pytorch to header convertor

c_reference/include/fastgrnn.h

@@ -21,8 +21,8 @@
  * @var       uRank        rank of U matrix
  * @var       Bg           pointer to bias for sigmoid
  * @var       Bh           pointer to bias for tanh
- * @var       zeta         first weight parameter for update from input from next step
- * @var       nu           second weight parameter for update from input from next step
+ * @var       sigmoid_zeta first weight parameter for update from input from next step
+ * @var       sigmoid_nu   second weight parameter for update from input from next step
  */
 typedef struct FastGRNN_LR_Params {
   float* mean;
@@ -35,8 +35,8 @@ typedef struct FastGRNN_LR_Params {
   unsigned uRank;
   float* Bg; 
   float* Bh;
-  float zeta;
-  float nu;
+  float sigmoid_zeta;
+  float sigmoid_nu;
 } FastGRNN_LR_Params;
 
 /**
@@ -82,8 +82,8 @@ int fastgrnn_lr(float* const hiddenState, unsigned hiddenDims,
  * @var       U            pointer U matrix
  * @var       Bg           pointer to bias for sigmoid
  * @var       Bh           pointer to bias for tanh
- * @var       zeta         first weight parameter for update from input from next step
- * @var       nu           second weight parameter for update from input from next step
+ * @var       sigmoid_zeta first weight parameter for update from input from next step
+ * @var       sigmoid_nu   second weight parameter for update from input from next step
  */
 typedef struct FastGRNN_Params {
   float* mean;
@@ -92,8 +92,8 @@ typedef struct FastGRNN_Params {
   float* U;
   float* Bg;
   float* Bh;
-  float zeta;
-  float nu;
+  float sigmoid_zeta;
+  float sigmoid_nu;
 } FastGRNN_Params;
 
 /**

c_reference/include/utils.h

@@ -36,14 +36,18 @@ void v_add(float scalar1, const float* const vec1,
   float scalar2, const float* const vec2,
   unsigned len, float* const ret);
 
-// point-wise vector division ret = vec2 / vec1 
+// point-wise vector multiplication ret = vec2 * vec1 
 void v_mult(const float* const vec1, const float* const vec2,
   unsigned len, float* const ret);
 
 // point-wise vector division ret = vec2 / vec1 
 void v_div(const float* const vec1, const float* const vec2,
   unsigned len, float* const ret);
 
+// Return squared Euclidean distance between vec1 and vec2
+float l2squared(const float* const vec1,
+  const float* const vec2, unsigned dim);
+
 // Return index with max value, if tied, return first tied index.
 unsigned argmax(const float* const vec, unsigned len);
 

c_reference/src/fastgrnn.c

@@ -19,7 +19,7 @@ int fastgrnn_lr(float* const hiddenState, unsigned hiddenDims,
   // #steps iterations of the RNN cell starting from hiddenState
   for (unsigned t = 0; t < steps; t++) {
     // Normalize the features
-    unsigned offset = backward ? steps - t : t;
+    unsigned offset = backward ? steps - 1 - t : t;
     if (normalize) {
       v_add(1.0f, input + offset * inputDims, -1.0f, tparams->mean + t * inputDims,
         inputDims, tbuffers->normFeatures);
@@ -45,7 +45,7 @@ int fastgrnn_lr(float* const hiddenState, unsigned hiddenDims,
     for (unsigned i = 0; i < hiddenDims; i++) {
       float gate = sigmoid(tbuffers->preComp[i] + tparams->Bg[i]);
       float update = tanh(tbuffers->preComp[i] + tparams->Bh[i]);
-      hiddenState[i] = gate * hiddenState[i] + (tparams->zeta * (1.0 - gate) + tparams->nu) * update;
+      hiddenState[i] = gate * hiddenState[i] + (tparams->sigmoid_zeta * (1.0 - gate) + tparams->sigmoid_nu) * update;
     }
   }
   return 0;
@@ -63,7 +63,7 @@ int fastgrnn(float* const hiddenState, unsigned hiddenDims,
 
   for (unsigned t = 0; t < steps; t++) {
     // Normalize the features
-    unsigned offset = backward ? steps - t : t;
+    unsigned offset = backward ? steps - 1 - t : t;
     if (normalize) {
       v_add(1.0f, input + offset * inputDims, -1.0f, tparams->mean + t * inputDims,
         inputDims, tbuffers->normFeatures);
@@ -86,7 +86,7 @@ int fastgrnn(float* const hiddenState, unsigned hiddenDims,
     for (unsigned i = 0; i < hiddenDims; i++) {
       float gate = sigmoid(tbuffers->preComp[i] + tparams->Bg[i]);
       float update = tanh(tbuffers->preComp[i] + tparams->Bh[i]);
-      hiddenState[i] = gate * hiddenState[i] + (tparams->zeta * (1.0 - gate) + tparams->nu) * update;
+      hiddenState[i] = gate * hiddenState[i] + (tparams->sigmoid_zeta * (1.0 - gate) + tparams->sigmoid_nu) * update;
     }
   }
   return 0;

c_reference/src/utils.c

@@ -90,6 +90,14 @@ void v_div(const float* const vec1, const float* const vec2,
     ret[i] = vec2[i] / vec1[i];
 }
 
+float l2squared(const float* const vec1,
+  const float* const vec2, unsigned dim) {
+  float sum = 0.0f;
+  for (unsigned i = 0; i < dim; i++)
+    sum += (vec1[i] - vec2[i]) * (vec1[i] - vec2[i]);
+  return sum;
+}
+
 unsigned argmax(const float* const vec, unsigned len) {
   unsigned maxId = 0;
   float maxScore = FLT_MIN;

c_reference/tests/Makefile

@@ -7,7 +7,7 @@ INCLUDE_DIR=../include
 SRC_DIR=../src
 IFLAGS = -I $(INCLUDE_DIR)
 
-all: test_fastgrnn_lr test_rnnpool
+all: test_fastgrnn_lr test_rnnpool test_rnnpool_face
 
 FASTGRNN_DIR=fastgrnn
 test_fastgrnn_lr: $(FASTGRNN_DIR)/test_fastgrnn_lr.c $(SRC_DIR)/utils.o $(SRC_DIR)/fastgrnn.o $(SRC_DIR)/classifier.o
@@ -17,6 +17,9 @@ RNNPOOL_DIR=rnnpool
 test_rnnpool: $(RNNPOOL_DIR)/test_rnnpool.c  $(SRC_DIR)/utils.o $(SRC_DIR)/fastgrnn.o $(SRC_DIR)/rnnpool.o
 	$(CC) -o $@ $^ $(IFLAGS) $(CFLAGS) -lm
 
+test_rnnpool_face: $(RNNPOOL_DIR)/test_rnnpool_face.c  $(SRC_DIR)/utils.o $(SRC_DIR)/fastgrnn.o $(SRC_DIR)/rnnpool.o
+	$(CC) -o $@ $^ $(IFLAGS) $(CFLAGS) -lm 
+
 
 .PHONY: clean cleanest
 

c_reference/tests/converters/ConvertNumpytoHeaderFilesRNNPool.py

@@ -0,0 +1,189 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT license.
+
+import numpy as np
+import argparse
+import os
+import sys
+
+def sigmoid(x):
+    return 1.0 / (1.0 + np.exp(-x))
+
+def saveTraces(tracesInputDir, tracesOutputDir):
+    if os.path.isdir(tracesOutputDir) is False:
+        try:
+            os.mkdir(tracesOutputDir)
+        except OSError:
+            print("Creation of the directory %s failed" % tracesOutputDir)
+            return
+    input_files = os.listdir(tracesInputDir+'/inputs/')
+
+    for file in input_files:
+        trace_input = np.load(tracesInputDir + "/inputs/" + file)
+        trace_output = np.load(tracesInputDir + "/outputs/" + file)
+
+        inputDims = trace_input.shape[-1]
+        patchDim = trace_input.shape[-2]
+        hiddenDims2 = int(trace_output.shape[0]/4)
+
+        f_input = open(tracesOutputDir+'/'+str(file)[:-4]+'_input.h','w')
+        f_output = open(tracesOutputDir+'/'+str(file)[:-4]+'_output.h','w')
+        f_input.write('#define INPUT_DIMS '+str(inputDims)+'\n#define PATCH_DIM '+ str(patchDim)+'\n\n')
+        f_output.write('#define HIDDEN_DIMS2 '+str(hiddenDims2)+'\n\n')
+
+        f_input.write('static float input[INPUT_DIMS * PATCH_DIM * PATCH_DIM] = ' + convertMatrixToVecString(trace_input) + ';')
+        f_output.write('static float output[4 * HIDDEN_DIMS2] = ' + convertMatrixToVecString(trace_output) + ';')
+
+        f_input.flush()
+        f_input.close()
+        f_output.flush()
+        f_output.close()
+
+
+
+
+def loadModel(modelDir):
+    model = {
+        "W1": np.transpose(np.load(modelDir + "/W1.npy")),
+        "W2": np.transpose(np.load(modelDir + "/W2.npy")),
+        "U1": np.transpose(np.load(modelDir + "/U1.npy")),
+        "U2": np.transpose(np.load(modelDir + "/U2.npy")),
+        "Bg1": np.load(modelDir + "/Bg1.npy"),
+        "Bh1": np.load(modelDir + "/Bh1.npy"),
+        "Bg2": np.load(modelDir + "/Bg2.npy"),
+        "Bh2": np.load(modelDir + "/Bh2.npy"),
+        "zeta1": sigmoid(np.load(modelDir + "/zeta1.npy")),
+        "nu1": sigmoid(np.load(modelDir + "/nu1.npy")),
+        "zeta2": sigmoid(np.load(modelDir + "/zeta2.npy")),
+        "nu2": sigmoid(np.load(modelDir + "/nu2.npy")),
+    }
+
+    return model
+
+
+def getArgs():
+    '''
+    Function to parse arguments for FastCells
+    '''
+    parser = argparse.ArgumentParser(
+        description='HyperParams for RNNPool inference')
+
+    parser.add_argument('-mdir', '--model-dir', required=False, default=None,
+                        help='Model directory containing' +
+                        'RNNPool model')
+
+    parser.add_argument('-tidir', '--trace-input-dir', required=False, default=None,
+                        help='Directory containing RnnPool input output numpy traces')
+
+    parser.add_argument('-todir', '--trace-output-dir', required=False, default=None,
+                        help='Output Directory for saving RnnPool input output .h traces')
+
+    parser.add_argument('-rnn1oF', '--rnn1-out-file', default=None,
+                        help='Give a output header file name for the model to dump rnn1 weights' +
+                        'default: stdout')
+
+    parser.add_argument('-rnn2oF', '--rnn2-out-file', default=None,
+                        help='Give a output header file name for the model to dump rnn2 weights' +
+                        'default: stdout')
+
+
+    return parser.parse_args()
+
+
+def saveReadableModel(modelDir, model):
+    if os.path.isdir(modelDir + '/ReadableModel') is False:
+        try:
+            os.mkdir(modelDir + '/ReadableModel')
+        except OSError:
+            print("Creation of the directory %s failed" %
+                  modelDir + '/ReadableModel')
+    currDir = modelDir + '/ReadableModel'
+
+    np.savetxt(currDir + "/W1.txt",
+               np.reshape(model["W1"], [1, -1]), delimiter=',')
+    np.savetxt(currDir + "/W2.txt",
+               np.reshape(model["W2"], [1, -1]), delimiter=',')
+    np.savetxt(currDir + "/U1.txt",
+               np.reshape(model["U1"], [1, -1]), delimiter=',')
+    np.savetxt(currDir + "/U2.txt",
+               np.reshape(model["U2"], [1, -1]), delimiter=',')
+    np.savetxt(currDir + "/Bg1.txt",
+               np.reshape(model["Bg1"], [1, -1]), delimiter=',')
+    np.savetxt(currDir + "/Bh1.txt",
+               np.reshape(model["Bh1"], [1, -1]), delimiter=',')
+    np.savetxt(currDir + "/Bg2.txt",
+               np.reshape(model["Bg2"], [1, -1]), delimiter=',')
+    np.savetxt(currDir + "/Bh2.txt",
+               np.reshape(model["Bh2"], [1, -1]), delimiter=',')
+    np.savetxt(currDir + "/zeta1.txt",
+               np.reshape(model["zeta1"], [1, -1]), delimiter=',')
+    np.savetxt(currDir + "/nu1.txt",
+               np.reshape(model["nu1"], [1, -1]), delimiter=',')
+    np.savetxt(currDir + "/zeta2.txt",
+               np.reshape(model["zeta2"], [1, -1]), delimiter=',')
+    np.savetxt(currDir + "/nu2.txt",
+               np.reshape(model["nu2"], [1, -1]), delimiter=',')
+
+    return currDir
+
+
+def convertMatrixToVecString(mat):
+    mat = str(np.reshape(mat, [1, -1])[0, :].tolist())
+    mat = '{' + mat[1:-1] + '}'
+    return mat
+
+
+def saveModelHeader(rnn1OutFile, rnn2OutFile, model):
+    rnn1OutFile = open(rnn1OutFile, 'w')
+    rnn2OutFile = open(rnn2OutFile, 'w')
+
+    print ("#define HIDDEN_DIMS1 8\n", file=rnn1OutFile)
+
+    print("static float W1[INPUT_DIMS * HIDDEN_DIMS1] = " + convertMatrixToVecString(model['W1']) + ";", file=rnn1OutFile)
+
+    print("static float U1[HIDDEN_DIMS1 * HIDDEN_DIMS1] = " + convertMatrixToVecString(model['U1']) + ";", file=rnn1OutFile)
+
+    print("static float Bg1[HIDDEN_DIMS1] = " + convertMatrixToVecString(model['Bg1']) + ";", file=rnn1OutFile)
+    print("static float Bh1[HIDDEN_DIMS1] = " +
+          convertMatrixToVecString(model['Bh1']) + ";\n", file=rnn1OutFile)
+
+    print("static float sigmoid_zeta1 = " + str(model['zeta1'][0][0]) + ";", file=rnn1OutFile)
+    print("static float sigmoid_nu1 = " + str(model['nu1'][0][0]) + ";\n", file=rnn1OutFile)
+
+
+    print("static float W2[HIDDEN_DIMS1 * HIDDEN_DIMS2] = " + convertMatrixToVecString(model['W2']) + ";", file=rnn2OutFile)
+
+    print("static float U2[HIDDEN_DIMS2 * HIDDEN_DIMS2] = " + convertMatrixToVecString(model['U2']) + ";\n", file=rnn2OutFile)
+
+    print("static float Bg2[HIDDEN_DIMS2] = " + convertMatrixToVecString(model['Bg2']) + ";", file=rnn2OutFile)
+    print("static float Bh2[HIDDEN_DIMS2] = " +
+          convertMatrixToVecString(model['Bh2']) + ";\n", file=rnn2OutFile)
+
+    print("static float sigmoid_zeta2 = " + str(model['zeta2'][0][0]) + ";", file=rnn2OutFile)
+    print("static float sigmoid_nu2 = " + str(model['nu2'][0][0]) + ";\n", file=rnn2OutFile)    
+
+
+    rnn1OutFile.flush()
+    rnn1OutFile.close()
+    rnn2OutFile.flush()
+    rnn2OutFile.close()
+
+
+def main():
+    args = getArgs()
+
+    if args.model_dir is not None:
+        model = loadModel(args.model_dir)
+        currDir = saveReadableModel(args.model_dir, model)
+        if args.rnn1_out_file is not None and args.rnn2_out_file is not None:
+            saveModelHeader(args.rnn1_out_file, args.rnn2_out_file, model)
+        else:
+            print('Not saving output header files as names are not specified')
+
+
+    if args.trace_input_dir is not None and args.trace_output_dir is not None:
+        saveTraces(args.trace_input_dir, args.trace_output_dir)
+
+
+
+main()

c_reference/tests/converters/README.md

@@ -0,0 +1,19 @@
+# Code for Converting Numpy Model Weights/Traces to Header Files
+
+
+## RNNPool
+
+Specify the path to model weights saved in numpy format in <your-numpy-model-weights-dir> and path to input output traces saved in numpy format in <your-numpy-traces-dir>. Specify folder to save converted header files for first rnn (which does the row-wise/column-wise traversal) of RNNPool in <your-path-to-rnn1-header-output-file> and that of second rnn (which does the bidirectional summarization of outputs of first rnn pass) in <your-path-to-rnn2-header-output-file>.
+
+```shell
+
+python3 ConvertNumpytoHeaderFilesRNNPool.py --model-dir <your-numpy-model-weights-dir> -tidir <your-numpy-traces-dir> -todir <your-traces-header-outputs-dir> -rnn1oF <your-path-to-rnn1-header-output-file> -rnn2oF <your-path-to-rnn2-header-output-file>                                 
+
+```
+
+eg.
+```shell
+
+python3 ConvertNumpytoHeaderFilesRNNPool.py --model-dir ./model_weights_face -tidir ./model_traces_face -todir ./traces_headers -rnn1oF rnn1.h -rnn2oF rnn2.h                                 
+
+```