''' Authors: Nishant Kumar. Copyright: Copyright (c) 2018 Microsoft Research Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. ''' import numpy import tensorflow as tf from tensorflow.tools.graph_transforms import TransformGraph def get_optimized_graph_def(output_tensor): #First save the graph def graph_def = tf.get_default_graph().as_graph_def() transforms = [ 'remove_nodes(op=Identity)', 'strip_unused_nodes', 'fold_batch_norms', 'fold_constants(ignore_errors=true)' # 'merge_duplicate_nodes', # Removing this otherwise in the output graph the topological ordering is broken - fix some other day #TODO_nishkum ] optimized_graph_def = TransformGraph(graph_def, [], [output_tensor.name], transforms) return optimized_graph_def def save_graph_metadata(output_tensor, sess, feed_dict): #First save the graph def graph_def = tf.get_default_graph().as_graph_def() transforms = [ 'remove_nodes(op=Identity)', 'strip_unused_nodes', 'fold_batch_norms', 'fold_constants(ignore_errors=true)' # 'merge_duplicate_nodes', # Removing this otherwise in the output graph the topological ordering is broken - fix some other day #TODO_nishkum ] optimized_graph_def = TransformGraph(graph_def, [], [output_tensor.name], transforms) with open('./graphDef.mtdata', 'w') as f: f.write(str(optimized_graph_def)) # Save size information for tensors on which output depends tensors_to_evaluate = [] tensors_to_evaluate_names = [] graph = tf.get_default_graph() for node in optimized_graph_def.node: cur_output = graph.get_operation_by_name(node.name).outputs[0] tensors_to_evaluate.append(cur_output) tensors_to_evaluate_names.append(node.name) tensors_evaluated = sess.run(tensors_to_evaluate, feed_dict) tensors_shape = list(map(lambda x : x.shape, tensors_evaluated)) # Write size info in a file with open('./sizeInfo.mtdata','w') as f: for ii, curr in enumerate(tensors_to_evaluate_names): curShape = tensors_shape[ii] f.write(tensors_to_evaluate_names[ii] + ' ') for dim in curShape: f.write(str(dim)+' ') f.write('\n') return optimized_graph_def def updateWeightsForBN(optimized_graph_def, sess, feed_dict): def findNodeInGraphDefWithName(graphDef, curName): for curNode in graphDef.node: if curNode.name == curName: return curNode return None print("Updating weights for BN...") graph = tf.get_default_graph() graphDef = optimized_graph_def for node in graphDef.node: if (node.op == 'FusedBatchNorm'): print("node.name = {0}".format(node.name)) gamma = graph.get_operation_by_name(node.input[1]).outputs[0] beta = graph.get_operation_by_name(node.input[2]).outputs[0] mu = graph.get_operation_by_name(node.input[3]).outputs[0] variance = graph.get_operation_by_name(node.input[4]).outputs[0] epsilon = 1e-5 # Taken from non-fused BN of TF rsigma = tf.rsqrt(variance + epsilon) sess.run(tf.assign(gamma, gamma*rsigma)) sess.run(tf.assign(beta, beta - gamma*mu)) sess.run(tf.assign(mu, tf.zeros(tf.shape(mu)))) sess.run(tf.assign(variance, tf.fill(tf.shape(variance), 1-epsilon))) print("BN weight updation done. Continuing...") def dumpImageDataInt(imgData, filename, scalingFac, writeMode): print("Dumping image data...") with open(filename, writeMode) as ff: for xx in numpy.nditer(imgData, order='C'): ff.write(str(int(xx * (1<<scalingFac))) + ' ') ff.write('\n\n') def dumpTrainedWeightsInt(sess, evalTensors, filename, scalingFac, writeMode, alreadyEvaluated=False): print("Dumping trained weights...") if alreadyEvaluated: finalParameters = evalTensors else: finalParameters = map(lambda x : sess.run(x), evalTensors) with open(filename, writeMode) as ff: for ii, curParameterVal in enumerate(finalParameters): for xx in numpy.nditer(curParameterVal, order='C'): ff.write(str(int(xx * (1<<scalingFac))) + ' ') ff.write('\n\n') def dumpTrainedWeightsFloat(sess, evalTensors, filename, writeMode, alreadyEvaluated=False): print("Dumping trained weights float...") if alreadyEvaluated: finalParameters = evalTensors else: finalParameters = map(lambda x : sess.run(x), evalTensors) with open(filename, writeMode) as ff: for ii, curParameterVal in enumerate(finalParameters): for xx in numpy.nditer(curParameterVal, order='C'): ff.write((str(xx)) + ' ') ff.write('\n\n') def dumpImgAndWeightsData(sess, imgData, evalTensors, filename, scalingFac, alreadyEvaluated=False): print("Starting to dump data...") dumpImageDataInt(imgData, filename, scalingFac, 'w') dumpTrainedWeightsInt(sess, evalTensors, filename, scalingFac, 'a', alreadyEvaluated=alreadyEvaluated) def dumpImgAndWeightsDataSeparate(sess, imgData, evalTensors, imgFileName, weightFileName, scalingFac, alreadyEvaluated=False): print("Starting to dump data...") dumpImageDataInt(imgData, imgFileName, scalingFac, 'w') dumpTrainedWeightsInt(sess, evalTensors, weightFileName, scalingFac, 'w', alreadyEvaluated=alreadyEvaluated)