def evaluate(self,
policy,
n_episodes=1,
initial_state=None,
transform_to_internal_state=None,
render=False,
results_dir='/tmp/'):
"""
Evaluate a policy on the mdp.
:param policy: a policy outputing an action
:type policy: Policy
:param n_episodes: number of episodes to evaluate
:type n_episodes: int
:param initial_state: the initial state to evaluate from
:type initial_state: int
:param transform_to_internal_state: transform the states or initial state to an internal state of the mdp
:type transform_to_internal_state: function
:param render: reder the environment
:type render: bool
:return: a Dataset with the trajectories
"""
dataset = Dataset(results_dir=results_dir)
for episode in range(n_episodes):
trajectory = []
state = self._env.reset()
if initial_state is not None:
state = initial_state
if self._quanser_robots:
self._env.env._sim_state = np.copy(
transform_to_internal_state(state))
else:
self._env.env.state = np.copy(
transform_to_internal_state(state))
for j in range(self._env._max_episode_steps):
with torch.no_grad():
state = torch.tensor(state,
device=policy.device,
dtype=TORCH_DTYPE)
action = policy(state).to('cpu').numpy().reshape((-1, ))
state_next, rew, done, _ = self._env.step(action)
trajectory.append((state.to('cpu').numpy(), action, rew,
state_next, done))
state = state_next
if render:
self._env.render()
if done:
break
dataset.add_trajectory(trajectory)
self._env.close()
return dataset
def build_dataset(self, examples_path, labels_path):
""" returns an instance of ParaphraseDatasat containing a list of examples"""
examples = self.get_examples(examples_path)
labels = self.get_labels(labels_path)
for example, label in zip(examples, labels):
example.label = label
return Dataset(examples)
def get_train_and_test():
# Read the data and split into train and test.
sc = SentenceAssigner("../data/datasetSplit.txt")
dataset = Dataset("../data/datasetSentences.txt", sc)
train_corpus = dataset.train_corpus
test_corpus = dataset.test_corpus
return train_corpus, test_corpus
def get_evaluation(self, sess, data_list, data_type, global_step=None):
_logger.add()
_logger.add('getting evaluation result')
logits_list, loss_list, accu_list = [], [], []
for sample_batch, _, _, _ in Dataset.generate_batch_sample_iter(
data_list):
feed_dict = self.model.get_feed_dict(sample_batch, 'dev')
logits, loss, accu = sess.run(
[self.model.logits, self.model.loss, self.model.accuracy],
feed_dict)
logits_list.append(np.argmax(logits, -1))
loss_list.append(loss)
accu_list.append(accu)
logits_array = np.concatenate(logits_list, 0)
loss_value = np.mean(loss_list)
accu_array = np.concatenate(accu_list, 0)
accu_value = np.mean(accu_array)
if global_step is not None:
if data_type == 'train':
summary_feed_dict = {
self.train_loss: loss_value,
self.train_accuracy: accu_value,
}
summary = sess.run(self.train_summaries, summary_feed_dict)
self.writer.add_summary(summary, global_step)
elif data_type == 'dev':
summary_feed_dict = {
self.dev_loss: loss_value,
self.dev_accuracy: accu_value,
}
summary = sess.run(self.dev_summaries, summary_feed_dict)
self.writer.add_summary(summary, global_step)
else:
summary_feed_dict = {
self.test_loss: loss_value,
self.test_accuracy: accu_value,
}
summary = sess.run(self.test_summaries, summary_feed_dict)
self.writer.add_summary(summary, global_step)
return loss_value, accu_value
kwargs={
'fs': 200.0,
'fs_ctrl': 200.0
})
env = gym.make('Qube-100-v1')
mdp = MDP(env)
##########################################################################################
# Gather an Off-Policy Dataset
results_dir = '/home/carvalho/Documents/projects/nopg/results/qube/nopgs/'
os.makedirs(results_dir, exist_ok=True)
# Load trajectories from file
filename = '/home/carvalho/Documents/projects/nopg/datasets/qube/15_trajectories.npy'
dataset = Dataset(results_dir=results_dir)
dataset.load_trajectories_from_file(filename, n_trajectories=8)
dataset.update_dataset_internal()
s_band_factor = [15., 15., 15, 15., 1., 1.]
s_n_band_factor = s_band_factor
a_band_factor = [7.]
dataset.kde_bandwidths_internal(s_band_factor=s_band_factor,
a_band_factor=a_band_factor,
s_n_band_factor=s_n_band_factor)
dataset.plot_data_kde(state_labels=mdp._env.observation_space.labels,
action_labels=mdp._env.action_space.labels)
##########################################################################################
# Define the Policy Network
policy_params = {
from src.dataset.dataset import Dataset
# Get daily data for a state in India
kl_time_series = Dataset("India:KL")
kl_confirmed = kl_time_series.get_confirmed()
print(kl_confirmed)
# Get daily data for a country
india_time_series = Dataset("India")
india_confirmed = india_time_series.get_confirmed()
print(india_confirmed)
# Get all states dataframe
all_states = Dataset("India:AllStates")
states_df = all_states.get_confirmed()
print(states_df)
def train():
output_model_params()
loadFile = True
ifLoad, data = False, None
if loadFile:
ifLoad, data = load_file(cfg.processed_path, 'data', 'pickle')
if not ifLoad or not loadFile:
data_object = Dataset(cfg.train_data_path, cfg.dev_data_path)
data_object.save_dict(cfg.dict_path)
save_file({'data_obj': data_object}, cfg.processed_path)
else:
data_object = data['data_obj']
emb_mat_token, emb_mat_glove = data_object.emb_mat_token, data_object.emb_mat_glove
with tf.variable_scope(network_type) as scope:
if network_type in model_set:
model = Model(emb_mat_token, emb_mat_glove,
len(data_object.dicts['token']),
len(data_object.dicts['char']),
data_object.max_lens['token'], scope.name)
graphHandler = GraphHandler(model)
evaluator = Evaluator(model)
performRecoder = PerformRecoder(5)
if cfg.gpu_mem < 1.:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem, allow_growth=True)
else:
gpu_options = tf.GPUOptions()
graph_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True)
sess = tf.Session(config=graph_config)
graphHandler.initialize(sess)
# begin training
steps_per_epoch = int(
math.ceil(1.0 * len(data_object.digitized_train_data_list) /
cfg.train_batch_size))
num_steps = steps_per_epoch * cfg.max_epoch or cfg.num_steps
global_step = 0
# debug or not
if cfg.debug:
sess = tf_debug.LocalCLIDebugWrapperSession(sess)
for sample_batch, batch_num, data_round, idx_b in Dataset.generate_batch_sample_iter(
data_object.digitized_train_data_list, num_steps):
global_step = sess.run(model.global_step) + 1
if_get_summary = global_step % (cfg.log_period or steps_per_epoch) == 0
loss, summary, train_op = model.step(sess,
sample_batch,
get_summary=if_get_summary)
if global_step % 10 == 0:
_logger.add('data round: %d: %d/%d, global step:%d -- loss: %.4f' %
(data_round, idx_b, batch_num, global_step, loss))
if if_get_summary:
graphHandler.add_summary(summary, global_step)
# Occasional evaluation
if global_step % (cfg.eval_period or steps_per_epoch) == 0:
# ---- dev ----
dev_loss, dev_accu = evaluator.get_evaluation(
sess, data_object.digitized_dev_data_list, 'dev', global_step)
_logger.add('==> for dev, loss: %.4f, accuracy: %.4f' %
(dev_loss, dev_accu))
# ---- test ----
if cfg.test_data_name != None:
test_loss, test_accu = evaluator.get_evaluation(
sess, data_object.digitized_test_data_list, 'test',
global_step)
_logger.add('~~> for test, loss: %.4f, accuracy: %.4f' %
(test_loss, test_accu))
is_in_top, deleted_step = performRecoder.update_top_list(
global_step, dev_accu, sess)
this_epoch_time, mean_epoch_time = cfg.time_counter.update_data_round(
data_round)
if this_epoch_time is not None and mean_epoch_time is not None:
_logger.add('##> this epoch time: %f, mean epoch time: %f' %
(this_epoch_time, mean_epoch_time))
def test():
output_model_params()
loadFile = True
ifLoad, data = False, None
if loadFile:
ifLoad, data = load_file(cfg.processed_path, 'processed data',
'pickle')
if not ifLoad or not loadFile:
train_data_obj = Dataset(cfg.train_data_path, 'train')
dev_data_obj = Dataset(cfg.dev_data_path, 'dev', train_data_obj.dicts)
test_data_obj = Dataset(cfg.test_data_path, 'test',
train_data_obj.dicts)
save_file(
{
'train_data_obj': train_data_obj,
'dev_data_obj': dev_data_obj,
'test_data_obj': test_data_obj
}, cfg.processed_path)
train_data_obj.save_dict(cfg.dict_path)
else:
train_data_obj = data['train_data_obj']
dev_data_obj = data['dev_data_obj']
test_data_obj = data['test_data_obj']
emb_mat_token, emb_mat_glove = train_data_obj.emb_mat_token, train_data_obj.emb_mat_glove
with tf.variable_scope(network_type) as scope:
if network_type in model_set:
model = Model(emb_mat_token, emb_mat_glove,
len(train_data_obj.dicts['token']),
len(train_data_obj.dicts['char']),
train_data_obj.max_lens['token'], scope.name)
graphHandler = GraphHandler(model)
evaluator = Evaluator(model)
if cfg.gpu_mem < 1.:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem, allow_growth=True)
else:
gpu_options = tf.GPUOptions()
graph_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True)
# graph_config.gpu_options.allow_growth = True
sess = tf.Session(config=graph_config)
graphHandler.initialize(sess)
# ---- dev ----
dev_loss, dev_accu = evaluator.get_evaluation(sess, dev_data_obj, 1)
_logger.add('==> for dev, loss: %.4f, accuracy: %.4f' %
(dev_loss, dev_accu))
# ---- test ----
test_loss, test_accu = evaluator.get_evaluation(sess, test_data_obj, 1)
_logger.add('~~> for test, loss: %.4f, accuracy: %.4f' %
(test_loss, test_accu))
# ---- train ----
train_loss, train_accu, train_sent_accu = evaluator.get_evaluation(
sess, train_data_obj, 1)
_logger.add('--> for test, loss: %.4f, accuracy: %.4f' %
(train_loss, train_accu))
def get_samples(self,
sampling_type=None,
states=None,
actions=None,
transform_to_internal_state=lambda x: x,
policy=None,
n_samples=None,
n_trajectories=None,
initial_state=None,
max_ep_transitions=100,
render=False,
press_enter_to_start=False,
dataset=None,
results_dir='/tmp/'):
"""
Collect samples from the mdp.
If sampling_type is 'uniform', provide the grid of states and actions to samples uniformly from, and
additionally transform_to_internal_state function.
If sampling_type is 'behavioral', provide at least the behavioral policy, and no states and actions.
NOTE: opt between either n_samples or n_trajectories. In the case of n_trajectories specify the maximum number
of transitions per episode max_ep_transitions.
:param sampling_type: can be uniform or behavioral
:type sampling_type: str
:param states: a batch of states to sample on
:type states: ndarray
:param actions: a batch of actions to sample each state on
:type actions: ndarray
:param transform_to_internal_state: transform the states or initial state to an internal state of the mdp
:type transform_to_internal_state: function
:param policy: behavioral policy outputing an action given a state
:type policy: callable function
:param n_samples: number of samples to collect (EXCLUSIVE between n_samples and n_trajectories)
:type n_samples: int
:param n_trajectories: number of trajectories to collect (EXCLUSIVE between n_samples and n_trajectories)
:type n_trajectories: int
:param initial_state: initial state to start sampling from (e.g., in the pendulum we want to start from the
upright position)
:type initial_state: ndarray
:param max_ep_transitions: maximum number of transitions before reseting to the initial state
:type max_ep_transitions: int
:param render: render the environment
:type render: int
:param press_enter_to_start: wait for user input to start collecting trajectories
:type press_enter_to_start: bool
:param dataset: if a dataset is provided add trajectories to it, else start an empty one
:type dataset: Dataset
:param results_dir: provide a directory path to save intermediate results
:type results_dir: str
:return: a dataset containing the collected trajectories
"""
dataset = dataset if dataset is not None else Dataset(
results_dir=results_dir)
if sampling_type == 'uniform':
for state in states:
for action in actions:
self._env.reset()
if self._quanser_robots:
self._env.env._sim_state = np.copy(
transform_to_internal_state(state))
else:
self._env.env.state = np.copy(
transform_to_internal_state(state))
state_next, rew, done, _ = self._env.step(action)
dataset.add_trajectory([(state, action, rew, state_next,
done)])
elif sampling_type == 'behavioral':
behavioral_policy = policy
i = 0
n_samples_collected = 0
criteria = 0
if n_samples is not None:
criteria = n_samples
elif n_trajectories is not None:
criteria = n_trajectories
while i < criteria:
if press_enter_to_start:
if render:
self._env.render()
input("Press ENTER to start")
state = self._env.reset()
if initial_state is not None:
state = initial_state
if self._quanser_robots:
self._env.env._sim_state = np.copy(
transform_to_internal_state(state))
else:
self._env.env.state = np.copy(
transform_to_internal_state(state))
trajectory = []
for j in range(max_ep_transitions):
if render:
self._env.render()
action = behavioral_policy(x=state)
state_next, rew, done, _ = self._env.step(action)
trajectory.append((state, action, rew, state_next, done))
n_samples_collected += 1
print("\r{}/{} - {}".format(i + 1, criteria,
n_samples_collected),
end=' ',
flush=True)
if n_samples is not None:
i += 1
if done or (i >= n_samples
if n_samples is not None else False):
break
state = state_next
dataset.add_trajectory(trajectory)
if n_trajectories is not None:
i += 1
if i >= n_trajectories:
break
self._env.close()
return dataset
def test_split(self):
return Dataset(self.test_portion, self.test_labels)
epochs=args.epochs,
device=torch.device(args.device),
tokenizer=transformers.AutoTokenizer.from_pretrained(args.model),
)
"""
model = TransformerWrapper.load_pretrained(
args.pretrained_model,
params=configuration,
pooler = BertPoolingStrategy(configuration),
loss = SoftmaxLoss(configuration))
"""
model = DistilBertForSequenceClassification.from_pretrained(
args.pretrained_model, num_labels=len(LABELS_TO_ID))
train_split, valid_split = dataset.split_dataset(test_perc=0.1)
train_dataset = Dataset(train_split)
valid_dataset = Dataset(valid_split)
print(f"train dataset size: {len(train_dataset)}")
print(f"valid dataset size: {len(valid_dataset)}")
valid_data_loader = SmartParaphraseDataloader.build_batches(
valid_dataset, 16, mode="sequence", config=configuration)
evaluator = ClassificationEvaluator(params=configuration,
model=model,
data_loader=valid_data_loader,
device=args.device,
metrics=metrics,
fp16=True,
verbose=True)
# Use the GPU if available, or if the memory is insufficient use only the CPU
# DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
DEVICE = torch.device('cpu')
##########################################################################################
# Create the Environment (MDP)
env = gym.make('MountainCarContinuous-v0')
mdp = MDP(env)
##########################################################################################
# Gather an Off-Policy Dataset
# Load trajectories from file
filename = '../datasets/mountaincar/10_trajectories.npy'
dataset = Dataset()
dataset.load_trajectories_from_file(filename, n_trajectories=5)
dataset.update_dataset_internal()
s_band_factor = [2., 2.]
s_n_band_factor = s_band_factor
a_band_factor = [50.]
dataset.kde_bandwidths_internal(s_band_factor=s_band_factor,
a_band_factor=a_band_factor,
s_n_band_factor=s_n_band_factor)
##########################################################################################
# Define the Policy Network
policy_params = {
'policy_class': 'stochastic',
'neurons': [mdp.s_dim, 50,
TEST_PATH = DATA / 'test'
TEST = TEST_PATH / 'Covid/0000.png'
CWD = Path.cwd()
OUTPUT_PATH = CWD / 'outputs'
CLEAR = False
LABELS = ['Covid','Normal','Pneumonia']
if CLEAR:
remove_folder([OUTPUT_PATH, Path('./logs'), Path('./build')])
nets_path = create_folders(name=OUTPUT_PATH, nets=NETS)
# %% [code]
# np.random.seed(seed=42)
labels = listdir(TRAIN_PATH)
ds_train = Dataset(path_data=TRAIN_PATH, train=False)
ds_test = Dataset(path_data=TEST_PATH, train=False)
part_param = {'tamanho': 10}
train, validation = ds_train.partition(val_size=0.2, **part_param)
test_values, _test_val_v = ds_test.partition(val_size=1e-5, **part_param)
params = {
'dim': DIM_SPLIT,
'batch_size': BATCH_SIZE,
'n_class': len(labels),
'channels': CHANNELS
}
train_generator = DataGenerator(x_set=train[0], y_set=train[1], **params)
val_generator = DataGenerator(x_set=validation[0], y_set=validation[1], **params)
test_generator = DataGenerator(x_set=test_values[0], y_set=test_values[1], **params)
centre=(0.0, 0.5),
axis_ratio=0.75,
phi=45.0,
intensity=0.1,
effective_radius=0.5,
sersic_index=1.0,
)
image = model.profile_image_from_grid(grid=grid)
# plt.figure()
# plt.imshow(image.in_2d)
# plt.show()
# exit()
noise_map = np.random.normal(loc=0.0, scale=10**-2.0, size=grid.shape_2d)
dataset = Dataset(data=np.add(image.in_2d, noise_map), noise_map=noise_map)
# plt.figure()
# plt.imshow(dataset.data)
# plt.show()
# exit()
# masked_dataset = MaskedDataset(
# dataset=dataset,
# mask=mask
# )
# likelihood = fit.likelihood_from_chi_squared_and_noise_normalization(
# chi_squared=fit.chi_squared_from_chi_squared_map_and_mask(
# chi_squared_map=fit.chi_squared_map_from_residual_map_noise_map_and_mask(
# residual_map=fit.residual_map_from_data_model_data_and_mask(
# data=dataset.data,
from quanser_robots import GentlyTerminating
from quanser_robots.qube import SwingUpCtrl
from src.dataset.dataset import Dataset
register(id='Qube-100-v1',
entry_point='quanser_robots.qube.qube:Qube',
max_episode_steps=1500,
kwargs={
'fs': 200.0,
'fs_ctrl': 200.0
})
env = GentlyTerminating(gym.make('Qube-100-v1'))
dataset = Dataset()
n_trajectories = 1
for traj in range(n_trajectories):
trajectory = []
ctrl = SwingUpCtrl()
obs = env.reset()
done = False
while not done:
env.render()
act = ctrl(obs)
obs_n, r, done, _ = env.step(act)
trajectory.append((obs, act, r, obs_n, done))
if done:
break
obs = np.copy(obs_n)
env.close()
shape_3d=grid_3d.shape_3d,
z_step_kms=z_step_kms
)
# plot_utils.plot_cube(
# cube=cube,
# ncols=8
# )
noise_map = np.random.normal(
loc=0.0, scale=10**-4.0, size=grid_3d.shape_3d
)
dataset = Dataset(
data=np.add(
cube,
noise_map
),
noise_map=noise_map,
z_step_kms=z_step_kms
)
# plot_utils.plot_cube(
# cube=dataset.data,
# ncols=8
# )
# exit()
masked_dataset = MaskedDataset(
dataset=dataset,
mask=mask_3d
)
transformer = transformer_class(uv_wavelengths=uv_wavelengths,
grid=grid.in_radians)
visibilities = tracer.profile_visibilities_from_grid_and_transformer(
grid=grid, transformer=transformer)
if data_with_phase_errors:
visibilities = corrupt_visibilities_from_f_matrix_and_phase_errors(
visibilities=visibilities, f=f, phase_errors=phase_errors)
noise_map = np.random.normal(loc=0.0,
scale=1.0 * 10**-2.0,
size=visibilities.shape)
dataset = Dataset(uv_wavelengths=uv_wavelengths,
visibilities=np.add(visibilities, noise_map),
noise_map=noise_map,
antennas=antennas)
xy_mask = al.Mask.unmasked(
shape_2d=grid.shape_2d,
pixel_scales=grid.pixel_scales,
sub_size=grid.sub_size,
)
def test(dataset,
xy_mask,
tracer,
self_calibration,
transformer_class=al.TransformerFINUFFT):
masked_dataset = MaskedDataset(dataset=dataset, xy_mask=xy_mask)
loss.backward()
optimizer.step()
return test(model, validate_loader)
if __name__ == '__main__':
random.seed(SEED)
os.environ['PYTHONHASHSEED'] = str(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(SEED)
reader = DataReader('isear_v2/')
reader.read_datas()
train_dataset = Dataset(reader.train_datas, reader.word_num)
validate_dataset = Dataset(reader.validate_datas, reader.word_num)
test_dataset = Dataset(reader.test_datas, reader.word_num)
train_loader = DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
collate_fn=collate_data)
validate_loader = DataLoader(dataset=validate_dataset,
batch_size=len(validate_dataset),
collate_fn=collate_data)
test_loader = DataLoader(dataset=test_dataset,
batch_size=len(test_dataset),
collate_fn=collate_data)
model = eval(args_ori.model)(reader.word_num, reader.label_num,
**vars(args))
# visibilities=visibilities,
# transformers=transformers,
# shape=cube.shape
# ),
# ncols=8,
# cube_contours=cube,
# )
# exit()
noise_map = np.random.normal(
loc=0.0, scale=5.0 * 10**-1.0, size=visibilities.shape
)
dataset = Dataset(
uv_wavelengths=uv_wavelengths,
visibilities=np.add(
visibilities, noise_map
),
noise_map=noise_map,
z_step_kms=z_step_kms
)
# plot_utils.plot_cube(
# cube=dirty_cube_from_visibilities(
# visibilities=dataset.visibilities,
# transformers=transformers,
# shape=cube.shape,
# invert=True
# ),
# ncols=8,
# cube_contours=cube,
#
# )
# exit()
def train_split(self):
return Dataset(self.examples, self.labels)
