[jacinto-ai/pytorch-jacinto-ai-devkit.git] / modules / pytorch_jacinto_ai / engine / engine_utils.py
diff --git a/modules/pytorch_jacinto_ai/engine/engine_utils.py b/modules/pytorch_jacinto_ai/engine/engine_utils.py
index e9b727df5a2adc76e912d02ecdc48ccf3219c4de..3d0f773f515ff2cf08f044de59ede003f889fdc4 100644 (file)
+# Copyright (c) 2018-2021, Texas Instruments
+# All Rights Reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of the copyright holder nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
import os
import numpy as np
import torch
def write_tensor_int(m=[], tensor=[], suffix='op', bitwidth=8, power2_scaling=True, file_format='bin',
- rnd_type='rnd_sym'):
+ rnd_type='rnd_sym', force_data_type=None,save_path=None):
mn = tensor.min()
mx = tensor.max()
'{:6}, {:32}, {:10}, {:7.2f}, {:7.2f}'.format(suffix, m.name, m.__class__.__name__, tensor.min(), tensor.max()),
end=" ")
- [tensor_scale, clamp_limits] = xnn.quantize.compute_tensor_scale(tensor, mn, mx, bitwidth, power2_scaling)
+ [tensor_scale, clamp_limits, tensor_signed] = xnn.utils.compute_tensor_scale(tensor, mn, mx, bitwidth, power2_scaling, force_data_type=force_data_type)
+ #tensor_signed = min(mn, mx) < 0
print("{:30} : {:15} : {:8.2f}".format(str(tensor.shape), str(tensor.dtype), tensor_scale), end=" ")
print_weight_bias = False
@@ -33,33 +62,41 @@ def write_tensor_int(m=[], tensor=[], suffix='op', bitwidth=8, power2_scaling=Tr
print("tensor_scale: ", tensor_scale)
print(tensor[no_idx])
if tensor.dtype != torch.int64:
- tensor = xnn.quantize.symmetric_round_tensor(tensor * tensor_scale)
+ tensor = xnn.utils.symmetric_round_tensor(tensor * tensor_scale)
if suffix == 'weight' and print_weight_bias:
print(tensor[no_idx])
else:
# for activation use HW friendly rounding
if tensor.dtype != torch.int64:
- tensor = xnn.quantize.upward_round_tensor(tensor * tensor_scale)
+ tensor = xnn.utils.upward_round_tensor(tensor * tensor_scale)
tensor = tensor.clamp(clamp_limits[0], clamp_limits[1]).float()
- if bitwidth == 8:
+ if bitwidth == 8 and tensor_signed:
data_type = np.int8
- elif bitwidth == 16:
+ str_data_type = 'int8'
+ elif bitwidth == 16 and tensor_signed:
data_type = np.int16
- elif bitwidth == 32:
+ str_data_type = 'int16'
+ elif bitwidth == 32 and tensor_signed:
data_type = np.int32
+ str_data_type = 'int32'
+ elif bitwidth == 8 and not tensor_signed:
+ data_type = np.uint8
+ str_data_type = 'uint8'
+ elif bitwidth == 16 and not tensor_signed:
+ data_type = np.uint16
+ str_data_type = 'uint16'
+ elif bitwidth == 32 and not tensor_signed:
+ data_type = np.uint32
+ str_data_type = 'uint32'
else:
exit("Bit width other 8,16,32 not supported for writing layer level op")
tensor = tensor.cpu().numpy().astype(data_type)
print("{:7} : {:7d} : {:7d}".format(str(tensor.dtype), tensor.min(), tensor.max()))
-
root = os.getcwd()
- tensor_dir = root + '/checkpoints/debug/test_vecs/' + '{}_{}_{}_scale_{:010.4f}'.format(m.name,
- m.__class__.__name__,
- suffix, tensor_scale)
-
+ tensor_dir = os.path.join(root, save_path, '{}_{}_{}_{}_scale_{:010.4f}'.format(m.name, m.__class__.__name__, suffix, str_data_type, tensor_scale))
if not os.path.exists(tensor_dir):
os.makedirs(tensor_dir)
@@ -71,9 +108,10 @@ def write_tensor_int(m=[], tensor=[], suffix='op', bitwidth=8, power2_scaling=Tr
np.save(tensor_name, tensor)
# utils_hist.comp_hist_tensor3d(x=tensor, name=m.name, en=True, dir=m.name, log=True, ch_dim=0)
+ return tensor_scale
-def write_tensor_float(m=[], tensor=[], suffix='op'):
+def write_tensor_float(m=[], tensor=[], suffix='op',save_path=None):
mn = tensor.min()
mx = tensor.max()
def write_tensor(data_type='int', m=[], tensor=[], suffix='op', bitwidth=8, power2_scaling=True, file_format='bin',
- rnd_type='rnd_sym'):
+ rnd_type='rnd_sym', force_data_type=None,save_path=None):
if data_type == 'int':
- write_tensor_int(m=m, tensor=tensor, suffix=suffix, rnd_type=rnd_type, file_format=file_format)
+ tensor_scale = write_tensor_int(m=m, tensor=tensor, suffix=suffix, rnd_type=rnd_type, bitwidth=bitwidth,
+ file_format=file_format,force_data_type=force_data_type,save_path=save_path)
elif data_type == 'float':
- write_tensor_float(m=m, tensor=tensor, suffix=suffix)
-
-
-enable_hook_function = True
-
+ write_tensor_float(m=m, tensor=tensor, suffix=suffix,save_path=save_path)
+ return tensor_scale
-def write_tensor_hook_function(m, inp, out, file_format='none'):
- if not enable_hook_function:
- return
+def write_tensor_hook_function(m, inp, out, save_path=None, file_format='bin'):
# Output
if isinstance(out, (torch.Tensor)):
- write_tensor(m=m, tensor=out, suffix='op', rnd_type='rnd_up', file_format=file_format)
+ tensor_scale_op = write_tensor(m=m, tensor=out, suffix='op', rnd_type='rnd_up', file_format=file_format, save_path=save_path)
# Input(s)
if type(inp) is tuple:
# if there are more than 1 inputs
for index, sub_ip in enumerate(inp[0]):
if isinstance(sub_ip, (torch.Tensor)):
- write_tensor(m=m, tensor=sub_ip, suffix='ip_{}'.format(index), rnd_type='rnd_up',
- file_format=file_format)
+ tensor_scale_ip = write_tensor(m=m, tensor=sub_ip, suffix='ip_{}'.format(index), rnd_type='rnd_up',
+ file_format=file_format, save_path=save_path)
elif isinstance(inp, (torch.Tensor)):
- write_tensor(m=m, tensor=inp, suffix='ip', rnd_type='rnd_up', file_format=file_format)
+ tensor_scale_ip = write_tensor(m=m, tensor=inp, suffix='ip', rnd_type='rnd_up', file_format=file_format, save_path=save_path)
# weights
if hasattr(m, 'weight'):
if isinstance(m.weight, torch.Tensor):
- write_tensor(m=m, tensor=m.weight, suffix='weight', rnd_type='rnd_sym', file_format=file_format)
+ tensor_scale_wt = write_tensor(m=m, tensor=m.weight, suffix='weight', rnd_type='rnd_sym', file_format=file_format, save_path=save_path)
# bias
if hasattr(m, 'bias'):
if m.bias is not None:
- write_tensor(m=m, tensor=m.bias, suffix='bias', rnd_type='rnd_sym', file_format=file_format)
+ tensor_scale_bias = write_tensor(m=m, tensor=m.bias, suffix='bias', rnd_type='rnd_sym', bitwidth=16,
+ force_data_type = 'signed', file_format=file_format, save_path=save_path)