ImageNet HF support and fixes (#30)

Summary:
This PR fixes some of the bugs in the original code that allows us to
match the FLAVA numbers with the original codebase. This was separately
tested on WinoGround dataset for zero-shot and ITM evaluation to match
with the original code.

Also, adds support for ImageNet dataset from HF which simplifies some of
the launch process.

Fixes #27

Pull Request resolved: #30

Reviewed By: ebsmothers

Differential Revision: D35970492

Pulled By: ankitade

fbshipit-source-id: a33af4c009751e46459b86a2cfb623dcd6c3e70e

examples/flava/callbacks/multimodal_eval.py

@@ -8,7 +8,7 @@
 
 import torch
 from data import default_text_transform, VL_MAX_LENGTH_DEFAULT
-from imagenet_zeroshot_data import imagenet_classnames, openai_imagenet_template
+from data.imagenet_zeroshot_data import imagenet_classnames, openai_imagenet_template
 from pytorch_lightning import Callback, LightningDataModule
 from pytorch_lightning.utilities import rank_zero_only
 from tqdm import tqdm
@@ -50,8 +50,10 @@ def run_imagenet_zero_shot(model, dataloader, device, text_transform, *args, **k
     classifier = _zero_shot_classifier(model, device, text_transform)
     logger.info("Classifier built")
     top1, top5, n = 0.0, 0.0, 0.0
-    for images, target in tqdm(dataloader):
-        images = images["image"].to(device)
+    for sample in tqdm(dataloader):
+        images = sample["image"]
+        target = sample["label"]
+        images = images.to(device)
         target = target.to(device)
 
         # predict