rename keras_frontend

hpvm/projects/onnx/frontend/quantize_utils.py

-
-
-from scipy import stats
-from numpy import linalg
-import numpy as np
-import sys
-
-
-# NOTE: enable/disable smart quantization of weights and activations
-smart_quantization = False
-
-
-def quantize_arr(input_arr, min_val, max_val):
-
-    quantize_range = 256.0
-    input_range = max_val - min_val
-    mul_factor = input_range / quantize_range
-
-    v1 = np.subtract(input_arr, min_val)
-    v2 = np.divide(v1, mul_factor)
-    v3 = v2.astype(int)
-    v4 = np.multiply(v3, mul_factor)
-    v5 = np.add(v4, min_val)
-    v6 = np.clip(v5, min_val, max_val)
-
-    return v6
-
-
-def compute_norm(a1, a2):
-
-    norm_inp = np.subtract(a1, a2)
-    #norm = linalg.norm(norm_inp, ord = 1)
-    norm = np.sum(np.abs(norm_inp))
-    print ("*** norm = ", norm)
-        
-    return norm
-    
-
-def get_best_quant_range(input_arr):
-
-    # For disabled smart quantization, skip expensive quant range computation
-    if smart_quantization == False:
-        min_val = np.percentile(input_arr, 0.1)
-        max_val = np.percentile(input_arr, 99.9)
-        return (min_val, max_val)
-
-
-    # Trying different threshold values for INT8 quantization    
-    min_percentiles = [0, 0.1, 0.2, 0.3]
-    max_percentiles = [100, 99.9, 99.8, 99.7]
-
-      
-    min_norm = 100000000
-    min_pair = (0, 100)
-    range_vals = (0, 0)
-    for i in min_percentiles:
-      for j in max_percentiles:
-        print (" i = ", i, " j = ", j, " \n")
-        min_val = np.percentile(input_arr, i)
-        max_val = np.percentile(input_arr, j)
-
-        res = quantize_arr(input_arr, min_val, max_val)    
-        norm = compute_norm(res, input_arr)
-
-        if norm < min_norm:
-          min_norm = norm
-          min_pair = (i, j)
-          range_vals = (min_val, max_val)
-
-    print ("--- min_norm = ", min_norm, " , min_pair = ", min_pair ,  "  range_vals = ", range_vals)
-
-
-    return range_vals                
-    
-    
-
-
-
-     
-if __name__ == "__main__":
-
-
-    vals = np.zeros((2,3))
-    vals[0][0] = 1.2
-    vals[0][1] = 0.48
-    vals[0][2] = 0.5
-    vals[1][0] = -0.3
-    vals[1][1] = 0.25
-    vals[1][2] = 0.46
-
-    input_arr = np.array(vals)
-
-    res = quantize_arr(input_arr, -0.3, 1.4)
-
-    print (res, "\n")
-
-    divergence = compute_norm(res, input_arr) 
-
-    print ("divergence = ", divergence, "\n")

hpvm/projects/onnx/frontend/setup.py

-
-from setuptools import setup
-
-setup(
-    name='frontend',
-    version='1.0',
-    description='ApproxHPVM frontend modules',
-    author='Hashim',
-    author_email='hsharif3@illinois.edu',
-    packages=['frontend'],
-    install_requires=[],
-)

hpvm/projects/onnx/frontend/utils.py

-
-
-
-
-def nodeHasBias(cur_node):
-    
-  if cur_node.layer_type == "Conv2D" or cur_node.layer_type == "DepthwiseConv2D" or cur_node.layer_type == "Dense":
-    #return True
-    return cur_node.use_bias
-  else:
-    return False
-
-  
-def layerHasActivationAttr(cur_node):
-    
-  if cur_node.layer_type == "Conv2D" or cur_node.layer_type == "DepthwiseConv2D" \
-     or cur_node.layer_type == "Dense" or cur_node.layer_type == "Activation":
-    return True
-  else:
-    return False
-
-
-def nodeHasActivation(cur_node):
-    
-  if cur_node.layer_type == "Conv2D" or cur_node.layer_type == "DepthwiseConv2D" \
-     or cur_node.layer_type == "Dense" or cur_node.layer_type == "Activation":
-    #return True
-    return cur_node.activation_type != "linear"
-  else:
-    return False
-
-
-def genActivationCallStr(input_var, output_var, activation_type):
- 
-  func_name = ""
-  if activation_type == "tanh":
-    func_name = "Tanh"
-
-  if activation_type == "relu":
-    func_name = "Relu"
-
-  if activation_type == "softmax":
-    func_name = "Softmax"
-
-  inst_str = "void* " + output_var + " = "
-  inst_str += "tensor" + func_name + "(" + input_var + "); \n"
-
-  print ("***** inst_str = ", inst_str, "\n")
-    
-  return inst_str
-
-  

hpvm/projects/onnx/frontend/weight_utils.py

-
-import numpy as np
-import struct
-import random
-
-
-def dumpLabels(file_name, Y_test):
-    
-    f = open(file_name, "wb")
-    
-    labels_map = {}    
-    for label in Y_test:
-        label_val = 0
-        if len(Y_test.shape) > 1:        
-          #label_val = np.int8(label[0])
-          label_val = np.int32(label[0])
-        else:
-          #label_val = np.int8(label)
-          label_val = np.int32(label)
-         
-        if label_val not in labels_map:
-            labels_map[label_val] = 0
-            labels_map[label_val] += 1
-
-        f.write(label_val)
-
-    f.close()
-    
-
-    
-"""def dumpData(file_name, X_test):
-
-    N = X_test.shape[0]
-    C = X_test.shape[1]
-    H = X_test.shape[2]
-    W = X_test.shape[3]
- 
-    print ("*DumpData")
-    print("-min_val = ", np.amin(X_test))
-    print("-max_val = ", np.amax(X_test))
-    
-    f = open(file_name, "wb")
-    for i in range(N):
-        for j in range(C):
-            for k in range(H):
-                for l in range(W):
-                    val = struct.unpack("f", struct.pack("f", X_test[i][j][k][l]))
-                    f.write(np.float32(val[0]))
-
-    f.close()
-"""
-
-
-def dumpData(file_name, X_test):
-
-    N = X_test.shape[0]
-    C = X_test.shape[1]
-    H = X_test.shape[2]
-    W = X_test.shape[3]
- 
-    print ("*DumpData")
-    print("-min_val = ", np.amin(X_test))
-    print("-max_val = ", np.amax(X_test))
-
-    f = open(file_name, "wb")
-
-    X_test.tofile(f)    
-
-    f.close()
-
-    
-    
-def dumpConvWeights(file_name, weights, N, C, H, W):
-
-    print (weights.shape)
-    print ("*DumpConvWeights")
-    print("-min_val = ", np.amin(weights))
-    print("-max_val = ", np.amax(weights))
-
-    
-    f = open(file_name, "wb")
-    for i in range(N):
-        for j in range(C):
-            for k in range(H):
-                for l in range(W):
-                    f.write(weights[k][l][j][i])
-
-    f.close()
-
-
-    
-def dumpFcWeights(file_name, weights, H, W):
-
-    print (weights.shape)
-    print ("*DumpFcWeights")
-    print("-min_val = ", np.amin(weights))
-    print("-max_val = ", np.amax(weights))
-
-
-    f = open(file_name, "wb")
-    for i in range(H):
-        for j in range(W):
-            f.write(weights[i][j])
-
-    f.close()        
-
-
-    
-def dumpFcBias(file_name, bias, W):
-
-    print (bias.shape)
-    print ("*DumpFcBias")
-    print("-min_val = ", np.amin(bias))
-    print("-max_val = ", np.amax(bias))
-
-
-    f = open(file_name, "wb")
-    for i in range(W):
-        f.write(bias[i])
-
-    f.close()
-
-
-
-def dumpCalibrationData(file_name, X_train, labels_fname, train_labels):
-
-  combined_list = []
-  for i in range(len(X_train)):
-    tup = (X_train[i], train_labels[i])
-    combined_list.append(tup)       
-  
-  np.random.shuffle(combined_list)
-  #X_calibration = X_train[0:5000]
-
-  data_list = []
-  labels_list = []
-  for i in range(5000):
-    tup = combined_list[i]
-    data_list.append(tup[0])
-    labels_list.append(tup[1])
-
-  data_list = np.array(data_list)
-  labels_list = np.array(labels_list)
-  
-  dumpData(file_name, data_list)
-  dumpLabels(labels_fname, labels_list)
-