from __future__ import division import warnings from extract_feature import build_model, run_image, get_img_feat # warnings.filterwarnings("ignore", category=FutureWarning) # warnings.filterwarnings("ignore", message="size changed") warnings.filterwarnings("ignore") import sys import os import time import math import random try: import Queue as queue except ImportError: import queue import threading import h5py import json import numpy as np import tensorflow as tf from termcolor import colored, cprint from config import config, loadDatasetConfig, parseArgs from preprocess import Preprocesser, bold, bcolored, writeline, writelist from model import MACnet from collections import defaultdict ############################################# loggers ############################################# # Writes log header to file def logInit(): with open(config.logFile(), "a+") as outFile: writeline(outFile, config.expName) headers = ["epoch", "trainAcc", "valAcc", "trainLoss", "valLoss"] if config.evalTrain: headers += ["evalTrainAcc", "evalTrainLoss"] if config.extra: if config.evalTrain: headers += ["thAcc", "thLoss"] headers += ["vhAcc", "vhLoss"] headers += ["time", "lr"] writelist(outFile, headers) # lr assumed to be last # Writes log record to file def logRecord(epoch, epochTime, lr, trainRes, evalRes, extraEvalRes): with open(config.logFile(), "a+") as outFile: record = [epoch, trainRes["acc"], evalRes["val"]["acc"], trainRes["loss"], evalRes["val"]["loss"]] if config.evalTrain: record += [evalRes["evalTrain"]["acc"], evalRes["evalTrain"]["loss"]] if config.extra: if config.evalTrain: record += [extraEvalRes["evalTrain"]["acc"], extraEvalRes["evalTrain"]["loss"]] record += [extraEvalRes["val"]["acc"], extraEvalRes["val"]["loss"]] record += [epochTime, lr] writelist(outFile, record) # Gets last logged epoch and learning rate def lastLoggedEpoch(): with open(config.logFile(), "r") as inFile: lastLine = list(inFile)[-1].split(",") epoch = int(lastLine[0]) lr = float(lastLine[-1]) return epoch, lr ################################## printing, output and analysis ################################## # Analysis by type analysisQuestionLims = [(0, 18), (19, float("inf"))] analysisProgramLims = [(0, 12), (13, float("inf"))] toArity = lambda instance: instance["programSeq"][-1].split("_", 1)[0] toType = lambda instance: instance["programSeq"][-1].split("_", 1)[1] def fieldLenIsInRange(field): return lambda instance, group: \ (len(instance[field]) >= group[0] and len(instance[field]) <= group[1]) # Groups instances based on a key def grouperKey(toKey): def grouper(instances): res = defaultdict(list) for instance in instances: res[toKey(instance)].append(instance) return res return grouper # Groups instances according to their match to condition def grouperCond(groups, isIn): def grouper(instances): res = {} for group in groups: res[group] = (instance for instance in instances if isIn(instance, group)) return res return grouper groupers = { "questionLength": grouperCond(analysisQuestionLims, fieldLenIsInRange("questionSeq")), "programLength": grouperCond(analysisProgramLims, fieldLenIsInRange("programSeq")), "arity": grouperKey(toArity), "type": grouperKey(toType) } # Computes average def avg(instances, field): if len(instances) == 0: return 0.0 return sum(instances[field]) / len(instances) # Prints analysis of questions loss and accuracy by their group def printAnalysis(res): if config.analysisType != "": print("Analysis by {type}".format(type=config.analysisType)) groups = groupers[config.analysisType](res["preds"]) for key in groups: instances = groups[key] avgLoss = avg(instances, "loss") avgAcc = avg(instances, "acc") num = len(instances) print("Group {key}: Loss: {loss}, Acc: {acc}, Num: {num}".format(key, avgLoss, avgAcc, num)) # Print results for a tier def printTierResults(tierName, res, color): if res is None: return print("{tierName} Loss: {loss}, {tierName} accuracy: {acc}".format(tierName=tierName, loss=bcolored(res["loss"], color), acc=bcolored(res["acc"], color))) printAnalysis(res) # Prints dataset results (for several tiers) def printDatasetResults(trainRes, evalRes): printTierResults("Training", trainRes, "magenta") printTierResults("Training EMA", evalRes["evalTrain"], "red") printTierResults("Validation", evalRes["val"], "cyan") # Writes predictions for several tiers def writePreds(preprocessor, evalRes): preprocessor.writePreds(evalRes, "_") ############################################# session ############################################# # Initializes TF session. Sets GPU memory configuration. def setSession(): sessionConfig = tf.ConfigProto(allow_soft_placement=True, log_device_placement=False) if config.allowGrowth: sessionConfig.gpu_options.allow_growth = True if config.maxMemory < 1.0: sessionConfig.gpu_options.per_process_gpu_memory_fraction = config.maxMemory return sessionConfig ############################################## savers ############################################# # Initializes savers (standard, optional exponential-moving-average and optional for subset of variables) def setSavers(model): saver = tf.train.Saver(max_to_keep=config.weightsToKeep) subsetSaver = None if config.saveSubset: isRelevant = lambda var: any(s in var.name for s in config.varSubset) relevantVars = [var for var in tf.global_variables() if isRelevant(var)] subsetSaver = tf.train.Saver(relevantVars, max_to_keep=config.weightsToKeep, allow_empty=True) emaSaver = None if config.useEMA: emaSaver = tf.train.Saver(model.emaDict, max_to_keep=config.weightsToKeep) return { "saver": saver, "subsetSaver": subsetSaver, "emaSaver": emaSaver } ################################### restore / initialize weights ################################## # Restores weights of specified / last epoch if on restore mod. # Otherwise, initializes weights. def loadWeights(sess, saver, init): if config.restoreEpoch > 0 or config.restore: # restore last epoch only if restoreEpoch isn't set if config.restoreEpoch == 0: # restore last logged epoch config.restoreEpoch, config.lr = lastLoggedEpoch() print(bcolored("Restoring epoch {} and lr {}".format(config.restoreEpoch, config.lr), "cyan")) print(bcolored("Restoring weights", "blue")) print(config.weightsFile(config.restoreEpoch)) saver.restore(sess, config.weightsFile(config.restoreEpoch)) epoch = config.restoreEpoch else: print(bcolored("Initializing weights", "blue")) sess.run(init) logInit() epoch = 0 return epoch ###################################### training / evaluation ###################################### # Chooses data to train on (main / extra) data. def chooseTrainingData(data): trainingData = data["main"]["train"] alterData = None if config.extra: if config.trainExtra: if config.extraVal: trainingData = data["extra"]["val"] else: trainingData = data["extra"]["train"] if config.alterExtra: alterData = data["extra"]["train"] return trainingData, alterData #### evaluation # Runs evaluation on train / val / test datasets. def runEvaluation(sess, model, data, epoch, evalTrain=True, evalTest=False, getAtt=None): if getAtt is None: getAtt = config.getAtt res = {"evalTrain": None, "val": None, "test": None} if data is not None: if evalTrain and config.evalTrain: res["evalTrain"] = runEpoch(sess, model, data["evalTrain"], train=False, epoch=epoch, getAtt=getAtt) res["val"] = runEpoch(sess, model, data["val"], train=False, epoch=epoch, getAtt=getAtt) if evalTest or config.test: res["test"] = runEpoch(sess, model, data["test"], train=False, epoch=epoch, getAtt=getAtt) return res ## training conditions (comparing current epoch result to prior ones) def improveEnough(curr, prior, lr): prevRes = prior["prev"]["res"] currRes = curr["res"] if prevRes is None: return True prevTrainLoss = prevRes["train"]["loss"] currTrainLoss = currRes["train"]["loss"] lossDiff = prevTrainLoss - currTrainLoss notImprove = ((lossDiff < 0.015 and prevTrainLoss < 0.5 and lr > 0.00002) or \ (lossDiff < 0.008 and prevTrainLoss < 0.15 and lr > 0.00001) or \ (lossDiff < 0.003 and prevTrainLoss < 0.10 and lr > 0.000005)) # (prevTrainLoss < 0.2 and config.lr > 0.000015) return not notImprove def better(currRes, bestRes): return currRes["val"]["acc"] > bestRes["val"]["acc"] ############################################## data ############################################### #### instances and batching # Trims sequences based on their max length. def trim2DVectors(vectors, vectorsLengths): maxLength = np.max(vectorsLengths) return vectors[:, :maxLength] # Trims batch based on question length. def trimData(data): data["questions"] = trim2DVectors(data["questions"], data["questionLengths"]) return data # Gets batch / bucket size. def getLength(data): return len(data["instances"]) # Selects the data entries that match the indices. def selectIndices(data, indices): def select(field, indices): if type(field) is np.ndarray: return field[indices] if type(field) is list: return [field[i] for i in indices] else: return field selected = {k: select(d, indices) for k, d in data.items()} return selected # Batches data into a a list of batches of batchSize. # Shuffles the data by default. def getBatches(data, batchSize=None, shuffle=True): batches = [] dataLen = getLength(data) if batchSize is None or batchSize > dataLen: batchSize = dataLen indices = np.arange(dataLen) if shuffle: np.random.shuffle(indices) for batchStart in range(0, dataLen, batchSize): batchIndices = indices[batchStart: batchStart + batchSize] # if len(batchIndices) == batchSize? if len(batchIndices) >= config.gpusNum: batch = selectIndices(data, batchIndices) batches.append(batch) # batchesIndices.append((data, batchIndices)) return batches #### image batches # Opens image files. def openImageFiles(images): images["imagesFile"] = h5py.File(images["imagesFilename"], "r") images["imagesIds"] = None if config.dataset == "NLVR": with open(images["imageIdsFilename"], "r") as imageIdsFile: images["imagesIds"] = json.load(imageIdsFile) # Closes image files. def closeImageFiles(images): images["imagesFile"].close() # Loads an images from file for a given data batch. def loadImageBatch(images, batch): imagesFile = images["imagesFile"] id2idx = images["imagesIds"] toIndex = lambda imageId: imageId if id2idx is not None: toIndex = lambda imageId: id2idx[imageId] imageBatch = np.stack([imagesFile["features"][toIndex(imageId)] for imageId in batch["imageIds"]], axis=0) return {"images": imageBatch, "imageIds": batch["imageIds"]} # Loads images for several num batches in the batches list from start index. def loadImageBatches(images, batches, start, num): batches = batches[start: start + num] return [loadImageBatch(images, batch) for batch in batches] #### data alternation # Alternates main training batches with extra data. def alternateData(batches, alterData, dataLen): alterData = alterData["data"][0] # data isn't bucketed for altered data # computes number of repetitions needed = math.ceil(len(batches) / config.alterNum) print(bold("Extra batches needed: %d") % needed) perData = math.ceil(getLength(alterData) / config.batchSize) print(bold("Batches per extra data: %d") % perData) repetitions = math.ceil(needed / perData) print(bold("reps: %d") % repetitions) # make alternate batches alterBatches = [] for _ in range(repetitions): repBatches = getBatches(alterData, batchSize=config.batchSize) random.shuffle(repBatches) alterBatches += repBatches print(bold("Batches num: %d") + len(alterBatches)) # alternate data with extra data curr = len(batches) - 1 for alterBatch in alterBatches: if curr < 0: # print(colored("too many" + str(curr) + " " + str(len(batches)),"red")) break batches.insert(curr, alterBatch) dataLen += getLength(alterBatch) curr -= config.alterNum return batches, dataLen ############################################ threading ############################################ imagesQueue = queue.Queue(maxsize=20) # config.tasksNum inQueue = queue.Queue(maxsize=1) outQueue = queue.Queue(maxsize=1) # Runs a worker thread(s) to load images while training . class StoppableThread(threading.Thread): # Thread class with a stop() method. The thread itself has to check # regularly for the stopped() condition. def __init__(self, images, batches): # i super(StoppableThread, self).__init__() # self.i = i self.images = images self.batches = batches self._stop_event = threading.Event() # def __init__(self, args): # super(StoppableThread, self).__init__(args = args) # self._stop_event = threading.Event() # def __init__(self, target, args): # super(StoppableThread, self).__init__(target = target, args = args) # self._stop_event = threading.Event() def stop(self): self._stop_event.set() def stopped(self): return self._stop_event.is_set() def run(self): while not self.stopped(): try: batchNum = inQueue.get(timeout=60) nextItem = loadImageBatches(self.images, self.batches, batchNum, int(config.taskSize / 2)) outQueue.put(nextItem) # inQueue.task_done() except: pass # print("worker %d done", self.i) def loaderRun(images, batches): batchNum = 0 # if config.workers == 2: # worker = StoppableThread(images, batches) # i, # worker.daemon = True # worker.start() # while batchNum < len(batches): # inQueue.put(batchNum + int(config.taskSize / 2)) # nextItem1 = loadImageBatches(images, batches, batchNum, int(config.taskSize / 2)) # nextItem2 = outQueue.get() # nextItem = nextItem1 + nextItem2 # assert len(nextItem) == min(config.taskSize, len(batches) - batchNum) # batchNum += config.taskSize # imagesQueue.put(nextItem) # worker.stop() # else: while batchNum < len(batches): nextItem = loadImageBatches(images, batches, batchNum, config.taskSize) assert len(nextItem) == min(config.taskSize, len(batches) - batchNum) batchNum += config.taskSize imagesQueue.put(nextItem) # print("manager loader done") ########################################## stats tracking ######################################### # Computes exponential moving average. def emaAvg(avg, value): if avg is None: return value emaRate = 0.98 return avg * emaRate + value * (1 - emaRate) # Initializes training statistics. def initStats(): return { "totalBatches": 0, "totalData": 0, "totalLoss": 0.0, "totalCorrect": 0, "loss": 0.0, "acc": 0.0, "emaLoss": None, "emaAcc": None, } # Updates statistics with training results of a batch def updateStats(stats, res, batch): stats["totalBatches"] += 1 stats["totalData"] += getLength(batch) stats["totalLoss"] += res["loss"] stats["totalCorrect"] += res["correctNum"] stats["loss"] = stats["totalLoss"] / stats["totalBatches"] stats["acc"] = stats["totalCorrect"] / stats["totalData"] stats["emaLoss"] = emaAvg(stats["emaLoss"], res["loss"]) stats["emaAcc"] = emaAvg(stats["emaAcc"], res["acc"]) return stats # auto-encoder ae = {:2.4f} autoEncLoss, # Translates training statistics into a string to print def statsToStr(stats, res, epoch, batchNum, dataLen, startTime): formatStr = "\reb {epoch},{batchNum} ({dataProcessed} / {dataLen:5d}), " + \ "t = {time} ({loadTime:2.2f}+{trainTime:2.2f}), " + \ "lr {lr}, l = {loss}, a = {acc}, avL = {avgLoss}, " + \ "avA = {avgAcc}, g = {gradNorm:2.4f}, " + \ "emL = {emaLoss:2.4f}, emA = {emaAcc:2.4f}; " + \ "{expname}" # {machine}/{gpu}" s_epoch = bcolored("{:2d}".format(epoch), "green") s_batchNum = "{:3d}".format(batchNum) s_dataProcessed = bcolored("{:5d}".format(stats["totalData"]), "green") s_dataLen = dataLen s_time = bcolored("{:2.2f}".format(time.time() - startTime), "green") s_loadTime = res["readTime"] s_trainTime = res["trainTime"] s_lr = bold(config.lr) s_loss = bcolored("{:2.4f}".format(res["loss"]), "blue") s_acc = bcolored("{:2.4f}".format(res["acc"]), "blue") s_avgLoss = bcolored("{:2.4f}".format(stats["loss"]), "blue") s_avgAcc = bcolored("{:2.4f}".format(stats["acc"]), "red") s_gradNorm = res["gradNorm"] s_emaLoss = stats["emaLoss"] s_emaAcc = stats["emaAcc"] s_expname = config.expName # s_machine = bcolored(config.dataPath[9:11],"green") # s_gpu = bcolored(config.gpus,"green") return formatStr.format(epoch=s_epoch, batchNum=s_batchNum, dataProcessed=s_dataProcessed, dataLen=s_dataLen, time=s_time, loadTime=s_loadTime, trainTime=s_trainTime, lr=s_lr, loss=s_loss, acc=s_acc, avgLoss=s_avgLoss, avgAcc=s_avgAcc, gradNorm=s_gradNorm, emaLoss=s_emaLoss, emaAcc=s_emaAcc, expname=s_expname) # machine = s_machine, gpu = s_gpu) # collectRuntimeStats, writer = None, ''' Runs an epoch with model and session over the data. 1. Batches the data and optionally mix it with the extra alterData. 2. Start worker threads to load images in parallel to training. 3. Runs model for each batch, and gets results (e.g. loss, accuracy). 4. Updates and prints statistics based on batch results. 5. Once in a while (every config.saveEvery), save weights. Args: sess: TF session to run with. model: model to process data. Has runBatch method that process a given batch. (See model.py for further details). data: data to use for training/evaluation. epoch: epoch number. saver: TF saver to save weights calle: a method to call every number of iterations (config.calleEvery) alterData: extra data to mix with main data while training. getAtt: True to return model attentions. ''' def main(question, image): with open(config.configFile(), "a+") as outFile: json.dump(vars(config), outFile) # set gpus if config.gpus != "": config.gpusNum = len(config.gpus.split(",")) os.environ["CUDA_VISIBLE_DEVICES"] = config.gpus tf.logging.set_verbosity(tf.logging.ERROR) # process data print(bold("Preprocess data...")) start = time.time() preprocessor = Preprocesser() cnn_model = build_model() imageData = get_img_feat(cnn_model, image) qData, embeddings, answerDict = preprocessor.preprocessData(question) data = {'data': qData, 'image': imageData} print("took {} seconds".format(bcolored("{:.2f}".format(time.time() - start), "blue"))) # build model print(bold("Building model...")) start = time.time() model = MACnet(embeddings, answerDict) print("took {} seconds".format(bcolored("{:.2f}".format(time.time() - start), "blue"))) # initializer init = tf.global_variables_initializer() # savers savers = setSavers(model) saver, emaSaver = savers["saver"], savers["emaSaver"] # sessionConfig sessionConfig = setSession() with tf.Session(config=sessionConfig) as sess: # ensure no more ops are added after model is built sess.graph.finalize() # restore / initialize weights, initialize epoch variable epoch = loadWeights(sess, saver, init) print(epoch) start = time.time() if epoch > 0: if config.useEMA: emaSaver.restore(sess, config.weightsFile(epoch)) else: saver.restore(sess, config.weightsFile(epoch)) evalRes = model.runBatch(sess, data['data'], data['image'], False) print("took {:.2f} seconds".format(time.time() - start)) print(evalRes) if __name__ == '__main__': parseArgs() loadDatasetConfig[config.dataset]() question = 'How many text objects are located at the bottom side of table?' imagePath = './mac-layoutLM-sample/PDF_val_64.png' main(question, imagePath)