def main():
args = parser.parse_args()
#model_md5 = file_md5(args.model)
model = load_model(args.model)
print("Previous alphabet",model.sublayers[-1].output_alphabet)
for attr in ["mod_bases","mod_labels","mod_name_conv","ordered_mod_long_names"]:
if "mod" in attr:
print(attr,getattr(model.sublayers[-1],attr))
#print("Previous alphabet",dir(model.sublayers[-1]))#).mod_long_names)
model.sublayers[-1].output_alphabet = args.alphabet
save_model(model,args.output)
def main():
args = parser.parse_args()
np.random.seed(args.seed)
device = torch.device(args.device)
if device.type == 'cuda':
try:
torch.cuda.set_device(device)
except AttributeError:
sys.stderr.write('ERROR: Torch not compiled with CUDA enabled ' +
'and GPU device set.')
sys.exit(1)
if not os.path.exists(args.output):
os.mkdir(args.output)
elif not args.overwrite:
sys.stderr.write('Error: Output directory {} exists but --overwrite ' +
'is false\n'.format(args.output))
exit(1)
if not os.path.isdir(args.output):
sys.stderr.write('Error: Output location {} is not directory\n'.format(
args.output))
exit(1)
copyfile(args.model, os.path.join(args.output, 'model.py'))
# Create a logging file to save details of chunks.
# If args.chunk_logging_threshold is set to 0 then we log all chunks
# including those rejected.
chunk_log = chunk_selection.ChunkLog(args.output)
log = helpers.Logger(os.path.join(args.output, 'model.log'), args.quiet)
log.write('* Taiyaki version {}\n'.format(__version__))
log.write('* Command line\n')
log.write(' '.join(sys.argv) + '\n')
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, _, _ = 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)
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_parameters = chunk_selection.sample_filter_parameters(
read_data,
args.sample_nreads_before_filtering,
sampling_chunk_len,
args,
log,
chunk_log=chunk_log)
medmd, madmd = filter_parameters
log.write(
"* Sampled {} chunks: median(mean_dwell)={:.2f}, mad(mean_dwell)={:.2f}\n"
.format(args.sample_nreads_before_filtering, medmd, madmd))
log.write('* Reading network from {}\n'.format(args.model))
nbase = len(alphabet)
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,
'outsize': flipflopfings.nstate_flipflop(nbase)
}
network = helpers.load_model(args.model, **model_kwargs).to(device)
log.write('* Network has {} parameters.\n'.format(
sum([p.nelement() for p in network.parameters()])))
optimizer = torch.optim.Adam(network.parameters(),
lr=args.lr_max,
betas=args.adam,
weight_decay=args.weight_decay)
lr_scheduler = optim.CosineFollowedByFlatLR(optimizer, args.lr_min,
args.lr_cosine_iters)
score_smoothed = helpers.WindowedExpSmoother()
log.write('* Dumping initial model\n')
helpers.save_model(network, args.output, 0)
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):
lr_scheduler.step()
# 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 batch size so that the size of the data in the batch
# is about the same as args.min_batch_size chunks of length
# args.chunk_len_max
target_batch_size = int(args.min_batch_size * args.chunk_len_max /
batch_chunk_len + 0.5)
# ...but it can't be more than the number of reads.
batch_size = min(target_batch_size, len(read_data))
# If the logging threshold is 0 then we log all chunks, including those
# rejected, so pass the log
# object into assemble_batch
if args.chunk_logging_threshold == 0:
log_rejected_chunks = chunk_log
else:
log_rejected_chunks = None
# Chunk_batch is a list of dicts.
chunk_batch, batch_rejections = chunk_selection.assemble_batch(
read_data,
batch_size,
batch_chunk_len,
filter_parameters,
args,
log,
chunk_log=log_rejected_chunks)
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 tensor must be:
# (timesteps) x (batch size) x (input channels)
# in this case:
# batch_chunk_len x batch_size x 1
stacked_current = np.vstack([d['current'] for d in chunk_batch]).T
indata = torch.tensor(stacked_current,
device=device,
dtype=torch.float32).unsqueeze(2)
# Sequence input tensor is just a 1D vector, and so is seqlens
seqs = torch.tensor(np.concatenate([
flipflopfings.flipflop_code(d['sequence'], nbase)
for d in chunk_batch
]),
device=device,
dtype=torch.long)
seqlens = torch.tensor([len(d['sequence']) for d in chunk_batch],
dtype=torch.long,
device=device)
optimizer.zero_grad()
outputs = network(indata)
lossvector = ctc.crf_flipflop_loss(outputs, seqs, seqlens,
args.sharpen)
loss = lossvector.sum() / (seqlens > 0.0).float().sum()
loss.backward()
optimizer.step()
fval = float(loss)
score_smoothed.update(fval)
# Check for poison chunk and save losses and chunk locations if we're
# poisoned If args.chunk_logging_threshold set to zero then we log
# everything
if fval / score_smoothed.value >= args.chunk_logging_threshold:
chunk_log.write_batch(i, chunk_batch, lossvector)
total_bases += int(seqlens.sum())
total_samples += int(indata.nelement())
# Doing this deletion leads to less CUDA memory usage.
del indata, seqs, seqlens, outputs, loss, lossvector
if device.type == 'cuda':
torch.cuda.empty_cache()
if (i + 1) % args.save_every == 0:
helpers.save_model(network, args.output,
(i + 1) // args.save_every)
log.write('C')
else:
log.write('.')
if (i + 1) % DOTROWLENGTH == 0:
# 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} {:5.3f} {:5.2f}s ({:.2f} ksample/s {:.2f} kbase/s) ' +
'lr={:.2e}')
log.write(
t.format((i + 1) // DOTROWLENGTH, score_smoothed.value, dt,
total_samples / 1000.0 / dt,
total_bases / 1000.0 / dt, learning_rate))
# Write summary of chunk rejection reasons
for k, v in rejection_dict.items():
log.write(" {}:{} ".format(k, v))
log.write("\n")
total_bases = 0
total_samples = 0
t0 = tn
helpers.save_model(network, args.output)
def main():
args = parser.parse_args()
np.random.seed(args.seed)
if not os.path.exists(args.output):
os.mkdir(args.output)
elif not args.overwrite:
sys.stderr.write(
'Error: Output directory {} exists but --overwrite is false\n'.
format(args.output))
exit(1)
if not os.path.isdir(args.output):
sys.stderr.write('Error: Output location {} is not directory\n'.format(
args.output))
exit(1)
log = helpers.Logger(os.path.join(args.output, 'model.log'), args.quiet)
log.write('# Taiyaki version {}\n'.format(__version__))
log.write('# Command line\n')
log.write(' '.join(sys.argv) + '\n')
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, _, _ = per_read_file.get_alphabet_information()
assert len(alphabet) == 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)))
# Create a logging file to save details of chunks.
# If args.chunk_logging_threshold is set to 0 then we log all chunks including those rejected.
chunk_log = chunk_selection.ChunkLog(args.output)
# 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,
log,
chunk_log=chunk_log)
medmd, madmd = filter_parameters
log.write(
"* Sampled {} chunks: median(mean_dwell)={:.2f}, mad(mean_dwell)={:.2f}\n"
.format(args.sample_nreads_before_filtering, medmd, madmd))
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)))
device = torch.device(args.device)
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)
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):
lr_scheduler.step()
# If the logging threshold is 0 then we log all chunks, including those rejected, so pass the log
# object into assemble_batch
if args.chunk_logging_threshold == 0:
log_rejected_chunks = chunk_log
else:
log_rejected_chunks = None
# 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,
args,
log,
chunk_log=log_rejected_chunks,
chunk_len_means_sequence_len=True)
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))
# Check for poison chunk and save losses and chunk locations if we're poisoned
# If args.chunk_logging_threshold set to zero then we log everything
if fval / score_smoothed.value >= args.chunk_logging_threshold:
chunk_log.write_batch(i, chunk_batch, batch_loss)
if (i + 1) % args.save_every == 0:
helpers.save_model(conv_net, args.output,
(i + 1) // args.save_every)
log.write('C')
else:
log.write('.')
if (i + 1) % DOTROWLENGTH == 0:
tn = time.time()
dt = tn - t0
t = ' {:5d} {:5.3f} {:5.2f}s'
log.write(
t.format((i + 1) // DOTROWLENGTH, score_smoothed.value, dt))
t0 = tn
# Write summary of chunk rejection reasons
for k, v in rejection_dict.items():
log.write(" {}:{} ".format(k, v))
log.write("\n")
helpers.save_model(conv_net, args.output)
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 train_model(train_params, net_info, optim_info, res_info, read_data,
alphabet_info, filter_params, mod_info, reporting_batch_list,
logs):
# Set cap at very large value (before we have any gradient stats).
grad_max_threshs = None
grad_max_thresh_str = 'NaN'
score_smoothed = helpers.WindowedExpSmoother()
total_bases = total_samples = total_chunks = 0
# To count the numbers of different sorts of chunk rejection
rejection_dict = defaultdict(int)
time_last = time.time()
logs.main.write('* Training\n')
for curr_iter in range(train_params.niteration):
sharpen = float(train_params.sharpen.min +
(train_params.sharpen.max - train_params.sharpen.min) *
min(1.0, curr_iter / train_params.sharpen.niter))
mod_factor = float(
mod_info.mod_factor.start +
(mod_info.mod_factor.final - mod_info.mod_factor.start) *
min(1.0, curr_iter / mod_info.mod_factor.niter))
# Chunk length is chosen randomly in the range given but forced to
# be a multiple of the stride
batch_chunk_len = (np.random.randint(train_params.chunk_len_min,
train_params.chunk_len_max + 1) //
net_info.stride) * net_info.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(train_params.min_sub_batch_size *
train_params.chunk_len_max / batch_chunk_len +
0.5)
main_batch_gen = prepare_random_batches(read_data, batch_chunk_len,
sub_batch_size,
train_params.sub_batches,
alphabet_info, filter_params,
net_info, logs.main)
# take optimiser step
optim_info.optimiser.zero_grad()
chunk_count, fval, chunk_samples, chunk_bases, batch_rejections = \
calculate_loss(
net_info, main_batch_gen, sharpen, mod_info.mod_cat_weights,
mod_factor, calc_grads=True)
grad_maxs = apply_clipping(net_info, grad_max_threshs)
optim_info.optimiser.step()
if optim_info.rolling_mads is not None:
grad_max_threshs = optim_info.rolling_mads.update(grad_maxs)
# record step information
if res_info.is_lead_process:
grad_max_thresh_str = ','.join(
('NA' if l_gmt is None else str(float(l_gmt))
for l_gmt in grad_maxs))
logs.batch.write(
BATCH_TMPLT.format(curr_iter + 1, fval,
','.join(map(str, grad_maxs)),
grad_max_thresh_str,
optim_info.lr_scheduler.get_last_lr()[0],
batch_chunk_len))
total_chunks += chunk_count
total_samples += chunk_samples
total_bases += chunk_bases
score_smoothed.update(fval)
# Update counts of reasons for rejection
for k, v in batch_rejections.items():
rejection_dict[k] += v
logs.main.write('.')
if (curr_iter + 1) % DOTROWLENGTH == 0:
log_polka(net_info, train_params, optim_info, time_last,
score_smoothed, curr_iter, total_samples, total_bases,
rejection_dict, res_info, logs.main)
time_last = time.time()
total_bases = total_samples = 0
if (curr_iter + 1) % train_params.save_every == 0:
# Save model and validate
if res_info.is_lead_process:
saved_filename = helpers.save_model(
net_info.net, train_params.outdir,
(curr_iter + 1) // train_params.save_every,
net_info.net_clone)
logs.main.write("Model saved to {}.\n".format(saved_filename))
log_validation(net_info, reporting_batch_list, train_params,
mod_info, curr_iter, logs)
time_last = time.time()
# step learning rate scheduler
optim_info.lr_scheduler.step()
if res_info.is_lead_process:
helpers.save_model(net_info.net,
train_params.outdir,
model_skeleton=net_info.net_clone)
def load_network(args, alphabet_info, res_info, log):
log.write('* Reading network from {}\n'.format(args.model))
if res_info.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.
model_kwargs = {
'stride': args.stride,
'winlen': args.winlen,
'insize': 1,
'size': args.size,
'alphabet_info': alphabet_info
}
model_metadata = {
'reverse': args.reverse,
'standardize': args.standardize
}
net_clone = helpers.load_model(args.model,
model_metadata=model_metadata,
**model_kwargs)
log.write('* Network has {} parameters.\n'.format(
sum(p.nelement() for p in net_clone.parameters())))
if not alphabet_info.is_compatible_model(net_clone):
sys.stderr.write(
'* ERROR: Model and mapped signal files contain ' +
'incompatible alphabet definitions (including modified ' +
'bases).')
sys.exit(1)
if layers.is_cat_mod_model(net_clone):
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)
if layers.is_delta_model(net_clone) and model_metadata.standardize:
log.write('*' * 60 + '\n* WARNING: Delta-scaling models trained ' +
'with --standardize are not compatible with Guppy.\n' +
'*' * 60)
log.write('* Dumping initial model\n')
helpers.save_model(net_clone, args.outdir, 0)
else:
net_clone = None
if res_info.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(res_info.device)
network_metadata = parse_network_metadata(network)
# 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:
log.write('* Loading model onto device\n')
network = net_clone.to(res_info.device)
network_metadata = parse_network_metadata(network)
net_clone = None
log.write('* Estimating filter parameters from training data\n')
stride = guess_model_stride(network)
optimiser = torch.optim.AdamW(network.parameters(),
lr=args.lr_max,
betas=args.adam,
weight_decay=args.weight_decay,
eps=args.eps)
lr_warmup = args.lr_min if args.lr_warmup is None else args.lr_warmup
adam_beta1, _ = args.adam
if args.warmup_batches >= args.niteration:
sys.stderr.write('* Error: --warmup_batches must be < --niteration\n')
sys.exit(1)
warmup_fraction = args.warmup_batches / args.niteration
# Pytorch OneCycleLR crashes if pct_start==1 (i.e. warmup_fraction==1)
lr_scheduler = torch.optim.lr_scheduler.OneCycleLR(
optimiser,
args.lr_max,
total_steps=args.niteration,
# pct_start is really fractional, not percent
pct_start=warmup_fraction,
div_factor=args.lr_max / lr_warmup,
final_div_factor=lr_warmup / args.lr_min,
cycle_momentum=(args.min_momentum is not None),
base_momentum=adam_beta1 if args.min_momentum is None \
else args.min_momentum,
max_momentum=adam_beta1
)
log.write(
('* Learning rate increases from {:.2e} to {:.2e} over {} ' +
'iterations using cosine schedule.\n').format(lr_warmup, args.lr_max,
args.warmup_batches))
log.write(('* Then learning rate decreases from {:.2e} to {:.2e} over ' +
'{} iterations using cosine schedule.\n').format(
args.lr_max, args.lr_min,
args.niteration - args.warmup_batches))
if args.gradient_clip_num_mads is None:
log.write('* No gradient clipping\n')
rolling_mads = None
else:
nparams = len([p for p in network.parameters() if p.requires_grad])
if nparams == 0:
rolling_mads = None
log.write('* No gradient clipping due to missing parameters\n')
else:
rolling_mads = maths.RollingMAD(nparams,
args.gradient_clip_num_mads)
log.write((
'* Gradients will be clipped (by value) at {:3.2f} MADs ' +
'above the median of the last {} gradient maximums.\n').format(
rolling_mads.n_mads, rolling_mads.window))
net_info = NETWORK_INFO(net=network,
net_clone=net_clone,
metadata=network_metadata,
stride=stride)
optim_info = OPTIM_INFO(optimiser=optimiser,
lr_warmup=lr_warmup,
lr_scheduler=lr_scheduler,
rolling_mads=rolling_mads)
return net_info, optim_info
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)