This repository has been archived by the owner on Dec 16, 2022. It is now read-only.
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
* max instances for debugging * b * printing devices * moving tensors? * self * p * p * l * l * fixing heap? * stop logging and printing * less prints * printing devices * p * . * devices * device * test * testing not sampling * testing not using model again * test not moving tensors * not printing * trying image subset * debugging model * going back to full (model is slow?) * right number of instances * distribut * more potential hard negatives * non-distributed * distributed + adding another seen set * fixing evaluation method * format * testing fixed eval * fixing variable * fixing training var * testing new eval again * fix * fix * fix? * changing k to 5 * float * moving labels to gpu * long * trying hopefully fixed loss function * fix * testing out the whole thing * setting max instances to debug in distributed * debug stuff * fixing num images * hopefully fixing the dataset reader in dist * full data * testing out brand new changes * deleting some old comments * fixing validation bug * testing on 1 gpu for now * feature cache broken? * switching to tensor fields and stuff * fix * print device * trying to not move the batch? * moving small batches to cpu * not printing device * deleting old tensor? * debug * printing memory allocation * moving tensor to cpu immediately? * deleting batch? * debug * debug * debug * does this work? * switching to eval and no grad * fix * mask list * backbone roll * typo * log * debug * notes * testing no grad * testing validation batch size of 1 * bug * didn't have the right variable? * don't need to softmax? * trying flickr30k with 8 batch and dummy captions * full flickr? * batch size 1 * testing training batches for validation * Testing out val stuff * updating reader (test will fail for now) * debug statements to figure out why val isn't worki * testing if top images always have same scores * getting rid of caption debugging step * using the right caption var * updating reader to mirror vilbert training setup * full dataset (dummy caption embeddings) * switching to real caption embeddings * testing caching hard negatives * log * limit instances to test caching * delete faiss * more cache tests * one more log statement * single epoch to calculate hard negatives * need to import logging * don't log misses anymore (too slow) * using consistent hash function (test # instances) * Flickr30k batching (#277) merge main + caching captions * test caching captions and hard negatives on full * don't log cache hits * logging training labels to debug * switching val to 4 way mc * can we overfit * not 1k instances * not logging + overfit * not overfit * even fewer instances * all instances * even more overfitting * back to normal * b * bkac to normal * log loss and stuff again * reset * don't include hard negatives in case there's a bug * batch size of 1 * more epochs * only correct answer and hard negatives * Cleanup * Fix error in caption caching * Find hard negatives even when we don't have enough instances * O(1) algorithm for finding a random number with one exception * Make sure the wrong caption comes from a different image * Cross entropy loss * trying overfitting with full instances * use full dataset without learning rate scheduler * don't limit instances and don't log * batch size, scheduler, wandb * comment out wandb * full dataset no hard negatives * don't log loss * giving the correct answer a cheat word * use local feature cache * logging cache stuff * different local feature cache dir * switching to cheat box * bug * something up with some boxes * no cheating and no hard negatives * seeing is a really big batch size works * bug * testing 64 bs * batch size 32 * batch size 48 * full training with 32 batch size no hard negatives * more gradient accumulation steps * trying to train with 10% of the data * fix * bumping up the learning rate, don't correct bias * gradient accumulation + hard negatives * use local feature cache * changing params back * trying real validation * no hard negatives * hard negatives and not real validation * no hard negatives + real validation * calc hn * fixing predictors * fix * fix * fix * fix * cleaning up PR (in progress) * cleaning things up * more cleanup * change warmup steps * only validate every ~5 epochs * printing shapes * more logging * fix log * try cat instead of stack * different logging * test * fix * try batches per epoch * bug * get rid of log statement * use local feature cache * log * logging cache miss * switching back to old captions to use cache * switching back to preprocesing captions * using nfs * Disabling hard negatives to test epoch strat * not logging cache misses * write to local cache (faster) * epoch multiplier * no hard negatives * hard negatives * lowering number of warmup steps * no hard negatives * hard negatives * no hard negatives * hard negatives * Trying Jiasen's featurizer (1x epoch mult) * null image stuff * null image * don't featurize captions (no hn) * adding vilbert ir model tests * cleanup + test distributed * cleanup + dist * test distributed * don't use shard_iterable * fix feature dir * changelog * reformat * log shapes * removing unused vars * using old features * style * lint * lint * don't log shapes * lint * fixing type * debug * changing test files to hopefully fix test * using cloud link for data dir * cleanup * delete print * comment * cleanup * fixing test assert * committing a bunch of fixes * not distributed * fixing metrics * Adding test files + upping max instances * fixes * Switching back to nfs cache * renaming n * update comment * fix * making test deterministic? * sorting files to hopefully achieve consistency Co-authored-by: Dirk Groeneveld <[email protected]>
- Loading branch information