def test_predictor_with_dtype():
prefix = 'test_predictor_simple_dense'
symbol_file = "%s-symbol.json" % prefix
param_file = "%s-0000.params" % prefix
input1 = np.random.uniform(size=(1, 3))
input1 = input1.astype(np.float16)
block = mx.gluon.nn.HybridSequential()
block.add(mx.gluon.nn.Dense(7))
block.add(mx.gluon.nn.Dense(3))
block.cast(np.float16)
block.hybridize()
block.initialize(ctx=mx.gpu(0))
tmp = mx.nd.array(input1, dtype=np.float16, ctx=mx.gpu(0))
out1 = block.forward(tmp)
block.export(prefix)
predictor = Predictor(open(symbol_file, "r").read(),
open(param_file, "rb").read(),
{"data": input1.shape},
dev_type="gpu",
dev_id=0,
type_dict={"data": input1.dtype})
predictor.forward(data=input1)
predictor_out1 = predictor.get_output(0)
assert_almost_equal(out1.asnumpy(), predictor_out1, rtol=1e-5, atol=1e-6)
def test_predictor():
prefix = 'test_predictor_simple_dense'
symbol_file = "%s-symbol.json" % prefix
param_file = "%s-0000.params" % prefix
# two inputs with different batch sizes
input1 = np.random.uniform(size=(1, 3))
input2 = np.random.uniform(size=(3, 3))
# define a simple model
block = gluon.nn.HybridSequential()
block.add(gluon.nn.Dense(7))
block.add(gluon.nn.Dense(3))
block.hybridize()
block.initialize()
out1 = block.forward(nd.array(input1))
out2 = block.forward(nd.array(input2))
block.export(prefix)
# create a predictor
predictor = Predictor(
open(symbol_file, "r").read(),
open(param_file, "rb").read(), {'data': input1.shape})
# forward and get output
predictor.forward(data=input1)
predictor_out1 = predictor.get_output(0)
assert_almost_equal(out1.asnumpy(), predictor_out1, rtol=1e-5, atol=1e-6)
# reshape
predictor.reshape({'data': input2.shape})
predictor.forward(data=input2)
predictor_out2 = predictor.get_output(0)
assert_almost_equal(out2.asnumpy(), predictor_out2, rtol=1e-5, atol=1e-6)
# destroy the predictor
del predictor
def compare_module_cpredict(result_sym,
result_arg_params,
result_aux_params,
monitor_callback=False):
# Dummmy inputs
input1 = np.ones((1, 3, 224, 224))
input1 = input1.astype(np.float32)
nd_dict = {}
def pred_mon_callback(name, arr):
nd_dict[name] = arr
mod = mx.mod.Module(result_sym,
data_names=["data"],
label_names=["softmax_label"],
context=mx.gpu())
mod.bind(data_shapes=[['data', (1, 3, 224, 224)]],
label_shapes=[['softmax_label', (1, )]],
for_training=False)
mod.set_params(result_arg_params, result_aux_params)
mod.forward(
mx.io.DataBatch(data=[mx.nd.array(input1, ctx=mx.gpu())],
label=[mx.nd.ones((1, ), ctx=mx.gpu())]))
prefix = "test_predictor_amp"
mod.save_checkpoint(prefix, 0, remove_amp_cast=False)
sym_file = "{}-symbol.json".format(prefix)
params_file = "{}-0000.params".format(prefix)
predictor = Predictor(open(sym_file, "r").read(),
open(params_file, "rb").read(), {
'data': (1, 3, 224, 224),
'softmax_label': (1, )
},
dev_type="gpu",
dev_id=0)
if monitor_callback:
predictor.set_monitor_callback(pred_mon_callback, monitor_all=True)
predictor.forward(data=input1, softmax_label=mx.nd.ones((1, )).asnumpy())
predictor_out1 = predictor.get_output(0)
if monitor_callback:
assert len(nd_dict) > 0, "Callback not called"
assert_almost_equal(mod.get_outputs()[0].asnumpy(),
predictor_out1,
atol=1e-1,
rtol=1e-1)
class lstm_ocr_model(object):
# Keep Zero index for blank. (CTC request it)
def __init__(self, path_of_json, path_of_params, classes, data_shape,
batch_size, num_label, num_hidden, num_lstm_layer):
super(lstm_ocr_model, self).__init__()
self.path_of_json = path_of_json
self.path_of_params = path_of_params
self.classes = classes
self.batch_size = batch_size
self.data_shape = data_shape
self.num_label = num_label
self.num_hidden = num_hidden
self.num_lstm_layer = num_lstm_layer
self.predictor = None
self.__init_ocr()
def __init_ocr(self):
init_c = [('l%d_init_c' % l, (self.batch_size, self.num_hidden))
for l in range(self.num_lstm_layer)]
init_h = [('l%d_init_h' % l, (self.batch_size, self.num_hidden))
for l in range(self.num_lstm_layer)]
init_states = init_c + init_h
all_shapes = [
('data',
(self.batch_size, self.data_shape[0] * self.data_shape[1]))
] + init_states + [('label', (self.batch_size, self.num_label))]
all_shapes_dict = {}
for _shape in all_shapes:
all_shapes_dict[_shape[0]] = _shape[1]
self.predictor = Predictor(
open(self.path_of_json).read(),
open(self.path_of_params).read(), all_shapes_dict)
def forward_ocr(self, img):
img = cv2.resize(img, self.data_shape)
img = img.transpose(1, 0)
img = img.reshape((self.data_shape[0] * self.data_shape[1]))
img = np.multiply(img, 1 / 255.0)
self.predictor.forward(data=img)
prob = self.predictor.get_output(0)
label_list = []
for p in prob:
max_index = np.argsort(p)[::-1][0]
label_list.append(max_index)
return self.__get_string(label_list)
def __get_string(self, label_list):
# Do CTC label rule
# CTC cannot emit a repeated symbol on consecutive timesteps
ret = []
label_list2 = [0] + list(label_list)
for i in range(len(label_list)):
c1 = label_list2[i]
c2 = label_list2[i + 1]
if c2 == 0 or c2 == c1:
continue
ret.append(c2)
# change to ascii
s = ''
for l in ret:
if l > 0 and l < (len(self.classes) + 1):
c = self.classes[l - 1]
else:
c = ''
s += c
return s
crop_img = img[yy:yy + short_egde, xx:xx + short_egde]
# resize to 224, 224
resized_img = transform.resize(crop_img, (224, 224))
if show_img:
io.imshow(resized_img)
# convert to numpy.ndarray
sample = np.asarray(resized_img) * 255
# swap axes to make image from (224, 224, 3) to (3, 224, 224)
sample = np.swapaxes(sample, 0, 2)
sample = np.swapaxes(sample, 1, 2)
# sub mean
normed_img = sample - mean_img
normed_img.resize(1, 3, 224, 224)
return normed_img
# Get preprocessed batch (single image batch)
batch = PreprocessImage('./download.png', True)
predictor.forward(data=batch)
prob = predictor.get_output(0)[0]
pred = np.argsort(prob)[::-1]
# Get top1 label
top1 = synset[pred[0]]
print("Top1: ", top1)
# Get top5 label
top5 = [synset[pred[i]] for i in range(5)]
print("Top5: ", top5)
class lstm_ocr_model(object):
"""LSTM network for predicting the Optical Character Recognition"""
# Keep Zero index for blank. (CTC request it)
CONST_CHAR = '0123456789'
def __init__(self, path_of_json, path_of_params):
super(lstm_ocr_model, self).__init__()
self.path_of_json = path_of_json
self.path_of_params = path_of_params
self.predictor = None
self.__init_ocr()
def __init_ocr(self):
num_label = 4 # Set your max length of label, add one more for blank
batch_size = 1
num_hidden = 100
num_lstm_layer = 2
init_c = [('l%d_init_c' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
init_h = [('l%d_init_h' % l, (batch_size, num_hidden))
for l in range(num_lstm_layer)]
init_states = init_c + init_h
init_state_arrays = np.zeros((batch_size, num_hidden), dtype="float32")
self.init_state_dict = {}
for x in init_states:
self.init_state_dict[x[0]] = init_state_arrays
all_shapes = [('data', (batch_size, 80, 30))
] + init_states + [('label', (batch_size, num_label))]
all_shapes_dict = {}
for _shape in all_shapes:
all_shapes_dict[_shape[0]] = _shape[1]
self.predictor = Predictor(
open(self.path_of_json, 'rb').read(),
open(self.path_of_params, 'rb').read(), all_shapes_dict)
def forward_ocr(self, img_):
"""Forward the image through the LSTM network model
Parameters
----------
img_: int of array
Returns
----------
label_list: string of list
"""
img_ = cv2.resize(img_, (80, 30))
img_ = img_.transpose(1, 0)
print(img_.shape)
img_ = img_.reshape((1, 80, 30))
print(img_.shape)
# img_ = img_.reshape((80 * 30))
img_ = np.multiply(img_, 1 / 255.0)
self.predictor.forward(data=img_, **self.init_state_dict)
prob = self.predictor.get_output(0)
label_list = []
for p in prob:
print(np.argsort(p))
max_index = np.argsort(p)[::-1][0]
label_list.append(max_index)
return self.__get_string(label_list)
@staticmethod
def __get_string(label_list):
# Do CTC label rule
# CTC cannot emit a repeated symbol on consecutive timesteps
ret = []
label_list2 = [0] + list(label_list)
for i in range(len(label_list)):
c1 = label_list2[i]
c2 = label_list2[i + 1]
if c2 in (0, c1):
continue
ret.append(c2)
# change to ascii
s = ''
for l in ret:
if l > 0 and l < (len(lstm_ocr_model.CONST_CHAR) + 1):
c = lstm_ocr_model.CONST_CHAR[l - 1]
else:
c = ''
s += c
return s
class lstm_ocr_model(object):
# Keep Zero index for blank. (CTC request it)
CONST_CHAR='0123456789'
def __init__(self, path_of_json, path_of_params):
super(lstm_ocr_model, self).__init__()
self.path_of_json = path_of_json
self.path_of_params = path_of_params
self.predictor = None
self.__init_ocr()
def __init_ocr(self):
num_label = 4 # Set your max length of label, add one more for blank
batch_size = 1
num_hidden = 100
num_lstm_layer = 2
init_c = [('l%d_init_c'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_h = [('l%d_init_h'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_states = init_c + init_h
all_shapes = [('data', (batch_size, 80 * 30))] + init_states + [('label', (batch_size, num_label))]
all_shapes_dict = {}
for _shape in all_shapes:
all_shapes_dict[_shape[0]] = _shape[1]
self.predictor = Predictor(open(self.path_of_json).read(),
open(self.path_of_params).read(),
all_shapes_dict)
def forward_ocr(self, img):
img = cv2.resize(img, (80, 30))
img = img.transpose(1, 0)
img = img.reshape((80 * 30))
img = np.multiply(img, 1/255.0)
self.predictor.forward(data=img)
prob = self.predictor.get_output(0)
label_list = []
for p in prob:
max_index = np.argsort(p)[::-1][0]
label_list.append(max_index)
return self.__get_string(label_list)
def __get_string(self, label_list):
# Do CTC label rule
# CTC cannot emit a repeated symbol on consecutive timesteps
ret = []
label_list2 = [0] + list(label_list)
for i in range(len(label_list)):
c1 = label_list2[i]
c2 = label_list2[i+1]
if c2 == 0 or c2 == c1:
continue
ret.append(c2)
# change to ascii
s = ''
for l in ret:
if l > 0 and l < (len(lstm_ocr_model.CONST_CHAR)+1):
c = lstm_ocr_model.CONST_CHAR[l-1]
else:
c = ''
s += c
return s
# we crop image from center
short_egde = min(img.shape[:2])
yy = int((img.shape[0] - short_egde) / 2)
xx = int((img.shape[1] - short_egde) / 2)
crop_img = img[yy : yy + short_egde, xx : xx + short_egde]
# resize to 224, 224
resized_img = transform.resize(crop_img, (224, 224))
# convert to numpy.ndarray
sample = np.asarray(resized_img) * 255
# swap axes to make image from (224, 224, 3) to (3, 224, 224)
sample = np.swapaxes(sample, 0, 2)
sample = np.swapaxes(sample, 1, 2)
# sub mean
return sample
# Get preprocessed batch (single image batch)
batch = PreprocessImage('./download.jpg', True)
predictor.forward(data=batch)
prob = predictor.get_output(0)[0]
pred = np.argsort(prob)[::-1]
# Get top1 label
top1 = synset[pred[0]]
print("Top1: ", top1)
# Get top5 label
top5 = [synset[pred[i]] for i in range(5)]
print("Top5: ", top5)
# swap axes to make image from (224, 224, 3) to (3, 224, 224)
sample = np.swapaxes(sample, 0, 2)
sample = np.swapaxes(sample, 1, 2)
# sub mean
normed_img = sample - mean_img
normed_img.resize(1, 3, 224, 224)
return normed_img
# Get preprocessed batch (single image batch)
img_path = './test/1.jpg'
print (img_path)
batch = PreprocessImage(img_path, True)
# gpu
predictor_gpu.forward(data=batch)
prob = predictor_gpu.get_output(0)[0]
pred = np.argsort(prob)[::-1]
# Get top1 label
top1 = synset[pred[0]]
print("Top1: ", top1, "Index: ", pred[0], "Prob: ", prob[pred[0]])
# Get top5 label
#top5 = [synset[pred[i]] for i in range(5)]
#print("Top5: ", top5)
# cpu
predictor_cpu.forward(data=batch)
prob = predictor_cpu.get_output(0)[0]
pred = np.argsort(prob)[::-1]
# Get top1 label
top1 = synset[pred[0]]
print("Top1: ", top1, "Index: ", pred[0], "Prob: ", prob[pred[0]])
class lstm_ocr_model(object):
"""LSTM network for predicting the Optical Character Recognition"""
# Keep Zero index for blank. (CTC request it)
CONST_CHAR = '0123456789'
def __init__(self, path_of_json, path_of_params):
super(lstm_ocr_model, self).__init__()
self.path_of_json = path_of_json
self.path_of_params = path_of_params
self.predictor = None
self.__init_ocr()
def __init_ocr(self):
num_label = 4 # Set your max length of label, add one more for blank
batch_size = 1
num_hidden = 100
num_lstm_layer = 2
init_c = [('l%d_init_c'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_h = [('l%d_init_h'%l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_states = init_c + init_h
init_state_arrays = np.zeros((batch_size, num_hidden), dtype="float32")
self.init_state_dict = {}
for x in init_states:
self.init_state_dict[x[0]] = init_state_arrays
all_shapes = [('data', (batch_size, 80, 30))] + init_states + [('label', (batch_size, num_label))]
all_shapes_dict = {}
for _shape in all_shapes:
all_shapes_dict[_shape[0]] = _shape[1]
self.predictor = Predictor(open(self.path_of_json, 'rb').read(),
open(self.path_of_params, 'rb').read(),
all_shapes_dict)
def forward_ocr(self, img_):
"""Forward the image through the LSTM network model
Parameters
----------
img_: int of array
Returns
----------
label_list: string of list
"""
img_ = cv2.resize(img_, (80, 30))
img_ = img_.transpose(1, 0)
print(img_.shape)
img_ = img_.reshape((1, 80, 30))
print(img_.shape)
# img_ = img_.reshape((80 * 30))
img_ = np.multiply(img_, 1 / 255.0)
self.predictor.forward(data=img_, **self.init_state_dict)
prob = self.predictor.get_output(0)
label_list = []
for p in prob:
print(np.argsort(p))
max_index = np.argsort(p)[::-1][0]
label_list.append(max_index)
return self.__get_string(label_list)
@staticmethod
def __get_string(label_list):
# Do CTC label rule
# CTC cannot emit a repeated symbol on consecutive timesteps
ret = []
label_list2 = [0] + list(label_list)
for i in range(len(label_list)):
c1 = label_list2[i]
c2 = label_list2[i+1]
if c2 in (0, c1):
continue
ret.append(c2)
# change to ascii
s = ''
for l in ret:
if l > 0 and l < (len(lstm_ocr_model.CONST_CHAR)+1):
c = lstm_ocr_model.CONST_CHAR[l-1]
else:
c = ''
s += c
return s
class lstm_ocr_model(object):
# Keep Zero index for blank. (CTC request it)
CONST_CHAR = '0123456789'
def __init__(self, path_of_json, path_of_params):
super(lstm_ocr_model, self).__init__()
self.path_of_json = path_of_json
self.path_of_params = path_of_params
self.predictor = None
self.__init_ocr()
def __init_ocr(self):
num_label = 4 # Set your max length of label, add one more for blank
batch_size = 1
num_hidden = 100
num_lstm_layer = 2
init_c = [('l%d_init_c' % l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_h = [('l%d_init_h' % l, (batch_size, num_hidden)) for l in range(num_lstm_layer)]
init_states = init_c + init_h
init_state_arrays = np.zeros((batch_size, num_hidden), dtype="float32")
self.init_state_dict = {}
for x in init_states:
self.init_state_dict[x[0]] = init_state_arrays
all_shapes = [('data', (batch_size, 80 * 30))] + init_states + [('label', (batch_size, num_label))]
all_shapes_dict = {}
for _shape in all_shapes:
all_shapes_dict[_shape[0]] = _shape[1]
self.predictor = Predictor(open(self.path_of_json).read(),
open(self.path_of_params).read(),
all_shapes_dict)
def forward_ocr(self, img):
img = cv2.resize(img, (80, 30))
img = img.transpose(1, 0)
img = img.reshape((80 * 30))
img = np.multiply(img, 1 / 255.0)
self.predictor.forward(data=img, **self.init_state_dict)
prob = self.predictor.get_output(0)
label_list = []
for p in prob:
max_index = np.argsort(p)[::-1][0]
label_list.append(max_index)
return self.__get_string(label_list)
def __get_string(self, label_list):
# Do CTC label rule
# CTC cannot emit a repeated symbol on consecutive timesteps
ret = []
label_list2 = [0] + list(label_list)
for i in range(len(label_list)):
c1 = label_list2[i]
c2 = label_list2[i + 1]
if c2 == 0 or c2 == c1:
continue
ret.append(c2)
# change to ascii
s = ''
for l in ret:
if l > 0 and l < (len(lstm_ocr_model.CONST_CHAR) + 1):
c = lstm_ocr_model.CONST_CHAR[l - 1]
else:
c = ''
s += c
return s
sample = np.swapaxes(sample, 0, 2)
sample = np.swapaxes(sample, 1, 2)
# sub mean
normed_img = sample - mean_img
normed_img.resize(1, 3, 224, 224)
return normed_img
# Get preprocessed batch (single image batch)
img_path = './test/1.jpg'
print(img_path)
batch = PreprocessImage(img_path, True)
# gpu
predictor_gpu.forward(data=batch)
prob = predictor_gpu.get_output(0)[0]
pred = np.argsort(prob)[::-1]
# Get top1 label
top1 = synset[pred[0]]
print("Top1: ", top1, "Index: ", pred[0], "Prob: ", prob[pred[0]])
# Get top5 label
#top5 = [synset[pred[i]] for i in range(5)]
#print("Top5: ", top5)
# cpu
predictor_cpu.forward(data=batch)
prob = predictor_cpu.get_output(0)[0]
pred = np.argsort(prob)[::-1]
# Get top1 label
top1 = synset[pred[0]]
print("Top1: ", top1, "Index: ", pred[0], "Prob: ", prob[pred[0]])
