openmv/ml/cmsisnn/nn_quantizer.py

639 lines
34 KiB
Python

# Copyright (C) 2018 Arm Limited or its affiliates. All rights reserved.
#
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the License); you may
# not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an AS IS BASIS, WITHOUT
# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# NN-Quantizer for Caffe models
import sys
# Include <Caffe installation path>/python in PYTHONPATH environment variable
import os
import caffe
from caffe.proto import caffe_pb2
import numpy as np
import argparse
from google.protobuf import text_format
import pickle
class Caffe_Quantizer(object):
"""\
Quantize a trained caffe model to 8-bits
"""
def __init__(self,model_file='',weight_file='',iterations=100,
accuracy_layer='accuracy',gpu=False):
self.model_file=model_file
self.weight_file=weight_file
self.quant_weight_file=""
self.conv_layer=[]
self.ip_layer=[]
self.start_layer=[]
self.end_layer=[]
self.layer=[]
self.layer_shape={}
self.layer_wt_shape={}
self.top_blob={}
self.bottom_blob={}
self.layer_type={}
self.kernel_size={}
self.stride={}
self.pad={}
self.group={}
self.pool_type={}
self.lrn_type={}
self.lrn_size={}
self.lrn_alpha={}
self.lrn_beta={}
self.num_ops={}
self.num_wts={}
self.wt_int_bits={}
self.wt_dec_bits={}
self.bias_int_bits={}
self.bias_dec_bits={}
self.act_int_bits={}
self.act_dec_bits={}
self.bias_lshift={}
self.act_rshift={}
self.data_layer=None
self.label_layer=None
self.accuracy_layer=accuracy_layer
self.iterations=iterations
self.gpu=gpu
def save_quant_params(self,model_info_file):
pickle.dump(self,open(model_info_file,'wb'))
def load_quant_params(self,model_info_file):
model_par=pickle.load(open(model_info_file,'rb'))
self.model_file=model_par.model_file
self.weight_file=model_par.weight_file
self.quant_weight_file=model_par.quant_weight_file
self.conv_layer=model_par.conv_layer
self.ip_layer=model_par.ip_layer
self.start_layer=model_par.start_layer
self.end_layer=model_par.end_layer
self.layer=model_par.layer
self.layer_shape=model_par.layer_shape
self.layer_wt_shape=model_par.layer_wt_shape
self.top_blob=model_par.top_blob
self.bottom_blob=model_par.bottom_blob
self.layer_type=model_par.layer_type
self.kernel_size=model_par.kernel_size
self.stride=model_par.stride
self.pad=model_par.pad
self.group=model_par.group
self.pool_type=model_par.pool_type
self.lrn_type=model_par.lrn_type
self.lrn_size=model_par.lrn_size
self.lrn_alpha=model_par.lrn_alpha
self.lrn_beta=model_par.lrn_beta
self.num_ops=model_par.num_ops
self.num_wts=model_par.num_wts
self.wt_int_bits=model_par.wt_int_bits
self.wt_dec_bits=model_par.wt_dec_bits
self.bias_int_bits=model_par.bias_int_bits
self.bias_dec_bits=model_par.bias_dec_bits
self.act_int_bits=model_par.act_int_bits
self.act_dec_bits=model_par.act_dec_bits
self.bias_lshift=model_par.bias_lshift
self.act_rshift=model_par.act_rshift
self.data_layer=model_par.data_layer
self.label_layer=model_par.label_layer
self.accuracy_layer=model_par.accuracy_layer
self.iterations=model_par.iterations
self.gpu=model_par.gpu
def run_full_network(self):
if self.gpu==True:
caffe.set_mode_gpu()
net = caffe.Net(self.model_file,self.weight_file,caffe.TEST)
acc = np.zeros(self.iterations)
for i in range(0,self.iterations):
out = net.forward()
acc[i] = out[self.accuracy_layer]*100
print("Full precision accuracy: %.2f%%" %(acc.mean()))
return acc.mean()
def run_quantized_network(self):
if self.gpu==True:
caffe.set_mode_gpu()
net = caffe.Net(self.model_file,self.quant_weight_file,caffe.TEST)
acc = np.zeros(self.iterations)
for i in range(0,self.iterations):
out = net.forward()
acc[i] = out[self.accuracy_layer]*100
print("Accuracy with quantized weights/biases: %.2f%%" %(acc.mean()))
for i in range(0,self.iterations):
for layer_no in range(0,len(self.start_layer)):
if layer_no==0:
net.forward(end=str(self.end_layer[layer_no]))
else:
net.forward(start=str(self.start_layer[layer_no]),end=str(self.end_layer[layer_no]))
if layer_no < len(self.start_layer)-1: # not quantizing accuracy layer
net.blobs[self.end_layer[layer_no]].data[:]=np.floor(net.blobs[self.end_layer[layer_no]].data*\
(2**self.act_dec_bits[self.end_layer[layer_no]]))
net.blobs[self.end_layer[layer_no]].data[net.blobs[self.end_layer[layer_no]].data>126]=127
net.blobs[self.end_layer[layer_no]].data[net.blobs[self.end_layer[layer_no]].data<-127]=-128
net.blobs[self.end_layer[layer_no]].data[:]=net.blobs[self.end_layer[layer_no]].data/\
(2**self.act_dec_bits[self.end_layer[layer_no]])
acc[i] = net.blobs[self.accuracy_layer].data*100
accuracy = acc.mean()
print("Accuracy with quantized weights/biases and activations: %.2f%%" %(accuracy))
return accuracy
def get_layer_info(self):
net=caffe_pb2.NetParameter()
text_format.Merge(open(self.model_file,'r').read(),net)
cnn = caffe.Net(self.model_file,self.weight_file,caffe.TEST)
if len(net.layer)==0: #some prototxts use "layer", some use "layers"
layers = net.layers
else:
layers = net.layer
for layer in layers:
layer_name=[]
for val in layer.top:
layer_name+=[str(val)]
self.top_blob[str(layer.name)]=layer_name
self.layer_shape[str(layer.name)]=cnn.blobs[self.top_blob[str(layer.name)][0]].data.shape
layer_name=[]
for val in layer.bottom:
layer_name+=[str(val)]
self.bottom_blob[str(layer.name)]=layer_name
self.layer_type[str(layer.name)] = str(layer.type).lower()
if str(layer.type).lower() == 'convolution' or str(layer.type)=='4':
self.layer_wt_shape[str(layer.name)]=cnn.params[self.top_blob[str(layer.name)][0]][0].data.shape
self.conv_layer.append(str(layer.name))
self.kernel_size[str(layer.name)] = layer.convolution_param.kernel_size[0]
self.stride[str(layer.name)] = 1;
self.pad[str(layer.name)] = 0;
if(len(layer.convolution_param.stride)!=0):
self.stride[str(layer.name)] = layer.convolution_param.stride[0]
if(len(layer.convolution_param.pad)!=0):
self.pad[str(layer.name)] = layer.convolution_param.pad[0]
self.group[str(layer.name)] = layer.convolution_param.group
elif str(layer.type).lower() == 'pooling' or str(layer.type)=='17':
self.pool_type[str(layer.name)] = layer.pooling_param.pool
self.kernel_size[str(layer.name)] = layer.pooling_param.kernel_size
self.stride[str(layer.name)] = layer.pooling_param.stride
self.pad[str(layer.name)] = layer.pooling_param.pad
elif str(layer.type).lower() == 'lrn' or str(layer.type)=='15':
self.lrn_type[str(layer.name)] = layer.lrn_param.norm_region
self.lrn_size[str(layer.name)] = layer.lrn_param.local_size
self.lrn_alpha[str(layer.name)] = layer.lrn_param.alpha
self.lrn_beta[str(layer.name)] = layer.lrn_param.beta
elif str(layer.type).lower() == 'innerproduct' or str(layer.type)=='14':
self.layer_wt_shape[str(layer.name)]=cnn.params[self.top_blob[str(layer.name)][0]][0].data.shape
self.ip_layer.append(str(layer.name))
elif str(layer.type).lower() == 'data' or str(layer.type)=='5':
included = False
for layer_phase in layer.include:
included = included or layer_phase.phase == caffe.TEST
if(included == True):
batch_size = layer.data_param.batch_size
self.data_layer = str(layer.top[0])
self.label_layer = str(layer.top[1])
def get_graph_connectivity(self):
# Extract network connectivity for running CNN functions in the correct order
# Traversing back from output layer (accuracy) to input layer (data) especially because
# googLeNet has many accuracy labels, which branch out and end at a different accuracy
# label with forward traversal
net=caffe_pb2.NetParameter()
text_format.Merge(open(self.model_file,'r').read(),net)
allowed_layer_types = ['data','convolution','innerproduct','pooling','lrn','relu',\
'accuracy','concat','5','4','14','17','15','18','1','3']
current_layer = self.accuracy_layer
traversed=[]
while current_layer != str(self.data_layer):
traversed += [current_layer]
num_branch = len(self.bottom_blob[current_layer])
current_blob = self.bottom_blob[current_layer][0]
has_unused_relu = 0
for key, value in self.top_blob.iteritems():
if (current_blob in value) and (key not in traversed) and \
(self.layer_type[key] == 'relu' or self.layer_type[key]=='18'):
has_unused_relu = 1
break
for key, value in self.top_blob.iteritems():
if(has_unused_relu == 1):
if (current_blob in value) and (key not in traversed) and \
(self.layer_type[key]=='relu' or self.layer_type[key]=='18'):
has_unused_relu = 0
current_layer = key
break
else:
if (current_blob in value) and (key not in traversed) and \
(self.layer_type[key] in allowed_layer_types):
current_layer = key
break
traversed += [current_layer]
traversed.reverse()
self.layer=traversed[:]
self.start_layer+=['']
for layer_no in range(0,len(self.layer)):
layer = self.layer[layer_no]
if layer == self.data_layer or layer in self.conv_layer or \
layer in self.ip_layer or layer in self.accuracy_layer or\
((self.layer_type[layer]=='pooling' or self.layer_type[layer]=='17') \
and self.pool_type[layer]==1):
self.end_layer+=[layer]
if layer_no < len(self.layer)-1:
self.start_layer+=[self.layer[layer_no+1]]
print(self.start_layer)
print(self.end_layer)
# Quantize weights to 8 bits
# Using min and max of weights as nearest power of 2, quantize to 8bits (QM.N) and check accuracy
# If accuracy is lost, try QM-1:N+1, QM-2,N+2,... with saturation to find out the best combination
# with least accuracy loss (Trading-off weights that occur infrequently for more precision)
#
# -2^(M+N) 0 2^(M+N)
# | ^ |
# | *|||* |
# <--------| *|||||* |------->
# Saturated| *|||||||* |Saturated
# | *|||||||||||* |
# | *|||||||||||||||||* |
# *| |*
# * |<-------------------------->| *
# Weight quantization and
# truncation with minimal
# loss of accuracy
#
def quantize_wts_8bit(self,tolerance=0.001,search_range=3):
if self.gpu==True:
caffe.set_mode_gpu()
net = caffe.Net(self.model_file,self.weight_file,caffe.TEST)
acc = np.zeros(self.iterations)
for i in range(0,self.iterations):
out = net.forward()
acc[i] = out[self.accuracy_layer]*100
target_accuracy = acc.mean()
print("Full precision accuracy: %.2f%%" %(target_accuracy))
self.quant_weight_file = self.weight_file
wfile = os.path.basename(self.weight_file)
qwfile = 'quantized_'+wfile
self.quant_weight_file = self.weight_file.replace(wfile,qwfile)
self.quant_weight_file = self.quant_weight_file.replace('.h5','')
net.save(self.quant_weight_file)
for layer_name in self.conv_layer+self.ip_layer:
#Start with min/max of weights to the rounded up to nearest power of 2.
wt_max = net.params[layer_name][0].data.max()
wt_min = net.params[layer_name][0].data.min()
self.wt_int_bits[layer_name] = int(np.ceil(np.log2(max(abs(wt_min),abs(wt_max)))))
self.wt_dec_bits[layer_name] = 7-self.wt_int_bits[layer_name]
max_int_bits = self.wt_int_bits[layer_name]-search_range
print('Layer: '+ layer_name + ' weights max: '+str(wt_max)+' min: '+str(wt_min)+\
' Format: Q'+str(self.wt_int_bits[layer_name])+'.'+str(self.wt_dec_bits[layer_name]))
net.params[layer_name][0].data[:]=np.round(net.params[layer_name][0].data*\
(2**self.wt_dec_bits[layer_name]))/(2**self.wt_dec_bits[layer_name])
for i in range(0,self.iterations):
out = net.forward()
acc[i] = out[self.accuracy_layer]*100
accuracy = acc.mean()
print("Accuracy: %.2f%%" %(accuracy))
best_int_bits = self.wt_int_bits[layer_name]
best_dec_bits = self.wt_dec_bits[layer_name]
best_accuracy = accuracy
while target_accuracy-accuracy>tolerance and self.wt_int_bits[layer_name]>max_int_bits:
self.wt_int_bits[layer_name] = self.wt_int_bits[layer_name]-1
self.wt_dec_bits[layer_name] = self.wt_dec_bits[layer_name]+1
net.copy_from(self.quant_weight_file)
net.params[layer_name][0].data[:]=np.round(net.params[layer_name][0].data*\
(2**self.wt_dec_bits[layer_name]))
net.params[layer_name][0].data[net.params[layer_name][0].data>126]=127
net.params[layer_name][0].data[net.params[layer_name][0].data<-127]=-128
net.params[layer_name][0].data[:]=net.params[layer_name][0].data/\
(2**self.wt_dec_bits[layer_name])
for i in range(0,self.iterations):
out = net.forward()
acc[i] = out[self.accuracy_layer]*100
accuracy = acc.mean()
print('Format Q'+str(self.wt_int_bits[layer_name])+'.'+\
str(self.wt_dec_bits[layer_name])+' Accuracy: %.2f%%' %(accuracy))
if accuracy>best_accuracy:
best_int_bits = self.wt_int_bits[layer_name]
best_dec_bits = self.wt_dec_bits[layer_name]
best_accuracy = accuracy
self.wt_int_bits[layer_name] = best_int_bits
self.wt_dec_bits[layer_name] = best_dec_bits
net.copy_from(self.quant_weight_file)
net.params[layer_name][0].data[:]=np.round(net.params[layer_name][0].data*\
(2**self.wt_dec_bits[layer_name]))
net.params[layer_name][0].data[net.params[layer_name][0].data>126]=127
net.params[layer_name][0].data[net.params[layer_name][0].data<-127]=-128
net.params[layer_name][0].data[:]=net.params[layer_name][0].data/\
(2**self.wt_dec_bits[layer_name])
print('Final '+layer_name+ ' weights format Q'+str(best_int_bits)+'.'+\
str(best_dec_bits)+' Accuracy: %.2f%%' %(best_accuracy))
net.save(self.quant_weight_file)
# Quantize activations (inter-layer data) to 8 bits
# Using min and max of activations as nearest power of 2, quantize to 8bits (QM.N) and check accuracy
# If accuracy is lost, try QM-1:N+1, QM-2,N+2,... with saturation to find out the best combination
# with least accuracy loss (Trading-off activations that occur infrequently for more precision)
def quantize_activations_8bit(self,tolerance=0.001,search_range=3):
if self.gpu==True:
caffe.set_mode_gpu()
net = caffe.Net(self.model_file,self.quant_weight_file,caffe.TEST)
acc = np.zeros(self.iterations)
for i in range(0,self.iterations):
out = net.forward()
acc[i] = out[self.accuracy_layer]*100
target_accuracy = acc.mean()
print("Accuracy with quantized weights: %.2f%%" %(target_accuracy))
max_val={}
min_val={}
quant_layer_flag={}
for layer in self.end_layer:
max_val[layer]=float('-inf')
min_val[layer]=float('inf')
quant_layer_flag[layer]=0
#Finding min max for output of all layers
for i in range(0,self.iterations):
for layer_no in range(0,len(self.start_layer)):
if layer_no==0:
net.forward(end=str(self.end_layer[layer_no]))
else:
net.forward(start=str(self.start_layer[layer_no]),end=str(self.end_layer[layer_no]))
layer_max = net.blobs[self.end_layer[layer_no]].data.max()
layer_min = net.blobs[self.end_layer[layer_no]].data.min()
if(layer_max>max_val[self.end_layer[layer_no]]):
max_val[self.end_layer[layer_no]]=layer_max
if(layer_min<min_val[self.end_layer[layer_no]]):
min_val[self.end_layer[layer_no]]=layer_min
#print("Running %s layer, max,min : %.2f,%.2f" %(self.end_layer[layer_no],layer_max,layer_min))
max_int_bits={}
for layer in self.end_layer:
self.act_int_bits[layer] = int(np.ceil(np.log2(max(abs(max_val[layer]),abs(min_val[layer])))))
self.act_dec_bits[layer] = 7-self.act_int_bits[layer]
max_int_bits[layer]=self.act_int_bits[layer]-search_range
print('Layer: '+layer+' max: '+ str(max_val[layer]) + ' min: '+str(min_val[layer])+ \
' Format: Q'+str(self.act_int_bits[layer])+'.'+str(self.act_dec_bits[layer]))
quant_max_val={}
quant_min_val={}
for layer in self.end_layer:
quant_max_val[layer]=float('-inf')
quant_min_val[layer]=float('inf')
for quant_layer_no in range(0,len(self.start_layer)-1): #No need to quantize accuracy layer
quant_layer=self.end_layer[quant_layer_no]
quant_layer_flag[quant_layer]=1
if((self.layer_type[quant_layer]=='pooling' or self.layer_type[quant_layer]=='17') and \
self.pool_type[quant_layer]==1):
prev_layer=self.end_layer[quant_layer_no-1]
self.act_int_bits[quant_layer]=self.act_int_bits[prev_layer]
self.act_dec_bits[quant_layer]=self.act_dec_bits[prev_layer]
continue
# quantize layer by layer
for i in range(0,self.iterations):
for layer_no in range(0,len(self.start_layer)):
if layer_no==0:
net.forward(end=str(self.end_layer[layer_no]))
else:
net.forward(start=str(self.start_layer[layer_no]),end=str(self.end_layer[layer_no]))
if quant_layer_flag[self.end_layer[layer_no]]==1: # quantize incrementally layer by layer
net.blobs[self.end_layer[layer_no]].data[:]=np.floor(net.blobs[self.end_layer[layer_no]].data*\
(2**self.act_dec_bits[self.end_layer[layer_no]]))/(2**self.act_dec_bits[self.end_layer[layer_no]])
layer_max = net.blobs[self.end_layer[layer_no]].data.max()
layer_min = net.blobs[self.end_layer[layer_no]].data.min()
if(layer_max>quant_max_val[self.end_layer[layer_no]]):
quant_max_val[self.end_layer[layer_no]]=layer_max
if(layer_min<quant_min_val[self.end_layer[layer_no]]):
quant_min_val[self.end_layer[layer_no]]=layer_min
acc[i] = net.blobs[self.accuracy_layer].data*100
accuracy=acc.mean()
print('Layer-'+quant_layer+' max: '+str(quant_max_val[quant_layer])+\
' min: '+str(quant_min_val[quant_layer])+' format: Q'+\
str(self.act_int_bits[quant_layer])+'.'+str(self.act_dec_bits[quant_layer])+\
' accuracy: %.2f%%' %(acc.mean()))
best_accuracy = accuracy
best_int_bits = self.act_int_bits[quant_layer]
best_dec_bits = self.act_dec_bits[quant_layer]
while target_accuracy-accuracy>tolerance and self.act_int_bits[quant_layer]>\
max_int_bits[quant_layer]:
for layer in self.end_layer:
quant_max_val[layer]=float('-inf')
quant_min_val[layer]=float('inf')
self.act_int_bits[quant_layer] = self.act_int_bits[quant_layer]-1
self.act_dec_bits[quant_layer] = self.act_dec_bits[quant_layer]+1
for i in range(0,self.iterations):
for layer_no in range(0,len(self.start_layer)):
if layer_no==0:
net.forward(end=str(self.end_layer[layer_no]))
else:
net.forward(start=str(self.start_layer[layer_no]),end=str(self.end_layer[layer_no]))
if quant_layer_flag[self.end_layer[layer_no]]==1:
net.blobs[self.end_layer[layer_no]].data[:]=np.floor(net.blobs[self.end_layer[layer_no]].data*\
(2**self.act_dec_bits[self.end_layer[layer_no]]))
net.blobs[self.end_layer[layer_no]].data[net.blobs[self.end_layer[layer_no]].data>126]=127
net.blobs[self.end_layer[layer_no]].data[net.blobs[self.end_layer[layer_no]].data<-127]=-128
net.blobs[self.end_layer[layer_no]].data[:]=net.blobs[self.end_layer[layer_no]].data/\
(2**self.act_dec_bits[self.end_layer[layer_no]])
layer_max = net.blobs[self.end_layer[layer_no]].data.max()
layer_min = net.blobs[self.end_layer[layer_no]].data.min()
if(layer_max>quant_max_val[self.end_layer[layer_no]]):
quant_max_val[self.end_layer[layer_no]]=layer_max
if(layer_min<quant_min_val[self.end_layer[layer_no]]):
quant_min_val[self.end_layer[layer_no]]=layer_min
acc[i] = net.blobs[self.accuracy_layer].data*100
accuracy=acc.mean()
if accuracy>best_accuracy:
best_int_bits = self.act_int_bits[quant_layer]
best_dec_bits = self.act_dec_bits[quant_layer]
best_accuracy = accuracy
print('Layer-'+quant_layer+' max: '+str(quant_max_val[quant_layer])+\
'min: '+str(quant_min_val[quant_layer])+' format: Q'+\
str(self.act_int_bits[quant_layer])+'.'+str(self.act_dec_bits[quant_layer])+\
' accuracy: %.2f%%' %(acc.mean()))
self.act_int_bits[quant_layer] = best_int_bits
self.act_dec_bits[quant_layer] = best_dec_bits
print('Layer-'+quant_layer+' final format: Q'+str(self.act_int_bits[quant_layer])+\
'.'+str(self.act_dec_bits[quant_layer])+ ' accuracy: %.2f%%' %(best_accuracy))
def quantize_bias_8bit(self,tolerance=0.001,search_range=3):
if self.gpu==True:
caffe.set_mode_gpu()
net = caffe.Net(self.model_file,self.quant_weight_file,caffe.TEST)
acc = np.zeros(self.iterations)
for i in range(0,self.iterations):
net.forward()
acc[i] = net.blobs[self.accuracy_layer].data*100
target_accuracy = acc.mean()
print("Accuracy with quantized weights: %.2f%%" %(target_accuracy))
for i in range(0,self.iterations):
for layer_no in range(0,len(self.start_layer)):
if layer_no==0:
net.forward(end=str(self.end_layer[layer_no]))
else:
net.forward(start=str(self.start_layer[layer_no]),end=str(self.end_layer[layer_no]))
if layer_no < len(self.start_layer)-1: # not quantizing accuracy layer
net.blobs[self.end_layer[layer_no]].data[:]=np.floor(net.blobs[self.end_layer[layer_no]].data*\
(2**self.act_dec_bits[self.end_layer[layer_no]]))
net.blobs[self.end_layer[layer_no]].data[net.blobs[self.end_layer[layer_no]].data>126]=127
net.blobs[self.end_layer[layer_no]].data[net.blobs[self.end_layer[layer_no]].data<-127]=-128
net.blobs[self.end_layer[layer_no]].data[:]=net.blobs[self.end_layer[layer_no]].data/\
(2**self.act_dec_bits[self.end_layer[layer_no]])
acc[i] = net.blobs[self.accuracy_layer].data*100
target_accuracy = acc.mean()
print("Accuracy with quantized weights and activations: %.2f%%" %(target_accuracy))
input_of={}
for i in range (1,len(self.end_layer)):
input_of[self.end_layer[i]]=self.end_layer[i-1]
for layer_name in self.conv_layer+self.ip_layer:
mac_dec_bits = self.wt_dec_bits[layer_name]+self.act_dec_bits[input_of[layer_name]]
bias_max = net.params[layer_name][1].data.max()
bias_min = net.params[layer_name][1].data.min()
int_bits = int(np.ceil(np.log2(max(abs(bias_min),abs(bias_max)))))
dec_bits = 7-int_bits
max_int_bits = int_bits-search_range
if(dec_bits>mac_dec_bits):
dec_bits=mac_dec_bits
int_bits=7-dec_bits
max_int_bits=int_bits #can't increase dec_bits any more as they will be shifted right anyway
print('Layer: '+ layer_name + ' biases max: '+str(bias_max)+' min: '+str(bias_min)+\
' Format: Q'+str(int_bits)+'.'+str(dec_bits))
net.params[layer_name][1].data[:]=np.round(net.params[layer_name][1].data*(2**dec_bits))/(2**dec_bits)
for i in range(0,self.iterations):
for layer_no in range(0,len(self.start_layer)):
if layer_no==0:
net.forward(end=str(self.end_layer[layer_no]))
else:
net.forward(start=str(self.start_layer[layer_no]),end=str(self.end_layer[layer_no]))
if layer_no < len(self.start_layer)-1: # not quantizing accuracy layer
net.blobs[self.end_layer[layer_no]].data[:]=np.floor(net.blobs[self.end_layer[layer_no]].data*\
(2**self.act_dec_bits[self.end_layer[layer_no]]))
net.blobs[self.end_layer[layer_no]].data[net.blobs[self.end_layer[layer_no]].data>126]=127
net.blobs[self.end_layer[layer_no]].data[net.blobs[self.end_layer[layer_no]].data<-127]=-128
net.blobs[self.end_layer[layer_no]].data[:]=net.blobs[self.end_layer[layer_no]].data/\
(2**self.act_dec_bits[self.end_layer[layer_no]])
acc[i] = net.blobs[self.accuracy_layer].data*100
accuracy = acc.mean()
print("Accuracy: %.2f%%" %(accuracy))
best_int_bits = int_bits
best_dec_bits = dec_bits
best_accuracy = accuracy
while target_accuracy-accuracy>tolerance and int_bits>max_int_bits:
int_bits = int_bits-1
dec_bits = dec_bits+1
net.copy_from(self.quant_weight_file)
net.params[layer_name][1].data[:]=np.round(net.params[layer_name][1].data*(2**dec_bits))
net.params[layer_name][1].data[net.params[layer_name][1].data>126]=127
net.params[layer_name][1].data[net.params[layer_name][1].data<-127]=-128
net.params[layer_name][1].data[:]=net.params[layer_name][1].data/(2**dec_bits)
for i in range(0,self.iterations):
for layer_no in range(0,len(self.start_layer)):
if layer_no==0:
net.forward(end=str(self.end_layer[layer_no]))
else:
net.forward(start=str(self.start_layer[layer_no]),end=str(self.end_layer[layer_no]))
if layer_no < len(self.start_layer)-1: # not quantizing accuracy layer
net.blobs[self.end_layer[layer_no]].data[:]=np.floor(net.blobs[self.end_layer[layer_no]].data*\
(2**self.act_dec_bits[self.end_layer[layer_no]]))
net.blobs[self.end_layer[layer_no]].data[net.blobs[self.end_layer[layer_no]].data>126]=127
net.blobs[self.end_layer[layer_no]].data[net.blobs[self.end_layer[layer_no]].data<-127]=-128
net.blobs[self.end_layer[layer_no]].data[:]=net.blobs[self.end_layer[layer_no]].data/\
(2**self.act_dec_bits[self.end_layer[layer_no]])
acc[i] = net.blobs[self.accuracy_layer].data*100
accuracy = acc.mean()
print('Format Q'+str(int_bits)+'.'+str(dec_bits)+' Accuracy: %.2f%%' %(accuracy))
if accuracy>best_accuracy:
best_int_bits = int_bits
best_dec_bits = dec_bits
best_accuracy = accuracy
self.bias_int_bits[layer_name] = best_int_bits
self.bias_dec_bits[layer_name] = best_dec_bits
self.bias_lshift[layer_name]=mac_dec_bits-best_dec_bits
self.act_rshift[layer_name]=mac_dec_bits-self.act_dec_bits[layer_name]
net.copy_from(self.quant_weight_file)
net.params[layer_name][1].data[:]=np.round(net.params[layer_name][1].data*(2**best_dec_bits))
net.params[layer_name][1].data[net.params[layer_name][1].data>126]=127
net.params[layer_name][1].data[net.params[layer_name][1].data<-127]=-128
net.params[layer_name][1].data[:]=net.params[layer_name][1].data/(2**best_dec_bits)
print('Final '+layer_name+ ' biases format Q'+str(best_int_bits)+'.'+str(best_dec_bits)+\
' Accuracy: %.2f%%' %(best_accuracy))
net.save(self.quant_weight_file)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', dest='gpu', action='store_true',
help='flag to enable gpu for quantization sweeps')
parser.set_defaults(gpu=False)
parser.add_argument('--accuracy', type=str, default="accuracy",
help='target accuracy')
parser.add_argument('--iterations', type=int, default=100,
help='number of iterations: data_size/batch_size')
parser.add_argument('--tolerance', type=float, default=0.001,
help='accuracy tolerance')
parser.add_argument('--model', type=str, default=\
"models/cifar10_m4_train_test.prototxt",
help='caffe model definition (.prototxt)')
parser.add_argument('--weights', type=str, default=\
"models/cifar10_m4_iter_70000.caffemodel.h5",
help='caffe model weights (.caffemodel)')
parser.add_argument('--save', type=str, default=\
"models/cifar10_m4.pkl",
help='save quantization parameters and connectivity')
cmd_args, _ = parser.parse_known_args()
gpu_flag = cmd_args.gpu
model_file=cmd_args.model
weight_file=cmd_args.weights
iterations=cmd_args.iterations
tolerance=cmd_args.tolerance
target_accuracy_layer=cmd_args.accuracy
my_model=Caffe_Quantizer(model_file,weight_file,iterations,accuracy_layer=target_accuracy_layer,gpu=gpu_flag)
my_model.get_layer_info()
my_model.get_graph_connectivity()
my_model.run_full_network()
#First quantize weights to 8 bits
my_model.quantize_wts_8bit()
#Then quantize activations to 8 bits
my_model.quantize_activations_8bit()
#Quantize biases to 8 bits based on the quantization outputs of weights and activations
my_model.quantize_bias_8bit()
my_model.run_quantized_network()
my_model.save_quant_params(cmd_args.save)
#To load the parameters use the following:
#my_model.load_quant_params('mymodel.p')
#Print dataformats
print('Input: '+my_model.data_layer+' Q'+str(my_model.act_int_bits[my_model.data_layer])+'.'+\
str(my_model.act_dec_bits[my_model.data_layer])+'(scaling factor:'+\
str(2**(my_model.act_dec_bits[my_model.data_layer]))+')')
for layer in my_model.conv_layer+my_model.ip_layer:
print('Layer: '+layer+' Q'+str(my_model.act_int_bits[layer])+'.'+str(my_model.act_dec_bits[layer])+\
' (scaling factor:'+str(2**(my_model.act_dec_bits[layer]))+') Wts: Q'+\
str(my_model.wt_int_bits[layer])+'.'+str(my_model.wt_dec_bits[layer])+\
' (scaling factor:'+str(2**(my_model.wt_dec_bits[layer]))+') Biases: Q'+\
str(my_model.bias_int_bits[layer])+'.'+str(my_model.bias_dec_bits[layer])+\
'(scaling factor:'+str(2**(my_model.bias_dec_bits[layer]))+')')
#Print data shifts to be used by ML kernels
for layer in my_model.conv_layer+my_model.ip_layer:
print('Layer: '+layer+' bias left shift: '+str(my_model.bias_lshift[layer])+\
' act_rshift: '+str(my_model.act_rshift[layer]))