def predict(net, labels, files, params):
print('starting inference')
device = torch.device(params.device)
predictions = []
probs = []
for i, file in enumerate(files):
filename = os.path.splitext(os.path.basename(file))[0]
processed = filename + '_proc.wav'
pre.preprocess(file, processed)
data = vggish_input.wavfile_to_examples(processed)
data = torch.from_numpy(data).unsqueeze(1).float()
data = data.to(device)
net.to(device)
out = net(data)
# # for each spectrogram/row index of max probability
# pred = np.argmax(out.detach().cpu().numpy(), axis=1)
# # find most frequent index over all spectrograms
# consensus = np.bincount(pred).argmax()
# print('file {} sounds like a {} to me'.format(i, labels[consensus]))
# mean probabilities for each col/class over all spectrograms
mean_probs = np.mean(out.detach().cpu().numpy(), axis=0)
# find index of max mean_probs
idx = np.argmax(mean_probs)
print('file {} sounds like a {} to me'.format(i, labels[idx]))
print('my guesses are: ')
for j, label in enumerate(labels):
print('{0}: {1:.04f}'.format(label, mean_probs[j]))
# predictions.append(labels[consensus])
predictions.append(labels[idx])
probs.append(mean_probs)
os.remove(processed)
return predictions, probs
def create_multiartifact_sample(self, depthmap):
depthmaps = np.zeros((240, 180, 5))
for i, depthmap_path in enumerate(depthmap[0]):
data, width, height, depth_scale, _max_confidence = preprocessing.load_depth(
depthmap_path)
depthmap = preprocessing.prepare_depthmap(data, width, height,
depth_scale)
depthmap = preprocessing.preprocess(depthmap)
depthmaps[:, :, i] = tf.squeeze(depthmap, axis=2)
depthmaps = tf.stack([depthmaps])
return depthmaps
def process_depthmaps(self):
depthmaps = []
for artifact in self.artifacts:
input_path = self.get_input_path(self.scan_directory,
artifact['file'])
data, width, height, depthScale, _max_confidence = preprocessing.load_depth(
input_path)
depthmap = preprocessing.prepare_depthmap(data, width, height,
depthScale)
depthmap = preprocessing.preprocess(depthmap)
depthmaps.append(depthmap)
depthmaps = np.array(depthmaps)
return depthmaps
def main(data, k_folds):
"""Trains a new ML model and uploads it to S3."""
df = pd.read_csv(data, sep=";")
df = preprocess(df)
results = cross_validate_performance(df=df, n=k_folds)
avg_results = average_evaluation_metrics(results)
print(f"== Model Evaluation Results (Folds: {k_folds}) ==")
for metric, value in avg_results.items():
print(f"{metric.upper()}: {value}")
print(f"=================================================")
X, y = Xy_split(df)
model = WineQualityModel()
model.fit(X, y)
model.save(path=S3_PATH)
return
def main(
df_path:
str = '/project/cq-training-1/project1/data/catalog.helios.public.20100101-20160101.pkl',
image_size: int = 32,
model: str = 'dummy',
epochs: int = 20,
optimizer: str = 'adam',
lr: float = 1e-4,
batch_size: int = 100,
subset_perc: float = 1,
subset_dates: bool = False,
saved_model_dir: str = None,
seq_len: int = 6,
seed: bool = True,
scale_label: bool = True,
use_csky: bool = False,
cache: bool = True,
timesteps_minutes: int = 15):
# Warning if no GPU detected
if len(tf.config.list_physical_devices('GPU')) == 0:
logger.warning('No GPU detected, training will run on CPU.')
elif len(tf.config.list_physical_devices('GPU')) > 1:
logger.warning(
'Multiple GPUs detected, training will run on only one GPU.')
if subset_dates and subset_perc != 1:
raise Exception(
f'Invalid configuration. Argument --subset_dates=True and --subset_perc={subset_perc}.'
)
# Set random seed
if seed:
tf.random.set_seed(SEED)
np.random.seed(SEED)
# Load dataframe
logger.info('Loading and preprocessing dataframe...')
df = pd.read_pickle(df_path)
df = preprocessing.preprocess(df, shuffle=False, scale_label=scale_label)
metadata = data.Metadata(df, scale_label)
# Pre-crop data
logger.info('Getting crops...')
images = data.Images(metadata, image_size)
# images.crop(dest=SLURM_TMPDIR)
images.crop(dest=images.shared_storage)
# Split into train and valid
if subset_dates:
metadata_train, metadata_valid = metadata.split_with_dates()
else:
metadata, _ = metadata.split(1 - subset_perc)
metadata_train, metadata_valid = metadata.split(VALID_PERC)
nb_train_examples = metadata_train.get_number_of_examples()
nb_valid_examples = metadata_valid.get_number_of_examples()
logger.info(
f'Number of training examples : {nb_train_examples}, number of validation examples : \
{nb_valid_examples}')
# Create model
if model == 'dummy':
model = baselines.DummyModel()
elif model == 'sunset':
model = baselines.SunsetModel()
elif model == 'cnndem':
model = baselines.ConvDemModel(image_size)
elif model == 'sunset3d':
model = baselines.Sunset3DModel()
elif model == 'convlstm':
model = baselines.ConvLSTM()
elif model == 'cnngru':
model = CnnGru(seq_len)
elif model == 'cnngruatt':
model = CnnGruAtt(seq_len)
elif model == 'cnnlstm':
model = LSTM_Resnet(seq_len)
elif model == 'resnet':
model = baselines.ResNetModel()
else:
raise Exception(f'Model "{model}" not recognized.')
# Load model weights
if saved_model_dir is not None:
model.load_weights(os.path.join(saved_model_dir, "model"))
# Loss and optimizer
mse = tf.keras.losses.MeanSquaredError()
if optimizer == 'adam':
optimizer = tf.keras.optimizers.Adam(lr)
elif optimizer == 'sgd':
optimizer = tf.keras.optimizers.SGD(lr)
else:
raise Exception(f'Optimizer "{optimizer}" not recognized.')
# Create data loader
dataloader_train = SequenceDataset(
metadata_train,
images,
seq_len,
batch_size,
timesteps=datetime.timedelta(minutes=timesteps_minutes),
cache=cache)
dataloader_valid = SequenceDataset(
metadata_valid,
images,
seq_len,
batch_size,
timesteps=datetime.timedelta(minutes=timesteps_minutes),
cache=cache)
# Training loop
logger.info('Training...')
losses = {'train': [], 'valid': []}
best_valid_loss = float('inf')
for epoch in range(epochs):
train_epoch(model, dataloader_train, batch_size, mse, optimizer,
nb_train_examples, scale_label, use_csky)
test_epoch(model, dataloader_valid, batch_size, mse, nb_valid_examples,
scale_label, use_csky)
train_loss = np.sqrt(train_mse_metric.result().numpy())
valid_loss = np.sqrt(valid_mse_metric.result().numpy())
csky_valid_loss = np.sqrt(valid_csky_mse_metric.result().numpy())
if valid_loss < best_valid_loss:
best_valid_loss = valid_loss
utils.save_model(model)
# Logs
logger.info(
f'Epoch {epoch} - Train Loss : {train_loss:.4f}, Valid Loss : {valid_loss:.4f}, Csky Valid Loss : \
{csky_valid_loss:.4f}')
losses['train'].append(train_loss)
losses['valid'].append(valid_loss)
with train_summary_writer.as_default():
tf.summary.scalar('loss', train_loss, step=epoch)
with test_summary_writer.as_default():
tf.summary.scalar('loss', valid_loss, step=epoch)
# Plot losses
plots.plot_loss(losses['train'], losses['valid'], csky_valid_loss)
def prepare_dataloader(
dataframe: pd.DataFrame,
target_datetimes: typing.List[datetime.datetime],
stations: typing.Dict[typing.AnyStr, typing.Tuple[float, float, float]],
target_time_offsets: typing.List[datetime.timedelta],
config: typing.Dict[typing.AnyStr, typing.Any],
) -> tf.data.Dataset:
"""This function should be modified in order to prepare & return your own data loader.
Note that you can use either the netCDF or HDF5 data. Each iteration over your data loader should return a
2-element tuple containing the tensor that should be provided to the model as input, and the target values.
In
this specific case, you will not be able to provide the latter since the dataframe contains no GHI, and we
are
only interested in predictions, not training. Therefore, you must return a placeholder (or ``None``) as the
second
tuple element.
Reminder: the dataframe contains imagery paths for every possible timestamp requested in
``target_datetimes``.
However, we expect that you will use some of the "past" imagery (i.e. imagery at T<=0) for any T in
``target_datetimes``, but you should NEVER rely on "future" imagery to generate predictions (for T>0). We
will be inspecting data loader implementations to ensure this is the case, and those who "cheat" will be
dramatically penalized.
See https://github.com/mila-iqia/ift6759/tree/master/projects/project1/evaluation.md for more information.
Args:
dataframe: a pandas dataframe that provides the netCDF file path (or HDF5 file path and offset) for all
relevant timestamp values over the test period.
target_datetimes: a list of timestamps that your data loader should use to provide imagery for your
model.
The ordering of this list is important, as each element corresponds to a sequence of GHI values
to predict. By definition, the GHI values must be provided for the offsets given by
``target_time_offsets``
which are added to each timestamp (T=0) in this datetimes list.
stations: a map of station names of interest paired with their coordinates (latitude, longitude,
elevation).
target_time_offsets: the list of timedeltas to predict GHIs for (by definition: [T=0, T+1h, T+3h,
T+6h]).
config: configuration dictionary holding any extra parameters that might be required by the user. These
parameters are loaded automatically if the user provided a JSON file in their submission.
Submitting
such a JSON file is completely optional, and this argument can be ignored if not needed.
Returns:
A ``tf.data.Dataset`` object that can be used to produce input tensors for your model. One tensor
must correspond to one sequence of past imagery data. The tensors must be generated in the order given
by ``target_sequences``.
"""
# Things to parse from the json config file
image_size = config['image_size']
seq_len = config['seq_len']
timesteps = datetime.timedelta(minutes=config['timesteps_minutes'])
scale_label = config['scale_label']
# Load dataframe
dataframe = preprocessing.preprocess(dataframe,
shuffle=False,
scale_label=scale_label)
metadata = data.Metadata(dataframe, scale_label)
# Build dataloader
data_loader = EvaluatorDataset(metadata,
image_size=image_size,
seq_len=seq_len,
timesteps=timesteps,
target_datetimes=target_datetimes,
stations=stations,
target_time_offsets=target_time_offsets)
return data_loader
def train():
discount_factor = 0.99
num_episodes = 100000000
exploration_rate_begin = 1
exploration_rate_end = 0.1
exploration_rate = exploration_rate_begin
exploration_decay = 100000
render = False
render_freq = 30
steps_done = 0
replay_size_start = 10000
batch_size = 64
update_model_freq = 16
update_target_freq = 1000
lr = 0.00025
momentum = 0.95
# init objects
buffer = Experience_buffer()
env = gym.make('FlappyBird-v0')
model = Model(env.action_space.n)
# model.apply(Model.weights_init)
target = Model(env.action_space.n)
optimizer = optim.RMSprop(params=model.parameters(), lr=lr, momentum=momentum)
loss = nn.SmoothL1Loss()
agent = DQNAgent(env, model, target, optimizer, loss, update_target_freq)
# let's play
for i in range(num_episodes):
# start a new episode
print('Episode #{}'.format(i))
done = False
episode_reward = 0
current_loss = 0
current_obs = env.reset()
current_obs = preprocess(current_obs)
# current_obs = Variable(torch.from_numpy(current_obs).unsqueezed(0), volatile=True)
if render:
env.render()
while not done:
action = agent.select_action(current_obs, exploration_rate)
next_obs, reward, done, _ = env.step(action)
next_obs = preprocess(next_obs)
if render:
env.render()
buffer.add_experience(current_obs, action, reward, next_obs, done)
# update data
current_obs = next_obs
episode_reward += reward
steps_done += 1
if steps_done > replay_size_start:
exploration_rate = exploration_rate_end + (exploration_rate_begin - exploration_rate_end) * math.exp(-1. * steps_done / exploration_decay)
# if the buffer is filled enough, periodically update the model
if len(buffer) > batch_size and steps_done % update_model_freq == 0 and steps_done > replay_size_start:
print('INFO: agent updating...')
batch = buffer.sample(batch_size)
current_loss = agent.update(batch, i, discount_factor)
if i % update_target_freq == 0:
agent.update_target()
if (i + 1 + render_freq) % render_freq == 0:
render = True
else:
render = False
print('Episode #{} reward:'.format(i), episode_reward)
print('Current loss:'.format(i), current_loss)
print()
# Model directory
model_dir = args.result_dir + '/models/' + args.model + '/' + args.dataset + '/' + args.loss + '/BS_' + str(
args.batch_size) + '/CI_' + str(args.critic_iter) + '/ND_' + str(
args.noise_dim) + '/L_' + str(args.LAMBDA)
# Load and prepare data
if args.dataset.lower() == 'stress_strain':
dataset = stress_strain.StressStrainDS()
preproc = preprocessing.standardize
elif args.dataset.lower() == 'mnist':
dataset = mnist.MNISTDS()
preproc = preprocessing.standardize_MNIST
train_dataset = dataset.load_dataset()
train_dataset, scaler = preprocessing.preprocess(train_dataset,
args.batch_size, preproc)
INPUT_SHAPE = tuple(
tf.compat.v1.data.get_output_shapes(train_dataset).as_list()[1:])
# Instantiate Generator and Discriminator
generator, discriminator = gans.get_models(args.model, args.loss,
INPUT_SHAPE, args.noise_dim)
print('\n\n######### GENERATOR #########\n')
generator.summary()
print('\n\n####### DISCRIMINATOR #######\n')
discriminator.summary()
# Optimizers
if args.optimizer.lower() == 'adam':
def train():
ctx = [mx.cpu(0), mx.cpu(1), mx.cpu(2), mx.cpu(3)]
discount_factor = 0.9
num_episodes = 10000
exploration_rate_begin = 0.9
exploration_rate_end = 0.05
exploration_rate = exploration_rate_begin
exploration_decay = 200
render = False
steps_done = 0
batch_size = 64
update_freq = 5
lr = 0.01
# init objects
buffer = Experience_buffer()
env = gym.make('FlappyBird-v0')
model = Model(env.action_space.n)
model.initialize(init=mx.initializer.Xavier(), ctx=ctx)
optimizer = mx.optimizer.Adam(learning_rate=lr)
trainer = gluon.Trainer(params=model.collect_params(), optimizer=optimizer)
loss = gluon.loss.HuberLoss()
agent = DQNAgent(env, model, trainer, loss)
# let's play !
for i in range(num_episodes):
# begin a new episode
print('Episode #{}'.format(i))
done = False
episode_reward = 0
current_loss = 0
current_obs = env.reset()
current_obs = nd.array(preprocess(current_obs))
if render:
env.render()
while not done:
action = agent.select_action(current_obs, exploration_rate)
next_obs, reward, done, _ = env.step(action)
next_obs = nd.array(preprocess(next_obs))
if render:
env.render()
buffer.add_experience(current_obs, action, reward, next_obs, done)
if buffer.is_full():
print('INFO: buffer is full')
# update information
current_obs = next_obs
episode_reward += reward
steps_done += 1
exploration_rate = exploration_rate_end + (
exploration_rate_begin - exploration_rate_end) * math.exp(
-1. * steps_done / exploration_decay)
# if the buffer is filled enough, periodically update the model
if len(buffer) > batch_size and update_freq % i == 0:
print('INFO: agent updating...')
batch = buffer.sample(batch_size)
current_loss = agent.update(batch, batch_size, i,
discount_factor)
render = True
else:
render = False
print('Episode #{} reward:'.format(i), episode_reward)
print('Current loss:'.format(i), current_loss)
def __read_data__(self):
scaler = StandardScaler()
print('loading dataset...')
# print(self.root_path)
# print(self.data_name)
data_path = Path(self.root_path)/self.data_name
pickle_path = Path(self.root_path)/f"{self.data_name}.pandas.pickle"
# print(data_path)
if not pickle_path.exists():
with pickle_path.open('wb') as p_fd:
df_raw = pd.read_csv(data_path)
features = [c for c in df_raw.columns if 'feature' in c]
print('preprocessing data...')
df_raw = preprocess(df_raw, self.scale)
pickle.dump(df_raw, p_fd)
with pickle_path.open('rb') as p_fd:
df_pickled = pickle.load(p_fd)
# df_pickled.info()
df_pickled = df_pickled[df_pickled.weight != 0]
# df_pickled = df_pickled[df_pickled.date > 399]
df_pickled = df_pickled[df_pickled.date > 85].reset_index(drop=True)
print('generate target...')
resp_cols = [c for c in df_pickled.columns if 'resp' in c]
# df_pickled['action'] = ((df_pickled['resp'] > 0) &
# (df_pickled['resp_1'] > 0) &
# (df_pickled['resp_2'] > 0) &
# (df_pickled['resp_3'] > 0) &
# (df_pickled['resp_4'] > 0)).astype('int')
# df_pickled['action'] = df_pickled['resp'].copy()
df_pickled['action'] = df_pickled[resp_cols].sum(axis=1)/len(resp_cols)
# df_pickled['action'] = df_pickled.apply(lambda row: row.weight * row.resp, axis='columns')
# df_pickled['action_1'] = df_pickled['resp_1']
# df_pickled['action_2'] = df_pickled['resp_2']
# df_pickled['action_3'] = df_pickled['resp_3']
# df_pickled['action_4'] = df_pickled['resp_4']
# df_pickled.info()
print("split train, valid...")
split_date = 400
train_df = df_pickled.loc[df_pickled.date <= split_date].reset_index(drop=True)
valid_df = df_pickled.loc[df_pickled.date > split_date].reset_index(drop=True)
# print(train_df)
# valid_df['weighted_resp'] = valid_df.apply(lambda row: row.weight * row.resp, axis='columns')
# target_cols = ['action', 'action_1', 'action_2', 'action_3', 'action_4']
# target_cols = ['action', 'action_1', 'action_2', 'action_3']
# target_cols = ['weighted_resp']
target_cols = ['action']
if self.scale:
train_df[target_cols] = scaler.fit_transform(train_df[target_cols].values)
valid_df[target_cols] = scaler.fit_transform(valid_df[target_cols].values)
print('organize values...')
features = [c for c in train_df.columns if 'feature' in c]
if self.set_type == 0:
self.weight = train_df.weight.values
self.resp = train_df.resp.values
self.data_x = train_df[features+target_cols].values
self.data_y = train_df[features+target_cols].values
self.data_stamp = train_df.date.values
elif self.set_type == 1:
self.weight = valid_df.weight.values
self.resp = valid_df.resp.values
self.data_x = valid_df[features+target_cols].values
self.data_y = valid_df[features+target_cols].values
self.data_stamp = valid_df.date.values
import numpy as np
from sklearn.model_selection import train_test_split
import utils.config as config
from utils.metrics import show_metrics, show_conf_matrix, dtree_viz
from utils.models import DecisionTree, RandomForest, KNeighbours, ArtificialNeuralNetwork, XGBoost, \
Results, VotingClassifier
from utils.preprocessing import preprocess
# Eliminate randomness
np.random.seed(1337)
random.seed(1337)
# Take sequence file as a BoW of bytes and make a probability distribution from that
X1, X2, y = preprocess()
# Cast lists to numpy array
X1 = np.array(X1)
X2 = np.array(X2)
y = np.array(y)
# Split train and test data
X1_train, X1_test, X2_train, X2_test, y_train, y_test = train_test_split(
X1, X2, y, test_size=config.TEST_SIZE)
# Make models and train
models_probability = {
'dt': DecisionTree,
'rf': RandomForest,
'knn': KNeighbours,