def calculate_loss(network,
network_is_catmod,
batch_gen,
sharpen,
can_mods_offsets=None,
mod_cat_weights=None,
mod_factor_t=None,
calc_grads=False):
total_chunk_count = 0
total_fval = 0
total_samples = 0
total_bases = 0
rejection_dict = defaultdict(int)
n_subbatches = 0
for (indata, seqs, seqlens, mod_cats, sub_batch_size,
batch_rejections) in batch_gen:
n_subbatches += 1
# Update counts of reasons for rejection
for k, v in batch_rejections.items():
rejection_dict[k] += v
total_chunk_count += sub_batch_size
with torch.set_grad_enabled(calc_grads):
outputs = network(indata)
if network_is_catmod:
lossvector = ctc.cat_mod_flipflop_loss(outputs, seqs, seqlens,
mod_cats,
can_mods_offsets,
mod_cat_weights,
mod_factor_t, sharpen)
else:
lossvector = ctc.crf_flipflop_loss(outputs, seqs, seqlens,
sharpen)
non_zero_seqlens = (seqlens > 0.0).float().sum()
# In multi-GPU mode, gradients are synchronised when
# loss.backward() is called. We need to make sure we are
# calculating a gradient that can be synchronised across processes
# - so loss must be per-block-in-batch
loss = lossvector.sum() / non_zero_seqlens
fval = float(loss)
total_fval += fval
total_samples += int(indata.nelement())
total_bases += int(seqlens.sum())
if calc_grads:
loss.backward()
if calc_grads:
for p in network.parameters():
if p.grad is not None:
p.grad /= n_subbatches
return total_chunk_count, total_fval / n_subbatches, \
total_samples, total_bases, rejection_dict
def calculate_loss(net_info,
batch_gen,
sharpen,
mod_cat_weights=None,
mod_factor=None,
calc_grads=False):
can_mods_offsets = net_info.metadata.can_mods_offsets
total_chunk_count = total_fval = total_samples = total_bases = \
n_subbatches = 0
rejection_dict = defaultdict(int)
for (indata, seqs, seqlens, mod_cats, sub_batch_size,
batch_rejections) in batch_gen:
n_subbatches += 1
# Update counts of reasons for rejection
for k, v in batch_rejections.items():
rejection_dict[k] += v
total_chunk_count += sub_batch_size
with torch.set_grad_enabled(calc_grads):
outputs = net_info.net(
indata.to(get_model_device(net_info.net), non_blocking=True))
nblk = float(outputs.shape[0])
ntrans = outputs.shape[2]
if net_info.metadata.is_cat_mod:
lossvector = ctc.cat_mod_flipflop_loss(
outputs, seqs, seqlens, mod_cats, can_mods_offsets,
mod_cat_weights * mod_factor, sharpen)
ntrans -= can_mods_offsets[-1]
else:
lossvector = ctc.crf_flipflop_loss(outputs, seqs, seqlens,
sharpen)
lossvector += layers.flipflop_logpartition(
outputs[:, :, :ntrans]) / nblk
# In multi-GPU mode, gradients are synchronised when
# loss.backward() is called. We need to make sure we are
# calculating a gradient that can be synchronised across processes
# - so loss must be per-block-in-batch
loss = lossvector.mean()
if calc_grads:
loss.backward()
fval = float(loss)
total_fval += fval
total_samples += int(indata.nelement())
total_bases += int(seqlens.sum())
if calc_grads:
for p in net_info.net.parameters():
if p.grad is not None:
p.grad /= n_subbatches
return total_chunk_count, total_fval / n_subbatches, \
total_samples, total_bases, rejection_dict
def main():
args = parser.parse_args()
log, loss_log, chunk_log, device = _setup_and_logs(args)
read_data, alphabet_info = _load_data(args, log)
# 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
filter_parameters = chunk_selection.sample_filter_parameters(
read_data,
args.sample_nreads_before_filtering,
(args.chunk_len_min + args.chunk_len_max) // 2,
args,
log,
chunk_log=chunk_log)
log.write(("* Sampled {} chunks: median(mean_dwell)={:.2f}, " +
"mad(mean_dwell)={:.2f}\n").format(
args.sample_nreads_before_filtering, *filter_parameters))
log.write('* Reading network from {}\n'.format(args.model))
model_kwargs = {
'insize': 1,
'winlen': args.winlen,
'stride': args.stride,
'size': args.size,
'alphabet_info': alphabet_info
}
network = helpers.load_model(args.model, **model_kwargs).to(device)
if not isinstance(network.sublayers[-1], layers.GlobalNormFlipFlopCatMod):
log.write(
'ERROR: Model must end with GlobalNormCatModFlipFlop layer, ' +
'not {}.\n'.format(str(network.sublayers[-1])))
sys.exit(1)
can_mods_offsets = network.sublayers[-1].can_mods_offsets
flipflop_can_labels = network.sublayers[-1].can_labels
flipflop_mod_labels = network.sublayers[-1].mod_labels
flipflop_ncan_base = network.sublayers[-1].ncan_base
log.write('* Loaded categorical modifications flip-flop model.\n')
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)
if args.scale_mod_loss:
try:
mod_cat_weights = alphabet_info.compute_mod_inv_freq_weights(
read_data, args.num_inv_freq_reads)
log.write('* Modified base weights: {}\n'.format(
str(mod_cat_weights)))
except NotImplementedError:
log.write(
'* WARNING: Some mods not found when computing inverse ' +
'frequency weights. Consider raising ' +
'[--num_inv_freq_reads].\n')
mod_cat_weights = np.ones(alphabet_info.nbase, dtype=np.float32)
else:
mod_cat_weights = np.ones(alphabet_info.nbase, dtype=np.float32)
log.write('* Dumping initial model\n')
save_model(network, args.outdir, 0)
total_bases = 0
total_chunks = 0
total_samples = 0
# To count the numbers of different sorts of chunk rejection
rejection_dict = defaultdict(int)
score_smoothed = helpers.WindowedExpSmoother()
t0 = time.time()
log.write('* Training\n')
for i in range(args.niteration):
lr_scheduler.step()
mod_factor_t = torch.tensor(args.mod_factor, dtype=torch.float32)
# 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)
seqs, mod_cats, seqlens = [], [], []
for chunk in chunk_batch:
chunk_labels = chunk['sequence']
seqlens.append(len(chunk_labels))
chunk_seq = flipflop_code(
np.ascontiguousarray(flipflop_can_labels[chunk_labels]),
flipflop_ncan_base)
chunk_mod_cats = np.ascontiguousarray(
flipflop_mod_labels[chunk_labels])
seqs.append(chunk_seq)
mod_cats.append(chunk_mod_cats)
seqs, mod_cats = np.concatenate(seqs), np.concatenate(mod_cats)
seqs = torch.tensor(seqs, dtype=torch.float32, device=device)
seqlens = torch.tensor(seqlens, dtype=torch.long, device=device)
mod_cats = torch.tensor(mod_cats, dtype=torch.long, device=device)
optimizer.zero_grad()
outputs = network(indata)
lossvector = ctc.cat_mod_flipflop_loss(outputs, seqs, seqlens,
mod_cats, can_mods_offsets,
mod_cat_weights, mod_factor_t,
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())
del indata, seqs, mod_cats, seqlens, outputs, loss, lossvector
if device.type == 'cuda':
torch.cuda.empty_cache()
loss_log.write('{}\t{:.10f}\t{:.10f}\n'.format(i, fval,
score_smoothed.value))
if (i + 1) % args.save_every == 0:
save_model(network, args.outdir, (i + 1) // args.save_every)
log.write('C')
else:
log.write('.')
if (i + 1) % 50 == 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
log.write((' {:5d} {:5.3f} {:5.2f}s ({:.2f} ksample/s ' +
'{:.2f} kbase/s) lr={:.2e}').format(
(i + 1) // 50, score_smoothed.value, dt,
total_samples / 1000 / dt, total_bases / 1000 / 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
save_model(network, args.outdir)
return