schihei/mixed-precision-neural-networks-on-risc-v-cores
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Adding new features

Browse Source
This commit is contained in:
alexmr09
2024-07-23 13:00:49 +03:00
parent 9e044fd7fc
commit 745cc4ed6d
28 changed files with 33632 additions and 106 deletions
Download Patch File Download Diff File Expand all files Collapse all files
mpq/configure_ibex.py
+171 -58
View File
@@ -27,26 +27,34 @@ def quantize_multiplier(real_multiplier):
return quantized_multiplier, right_shift
def get_int_params(quant_net):
int_weights = []
int_bias = []
in_scales = []
act_scales = []
for _, module in quant_net.sequential.named_children():
if hasattr(module, 'weight') and module.weight is not None:
int_weights.append(module.int_weight().cpu().numpy())
int_bias.append(module.int_bias().cpu().numpy())
in_scales.append(module.quant_bias_scale().cpu().detach().numpy())
def extract_quant_params(module):
for name, submodule in module.named_children():
# Check if the submodule has weights and append them if present
if hasattr(submodule, 'weight') and submodule.weight is not None:
int_weights.append(submodule.int_weight().cpu().detach().numpy())
int_bias.append(submodule.int_bias().cpu().detach().numpy())
in_scales.append(submodule.quant_bias_scale().cpu().detach().numpy())
if hasattr(module, 'quant_act_scale') and module.quant_act_scale is not None:
act_scales.append(module.quant_act_scale().cpu().detach().numpy())
# Check if the submodule has activation scale and append it if present
if hasattr(submodule, 'quant_act_scale') and submodule.quant_act_scale is not None:
act_scales.append(submodule.quant_act_scale().cpu().detach().numpy())
act_scales.append(quant_net.o_quant.quant_act_scale().cpu().detach().numpy())
# Recursively extract parameters from the children modules
extract_quant_params(submodule)
# Start extraction from the top-level module
extract_quant_params(quant_net)
mul_vals, shift_vals = [], []
for i in range(len(act_scales)):
M = in_scales[i]/act_scales[i]
for i in range(len(act_scales)-1):
M = in_scales[i]/act_scales[i+1]
mul, shift = quantize_multiplier(M[0])
mul_vals.append(mul)
shift_vals.append(shift)
@@ -87,22 +95,27 @@ def decide_mode(network, weight_bit_width, input_uint8 = True):
for name, module in network.named_modules():
if isinstance(module, layer_types_py):
layer_type_name = module.__class__.__name__
if(layer_type_name == 'Conv2d' or layer_type_name == 'Linear' or layer_type_name == 'DepthwiseConv2d'):
if(layer_type_name == 'Linear'):
layer_type.append(layer_type_name)
if(layer_type_name == 'Conv2d'):
if(module.groups == module.in_channels):
layer_type.append('DepthwiseConv2d')
else:
layer_type.append(layer_type_name)
else:
if(layer_type_name == 'ReLU' or layer_type_name == 'Sigmoid'):
input_sign[ins] = 0
ins += 1
for i in range(len(weight_bit_width)):
signed_input = 4 * input_sign[i]
if(weight_bit_width[i] == 2):
mode_per_layer.append(signed_input + 3)
elif(weight_bit_width[i] == 4):
mode_per_layer.append(signed_input + 2)
else:
if(layer_type[i] == 'DepthwiseConv2d'):
if(layer_type[i] == 'DepthwiseConv2d'):
mode_per_layer.append(signed_input + 1)
else:
if(weight_bit_width[i] == 2):
mode_per_layer.append(signed_input + 3)
elif(weight_bit_width[i] == 4):
mode_per_layer.append(signed_input + 2)
else:
mode_per_layer.append(signed_input)
@@ -161,15 +174,22 @@ def pad_inputs_weights(quant_net, test_loader, mode_per_layer,
else:
new_size_0 = a * 4
b = w.shape[1] // 4
if(w.shape[1] % 4 != 0):
new_size_1 = (b + 1) * 4
if((mode_per_layer[i] != 1) and (mode_per_layer[i] != 5)):
b = w.shape[1] // 4
if(w.shape[1] % 4 != 0):
new_size_1 = (b + 1) * 4
else:
new_size_1 = b * 4
new_w = np.zeros((new_size_0, new_size_1, w.shape[2], w.shape[3])).astype(np.int8)
new_w[:w.shape[0], :w.shape[1], :, :] = w
else:
new_size_1 = b * 4
new_w = np.zeros((new_size_0, new_size_1, w.shape[2], w.shape[3])).astype(np.int8)
new_w[:w.shape[0], :w.shape[1], :, :] = w
new_size_1 = 1
new_w = np.zeros((new_size_0, new_size_1, w.shape[2], w.shape[3])).astype(np.int8)
new_w[:w.shape[0], :w.shape[1], :, :] = w
new_w = np.squeeze(new_w, axis = 1)
padded_int_weights.append(new_w)
padded_int_biases = []
@@ -325,6 +345,15 @@ def concat_inputs_weights(mode_per_layer, padded_input, padded_int_weights, padd
comb = combine_values(vector)
new_mat[i][j] = comb
elif(len(dims) == 3):
new_mat = np.zeros((int(dims[0]//4), dims[1], dims[2]), dtype = np.int64)
for i in range(int(dims[0]//4)):
for j in range(dims[1]):
for k in range(dims[2]):
vector = layer_weight[4*i : 4*(i+1), j, k]
comb = combine_values(vector)
new_mat[i][j][k] = comb
elif(len(dims) == 4):
if((mode_per_layer[iter] == 0) | (mode_per_layer[iter] == 4)):
new_mat = np.zeros((int(dims[0]//4), dims[1], dims[2], dims[3]), dtype = np.int64)
@@ -602,9 +631,17 @@ def save_cnn_net_params(path, int_weights, int_biases, mul_vals, shift_vals, shi
dims = np.shape(int_weights[k])
mat = int_weights[k]
if(len(dims) == 2):
wi += 1
st = 'static const int W' + str(wi) + '[' + str(dims[0]) + ']' + '[' + str(dims[1]) + '] = {\n'
if(len(dims) == 2 or ((len(dims) == 4) and dims[2] == dims[3] == 1)):
f.write('static const int ')
if(len(dims) == 2):
wi += 1
f.write('W' + str(wi))
else:
mat = np.squeeze(mat, axis = (2,3))
fi += 1
f.write('F' + str(fi))
st = '[' + str(dims[0]) + ']' + '[' + str(dims[1]) + '] = {\n'
f.write(st)
for n in range(dims[0]):
f.write('\t{')
@@ -618,7 +655,33 @@ def save_cnn_net_params(path, int_weights, int_biases, mul_vals, shift_vals, shi
f.write(',')
f.write('\n')
f.write('};\n\n')
elif (len(dims) == 3):
dims = np.shape(mat)
fi += 1
st = 'static const int F' + str(fi) + '[' + str(dims[0]) + '][' + str(dims[1])
st += '][' + str(dims[2]) + '] = {\n'
f.write(st)
for n in range(dims[0]):
f.write('\t{\n')
for l in range(dims[1]):
f.write('\t\t{')
for h in range(dims[2] - 1):
f.write(str(mat[n][l][h]) + ', ')
if dims[2] != 1:
f.write(str(mat[n][l][dims[2] - 1]) + '}')
else:
f.write(str(mat[n][l][0]) + '}')
if (l != dims[1] - 1):
f.write(',')
f.write('\n')
f.write('\t}')
if n != dims[0] - 1:
f.write(',')
f.write('\n')
f.write('};\n\n')
elif(len(dims) == 4):
mat = np.transpose(mat, (0, 2, 3, 1))
dims = np.shape(mat)
@@ -856,9 +919,11 @@ def generate_opt_c_code_mlp(path, name, int_weights, optimal_config, type_of_lay
f.write('\t' + name + '();\n\n')
f.write('\treturn 0;\n}')
def get_cnn_details(model):
details = []
for layer in model.children():
def get_cnn_details(module, details = None):
if details is None:
details = []
for layer in module.children():
if isinstance(layer, nn.Conv2d):
details.append({
"layer_type": "Conv2d",
@@ -866,18 +931,19 @@ def get_cnn_details(model):
"out_channels": layer.out_channels,
"kernel_size": layer.kernel_size,
"stride": layer.stride,
"padding": layer.padding
"padding": layer.padding,
"groups": layer.groups
})
elif (isinstance(layer, nn.MaxPool2d)):
elif isinstance(layer, nn.MaxPool2d):
details.append({
"layer_type": "MaxPool2d",
"kernel_size": layer.kernel_size,
"stride": layer.stride,
"padding": layer.padding
})
elif (isinstance(layer, nn.AvgPool2d)):
elif isinstance(layer, nn.AvgPool2d):
details.append({
"layer_type": "AvgPool2d",
"kernel_size": layer.kernel_size,
@@ -891,6 +957,10 @@ def get_cnn_details(model):
"in_features": layer.in_features,
"out_features": layer.out_features
})
# Recursively apply to children modules
get_cnn_details(layer, details)
return details
def generate_og_c_code_cnn(path, name, input, cnn_details, int_weights):
@@ -900,10 +970,17 @@ def generate_og_c_code_cnn(path, name, input, cnn_details, int_weights):
f.write('#include "fully_connected.h"\n')
f.write('#include "ibex_cnn_params.h"\n')
f.write('#include "ibex_inputs.h"\n')
f.write('#include "conv2d.h"\n\n')
f.write('#include "conv2d.h"\n')
f.write('#define IMG_SZ ' + str(input.shape[2]) + '\n')
f.write('#define NUM_FIL0 ' + str(int_weights[0].shape[1]) + '\n\n')
for detail in cnn_details[:-1]:
if detail["layer_type"] == "Conv2d":
if(detail["in_channels"] == detail["out_channels"] == detail["groups"] != 1):
f.write('#include "dws_conv.h"\n')
break
f.write('\n')
f.write('#define IMG_SZ ' + str(np.shape(input)[2]) + '\n')
f.write('#define NUM_FIL0 ' + str(np.shape(input)[1]) + '\n\n')
i = 1
for w in int_weights:
if(len(np.shape(w)) == 4):
@@ -1050,11 +1127,17 @@ def generate_og_c_code_cnn(path, name, input, cnn_details, int_weights):
for detail in cnn_details[:-1]:
if detail["layer_type"] == "Conv2d":
if(detail["in_channels"] == detail["out_channels"] == detail["groups"] != 1):
conv_type = 'dw_conv'
elif(detail["kernel_size"][0] == 1):
conv_type = 'pw_conv'
else:
conv_type = "conv2"
if(i == 1):
f.write('\t\tconv2(inp_dim, f_dim1, outp_dim1, in, F1, B1, ')
f.write('\t\t' + conv_type + '(inp_dim, f_dim1, outp_dim1, in, F1, B1, ')
f.write('out1, STRIDE1, pad_1, SB1, MV1, SV1);')
else:
f.write('\t\tconv2(outp_dim' + str(i-1) + ', f_dim' + str(i) + ', outp_dim' + str(i))
f.write('\t\t' + conv_type + '(outp_dim' + str(i-1) + ', f_dim' + str(i) + ', outp_dim' + str(i))
f.write(', out' + str(i-1) + ', F' + str(fi) + ', B' + str(fi) + ', out' + str(i))
f.write(', STRIDE' + str(fi) + ', pad_' + str(i) + ', SB' + str(fi))
f.write(', MV' + str(fi) + ', SV' + str(fi) + ');')
@@ -1091,10 +1174,17 @@ def generate_og_c_code_cnn(path, name, input, cnn_details, int_weights):
f.write('\n')
i += 1
f.write('\t\tmlp_layer(out' + str(i-1) + ', out, DENSE_DIM' + str(dn-1))
f.write(', OUT_DIM, W' + str(dn) + ', B' + str(fi + dn - 1))
f.write(', SB' + str(fi + dn - 1) + ', MV' + str(fi + dn - 1))
f.write(', SV' + str(fi + dn - 1) + ');\n')
if flatten == 0:
f.write('\t\tflatten(outp_dim' + str(i-1) + ', out' + str(i-1) + ', out' + str(i) + ');\n\n')
i += 1
f.write('\t\tmlp_layer(out' + str(i-1) + ', out, flatten_dim, OUT_DIM, ')
f.write('W1, B' + str(fi + dn - 1) + ', SB' + str(fi + dn - 1) + ', MV' + str(fi + dn - 1))
f.write(', SV' + str(fi + dn - 1) + ');')
else:
f.write('\t\tmlp_layer(out' + str(i-1) + ', out, DENSE_DIM' + str(dn-1))
f.write(', OUT_DIM, W' + str(dn) + ', B' + str(fi + dn - 1))
f.write(', SB' + str(fi + dn - 1) + ', MV' + str(fi + dn - 1))
f.write(', SV' + str(fi + dn - 1) + ');\n')
f.write('\n\t\tpcount_enable(0);\n\n')
f.write('\t\tputs("Output Layer Values:\\n");\n')
@@ -1119,13 +1209,21 @@ def generate_opt_c_code_cnn(path, name, input, cnn_details, int_weights, optimal
f.write('#include "fully_connected_opt.h"\n')
f.write('#include "ibex_cnn_params.h"\n')
f.write('#include "ibex_inputs.h"\n')
f.write('#include "conv2d_opt.h"\n\n')
f.write('#include "conv2d_opt.h"\n')
for detail in cnn_details[:-1]:
if detail["layer_type"] == "Conv2d":
if(detail["in_channels"] == detail["out_channels"] == detail["groups"] != 1):
f.write('#include "dws_conv_opt.h"\n')
break
f.write('\n')
f.write('#define IMG_SZ ' + str(np.shape(input)[2]) + '\n')
f.write('#define NUM_FIL0 ' + str(np.shape(input)[0]) + '\n\n')
f.write('#define NUM_FIL0 ' + str(np.shape(input)[1]) + '\n\n')
i = 1
for w in int_weights:
if(len(np.shape(w)) == 4):
if(len(np.shape(w)) == 4 or len(np.shape(w)) == 3):
f.write('#define FILTER' + str(i) + ' ' + str(w.shape[2]) + '\n')
i += 1
@@ -1133,7 +1231,7 @@ def generate_opt_c_code_cnn(path, name, input, cnn_details, int_weights, optimal
i = 1
for w in int_weights:
if(len(np.shape(w)) == 4):
if(len(np.shape(w)) == 4 or len(np.shape(w)) == 3):
f.write('#define NUM_FIL' + str(i) + ' ' + str(w.shape[0]) + '\n')
i += 1
@@ -1270,14 +1368,21 @@ def generate_opt_c_code_cnn(path, name, input, cnn_details, int_weights, optimal
for detail in cnn_details[:-1]:
if detail["layer_type"] == "Conv2d":
if(detail["in_channels"] == detail["out_channels"] == detail["groups"] != 1):
conv_type = 'dw_conv_opt'
elif(detail["kernel_size"][0] == 1):
conv_type = 'pw_conv_' + str(optimal_config[j]) + 'bits'
else:
conv_type = 'conv2_' + str(optimal_config[j]) + 'bits'
if(i == 1):
f.write('\t\tconv2_' + str(optimal_config[j]) + 'bits')
if(np.shape(input)[0] == 1):
f.write('\t\t' + conv_type)
if(np.shape(input)[1] == 1):
f.write('_1ch')
f.write('(inp_dim, f_dim1, outp_dim1, in, F1, B1, ')
f.write('out1, STRIDE1, pad_1, SB1, MV1, SV1);')
else:
f.write('\t\tconv2_' + str(optimal_config[j]) + 'bits(outp_dim' + str(i-1) + ', f_dim' + str(i))
f.write('\t\t' + conv_type + '(outp_dim' + str(i-1) + ', f_dim' + str(i))
f.write(', outp_dim' + str(i) + ', out' + str(i-1) + ', F' + str(fi) + ', B' + str(fi) + ', out')
f.write(str(i) + ', STRIDE' + str(fi) + ', pad_' + str(i) + ', SB' + str(fi))
f.write(', MV' + str(fi) + ', SV' + str(fi) + ');')
@@ -1314,11 +1419,19 @@ def generate_opt_c_code_cnn(path, name, input, cnn_details, int_weights, optimal
f.write('\n')
i += 1
f.write('\t\tmlp_layer_' + str(optimal_config[-1]) + 'bits(out' + str(i-1) + ', out, DENSE_DIM' + str(dn-1))
f.write(', OUT_DIM, W' + str(dn) + ', B' + str(fi + dn - 1))
f.write(', SB' + str(fi + dn - 1) + ', MV' + str(fi + dn - 1))
f.write(', SV' + str(fi + dn - 1) + ');\n')
if flatten == 0:
f.write('\t\tflatten(outp_dim' + str(i-1) + ', out' + str(i-1) + ', out' + str(i) + ');\n\n')
i += 1
f.write('\t\tmlp_layer_' + str(optimal_config[j]) + 'bits(out' + str(i-1) + ', out, ')
f.write('flatten_dim, OUT_DIM, W1, B' + str(fi + dn - 1) + ', SB' + str(fi + dn - 1) + ', MV')
f.write(str(fi + dn - 1) + ', SV' + str(fi + dn - 1) + ');\n')
else:
f.write('\t\tmlp_layer_' + str(optimal_config[-1]) + 'bits(out' + str(i-1) + ', out, DENSE_DIM' + str(dn-1))
f.write(', OUT_DIM, W' + str(dn) + ', B' + str(fi + dn - 1))
f.write(', SB' + str(fi + dn - 1) + ', MV' + str(fi + dn - 1))
f.write(', SV' + str(fi + dn - 1) + ');\n')
f.write('\n\t\tpcount_enable(0);\n\n')
f.write('\t\tputs("Output Layer Values:\\n");\n')
f.write('\t\tfor(int i = 0; i < OUT_DIM; i++) {\n')
f.write('\t\t\tputhex((out[i] & 0xFF000000) >> 24);\n')