def predict_loop(model, data, batch_size=128, callbacks=[], log=print, f=None):
if f is None:
f = model._predict
ins = [data[name] for name in model.input_order]
nb_sample = len(ins[0])
outs = []
batches = km.make_batches(nb_sample, batch_size)
index_array = np.arange(nb_sample)
nb_batch = len(batches)
for batch_index, (batch_start, batch_end) in enumerate(batches):
if log is not None:
s = progress(batch_index, nb_batch)
if s is not None:
log(s)
for callback in callbacks:
callback(batch_index, len(batches))
batch_ids = list(index_array[batch_start:batch_end])
ins_batch = km.slice_X(ins, batch_ids)
batch_outs = f(*ins_batch)
if type(batch_outs) != list:
batch_outs = [batch_outs]
if batch_index == 0:
for batch_out in batch_outs:
shape = (nb_sample, ) + batch_out.shape[1:]
outs.append(np.zeros(shape))
for i, batch_out in enumerate(batch_outs):
outs[i][batch_start:batch_end] = batch_out
return dict(zip(model.output_order, outs))
def predict_loop(model, data, batch_size=128, callbacks=[], log=print, f=None):
if f is None:
f = model._predict
ins = [data[name] for name in model.input_order]
nb_sample = len(ins[0])
outs = []
batches = km.make_batches(nb_sample, batch_size)
index_array = np.arange(nb_sample)
nb_batch = len(batches)
for batch_index, (batch_start, batch_end) in enumerate(batches):
if log is not None:
s = progress(batch_index, nb_batch)
if s is not None:
log(s)
for callback in callbacks:
callback(batch_index, len(batches))
batch_ids = list(index_array[batch_start:batch_end])
ins_batch = km.slice_X(ins, batch_ids)
batch_outs = f(*ins_batch)
if type(batch_outs) != list:
batch_outs = [batch_outs]
if batch_index == 0:
for batch_out in batch_outs:
shape = (nb_sample,) + batch_out.shape[1:]
outs.append(np.zeros(shape))
for i, batch_out in enumerate(batch_outs):
outs[i][batch_start:batch_end] = batch_out
return dict(zip(model.output_order, outs))
def predict_on_batch(model, X, normalize=None, batch_size=32, shuffle=False, verbose=0):
predictions = []
nb_samples = X.shape[0]
if shuffle:
X = shuffle_data(X)
# predict
batches = make_batches(nb_samples, batch_size)
for batch_index, (batch_start, batch_end) in enumerate(batches):
batch_logs = {}
batch_logs['batch'] = batch_index
batch_logs['size'] = batch_end - batch_start
# load the actual images; X only contains paths
X_batch = load_samples(X[batch_start:batch_end], batch_end - batch_start)
X_batch = X_batch.astype("float32") / 255
if normalize:
X_batch = X_batch - normalize[0] # mean
X_batch /= normalize[1] # std
predictions += [model.predict_classes(X_batch, verbose=verbose).tolist()]
predictions = np.hstack(predictions).tolist()
return predictions
def predict_on_batch(model,
X,
normalize=None,
batch_size=32,
shuffle=False,
verbose=0):
predictions = []
nb_samples = X.shape[0]
if shuffle:
X = shuffle_data(X)
# predict
batches = make_batches(nb_samples, batch_size)
for batch_index, (batch_start, batch_end) in enumerate(batches):
batch_logs = {}
batch_logs['batch'] = batch_index
batch_logs['size'] = batch_end - batch_start
# load the actual images; X only contains paths
X_batch = load_samples(X[batch_start:batch_end],
batch_end - batch_start)
X_batch = X_batch.astype("float32") / 255
if normalize:
X_batch = X_batch - normalize[0] # mean
X_batch /= normalize[1] # std
predictions += [
model.predict_classes(X_batch, verbose=verbose).tolist()
]
predictions = np.hstack(predictions).tolist()
return predictions
def train_on_batch(model, X, y, nb_classes,
callbacks=None, normalize=None, batch_size=32, class_weight=None, class_acc=True, shuffle=False):
loss = []
acc = []
size = []
nb_samples = X.shape[0]
out_labels = ['loss', 'acc']
if shuffle:
X, y = shuffle_data(X, y)
# batch train
batches = make_batches(nb_samples, batch_size)
for batch_index, (batch_start, batch_end) in enumerate(batches):
batch_logs = {}
batch_logs['batch'] = batch_index
batch_logs['size'] = batch_end - batch_start
if callbacks:
callbacks.on_batch_begin(batch_index, batch_logs)
# load the actual images; X only contains paths
X_batch = load_samples(X[batch_start:batch_end], batch_end - batch_start)
X_batch = X_batch.astype("float32") / 255
y_batch = y[batch_start:batch_end]
y_batch = to_categorical(y_batch, nb_classes)
if normalize:
X_batch = X_batch - normalize[0] # mean
X_batch /= normalize[1] # std
# calculates the overall loss and accuracy
outs = model.train_on_batch(X_batch, y_batch, accuracy=True, class_weight=class_weight)
if type(outs) != list:
outs = [outs]
for l, o in zip(out_labels, outs):
batch_logs[l] = o
# calculates the accuracy per class
if class_acc:
result = calc_class_acc(model, X[batch_start:batch_end], y[batch_start:batch_end], nb_classes,
normalize=normalize,
batch_size=batch_size,
keys=['acc'])
batch_logs['class_acc'] = result['acc']
if callbacks:
callbacks.on_batch_end(batch_index, batch_logs)
return loss, acc, size
def write_loop(ins, fun, write_fun, batch_size=128, callbacks=[], log=print):
nb_sample = len(ins[0])
batches = km.make_batches(nb_sample, batch_size)
index_array = np.arange(nb_sample)
nb_batch = len(batches)
for batch_index, (batch_start, batch_end) in enumerate(batches):
if log is not None:
s = progress(batch_index, nb_batch)
if s is not None:
log(s)
for callback in callbacks:
callback(batch_index, len(batches))
batch_ids = list(index_array[batch_start:batch_end])
ins_batch = km.slice_X(ins, batch_ids)
batch_outs = fun(*ins_batch)
write_fun(batch_outs, batch_start, batch_end)
def write_loop(ins, fun, write_fun, batch_size=128, callbacks=[], log=print):
nb_sample = len(ins[0])
batches = km.make_batches(nb_sample, batch_size)
index_array = np.arange(nb_sample)
nb_batch = len(batches)
for batch_index, (batch_start, batch_end) in enumerate(batches):
if log is not None:
s = progress(batch_index, nb_batch)
if s is not None:
log(s)
for callback in callbacks:
callback(batch_index, len(batches))
batch_ids = list(index_array[batch_start:batch_end])
ins_batch = km.slice_X(ins, batch_ids)
batch_outs = fun(*ins_batch)
write_fun(batch_outs, batch_start, batch_end)
def test_on_batch(model,
X,
y,
nb_classes,
normalize=None,
batch_size=32,
shuffle=False):
loss = []
acc = []
size = []
nb_samples = X.shape[0]
if shuffle:
X, y = shuffle_data(X, y)
# batch test
batches = make_batches(nb_samples, batch_size)
for batch_index, (batch_start, batch_end) in enumerate(batches):
batch_logs = {}
batch_logs['batch'] = batch_index
batch_logs['size'] = batch_end - batch_start
# load the actual images; X only contains paths
X_batch = load_samples(X[batch_start:batch_end],
batch_end - batch_start)
X_batch = X_batch.astype("float32") / 255
y_batch = y[batch_start:batch_end]
y_batch = to_categorical(y_batch, nb_classes)
if normalize:
X_batch = X_batch - normalize[0] # mean
X_batch /= normalize[1] # std
outs = model.test_on_batch(X_batch, y_batch, accuracy=True)
# logging of the loss, acc and batch_size
loss += [float(outs[0])]
acc += [float(outs[1])]
size += [batch_end - batch_start]
return loss, acc, size
def test_on_batch(model, X, y, nb_classes, normalize=None, batch_size=32, shuffle=False):
loss = []
acc = []
size = []
nb_samples = X.shape[0]
if shuffle:
X, y = shuffle_data(X, y)
# batch test
batches = make_batches(nb_samples, batch_size)
for batch_index, (batch_start, batch_end) in enumerate(batches):
batch_logs = {}
batch_logs['batch'] = batch_index
batch_logs['size'] = batch_end - batch_start
# load the actual images; X only contains paths
X_batch = load_samples(X[batch_start:batch_end], batch_end - batch_start)
X_batch = X_batch.astype("float32") / 255
y_batch = y[batch_start:batch_end]
y_batch = to_categorical(y_batch, nb_classes)
if normalize:
X_batch = X_batch - normalize[0] # mean
X_batch /= normalize[1] # std
outs = model.test_on_batch(X_batch, y_batch, accuracy=True)
# logging of the loss, acc and batch_size
loss += [float(outs[0])]
acc += [float(outs[1])]
size += [batch_end - batch_start]
return loss, acc, size
def batch_generator(word_crop, max_len, batch_size, doc_ids, words, word_metas, pos_tags, dependencies, parsetrees, rel_ids, rel_parts, rel_types, rel_senses, words2id, words2id_size, skipgram_offsets, pos_tags2id, pos_tags2id_size, rel_types2id, rel_types2id_size, rel_senses2id, rel_senses2id_size, rel_marking2id, rel_marking2id_size):
"""Batch generator where each sample represents a different discourse relation."""
rel_ids = list(rel_ids) # copy list
while True:
# shuffle relations on each epoch
random.shuffle(rel_ids)
for batch_start, batch_end in make_batches(len(rel_ids), batch_size):
# prepare batch data
_rel_id = []
_token_start = []
_token_end = []
x_words_pad = []
x_words_rand = []
x_skipgram = []
x_pos_tags = []
x_rel_types = []
x_rel_senses = []
x_rel_focus = []
x_rel_marking = []
x_rel_types_one = []
x_rel_senses_one = []
for rel_id in rel_ids[batch_start:batch_end]:
doc_id = rel_parts[rel_id]['DocID']
words_len = len(words[doc_id])
token_min = rel_parts[rel_id]['TokenMin']
token_max = rel_parts[rel_id]['TokenMax']
token_start, token_end = token_boundary_random(token_min, token_max, word_crop, words_len)
_rel_id.append(rel_id)
_token_start.append(token_start)
_token_end.append(token_end)
x1_words_pad, x1_words_rand, x1_skipgram, x1_pos_tags, x1_rel_types, x1_rel_senses, x1_rel_focus, x1_rel_marking, x1_rel_types_one, x1_rel_senses_one = relation_sample(rel_id, token_start, token_end, max_len, doc_ids, words, word_metas, pos_tags, dependencies, parsetrees, rel_ids, rel_parts, rel_types, rel_senses, words2id, words2id_size, skipgram_offsets, pos_tags2id, pos_tags2id_size, rel_types2id, rel_types2id_size, rel_senses2id, rel_senses2id_size, rel_marking2id, rel_marking2id_size)
x_words_pad.append(x1_words_pad)
x_words_rand.append(x1_words_rand)
x_skipgram.append(x1_skipgram)
x_pos_tags.append(x1_pos_tags)
x_rel_types.append(x1_rel_types)
x_rel_senses.append(x1_rel_senses)
x_rel_focus.append(x1_rel_focus)
x_rel_marking.append(x1_rel_marking)
x_rel_types_one.append(x1_rel_types_one)
x_rel_senses_one.append(x1_rel_senses_one)
# yield batch
yield {
'_rel_id': _rel_id,
'_token_start': _token_start,
'_token_end': _token_end,
'x_words_pad': np.asarray(x_words_pad, dtype=np.int32),
'x_words_rand': np.asarray(x_words_rand, dtype=np.int32),
'x_rel_focus': np.asarray(x_rel_focus, dtype=np.int8),
'x_rel_marking': np.asarray(x_rel_marking, dtype=np.int8),
'x_skipgram': np.asarray(x_skipgram, dtype=np.int8),
'x_pos_tags': np.asarray(x_pos_tags, dtype=np.int8),
'x_rel_types': np.asarray(x_rel_types, dtype=np.float32),
'x_rel_senses': np.asarray(x_rel_senses, dtype=np.float32),
'x_rel_types_one': np.asarray(x_rel_types_one, dtype=np.int8),
'x_rel_senses_one': np.asarray(x_rel_senses_one, dtype=np.int8),
'x_rel_focus_out': np.asarray(x_rel_focus, dtype=np.int8),
}
def batch_generator(word_crop, max_len, batch_size, doc_ids, words, word_metas,
pos_tags, dependencies, parsetrees, rel_ids, rel_parts,
rel_types, rel_senses, words2id, words2id_size,
skipgram_offsets, pos_tags2id, pos_tags2id_size,
rel_types2id, rel_types2id_size, rel_senses2id,
rel_senses2id_size, rel_marking2id, rel_marking2id_size):
"""Batch generator where each sample represents a different discourse relation."""
rel_ids = list(rel_ids) # copy list
while True:
# shuffle relations on each epoch
random.shuffle(rel_ids)
for batch_start, batch_end in make_batches(len(rel_ids), batch_size):
# prepare batch data
_rel_id = []
_token_start = []
_token_end = []
x_words_pad = []
x_words_rand = []
x_skipgram = []
x_pos_tags = []
x_rel_types = []
x_rel_senses = []
x_rel_focus = []
x_rel_marking = []
x_rel_types_one = []
x_rel_senses_one = []
for rel_id in rel_ids[batch_start:batch_end]:
doc_id = rel_parts[rel_id]['DocID']
words_len = len(words[doc_id])
token_min = rel_parts[rel_id]['TokenMin']
token_max = rel_parts[rel_id]['TokenMax']
token_start, token_end = token_boundary_random(
token_min, token_max, word_crop, words_len)
_rel_id.append(rel_id)
_token_start.append(token_start)
_token_end.append(token_end)
x1_words_pad, x1_words_rand, x1_skipgram, x1_pos_tags, x1_rel_types, x1_rel_senses, x1_rel_focus, x1_rel_marking, x1_rel_types_one, x1_rel_senses_one = relation_sample(
rel_id, token_start, token_end, max_len, doc_ids, words,
word_metas, pos_tags, dependencies, parsetrees, rel_ids,
rel_parts, rel_types, rel_senses, words2id, words2id_size,
skipgram_offsets, pos_tags2id, pos_tags2id_size,
rel_types2id, rel_types2id_size, rel_senses2id,
rel_senses2id_size, rel_marking2id, rel_marking2id_size)
x_words_pad.append(x1_words_pad)
x_words_rand.append(x1_words_rand)
x_skipgram.append(x1_skipgram)
x_pos_tags.append(x1_pos_tags)
x_rel_types.append(x1_rel_types)
x_rel_senses.append(x1_rel_senses)
x_rel_focus.append(x1_rel_focus)
x_rel_marking.append(x1_rel_marking)
x_rel_types_one.append(x1_rel_types_one)
x_rel_senses_one.append(x1_rel_senses_one)
# yield batch
yield {
'_rel_id': _rel_id,
'_token_start': _token_start,
'_token_end': _token_end,
'x_words_pad': np.asarray(x_words_pad, dtype=np.int32),
'x_words_rand': np.asarray(x_words_rand, dtype=np.int32),
'x_rel_focus': np.asarray(x_rel_focus, dtype=np.int8),
'x_rel_marking': np.asarray(x_rel_marking, dtype=np.int8),
'x_skipgram': np.asarray(x_skipgram, dtype=np.int8),
'x_pos_tags': np.asarray(x_pos_tags, dtype=np.int8),
'x_rel_types': np.asarray(x_rel_types, dtype=np.float32),
'x_rel_senses': np.asarray(x_rel_senses, dtype=np.float32),
'x_rel_types_one': np.asarray(x_rel_types_one, dtype=np.int8),
'x_rel_senses_one': np.asarray(x_rel_senses_one,
dtype=np.int8),
'x_rel_focus_out': np.asarray(x_rel_focus, dtype=np.int8),
}
loss_best = loss_avg
loss_min = float(stats.history[-1]['loss_min'])
loss_max = float(stats.history[-1]['loss_max'])
log.info(" continue from epoch {}, loss avg: {:.4f}, min: {:.4f}, max: {:.4f}".format(epoch_i, loss_avg, loss_min, loss_max))
# train model
while epoch_i < epochs:
epoch_i += 1
time_0 = time.time()
log.info("train epoch {}/{} ({} docs)".format(epoch_i, epochs, len(train_doc_ids)))
# one document per batch update
loss_avg = 0.
loss_min = np.inf
loss_max = -np.inf
batches = make_batches(len(train_doc_ids), batch_size)
for batch_i, (batch_start, batch_end) in enumerate(batches):
# prepare batch data
doc_ids = train_doc_ids[batch_start:batch_end]
x_word_pad, x_word_rand = build_x_word(doc_ids, train_words, word2id, word_crop, max_len)
#print("x_word_pad:", x_word_pad.shape); pprint(x_word_pad)
#print("x_word_rand:", x_word_rand.shape); pprint(x_word_rand)
y_skipgram = build_y_skipgram(x_word_pad, skipgram_offsets, max_len)
#print("y_skipgram:", y_skipgram.shape); pprint(y_skipgram)
y_pos = build_y_pos(doc_ids, train_words, pos2id, word_crop, max_len)
#print("y_pos:", y_pos.shape); pprint(y_pos)
y_pdtbpair = build_y_pdtbpair(doc_ids, train_words, pdtbpair_offsets, pdtbpair2id, pdtbpair2id_weights, word_crop, max_len)
log.info(
" continue from epoch {}, loss avg: {:.4f}, min: {:.4f}, max: {:.4f}"
.format(epoch_i, loss_avg, loss_min, loss_max))
# train model
while epoch_i < epochs:
epoch_i += 1
time_0 = time.time()
log.info("train epoch {}/{} ({} docs)".format(epoch_i, epochs,
len(train_doc_ids)))
# one document per batch update
loss_avg = 0.
loss_min = np.inf
loss_max = -np.inf
batches = make_batches(len(train_doc_ids), batch_size)
for batch_i, (batch_start, batch_end) in enumerate(batches):
# prepare batch data
doc_ids = train_doc_ids[batch_start:batch_end]
x_word_pad, x_word_rand = build_x_word(doc_ids, train_words,
word2id, word_crop, max_len)
#print("x_word_pad:", x_word_pad.shape); pprint(x_word_pad)
#print("x_word_rand:", x_word_rand.shape); pprint(x_word_rand)
y_skipgram = build_y_skipgram(x_word_pad, skipgram_offsets,
max_len)
#print("y_skipgram:", y_skipgram.shape); pprint(y_skipgram)
y_pos = build_y_pos(doc_ids, train_words, pos2id, word_crop,
max_len)
def train_on_batch(model,
X,
y,
nb_classes,
callbacks=None,
normalize=None,
batch_size=32,
class_weight=None,
class_acc=True,
shuffle=False):
loss = []
acc = []
size = []
nb_samples = X.shape[0]
out_labels = ['loss', 'acc']
if shuffle:
X, y = shuffle_data(X, y)
# batch train
batches = make_batches(nb_samples, batch_size)
for batch_index, (batch_start, batch_end) in enumerate(batches):
batch_logs = {}
batch_logs['batch'] = batch_index
batch_logs['size'] = batch_end - batch_start
if callbacks:
callbacks.on_batch_begin(batch_index, batch_logs)
# load the actual images; X only contains paths
X_batch = load_samples(X[batch_start:batch_end],
batch_end - batch_start)
X_batch = X_batch.astype("float32") / 255
y_batch = y[batch_start:batch_end]
y_batch = to_categorical(y_batch, nb_classes)
if normalize:
X_batch = X_batch - normalize[0] # mean
X_batch /= normalize[1] # std
# calculates the overall loss and accuracy
outs = model.train_on_batch(X_batch,
y_batch,
accuracy=True,
class_weight=class_weight)
if type(outs) != list:
outs = [outs]
for l, o in zip(out_labels, outs):
batch_logs[l] = o
# calculates the accuracy per class
if class_acc:
result = calc_class_acc(model,
X[batch_start:batch_end],
y[batch_start:batch_end],
nb_classes,
normalize=normalize,
batch_size=batch_size,
keys=['acc'])
batch_logs['class_acc'] = result['acc']
if callbacks:
callbacks.on_batch_end(batch_index, batch_logs)
return loss, acc, size
