def worker_init(device, modelname, chunk_size, overlap,
read_params, alphabet, max_concurrent_chunks,
fastq, qscore_scale, qscore_offset, beam, posterior,
temperature):
global all_read_params
global process_read_partial
all_read_params = read_params
device = helpers.set_torch_device(device)
model = load_model(modelname).to(device)
stride = guess_model_stride(model)
chunk_size = chunk_size * stride
overlap = overlap * stride
n_can_base = len(alphabet)
n_can_state = nstate_flipflop(n_can_base)
def process_read_partial(read_filename, read_id, read_params):
res = process_read(read_filename, read_id,
model, chunk_size, overlap, read_params,
n_can_state, stride, alphabet,
max_concurrent_chunks, fastq, qscore_scale,
qscore_offset, beam, posterior, temperature)
return (read_id, *res)
def main():
args = parser.parse_args()
device = helpers.set_torch_device(args.device)
# TODO convert to logging
sys.stderr.write("* Loading model.\n")
model = load_model(args.model).to(device)
is_cat_mod = isinstance(model.sublayers[-1],
layers.GlobalNormFlipFlopCatMod)
do_output_mods = args.modified_base_output is not None
if do_output_mods and not is_cat_mod:
sys.stderr.write(
"Cannot output modified bases from canonical base only model.")
sys.exit()
n_can_states = nstate_flipflop(model.sublayers[-1].nbase)
stride = guess_model_stride(model)
chunk_size = args.chunk_size * stride
chunk_overlap = args.overlap * stride
sys.stderr.write("* Initializing reads file search.\n")
fast5_reads = list(
fast5utils.iterate_fast5_reads(args.input_folder,
limit=args.limit,
strand_list=args.input_strand_list,
recursive=args.recursive))
sys.stderr.write("* Found {} reads.\n".format(len(fast5_reads)))
if args.scaling is not None:
sys.stderr.write("* Loading read scaling parameters from {}.\n".format(
args.scaling))
all_read_params = get_per_read_params_dict_from_tsv(args.scaling)
input_read_ids = frozenset(rec[1] for rec in fast5_reads)
scaling_read_ids = frozenset(all_read_params.keys())
sys.stderr.write("* {} / {} reads have scaling information.\n".format(
len(input_read_ids & scaling_read_ids), len(input_read_ids)))
fast5_reads = [
rec for rec in fast5_reads if rec[1] in scaling_read_ids
]
else:
all_read_params = {}
mods_fp = None
if do_output_mods:
mods_fp = h5py.File(args.modified_base_output)
mods_fp.create_group('Reads')
mod_long_names = model.sublayers[-1].ordered_mod_long_names
sys.stderr.write("* Preparing modified base output: {}.\n".format(
', '.join(map(str, mod_long_names))))
mods_fp.create_dataset('mod_long_names',
data=np.array(mod_long_names, dtype='S'),
dtype=h5py.special_dtype(vlen=str))
sys.stderr.write("* Calling reads.\n")
nbase, ncalled, nread, nsample = 0, 0, 0, 0
t0 = time.time()
progress = Progress(quiet=args.quiet)
startcharacter = '@' if args.fastq else '>'
try:
with open_file_or_stdout(args.output) as fh:
for read_filename, read_id in fast5_reads:
read_params = all_read_params[
read_id] if read_id in all_read_params else None
basecall, qstring, read_nsample = process_read(
read_filename, read_id, model, chunk_size, chunk_overlap,
read_params, n_can_states, stride, args.alphabet,
is_cat_mod, mods_fp, args.max_concurrent_chunks,
args.fastq, args.qscore_scale, args.qscore_offset)
if basecall is not None:
fh.write("{}{}\n{}\n".format(
startcharacter, read_id,
basecall[::-1] if args.reverse else basecall))
nbase += len(basecall)
ncalled += 1
if args.fastq:
fh.write("+\n{}\n".format(
qstring[::-1] if args.reverse else qstring))
nread += 1
nsample += read_nsample
progress.step()
finally:
if mods_fp is not None:
mods_fp.close()
total_time = time.time() - t0
sys.stderr.write("* Called {} reads in {:.2f}s\n".format(
nread, int(total_time)))
sys.stderr.write("* {:7.2f} kbase / s\n".format(nbase / total_time /
1000.0))
sys.stderr.write("* {:7.2f} ksample / s\n".format(nsample / total_time /
1000.0))
sys.stderr.write("* {} reads failed.\n".format(nread - ncalled))
return
basename = 'model_final'
else:
basename = 'model_checkpoint_{:05d}'.format(index)
model_file = os.path.join(outdir, basename + '.checkpoint')
torch.save(network, model_file)
params_file = os.path.join(outdir, basename + '.params')
torch.save(network.state_dict(), params_file)
if __name__ == '__main__':
args = parser.parse_args()
np.random.seed(args.seed)
device = helpers.set_torch_device(args.device)
helpers.prepare_outdir(args.outdir, args.overwrite)
copyfile(args.model, os.path.join(args.outdir, 'model.py'))
log = helpers.Logger(os.path.join(args.outdir, 'model.log'), args.quiet)
log.write(helpers.formatted_env_info(device))
log.write('* Loading data from {}\n'.format(args.chunks))
log.write('* Per read file MD5 {}\n'.format(helpers.file_md5(args.chunks)))
if args.limit is not None:
log.write('* Limiting number of strands to {}\n'.format(args.limit))
with h5py.File(args.chunks, 'r') as h5:
def main():
args = parser.parse_args()
is_multi_gpu = (args.local_rank is not None)
is_lead_process = (not is_multi_gpu) or args.local_rank == 0
if is_multi_gpu:
#Use distributed parallel processing to run one process per GPU
try:
torch.distributed.init_process_group(backend='nccl')
except:
raise Exception(
"Unable to start multiprocessing group. " +
"The most likely reason is that the script is running with " +
"local_rank set but without the set-up for distributed " +
"operation. local_rank should be used " +
"only by torch.distributed.launch. See the README.")
device = helpers.set_torch_device(args.local_rank)
if args.seed is not None:
#Make sure processes get different random picks of training data
np.random.seed(args.seed + args.local_rank)
else:
device = helpers.set_torch_device(args.device)
np.random.seed(args.seed)
if is_lead_process:
helpers.prepare_outdir(args.outdir, args.overwrite)
if args.model.endswith('.py'):
copyfile(args.model, os.path.join(args.outdir, 'model.py'))
batchlog = helpers.BatchLog(args.outdir)
logfile = os.path.join(args.outdir, 'model.log')
else:
logfile = None
log = helpers.Logger(logfile, args.quiet)
log.write(helpers.formatted_env_info(device))
log.write('* Loading data from {}\n'.format(args.input))
log.write('* Per read file MD5 {}\n'.format(helpers.file_md5(args.input)))
if args.input_strand_list is not None:
read_ids = list(set(helpers.get_read_ids(args.input_strand_list)))
log.write(('* Will train from a subset of {} strands, determined ' +
'by read_ids in input strand list\n').format(len(read_ids)))
else:
log.write('* Reads not filtered by id\n')
read_ids = 'all'
if args.limit is not None:
log.write('* Limiting number of strands to {}\n'.format(args.limit))
with mapped_signal_files.HDF5Reader(args.input) as per_read_file:
alphabet_info = per_read_file.get_alphabet_information()
read_data = per_read_file.get_multiple_reads(read_ids,
max_reads=args.limit)
# read_data now contains a list of reads
# (each an instance of the Read class defined in
# mapped_signal_files.py, based on dict)
log.write('* Using alphabet definition: {}\n'.format(str(alphabet_info)))
if len(read_data) == 0:
log.write('* No reads remaining for training, exiting.\n')
exit(1)
log.write('* Loaded {} reads.\n'.format(len(read_data)))
# Get parameters for filtering by sampling a subset of the reads
# Result is a tuple median mean_dwell, mad mean_dwell
# Choose a chunk length in the middle of the range for this
sampling_chunk_len = (args.chunk_len_min + args.chunk_len_max) // 2
filter_params = chunk_selection.sample_filter_parameters(
read_data, args.sample_nreads_before_filtering, sampling_chunk_len,
args.filter_mean_dwell, args.filter_max_dwell)
log.write("* Sampled {} chunks".format(
args.sample_nreads_before_filtering))
log.write(": median(mean_dwell)={:.2f}".format(
filter_params.median_meandwell))
log.write(", mad(mean_dwell)={:.2f}\n".format(filter_params.mad_meandwell))
log.write('* Reading network from {}\n'.format(args.model))
model_kwargs = {
'stride': args.stride,
'winlen': args.winlen,
# Number of input features to model e.g. was >1 for event-based
# models (level, std, dwell)
'insize': 1,
'size': args.size,
'alphabet_info': alphabet_info
}
if is_lead_process:
# Under pytorch's DistributedDataParallel scheme, we
# need a clone of the start network to use as a template for saving
# checkpoints. Necessary because DistributedParallel makes the class
# structure different.
network_save_skeleton = helpers.load_model(args.model, **model_kwargs)
log.write('* Network has {} parameters.\n'.format(
sum([p.nelement() for p in network_save_skeleton.parameters()])))
if not alphabet_info.is_compatible_model(network_save_skeleton):
sys.stderr.write(
'* ERROR: Model and mapped signal files contain incompatible '
+ 'alphabet definitions (including modified bases).')
sys.exit(1)
if is_cat_mod_model(network_save_skeleton):
log.write('* Loaded categorical modified base model.\n')
if not alphabet_info.contains_modified_bases():
sys.stderr.write(
'* ERROR: Modified bases model specified, but mapped ' +
'signal file does not contain modified bases.')
sys.exit(1)
else:
log.write('* Loaded standard (canonical bases-only) model.\n')
if alphabet_info.contains_modified_bases():
sys.stderr.write(
'* ERROR: Standard (canonical bases only) model ' +
'specified, but mapped signal file does contains ' +
'modified bases.')
sys.exit(1)
log.write('* Dumping initial model\n')
helpers.save_model(network_save_skeleton, args.outdir, 0)
if is_multi_gpu:
#so that processes 1,2,3.. don't try to load before process 0 has saved
torch.distributed.barrier()
log.write('* MultiGPU process {}'.format(args.local_rank))
log.write(': loading initial model saved by process 0\n')
saved_startmodel_path = os.path.join(
args.outdir, 'model_checkpoint_00000.checkpoint')
network = helpers.load_model(saved_startmodel_path).to(device)
# Wrap network for training in the DistributedDataParallel structure
network = torch.nn.parallel.DistributedDataParallel(
network,
device_ids=[args.local_rank],
output_device=args.local_rank)
else:
network = network_save_skeleton.to(device)
network_save_skeleton = None
optimizer = torch.optim.Adam(network.parameters(),
lr=args.lr_max,
betas=args.adam,
weight_decay=args.weight_decay,
eps=args.eps)
if args.lr_warmup is None:
lr_warmup = args.lr_min
else:
lr_warmup = args.lr_warmup
if args.lr_frac_decay is not None:
lr_scheduler = optim.ReciprocalLR(optimizer, args.lr_frac_decay,
args.warmup_batches, lr_warmup)
log.write('* Learning rate schedule lr_max*k/(k+t)')
log.write(', k={}, t=iterations.\n'.format(args.lr_frac_decay))
else:
lr_scheduler = optim.CosineFollowedByFlatLR(optimizer, args.lr_min,
args.lr_cosine_iters,
args.warmup_batches,
lr_warmup)
log.write('* Learning rate goes like cosine from lr_max to lr_min ')
log.write('over {} iterations.\n'.format(args.lr_cosine_iters))
log.write('* At start, train for {} '.format(args.warmup_batches))
log.write('batches at warm-up learning rate {:3.2}\n'.format(lr_warmup))
score_smoothed = helpers.WindowedExpSmoother()
# prepare modified base paramter tensors
network_is_catmod = is_cat_mod_model(network)
mod_factor_t = torch.tensor(args.mod_factor,
dtype=torch.float32).to(device)
can_mods_offsets = (network.sublayers[-1].can_mods_offsets
if network_is_catmod else None)
# mod cat inv freq weighting is currently disabled. Compute and set this
# value to enable mod cat weighting
mod_cat_weights = np.ones(alphabet_info.nbase, dtype=np.float32)
#Generating list of batches for standard loss reporting
reporting_chunk_len = (args.chunk_len_min + args.chunk_len_max) // 2
reporting_batch_list = list(
prepare_random_batches(device, read_data, reporting_chunk_len,
args.min_sub_batch_size,
args.reporting_sub_batches, alphabet_info,
filter_params, network, network_is_catmod, log))
log.write(
('* Standard loss report: chunk length = {} & sub-batch size ' +
'= {} for {} sub-batches. \n').format(reporting_chunk_len,
args.min_sub_batch_size,
args.reporting_sub_batches))
#Set cap at very large value (before we have any gradient stats).
gradient_cap = constants.LARGE_VAL
if args.gradient_cap_fraction is None:
log.write('* No gradient capping\n')
else:
rolling_quantile = maths.RollingQuantile(args.gradient_cap_fraction)
log.write('* Gradient L2 norm cap will be upper' +
' {:3.2f} quantile of the last {} norms.\n'.format(
args.gradient_cap_fraction, rolling_quantile.window))
total_bases = 0
total_samples = 0
total_chunks = 0
# To count the numbers of different sorts of chunk rejection
rejection_dict = defaultdict(int)
t0 = time.time()
log.write('* Training\n')
for i in range(args.niteration):
# Chunk length is chosen randomly in the range given but forced to
# be a multiple of the stride
batch_chunk_len = (
np.random.randint(args.chunk_len_min, args.chunk_len_max + 1) //
args.stride) * args.stride
# We choose the size of a sub-batch so that the size of the data in
# the sub-batch is about the same as args.min_sub_batch_size chunks of
# length args.chunk_len_max
sub_batch_size = int(args.min_sub_batch_size * args.chunk_len_max /
batch_chunk_len + 0.5)
optimizer.zero_grad()
main_batch_gen = prepare_random_batches(
device, read_data, batch_chunk_len, sub_batch_size,
args.sub_batches, alphabet_info, filter_params, network,
network_is_catmod, log)
chunk_count, fval, chunk_samples, chunk_bases, batch_rejections = \
calculate_loss( network, network_is_catmod,
main_batch_gen, args.sharpen,
can_mods_offsets, mod_cat_weights,
mod_factor_t, calc_grads = True )
gradnorm_uncapped = torch.nn.utils.clip_grad_norm_(
network.parameters(), gradient_cap)
if args.gradient_cap_fraction is not None:
gradient_cap = rolling_quantile.update(gradnorm_uncapped)
optimizer.step()
if is_lead_process:
batchlog.record(
fval, gradnorm_uncapped,
None if args.gradient_cap_fraction is None else gradient_cap)
total_chunks += chunk_count
total_samples += chunk_samples
total_bases += chunk_bases
# Update counts of reasons for rejection
for k, v in batch_rejections.items():
rejection_dict[k] += v
score_smoothed.update(fval)
if (i + 1) % args.save_every == 0 and is_lead_process:
helpers.save_model(network, args.outdir,
(i + 1) // args.save_every,
network_save_skeleton)
log.write('C')
else:
log.write('.')
if (i + 1) % DOTROWLENGTH == 0:
_, rloss, _, _, _ = calculate_loss(network, network_is_catmod,
reporting_batch_list,
args.sharpen, can_mods_offsets,
mod_cat_weights, mod_factor_t)
# In case of super batching, additional functionality must be
# added here
learning_rate = lr_scheduler.get_lr()[0]
tn = time.time()
dt = tn - t0
t = (' {:5d} {:7.5f} {:7.5f} {:5.2f}s ({:.2f} ksample/s {:.2f} ' +
'kbase/s) lr={:.2e}')
log.write(
t.format((i + 1) // DOTROWLENGTH, score_smoothed.value, rloss,
dt, total_samples / 1000.0 / dt,
total_bases / 1000.0 / dt, learning_rate))
# Write summary of chunk rejection reasons
if args.full_filter_status:
for k, v in rejection_dict.items():
log.write(" {}:{} ".format(k, v))
else:
n_tot = n_fail = 0
for k, v in rejection_dict.items():
n_tot += v
if k != 'pass':
n_fail += v
log.write(" {:.1%} chunks filtered".format(n_fail / n_tot))
log.write("\n")
total_bases = 0
total_samples = 0
t0 = tn
# Uncomment the lines below to check synchronisation of models
# between processes in multi-GPU operation
#for p in network.parameters():
# v = p.data.reshape(-1)[:5].to('cpu')
# u = p.data.reshape(-1)[-5:].to('cpu')
# break
#if args.local_rank is not None:
# log.write("* GPU{} params:".format(args.local_rank))
#log.write("{}...{}\n".format(v,u))
lr_scheduler.step()
if is_lead_process:
helpers.save_model(network,
args.outdir,
model_skeleton=network_save_skeleton)
def parse_init_args(args):
is_multi_gpu = (args.local_rank is not None)
is_lead_process = (not is_multi_gpu) or args.local_rank == 0
# if seed is provided use this else generate random seed value
seed = (np.random.randint(0, np.iinfo(np.uint32).max, dtype=np.uint32)
if args.seed is None else args.seed)
main_log_fn = os.path.join(args.outdir, MODEL_LOG_FILENAME)
if is_lead_process:
helpers.prepare_outdir(args.outdir, args.overwrite)
if args.model.endswith('.py'):
copyfile(args.model, os.path.join(args.outdir, 'model.py'))
# note buffering=1 to enforce line buffering and enable
# inspection/plotting during a run
logs = LOGS(main=helpers.Logger(main_log_fn, args.quiet),
batch=open(os.path.join(args.outdir, BATCH_LOG_FILENAME),
'w',
buffering=1),
validation=open(os.path.join(args.outdir,
VAL_LOG_FILENAME),
'w',
buffering=1))
logs.batch.write(BATCH_HEADER)
logs.validation.write(VAL_HEADER)
if args.save_every % DOTROWLENGTH != 0:
# Illegal save_every, change
se2 = int(math.ceil(args.save_every / DOTROWLENGTH)) * DOTROWLENGTH
logs.main.write('* --save_every {} not a multiple of {}, rounding '
'to {}'.format(args.save_every, DOTROWLENGTH, se2))
args.save_every = se2
if args.chunk_len_min > args.chunk_len_max:
# Illegal chunk length parameters
raise ValueError('--chunk_len_min greater than --chunk_len_max')
logs.main.write('* Using random seed: {}\n'.format(seed))
else:
logs = LOGS(main=helpers.Logger(main_log_fn, args.quiet))
if is_multi_gpu:
# Use distributed parallel processing to run one process per GPU
try:
torch.distributed.init_process_group(backend='nccl')
except Exception:
raise Exception(
'Unable to start multiprocessing group. The most likely ' +
'reason is that the script is running with local_rank set ' +
'but without the set-up for distributed operation. ' +
'local_rank should be used only by torch.distributed.' +
'launch. See the README.')
device = helpers.set_torch_device(args.local_rank)
# offset seeds so different GPUs get different data streams
seed += args.local_rank
else:
device = helpers.set_torch_device(args.device)
logs.main.write(helpers.formatted_env_info(device))
# set random seed for this process
np.random.seed(seed)
torch.manual_seed(seed)
if _MAKE_TORCH_DETERMINISTIC and device.type == 'cuda':
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
return RESOURCE_INFO(is_multi_gpu, is_lead_process, device), logs
def main():
args = parser.parse_args()
np.random.seed(args.seed)
helpers.prepare_outdir(args.outdir, args.overwrite)
device = helpers.set_torch_device(args.device)
log = helpers.Logger(os.path.join(args.outdir, 'model.log'), args.quiet)
log.write(helpers.formatted_env_info(device))
if args.input_strand_list is not None:
read_ids = list(set(helpers.get_read_ids(args.input_strand_list)))
log.write('* Will train from a subset of {} strands\n'.format(
len(read_ids)))
else:
log.write('* Reads not filtered by id\n')
read_ids = 'all'
if args.limit is not None:
log.write('* Limiting number of strands to {}\n'.format(args.limit))
with mapped_signal_files.HDF5Reader(args.input) as per_read_file:
alphabet_info = per_read_file.get_alphabet_information()
assert alphabet_info.nbase == 4, (
'Squiggle prediction with modified base training data is ' +
'not currenly supported.')
read_data = per_read_file.get_multiple_reads(read_ids,
max_reads=args.limit)
# read_data now contains a list of reads
# (each an instance of the Read class defined in mapped_signal_files.py, based on dict)
if len(read_data) == 0:
log.write('* No reads remaining for training, exiting.\n')
exit(1)
log.write('* Loaded {} reads.\n'.format(len(read_data)))
# Get parameters for filtering by sampling a subset of the reads
# Result is a tuple median mean_dwell, mad mean_dwell
filter_parameters = chunk_selection.sample_filter_parameters(
read_data, args.sample_nreads_before_filtering, args.target_len,
args.filter_mean_dwell, args.filter_max_dwell)
log.write(
"* Sampled {} chunks: median(mean_dwell)={:.2f}, mad(mean_dwell)={:.2f}\n"
.format(args.sample_nreads_before_filtering,
filter_parameters.median_meandwell,
filter_parameters.mad_meandwell))
conv_net = create_convolution(args.size, args.depth, args.winlen)
nparam = sum([p.data.detach().numpy().size for p in conv_net.parameters()])
log.write('* Created network. {} parameters\n'.format(nparam))
log.write('* Depth {} layers ({} residual layers)\n'.format(
args.depth + 2, args.depth))
log.write('* Window width {}\n'.format(args.winlen))
log.write('* Context +/- {} bases\n'.format(
(args.depth + 2) * (args.winlen // 2)))
conv_net = conv_net.to(device)
optimizer = torch.optim.Adam(conv_net.parameters(),
lr=args.lr_max,
betas=args.adam,
weight_decay=args.weight_decay,
eps=args.eps)
lr_scheduler = optim.ReciprocalLR(optimizer, args.lr_decay)
rejection_dict = defaultdict(
lambda: 0
) # To count the numbers of different sorts of chunk rejection
t0 = time.time()
score_smoothed = helpers.WindowedExpSmoother()
total_chunks = 0
for i in range(args.niteration):
# If the logging threshold is 0 then we log all chunks, including those rejected, so pass the log
# object into assemble_batch
# chunk_batch is a list of dicts.
chunk_batch, batch_rejections = chunk_selection.assemble_batch(
read_data,
args.batch_size,
args.target_len,
filter_parameters,
chunk_len_means_sequence_len=True)
if len(chunk_batch) < args.batch_size:
log.write('* Warning: only {} chunks passed filters.\n'.format(
len(chunk_batch)))
total_chunks += len(chunk_batch)
# Update counts of reasons for rejection
for k, v in batch_rejections.items():
rejection_dict[k] += v
# Shape of input needs to be seqlen x batchsize x embedding_dimension
embedded_matrix = [
embed_sequence(d['sequence'], alphabet=None) for d in chunk_batch
]
seq_embed = torch.tensor(embedded_matrix).permute(1, 0, 2).to(device)
# Shape of labels is a flat vector
batch_signal = torch.tensor(
np.concatenate([d['current'] for d in chunk_batch])).to(device)
# Shape of lens is also a flat vector
batch_siglen = torch.tensor([len(d['current'])
for d in chunk_batch]).to(device)
#print("First 10 elements of first sequence in batch",seq_embed[:10,0,:])
#print("First 10 elements of signal batch",batch_signal[:10])
#print("First 10 lengths",batch_siglen[:10])
optimizer.zero_grad()
predicted_squiggle = conv_net(seq_embed)
batch_loss = squiggle_match_loss(predicted_squiggle, batch_signal,
batch_siglen, args.back_prob)
fval = batch_loss.sum() / float(batch_siglen.sum())
fval.backward()
optimizer.step()
score_smoothed.update(float(fval))
if (i + 1) % args.save_every == 0:
helpers.save_model(conv_net, args.outdir,
(i + 1) // args.save_every)
log.write('C')
else:
log.write('.')
if (i + 1) % DOTROWLENGTH == 0:
tn = time.time()
dt = tn - t0
t = ' {:5d} {:7.5f} {:5.2f}s'
log.write(
t.format((i + 1) // DOTROWLENGTH, score_smoothed.value, dt))
t0 = tn
# Write summary of chunk rejection reasons
if args.full_filter_status:
for k, v in rejection_dict.items():
log.write(" {}:{} ".format(k, v))
else:
n_tot = n_fail = 0
for k, v in rejection_dict.items():
n_tot += v
if k != 'pass':
n_fail += v
log.write(" {:.1%} chunks filtered".format(n_fail / n_tot))
log.write("\n")
lr_scheduler.step()
helpers.save_model(conv_net, args.outdir)