def preprocess_coco_for_ssdresnet34(data_dir,
preprocessed_data_dir,
formats,
overwrite=False):
def loader(file):
image = cv2.imread(file)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = np.array(
cv2.resize(image, (1200, 1200),
interpolation=cv2.INTER_LINEAR)).astype(np.float32)
image = image.transpose((2, 0, 1))
return image
def quantizer(image):
# Dynamic range of image is [-2.64064, 2.64064] based on calibration cache.
max_abs = 2.64064
image_int8 = image.clip(-max_abs, max_abs) / max_abs * 127.0
return image_int8.astype(dtype=np.int8, order='C')
preprocessor = ImagePreprocessor(loader, quantizer)
# Preprocess validation set.
preprocessor.run(
os.path.join(data_dir, "coco", "val2017"),
os.path.join(preprocessed_data_dir, "coco", "val2017", "SSDResNet34"),
"data_maps/coco/val_map.txt", formats, overwrite)
def preprocess_imagenet_for_resnet50(data_dir,
preprocessed_data_dir,
formats,
overwrite=False):
"""Proprocess the raw images for inference."""
def loader(file):
"""Resize and crop image to required dims and return as FP32 array."""
image = cv2.imread(file)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
w, h = (224, 224)
image = resize_with_aspectratio(image, h, w)
image = center_crop(image, h, w)
image = np.asarray(image, dtype='float32')
# Transpose.
image = image.transpose([2, 0, 1])
return image
def quantizer(image):
"""Return quantized INT8 image of input FP32 image."""
return np.clip(image, -128.0, 127.0).astype(dtype=np.int8, order='C')
preprocessor = ImagePreprocessor(loader, quantizer)
# Preprocess validation set.
preprocessor.run(
os.path.join(data_dir, "imagenet"),
os.path.join(preprocessed_data_dir, "imagenet", "ResNet50"),
"data_maps/imagenet/val_map.txt", formats, overwrite)
def preprocess_imagenet_for_resnet50(data_dir, preprocessed_data_dir, formats, overwrite=False, cal_only=False, val_only=False):
def loader(file):
img = Image.open(file)
img = img.convert('RGB')
img = resize_with_aspectratio(img, 256)
img = center_crop(img, (224, 224))
img = np.asarray(img, dtype='float32')
img /= 255.0
mean = np.array([0.485,0.456,0.406], dtype=np.float32)
std = np.array([0.229,0.224,0.225], dtype=np.float32)
img = (img - mean) / std
img = img.transpose([2, 0, 1])
#img = np.asarray(img.reshape((3,224,224)), dtype='float32')
return img
def quantizer(image):
max_abs = 2.64064
image_int8 = image.clip(-max_abs, max_abs) / max_abs * 127.0
return image_int8.astype(dtype=np.int8, order='C')
preprocessor = ImagePreprocessor(loader, quantizer)
if not val_only:
# Preprocess calibration set. FP32 only because calibrator always takes FP32 input.
preprocessor.run(os.path.join(data_dir, "imagenet"), os.path.join(preprocessed_data_dir, "imagenet", "ResNet50"),
"data_maps/imagenet/cal_map.txt", ["fp32"], overwrite)
if not cal_only:
# Preprocess validation set.
preprocessor.run(os.path.join(data_dir, "imagenet"), os.path.join(preprocessed_data_dir, "imagenet", "ResNet50"),
"data_maps/imagenet/val_map.txt", formats, overwrite)
def preprocess_coco_for_ssdresnet34(data_dir,
preprocessed_data_dir,
formats,
overwrite=False,
cal_only=False,
val_only=False):
def loader(file):
image = cv2.imread(file)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = np.array(
cv2.resize(image, (1200, 1200),
interpolation=cv2.INTER_LINEAR)).astype(np.float32)
image = image.transpose((2, 0, 1))
image /= 255.0
# Normalize image.
means = np.array([0.485, 0.456, 0.406],
dtype=np.float32)[:, np.newaxis, np.newaxis]
std = np.array([0.229, 0.224, 0.225], dtype=np.float32)[:, np.newaxis,
np.newaxis]
image = (image - means) / std
return image
def quantizer(image):
# Dynamic range of image is [-2.64064, 2.64064] based on calibration cache.
max_abs = 2.64064
image_int8 = image.clip(-max_abs, max_abs) / max_abs * 127.0
return image_int8.astype(dtype=np.int8, order='C')
preprocessor = ImagePreprocessor(loader, quantizer)
if not val_only:
# Preprocess calibration set. FP32 only because calibrator always takes FP32 input.
preprocessor.run(
os.path.join(data_dir, "coco", "train2017"),
os.path.join(preprocessed_data_dir, "coco", "train2017",
"SSDResNet34"), "data_maps/coco/cal_map.txt",
["fp32"], overwrite)
if not cal_only:
# Preprocess validation set.
preprocessor.run(
os.path.join(data_dir, "coco", "val2017"),
os.path.join(preprocessed_data_dir, "coco", "val2017",
"SSDResNet34"), "data_maps/coco/val_map.txt", formats,
overwrite)
def preprocess_imagenet_for_resnet50(data_dir,
preprocessed_data_dir,
formats,
overwrite=False,
cal_only=False,
val_only=False):
"""Proprocess the raw images for inference."""
def loader(file):
"""Resize and crop image to required dims and return as FP32 array."""
img = Image.open(file)
img = img.convert('RGB')
img = resize_with_aspectratio(img, 202)
img = center_crop(img, (176, 176))
img = np.asarray(img, dtype='float32')
img /= 255.0
mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
std = np.array([0.229, 0.224, 0.225], dtype=np.float32)
img = (img - mean) / std
img = img.transpose([2, 0, 1])
#img = np.asarray(img.reshape((3,224,224)), dtype='float32')
return img
def quantizer(image):
"""Return quantized INT8 image of input FP32 image."""
scale = struct.unpack('!f', bytes.fromhex('3caa5293'))[0]
image_int8 = (image / scale).clip(-127.0, 127.0)
return image_int8.astype(dtype=np.int8, order='C')
preprocessor = ImagePreprocessor(loader, quantizer)
if not val_only:
# Preprocess calibration set. FP32 only because calibrator always takes FP32 input.
preprocessor.run(
os.path.join(data_dir, "imagenet"),
os.path.join(preprocessed_data_dir, "imagenet", "ResNet50"),
"data_maps/imagenet/cal_map.txt", ["fp32"], overwrite)
if not cal_only:
# Preprocess validation set.
preprocessor.run(
os.path.join(data_dir, "imagenet"),
os.path.join(preprocessed_data_dir, "imagenet", "ResNet50"),
"data_maps/imagenet/val_map.txt", formats, overwrite)
def preprocess_imagenet_for_resnet50(data_dir,
preprocessed_data_dir,
formats,
overwrite=False,
cal_only=False,
val_only=False):
def loader(file):
image = cv2.imread(file)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
w, h = (224, 224)
image = resize_with_aspectratio(image, h, w)
image = center_crop(image, h, w)
image = np.asarray(image, dtype='float32')
# Normalize image.
means = np.array([123.68, 116.78, 103.94], dtype=np.float32)
image -= means
# Transpose.
image = image.transpose([2, 0, 1])
return image
def quantizer(image):
return np.clip(image, -128.0, 127.0).astype(dtype=np.int8, order='C')
preprocessor = ImagePreprocessor(loader, quantizer)
if not val_only:
# Preprocess calibration set. FP32 only because calibrator always takes FP32 input.
preprocessor.run(
os.path.join(data_dir, "imagenet"),
os.path.join(preprocessed_data_dir, "imagenet", "ResNet50"),
"data_maps/imagenet/cal_map.txt", ["fp32"], overwrite)
if not cal_only:
# Preprocess validation set.
preprocessor.run(
os.path.join(data_dir, "imagenet"),
os.path.join(preprocessed_data_dir, "imagenet", "ResNet50"),
"data_maps/imagenet/val_map.txt", formats, overwrite)
def preprocess_coco_for_ssdmobilenet(data_dir,
preprocessed_data_dir,
formats,
overwrite=False,
cal_only=False,
val_only=False):
def loader(file):
image = cv2.imread(file)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = np.array(
cv2.resize(image, (300, 300),
interpolation=cv2.INTER_LINEAR)).astype(np.float32)
image = image.transpose((2, 0, 1))
image = (2.0 / 255.0) * image - 1.0
return image
def quantizer(image):
# Dynamic range of image is [-1.0, 1.0]
image_int8 = image * 127.0
return image_int8.astype(dtype=np.int8, order='C')
preprocessor = ImagePreprocessor(loader, quantizer)
if not val_only:
# Preprocess calibration set. FP32 only because calibrator always takes FP32 input.
preprocessor.run(
os.path.join(data_dir, "coco", "train2017"),
os.path.join(preprocessed_data_dir, "coco", "train2017",
"SSDMobileNet"), "data_maps/coco/cal_map.txt",
["fp32"], overwrite)
if not cal_only:
# Preprocess validation set.
preprocessor.run(
os.path.join(data_dir, "coco", "val2017"),
os.path.join(preprocessed_data_dir, "coco", "val2017",
"SSDMobileNet"), "data_maps/coco/val_map.txt",
formats, overwrite)