.name = "Bob"
iteration = int(args.iteration) if args.iteration is not None else 10
deck_list_len = len(deck_lists)
iter_data = [(deck_list_len*i+j,Player1,
Player2,(i,j),(deck_lists[i],deck_lists[j]),iteration) for i,j in itertools.product(list_range,list_range)]
pool = Pool(3) # 最大プロセス数:8
# memory = pool.map(preparation, iter_data)
result = pool.map(multi_battle, iter_data)
#result = list(tqdm(result, total=len(list_range)**2))
pool.close() # add this.
pool.terminate() # add this.
for data in result:
Results[data[0]] = data[1]
print(Results)
def _invoke_pool(self, pool: TorchPool, fn: Callable, data: iter) -> \
List[int]:
"""Invoke on a torch pool (rather than a :class:`multiprocessing.Pool`).
"""
if pool is None:
return tuple(map(fn, data))
else:
return pool.map(fn, data)
def sample(self, observations, num_samples, thetas=None):
assert (thetas is None or len(thetas) is self.chains)
self.sampler.reset()
if thetas is None:
inputs = self._prepare_inputs()
pool = Pool(processes=self.workers)
arguments = self._prepare_arguments(observations, inputs, num_samples)
chains = pool.map(self.sample_chain, arguments)
del pool
return chains
def final_experiments():
for i in range(2, 5):
x_train, train, x_valid, valid, x_test, test, name = get_dataset(i)
print("Running for", name, "dataset")
x_train, e_train, t_train, x_valid, e_valid, t_valid, x_test, e_test, t_test = scale_data_to_torch(
x_train, train, x_valid, valid, x_test, test)
print("Dataset loaded and scaled")
risk_set, risk_set_valid, risk_set_test = compute_risk_set(
t_train, t_valid, t_test)
print("Risk set computed")
data_dict = {
"x_train": x_train,
"e_train": e_train,
"t_train": t_train,
"x_valid": x_valid,
"e_valid": e_valid,
"t_valid": t_valid,
"x_test": x_test,
"e_test": e_test,
"t_test": t_test,
"risk_set": risk_set,
"risk_set_valid": risk_set_valid,
"risk_set_test": risk_set_test
}
n_in = x_train.shape[1]
linear_models = [2, 5, 10, 12]
learning_rates = [1e-4, 1e-3]
layer_sizes = [[n_in], [n_in, n_in], [n_in, n_in, n_in],
[n_in, 20, 15]]
data = [data_dict]
hyperparams = [(linear_model, learning_rate, layer_size, seed, d)
for layer_size in layer_sizes
for learning_rate in learning_rates
for linear_model in linear_models for seed in range(3)
for d in data]
print("Hyperparams initialized")
p = Pool(50)
print("Pool created")
output = p.map(run_experiment, hyperparams)
p.close()
p.join()
print("Models trained. Writing to file")
filename = name + "_results.pkl"
f = open(filename, "wb")
pkl.dump(output, f)
f.flush()
f.close()
print(name, "done")
print("")
def make_dataset_XNAT(project, opt, phase):
session = xnat.connect('http://rufus.stanford.edu', user='******', password='******') #make XNAT connection
#xnat version of make_dataset_dicom()
images = []
#parsing xnat hierarchy: project -> subject -> experiment -> scan -> dicom img
subjs = session.projects[project].subjects
try:
set_start_method('spawn')
except RuntimeError:
pass
#multiprocessing code (not working) ------------
pool = Pool(os.cpu_count())
images = pool.map(process_subj, [subjs[s] for s in subjs])
images = [i for im in images for i in im]
#-----------------------------------------------
#original code (before multiproc, works fine) -------------------
'''
for s in subjs:
exps = subjs[s].experiments
for e in exps:
scans = exps[e].scans
for sc in scans:
my_file = scans[sc].resources['DICOM'].files[0]
fname = my_file.data['Name']
path = my_file.uri
with my_file.open() as f:
dicom = pydicom.read_file(f)
#save images whose filenames are in phase
if fname.endswith('.dcm'):
try:
opt.names_phase[fname]
except:
print('Error: we donot find the phase for file name' , fname, 'ignore this file')
continue
# sys.exit(3)
if opt.names_phase[fname] == phase:
Img = process_dicom(dicom)
if Img is not None:
print("x" + fname)
images.append([path, dicom]) #each image list item has filename + FileData obj
'''
#-------------------------------------------------
session.disconnect()
return images
def forward(
self,
images,
):
parameters = []
num_workers = self.num_workers
pool = Pool(num_workers)
for bx in range(len(images)):
bx_params = [bx, images[bx], True]
parameters.append(bx_params, )
predictions = pool.map(worker_distance_transform, parameters)
predictions = torch.stack(predictions)
pool.close()
pool.join()
return predictions
class ParallelApproximateBayesianComputation:
def __init__(self, abc, workers=2):
super(ParallelApproximateBayesianComputation, self).__init__()
self.abc = abc
self.pool = Pool(processes=workers)
self.workers = workers
def _prepare_arguments(self, observation, num_samples):
arguments = []
inputs = torch.arange(num_samples)
num_chunks = num_samples // self.workers
if num_chunks == 0:
num_chunks = 1
chunks = inputs.split(num_chunks, dim=0)
for chunk in chunks:
a = (self.abc, observation, len(chunk))
arguments.append(a)
return arguments
def sample(self, observation, num_samples=1):
arguments = self._prepare_arguments(observation, num_samples)
outputs = self.pool.map(self._sample, arguments)
outputs = torch.cat(outputs, dim=0)
return outputs
def __del__(self):
self.pool.close()
del self.pool
self.pool = None
@staticmethod
def _sample(arguments):
abc, observation, n = arguments
return abc.sample(observation, num_samples=n)
class ParallelSimulator(Simulator):
def __init__(self, simulator, workers=2):
super(ParallelSimulator, self).__init__()
self.pool = Pool(processes=workers)
self.simulator = simulator
self.workers = workers
def _prepare_arguments(self, inputs):
arguments = []
chunks = inputs.shape[0] // self.workers
if chunks == 0:
chunks = 1
chunks = inputs.split(chunks, dim=0)
for chunk in chunks:
a = (self.simulator, chunk)
arguments.append(a)
return arguments
def forward(self, inputs):
arguments = self._prepare_arguments(inputs)
outputs = self.pool.map(self._simulate, arguments)
outputs = torch.cat(outputs, dim=0)
return outputs
def terminate(self):
self.pool.close()
del self.pool
self.pool = None
self.simulator.terminate()
@staticmethod
def _simulate(arguments):
simulator, inputs = arguments
return simulator(inputs)
def fitness_grads(
self,
n_samples: int,
pool: Pool = None,
fitness_shaping_fn: Callable[[Iterable[float]], Iterable[float]] = lambda x: x
):
"""
Computes the (approximate) gradients of the expected fitness of the population.
Uses torch autodiff to compute the gradients. The Individual.fitness does NOT need to be differentiable,
but the log probability computations in Population.sample MUST be.
:param n_samples: How many individuals to sample to approximate the gradient
:param pool: Optional process pool to use when computing the fitness of the sampled individuals.
:param fitness_shaping_fn: Optional function to modify the fitness, e.g. normalization, etc. Input is a list of n raw fitness floats. Output must also be n floats.
:return: A (n,) tensor containing the raw fitness (before fitness_shaping_fn) for the n individuals.
"""
samples = self.sample(n_samples) # Generator
individuals = []
grads = []
for individual, log_prob in samples: # Compute gradients one at a time so only one log prob computational graph needs to be kept in memory at a time.
assert log_prob.ndim == 0 and log_prob.isfinite() and log_prob.grad_fn is not None, "log_probs must be differentiable finite scalars"
individuals.append(individual)
grads.append(t.autograd.grad(log_prob, self.parameters()))
if pool is not None:
raw_fitness = pool.map(_fitness_fn_no_grad, individuals)
else:
raw_fitness = list(map(_fitness_fn_no_grad, individuals))
fitness = fitness_shaping_fn(raw_fitness)
for i, p in enumerate(self.parameters()):
p.grad = -t.mean(t.stack([ind_fitness * grad[i] for grad, ind_fitness in zip(grads, fitness)]), dim=0)
return t.tensor(raw_fitness)
def fitness_grads(
self,
n_samples: int,
pool: Pool = None,
fitness_shaping_fn: Callable[[Iterable[float]],
Iterable[float]] = lambda x: x):
"""
Computes the (approximate) gradients of the expected fitness of the population.
Uses torch autodiff to compute the gradients. The Individual.fitness does NOT need to be differentiable,
but the log probability computations in Population.sample MUST be.
:param n_samples: How many individuals to sample to approximate the gradient
:param pool: Optional process pool to use when computing the fitness of the sampled individuals.
:param fitness_shaping_fn: Optional function to modify the fitness, e.g. normalization, etc. Input is a list of n raw fitness floats. Output must also be n floats.
:return: A (n,) tensor containing the raw fitness (before fitness_shaping_fn) for the n individuals.
"""
individuals, log_probs = zip(*self.sample(n_samples))
assert all(
lp.ndim == 0 and lp.isfinite() and lp.grad_fn is not None for lp in
log_probs), "log_probs must be differentiable finite scalars"
if pool is not None:
raw_fitness = pool.map(_fitness_fn_no_grad, individuals)
else:
raw_fitness = list(map(_fitness_fn_no_grad, individuals))
fitness = fitness_shaping_fn(raw_fitness)
t.mean(
t.stack([(-ind_fitness * log_prob)
for log_prob, ind_fitness in zip(log_probs, fitness)
])).backward()
return t.tensor(raw_fitness)
def main(input_fname, output_dir):
# open input file, read one filename per line
img_fnames = []
with open(input_fname, 'r') as ifd:
for line in ifd:
img_fname = line.strip()
# handle CSVs where the first column is the filename.
# This will obviously break if the filenames contain commas.
img_fname = img_fname.split(',')[0]
img_fnames.append(img_fname)
print('Read %d image filenames' % len(img_fnames))
chunk_size = 128
chunks = [
img_fnames[i:i + chunk_size]
for i in range(0, len(img_fnames), chunk_size)
]
if num_threads > 1:
pool = Pool(num_threads)
_ = pool.map(process_chunk, chunks)
else:
for chunk in chunks:
process_chunk(chunk)
print('done.')
class CTC(nn.Module):
def __init__(self, blank=0):
self.blank = blank
self.pool = Pool(8)
def create_target_p(self, target):
target_p = [self.blank]
for c in target:
target_p.append(c)
target_p.append(self.blank)
return target_p
def empty_dp(self, T, target_p):
return [[torch.tensor(-1.0).float() for _ in range(len(target_p))]
for _ in range(T)]
def compute_probs_dp(self, log_prob, dp, target_p):
ninf = torch.log(torch.from_numpy(np.array(0.0)))
for t in range(len(dp)):
for s in range(len(dp[0])):
if t == 0:
if s == 0:
res = log_prob[0, self.blank]
elif s == 1:
res = log_prob[0, target_p[1]]
else:
res = ninf
dp[t][s] = res
continue
a = dp[t - 1][s].float()
b = dp[t - 1][s - 1].float() if s - 1 >= 0 else ninf
# in case of a blank or a repeated label, we only consider s and s-1 at t-1, so we're done
if target_p[s] == self.blank or (s >= 2 and target_p[s - 2]
== target_p[s]):
if a == ninf and b == ninf: res = ninf
else:
res = max(a, b) + torch.log(1 + torch.exp(
-torch.abs(a - b))) + log_prob[t, target_p[s]]
dp[t][s] = res
continue
# otherwise, in case of a non-blank and non-repeated label, we additionally add s-2 at t-1
c = dp[t - 1][s - 2] if s - 2 >= 0 else ninf
m = max([a, b, c])
if a == ninf and b == ninf and c == ninf: res = ninf
else:
res = m + torch.log(
torch.exp(-torch.abs(a - m)) +
torch.exp(-torch.abs(b - m)) +
torch.exp(-torch.abs(c - m))) + log_prob[t,
target_p[s]]
dp[t][s] = res
return dp
def compute_instance_loss(self, inps):
log_prob, target = inps
T = log_prob.shape[0]
target_p = self.create_target_p(target)
dp = self.empty_dp(T, target_p)
log_prob.requires_grad = True
dp = self.compute_probs_dp(log_prob, dp, target_p)
a, b = dp[T - 1][len(target_p) - 1], dp[T - 1][len(target_p) - 2]
loss = max(a, b) + torch.log(1 + torch.exp(-torch.abs(a - b)))
loss.backward()
return log_prob.grad
def compute_loss(self, log_probs, targets, input_lengths, target_lengths):
grads = self.pool.map(self.compute_instance_loss,
[(log_probs[:input_lengths[i], i, :].detach(),
targets[i][:target_lengths[i]])
for i in range(log_probs.shape[1])])
dlog_probs = torch.stack(grads).permute((1, 0, 2))
loss = (dlog_probs * log_probs).sum() / len(grads) / -4
return loss
def __call__(self,
log_probs,
targets,
input_lengths,
target_lengths,
forget_rate,
w=None):
return self.compute_loss(log_probs, targets, input_lengths,
target_lengths)
def __getstate__(self):
self_dict = self.__dict__.copy()
del self_dict['pool']
return self_dict
def __setstate__(self, state):
self.__dict__.update(state)
class HyperOptArgumentParser(ArgumentParser):
"""
Subclass of argparse ArgumentParser which adds optional calls to sample from lists or ranges
Also enables running optimizations across parallel processes
"""
# these are commands injected by test tube from cluster operations
TRIGGER_CMD = 'test_tube_from_cluster_hopt'
SLURM_CMD_PATH = 'test_tube_slurm_cmd_path'
SLURM_EXP_CMD = 'hpc_exp_number'
SLURM_LOAD_CMD = 'test_tube_do_checkpoint_load'
CMD_MAP = {
TRIGGER_CMD: bool,
SLURM_CMD_PATH: str,
SLURM_EXP_CMD: int,
SLURM_LOAD_CMD: bool
}
def __init__(self, strategy='grid_search', **kwargs):
"""
:param strategy: 'grid_search', 'random_search'
:param enabled:
:param experiment:
:param kwargs:
"""
ArgumentParser.__init__(self, **kwargs)
self.strategy = strategy
self.trials = []
self.parsed_args = None
self.opt_args = {}
self.json_config_arg_name = None
self.pool = None
def __getstate__(self):
# capture what is normally pickled
state = self.__dict__.copy()
# remove all functions from the namespace
clean_state = {}
for k, v in state.items():
if not hasattr(v, '__call__'):
clean_state[k] = v
# what we return here will be stored in the pickle
return clean_state
def __setstate__(self, newstate):
# re-instate our __dict__ state from the pickled state
self.__dict__.update(newstate)
def add_argument(self, *args, **kwargs):
super(HyperOptArgumentParser, self).add_argument(*args, **kwargs)
def opt_list(self, *args, **kwargs):
options = kwargs.pop("options", None)
tunable = kwargs.pop("tunable", False)
self.add_argument(*args, **kwargs)
for i in range(len(args)):
arg_name = args[i]
self.opt_args[arg_name] = OptArg(obj_id=arg_name, opt_values=options, tunable=tunable)
def opt_range(
self,
*args,
**kwargs
):
low = kwargs.pop("low", None)
high = kwargs.pop("high", None)
arg_type = kwargs["type"]
nb_samples = kwargs.pop("nb_samples", 10)
tunable = kwargs.pop("tunable", False)
log_base = kwargs.pop("log_base", None)
self.add_argument(*args, **kwargs)
arg_name = args[-1]
self.opt_args[arg_name] = OptArg(
obj_id=arg_name,
opt_values=[low, high],
arg_type=arg_type,
nb_samples=nb_samples,
tunable=tunable,
log_base=log_base,
)
def json_config(self, *args, **kwargs):
self.add_argument(*args, **kwargs)
self.json_config_arg_name = re.sub('-', '', args[-1])
def __parse_args(self, args=None, namespace=None):
# allow bypassing certain missing params which other parts of test tube may introduce
args, argv = self.parse_known_args(args, namespace)
args, argv = self.__whitelist_cluster_commands(args, argv)
if argv:
msg = _('unrecognized arguments: %s')
self.error(msg % ' '.join(argv))
return args
def __whitelist_cluster_commands(self, args, argv):
parsed = {}
# build a dict where key = arg, value = value of the arg or None if just a flag
for i, arg_candidate in enumerate(argv):
arg = None
value = None
# only look at --keys
if '--' not in arg_candidate:
continue
# skip items not on the white list
if arg_candidate[2:] not in HyperOptArgumentParser.CMD_MAP:
continue
arg = arg_candidate[2:]
# pull out the value of the argument if given
if i + 1 <= len(argv) - 1:
if '--' not in argv[i + 1]:
value = argv[i + 1]
if arg is not None:
parsed[arg] = value
else:
if arg is not None:
parsed[arg] = value
# add the whitelist cmds to the args
all_values = set()
for k, v in args.__dict__.items():
all_values.add(k)
all_values.add(v)
for arg, v in parsed.items():
v_parsed = self.__parse_primitive_arg_val(v)
all_values.add(v)
all_values.add(arg)
args.__setattr__(arg, v_parsed)
# make list with only the unknown args
unk_args = []
for arg in argv:
arg_candidate = re.sub('--', '', arg)
is_bool = arg_candidate == 'True' or arg_candidate == 'False'
if is_bool: continue
if arg_candidate not in all_values:
unk_args.append(arg)
# when no bad args are left, return none to be consistent with super api
if len(unk_args) == 0:
unk_args = None
# add hpc_exp_number if not passed in so we can never get None
if HyperOptArgumentParser.SLURM_EXP_CMD not in args:
args.__setattr__(HyperOptArgumentParser.SLURM_EXP_CMD, None)
return args, unk_args
def __parse_primitive_arg_val(self, val):
if val is None:
return True
try:
return int(val)
except ValueError:
try:
return float(val)
except ValueError:
return val
def parse_args(self, args=None, namespace=None):
# call superclass arg first
results = self.__parse_args(args, namespace)
# extract vals
old_args = vars(results)
# override with json args if given
if self.json_config_arg_name and old_args[self.json_config_arg_name]:
for arg, v in self.__read_json_config(old_args[self.json_config_arg_name]).items():
old_args[arg] = v
# track args
self.parsed_args = deepcopy(old_args)
# attach optimization fx
old_args['trials'] = self.opt_trials
old_args['optimize_parallel'] = self.optimize_parallel
old_args['optimize_parallel_gpu'] = self.optimize_parallel_gpu
old_args['optimize_parallel_cpu'] = self.optimize_parallel_cpu
old_args['generate_trials'] = self.generate_trials
old_args['optimize_trials_parallel_gpu'] = self.optimize_trials_parallel_gpu
return TTNamespace(**old_args)
def __read_json_config(self, file_path):
with open(file_path) as json_data:
json_args = json.load(json_data)
return json_args
def opt_trials(self, num):
self.trials = strategies.generate_trials(
strategy=self.strategy,
flat_params=self.__flatten_params(self.opt_args),
nb_trials=num,
)
for trial in self.trials:
ns = self.__namespace_from_trial(trial)
yield ns
def generate_trials(self, nb_trials):
trials = strategies.generate_trials(
strategy=self.strategy,
flat_params=self.__flatten_params(self.opt_args),
nb_trials=nb_trials,
)
trials = [self.__namespace_from_trial(x) for x in trials]
return trials
def optimize_parallel_gpu(
self,
train_function,
gpu_ids,
max_nb_trials=None,
):
"""
Runs optimization across gpus with cuda drivers
:param train_function:
:param max_nb_trials:
:param gpu_ids: List of strings like: ['0', '1, 3']
:return:
"""
self.trials = strategies.generate_trials(
strategy=self.strategy,
flat_params=self.__flatten_params(self.opt_args),
nb_trials=max_nb_trials,
)
self.trials = [(self.__namespace_from_trial(x), train_function) for x in self.trials]
# build q of gpu ids so we can use them in each process
# this is thread safe so each process can pull out a gpu id, run its task and put it back when done
if self.pool is None:
gpu_q = Queue()
for gpu_id in gpu_ids:
gpu_q.put(gpu_id)
# init a pool with the nb of worker threads we want
nb_workers = len(gpu_ids)
self.pool = Pool(processes=nb_workers, initializer=init, initargs=(gpu_q,))
# apply parallelization
results = self.pool.map(optimize_parallel_gpu_private, self.trials)
return results
def optimize_trials_parallel_gpu(
self,
train_function,
nb_trials,
trials,
gpu_ids,
nb_workers=4,
):
"""
Runs optimization across gpus with cuda drivers
:param train_function:
:param nb_trials:
:param gpu_ids: List of strings like: ['0', '1, 3']
:param nb_workers:
:return:
"""
self.trials = trials
self.trials = [(x, train_function) for x in self.trials]
# build q of gpu ids so we can use them in each process
# this is thread safe so each process can pull out a gpu id, run its task and put it back when done
if self.pool is None:
gpu_q = Queue()
for gpu_id in gpu_ids:
gpu_q.put(gpu_id)
# init a pool with the nb of worker threads we want
self.pool = Pool(processes=nb_workers, initializer=init, initargs=(gpu_q,))
# apply parallelization
results = self.pool.map(optimize_parallel_gpu_private, self.trials)
return results
def optimize_parallel_cpu(
self,
train_function,
nb_trials,
nb_workers=4,
):
"""
Runs optimization across n cpus
:param train_function:
:param nb_trials:
:param nb_workers:
:return:
"""
self.trials = strategies.generate_trials(
strategy=self.strategy,
flat_params=self.__flatten_params(self.opt_args),
nb_trials=nb_trials
)
self.trials = [(self.__namespace_from_trial(x), train_function) for x in self.trials]
# init a pool with the nb of worker threads we want
if self.pool is None:
self.pool = Pool(processes=nb_workers)
# apply parallelization
results = self.pool.map(optimize_parallel_cpu_private, self.trials)
return results
def optimize_parallel(
self,
train_function,
nb_trials,
nb_parallel=4,
):
self.trials = strategies.generate_trials(
strategy=self.strategy,
flat_params=self.__flatten_params(self.opt_args),
nb_trials=nb_trials
)
# nb of runs through all parallel systems
fork_batches = [
self.trials[i:i + nb_parallel] for i in range(0, len(self.trials), nb_parallel)
]
for fork_batch in fork_batches:
children = []
# run n parallel forks
for parallel_nb, trial in enumerate(fork_batch):
# q up the trial and convert to a namespace
ns = self.__namespace_from_trial(trial)
# split new fork
pid = os.fork()
# when the process is a parent
if pid:
children.append(pid)
# when process is a child
else:
# slight delay to make sure we don't overwrite over test tube log versions
sleep(parallel_nb * 0.5)
train_function(ns, parallel_nb)
os._exit(0)
for i, child in enumerate(children):
os.waitpid(child, 0)
def __namespace_from_trial(self, trial):
trial_dict = {d['name']: d['val'] for d in trial}
for k, v in self.parsed_args.items():
if k not in trial_dict:
trial_dict[k] = v
return TTNamespace(**trial_dict)
def __flatten_params(self, params):
"""
Turns a list of parameters with values into a flat tuple list of lists
so we can permute
:param params:
:return:
"""
flat_params = []
for i, (opt_name, opt_arg) in enumerate(params.items()):
if opt_arg.tunable:
clean_name = opt_name.strip('-')
clean_name = re.sub('-', '_', clean_name)
param_groups = []
for val in opt_arg.opt_values:
param_groups.append({'idx': i, 'val': val, 'name': clean_name})
flat_params.append(param_groups)
return flat_params
str) + ".jpg"
return df
path = untar_data(URLs.FOOD)
train_path = path / 'train.txt'
test_path = path / 'test.txt'
def load_data(index):
train_df = filelist2df(train_path)
test_df = filelist2df(test_path)
food = DataBlock(blocks=(ImageBlock, CategoryBlock),
get_x=ColReader(1, pref=path / 'images'),
splitter=RandomSplitter(),
get_y=ColReader(cols=0),
item_tfms=Resize(224))
dls = food.dataloaders(train_df.values, bs=64)
if __name__ == '__main__':
set_start_method('spawn', force=True)
try:
pool = Pool(8)
pool.map(load_data, [1, 2, 3, 4, 5, 6, 7, 8])
except KeyboardInterrupt:
exit()
finally:
pool.terminate()
pool.join()
torch.set_default_tensor_type(torch.DoubleTensor)
experiment = experiments.cart_pole
pool = Pool(experiment.thread_count)
#torch.multiprocessing.set_start_method('spawn')
thread_count = experiment.thread_count
pop_size = experiment.population
generation_count = experiment.generations
mutate_range = experiment.mutate_range
population = Population()
sys.stdout.write("Evaluating Intial Fitness:")
sys.stdout.flush()
thread_list = []
new_nets = []
for i in range(pop_size):
new_net = Genome(experiment)
new_net.evalFitness()
new_nets.append(new_net)
iters_required = math.ceil(pop_size / thread_count)
for _ in range(iters_required):
threads = min(thread_count, len(new_nets))
unevaled_nets = []
for i in range(threads):
unevaled_nets.append(3) #(new_nets[i])
for _ in range(threads):
del new_nets[0] #Check for bug/change line? inefficient at best
fitnesses = pool.map(multiEvalFitness, unevaled_nets)
#for i in range(threads):
# unevaled_nets[i].fitness = fitnesses[i]
#for net in unevaled_nets:
# population.add(net)
print("Done!")
def ars(env_name, n_epochs, env_config, step_size, n_delta, n_top, exp_noise,
n_workers, policy, seed):
torch.autograd.set_grad_enabled(False) # Gradient free baby!
pool = Pool(processes=n_workers)
W = torch.nn.utils.parameters_to_vector(policy.parameters())
n_param = W.shape[0]
if env_config is None:
env_config = {}
env = gym.make(env_name, **env_config)
env.seed(seed)
torch.manual_seed(seed)
np.random.seed(seed)
total_steps = 0
r_hist = []
exp_dist = torch.distributions.Normal(torch.zeros(n_delta, n_param),
torch.ones(n_delta, n_param))
do_rollout_partial = partial(do_rollout_train, env_name, policy)
for _ in range(n_epochs):
deltas = exp_dist.sample()
###
pm_W = torch.cat((W + (deltas * exp_noise), W - (deltas * exp_noise)))
results = pool.map(do_rollout_partial, pm_W)
states = torch.empty(0)
p_returns = []
m_returns = []
l_returns = []
top_returns = []
for p_result, m_result in zip(results[:n_delta], results[n_delta:]):
ps, pr, plr = p_result
ms, mr, mlr = m_result
states = torch.cat((states, ms, ps), dim=0)
p_returns.append(pr)
m_returns.append(mr)
l_returns.append(plr)
l_returns.append(mlr)
top_returns.append(max(pr, mr))
top_idx = sorted(range(len(top_returns)),
key=lambda k: top_returns[k],
reverse=True)[:n_top]
p_returns = torch.stack(p_returns)[top_idx]
m_returns = torch.stack(m_returns)[top_idx]
l_returns = torch.stack(l_returns)[top_idx]
r_hist.append(l_returns.mean())
###
W = W + (step_size / (n_delta * torch.cat(
(p_returns, m_returns)).std() + 1e-6)) * torch.sum(
(p_returns - m_returns) * deltas[top_idx].T, dim=1)
ep_steps = states.shape[0]
policy.state_means = update_mean(states, policy.state_means,
total_steps)
policy.state_std = update_std(states, policy.state_std, total_steps)
do_rollout_partial = partial(do_rollout_train, env_name, policy)
total_steps += ep_steps
torch.nn.utils.vector_to_parameters(W, policy.parameters())
return policy, r_hist
json_decoded[county] = [[
county_dict[county][0], county_dict[county][1], num_cases,
date.today().strftime("%m/%d/%y")
]]
with lock:
with open(path, 'w') as json_file:
json.dump(json_decoded, json_file)
print(county, num_cases)
return max(scraped_cases)
else:
print("Blocked")
return 0
if __name__ == '__main__':
counties = list(county_dict.keys())
print(len(counties))
freeze_support()
try:
set_start_method('spawn')
except RuntimeError:
pass
p = Pool(4)
records = p.map(get_cases, counties)
p.close()
p.join()
import torch
from torch.multiprocessing import Pool, Process, set_start_method
from torch.autograd import Variable
import numpy as np
from scipy.ndimage import zoom
def get_pred(args):
img = args[0]
scale = args[1]
# feed input data
input_img = Variable(torch.from_numpy(img),
volatile=True).cuda()
return input_img
if __name__ == '__main__':
try:
set_start_method('spawn')
except RuntimeError:
pass
img = np.float32(np.random.randint(0, 2, (300, 300, 3)))
scales = [1,2,3,4,5]
scale_list = []
for scale in scales:
scale_list.append([img,scale])
multi_pool = Pool(processes=5)
predictions = multi_pool.map(get_pred,scale_list)
multi_pool.close()
multi_pool.join()
class LibriSelectionDataset(Dataset):
"""LibriSpeech Selection data from sincnet paper."""
def __init__(self,
sizeWindow=20480,
db_wav_root=DB_WAV_ROOT,
fps_list=str(),
label_path=str(),
nSpeakers=-1,
n_process_loader=50,
MAX_SIZE_LOADED=4000000000):
"""Init.
Args:
- sizeWindow (int): size of the sliding window
- db_wav_path (str):
- fps_list_path (str):
- label_path (str):
- n_process_loader (int):
- MAX_SIZE_LOADED (int): target maximal size of the floating array
containing all loaded data.
"""
self.MAX_SIZE_LOADED = MAX_SIZE_LOADED
self.n_process_loader = n_process_loader
self.db_wav_root = Path(db_wav_root)
self.sizeWindow = sizeWindow
"""Parsing customized to Libri-selection dataset."""
fps_name_only = get_fps_from_txt(fps_list)
label_dict = np.load(label_path, allow_pickle=True)[()]
self.all_labels_fps = [(label_dict[x], Path(db_wav_root) / Path(x))
for x in fps_name_only]
self.reload_pool = Pool(n_process_loader)
self.prepare(
) # Split large number of files into packages, and set {self.currentPack=-1, self.nextPack=0}
if nSpeakers == -1:
nSpeakers = len(set(label_dict.values()))
self.speakers = list(range(nSpeakers))
self.data = []
self.loadNextPack(first=True)
self.loadNextPack()
def __len__(self):
"""Get length."""
return self.totSize // self.sizeWindow
def prepare(self):
"""Prepare."""
random.shuffle(self.all_labels_fps)
start_time = time.time()
print("Checking length...")
allLength = self.reload_pool.map(extractLength, self.all_labels_fps)
self.packageIndex, self.totSize = [], 0
start, packageSize = 0, 0
for index, length in tqdm.tqdm(enumerate(allLength)):
packageSize += length
if packageSize > self.MAX_SIZE_LOADED:
self.packageIndex.append([start, index])
self.totSize += packageSize
start, packageSize = index, 0
if packageSize > 0:
self.packageIndex.append([start, len(self.all_labels_fps)])
self.totSize += packageSize
print(f"Done, elapsed: {time.time() - start_time:.3f} seconds")
print(f'Scanned {len(self.all_labels_fps)} sequences '
f'in {time.time() - start_time:.2f} seconds')
print(f"{len(self.packageIndex)} chunks computed")
self.currentPack = -1
self.nextPack = 0
def clear(self):
"""Clear."""
if 'data' in self.__dict__:
del self.data
if 'speakerLabel' in self.__dict__:
del self.speakerLabel
if 'seqLabel' in self.__dict__:
del self.seqLabel
def getNPacks(self):
"""Get N packs."""
return len(self.packageIndex)
def getNSeqs(self):
"""Get N seqs."""
return len(self.seqLabel) - 1
def getNLoadsPerEpoch(self):
"""Get N loads per epoch."""
return len(self.packageIndex)
def getSpeakerLabel(self, idx):
idSpeaker = next(
x[0] for x in enumerate(self.speakerLabel) if x[1] > idx) - 1
return idSpeaker
def loadNextPack(self, first=False):
"""Load next pack."""
self.clear()
if not first:
self.currentPack = self.nextPack
start_time = time.time()
print('Joining pool')
self.r.wait()
print(f'Joined process, elapsed={time.time()-start_time:.3f} secs')
self.nextData = self.r.get()
self.parseNextDataBlock()
del self.nextData
self.nextPack = (self.currentPack + 1) % len(self.packageIndex)
seqStart, seqEnd = self.packageIndex[self.nextPack]
if self.nextPack == 0 and len(self.packageIndex) > 1:
self.prepare()
"""map() blocks until complete, map_async() returns immediately and
schedules a callback to be run on the result."""
self.r = self.reload_pool.map_async(
loadFile, self.all_labels_fps[seqStart:seqEnd])
"""loadFile: return speaker, seqName, seq"""
def parseNextDataBlock(self):
"""Parse next data block."""
# Labels
self.speakerLabel = [0]
self.seqLabel = [0]
speakerSize = 0
indexSpeaker = 0
# To accelerate the process a bit
self.nextData.sort(key=lambda x: (x[0], x[1]))
"""
nextData[0] = (1243, '4910-14124-0001-1',
tensor([-0.0089, -0.0084, -0.0079, ..., -0.0015, -0.0056, 0.0047]))
"""
tmpData = []
for speaker, seqName, seq in self.nextData:
while self.speakers[indexSpeaker] < speaker:
indexSpeaker += 1
self.speakerLabel.append(speakerSize)
if self.speakers[indexSpeaker] != speaker:
raise ValueError(f'{speaker} invalid speaker')
sizeSeq = seq.size(0)
tmpData.append(seq)
self.seqLabel.append(self.seqLabel[-1] + sizeSeq)
speakerSize += sizeSeq
del seq
self.speakerLabel.append(speakerSize)
self.data = torch.cat(tmpData, dim=0)
def __getitem__(self, idx):
"""Get item."""
if idx < 0 or idx >= len(self.data) - self.sizeWindow - 1:
print(idx)
outData = self.data[idx:(self.sizeWindow + idx)].view(1, -1)
label = torch.tensor(self.getSpeakerLabel(idx), dtype=torch.long)
return outData, label
def getBaseSampler(self, type, batchSize, offset):
"""Get base sampler."""
if type == "samespeaker":
return SameSpeakerSampler(batchSize, self.speakerLabel,
self.sizeWindow, offset)
if type == "samesequence":
return SameSpeakerSampler(batchSize, self.seqLabel,
self.sizeWindow, offset)
if type == "sequential":
return SequentialSampler(len(self.data), self.sizeWindow, offset,
batchSize)
sampler = UniformAudioSampler(len(self.data), self.sizeWindow, offset)
return BatchSampler(sampler, batchSize, True)
def getDataLoader(self,
batchSize,
type,
randomOffset,
numWorkers=0,
onLoop=-1):
"""Get a batch sampler for the current dataset.
Args:
- batchSize (int): batch size
- groupSize (int): in the case of type in ["samespeaker", "samesequence"]
number of items sharing a same label in the group
(see AudioBatchSampler)
- type (string):
type == "samespeaker": grouped sampler speaker-wise
type == "samesequence": grouped sampler sequence-wise
type == "sequential": sequential sampling
else: uniform random sampling of the full audio
vector
- randomOffset (bool): if True add a random offset to the sampler
at the begining of each iteration
"""
nLoops = len(self.packageIndex)
totSize = self.totSize // (self.sizeWindow * batchSize)
if onLoop >= 0:
self.currentPack = onLoop - 1
self.loadNextPack()
nLoops = 1
def samplerCall():
offset = random.randint(0, self.sizeWindow // 2) \
if randomOffset else 0
return self.getBaseSampler(type, batchSize, offset)
return AudioLoader(self, samplerCall, nLoops, self.loadNextPack,
totSize, numWorkers)
if loss.requires_grad:
lbfgs_g.zero_grad()
loss.backward()
print("total train loss", loss.item()**(1/train_exp))
return loss
if True:
for t in range(25):
lbfgs_g.step(closure)
if t % 1 == 0:
val_loss, test_loss = val_test_loss()
print("Iteration", t, best_loss, val_loss.item(), test_loss.item())
if val_loss.item() < best_loss:
best_loss = val_loss.item()
torch.save(model.state_dict(), "trained_models/%s_game_model_vdelta%d_spacing%d_scaleaccel%d_vel_predsize%d_fold%d.torch" % (dataset_name,use_deltav,spacing,scale_accel,pred_size,fold,))
# load best model
model.load_state_dict(torch.load("trained_models/%s_game_model_vdelta%d_spacing%d_scaleaccel%d_vel_predsize%d_fold%d.torch" % (dataset_name,use_deltav,spacing,scale_accel,pred_size,fold,)))
val_loss, test_loss = val_test_loss()
print("Final Result spacing%d pred_size%d dataset %s fold%d" % (spacing,pred_size, dataset_name, fold,), val_loss.item(), test_loss.item())
multi_pool = Pool(processes=5)
print("Running training on dataset %s ..." % (dataset_name,))
predictions = multi_pool.map(train_fold,list(range(4)))
multi_pool.close()
multi_pool.join()
def evolve(experiment):
total_frames = 0
#If CUDA is breaking try adding this and removing it from main/run_experiment
#try:
# set_start_method('spawn')
#except RuntimeError:
# pass
pool = Pool(experiment.thread_count)
thread_count = experiment.thread_count
time_start = time.perf_counter()
#Set params based on the current experiment
pop_size = experiment.population
generation_count = experiment.generations
#Create new random population, sorted by starting fitness
population = Population(experiment)
new_nets = []
saved = [] #Saving fittest from each gen to pickle file
#sys.stdout.write("Evaluating Intial Fitness:")
#sys.stdout.flush()
for i in range(pop_size):
new_net = "Maybe I can write a function to make a new net of type specified by the experiment"
if experiment.genome == 'NEAT':
new_net = NEATGenome(experiment)
elif experiment.genome == 'TensorNEAT':
new_net = TensorNEATGenome(experiment)
else:
new_net = Genome(experiment)
new_nets.append(new_net)
#Multithreaded fitness evaluation
"""
net_copies = []
for _ in range(thread_count):
net_copies.append([])
for i in range(pop_size):
net_copies[i%thread_count].append(copy.deepcopy(new_nets[i]))
multiReturn = pool.map(multiEvalFitness, net_copies)
fitnesses = []
for thread in multiReturn:
fitnesses.append(thread[0])
total_frames += thread[1]
for i in range(pop_size):
new_nets[i].fitness = fitnesses[i%thread_count][i//thread_count]
for net in new_nets:
population.add(net)
"""
#net_copies = []
#for i in range(pop_size):
# net_copies.append(copy.deepcopy(new_nets[i]))
multiReturn = pool.map(multiEvalFitness, new_nets)
for i in range(pop_size):
#print(new_nets[i].fitness)
new_nets[i].fitness = multiReturn[i][0]
#sys.stdout.write(str(multiReturn[i][0]) + "\n")
#sys.stdout.flush()
total_frames += multiReturn[i][1]
for net in new_nets:
population.add(net)
#Run the main algorithm over many generations
#for g in range(generation_count):
generation = 0
while total_frames < experiment.max_frames:
#First print reports on generation:
#Debugging report I hope to remove soon
"""
if generation%20 == 0:
print("Top genomes:")
population[0].printToTerminal()
population[1].printToTerminal()
population[2].printToTerminal()
print("Species Report: size, gens_since_improvement, record fitness, current fitness")
for s in population.species:
if s.size() > 0:
print(s.size(), s.gens_since_improvement, s.last_fittest, s.genomes[0].fitness)
#print(torch.cuda.memory_summary())
gen_report_string = "\nGeneration " + str(generation) + "\nTotal frames used: " + str(total_frames) + "\nHighest Fitness: "+ str(population.fittest().fitness) + "\nTotal elapsed time:" + str(time.perf_counter() - time_start) + " seconds\n"
sys.stdout.write(gen_report_string)
sys.stdout.flush()
"""
#Next do the speciation work
#Check all species and remove those that haven't improved in so many generations
population.checkSpeciesForImprovement(experiment.gens_to_improve)
#Adjust fitness of each individual with fitness sharing
#Done here so it is skipped for the final population, where plain maximum fitness is desired
if experiment.fitness_sharing:
for species in population.species:
for genome in species:
genome.fitness = genome.fitness / species.size()
#Population is re-ordered afterwards based on new fitness
population.reorder() #make sure this is only called when necessary
#Assign how many offspring each species gets based on fitness of the species
population.assignOffspringProportions()
#Now we set the crossover/mutate counts for NEAT; since speciated evolution has a varying count of elite genomes retained
elite_count = 0
for species in population.species:
if species.size(
) >= experiment.elite_threshold and species.gens_since_improvement < experiment.gens_to_improve:
elite_count += experiment.elite_per_species
experiment.elite_count = elite_count
experiment.mutate_count = math.floor(
experiment.mutate_ratio *
(experiment.population - experiment.elite_count))
experiment.crossover_count = experiment.population - experiment.mutate_count - experiment.elite_count
#Make the new population to fill this generation
new_pop = Population(experiment)
#Now we select species reps for the new pop based on the old one
for s in population.species:
rep = population.randOfSpecies(s)
new_species = Species(
experiment, rep, False, s.gens_since_improvement,
s.last_fittest, s.can_reproduce
) #The genome is copied over as a rep but not added
new_pop.species.append(new_species)
new_nets = []
#Crossover is done first
#Roll the dice for interspecies; limit to one per gen to make this calculation simpler
if random.random(
) < experiment.interspecies_crossover * experiment.crossover_count:
parent1 = population.select()
parent2 = population.select()
while parent1 == parent2:
parent2 = population.select()
new_net = parent1.crossover(parent2)
new_nets.append(new_net)
experiment.crossover_count -= 1
#Create and add them to the pop, subtract from crossover count
#Since we round up to the nearest integer to find how many to take from each species, the total number at the end will likely exceed the desired number
#We fix this at the end through random pruning
offspring = []
for species in population.species:
count = math.ceil(experiment.crossover_count *
species.offspring_proportion)
for _ in range(count):
if species.size() == 1:
#Just do mutation
parent = species.select()
new_net = parent.mutate()
offspring.append(new_net)
else: #It's possible that this has to repeat a lot if there's only a few with a large fit diff, but unlikely
parent1 = species.select()
parent2 = species.select()
while parent1 == parent2:
parent2 = species.select()
new_net = parent1.crossover(parent2)
offspring.append(new_net)
#Do the crossover here
#Now remove from mutated at random until we have the right number
to_remove = len(offspring) - experiment.crossover_count
if to_remove < 0:
print(to_remove)
assert False
for _ in range(to_remove):
del offspring[random.randint(0, len(offspring) - 1)]
for n in offspring:
new_nets.append(n)
#Mutation to create offspring is done next; the same pruning method is used
mutated = []
for species in population.species:
for _ in range(
math.ceil(experiment.mutate_count *
species.offspring_proportion)):
parent = species.select()
new_net = parent.mutate()
mutated.append(new_net)
#Now remove from mutated at random until we have the right number
to_remove = len(mutated) - experiment.mutate_count
assert (to_remove >= 0)
for _ in range(to_remove):
del mutated[random.randint(0, len(mutated) - 1)]
for n in mutated:
new_nets.append(n)
#net_copies = []
#for i in range(pop_size-elite_count):
# net_copies.append(copy.deepcopy(new_nets[i]))
multiReturn = pool.map(multiEvalFitness, new_nets)
for i in range(pop_size - elite_count):
new_nets[i].fitness = multiReturn[i][0]
total_frames += multiReturn[i][1]
for net in new_nets:
new_pop.add(net)
#Elite Carry-over; re-evaluates fitness first before selection
#Currently not built to carry best of each species over; this should be handled by fitness sharing
#And since this is typically only 1, we just want the fittest genome regardless of species
#Eval all the elite nets many times
elite_nets = []
for species in population.species:
if species.size(
) >= experiment.elite_threshold and species.gens_since_improvement < experiment.gens_to_improve:
for i in range(experiment.elite_range):
elite_nets.append(species[i])
#net_copies = []
#for i in range(len(elite_nets)):
# net_copies.append(copy.deepcopy(elite_nets[i]))
multiReturn = pool.map(multiEvalFitnessElite, elite_nets)
#print(fitnesses)
for i in range(len(elite_nets)):
elite_nets[i].fitness = multiReturn[i][0]
total_frames += multiReturn[i][1]
elite_max = float("-inf")
top_elite = None
for species in population.species:
if species.size(
) >= experiment.elite_threshold and species.gens_since_improvement < experiment.gens_to_improve:
for i in range(
experiment.elite_per_species
): #This needs to be redone for elite_count > 1; currently would just take best genome twice
best_fitness = float('-inf')
fittest = None
for i in range(experiment.elite_range):
if species[i].fitness > best_fitness:
best_fitness = species[i].fitness
fittest = species[i]
if best_fitness > elite_max:
elite_max = best_fitness
top_elite = fittest
new_pop.add(fittest)
#If the experiment ran enough trials, we can just use the top elite. Generally this should be reserved for single-species algorithms
if not experiment.save_elite: #If not, run our own trials, not counting frames, to find the actual best genome from this generation to save for final evaluation
elite_nets = []
for i in range(10):
elite_nets.append(population[i])
fitnesses = pool.map(multiEvalFitnessThirty, elite_nets)
for i in range(10):
elite_nets[i].fitness = fitnesses[i]
elite_max = float("-inf")
top_elite = None
for net in elite_nets:
if net.fitness > elite_max:
elite_max = net.fitness
top_elite = net
save_copy = top_elite.newCopy(
) #Still need to modify/add this for regular/tensor genomes
saved.append([save_copy, elite_max])
"""
if top_elite is None:
sys.stdout.write("No elite could be found, using pop.fittest instead\n")
for s in population.species:
sys.stdout.write(str(species.size()) + str(species.gens_since_improvement) + str(species.can_reproduce))
top_elite = population.fittest()
top_elite.evalFitness(iters=top_elite.experiment.elite_evals)#These frames are not counted since they are only for reporting purposes and do not affect the actual algorithm
elite_max = top_elite.fitness
if elite_max < 5.0:
for s in population.species:
print(s.can_reproduce)
sys.stdout.write("Strange elite behavior. Fitness is " + str(top_elite.fitness) + " Trial is " + str(top_elite.evalFitness()) + " Top fitness is: " + str(population.fittest().fitness) + "\n")
sys.stdout.flush()
#Save top elite carryover to pickle file
save_copy = copy.deepcopy(top_elite)
saved.append([save_copy, elite_max])
"""
#total_layers = 0
#for genome in new_pop:
# total_layers += genome.layer_count
#avg_layers = total_layers/new_pop.size()
elapsed = int(time.perf_counter() - time_start)
time_string = str(elapsed // 3600) + ":" + str(
(elapsed % 3600) // 60) + ":" + str(elapsed % 60)
#sys.stdout.write(str(100*total_frames/experiment.max_frames) + "% complete | " + time_string + " elapsed | " + str(elite_max) + " recent score | " + str(save_copy.layer_size) + " " + str(save_copy.layer_count) + " layer size/count | " + str(len(new_pop.species)) + " total species | " + str(avg_layers) + " average layers\n")
sys.stdout.write(
str(100 * total_frames / experiment.max_frames) + "% complete | " +
time_string + " elapsed | " + str(elite_max) + " recent score | " +
str(len(new_pop.species)) + " total species \n")
sys.stdout.flush()
population.species = []
population.genomes = []
population = new_pop
generation += 1
print("Final frame count:", str(total_frames))
print("Total generations:", generation)
print("Time Elapsed:", time.perf_counter() - time_start)
return population, saved
class AudioBatchData(Dataset):
def __init__(self,
path,
sizeWindow,
seqNames,
phoneLabelsDict,
nSpeakers,
nProcessLoader=50,
MAX_SIZE_LOADED=4000000000):
"""
Args:
- path (string): path to the training dataset
- sizeWindow (int): size of the sliding window
- seqNames (list): sequences to load
- phoneLabelsDict (dictionnary): if not None, a dictionnary with the
following entries
"step": size of a labelled window
"$SEQ_NAME": list of phonem labels for
the sequence $SEQ_NAME
- nSpeakers (int): number of speakers to expect.
- nProcessLoader (int): number of processes to call when loading the
data from the disk
- MAX_SIZE_LOADED (int): target maximal size of the floating array
containing all loaded data.
"""
self.MAX_SIZE_LOADED = MAX_SIZE_LOADED
self.nProcessLoader = nProcessLoader
self.dbPath = Path(path)
self.sizeWindow = sizeWindow
self.seqNames = [(s, self.dbPath / x) for s, x in seqNames]
self.reload_pool = Pool(nProcessLoader)
self.prepare()
self.speakers = list(range(nSpeakers))
self.data = []
self.phoneSize = 0 if phoneLabelsDict is None else \
phoneLabelsDict["step"]
self.phoneStep = 0 if phoneLabelsDict is None else \
self.sizeWindow // self.phoneSize
self.phoneLabelsDict = deepcopy(phoneLabelsDict)
self.loadNextPack(first=True)
self.loadNextPack()
self.doubleLabels = False
def resetPhoneLabels(self, newPhoneLabels, step):
self.phoneSize = step
self.phoneStep = self.sizeWindow // self.phoneSize
self.phoneLabelsDict = deepcopy(newPhoneLabels)
self.loadNextPack()
def splitSeqTags(seqName):
path = os.path.normpath(seqName)
return path.split(os.sep)
def getSeqNames(self):
return [str(x[1]) for x in self.seqNames]
def clear(self):
if 'data' in self.__dict__:
del self.data
if 'speakerLabel' in self.__dict__:
del self.speakerLabel
if 'phoneLabels' in self.__dict__:
del self.phoneLabels
if 'seqLabel' in self.__dict__:
del self.seqLabel
def prepare(self):
random.shuffle(self.seqNames)
start_time = time.time()
print("Checking length...")
allLength = self.reload_pool.map(extractLength, self.seqNames)
self.packageIndex, self.totSize = [], 0
start, packageSize = 0, 0
for index, length in tqdm.tqdm(enumerate(allLength)):
packageSize += length
if packageSize > self.MAX_SIZE_LOADED:
self.packageIndex.append([start, index])
self.totSize += packageSize
start, packageSize = index, 0
if packageSize > 0:
self.packageIndex.append([start, len(self.seqNames)])
self.totSize += packageSize
print(f"Done, elapsed: {time.time() - start_time:.3f} seconds")
print(f'Scanned {len(self.seqNames)} sequences '
f'in {time.time() - start_time:.2f} seconds')
print(f"{len(self.packageIndex)} chunks computed")
self.currentPack = -1
self.nextPack = 0
def getNPacks(self):
return len(self.packageIndex)
def loadNextPack(self, first=False):
self.clear()
if not first:
self.currentPack = self.nextPack
start_time = time.time()
print('Joining pool')
self.r.wait()
print(f'Joined process, elapsed={time.time()-start_time:.3f} secs')
self.nextData = self.r.get()
self.parseNextDataBlock()
del self.nextData
self.nextPack = (self.currentPack + 1) % len(self.packageIndex)
seqStart, seqEnd = self.packageIndex[self.nextPack]
if self.nextPack == 0 and len(self.packageIndex) > 1:
self.prepare()
self.r = self.reload_pool.map_async(loadFile,
self.seqNames[seqStart:seqEnd])
def parseNextDataBlock(self):
# Labels
self.speakerLabel = [0]
self.seqLabel = [0]
self.phoneLabels = []
speakerSize = 0
indexSpeaker = 0
# To accelerate the process a bit
self.nextData.sort(key=lambda x: (x[0], x[1]))
tmpData = []
for speaker, seqName, seq in self.nextData:
while self.speakers[indexSpeaker] < speaker:
indexSpeaker += 1
self.speakerLabel.append(speakerSize)
if self.speakers[indexSpeaker] != speaker:
raise ValueError(f'{speaker} invalid speaker')
if self.phoneLabelsDict is not None:
self.phoneLabels += self.phoneLabelsDict[seqName]
newSize = len(self.phoneLabelsDict[seqName]) * self.phoneSize
seq = seq[:newSize]
sizeSeq = seq.size(0)
tmpData.append(seq)
self.seqLabel.append(self.seqLabel[-1] + sizeSeq)
speakerSize += sizeSeq
del seq
self.speakerLabel.append(speakerSize)
self.data = torch.cat(tmpData, dim=0)
def getPhonem(self, idx):
idPhone = idx // self.phoneSize
return self.phoneLabels[idPhone:(idPhone + self.phoneStep)]
def getSpeakerLabel(self, idx):
idSpeaker = next(
x[0] for x in enumerate(self.speakerLabel) if x[1] > idx) - 1
return idSpeaker
def __len__(self):
return self.totSize // self.sizeWindow
def __getitem__(self, idx):
if idx < 0 or idx >= len(self.data) - self.sizeWindow - 1:
print(idx)
outData = self.data[idx:(self.sizeWindow + idx)].view(1, -1)
label = torch.tensor(self.getSpeakerLabel(idx), dtype=torch.long)
if self.phoneSize > 0:
label_phone = torch.tensor(self.getPhonem(idx), dtype=torch.long)
if not self.doubleLabels:
label = label_phone
else:
label_phone = torch.zeros(1)
if self.doubleLabels:
return outData, label, label_phone
return outData, label
def getNSpeakers(self):
return len(self.speakers)
def getNSeqs(self):
return len(self.seqLabel) - 1
def getNLoadsPerEpoch(self):
return len(self.packageIndex)
def getBaseSampler(self, type, batchSize, offset):
if type == "samespeaker":
return SameSpeakerSampler(batchSize, self.speakerLabel,
self.sizeWindow, offset)
if type == "samesequence":
return SameSpeakerSampler(batchSize, self.seqLabel,
self.sizeWindow, offset)
if type == "sequential":
return SequentialSampler(len(self.data), self.sizeWindow, offset,
batchSize)
sampler = UniformAudioSampler(len(self.data), self.sizeWindow, offset)
return BatchSampler(sampler, batchSize, True)
def getDataLoader(self,
batchSize,
type,
randomOffset,
numWorkers=0,
onLoop=-1):
r"""
Get a batch sampler for the current dataset.
Args:
- batchSize (int): batch size
- groupSize (int): in the case of type in ["samespeaker", "samesequence"]
number of items sharing a same label in the group
(see AudioBatchSampler)
- type (string):
type == "speaker": grouped sampler speaker-wise
type == "sequence": grouped sampler sequence-wise
type == "sequential": sequential sampling
else: uniform random sampling of the full audio
vector
- randomOffset (bool): if True add a random offset to the sampler
at the begining of each iteration
"""
nLoops = len(self.packageIndex)
totSize = self.totSize // (self.sizeWindow * batchSize)
if onLoop >= 0:
self.currentPack = onLoop - 1
self.loadNextPack()
nLoops = 1
def samplerCall():
offset = random.randint(0, self.sizeWindow // 2) \
if randomOffset else 0
return self.getBaseSampler(type, batchSize, offset)
return AudioLoader(self, samplerCall, nLoops, self.loadNextPack,
totSize, numWorkers)
def trawl(dset,dtype,epochs,parallel=False,batch_size=1,max_n=30,max_rankings=1000,opt='Adam',num_dsets=10,seed=True,RE=False,K=5):
"""
trawls over a directory and fits models to all data files
Args:
dset- name of dataset(s) considered
dtype- 'soi' for partial rankings, 'soc' for complete rankings
epochs- number of times to loop over the data
parallel- whether to train models in parallel over the datasets in the directory
batch_size- number of choices to train on at a time
max_n- largest number of alternatives allowed to train on a dataset
max_rankings- maximum number of rankings to fit a dataset
opt- which optimizer to use
num_dsets- number of datasets to fit
seed- whether to seed PCMC
RE- whether to compute repeated elimianation (RS if false)
K- number of CV folds for each dataset
"""
#we will loop over the datasets stored in this directory
path = os.getcwd()+os.sep+'data'+os.sep+dset
files = os.listdir(path)
#shuffle(files)
#this is where we'll save the output models
save_path = os.getcwd()+os.sep+'cache'+os.sep+'learned_models'+os.sep+dset+os.sep
job_list = []
batch = (batch_size>1)
for filename in files:#loop over the directory
print(filename)
if filename.endswith(dtype):#will
filepath = path+os.sep+filename
if dtype=='soi':
L,n = scrape_soi(filepath)
else:
L,n = scrape_soc(filepath)
if len(L)<=10 or len(L)>max_rankings or n>max_n:
if len(L)<=10:
reason = 'too few rankings- '+str(len(L))+', min is 10'
elif len(L)>max_rankings:
reason = 'too many rankings- '+str(len(L))+', max is '+str(max_rankings)
else:
reason = 'too many alternatives- '+str(n)+', max is '+str(max_n)
print(filename+' skipped, '+reason)
continue
else:
print(filename+' added')
#collect models
models = []
for d in [1,4,8]:
if d>n:
continue
models.append((f'CRS, r = {d}', cp.CDM, {'embedding_dim': d}))
models.append((f'PL', cp.MNL, {}))
#models.append(PCMC(n,batch=batch))
#models.append((f'RS-PCMC', cp.PCMC, {'batch': batch}))
#append tuple containing all the ojects needed to train the model on the dataset
job_list.append((L,n,models,save_path+filename[:-4]+'-'+dtype,epochs,batch_size,opt,seed,False,K))
if RE:
job_list.append((map(lambda x:x[::-1],L),n,models,save_path+filename[:-4]+'-'+dtype,epochs,batch_size,opt,seed,True,K))
if len(job_list)>=num_dsets:
print('maximum number of datasets reached')
continue
print(str(len(job_list))+' datasets total')
print(str(sum(map(lambda x: len(x[0]),job_list)))+ ' total rankings')
#sorts the jobs by number of alternatives*number of (partial) rankings
#will roughly be the number of choices, up to partial ranking length
sorted(job_list,key=lambda x: x[1]*len(x[0]))
#training for each dataset can be done in parallel with this
if parallel:
p = Pool(4)
p.map(parallel_helper,job_list)
else:
[x for x in map(parallel_helper,job_list)]
def ars(env_name,
policy,
n_epochs,
n_workers=8,
step_size=.02,
n_delta=32,
n_top=16,
exp_noise=0.03,
zero_policy=True,
postprocess=postprocess_default):
torch.autograd.set_grad_enabled(False)
"""
Augmented Random Search
https://arxiv.org/pdf/1803.07055
Args:
Returns:
Example:
"""
pool = Pool(processes=n_workers)
env = gym.make(env_name)
W = torch.nn.utils.parameters_to_vector(policy.parameters())
n_param = W.shape[0]
if zero_policy:
W = torch.zeros_like(W)
r_hist = []
s_mean = torch.zeros(env.observation_space.shape[0])
s_stdv = torch.ones(env.observation_space.shape[0])
total_steps = 0
exp_dist = torch.distributions.Normal(torch.zeros(n_delta, n_param),
torch.ones(n_delta, n_param))
do_rollout_partial = partial(do_rollout_train, env_name, policy,
postprocess)
for _ in range(n_epochs):
deltas = exp_dist.sample()
pm_W = torch.cat((W + (deltas * exp_noise), W - (deltas * exp_noise)))
results = pool.map(do_rollout_partial, pm_W)
states = torch.empty(0)
p_returns = []
m_returns = []
l_returns = []
top_returns = []
for p_result, m_result in zip(results[:n_delta], results[n_delta:]):
ps, pr, plr = p_result
ms, mr, mlr = m_result
states = torch.cat((states, ms, ps), dim=0)
p_returns.append(pr)
m_returns.append(mr)
l_returns.append(plr)
l_returns.append(mlr)
top_returns.append(max(pr, mr))
top_idx = sorted(range(len(top_returns)),
key=lambda k: top_returns[k],
reverse=True)[:n_top]
p_returns = torch.stack(p_returns)[top_idx]
m_returns = torch.stack(m_returns)[top_idx]
l_returns = torch.stack(l_returns)[top_idx]
r_hist.append(l_returns.mean())
ep_steps = states.shape[0]
s_mean = update_mean(states, s_mean, total_steps)
s_stdv = update_std(states, s_stdv, total_steps)
total_steps += ep_steps
policy.state_means = s_mean
policy.state_std = s_stdv
do_rollout_partial = partial(do_rollout_train, env_name, policy,
postprocess)
W = W + (step_size / (n_delta * torch.cat(
(p_returns, m_returns)).std() + 1e-6)) * torch.sum(
(p_returns - m_returns) * deltas[top_idx].T, dim=1)
pool.terminate()
torch.nn.utils.vector_to_parameters(W, policy.parameters())
return policy, r_hist
def check_score():
print(args)
p_size = cpu_num
print("use cpu num:{}".format(p_size))
loss_history = []
check_deck_id = int(
args.check_deck_id) if args.check_deck_id is not None else None
cuda_flg = args.cuda == "True"
#node_num = int(args.node_num)
#net = New_Dual_Net(node_num)
model_name = args.model_name
existed_output_list = os.listdir(path="./Battle_Result")
existed_output_list = [
f for f in existed_output_list
if os.path.isfile(os.path.join("./Battle_Result", f))
]
result_name = "{}:{}".format(model_name.split(".")[0], args.deck_list)
same_name_count = len(
[1 for cell in existed_output_list if result_name in cell])
print("same_name_count:", same_name_count)
result_name += "_{:0>3}".format(same_name_count + 1)
PATH = 'model/' + model_name
model_dict = torch.load(PATH)
n_size = model_dict["final_layer.weight"].size()[1]
net = New_Dual_Net(n_size, hidden_num=args.hidden_num[0])
net.load_state_dict(model_dict)
opponent_net = None
if args.opponent_model_name is not None:
# opponent_net = New_Dual_Net(node_num)
o_model_name = args.opponent_model_name
PATH = 'model/' + o_model_name
model_dict = torch.load(PATH)
n_size = model_dict["final_layer.weight"].size()[1]
opponent_net = New_Dual_Net(n_size, hidden_num=args.hidden_num[1])
opponent_net.load_state_dict(model_dict)
if torch.cuda.is_available() and cuda_flg:
net = net.cuda()
opponent_net = opponent_net.cuda(
) if opponent_net is not None else None
print("cuda is available.")
#net.zero_grad()
deck_sampling_type = False
if args.deck is not None:
deck_sampling_type = True
G = Game()
net.cpu()
t3 = datetime.datetime.now()
if args.greedy_mode is not None:
p1 = Player(9, True, policy=Dual_NN_GreedyPolicy(origin_model=net))
else:
p1 = Player(9,
True,
policy=New_Dual_NN_Non_Rollout_OM_ISMCTSPolicy(
origin_model=net,
cuda=cuda_flg,
iteration=args.step_iter),
mulligan=Min_cost_mulligan_policy())
#p1 = Player(9, True, policy=AggroPolicy())
p1.name = "Alice"
if fixed_opponent is not None:
if fixed_opponent == "Aggro":
p2 = Player(9,
False,
policy=AggroPolicy(),
mulligan=Min_cost_mulligan_policy())
elif fixed_opponent == "OM":
p2 = Player(9, False, policy=Opponent_Modeling_ISMCTSPolicy())
elif fixed_opponent == "NR_OM":
p2 = Player(9,
False,
policy=Non_Rollout_OM_ISMCTSPolicy(iteration=200),
mulligan=Min_cost_mulligan_policy())
elif fixed_opponent == "ExItGreedy":
tmp = opponent_net if opponent_net is not None else net
p2 = Player(9,
False,
policy=Dual_NN_GreedyPolicy(origin_model=tmp))
elif fixed_opponent == "Greedy":
p2 = Player(9,
False,
policy=New_GreedyPolicy(),
mulligan=Simple_mulligan_policy())
elif fixed_opponent == "Random":
p2 = Player(9,
False,
policy=RandomPolicy(),
mulligan=Simple_mulligan_policy())
else:
assert opponent_net is not None
p2 = Player(9,
False,
policy=New_Dual_NN_Non_Rollout_OM_ISMCTSPolicy(
origin_model=opponent_net, cuda=cuda_flg),
mulligan=Min_cost_mulligan_policy())
# p2 = Player(9, False, policy=RandomPolicy(), mulligan=Min_cost_mulligan_policy())
p2.name = "Bob"
Battle_Result = {}
deck_list = tuple(map(int, args.deck_list.split(",")))
print(deck_list)
test_deck_list = deck_list # (0,1,4,10,13)
test_deck_list = tuple(itertools.product(test_deck_list, test_deck_list))
test_episode_len = evaluate_num #100
episode_num = evaluate_num
match_num = len(test_deck_list)
manager = Manager()
shared_array = manager.Array("i",
[0 for _ in range(3 * len(test_deck_list))])
iter_data = [(p1, p2, shared_array, episode_num, p_id, test_deck_list)
for p_id in range(p_size)]
freeze_support()
p1_name = p1.policy.name.replace("origin", args.model_name)
if args.opponent_model_name is not None:
p2_name = p2.policy.name.replace("origin", args.opponent_model_name)
else:
p2_name = p2.policy.name.replace("origin", args.model_name)
print(p1_name)
print(p2_name)
pool = Pool(p_size, initializer=tqdm.set_lock,
initargs=(RLock(), )) # 最大プロセス数:8
memory = pool.map(multi_battle, iter_data)
pool.close() # add this.
pool.terminate() # add this.
print("\n" * (match_num + 1))
memory = list(memory)
min_WR = 1.0
Battle_Result = {(deck_id[0], deck_id[1]): \
tuple(shared_array[3*index+1:3*index+3]) for index, deck_id in enumerate(test_deck_list)}
print(shared_array)
txt_dict = {}
for key in sorted(list((Battle_Result.keys()))):
cell = "{}:WR:{:.2%},first_WR:{:.2%}"\
.format(key,Battle_Result[key][0]/test_episode_len,2*Battle_Result[key][1]/test_episode_len)
print(cell)
txt_dict[key] = cell
print(Battle_Result)
# result_name = "{}:{}_{}".format(model_name.split(".")[0],args.deck_list,)
# result_name = model_name.split(".")[0] + ":" + args.deck_list + ""
deck_num = len(deck_list)
# os.makedirs("Battle_Result", exist_ok=True)
with open("Battle_Result/" + result_name, "w") as f:
writer = csv.writer(f, delimiter='\t', lineterminator='\n')
row = ["{} vs {}".format(p1_name, p2_name)]
deck_names = [deck_id_2_name[deck_list[i]] for i in range(deck_num)]
row = row + deck_names
writer.writerow(row)
for i in deck_list:
row = [deck_id_2_name[i]]
for j in deck_list:
row.append(Battle_Result[(i, j)])
writer.writerow(row)
for key in list(txt_dict.keys()):
writer.writerow([txt_dict[key]])
table_prep_params['LENGTH_PER_CELL'])
for i in range(0, table_prep_params['MAX_ROW_LEN'] - rows):
table.append(
[['<PAD>'] * table_prep_params['LENGTH_PER_CELL']] *
table_prep_params['MAX_COL_LEN'])
return table
def table_words2index(tables):
w2i = {w: i for i, w in enumerate(vocab)}
for i, t in enumerate(tables):
tables[i] = np.vectorize(lambda y: w2i[y])(
np.array(t)).tolist()
return tables
p = Pool(processes=40)
X = p.map(pad_table, X)
p.close()
p.join()
X = table_words2index(X)
X = np.array(X)
print(X.shape)
savepkl('./data/xp_2D_10-50_pad.pkl', X)
else:
device = torch.device(
f"cuda:{1}" if torch.cuda.is_available() else 'cpu')
dataset = T2VDataset(X, y, vocab, device, config)
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)
X_, y_ = next(iter(dataloader))
print(X_.shape, y_.shape)
print(time.time() - start)
