def data_generator_val(directories, sampler, batch_size=16, slices=None, **kwargs):
'''Generate a data stream from a collection of tracks and a sampler'''
seeds = []
for working in directories:
for track in tqdm(find_files(working,ext='h5')):
fname = os.path.join(working,track)
seeds.append(data_sampler(fname, sampler, slices))
# Send it all to a mux
mux = pescador.ChainMux(seeds, mode='cycle', **kwargs)
return pescador.buffer_stream(mux, batch_size, axis=0)
for model in models:
experiment_folder = config_file.DATA_FOLDER + 'experiments/' + str(
model) + '/'
config = json.load(open(experiment_folder + 'config.json'))
print('Experiment: ' + str(model))
print('\n' + str(config))
# pescador: define (finite, batched & parallel) streamer
pack = [config, 'overlap_sampling', config['n_frames'], False]
streams = [
pescador.Streamer(train.data_gen, id, id2audio_repr_path[id],
id2gt[id], pack) for id in ids
]
mux_stream = pescador.ChainMux(streams, mode='exhaustive')
batch_streamer = pescador.Streamer(pescador.buffer_stream,
mux_stream,
buffer_size=TEST_BATCH_SIZE,
partial=True)
batch_streamer = pescador.ZMQStreamer(batch_streamer)
# tensorflow: define model and cost
fuckin_graph = tf.Graph()
with fuckin_graph.as_default():
sess = tf.Session()
[x, y_, is_train, y, normalized_y,
cost] = train.tf_define_model_and_cost(config)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
results_folder = experiment_folder
def eval(config, ids, id2audio_repr_path, support_set, id2gt, id2label,
tf_vars, vis_vars):
[
id_string, save_latents, track_accuracies, printing, transfer_learning,
model_folder
] = vis_vars
if transfer_learning:
[sess, x, q, log_p_y, emb_q, emb_prototypes] = tf_vars
pack = [config, 'overlap_sampling', 1]
eval_streams = [
pescador.Streamer(transfer_train.data_gen, id,
id2audio_repr_path[id], id2gt[id], pack)
for id in ids
]
else:
[sess, x, q, is_train, log_p_y, emb_q, emb_prototypes] = tf_vars
pack = [config, 'overlap_sampling', 42] # 42 being a second of audio
eval_streams = [
pescador.Streamer(sl_train.data_gen, id, id2audio_repr_path[id],
id2gt[id], pack) for id in ids
]
eval_mux_stream = pescador.ChainMux(eval_streams, mode='exhaustive')
eval_batch_streamer = pescador.Streamer(
pescador.buffer_stream,
eval_mux_stream,
buffer_size=config['test_batch_size'],
partial=True)
first_eval = True
count = 0
for eval_batch in eval_batch_streamer:
if transfer_learning:
[probabilities, embeddings,
prototypes] = sess.run([log_p_y, emb_q, emb_prototypes],
feed_dict={
x: support_set,
q: np.expand_dims(eval_batch['X'],
axis=-1)
})
else:
[probabilities, embeddings, prototypes] = sess.run(
[log_p_y, emb_q, emb_prototypes],
feed_dict={
x: support_set,
q: np.expand_dims(eval_batch['X'], axis=-1),
is_train: False
})
if first_eval:
first_eval = False
pred_array = probabilities
id_array = eval_batch['ID']
if save_latents:
embed_array = embeddings
gt_array = eval_batch['Y']
else:
count = count + 1
pred_array = np.concatenate((pred_array, probabilities), axis=0)
id_array = np.append(id_array, eval_batch['ID'])
if save_latents:
embed_array = np.concatenate((embed_array, embeddings), axis=0)
gt_array = np.concatenate((gt_array, eval_batch['Y']), axis=0)
epoch_acc = shared.accuracy_with_aggergated_predictions(
pred_array, id_array, ids, id2label)
if printing:
print(id_string + ' Number of audios: ' + str(len(ids)))
print(id_string + ' Accuracy: ' + str(epoch_acc))
print(id_string + ' Prototypes: ' + str(prototypes.shape))
if track_accuracies:
fac = open(model_folder + 'epoch_accuracies.tsv', 'a')
fac.write(str(epoch_acc) + '\n')
fac.close()
if save_latents:
print(id_string + ' Embed_array: ' + str(embed_array.shape))
print(id_string + ' GT: ' + str(gt_array.shape))
np.savez(model_folder + 'embeddings_' + id_string + '.npz',
embed_array)
np.savez(model_folder + 'prototypes.npz', prototypes)
np.savez(model_folder + 'gt_' + id_string + '.npz', gt_array)
print('Storing latents for visualization..')
print('\nPrototypes: ')
print(prototypes)
return epoch_acc
def fetch_data(classes_vector,
label2selectedIDs,
id2audio_repr_path,
id2gt,
config,
transfer_learning=False):
set_dic = {}
gt_dic = {}
id_dic = {}
minimum_number_of_patches = np.inf
total_number_of_patches = 0
for c in classes_vector:
# pescador: to batch the computations
preprocess_batch_size = np.min(
[len(label2selectedIDs[c]), config['preprocess_batch_size']])
print('Batch size: ' + str(preprocess_batch_size))
print('IDs used for computing the category ' + str(c) +
' prototype: ' + str(label2selectedIDs[c]))
pack = [
config, config['train_sampling'], config['param_train_sampling']
]
if transfer_learning:
streams = [
pescador.Streamer(transfer_train.data_gen, id,
id2audio_repr_path[id], id2gt[id], pack)
for id in label2selectedIDs[c]
]
else:
streams = [
pescador.Streamer(sl_train.data_gen, id,
id2audio_repr_path[id], id2gt[id], pack)
for id in label2selectedIDs[c]
]
mux_stream = pescador.ChainMux(streams, mode='exhaustive')
batch_streamer = pescador.Streamer(pescador.buffer_stream,
mux_stream,
buffer_size=preprocess_batch_size,
partial=True)
# construct data vectors
first = True
gt = []
for batch in batch_streamer:
if first:
class_set = batch['X']
class_gt = batch['Y']
class_id = batch['ID']
first = False
else:
class_set = np.concatenate((class_set, batch['X']), axis=0)
class_gt = np.concatenate((class_gt, batch['Y']), axis=0)
class_id = np.concatenate((class_id, batch['ID']), axis=0)
print(class_set.shape)
print(class_gt.shape)
print(class_id.shape)
set_dic[c] = class_set
gt_dic[c] = class_gt
id_dic[c] = class_id
minimum_number_of_patches = min(minimum_number_of_patches,
class_set.shape[0])
total_number_of_patches += class_set.shape[0]
return [
set_dic, gt_dic, id_dic, minimum_number_of_patches,
total_number_of_patches
]
data = np.load(filename)
X = data['X']
Y = data['Y']
for i in range(m):
yield dict(X=X[i], y=Y[i])
############################
# Constructing the streamers
############################
# First, we'll make a streamer for each validation example.
#
val_streams = [data_gen(fn, M) for fn in val_files]
############################
# Building the mux
############################
# The `ChainMux` can be used to combine data from all val_streams
# in order.
# We'll use `cycle` mode here, so that the chain restarts after
# all of the streamers have been exhausted.
# This produces an infinite stream of data from a finite sequence
# that repeats every `N*M` steps.
# This can be used in `keras`'s `fit_generator` function
# with `validation_steps=N*M` to ensure that the validation set is
# constant at each epoch.
val_stream = pescador.ChainMux(val_streams, mode='cycle')
def train(self, train_dir, kk=0, folds=10000, grid=False, grid_file=None):
p = self.p
net = self.build()
net.to(device)
net.train()
# Loss and Optimizer
criterion = nn.NLLLoss()
optimizer = torch.optim.Adam(net.parameters(),
lr=p["learning_rate"],
weight_decay=p["reg_lambda"])
# Pescador streams
train_files = [
os.path.join(train_dir, f) for f in os.listdir(train_dir)
]
streams_train = [
pescador.Streamer(indexes_gen,
ff,
1,
p["utter_len"],
p["char_len"],
k=kk,
mode="train",
folds=folds) for ff in train_files
]
mux_stream_train = pescador.ShuffledMux(streams_train, random_state=33)
word_idxs = np.empty(shape=(p["batch_size"],
p["utter_len"] * p["char_len"]),
dtype=int)
labels = np.empty(shape=(p["batch_size"]), dtype=int)
# Train the Model
for epoch in range(p["num_epochs"]):
print("Epoch " + str(epoch))
for i, (word_idx, label, _, _) in enumerate(mux_stream_train):
np.copyto(word_idxs[i % p["batch_size"]], word_idx)
labels[i % p["batch_size"]] = label
if i % p["batch_size"] == 0 and i != 0:
answers = autograd.Variable(torch.LongTensor(labels))
samples = torch.LongTensor(word_idxs)
answers = answers.to(device)
samples = samples.to(device)
optimizer.zero_grad()
outputs = net(samples)
loss = criterion(outputs, answers)
loss.backward()
optimizer.step()
if (i + 1) % 20 == 0:
print("Epoch [%d/%d], Batch [%d], Loss: %.4f" %
(epoch + 1, p["num_epochs"], i + 1, loss.item()))
if i // p["batch_size"] > p["max_batch_epoch"]:
break
# Estimate intermediate
if grid and (epoch + 1) % 10 == 0:
net.eval()
results_file = tempfile.NamedTemporaryFile(mode="w",
delete=False)
streams_test = [
pescador.Streamer(indexes_gen,
ff,
1,
p["utter_len"],
p["char_len"],
k=kk,
mode="test",
folds=folds) for ff in train_files
]
mux_stream_test = pescador.ChainMux(streams_test)
for i, (word_idx, label, character,
_) in enumerate(mux_stream_test):
samples = torch.LongTensor(
word_idx.reshape((1, p["utter_len"] * p["char_len"])))
samples = samples.to(device)
output = net(samples)
entry = output.cpu().data.numpy()[0]
results_file.write(
str(character[0]) + "\t" + str(label[0]) + "\t" +
"\t".join([
str(y) for y in sorted(enumerate(np.exp(entry)),
key=lambda x: x[1],
reverse=True)
]) + "\n")
results_file.close()
mrr_character = compute_MRR_per_character(results_file.name)
macro_mrr = compute_MRR_per_prof(results_file.name, 1)
auroc = compute_auroc(results_file.name, 1)
grid_file.write(
str(epoch + 1) + "\t" + str(mrr_character) + "\t" +
str(macro_mrr) + "\t" + str(auroc[0]) + "\n")
grid_file.flush()
os.remove(results_file.name)
net.train()
# Save the Model
if grid == False:
self.save(p["model_path"])
def multiplex_tfr(data_dir, n_hops, batch_size, mode="inference", aug_kind_str="none", tfr_str="logmelspec",
label_inputs=False, partial_labels=True, structured=True, active_streamers=32, streamer_rate=1024,
num_cpus=1, multi_label=False, align_perturb=False, single_output="fine"):
tfr_dir = os.path.join(data_dir, tfr_str)
streams = []
# Parse augmentation kind string (aug_kind_str).
if mode == "train":
if aug_kind_str == "none":
augs = ["original"]
elif aug_kind_str == "pitch":
augs = ["original", "pitch"]
elif aug_kind_str == "stretch":
augs = ["original", "stretch"]
elif aug_kind_str == "all-but-noise":
augs = ["original", "pitch", "stretch"]
else:
if aug_kind_str == "all":
augs = ["original", "pitch", "stretch", "noise"]
elif aug_kind_str == "noise":
augs = ["original", "noise"]
else:
raise ValueError('Invalid augmentation kind: {}'.format(aug_kind_str))
# Generate a Pescador streamer for every HDF5 container, that is,
# every unit-augmentation-instance triplet.
aug_dict = get_augmentations()
aug_list = []
class_list = []
class_count = Counter()
for aug_str in augs:
if aug_str == "original":
instances = [aug_str]
else:
n_instances = aug_dict[aug_str]
instances = ["-".join([aug_str, str(instance_id+1)])
for instance_id in range(n_instances)]
if aug_str == "noise" and tfr_str == "logmelspec":
bias = np.float32(-17.0)
else:
bias = np.float32(0.0)
for instanced_aug_str in instances:
aug_dir = os.path.join(tfr_dir, instanced_aug_str)
lms_name = "_".join(["*", instanced_aug_str])
lms_pattern = os.path.join(aug_dir, lms_name + ".h5*")
for lms_path in glob.glob(lms_pattern):
if not is_valid_data_hdf5(lms_path, partial_labels):
continue
taxonomy_code = os.path.splitext(os.path.basename(lms_path))[0].split('_')[1].replace('-', '.')
triplet = annotations.get_taxonomy_code_idx_triplet(taxonomy_code)
coarse_idx, medium_idx, fine_idx = triplet
if structured or single_output == "fine":
bal_idx = fine_idx
elif single_output == "medium":
bal_idx = medium_idx
elif single_output == "coarse":
bal_idx = coarse_idx
else:
raise ValueError("Invalid single output mode:{}".format(single_output))
class_list.append(bal_idx)
class_count[bal_idx] += 1
aug_list.append(aug_str)
stream = pescador.Streamer(yield_tfr, lms_path, n_hops, bias, tfr_str, mode, label_inputs, multi_label, align_perturb)
streams.append(stream)
num_streamers = len(streams)
num_fine_classes = len(class_count)
num_aug = len([k for k in aug_dict.keys() if k != "original"])
class_weights = {cls: (num_streamers / float(num_fine_classes * count)) for cls, count in class_count.items()}
aug_weights = {aug: 1.0 if aug == "original" else 1.0 / num_aug for aug, n_inst in aug_dict.items()}
# Weight examples to balance for class, such that each file is sampled from evenly per class. Additionally,
# Balance so sampling any augmentation type (or original) is equally likely, despite the number of instances
# per augmentation. Within augmentation types, instances are equally likely.
weights = [class_weights[cls] * aug_weights[aug] for cls, aug in zip(class_list, aug_list)]
# Multiplex streamers together.
if num_cpus > 1:
mux = create_zmq_mux(streams, num_cpus, active_streamers, streamer_rate, weights=weights)
else:
mux = pescador.StochasticMux(streams, n_active=active_streamers, rate=streamer_rate, weights=weights)
# Create buffered streamer with specified batch size.
buffered_streamer = pescador.maps.buffer_stream(mux, batch_size)
else:
# If not dealing with augmentations, just go through all HDF5 files
weights = None
bias = np.float32(0.0)
for fname in os.listdir(data_dir):
lms_path = os.path.join(data_dir, fname)
if not is_valid_data_hdf5(lms_path, partial_labels):
continue
stream = pescador.Streamer(yield_tfr, lms_path, n_hops, bias, tfr_str, mode, label_inputs, multi_label, align_perturb)
streams.append(stream)
# Multiplex streamers together, but iterate exhaustively.
mux = pescador.ChainMux(streams, mode='exhaustive')
# Create buffered streamer with specified batch size.
buffered_streamer = cycle_partial_buffer_stream(mux, batch_size)
inputs = ["tfr_input"]
if mode in ('train', 'valid') and structured and label_inputs:
inputs += ["coarse_label_input", "medium_label_input"]
if structured:
outputs = ["y_coarse", "y_medium", "y_fine"]
else:
outputs = ["y_" + single_output]
return pescador.maps.keras_tuples(buffered_streamer,
inputs=inputs,
outputs=outputs)