def construct(self, gradients, overflow):
"""AdamWeightDecayForBert"""
lr = self.get_lr()
cond = self.op_cast(P.Fill()(ts.int32, self.op_shape(self.beta1), 1) *\
self.op_reshape(overflow, (())), ts.bool_)
beta1 = self.op_select(cond, self.op_cast(ts.array((1.0, )),
ts.float32), self.beta1)
beta2 = self.op_select(cond, self.op_cast(ts.array((1.0, )),
ts.float32), self.beta2)
if self.is_group:
if self.is_group_lr:
optim_result = self.hyper_map(
P.Partial()(_adam_opt, self.beta1, self.beta2,
self.eps), lr, self.weight_decay,
self.parameters, self.moments1, self.moments2, gradients,
self.decay_flags, self.optim_filter)
else:
optim_result = self.hyper_map(
P.Partial()(_adam_opt, beta1, beta2, self.eps, lr,
overflow), self.weight_decay, self.parameters,
self.moments1, self.moments2, gradients, self.decay_flags,
self.optim_filter)
else:
optim_result = self.hyper_map(
P.Partial()(_adam_opt, self.beta1, self.beta2, self.eps, lr,
self.weight_decay), self.parameters, self.moments1,
self.moments2, gradients, self.decay_flags, self.optim_filter)
if self.use_parallel:
self.broadcast_params(optim_result)
return optim_result
def __init__(self, learning_rate, end_learning_rate, warmup_steps,
decay_steps, power):
super(BertLearningRate, self).__init__()
self.warmup_flag = False
if warmup_steps > 0:
self.warmup_flag = True
self.warmup_lr = WarmUpLR(learning_rate, warmup_steps)
self.decay_lr = PolynomialDecayLR(learning_rate, end_learning_rate,
decay_steps, power)
self.warmup_steps = ts.array([warmup_steps], dtype=ts.float32)
self.greater = P.Greater()
self.one = ts.array([1.0], dtype=ts.float32)
self.cast = P.Cast()
def __init__(self,
params,
learning_rate=1e-3,
beta1=0.9,
beta2=0.999,
eps=1e-6,
weight_decay=0.0):
super(AdamWeightDecayOp, self).__init__(learning_rate, params,
weight_decay)
_check_param_value(beta1, beta2, eps, self.cls_name)
self.beta1 = ts.array([beta1], dtype=ts.float32)
self.beta2 = ts.array([beta2], dtype=ts.float32)
self.eps = ts.array([eps], dtype=ts.float32)
self.moments1 = self.parameters.clone(prefix="adam_m", init='zeros')
self.moments2 = self.parameters.clone(prefix="adam_v", init='zeros')
self.hyper_map = P.HyperMap()
def convert2tensor(self, transform_input):
r"""
Convert the numpy data to the tensor format.
Args:
transform_input (numpy.ndarray): the preprocessing image.
Returns:
Tensor, the converted image.
"""
if not isinstance(transform_input, np.ndarray):
err_msg = 'The transform_input type should be numpy.ndarray, got {}.'.format(
type(transform_input))
raise TypeError(err_msg)
input_tensor = ts.expand_dims(ts.array(list(transform_input)), 0)
return input_tensor
def construct(self):
"""Generates matrix of relative positions between inputs."""
range_vec_row_out = self.cast(ts.array(ts.arange(self._length)), int32)
range_vec_col_out = self.range_mat(range_vec_row_out, (self._length, -1))
tile_row_out = self.tile(range_vec_row_out, (self._length,))
tile_col_out = self.tile(range_vec_col_out, (1, self._length))
range_mat_out = self.range_mat(tile_row_out, (self._length, self._length))
transpose_out = self.range_mat(tile_col_out, (self._length, self._length))
distance_mat = self.sub(range_mat_out, transpose_out)
distance_mat_clipped = P.clip_by_value(distance_mat,
self._min_relative_position,
self._max_relative_position)
# Shift values to be >=0. Each integer still uniquely identifies a
# relative position difference.
final_mat = distance_mat_clipped + self._max_relative_position
return final_mat
def cyclegan_lr(max_epoch, n_epoch, dataset_size):
"""
Generate learning rate for cycle_gan.
Args:
max_epoch (int): Epoch size for training.
n_epoch (int): Number of epochs with the initial learning rate.
dataset_size (int): Total size of dataset.
Returns:
Tensor, learning rate.
"""
n_epochs_decay = max_epoch - n_epoch
lrs = [0.0002] * dataset_size * n_epoch
lr_epoch = 0
for epoch in range(n_epochs_decay):
lr_epoch = 0.0002 * (n_epochs_decay - epoch) / n_epochs_decay
lrs += [lr_epoch] * dataset_size
lrs += [lr_epoch] * dataset_size * (max_epoch - n_epochs_decay - n_epoch)
return ts.array(lrs, dtype=ts.float32)
def postprocess(self, input, strategy='TOP1_CLASS'):
r'''
Apply postprocess operation for prediction result.
Args:
input (numpy.ndarray): Prediction result.
strategy (str): Specifies the postprocess strategy. Default: TOP1_CLASS.
Returns:
str, the postprocess result.
'''
if not isinstance(input, np.ndarray):
raise TypeError("Input should be NumPy, got {}.".format(
type(input)))
if not input.ndim == 2:
raise TypeError("Input should be 2-D Numpy, got {}.".format(
input.ndim))
if strategy not in self.transform_strategy:
raise ValueError("Strategy should be one of {}, got {}.".format(
self.transform_strategy, strategy))
softmax = Softmax()
score_list = softmax(ts.array(input)).asnumpy()
if strategy == 'TOP1_CLASS':
score = max(score_list[0])
return ('TOP1: ' + str(self.labels[input[0].argmax()]) +
', score: ' + str(format(score, '.20f')))
else:
label_index = np.argsort(input[0])[::-1]
score_index = np.sort(score_list[0])[::-1]
top5_labels = []
res = ''
top5_scores = score_index[:5].tolist()
for i in range(5):
top5_labels.append(self.labels[label_index[i]])
res += 'TOP' + str(i+1) + ": " + str(top5_labels[i]) + \
", score: " + str(format(top5_scores[i], '.20f')) + '\n'
return res
def mobilenetv2_lr(global_step, lr_init, lr_end, lr_max, warmup_epochs,
total_epochs, steps_per_epoch):
"""
Generate learning rate for mobilenetv2.
Args:
global_step (int): Total steps of the training.
lr_init (float): Init learning rate.
lr_end (float): End learning rate.
lr_max (float): Max learning rate.
warmup_epochs (int): Number of warmup epochs.
total_epochs (int): Total epoch of training.
steps_per_epoch (int): Steps of one epoch.
Returns:
Tensor, learning rate.
"""
lr_each_step = []
total_steps = steps_per_epoch * total_epochs
warmup_steps = steps_per_epoch * warmup_epochs
for i in range(total_steps):
if i < warmup_steps:
lr = lr_init + (lr_max - lr_init) * i / warmup_steps
else:
lr = lr_end + \
(lr_max - lr_end) * \
(1. + math.cos(math.pi * (i - warmup_steps) / (total_steps - warmup_steps))) / 2.
if lr < 0.0:
lr = 0.0
lr_each_step.append(lr)
current_step = global_step
lr_each_step = ts.array(lr_each_step, dtype=ts.float32)
learning_rate = lr_each_step[current_step:]
return learning_rate
def web_predict(instance, servable_name, servable_model, dataset_name, strategy):
"""
Predict the result based on the input data.
A network will be constructed based on the input and servable data, then load the checkpoint and do the predict.
Args:
instance (dict): the dict of input image after transformation, with keys of `shape`, `dtype` and `data`(Image object).
servable_name (str): servable name
servable_model (str): name of the model
strategy (str): output strategy, usually select between `TOP1_CLASS` and `TOP5_CLASS`, for cyclegan, select between `gray2color` and `color2gray`
Returns:
The dict object of predicted result after post process.
Examples:
>>> # In the server part, after servable_search
>>> res = web_predict(instance, servable_name, servable['model'], strategy)
>>> return jsonify(res)
"""
# check if servable model name is valid
model_name = servable_model['name']
net_func = model_checker.get(model_name)
if net_func is None:
err_msg = "Currently model_name only supports " + str(list(model_checker.keys())) + "!"
return {"status": 1, "err_msg": err_msg}
# check if model_format is valid
model_format = servable_model['format']
if model_format not in ("ckpt"):
err_msg = "Currently model_format only supports `ckpt`!"
return {"status": 1, "err_msg": err_msg}
# Check if dataset supports
trans_func = transform_checker.get(dataset_name)
if trans_func is None:
print("Currently dataset_name only supports {}!".format(list(transform_checker.keys())))
sys.exit(0)
# process the original data
ori_img = np.array(json.loads(instance['data']), dtype=instance['dtype'])
if dataset_name in ['mnist']:
image = trans_func(ori_img)
else:
cvt_image = cv2.cvtColor(ori_img, cv2.COLOR_BGR2RGB)
image = trans_func(cvt_image)
input_data = ts.array(image.tolist(), dtype=image.dtype.name)
res_msg = ''
if model_name == "cycle_gan":
g_model = servable_model['g_model']
if strategy == 'gray2color':
# build the network
G_generator, _ = net_func(g_model=g_model)
ckpt_name = 'G_A'
elif strategy == 'color2gray':
_, G_generator = net_func(g_model=g_model)
ckpt_name = 'G_B'
else:
err_msg = "Currently cycle_gan strategy only supports `gray2color` and `color2gray`!"
return {"status": 1, "err_msg": err_msg}
ckpt_path = os.path.join(serving_path, servable_name, ckpt_name + "." + model_format)
out_img = cyclegan_predict(G_generator, input_data, ckpt_path)
res_msg = '原图使用{}风格迁移效果'.format(strategy)
data = numpy2base64(out_img)
else:
# build the network
class_num = servable_model['class_num']
net = net_func(class_num=class_num, is_training=False)
serve_model = model.Model(net)
# load checkpoint
ckpt_path = os.path.join(serving_path, servable_name, model_name + "." + model_format)
if not os.path.isfile(ckpt_path):
err_msg = "The model path " + ckpt_path + " not exist!"
return {"status": 1, "err_msg": err_msg}
serve_model.load_checkpoint(ckpt_path)
# execute the network to perform model prediction
output = serve_model.predict(ts.expand_dims(input_data, 0))
if model_name == "ssd300":
output_np = (ts.concatenate((output[0], output[1]), axis=-1).asnumpy())
ih, iw, _ = instance['shape']
bbox_data = trans_func.postprocess(output_np, (ih, iw), strategy)
print(bbox_data)
bbox_num = len(bbox_data)
if not bbox_num:
err_msg = "抱歉！未检测到任何种类，无法标注。"
return {"status": 1, "err_msg": err_msg}
out_img = draw_boxes_in_image(bbox_data, ori_img)
max_det = max(bbox_data, key=lambda k: k['score'])
max_score = max_det['score']
category = bbox_data[bbox_data.index(max_det)]['category_id']
res_msg = '图中共标注了：{}个框，其中物种{}的得分最高, 为{}。'.format(bbox_num, category, round(max_score, 3))
data = numpy2base64(cv2.cvtColor(out_img, cv2.COLOR_BGR2RGB))
else:
output_np = output.asnumpy()
res_msg = trans_func.postprocess(output_np, strategy)
data = numpy2base64(ori_img)
res = {
"status": 0,
"instance": {
"res_msg": res_msg,
"data": data
}
}
return res
def predict(instance, servable_name, servable_model, strategy):
"""
Predict the result based on the input data.
A network will be constructed based on the input and servable data, then load the checkpoint and do the predict.
Args:
instance (dict): the dict of input image after transformation, with keys of `shape`, `dtype` and `data`(Image object).
servable_name (str): servable name
servable_model (str): name of the model
strategy (str): output strategy, usually select between `TOP1_CLASS` and `TOP5_CLASS`, for cyclegan, select between `gray2color` and `color2gray`
Returns:
The dict object of predicted result after post process.
Examples:
>>> # In the server part, after servable_search
>>> res = predict(instance, servable_name, servable['model'], strategy)
>>> return jsonify(res)
"""
# check if servable model name is valid
model_name = servable_model['name']
net_func = model_checker.get(model_name)
if net_func is None:
err_msg = "Currently model_name only supports " + str(list(model_checker.keys())) + "!"
return {"status": 1, "err_msg": err_msg}
# check if model_format is valid
model_format = servable_model['format']
if model_format not in ("ckpt"):
err_msg = "Currently model_format only supports `ckpt`!"
return {"status": 1, "err_msg": err_msg}
# parse the input data
input_data = ts.array(json.loads(instance['data']), dtype=instance['dtype'])
if model_name == "cycle_gan":
g_model = servable_model['g_model']
if strategy == 'gray2color':
# build the network
G_generator, _ = net_func(g_model=g_model)
ckpt_name = 'G_A'
elif strategy == 'color2gray':
_, G_generator = net_func(g_model=g_model)
ckpt_name = 'G_B'
else:
err_msg = "Currently cycle_gan strategy only supports `gray2color` and `color2gray`!"
return {"status": 1, "err_msg": err_msg}
ckpt_path = os.path.join(serving_path, servable_name, ckpt_name + "." + model_format)
data = cyclegan_predict(G_generator, input_data, ckpt_path)
else:
# build the network
class_num = servable_model['class_num']
net = net_func(class_num=class_num, is_training=False)
serve_model = model.Model(net)
# load checkpoint
ckpt_path = os.path.join(serving_path, servable_name, model_name + "." + model_format)
if not os.path.isfile(ckpt_path):
err_msg = "The model path " + ckpt_path + " not exist!"
return {"status": 1, "err_msg": err_msg}
serve_model.load_checkpoint(ckpt_path)
# execute the network to perform model prediction
output = serve_model.predict(ts.expand_dims(input_data, 0))
data = (ts.concatenate((output[0], output[1]), axis=-1).asnumpy() if model_name == "ssd300"
else output.asnumpy())
return {
"status": 0,
"instance": {
"shape": data.shape,
"dtype": data.dtype.name,
"data": json.dumps(data.tolist())
}
}
def predict(instance, servable_name, servable_model, strategy):
# check if servable model name is valid
model_name = servable_model['name']
net_func = model_checker.get(model_name)
if net_func is None:
err_msg = "Currently model_name only supports " + str(
list(model_checker.keys())) + "!"
return {"status": 1, "err_msg": err_msg}
# check if model_format is valid
model_format = servable_model['format']
if model_format not in ("ckpt"):
err_msg = "Currently model_format only supports `ckpt`!"
return {"status": 1, "err_msg": err_msg}
# parse the input data
input_data = ts.array(json.loads(instance['data']),
dtype=instance['dtype'])
if model_name == "cycle_gan":
g_model = servable_model['g_model']
if strategy == 'gray2color':
# build the network
G_generator, _ = net_func(g_model=g_model)
ckpt_name = 'G_A'
elif strategy == 'color2gray':
_, G_generator = net_func(g_model=g_model)
ckpt_name = 'G_B'
else:
err_msg = "Currently cycle_gan strategy only supports `gray2color` and `color2gray`!"
return {"status": 1, "err_msg": err_msg}
ckpt_path = os.path.join("/etc/tinyms/serving", servable_name,
ckpt_name + "." + model_format)
data = cyclegan_predict(G_generator, input_data, ckpt_path)
else:
# build the network
class_num = servable_model['class_num']
net = net_func(class_num=class_num)
serve_model = model.Model(net)
# load checkpoint
ckpt_path = os.path.join("/etc/tinyms/serving", servable_name,
model_name + "." + model_format)
if not os.path.isfile(ckpt_path):
err_msg = "The model path " + ckpt_path + " not exist!"
return {"status": 1, "err_msg": err_msg}
serve_model.load_checkpoint(ckpt_path)
# execute the network to perform model prediction
output = serve_model.predict(ts.expand_dims(input_data, 0))
data = (ts.concatenate((output[0], output[1]), axis=-1).asnumpy()
if model_name == "ssd300" else output.asnumpy())
return {
"status": 0,
"instance": {
"shape": data.shape,
"dtype": data.dtype.name,
"data": json.dumps(data.tolist())
}
}
def construct(self, from_tensor, to_tensor, attention_mask):
"""reshape 2d/3d input tensors to 2d"""
from_tensor_2d = self.reshape(from_tensor, self.shape_from_2d)
to_tensor_2d = self.reshape(to_tensor, self.shape_to_2d)
query_out = self.query_layer(from_tensor_2d)
key_out = self.key_layer(to_tensor_2d)
value_out = self.value_layer(to_tensor_2d)
query_layer = self.reshape(query_out, self.shape_from)
query_layer = self.transpose(query_layer, self.trans_shape)
key_layer = self.reshape(key_out, self.shape_to)
key_layer = self.transpose(key_layer, self.trans_shape)
attention_scores = self.matmul_trans_b(query_layer, key_layer)
# use_relative_position, supplementary logic
if self.use_relative_positions:
# relations_keys is [F|T, F|T, H]
relations_keys = self._generate_relative_positions_embeddings()
relations_keys = self.cast_compute_type(relations_keys)
# query_layer_t is [F, B, N, H]
query_layer_t = self.transpose(query_layer, self.trans_shape_relative)
# query_layer_r is [F, B * N, H]
query_layer_r = self.reshape(query_layer_t,
(self.from_seq_length,
-1,
self.size_per_head))
# key_position_scores is [F, B * N, F|T]
key_position_scores = self.matmul_trans_b(query_layer_r,
relations_keys)
# key_position_scores_r is [F, B, N, F|T]
key_position_scores_r = self.reshape(key_position_scores,
(self.from_seq_length,
-1,
self.num_attention_heads,
self.from_seq_length))
# key_position_scores_r_t is [B, N, F, F|T]
key_position_scores_r_t = self.transpose(key_position_scores_r,
self.trans_shape_position)
attention_scores = attention_scores + key_position_scores_r_t
attention_scores = self.multiply(self.scores_mul, attention_scores)
if self.has_attention_mask:
attention_mask = self.expand_dims(attention_mask, 1)
multiply_out = self.sub(self.cast(ts.array((1.0,)), self.get_dtype(attention_scores)),
self.cast(attention_mask, self.get_dtype(attention_scores)))
adder = self.multiply(multiply_out, self.multiply_data)
attention_scores = self.add(adder, attention_scores)
attention_probs = self.softmax(attention_scores)
attention_probs = self.dropout(attention_probs)
value_layer = self.reshape(value_out, self.shape_to)
value_layer = self.transpose(value_layer, self.trans_shape)
context_layer = self.matmul(attention_probs, value_layer)
# use_relative_position, supplementary logic
if self.use_relative_positions:
# relations_values is [F|T, F|T, H]
relations_values = self._generate_relative_positions_embeddings()
relations_values = self.cast_compute_type(relations_values)
# attention_probs_t is [F, B, N, T]
attention_probs_t = self.transpose(attention_probs, self.trans_shape_relative)
# attention_probs_r is [F, B * N, T]
attention_probs_r = self.reshape(
attention_probs_t,
(self.from_seq_length,
-1,
self.to_seq_length))
# value_position_scores is [F, B * N, H]
value_position_scores = self.matmul(attention_probs_r,
relations_values)
# value_position_scores_r is [F, B, N, H]
value_position_scores_r = self.reshape(value_position_scores,
(self.from_seq_length,
-1,
self.num_attention_heads,
self.size_per_head))
# value_position_scores_r_t is [B, N, F, H]
value_position_scores_r_t = self.transpose(value_position_scores_r,
self.trans_shape_position)
context_layer = context_layer + value_position_scores_r_t
context_layer = self.transpose(context_layer, self.trans_shape)
context_layer = self.reshape(context_layer, self.shape_return)
return context_layer