def tf_is_emotion(image_file, emotion, match_emotion):
[
match,
] = tf.py_function(lambda x: is_emotion(x, match_emotion), [emotion],
[tf.bool])
return match
def tf_serialize_example(f0, f1):
tf_string = tf.py_function(
serialize_example,
(f0, f1), # pass these args to the above function.
tf.string) # the return type is `tf.string`.
return tf.reshape(tf_string, ()) # The result is a scalar
def load_from_txt(X_fpath,
Y_fpath,
word2vec_fpath,
vocab_fpath,
reverse_context=True,
buffer_size=10000,
batch_size=32):
def txt_generator():
with open(X_fpath, 'r') as f:
X = f.readlines()
with open(Y_fpath, 'r') as f:
Y = f.readlines()
print('Loaded data from txt file')
X = [x.strip('\n') for x in X]
Y = [y.strip('\n') for y in Y]
print('Removed new line characters')
# Pipeline elements
ws = WhiteSpaceTokenizer()
padder = Padder('<pad>')
pipe = Pipeline(steps=[('ws', ws), ('pad', padder)])
X_processed = pipe.transform(X)
Y_processed = pipe.transform(Y)
print('ws tokenized and padded')
for i, x in enumerate(X_processed):
yield (x, Y_processed[i])
# Word2Vec used to vectorise the encoder and decoder inputs
word2vec = Word2Vec(special_vectors={'unknown': 0, '<sos>': 1})
model = KeyedVectors.load_word2vec_format(word2vec_fpath)
word2vec.set_model(model)
print('Loaded word2vec mapping')
inttk = IntegerTokenizer(vocab_fpath, add_to_vocab_if_not_present=False)
ft = featurize(ftz=[word2vec.tf_map, inttk.tf_map])
dataset = tf.data.Dataset.from_generator(
txt_generator,
output_types=(tf.string, tf.string),
output_shapes=(tf.TensorShape((None, )), tf.TensorShape((None, ))),
)
print('Created tf.data.Dataset')
dataset = dataset.map(create_decoder_target)
dataset = dataset.map(
lambda x1, x2, y: tf.py_function(func=ft.tf_map,
inp=[x1, x2, y],
Tout=(tf.float32, tf.float32, tf.int32
)),
num_parallel_calls=tf.data.experimental.AUTOTUNE,
)
dataset = dataset.map(lambda x1, x2, y: ((x1, x2), y))
if reverse_context:
dataset = dataset.map(flip_context_func)
# dataset is fairly small (c. 200mb) so use the cache
dataset.cache()
# Shuffle and batch the dataset,
dataset = dataset.shuffle(buffer_size).batch(batch_size,
drop_remainder=True)
# Use prefetch to allow model to get batches in the background while training
dataset.prefetch(tf.data.experimental.AUTOTUNE)
return dataset
def wrapper(steps, times, values):
# Use a tf.py_function to prevent auto-graph from compiling the method
return tf.py_function(func,
inp=(steps, times, values),
Tout=(steps.dtype, times.dtype, values.dtype))
def tf_write_summary(tar_real, predictions, step): return tf.py_function(write_summary, [tar_real, predictions, step], Tout=[tf.float32, tf.float32])
if lib in ["tf_decode_wav"]:
dataset = dataset.map(
lambda x: loaders.load_tf_decode_wav(x),
num_parallel_calls=1)
elif lib in ["tfio_fromaudio"]:
dataset = dataset.map(
lambda x: loaders.load_tfio_fromaudio(x, args.ext),
num_parallel_calls=1)
elif lib in ["tfio_fromffmpeg"]:
dataset = dataset.map(
lambda x: loaders.load_tfio_fromffmpeg(x),
num_parallel_calls=1)
else:
loader_function = getattr(loaders, 'load_' + lib)
dataset = dataset.map(lambda filename: tf.py_function(
_make_py_loader_function(loader_function), [filename],
[tf.float32]),
num_parallel_calls=1)
dataset = dataset.apply(tf.data.experimental.ignore_errors())
start = time.time()
for i in range(repeat):
for audio in dataset:
value = tf.reduce_max(audio)
if value:
append = True
end = time.time()
def _process_text_map_fn(self, text, label):
processed_text, label = tf.py_function(self._process_text,
inp=[text, label],
Tout=(tf.float32, tf.int64))
return processed_text, label
def tf_file_exists(filepath):
return tf.py_function(
lambda x: os.path.exists(x.numpy()),
inp=[filepath],
Tout=tf.bool)
def tf_encode(pt, en):
result_pt, result_en = tf.py_function(encode, [pt, en], [tf.int64, tf.int64])
result_pt.set_shape([None])
result_en.set_shape([None])
return result_pt, result_en
def model_part3(results):
'''
apply nms on the rois
:param results:
:return:
'''
per_roi = results[4]
per_roi = per_roi[:, 1:]
score = tf.reshape(results[2], (-1, 1))
roi = tf.concat((per_roi, score), axis=-1)
# apply nms on rois to get boxes with highest scores
# keep = tf.py_function(nms, [roi, 0.3], tf.int32)
# keep = nms(roi, 0.3)
score = tf.reshape(results[2], (-1, ))
keep = tf.image.non_max_suppression(per_roi,
score,
max_output_size=20,
iou_threshold=0.3)
ovr_threshold = tf.convert_to_tensor(0.5)
for i in keep:
keep1 = tf.numpy_function(nms1, [i, keep, roi, ovr_threshold],
tf.int32)
box_up_num = len(keep1) + 1
# get an unsupressed box
mask_i = results[3][i] # (h,w,2)
roi_i = results[4][i] # (n,4)
score_i = results[2][i] # (n,)
mask_i, roi_i = tf.py_function(mask_transform, [mask_i, roi_i],
[tf.float32, tf.float32])
mask_i = tf.pad(tensor=mask_i,
paddings=[[roi_i[-4], 0], [roi_i[-3], 0], [0, 0]])
mask_i *= score_i
# fuse the unsupressed box with supressed boxes by weighted averaging
for j in keep1:
mask_j = results[3][j]
roi_j = results[4][j]
score_j = results[3][j]
mask_j, roi_j = tf.py_function(mask_transform, [mask_j, roi_j],
[tf.float32, tf.float32])
rb = tf.maximum(roi_i[-2:], roi_j[-2:])
pad_rb = rb - roi_j[-2:]
mask_j = tf.pad(tensor=mask_j,
paddings=[[roi_j[-4], pad_rb[0]],
[roi_j[-3], pad_rb[1]], [0, 0]])
mask_i = mask_i + mask_j * score_j
mask_i /= box_up_num
mask = mask_i[roi_i[-4]:, roi_i[-3]:, :]
results[3][i] = mask
# compute the postive boxes and positive boxes
# compute the boxes' inside_weights and out_side weights
roi_item = result[4][keep[0]] # (4,)
offset_item = result[5][keep[0]] # (4,)
cls_item = result[0][keep[0]]
rois = tf.reshape(roi_item, (1, -1))
offset = tf.reshape(offset_item, (1, -1))
cls = tf.reshape(cls_item, (1, -1))
roi_mask_item = result[3][keep[0]]
roi_mask = tf.reshape(roi_mask_item, (1, 5, 5, 2))
for i in range(1, len(keep)):
roi_item = result[4][keep[i]]
roi_item = tf.reshape(roi_item, (1, -1))
rois = tf.concat((rois, roi_item), axis=0) # (n,4)
offset_item = result[5][keep[i]]
offset_item = tf.reshape(offset_item, (1, -1))
offset = tf.concat((offset, offset_item), axis=0) # (n,4)
cls_item = result[0][keep[i]]
cls_item = tf.reshape(cls_item, (1, -1))
cls = tf.concat((cls, cls_item), axis=0) # (n,k+1)
roi_mask_item = result[3][keep[i]]
roi_mask_item = tf.reshape(roi_mask_item, (1, 5, 5, 2))
roi_mask = tf.concat((roi_mask, roi_mask_item), axis=0)
return rois, cls, offset, roi_mask
def tf_load_audio(filepath):
return tf.py_function(
lambda x: load_audio(x.numpy()),
inp=[filepath],
Tout=[tf.float32, tf.int32])
def load_image(path, label):
return tf.py_function(_load_image, (path, label), (tf.float32, tf.int32))
def tf_serialize_example(f0, f1, f2, f3):
tf_string = tf.py_function(serialize_example, (f0, f1, f2, f3), tf.string)
return tf.reshape(tf_string, ())
def f(x):
tf.py_function(side_effect, inp=[x], Tout=[])
def load_image(path, label):
image, label = tf.py_function(_load_image, (path, label),
(tf.float32, tf.int32))
image.set_shape([None, None, None])
label.set_shape([])
return tf.image.resize(image, INPUT_IMAGE_SIZE), label
def inTrainFilter(ident, _dummy_1, _dummy_2):
return not (tf.py_function(inValidationFilter, [ident], (tf.bool)))
def alaska_tf(y_true, y_val):
"""Wrapper for the above function"""
return tf.py_function(func=alaska_wuac_metric,
inp=[y_true, y_val],
Tout=tf.float32)
def _augment_and_encode(x, y):
if augment:
x = aug_func(x)
y = tf.py_function(_encode_text, (y, ), Tout=tf.int64)
return x, y
return z
#（2）建立网络模型
# 创建模型
# 占位符
X = tf.compat.v1.placeholder("float")
Y = tf.compat.v1.placeholder("float")
# 模型参数
W = tf.Variable(tf.random.normal([1]), name="weight")
b = tf.Variable(tf.zeros([1]), name="bias")
# 前向结构
#z = tf.multiply(X, W)+ b
z = tf.py_function(my_py_func, [X, W, b], tf.float32)
global_step = tf.Variable(0, name='global_step', trainable=False)
#反向优化
cost = tf.reduce_mean(input_tensor=tf.square(Y - z))
learning_rate = 0.01
optimizer = tf.compat.v1.train.GradientDescentOptimizer(
learning_rate).minimize(cost, global_step) #梯度下降
# 定义学习参数
training_epochs = 34
display_step = 2
savedir = "log/"
saver = tf.compat.v1.train.Saver(
tf.compat.v1.global_variables(),
max_to_keep=1) #生成saver。 max_to_keep=1，表明最多只保存一个检查点文件
def tf_serialize_example(image_str, label):
tf_string = tf.py_function(
serialize_example,
(image_str, label),
tf.string)
return tf.reshape(tf_string, ()) # The result is a scalar
def mapped_function(s):
# Do some hard pre-processing
tf.py_function(lambda: time.sleep(0.03), [], ())
return s
def load_file_and_preprocess(path):
shape = (tf.float16, tf.int16) if config['float16'] else (tf.float32,
tf.int32)
pyf = tf.py_function(wrapped_loader, [path], shape)
return pyf
def tf_encode(data):
result_input, result_label = tf.py_function(encode, [data[0], data[1]],
[tf.int64, tf.int64])
return result_input, result_label
def tf_iou(y_true, y_pred):
iou = tf.py_function(get_iou, [y_true, y_pred], tf.float32)
return iou
def tf_encode(doc, summary):
return tf.py_function(encode, [doc, summary], [tf.int64, tf.int64])
def tf_encode(txt, eq):
return tf.py_function(encode, [txt, eq], [tf.int64, tf.int64])
return blocks, masks, label
listOfFiles = os.listdir("/tf/kaggle/tf_records")
pattern = "*.tfrecords"
listOfFiles = [
"/tf/kaggle/tf_records/" + x for x in listOfFiles
if fnmatch.fnmatch(x, pattern)
]
print(listOfFiles)
dataset = tf.data.TFRecordDataset(listOfFiles)
print(dataset)
#train=dataset.map(parse_tfrecord)
train = dataset.map(lambda x1: tf.py_function(
func=parse_tfrecord, inp=[x1], Tout=[tf.float32, tf.int32, tf.int64]))
train = train.batch(4)
print(train)
layer_1 = tf.keras.layers.Conv2D(6, (7, 7),
activation='relu',
input_shape=(7, 7, 1280))
layer_12 = tf.keras.layers.Conv2D(6, (1, 1),
activation='relu',
input_shape=(1, 1, 6))
layer_2 = tf.keras.layers.Conv3D(6, (max_num_blocks, 1, 1),
activation='relu',
input_shape=(7, 7, 1280))
for blocks, masks, labels in train:
def _deploy_exe_info(self, losses, info):
with tf.name_scope("deploy_exe_info"):
hp = self.hparams
if self.trainable: # Train
self.train_loss = losses
params = tf.trainable_variables()
if hp.tunable:
learning_rate = hp.tune_rate
else:
learning_rate = hp.learning_rate
self.learning_rate = tf.constant(learning_rate, dtype=tf.float32)
# Warm-up
self.learning_rate = self._get_learning_rate_warmup()
# Decay
self.learning_rate = self._get_learning_rate_decay()
# Optimizer
opt = tf.train.MomentumOptimizer(self.learning_rate, hp.momentum_factor)
# Gradient
gradients = tf.gradients(self.train_loss, params)
# Gradient clip
clipped_grads, grad_norm_summaries, grad_norm = helper.gradient_clip(
gradients, max_gradient_norm=hp.max_grad_norm)
# Gradient norm
for summary in grad_norm_summaries:
self._add_to_summaries(summary)
self.grad_norm = grad_norm
# Apply update to params
self.update = opt.apply_gradients(
zip(clipped_grads, params), global_step=self.global_step)
# Trainable params summary
print("# Trainable variables")
print("Format: <name>, <shape>, <(soft) device placement>")
for param in params:
self.histogram.update({param.name: param})
print(" %s, %s, %s" % (param.name, str(param.get_shape()),
param.op.device))
self.histogram.update(train_loss=self.train_loss,
learning_rate=self.learning_rate)
if hp.forward_rcnn:
self.class_predicts = self.reverse_cate_table.lookup(
tf.to_int64(info["class_predicts"]))
self.detected_images = tf.py_function(
misc.draw_boxes_on_image,
[self.images_data, info["bbox_labels"],
info["class_scores"], self.class_predicts,
self.im_info, hp.pixel_mean], Tout=tf.float32)
self.train_summary = self._config_train_summary()
elif self.predicable: # Infer
stddevs = tf.tile(tf.constant(hp.bbox_norm_stddevs), multiples=hp.num_class)
means = tf.tile(tf.constant(hp.bbox_norm_means), multiples=hp.num_class)
deltas = info["bbox_predicts"]
# Restore bbox predicts
deltas = tf.add(tf.multiply(deltas, stddevs), means)
info["bbox_predicts"] = deltas
rois = info["rois"]
self.class_scores = info["class_scores"]
self.class_predicts = self.reverse_cate_table.lookup(
tf.to_int64(info["class_predicts"]))
# Get predicted ground-truth bbox
self.bboxes = proposal_util.bboxes_regression(rois, deltas)
self.detected_images = tf.py_function(
misc.draw_boxes_on_image,
[self.images_data, self.bboxes,
self.class_scores, self.class_predicts,
self.im_info, hp.pixel_mean], Tout=tf.float32)
self.infer_summary = self._config_infer_summary()
else: # Eval
rois = info["rois"]
deltas = info["bbox_predicts"]
self.eval_loss = losses
bboxes = proposal_util.bboxes_regression(rois, deltas)
self.accuracy = misc.mean_avg_overlap(
bboxes, self.bbox_labels)
self.eval_summary = self._config_eval_summary()
train_ds, test_ds, val_ds = get_open_shelf_dataset()
def resize_eager_test(image, label):
image = cv2.resize(image, (96, 96))
image = (image / 255).astype(dtype=np.float32)
return image, label
def resize(image, label):
image = cv2.resize(image.numpy(), (96, 96))
return image, label
# val_ds = val_ds.map(lambda item: tf.numpy_function(
# resize_eager_test, [item['image'], item['label']], [tf.float32, tf.int64])
# )
val_ds = val_ds \
.map(lambda item: tf.py_function(resize, [item['image'], item['label']], [tf.uint8, tf.int64])) \
.map(lambda image, label: tf.py_function(data_aug_v2, [image, label], [tf.float32, tf.int64])) \
.batch(batch_size=1)
index = 0
for i in val_ds:
matplotlib.image.imsave('./images/{}-augmentation.jpg'.format(index),
i[0][0].numpy())
index += 1
def tf_file_exists(image_file, label):
[
exists,
] = tf.py_function(file_exists, [image_file], [tf.bool])
return exists