#!/usr/bin/env python3 # # Copyright 2017 Robert Csordas. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # # ============================================================================== import functools import os import torch.utils.data import Utils.Debug as debug from Dataset.Bitmap.AssociativeRecall import AssociativeRecall from Dataset.Bitmap.BitmapTaskRepeater import BitmapTaskRepeater from Dataset.Bitmap.KeyValue import KeyValue from Dataset.Bitmap.CopyTask import CopyData from Dataset.Bitmap.KeyValue2Way import KeyValue2Way from Dataset.NLP.bAbi import bAbiDataset from Models.DNC import DNC, LSTMController, FeedforwardController from Utils import Visdom from Utils.ArgumentParser import ArgumentParser from Utils.Index import index_by_dim from Utils.Saver import Saver, GlobalVarSaver, StateSaver from Utils.Collate import MetaCollate from Utils import gpu_allocator from Dataset.NLP.NLPTask import NLPTask from tqdm import tqdm from Visualize.preview import preview from Utils.timer import OnceEvery from Utils import Seed import time import sys import signal import math from Utils import Profile Profile.ENABLED=False def main(): global i global loss_sum global running parser = ArgumentParser() parser.add_argument("-bit_w", type=int, default=8, help="Bit vector length for copy task") parser.add_argument("-block_w", type=int, default=3, help="Block width to associative recall task") parser.add_argument("-len", type=str, default="4", help="Sequence length for copy task", parser=lambda x: [int(a) for a in x.split("-")]) parser.add_argument("-repeat", type=str, default="1", help="Sequence length for copy task", parser=lambda x: [int(a) for a in x.split("-")]) parser.add_argument("-batch_size", type=int, default=16, help="Sequence length for copy task") parser.add_argument("-n_subbatch", type=str, default="auto", help="Average this much forward passes to a backward pass") parser.add_argument("-max_input_count_per_batch", type=int, default=6000, help="Max batch_size*len that can fit into memory") parser.add_argument("-lr", type=float, default=0.0001, help="Learning rate") parser.add_argument("-wd", type=float, default=1e-5, help="Weight decay") parser.add_argument("-optimizer", type=str, default="rmsprop", help="Optimizer algorithm") parser.add_argument("-name", type=str, help="Save training to this directory") parser.add_argument("-preview_interval", type=int, default=10, help="Show preview every nth iteration") parser.add_argument("-info_interval", type=int, default=10, help="Show info every nth iteration") parser.add_argument("-save_interval", type=int, default=500, help="Save network every nth iteration") parser.add_argument("-masked_lookup", type=bool, default=1, help="Enable masking in content lookups") parser.add_argument("-visport", type=int, default=-1, help="Port to run Visdom server on. -1 to disable") parser.add_argument("-gpu", default="auto", type=str, help="Run on this GPU.") parser.add_argument("-debug", type=bool, default=1, help="Enable debugging") parser.add_argument("-task", type=str, default="copy", help="Task to learn") parser.add_argument("-mem_count", type=int, default=16, help="Number of memory cells") parser.add_argument("-data_word_size", type=int, default=128, help="Memory word size") parser.add_argument("-n_read_heads", type=int, default=1, help="Number of read heads") parser.add_argument("-layer_sizes", type=str, default="256", help="Controller layer sizes. Separate with ,. For example 512,256,256", parser=lambda x: [int(y) for y in x.split(",") if y]) parser.add_argument("-debug_log", type=bool, default=0, help="Enable debug log") parser.add_argument("-controller_type", type=str, default="lstm", help="Controller type: lstm or linear") parser.add_argument("-lstm_use_all_outputs", type=bool, default=1, help="Use all LSTM outputs as controller output vs use only the last layer") parser.add_argument("-momentum", type=float, default=0.9, help="Momentum for optimizer") parser.add_argument("-embedding_size", type=int, default=256, help="Size of word embedding for NLP tasks") parser.add_argument("-test_interval", type=int, default=10000, help="Run test in this interval") parser.add_argument("-dealloc_content", type=bool, default=1, help="Deallocate memory content, unlike DNC, which leaves it unchanged, just decreases the usage counter, causing problems with lookup") parser.add_argument("-sharpness_control", type=bool, default=1, help="Distribution sharpness control for forward and backward links") parser.add_argument("-think_steps", type=int, default=0, help="Iddle steps before requiring the answer (for bAbi)") parser.add_argument("-dump_profile", type=str, save=False) parser.add_argument("-test_on_start", default="0", save=False) parser.add_argument("-dump_heatmaps", default=False, save=False) parser.add_argument("-test_batch_size", default=16) parser.add_argument("-mask_min", default=0.0) parser.add_argument("-load", type=str, save=False) parser.add_argument("-dataset_path", type=str, default="none", parser=ArgumentParser.str_or_none(), help="Specify babi path manually") parser.add_argument("-babi_train_tasks", type=str, default="none", parser=ArgumentParser.list_or_none(type=str), help="babi task list to use for training") parser.add_argument("-babi_test_tasks", type=str, default="none", parser=ArgumentParser.list_or_none(type=str), help="babi task list to use for testing") parser.add_argument("-babi_train_sets", type=str, default="train", parser=ArgumentParser.list_or_none(type=str), help="babi train sets to use") parser.add_argument("-babi_test_sets", type=str, default="test", parser=ArgumentParser.list_or_none(type=str), help="babi test sets to use") parser.add_argument("-noargsave", type=bool, default=False, help="Do not save modified arguments", save=False) parser.add_argument("-demo", type=bool, default=False, help="Do a single step with fixed seed", save=False) parser.add_argument("-exit_after", type=int, help="Exit after this amount of steps. Useful for debugging.", save=False) parser.add_argument("-grad_clip", type=float, default=10.0, help="Max gradient norm") parser.add_argument("-clip_controller", type=float, default=20.0, help="Max gradient norm") parser.add_argument("-print_test", default=False, save=False) parser.add_profile([ ArgumentParser.Profile("babi", { "preview_interval": 10, "save_interval": 500, "task": "babi", "mem_count": 256, "data_word_size": 64, "n_read_heads": 4, "layer_sizes": "256", "controller_type": "lstm", "lstm_use_all_outputs": True, "momentum": 0.9, "embedding_size": 128, "test_interval": 5000, "think_steps": 3, "batch_size": 2 }, include=["dnc-msd"]), ArgumentParser.Profile("repeat_copy", { "bit_w": 8, "repeat": "1-8", "len": "2-14", "task": "copy", "think_steps": 1, "preview_interval": 10, "info_interval": 10, "save_interval": 100, "data_word_size": 16, "layer_sizes": "32", "n_subbatch": 1, "controller_type": "lstm", }), ArgumentParser.Profile("repeat_copy_simple", { "repeat": "1-3", }, include="repeat_copy"), ArgumentParser.Profile("dnc", { "masked_lookup": False, "sharpness_control": False, "dealloc_content": False }), ArgumentParser.Profile("dnc-m", { "masked_lookup": True, "sharpness_control": False, "dealloc_content": False }), ArgumentParser.Profile("dnc-s", { "masked_lookup": False, "sharpness_control": True, "dealloc_content": False }), ArgumentParser.Profile("dnc-d", { "masked_lookup": False, "sharpness_control": False, "dealloc_content": True }), ArgumentParser.Profile("dnc-md", { "masked_lookup": True, "sharpness_control": False, "dealloc_content": True }), ArgumentParser.Profile("dnc-ms", { "masked_lookup": True, "sharpness_control": True, "dealloc_content": False }), ArgumentParser.Profile("dnc-sd", { "masked_lookup": False, "sharpness_control": True, "dealloc_content": True }), ArgumentParser.Profile("dnc-msd", { "masked_lookup": True, "sharpness_control": True, "dealloc_content": True }), ArgumentParser.Profile("keyvalue", { "repeat": "1", "len": "2-16", "mem_count": 16, "task": "keyvalue", "think_steps": 1, "preview_interval": 10, "info_interval": 10, "data_word_size": 32, "bit_w": 12, "save_interval": 1000, "layer_sizes": "32" }), ArgumentParser.Profile("keyvalue2way", { "task": "keyvalue2way", }, include="keyvalue"), ArgumentParser.Profile("associative_recall",{ "task": "recall", "bit_w": 8, "len": "2-16", "mem_count": 64, "data_word_size": 32, "n_read_heads": 1, "layer_sizes": "128", "controller_type": "lstm", "lstm_use_all_outputs": 1, "think_steps": 1, "mask_min": 0.1, "info_interval": 10, "save_interval": 1000, "preview_interval": 10, "n_subbatch": 1, }) ]) opt = parser.parse() assert opt.name is not None, "Training dir (-name parameter) not given" opt = parser.sync(os.path.join(opt.name, "args.json"), save=not opt.noargsave) if opt.demo: Seed.fix() os.makedirs(os.path.join(opt.name,"save"), exist_ok=True) os.makedirs(os.path.join(opt.name,"preview"), exist_ok=True) gpu_allocator.use_gpu(opt.gpu) debug.enableDebug = opt.debug_log if opt.visport>0: Visdom.start(opt.visport) Visdom.Text("Name").set(opt.name) class LengthHackSampler: def __init__(self, batch_size, length): self.length = length self.batch_size = batch_size def __iter__(self): while True: len = self.length() if callable(self.length) else self.length yield [len] * self.batch_size def __len__(self): return 0x7FFFFFFF embedding = None test_set = None curriculum = None loader_reset = False if opt.task=="copy": dataset = CopyData(bit_w=opt.bit_w) in_size = opt.bit_w + 1 out_size = in_size elif opt.task=="recall": dataset = AssociativeRecall(bit_w=opt.bit_w, block_w=opt.block_w) in_size = opt.bit_w + 2 out_size = in_size elif opt.task=="keyvalue": assert opt.bit_w % 2==0, "Key-value datasets works only with even bit_w" dataset = KeyValue(bit_w=opt.bit_w) in_size = opt.bit_w + 1 out_size = opt.bit_w//2 elif opt.task=="keyvalue2way": assert opt.bit_w % 2==0, "Key-value datasets works only with even bit_w" dataset = KeyValue2Way(bit_w=opt.bit_w) in_size = opt.bit_w + 2 out_size = opt.bit_w//2 elif opt.task=="babi": dataset = bAbiDataset(think_steps=opt.think_steps, dir_name=opt.dataset_path) test_set = bAbiDataset(think_steps=opt.think_steps, dir_name=opt.dataset_path, name="test") dataset.use(opt.babi_train_tasks, opt.babi_train_sets) in_size = opt.embedding_size print("bAbi: loaded total of %d sequences." % len(dataset)) test_set.use(opt.babi_test_tasks, opt.babi_test_sets) out_size = len(dataset.vocabulary) print("bAbi: using %d sequences for training, %d for testing" % (len(dataset), len(test_set))) else: assert False, "Invalid task: %s" % opt.task if opt.task in ["babi"]: data_loader = torch.utils.data.DataLoader(dataset, batch_size=opt.batch_size, num_workers=4, pin_memory=True, shuffle=True, collate_fn=MetaCollate()) test_loader = torch.utils.data.DataLoader(test_set, batch_size=opt.test_batch_size, num_workers=opt.test_batch_size, pin_memory=True, shuffle=False, collate_fn=MetaCollate()) if test_set is not None else None else: dataset = BitmapTaskRepeater(dataset) data_loader = torch.utils.data.DataLoader(dataset, batch_sampler=LengthHackSampler(opt.batch_size, BitmapTaskRepeater.key_sampler(opt.len, opt.repeat)), num_workers=1, pin_memory=True) if opt.controller_type == "lstm": controller_constructor = functools.partial(LSTMController, out_from_all_layers=opt.lstm_use_all_outputs) elif opt.controller_type == "linear": controller_constructor = FeedforwardController else: assert False, "Invalid controller: %s" % opt.controller_type model = DNC(in_size, out_size, opt.data_word_size, opt.mem_count, opt.n_read_heads, controller_constructor(opt.layer_sizes), batch_first=True, mask=opt.masked_lookup, dealloc_content=opt.dealloc_content, link_sharpness_control=opt.sharpness_control, mask_min=opt.mask_min, clip_controller=opt.clip_controller) params = [ {'params': [p for n, p in model.named_parameters() if not n.endswith(".bias")]}, {'params': [p for n, p in model.named_parameters() if n.endswith(".bias")], 'weight_decay': 0} ] device = torch.device('cuda') if opt.gpu!="none" else torch.device("cpu") print("DEVICE: ", device) if isinstance(dataset, NLPTask): embedding = torch.nn.Embedding(len(dataset.vocabulary), opt.embedding_size).to(device) params.append({'params': embedding.parameters(), 'weight_decay': 0}) if opt.optimizer=="sgd": optimizer = torch.optim.SGD(params, lr=opt.lr, weight_decay=opt.wd, momentum=opt.momentum) elif opt.optimizer=="adam": optimizer = torch.optim.Adam(params, lr=opt.lr, weight_decay=opt.wd) elif opt.optimizer == "rmsprop": optimizer = torch.optim.RMSprop(params, lr=opt.lr, weight_decay=opt.wd, momentum=opt.momentum, eps=1e-10) else: assert "Invalid optimizer: %s" % opt.optimizer n_params = sum([sum([t.numel() for t in d['params']]) for d in params]) print("Number of parameters: %d" % n_params) model = model.to(device) if embedding is not None and hasattr(embedding, "to"): embedding = embedding.to(device) i=0 loss_sum = 0 loss_plot = Visdom.Plot2D("loss", store_interval=opt.info_interval, xlabel="iterations", ylabel="loss") if curriculum is not None: curriculum_plot = Visdom.Plot2D("curriculum lesson" + (" (last %d)" % (curriculum.n_lessons-1) if curriculum.n_lessons is not None else ""), xlabel="iterations", ylabel="lesson") curriculum_accuracy = Visdom.Plot2D("curriculum accuracy", xlabel="iterations", ylabel="accuracy") saver = Saver(os.path.join(opt.name, "save"), short_interval=opt.save_interval) saver.register("model", StateSaver(model)) saver.register("optimizer", StateSaver(optimizer)) saver.register("i", GlobalVarSaver("i")) saver.register("loss_sum", GlobalVarSaver("loss_sum")) saver.register("loss_plot", StateSaver(loss_plot)) saver.register("dataset", StateSaver(dataset)) if test_set: saver.register("test_set", StateSaver(test_set)) if curriculum is not None: saver.register("curriculum", StateSaver(curriculum)) saver.register("curriculum_plot", StateSaver(curriculum_plot)) saver.register("curriculum_accuracy", StateSaver(curriculum_accuracy)) if isinstance(dataset, NLPTask): saver.register("word_embeddings", StateSaver(embedding)) elif embedding is not None: saver.register("embeddings", StateSaver(embedding)) if not saver.load(opt.load): model.reset_parameters() if embedding is not None: embedding.reset_parameters() visualizers = {} debug_schemas={ "read_head" : { "list_dim" : 2 }, "temporal_links/forward_dists" : { "list_dim" : 2 }, "temporal_links/backward_dists" : { "list_dim" : 2 } } def plot_debug(debug, prefix="", schema={}): if debug is None: return for k, v in debug.items(): curr_name = prefix+k if curr_name in debug_schemas: curr_schema = schema.copy() curr_schema.update(debug_schemas[curr_name]) else: curr_schema = schema if isinstance(v, dict): plot_debug(v, curr_name+"/", curr_schema) continue data = v[0] if curr_schema.get("list_dim",-1) > 0: if data.ndim != 3: print("WARNING: unknown data shape for array display: %s, tensor %s" % (data.shape, curr_name)) continue n_steps = data.shape[curr_schema["list_dim"]-1] if curr_name not in visualizers: visualizers[curr_name] = [Visdom.Heatmap(curr_name+"_%d" % i, dumpdir=os.path.join(opt.name, "preview") if opt.dump_heatmaps else None) for i in range(n_steps)] for i in range(n_steps): visualizers[curr_name][i].draw(index_by_dim(data, curr_schema["list_dim"]-1, i)) else: if data.ndim != 2: print("WARNING: unknown data shape for simple display: %s, tensor %s" % (data.shape, curr_name)) continue if curr_name not in visualizers: visualizers[curr_name] = Visdom.Heatmap(curr_name, dumpdir=os.path.join(opt.name, "preview") if opt.dump_heatmaps else None) visualizers[curr_name].draw(data) def run_model(input, debug=None): if isinstance(dataset, NLPTask): input = embedding(input["input"]) else: input = input["input"] * 2.0 - 1.0 return model(input, debug=debug) def multiply_grads(params, mul): if mul==1: return for pa in params: for p in pa["params"]: p.grad.data *= mul def test(): if test_set is None: return print("TESTING...") start_time=time.time() t = test_set.start_test() with torch.no_grad(): for data in tqdm(test_loader): data = {k: v.to(device) if torch.is_tensor(v) else v for k, v in data.items()} if hasattr(dataset, "prepare"): data = dataset.prepare(data) net_out = run_model(data) test_set.veify_result(t, data, net_out) test_set.show_test_results(i, t) print("Test done in %gs" % (time.time() - start_time)) if opt.test_on_start.lower() in ["on", "1", "true", "quit"]: test() if opt.test_on_start.lower() == "quit": saver.write(i) sys.exit(-1) if opt.print_test: model.eval() total = 0 correct = 0 with torch.no_grad(): for data in tqdm(test_loader): if not running: return data = {k: v.to(device) if torch.is_tensor(v) else v for k, v in data.items()} if hasattr(test_set, "prepare"): data = test_set.prepare(data) net_out = run_model(data) c,t = test_set.curriculum_measure(net_out, data["output"]) total += t correct += c print("Test result: %2.f%% (%d out of %d correct)" % (100.0*correct/total, correct, total)) model.train() return iter_start_time = time.time() if i % opt.info_interval == 0 else None data_load_total_time = 0 start_i = i if opt.dump_profile: profiler = torch.autograd.profiler.profile(use_cuda=True) if opt.dump_heatmaps: dataset.set_dump_dir(os.path.join(opt.name, "preview")) @preview() def do_visualize(raw_data, output, pos_map, debug): if pos_map is not None: output = embedding.backmap_output(output, pos_map, raw_data["output"].shape[1]) dataset.visualize_preview(raw_data, output) if debug is not None: plot_debug(debug) preview_timer=OnceEvery(opt.preview_interval) pos_map = None start_iter = i if curriculum is not None: curriculum.init() while running: data_load_timer = time.time() for data in data_loader: if not running: break if loader_reset: print("Loader reset requested. Resetting...") loader_reset = False if curriculum is not None: curriculum.lesson_started() break if opt.dump_profile: if i==start_i+1: print("Starting profiler") profiler.__enter__() elif i==start_i+5+1: print("Stopping profiler") profiler.__exit__(None, None, None) print("Average stats") print(profiler.key_averages().table("cpu_time_total")) print("Writing trace to file") profiler.export_chrome_trace(opt.dump_profile) print("Done.") sys.exit(0) else: print("Step %d out of 5" % (i-start_i)) debug.dbg_print("-------------------------------------") raw_data = data data = {k: v.to(device) if torch.is_tensor(v) else v for k,v in data.items()} if hasattr(dataset, "prepare"): data = dataset.prepare(data) data_load_total_time += time.time() - data_load_timer need_preview = preview_timer() debug_data = {} if opt.debug and need_preview else None optimizer.zero_grad() if opt.n_subbatch=="auto": n_subbatch = math.ceil(data["input"].numel() / opt.max_input_count_per_batch) else: n_subbatch = int(opt.n_subbatch) real_batch = max(math.floor(opt.batch_size/n_subbatch),1) n_subbatch = math.ceil(opt.batch_size/real_batch) remaning_batch = opt.batch_size % real_batch for subbatch in range(n_subbatch): if not running: break input = data["input"] target = data["output"] if n_subbatch!=1: input = input[subbatch * real_batch: (subbatch + 1) * real_batch] target = target[subbatch * real_batch:(subbatch + 1) * real_batch] f2 = data.copy() f2["input"] = input output = run_model(f2, debug=debug_data if subbatch==n_subbatch-1 else None) l = dataset.loss(output, target) debug.nan_check(l, force=True) l.backward() if curriculum is not None: curriculum.update(*dataset.curriculum_measure(output, target)) if remaning_batch!=0 and subbatch == n_subbatch-2: multiply_grads(params, real_batch/remaning_batch) if n_subbatch!=1: if remaning_batch==0: multiply_grads(params, 1/n_subbatch) else: multiply_grads(params, remaning_batch / opt.batch_size) for p in params: torch.nn.utils.clip_grad_norm_(p["params"], opt.grad_clip) optimizer.step() i += 1 curr_loss = l.data.item() loss_plot.add_point(i, curr_loss) loss_sum += curr_loss if i % opt.info_interval == 0: tim = time.time() loss_avg = loss_sum / opt.info_interval if curriculum is not None: curriculum_accuracy.add_point(i, curriculum.get_accuracy()) curriculum_plot.add_point(i, curriculum.step) message = "Iteration %d, loss: %.4f" % (i, loss_avg) if iter_start_time is not None: message += " (%.2f ms/iter, load time %.2g ms/iter, visport: %s)" % ( (tim - iter_start_time) / opt.info_interval * 1000.0, data_load_total_time / opt.info_interval * 1000.0, Visdom.port) print(message) iter_start_time = tim loss_sum = 0 data_load_total_time = 0 debug.dbg_print("Iteration %d, loss %g" % (i, curr_loss)) if need_preview: do_visualize(raw_data, output, pos_map, debug_data) if i % opt.test_interval==0: test() saver.tick(i) if opt.demo and opt.exit_after is None: running = False input("Press enter to quit.") if opt.exit_after is not None and (i-start_iter)>=opt.exit_after: running=False data_load_timer = time.time() if __name__ == "__main__": global running running = True def signal_handler(signal, frame): global running print('You pressed Ctrl+C!') running = False signal.signal(signal.SIGINT, signal_handler) main()