def main(sim_path, out):
# Define wildcard path to contact matrix data
cm_filepath = os.path.join(sim_path, 'output-cm-*.h5')
# Collect contact matrix file names sorted by sim_id
cm_files = sorted(glob(cm_filepath))
if not cm_files:
raise FileNotFoundError(f'No h5 files found, recheck your input path {sim_path}')
with ExitStack() as stack:
# Open all h5 files and add them to exit stack
open_cm_files = map(lambda file: stack.enter_context(open_h5(file)),
cm_files)
# Iterate through open h5 files and get contact_map datasets
cm_data = list(map(lambda file: file['contact_maps'], open_cm_files))
# Compress all .h5 files into one in cvae format
cvae_input = cm_to_cvae(cm_data)
# Create .h5 as cvae input
cvae_input_file = os.path.join(out, 'cvae-input.h5')
# Create and open contact map aggregation output file
cvae_input_file = stack.enter_context(h5py.File(cvae_input_file, 'w'))
# Write aggregated contact map dataset to file
cvae_input_file.create_dataset('contact_maps', data=cvae_input)
def main(input_path, out_path):
f = open_h5(input_path)
contact_maps = np.array(f['contact_maps'][:])
f.close()
sparse_contact_maps_from_matrices(contact_maps, out_path)
def __init__(self, path, split_ptc=0.8, split='train', squeeze=False):
with open_h5(path) as input_file:
# Access contact matrix data from h5 file
data = np.array(input_file['contact_maps'])
# 80-20 train validation split index
split_ind = int(split_ptc * len(data))
if split == 'train':
self.data = data[:split_ind]
elif split == 'valid':
self.data = data[split_ind:]
else:
raise ValueError(f'Parameter split={split} is invalid.')
# TODO: in future contact map Dataset, pass in device to precompute
# the operation
# TODO: this reshape code may not be the best solution. revisit
num_residues = self.data.shape[2]
assert num_residues == 22
if squeeze:
shape = (-1, num_residues, num_residues)
else:
shape = (-1, 1, num_residues, num_residues)
self.data = torch.from_numpy(self.data.reshape(shape)).to(torch.float32)
def __init__(self,
path,
out_dir,
squeeze,
sample_interval=20,
batch_size=128,
writer=None):
"""
Parameters
----------
path : str
Path to h5 file containing contact matrices.
out_dir : str
Directory to store output plots.
squeeze : bool
If True, data is reshaped to (H, W) else if False data
is reshaped to (1, H, W).
sample_interval : int
Plots every sample_interval'th point in the data set
batch_size : int
Batch size to load raw contact matrices into memory.
Batches are loaded into memory, encoded to a latent
dimension and then collected in a np.ndarray. The
np.ndarray is then passed to the TSNE algorithm.
NOTE: Not a learning hyperparameter, simply needs to
be small enough to load batch into memory.
writer : torch.utils.tensorboard.SummaryWriter
"""
os.makedirs(out_dir, exist_ok=True)
# Open h5 file. Python's garbage collector closes the
# file when class is destructed.
h5_file = open_h5(path)
self.dset = h5_file['contact_maps']
self.out_dir = out_dir
self.sample_interval = sample_interval
self.batch_size = batch_size
self.writer = writer
self.device = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
if squeeze:
self.shape = (self.dset.shape[1], self.dset.shape[2])
else:
self.shape = (1, self.dset.shape[1], self.dset.shape[2])
self._init_plot(h5_file)
def generate_embeddings(encoder_hparams_path, encoder_weight_path, cm_path):
encoder_hparams = EncoderHyperparams.load(encoder_hparams_path)
with open_h5(cm_path) as file:
# Access contact matrix data from h5 file
data = file['contact_maps']
# Get shape of an individual contact matrix
# (ignore total number of matrices)
input_shape = data.shape[1:]
encoder = EncoderConvolution2D(input_shape=input_shape,
hyperparameters=encoder_hparams)
# Load best model weights
encoder.load_weights(encoder_weight_path)
# Create contact matrix embeddings
cm_embeddings, *_ = encoder.embed(data)
return cm_embeddings
def main(input_path, out_path, model_id, gpu, epochs, batch_size, latent_dim):
# Set CUDA environment variables
os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu)
with open_h5(input_path) as input_file:
# Access contact matrix data from h5 file
data = np.array(input_file['contact_maps'])
# Shuffle data before train validation split
np.random.shuffle(data)
# 80-20 train validation split index
split = int(0.8 * len(data))
# Partition input data into 80-20 train valid split
train, valid = data[:split], data[split:]
# Get shape of an individual contact matrix
# (ignore total number of matrices)
input_shape = train.shape[1:]
# Set model hyperparameters for encoder and decoder
shared_hparams = {'num_conv_layers': 4,
'filters': [64, 64, 64, 64],
'kernels': [3, 3, 3, 3],
'strides': [1, 2, 1, 1],
'num_affine_layers': 1,
'affine_widths': [128],
'latent_dim': latent_dim
}
affine_dropouts = [0]
encoder_hparams = EncoderHyperparams(affine_dropouts=affine_dropouts,
**shared_hparams)
decoder_hparams = DecoderHyperparams(**shared_hparams)
encoder = EncoderConvolution2D(input_shape=input_shape,
hyperparameters=encoder_hparams)
# Get shape attributes of the last encoder layer to define the decoder
encode_conv_shape, num_conv_params = encoder.get_final_conv_params()
decoder = DecoderConvolution2D(output_shape=input_shape,
enc_conv_params=num_conv_params,
enc_conv_shape=encode_conv_shape,
hyperparameters=decoder_hparams)
optimizer = RMSprop(lr=0.001, rho=0.9, epsilon=1e-08, decay=0.0)
cvae = VAE(input_shape=input_shape,
encoder=encoder,
decoder=decoder,
optimizer=optimizer)
# Define callbacks to report model performance for analysis
embed_callback = EmbeddingCallback(train, cvae)
loss_callback = LossHistory()
cvae.train(data=train, validation_data=valid,
batch_size=batch_size, epochs=epochs,
callbacks=[embed_callback, loss_callback])
# Define file paths to store model performance and weights
ae_weight_path = os.path.join(out_path, f'ae-weight-{model_id}.h5')
encoder_weight_path = os.path.join(out_path, f'encoder-weight-{model_id}.h5')
encoder_hparams_path = os.path.join(out_path, f'encoder-hparams-{model_id}.pkl')
decoder_hparams_path = os.path.join(out_path, f'decoder-hparams-{model_id}.pkl')
embed_path = os.path.join(out_path, f'embed-{model_id}.npy')
idx_path = os.path.join(out_path, f'embed-idx-{model_id}.npy')
loss_path = os.path.join(out_path, f'loss-{model_id}.npy')
val_loss_path = os.path.join(out_path, f'val-loss-{model_id}.npy')
# Save weights, hyperparameters, and model performance.
# Save encoder weights seperately so the full model doesn't need to be
# loaded during the outlier detection stage.
cvae.save_weights(ae_weight_path)
encoder.save_weights(encoder_weight_path)
encoder_hparams.save(encoder_hparams_path)
decoder_hparams.save(decoder_hparams_path)
embed_callback.save(embed_path=embed_path, idx_path=idx_path)
loss_callback.save(loss_path=loss_path, val_loss_path=val_loss_path)
def __init__(self,
path,
input_shape,
split_ptc=0.8,
split='train',
sparse=False,
gpu=None):
"""
Parameters
----------
path : str
Path to h5 file containing contact matrices.
input_shape : tuple
Shape of contact matrices (H, W), may be (1, H, W).
split_ptc : float
Percentage of total data to be used as training set.
split : str
Either 'train' or 'valid', specifies whether this
dataset returns train or validation data.
sparse : bool
If True, process data as sparse row/col COO format. Data
should not contain any values because they are all 1's and
generated on the fly. If False, input data is normal tensor.
gpu : int, None
If None, then data will be put onto the default GPU if CUDA
is available and otherwise is put onto a CPU. If gpu is int
type, then data is put onto the specified GPU.
"""
if split not in ('train', 'valid'):
raise ValueError("Parameter split must be 'train' or 'valid'.")
if split_ptc < 0 or split_ptc > 1:
raise ValueError(
'Parameter split_ptc must satisfy 0 <= split_ptc <= 1.')
# Open h5 file. Python's garbage collector closes the
# file when class is destructed.
h5_file = open_h5(path)
if sparse:
group = h5_file['contact_maps']
self.row_dset = group.get('row')
self.col_dset = group.get('col')
self.len = len(self.row_dset)
else:
# contact_maps dset has shape (N, W, H, 1)
self.dset = h5_file['contact_maps']
self.len = len(self.dset)
# train validation split index
self.split_ind = int(split_ptc * self.len)
self.split = split
self.sparse = sparse
self.shape = input_shape
if gpu is None:
self.device = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
else:
self.device = torch.device(f'cuda:{gpu}')
