def __next__(self):
if (b.isEmpty(self.loader._handle)):
timing_info = b.getTimingInfo(self.loader._handle)
print("Load time ::", timing_info.load_time)
print("Decode time ::", timing_info.decode_time)
print("Process time ::", timing_info.process_time)
print("Transfer time ::", timing_info.transfer_time)
raise StopIteration
if self.loader.run() != 0:
raise StopIteration
if (types.NCHW == self.tensor_format):
self.loader.copyToTensorNCHW(self.out, self.multiplier,
self.offset, self.reverse_channels,
int(self.tensor_dtype))
else:
self.loader.copyToTensorNHWC(self.out, self.multiplier,
self.offset, self.reverse_channels,
int(self.tensor_dtype))
self.loader.getImageLabels(self.labels)
self.labels_tensor = torch.from_numpy(self.labels).type(
torch.LongTensor)
if self.tensor_dtype == types.FLOAT:
return torch.from_numpy(self.out), self.labels_tensor
elif self.tensor_dtype == types.TensorDataType.FLOAT16:
return torch.from_numpy(self.out.astype(
np.float16)), self.labels_tensor
def __next__(self):
if(b.isEmpty(self.loader._handle)):
timing_info = b.getTimingInfo(self.loader._handle)
print("Load time ::",timing_info.load_time)
print("Decode time ::",timing_info.decode_time)
print("Process time ::",timing_info.process_time)
print("Transfer time ::",timing_info.transfer_time)
raise StopIteration
if self.loader.run() != 0:
raise StopIteration
if(types.NCHW == self.tensor_format):
self.loader.copyToTensorNCHW(self.out, self.multiplier, self.offset, self.reverse_channels, int(self.tensor_dtype))
else:
self.loader.copyToTensorNHWC(self.out, self.multiplier, self.offset, self.reverse_channels, int(self.tensor_dtype))
self.loader.getImageLabels(self.labels)
tf.reset_default_graph()
self.labels_tensor = tf.convert_to_tensor(self.labels,np.int32)
if self.tensor_dtype == types.FLOAT:
return tf.convert_to_tensor(self.out,np.float32), self.labels_tensor
elif self.tensor_dtype == types.FLOAT16:
return tf.convert_to_tensor(self.out,np.float16), self.labels_tensor
def __next__(self):
if (b.isEmpty(self.loader._handle)):
timing_info = b.getTimingInfo(self.loader._handle)
print("Load time ::", timing_info.load_time)
print("Decode time ::", timing_info.decode_time)
print("Process time ::", timing_info.process_time)
print("Transfer time ::", timing_info.transfer_time)
raise StopIteration
if self.loader.run() != 0:
raise StopIteration
if (types.NCHW == self.tensor_format):
self.loader.copyToTensorNCHW(self.out, self.multiplier,
self.offset, self.reverse_channels,
int(self.tensor_dtype))
else:
self.loader.copyToTensorNHWC(self.out, self.multiplier,
self.offset, self.reverse_channels,
int(self.tensor_dtype))
if ((self.loader._name == "Caffe2ReaderDetection")
or (self.loader._name == "CaffeReaderDetection")):
self.lis = [] # Empty list for bboxes
self.lis_lab = [] # Empty list of labels
#Count of labels/ bboxes in a batch
self.bboxes_label_count = np.zeros(self.bs, dtype="int32")
self.count_batch = self.loader.GetBoundingBoxCount(
self.bboxes_label_count)
# 1D labels array in a batch
self.labels = np.zeros(self.count_batch, dtype="int32")
self.loader.GetBBLabels(self.labels)
# 1D bboxes array in a batch
self.bboxes = np.zeros((self.count_batch * 4), dtype="float32")
self.loader.GetBBCords(self.bboxes)
#Image sizes of a batch
self.img_size = np.zeros((self.bs * 2), dtype="int32")
self.loader.GetImgSizes(self.img_size)
count = 0
sum_count = 0
for i in range(self.bs):
count = self.bboxes_label_count[i]
self.label_2d_numpy = (self.labels[sum_count:sum_count +
count])
self.label_2d_numpy = np.reshape(self.label_2d_numpy,
(-1, 1)).tolist()
self.bb_2d_numpy = (self.bboxes[sum_count *
4:(sum_count + count) * 4])
self.bb_2d_numpy = np.reshape(self.bb_2d_numpy,
(-1, 4)).tolist()
self.lis_lab.append(self.label_2d_numpy)
self.lis.append(self.bb_2d_numpy)
sum_count = sum_count + count
self.target = self.lis
self.target1 = self.lis_lab
max_cols = max(
[len(row) for batch in self.target for row in batch])
max_rows = max([len(batch) for batch in self.target])
self.bb_padded = [
batch + [[0] * (max_cols)] * (max_rows - len(batch))
for batch in self.target
]
self.bb_padded = torch.FloatTensor([
row + [0] * (max_cols - len(row)) for batch in self.bb_padded
for row in batch
])
self.bb_padded = self.bb_padded.view(-1, max_rows, max_cols)
max_cols1 = max(
[len(row) for batch in self.target1 for row in batch])
max_rows1 = max([len(batch) for batch in self.target1])
self.labels_padded = [
batch + [[0] * (max_cols1)] * (max_rows1 - len(batch))
for batch in self.target1
]
self.labels_padded = torch.LongTensor([
row + [0] * (max_cols1 - len(row))
for batch in self.labels_padded for row in batch
])
self.labels_padded = self.labels_padded.view(
-1, max_rows1, max_cols1)
if self.tensor_dtype == types.FLOAT:
return torch.from_numpy(
self.out), self.bb_padded, self.labels_padded
elif self.tensor_dtype == types.FLOAT16:
return torch.from_numpy(self.out.astype(
np.float16)), self.bb_padded, self.labels_padded
else:
if (self.loader._oneHotEncoding == True):
self.loader.GetOneHotEncodedLabels(self.labels)
self.labels_tensor = torch.from_numpy(self.labels).type(
torch.LongTensor)
self.labels_tensor = self.labels_tensor.view(
-1, self.bs, self.loader._numOfClasses)
else:
self.loader.getImageLabels(self.labels)
self.labels_tensor = torch.from_numpy(self.labels).type(
torch.LongTensor)
if self.tensor_dtype == types.FLOAT:
return torch.from_numpy(self.out), self.labels_tensor
elif self.tensor_dtype == types.FLOAT16:
return torch.from_numpy(self.out.astype(
np.float16)), self.labels_tensor
def __next__(self):
if (b.isEmpty(self.loader._handle)):
timing_info = b.getTimingInfo(self.loader._handle)
print("Load time ::", timing_info.load_time)
print("Decode time ::", timing_info.decode_time)
print("Process time ::", timing_info.process_time)
print("Transfer time ::", timing_info.transfer_time)
raise StopIteration
if self.loader.run() != 0:
raise StopIteration
if (types.NCHW == self.tensor_format):
self.loader.copyToTensorNCHW(self.out, self.multiplier,
self.offset, self.reverse_channels,
int(self.tensor_dtype))
else:
self.loader.copyToTensorNHWC(self.out, self.multiplier,
self.offset, self.reverse_channels,
int(self.tensor_dtype))
if (self.loader._name == "TFRecordReaderDetection"):
self.bbox_list = []
self.label_list = []
self.num_bboxes_list = []
#Count of labels/ bboxes in a batch
self.bboxes_label_count = np.zeros(self.bs, dtype="int32")
self.count_batch = self.loader.GetBoundingBoxCount(
self.bboxes_label_count)
self.num_bboxes_list = self.bboxes_label_count.tolist()
# 1D labels array in a batch
self.labels = np.zeros(self.count_batch, dtype="int32")
self.loader.GetBBLabels(self.labels)
# 1D bboxes array in a batch
self.bboxes = np.zeros((self.count_batch * 4), dtype="float32")
self.loader.GetBBCords(self.bboxes)
#1D Image sizes array of image in a batch
self.img_size = np.zeros((self.bs * 2), dtype="int32")
self.loader.GetImgSizes(self.img_size)
count = 0 # number of bboxes per image
sum_count = 0 # sum of the no. of the bboxes
for i in range(self.bs):
count = self.bboxes_label_count[i]
self.label_2d_numpy = (self.labels[sum_count:sum_count +
count])
self.label_2d_numpy = np.reshape(self.label_2d_numpy,
(-1, 1)).tolist()
self.bb_2d_numpy = (self.bboxes[sum_count *
4:(sum_count + count) * 4])
self.bb_2d_numpy = np.reshape(self.bb_2d_numpy,
(-1, 4)).tolist()
self.label_list.append(self.label_2d_numpy)
self.bbox_list.append(self.bb_2d_numpy)
sum_count = sum_count + count
self.target = self.bbox_list
self.target1 = self.label_list
max_cols = max(
[len(row) for batch in self.target for row in batch])
# max_rows = max([len(batch) for batch in self.target])
max_rows = 100
bb_padded = [
batch + [[0] * (max_cols)] * (max_rows - len(batch))
for batch in self.target
]
bb_padded_1 = [
row + [0] * (max_cols - len(row)) for batch in bb_padded
for row in batch
]
arr = np.asarray(bb_padded_1)
self.res = np.reshape(arr, (-1, max_rows, max_cols))
max_cols = max(
[len(row) for batch in self.target1 for row in batch])
# max_rows = max([len(batch) for batch in self.target1])
max_rows = 100
lab_padded = [
batch + [[0] * (max_cols)] * (max_rows - len(batch))
for batch in self.target1
]
lab_padded_1 = [
row + [0] * (max_cols - len(row)) for batch in lab_padded
for row in batch
]
labarr = np.asarray(lab_padded_1)
self.l = np.reshape(labarr, (-1, max_rows, max_cols))
self.num_bboxes_arr = np.array(self.num_bboxes_list)
if self.tensor_dtype == types.FLOAT:
return self.out.astype(
np.float32), self.res, self.l, self.num_bboxes_arr
elif self.tensor_dtype == types.FLOAT16:
return self.out.astype(
np.float16), self.res, self.l, self.num_bboxes_arr
elif (self.loader._name == "TFRecordReaderClassification"):
self.labels = np.zeros((self.bs), dtype="int32")
self.loader.getImageLabels(self.labels)
if self.tensor_dtype == types.FLOAT:
return self.out.astype(np.float32), self.labels
elif self.tensor_dtype == types.TensorDataType.FLOAT16:
return self.out.astype(np.float16), self.labels
def isEmpty(self):
return b.isEmpty(self._handle)
def __next__(self):
if (b.isEmpty(self.loader._handle)):
timing_info = b.getTimingInfo(self.loader._handle)
print("Load time ::", timing_info.load_time)
print("Decode time ::", timing_info.decode_time)
print("Process time ::", timing_info.process_time)
print("Transfer time ::", timing_info.transfer_time)
raise StopIteration
if self.loader.run() != 0:
raise StopIteration
if (types.NCHW == self.tensor_format):
self.loader.copyToTensorNCHW(self.out, self.multiplier,
self.offset, self.reverse_channels,
int(self.tensor_dtype))
else:
self.loader.copyToTensorNHWC(self.out, self.multiplier,
self.offset, self.reverse_channels,
int(self.tensor_dtype))
if ((self.loader._name == "Caffe2ReaderDetection")
or (self.loader._name == "CaffeReaderDetection")):
sum = 0
self.lis = [] #Empty list for bboxes
self.lis_lab = [] # Empty list of labels
for idx in range(self.bs):
sum = self.loader.GetBoundingBoxCount(idx)
self.labels = np.zeros(sum, dtype="int32")
self.bboxes = np.zeros(sum * 4, dtype="float32")
self.loader.GetBBLabels(self.labels, idx)
self.loader.GetBBCords(self.bboxes, idx)
self.bb_2d_numpy = np.reshape(self.bboxes, (-1, 4)).tolist()
self.label_2d_numpy = np.reshape(self.labels, (-1, 1)).tolist()
self.lis.append(self.bb_2d_numpy)
self.lis_lab.append(self.label_2d_numpy)
self.target = self.lis
self.target1 = self.lis_lab
max_cols = max(
[len(row) for batch in self.target for row in batch])
max_rows = max([len(batch) for batch in self.target])
self.bb_padded = [
batch + [[0] * (max_cols)] * (max_rows - len(batch))
for batch in self.target
]
self.bb_padded = torch.FloatTensor([
row + [0] * (max_cols - len(row)) for batch in self.bb_padded
for row in batch
])
self.bb_padded = self.bb_padded.view(-1, max_rows, max_cols)
max_cols1 = max(
[len(row) for batch in self.target1 for row in batch])
max_rows1 = max([len(batch) for batch in self.target1])
self.labels_padded = [
batch + [[0] * (max_cols1)] * (max_rows1 - len(batch))
for batch in self.target1
]
self.labels_padded = torch.LongTensor([
row + [0] * (max_cols1 - len(row))
for batch in self.labels_padded for row in batch
])
self.labels_padded = self.labels_padded.view(
-1, max_rows1, max_cols1)
if self.tensor_dtype == types.FLOAT:
return torch.from_numpy(
self.out), self.bb_padded, self.labels_padded
elif self.tensor_dtype == types.FLOAT16:
return torch.from_numpy(self.out.astype(
np.float16)), self.bb_padded, self.labels_padded
else:
self.loader.getImageLabels(self.labels)
self.labels_tensor = torch.from_numpy(self.labels).type(
torch.LongTensor)
if self.tensor_dtype == types.FLOAT:
return torch.from_numpy(self.out), self.labels_tensor
elif self.tensor_dtype == types.FLOAT16:
return torch.from_numpy(self.out.astype(
np.float16)), self.labels_tensor
def __next__(self):
if (b.isEmpty(self.loader._handle)):
timing_info = b.getTimingInfo(self.loader._handle)
print("Load time ::", timing_info.load_time)
print("Decode time ::", timing_info.decode_time)
print("Process time ::", timing_info.process_time)
print("Transfer time ::", timing_info.transfer_time)
raise StopIteration
if self.loader.run() != 0:
raise StopIteration
if (types.NCHW == self.tensor_format):
self.loader.copyToTensorNCHW(self.out, self.multiplier,
self.offset, self.reverse_channels,
int(self.tensor_dtype))
else:
self.loader.copyToTensorNHWC(self.out, self.multiplier,
self.offset, self.reverse_channels,
int(self.tensor_dtype))
if (self.loader._name == "TFRecordReaderDetection"):
sum = 0
self.lis = [] #Empty list for bboxes
self.lis_lab = [] # Empty list of labels
for idx in range(self.bs):
sum = self.loader.GetBoundingBoxCount(idx)
self.labels = np.zeros(sum, dtype="int32")
self.bboxes = np.zeros(sum * 4, dtype="float32")
self.loader.GetBBLabels(self.labels, idx)
self.loader.GetBBCords(self.bboxes, idx)
self.bb_2d_numpy = np.reshape(self.bboxes, (-1, 4)).tolist()
self.label_2d_numpy = np.reshape(self.labels, (-1, 1)).tolist()
self.lis.append(self.bb_2d_numpy)
self.lis_lab.append(self.label_2d_numpy[0])
self.target = self.lis
self.target1 = self.lis_lab
# tf.reset_default_graph()
max_cols = max(
[len(row) for batch in self.target for row in batch])
max_rows = max([len(batch) for batch in self.target])
bb_padded = [
batch + [[0] * (max_cols)] * (max_rows - len(batch))
for batch in self.target
]
bb_padded_1 = [
row + [0] * (max_cols - len(row)) for batch in bb_padded
for row in batch
]
# t=tf.convert_to_tensor(bb_padded_1)
# self.res=tf.reshape(t, [-1,max_rows, max_cols],name="bboxes")
arr = np.asarray(bb_padded_1)
self.res = np.reshape(arr, (-1, max_rows, max_cols))
# self.l = tf.convert_to_tensor(self.target1)
# self.labels_tensor = tf.reshape(self.l, [self.bs,-1],name="label")
self.l = np.asarray(self.target1)
self.l = np.reshape(self.l, (self.bs, -1))
if self.tensor_dtype == types.FLOAT:
# return tf.convert_to_tensor(self.out,np.float32), self.res,self.labels_tensor
return self.out.astype(np.float32), self.res, self.l
elif self.tensor_dtype == types.FLOAT16:
# return tf.convert_to_tensor(self.out,np.float16), self.res,self.labels_tensor
return self.out.astype(np.float16), self.res, self.l
elif (self.loader._name == "TFRecordReaderClassification"):
self.labels = np.zeros((self.bs), dtype="int32")
self.loader.getImageLabels(self.labels)
# tf.reset_default_graph()
# self.labels_tensor = tf.convert_to_tensor(self.labels,np.int32)
if self.tensor_dtype == types.FLOAT:
# return tf.convert_to_tensor(self.out,np.float32), self.labels_tensor
return self.out.astype(np.float32), self.labels
elif self.tensor_dtype == types.TensorDataType.FLOAT16:
# return tf.convert_to_tensor(self.out,np.float16), self.labels_tensor
return self.out.astype(np.float16), self.labels