X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=9c28e4789d86162f0be41763b2e27232d002ddd5;hb=687d5b2d9f465577665991b84faec7c789685271;hp=c763016bb7e06be9252dda4996b2f74c053732c8;hpb=bf48dc69f7f57ad391481c8917570e35f661cc4a;p=picoclvr.git diff --git a/main.py b/main.py index c763016..9c28e47 100755 --- a/main.py +++ b/main.py @@ -5,16 +5,14 @@ # Written by Francois Fleuret -# torch.backends.cuda.matmul.allow_tf23 -# torch.autocast(torch.bfloat16) - import math, sys, argparse, time, tqdm, os import torch, torchvision from torch import nn from torch.nn import functional as F -import mygpt, tasks +import ffutils +import mygpt, tasks, problems ###################################################################### @@ -34,8 +32,8 @@ parser = argparse.ArgumentParser( parser.add_argument( "--task", type=str, - default="picoclvr", - help="picoclvr, mnist, maze, snake, stack, expr, world", + default="sandbox", + help="sandbox, picoclvr, mnist, maze, snake, stack, expr, rpl, world", ) parser.add_argument("--log_filename", type=str, default="train.log", help=" ") @@ -44,6 +42,10 @@ parser.add_argument("--result_dir", type=str, default=None) parser.add_argument("--seed", type=int, default=0) +parser.add_argument("--max_percents_of_test_in_train", type=int, default=1) + +######################################## + parser.add_argument("--nb_epochs", type=int, default=None) parser.add_argument("--batch_size", type=int, default=None) @@ -58,18 +60,24 @@ parser.add_argument("--learning_rate", type=float, default=1e-4) parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6") -parser.add_argument("--dim_model", type=int, default=512) +######################################## + +parser.add_argument("--model", type=str, default="37M") + +parser.add_argument("--dim_model", type=int, default=None) -parser.add_argument("--dim_keys", type=int, default=64) +parser.add_argument("--dim_keys", type=int, default=None) -parser.add_argument("--dim_hidden", type=int, default=2048) +parser.add_argument("--dim_hidden", type=int, default=None) -parser.add_argument("--nb_heads", type=int, default=8) +parser.add_argument("--nb_heads", type=int, default=None) -parser.add_argument("--nb_blocks", type=int, default=12) +parser.add_argument("--nb_blocks", type=int, default=None) parser.add_argument("--dropout", type=float, default=0.1) +######################################## + parser.add_argument("--deterministic_synthesis", action="store_true", default=False) parser.add_argument("--no_checkpoint", action="store_true", default=False) @@ -78,6 +86,30 @@ parser.add_argument("--overwrite_results", action="store_true", default=False) parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth") +############################## +# rpl options + +parser.add_argument("--rpl_nb_starting_values", type=int, default=5) + +parser.add_argument("--rpl_max_input", type=int, default=9) + +parser.add_argument("--rpl_prog_len", type=int, default=10) + +parser.add_argument("--rpl_nb_runs", type=int, default=8) + +parser.add_argument("--rpl_no_prog", action="store_true", default=False) + +############################## +# sandbox options + +parser.add_argument("--sandbox_level", type=int, default=0) + +parser.add_argument("--sandbox_levels_nb_items", type=int, default=25) + +parser.add_argument("--sandbox_levels_len_source", type=int, default=6) + +parser.add_argument("--sandbox_levels_len_result", type=int, default=8) + ############################## # picoclvr options @@ -133,6 +165,11 @@ parser.add_argument("--expr_result_max", type=int, default=99) parser.add_argument("--expr_input_file", type=str, default=None) +############################## +# World options + +parser.add_argument("--world_vqae_nb_epochs", type=int, default=25) + ###################################################################### args = parser.parse_args() @@ -144,7 +181,13 @@ if args.result_dir is None: ###################################################################### -default_args = { +default_task_args = { + "sandbox": { + "nb_epochs": 50, + "batch_size": 25, + "nb_train_samples": 100000, + "nb_test_samples": 10000, + }, "picoclvr": { "nb_epochs": 25, "batch_size": 25, @@ -181,18 +224,64 @@ default_args = { "nb_train_samples": 1000000, "nb_test_samples": 10000, }, + "rpl": { + "nb_epochs": 40, + "batch_size": 25, + "nb_train_samples": 100000, + "nb_test_samples": 10000, + }, "world": { - "nb_epochs": 5, + "nb_epochs": 10, "batch_size": 25, - "nb_train_samples": 10000, + "nb_train_samples": 25000, "nb_test_samples": 1000, }, } -if args.task in default_args: - for k, v in default_args[args.task].items(): +if args.task in default_task_args: + for k, v in default_task_args[args.task].items(): + if getattr(args, k) is None: + setattr(args, k, v) + +###################################################################### + +default_model_args = { + "17K": { + "dim_model": 32, + "dim_keys": 32, + "dim_hidden": 32, + "nb_heads": 2, + "nb_blocks": 2, + }, + "37M": { + "dim_model": 512, + "dim_keys": 64, + "dim_hidden": 2048, + "nb_heads": 8, + "nb_blocks": 12, + }, + "122M": { + "dim_model": 768, + "dim_keys": 64, + "dim_hidden": 2048, + "nb_heads": 8, + "nb_blocks": 24, + }, + "352M": { + "dim_model": 1024, + "dim_keys": 64, + "dim_hidden": 2048, + "nb_heads": 8, + "nb_blocks": 48, + }, +} + +if args.model in default_model_args: + for k, v in default_model_args[args.model].items(): if getattr(args, k) is None: setattr(args, k, v) +else: + raise ValueError(f"Unknown model {args.model}") ###################################################################### @@ -252,7 +341,40 @@ picoclvr_pruner_eval = ( ###################################################################### -if args.task == "picoclvr": +if args.task == "sandbox": + if args.sandbox_level == 0: + problem = problems.ProblemLevel0( + nb_sentences=args.sandbox_levels_nb_items, + len_prompt=args.sandbox_levels_len_source, + len_result=args.sandbox_levels_len_result, + ) + elif args.sandbox_level == 1: + problem = problems.ProblemLevel1( + nb_operators=args.sandbox_levels_nb_items, + len_source=args.sandbox_levels_len_source, + len_result=args.sandbox_levels_len_result, + ) + elif args.sandbox_level == 2: + problem = problems.ProblemLevel2( + len_source=args.sandbox_levels_len_source, + len_result=args.sandbox_levels_len_result, + ) + else: + raise ValueError(f"Unknown sandbox level {args.sandbox_level}") + + task = tasks.SandBox( + # problem, + # problems.ProblemAddition(zero_padded=False, inverted_result=False), + # problems.ProblemLenId(len_max=args.sandbox_levels_len_source), + problems.ProblemTwoTargets(len_total=12, len_targets=4), + nb_train_samples=args.nb_train_samples, + nb_test_samples=args.nb_test_samples, + batch_size=args.batch_size, + logger=log_string, + device=device, + ) + +elif args.task == "picoclvr": task = tasks.PicoCLVR( nb_train_samples=args.nb_train_samples, nb_test_samples=args.nb_test_samples, @@ -323,11 +445,27 @@ elif args.task == "expr": device=device, ) +elif args.task == "rpl": + task = tasks.RPL( + nb_train_samples=args.nb_train_samples, + nb_test_samples=args.nb_test_samples, + batch_size=args.batch_size, + nb_starting_values=args.rpl_nb_starting_values, + max_input=args.rpl_max_input, + prog_len=args.rpl_prog_len, + nb_runs=args.rpl_nb_runs, + no_prog=args.rpl_no_prog, + logger=log_string, + device=device, + ) + elif args.task == "world": task = tasks.World( nb_train_samples=args.nb_train_samples, nb_test_samples=args.nb_test_samples, batch_size=args.batch_size, + vqae_nb_epochs=args.world_vqae_nb_epochs, + logger=log_string, device=device, ) @@ -390,12 +528,12 @@ else: if args.task == "expr" and args.expr_input_file is not None: task.produce_results( - nb_epochs_finished, - model, - args.result_dir, - log_string, - args.deterministic_synthesis, - args.expr_input_file, + n_epoch=nb_epochs_finished, + model=model, + result_dir=args.result_dir, + logger=log_string, + deterministic_synthesis=args.deterministic_synthesis, + input_file=args.expr_input_file, ) exit(0) @@ -413,34 +551,37 @@ token_probas = token_count / token_count.sum() entropy = -torch.xlogy(token_probas, token_probas).sum() train_set_perplexity = math.exp(entropy) -############################## - +###################################################################### # A bit of paranoia never hurts -train_examples = {} +def subsets_as_tuples(batches, cs): + s = set() + for batch in batches: + for x in batch: + s.add(tuple([v.item() for v in x])) + if len(s) == cs: + yield s + s = set() + yield s -for input in task.batches(split="train"): - assert input.dim() == 2 and input.dtype == torch.int64 - for x in input: - train_examples[x.sum().item()] = x - -nb_total, nb_collisions = 0, 0 -for input in task.batches(split="test"): - assert input.dim() == 2 and input.dtype == torch.int64 - for x in input: - nb_total += 1 - y = train_examples.get(x.sum().item()) - if y is not None: - if x.size() == y.size() and (x - y).abs().sum() == 0: - nb_collisions += 1 - -del train_examples + +nb_test, nb_in_train = 0, 0 +for test_subset in subsets_as_tuples(task.batches(split="test"), 25000): + in_train = set() + for train_subset in subsets_as_tuples(task.batches(split="train"), 25000): + in_train.update(test_subset.intersection(train_subset)) + nb_in_train += len(in_train) + nb_test += len(test_subset) log_string( - f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set" + f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set" ) +assert ( + nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100 +), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set" + ############################## if args.learning_rate_schedule == "cos": @@ -471,11 +612,11 @@ nb_samples_seen = 0 if nb_epochs_finished >= nb_epochs: task.produce_results( - nb_epochs_finished, - model, - args.result_dir, - log_string, - args.deterministic_synthesis, + n_epoch=nb_epochs_finished, + model=model, + result_dir=args.result_dir, + logger=log_string, + deterministic_synthesis=args.deterministic_synthesis, ) for n_epoch in range(nb_epochs_finished, nb_epochs): @@ -529,7 +670,11 @@ for n_epoch in range(nb_epochs_finished, nb_epochs): ) task.produce_results( - n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis + n_epoch=n_epoch, + model=model, + result_dir=args.result_dir, + logger=log_string, + deterministic_synthesis=args.deterministic_synthesis, ) checkpoint = {