diff --git a/llvm/projects/keras/frontend/approxhpvm_translator.py b/llvm/projects/keras/frontend/approxhpvm_translator.py
index 2c964db34b60f7bb67ceda8d9cba464d7856ef82..e60d6adb994ee4da56765a1fa365b14a792863d2 100644
--- a/llvm/projects/keras/frontend/approxhpvm_translator.py
+++ b/llvm/projects/keras/frontend/approxhpvm_translator.py
@@ -21,6 +21,7 @@ class DFG:
def hasSingleInput(self, layer):
+
layer_name = layer.__class__.__name__
singleInLayers = {}
@@ -57,12 +58,12 @@ class DFG:
inbound_node_name = inbound_node_name.split("/")[0]
if inbound_node_name in self.node_map:
inbound_node = self.node_map[inbound_node_name]
- print (inbound_node_name, " found!")
+ DEBUG (inbound_node_name, " found!")
inbound_node.add_output(dfg_node)
dfg_node.add_input(inbound_node)
else:
- print ("--inbound node NOT FOUND!")
+ DEBUG ("--inbound node NOT FOUND!")
@@ -75,12 +76,12 @@ class DFG:
if self.hasMultipleInputs(layer):
for j in range(len(layer.input)):
- print(type(layer.input[j]))
- print(layer.input[j].op.name)
+ DEBUG (type(layer.input[j]))
+ DEBUG (layer.input[j].op.name)
self.add_dfg_edge(layer.input[j].op.name, dfg_node)
else:
- print (layer.input.name)
+ DEBUG (layer.input.name)
self.add_dfg_edge(layer.input.name, dfg_node)
# Adding DFG node to name mapping
@@ -104,8 +105,8 @@ class DFG:
return
if self.predVisited(cur_node, visited_nodes):
- print(cur_node.layer_type)
- print(cur_node.layer_name)
+ DEBUG (cur_node.layer_type)
+ DEBUG (cur_node.layer_name)
visited_nodes[cur_node.layer_name] = True
# Invoking traversal on outbound nodes
@@ -119,7 +120,7 @@ class DFG:
#Build and Print the DFG in reverse postorder
def buildDFG(self):
- print ("\n\n ****** Traversing and Printing DFG ******* \n\n")
+ DEBUG ("\n\n ****** Traversing and Printing DFG ******* \n\n")
visited_nodes = {}
# Starting traversal at the DFG root node
self.traverseNode(self.root_node, visited_nodes)
@@ -144,42 +145,42 @@ class DFGNode:
layer_type = layer.__class__.__name__
self.layer_type = layer_type # layer type e.g., conv2d, add, dense
self.layer_name = layer.name # unique layer identifier
- print (self.layer_name)
+ DEBUG (self.layer_name)
if layer_type == "Conv2D" or layer_type == "DepthwiseConv2D" or layer_type == "Dense":
self.weights = layer.get_weights()[0]
- print("\t", self.weights.shape)
+ DEBUG ("\t", self.weights.shape)
self.use_bias = layer.use_bias
if layer.use_bias:
self.use_bias = layer.use_bias
self.bias_weights = layer.get_weights()[1]
- print("\t", self.bias_weights.shape)
+ DEBUG ("\t", self.bias_weights.shape)
if layer_type == "Conv2D" or layer_type == "DepthwiseConv2D":
self.padding = layer.padding
self.strides = layer.strides
- print("\t", self.strides)
- print("\tPadding = ", self.padding)
+ DEBUG ("\t", self.strides)
+ DEBUG ("\tPadding = ", self.padding)
if layer_type == "MaxPooling2D" or layer_type == "AveragePooling2D":
self.pool_size = layer.pool_size
self.strides = layer.strides
- print("\t pool_size = ", self.pool_size)
- print("\t strides = ", self.strides)
+ DEBUG ("\t pool_size = ", self.pool_size)
+ DEBUG ("\t strides = ", self.strides)
if layerHasActivationAttr(self):
self.activation_type = layer.activation.__name__
- print ("\t Activation = ", self.activation_type)
+ DEBUG ("\t Activation = ", self.activation_type)
if layer_type == "ZeroPadding2D":
- print ("***ZeroPaddding \n");
+ DEBUG ("***ZeroPaddding \n");
self.padding = layer.padding
- print ("padding = ", self.padding);
+ DEBUG ("padding = ", self.padding);
if layer_type == "BatchNormalization":
self.epsilon = layer.epsilon
@@ -264,12 +265,12 @@ class TensorRtTranslator:
def getSingleInputName(self, cur_node):
- print (cur_node.layer_name)
+ DEBUG (cur_node.layer_name)
# Assumption: If no inputs, the previous layer must be input layer
if len(cur_node.inputs) == 0:
return "input"
- print ("Input_type = ", cur_node.inputs[0].layer_type)
+ DEBUG ("Input_type = ", cur_node.inputs[0].layer_type)
# NOTE: Assuming the 'inference' phase - hence skipping Dropout
pred_layer_type = cur_node.inputs[0].layer_type
@@ -295,12 +296,12 @@ class TensorRtTranslator:
def getPrevLayerPadding(self, cur_node):
- print (cur_node.layer_name)
+ DEBUG (cur_node.layer_name)
# Assumption: If no inputs, the previous layer must be input layer
if len(cur_node.inputs) == 0:
return None
- print ("Input_type = ", cur_node.inputs[0].layer_type)
+ DEBUG ("Input_type = ", cur_node.inputs[0].layer_type)
if cur_node.inputs[0].layer_type == "ZeroPadding2D":
pred_padding = cur_node.inputs[0].padding
return pred_padding
@@ -483,9 +484,7 @@ class TensorRtTranslator:
self.program_str += inst_str
-
-
-
+
def codegenNode(self, dfg, cur_node, visited_nodes):
@@ -494,12 +493,11 @@ class TensorRtTranslator:
if cur_node.layer_name in visited_nodes:
return
- print ("-visiting = ", cur_node.layer_name, "\n")
+ DEBUG ("-visiting = ", cur_node.layer_name, "\n")
if dfg.predVisited(cur_node, visited_nodes):
visited_nodes[cur_node.layer_name] = True
-
self.genNodeCalls(cur_node)
# Invoking traversal on outbound nodes
@@ -510,11 +508,13 @@ class TensorRtTranslator:
# Print the DFG in reverse postorder
def codegen(self, dfg):
- print ("\n\n ****** Codegen for HPVM Tensor Rt ******* \n\n")
+ print ("\n *** Starting Codegen for HPVM Tensor Rt *** \n")
visited_nodes = {}
# Starting traversal at the DFG root node
self.codegenNode(dfg, dfg.root_node, visited_nodes)
+ print ("\n\n --- Codegen Completed --- \n\n")
+
@@ -531,7 +531,7 @@ class TensorRtTranslator:
w_name = layer_name + "_w"
self.filter_names[w_name] = 1
- print (weights.shape, w_name)
+ DEBUG (weights.shape, w_name)
N = weights.shape[3]
C = weights.shape[2]
@@ -563,7 +563,7 @@ class TensorRtTranslator:
b_name = layer_name + "_b"
self.filter_names[b_name] = 1
- print (bias_weights.shape, b_name)
+ DEBUG (bias_weights.shape, b_name)
unique_file_name = b_name + ".bin"
dumpFcBias(prefix + unique_file_name, bias_weights, bias_weights.shape[0])
@@ -584,7 +584,7 @@ class TensorRtTranslator:
w_name = layer_name + "_w"
self.filter_names[w_name] = 1
- print (weights.shape, w_name)
+ DEBUG (weights.shape, w_name)
H = weights.shape[0]
W = weights.shape[1]
@@ -606,7 +606,7 @@ class TensorRtTranslator:
b_name = layer_name + "_b"
self.filter_names[b_name] = 1
- print (bias_weights.shape, b_name)
+ DEBUG (bias_weights.shape, b_name)
unique_file_name = b_name + ".bin"
dumpFcBias(prefix + unique_file_name, bias_weights, bias_weights.shape[0])
@@ -862,7 +862,7 @@ def reloadModelParams(model, reload_dir, x_test, y_test):
for i in range(len(model.layers)):
layer = model.layers[i]
layer_name = layer.name
- print ("*layer_name = ", layer_name)
+ DEBUG ("*layer_name = ", layer_name)
if "conv" not in layer_name and "dense" not in layer_name:
continue
@@ -870,9 +870,6 @@ def reloadModelParams(model, reload_dir, x_test, y_test):
w_path = reload_dir + layer_name + "_w.bin"
b_path = reload_dir + layer_name + "_b.bin"
- print ("** w_path = ", w_path)
- print ("** b_path = ", b_path)
-
w_arr = np.fromfile(w_path, dtype='float32')
b_arr = np.fromfile(b_path, dtype='float32')
@@ -885,7 +882,7 @@ def reloadModelParams(model, reload_dir, x_test, y_test):
b_arr = np.reshape(b_arr, b_shape)
w_arr = np.transpose(w_arr, (2,3,1,0))
- print ("old_shape = ", w_shape, " new_shape = ", w_arr.shape)
+ DEBUG ("old_shape = ", w_shape, " new_shape = ", w_arr.shape)
if "dense" in layer_name:
w_arr = np.reshape(w_arr, w_shape)
@@ -921,7 +918,6 @@ def getUniquePath(weights_dir):
weights_dir = getUniquePath(weights_dir)
- #print (weights_dir)
return weights_dir
@@ -950,8 +946,8 @@ def translate_to_approxhpvm(model, weights_dir, test_data=None, test_labels=None
dfg.buildDFG()
- print ("test_data.shape = ", test_data.shape, "\n")
- print ("test_labels.shape = ", test_labels.shape, "\n")
+ DEBUG ("test_data.shape = ", test_data.shape, "\n")
+ DEBUG ("test_labels.shape = ", test_labels.shape, "\n")
tensorRtTranslator = TensorRtTranslator(dfg)
tensorRtTranslator.translate(model, weights_dir, test_data, test_labels)
@@ -965,7 +961,11 @@ def translate_to_approxhpvm(model, weights_dir, test_data=None, test_labels=None
filter_names = tensorRtTranslator.getFilterNames()
hpvmTranslator = HPVMTranslator(dfg, weight_str, input_str, filter_names)
hpvmTranslator.translate(model, weights_dir, test_data)
-
+
+ print ("-- Weight Files Under : ", weights_dir)
+ print ("-- TensorRT src : ", weights_dir + "/src.cc")
+ print ("-- ApproxHPVM src : ", weights_dir + "approxhpvm_src.cc")
+
return weights_dir
diff --git a/llvm/projects/keras/frontend/hpvm_dfg_translator.py b/llvm/projects/keras/frontend/hpvm_dfg_translator.py
index eda8ed9da8ab9e34f77a6165ca79d2e0a5e4e4f8..65574a98881f010bf7cd67df344517803de8c67c 100644
--- a/llvm/projects/keras/frontend/hpvm_dfg_translator.py
+++ b/llvm/projects/keras/frontend/hpvm_dfg_translator.py
@@ -59,12 +59,12 @@ class HPVMTranslator:
def getSingleInputName(self, cur_node):
- print (cur_node.layer_name)
+ DEBUG (cur_node.layer_name)
# Assumption: If no inputs, the previous layer must be input layer
if len(cur_node.inputs) == 0:
return "input"
- print ("Input_type = ", cur_node.inputs[0].layer_type)
+ DEBUG ("Input_type = ", cur_node.inputs[0].layer_type)
pred_layer_type = cur_node.inputs[0].layer_type
# NOTE: Assuming the 'inference' phase - hence skipping Dropout
@@ -90,12 +90,12 @@ class HPVMTranslator:
def getPrevLayerPadding(self, cur_node):
- print (cur_node.layer_name)
+ DEBUG (cur_node.layer_name)
# Assumption: If no inputs, the previous layer must be input layer
if len(cur_node.inputs) == 0:
return None
- print ("Input_type = ", cur_node.inputs[0].layer_type)
+ DEBUG ("Input_type = ", cur_node.inputs[0].layer_type)
if cur_node.inputs[0].layer_type == "ZeroPadding2D":
pred_padding = cur_node.inputs[0].padding
return pred_padding
@@ -215,8 +215,8 @@ class HPVMTranslator:
# Fix: replace deprecated genHpvmNodeEdges with genHpvmEdges
def genHpvmNodeEdges(self, out_var_name, input_var_name, input_var_name2):
- print ("input_var_name2 = ", input_var_name2)
- print ("input_var_name = ", input_var_name)
+ DEBUG ("input_var_name2 = ", input_var_name2)
+ DEBUG ("input_var_name = ", input_var_name)
hpvm_edge_str = "\n void* " + out_var_name + " = "
hpvm_edge_str += "__visc__createNodeND(0, " + out_var_name + "_node); \n\n"
@@ -531,11 +531,13 @@ class HPVMTranslator:
# Print the DFG in reverse postorder
def codegen(self, dfg):
- print ("\n\n ****** Codegen for ApproxHPVM DFG Representation ******* \n\n")
+ print ("\n *** Starting Codegen for ApproxHPVM DFG Representation *** \n")
visited_nodes = {}
# Starting traversal at the DFG root node
self.codegenNode(dfg, dfg.root_node, visited_nodes)
+ print ("\n --- Codegen Completed --- \n")
+
@@ -648,7 +650,7 @@ class HPVMTranslator:
program_str = self.file_header_str + self.node_str + self.root_str
program_str += self.root_struct_str + self.main_func_str
- print (program_str)
+ DEBUG (program_str)
f = open(dir_prefix + "/approxhpvm_src.cc", "w+")
f.write(program_str)
diff --git a/llvm/projects/keras/frontend/utils.py b/llvm/projects/keras/frontend/utils.py
index ffc338c19ea60df7c53430ac38b613c2daef402e..9a2c13715ee351b61f8d7aaffc148a90a45ee233 100644
--- a/llvm/projects/keras/frontend/utils.py
+++ b/llvm/projects/keras/frontend/utils.py
@@ -1,6 +1,13 @@
+def DEBUG(str, *args):
+
+ debug = False
+ if debug:
+ print (str, *args)
+
+
def nodeHasBias(cur_node):
@@ -45,7 +52,6 @@ def genActivationCallStr(input_var, output_var, activation_type):
inst_str = "void* " + output_var + " = "
inst_str += "tensor" + func_name + "(" + input_var + "); \n"
- print ("***** inst_str = ", inst_str, "\n")
return inst_str
diff --git a/llvm/projects/keras/frontend/weight_utils.py b/llvm/projects/keras/frontend/weight_utils.py
index 4a9e3dbe0a1947fadda46704c984bbb087ce1545..dd22765386e2172572ad0feec201c7dec407a909 100644
--- a/llvm/projects/keras/frontend/weight_utils.py
+++ b/llvm/projects/keras/frontend/weight_utils.py
@@ -6,9 +6,10 @@ from keras.optimizers import Adam
def dumpLabels(file_name, Y_test):
+
+ print ("Dumping Labels File = ", file_name)
- f = open(file_name, "wb")
-
+ f = open(file_name, "wb")
labels_map = {}
for label in Y_test:
label_val = 0
@@ -54,14 +55,12 @@ def dumpData(file_name, X_test):
def dumpData(file_name, X_test):
- print ("*DumpData")
+ print ("*Dumping Input File = ", file_name)
#print("-min_val = ", np.amin(X_test))
#print("-max_val = ", np.amax(X_test))
f = open(file_name, "wb")
- print ("X.shape = ", X_test.shape)
- print ("X_test.size", len(X_test))
X_test = X_test.flatten()
X_test = X_test.astype(np.float32)
X_test.tofile(f)
@@ -94,18 +93,13 @@ def dumpConvWeights(file_name, weights, N, C, H, W):
def dumpConvWeights(file_name, X_test, N, C, H, W):
- print ("*DumpConvWeights")
+ print ("*Dumping Conv Weights to file = ", file_name)
#print("-min_val = ", np.amin(X_test))
#print("-max_val = ", np.amax(X_test))
f = open(file_name, "wb")
- print ("X_test.shape", X_test.shape)
- print ("X_test.size", len(X_test))
-
X_test = np.transpose(X_test, (3, 2, 0, 1))
-
- print ("X_test.shape", X_test.shape)
X_test = X_test.flatten()
X_test = X_test.astype(np.float32)
X_test.tofile(f)
@@ -117,11 +111,10 @@ def dumpConvWeights(file_name, X_test, N, C, H, W):
def dumpFcWeights(file_name, weights, H, W):
- print (weights.shape)
- print ("*DumpFcWeights")
- print("-min_val = ", np.amin(weights))
- print("-max_val = ", np.amax(weights))
-
+ print ("*Dumping FC weights to = ", file_name)
+
+ #print("-min_val = ", np.amin(weights))
+ #print("-max_val = ", np.amax(weights))
f = open(file_name, "wb")
for i in range(H):
@@ -134,11 +127,7 @@ def dumpFcWeights(file_name, weights, H, W):
def dumpFcBias(file_name, bias, W):
- print (bias.shape)
- print ("*DumpFcBias")
- print("-min_val = ", np.amin(bias))
- print("-max_val = ", np.amax(bias))
-
+ print ("*Dump Bias Weights = ", file_name)
f = open(file_name, "wb")
for i in range(W):
@@ -189,12 +178,12 @@ def reloadHPVMWeights(model, reload_dir, output_model, X_test, Y_test):
for i in range(len(model.layers)):
layer = model.layers[i]
layer_name = layer.name
- print ("*layer_name = ", layer_name)
+ #-- print ("*layer_name = ", layer_name)
if "conv" not in layer_name and "dense" not in layer_name:
continue
w_path = reload_dir + layer_name + "_w.bin"
- print ("** w_path = ", w_path)
+ #-- print ("** w_path = ", w_path)
w_arr = np.fromfile(w_path, dtype='float32')
b_path = reload_dir + layer_name + "_b.bin"
diff --git a/llvm/projects/keras/src/Benchmark.py b/llvm/projects/keras/src/Benchmark.py
index a275d103e253e842057dd0de9b98b1fecb625aeb..e83a78f4ea5d776a21ed6a6b47ccdb840f42c129 100644
--- a/llvm/projects/keras/src/Benchmark.py
+++ b/llvm/projects/keras/src/Benchmark.py
@@ -91,7 +91,7 @@ class Benchmark:
# Main call to ApproxHPVM-Keras Frontend
working_dir = translate_to_approxhpvm(model, self.hpvm_dir, X_test, Y_test, self.num_classes)
- print ("*** working_dir = ", working_dir)
+ #-- print ("*** working_dir = ", working_dir)
if len(argv) > 3 and argv[3] == "compile":
self.compileSource(working_dir)