[feat] add M4C model for TextVQA #213

ronghanghu · 2020-01-06T01:44:44Z

Merge the M4C model (https://arxiv.org/pdf/1911.06258.pdf) for TextVQA into Pythia.

Summary of changes:

Adding README.md under projects/M4C
Adding new models: M4C under pythia/models/m4c.py
Adding new dataset classes: m4c_textvqa, m4c_stvqa, and m4c_ocrvqa under pythia/datasets/vqa/
Adding new config files under configs/vqa
Adding new processors, metrics and losses for M4C training and evaluation.
Adding other utilities (such as PHOC feature extraction).

Introducing new dependencies (added to requirements.txt):

pytorch-transformers
editdistance

M4C for the TextVQA Task

R. Hu, A. Singh, T. Darrell, M. Rohrbach, Iterative Answer Prediction with Pointer-Augmented Multimodal Transformers for TextVQA. arXiv preprint arXiv:1911.06258, 2019 (PDF)

@article{hu2019iterative,
  title={Iterative Answer Prediction with Pointer-Augmented Multimodal Transformers for TextVQA},
  author={Hu, Ronghang and Singh, Amanpreet and Darrell, Trevor and Rohrbach, Marcus},
  journal={arXiv preprint arXiv:1911.06258},
  year={2019}
}

Vocabs, ImDBs and Features:

Datasets	M4C Vocabs	M4C ImDBs	Object Faster R-CNN Features	OCR Faster R-CNN Features
TextVQA	All Vocabs	TextVQA ImDB	OpenImages	TextVQA Rosetta-en OCRs, TextVQA Rosetta-ml OCRs
ST-VQA	All Vocabs	ST-VQA ImDB	ST-VQA Objects	ST-VQA Rosetta-en OCRs
OCR-VQA	All Vocabs	OCR-VQA ImDB	OCR-VQA Objects	OCR-VQA Rosetta-en OCRs

Pretrained models:

Datasets	Configs (under `configs/vqa/`)	Pretrained Models	Metrics	Notes
TextVQA (`m4c_textvqa`)	`m4c_textvqa/m4c_with_stvqa.yml`	`download`	val accuracy - 40.55%; test accuracy - 40.46%	Rosetta-en OCRs; ST-VQA as additional data
TextVQA (`m4c_textvqa`)	`m4c_textvqa/m4c.yml`	`download`	val accuracy - 39.40%; test accuracy - 39.01%	Rosetta-en OCRs
TextVQA (`m4c_textvqa`)	`m4c_textvqa/m4c_ocr_ml.yml`	`download`	val accuracy - 37.06%	Rosetta-ml OCRs
ST-VQA (`m4c_stvqa`)	`m4c_stvqa/m4c.yml`	`download`	val ANLS - 0.472 (accuracy - 38.05%); test ANLS - 0.462	Rosetta-en OCRs
OCR-VQA (`m4c_ocrvqa`)	`m4c_ocrvqa/m4c.yml`	`download`	val accuracy - 63.52%; test accuracy - 63.87%	Rosetta-en OCRs

apsdehal · 2020-01-12T06:28:55Z

Any plans to fix the build?

ronghanghu · 2020-01-12T16:47:25Z

Any plans to fix the build?

Yes! After finishing integration of captioning experiments, I'm addressing the issues you mentioned offline last week (including the CI errors above), and will update this PR.

vedanuj

Please make sure all new files have the required copyright headers.

vedanuj · 2020-01-13T20:07:44Z

pythia/utils/m4c_evaluators.py

+        "wouldnt": "wouldn't",
+        "wouldnt've": "wouldn't've",
+        "wouldn'tve": "wouldn't've",
+        "yall": "y'all",
+        "yall'll": "y'all'll",
+        "y'allll": "y'all'll",
+        "yall'd've": "y'all'd've",
+        "y'alld've": "y'all'd've",
+        "y'all'dve": "y'all'd've",
+        "youd": "you'd",
+        "youd've": "you'd've",
+        "you'dve": "you'd've",
+        "youll": "you'll",
+        "youre": "you're",
+        "youve": "you've",
+    }
+
+    NUMBER_MAP = {
+        "none": "0",
+        "zero": "0",
+        "one": "1",
+        "two": "2",
+        "three": "3",
+        "four": "4",
+        "five": "5",
+        "six": "6",
+        "seven": "7",
+        "eight": "8",
+        "nine": "9",
+        "ten": "10",
+    }
+    ARTICLES = ["a", "an", "the"]
+    PERIOD_STRIP = re.compile("(?!<=\d)(\.)(?!\d)")
+    COMMA_STRIP = re.compile("(?<=\d)(\,)+(?=\d)")
+    PUNCTUATIONS = [
+        ";",
+        r"/",
+        "[",
+        "]",
+        '"',
+        "{",
+        "}",
+        "(",
+        ")",
+        "=",
+        "+",
+        "\\",
+        "_",
+        "-",
+        ">",
+        "<",
+        "@",
+        "`",
+        ",",
+        "?",
+        "!",
+    ]
+
+    def __init__(self, *args, **kwargs):
+        pass
+
+    def word_tokenize(self, word):
+        word = word.lower()
+        word = word.replace(",", "").replace("?", "").replace("'s", " 's")
+        return word.strip()
+
+    def process_punctuation(self, in_text):
+        out_text = in_text
+        for p in self.PUNCTUATIONS:
+            if (p + " " in in_text or " " + p in in_text) or (
+                re.search(self.COMMA_STRIP, in_text) is not None
+            ):
+                out_text = out_text.replace(p, "")
+            else:
+                out_text = out_text.replace(p, " ")
+        out_text = self.PERIOD_STRIP.sub("", out_text, re.UNICODE)
+        return out_text
+
+    def process_digit_article(self, in_text):
+        out_text = []
+        temp_text = in_text.lower().split()
+        for word in temp_text:
+            word = self.NUMBER_MAP.setdefault(word, word)
+            if word not in self.ARTICLES:
+                out_text.append(word)
+            else:
+                pass
+        for word_id, word in enumerate(out_text):
+            if word in self.CONTRACTIONS:
+                out_text[word_id] = self.CONTRACTIONS[word]
+        out_text = " ".join(out_text)
+        return out_text
+
+    def __call__(self, item):
+        item = self.word_tokenize(item)
+        item = item.replace("\n", " ").replace("\t", " ").strip()
+        item = self.process_punctuation(item)
+        item = self.process_digit_article(item)
+        return item
+


Please reuse the existing code and remove this duplication.

@vedanuj Thanks for the comments! I'll fix it later this week.

ronghanghu · 2020-01-17T02:11:18Z

@vedanuj thanks for your review!

Any plans to fix the build?

It's fixed now, by pytest==5.2.0 in requirements.txt

Please make sure all new files have the required copyright headers.

The header Copyright (c) Facebook, Inc. and its affiliates. is added to the new files.

Please reuse the existing code and remove this duplication.

I found that this is non-trivial to fix. This PR is made against v0.4. However, the EvalAIAnswerProcessor only exists in the current master branch. I tried rebasing against master. However, I found that v0.4 has a major commit ahead of master (926d3b0) that address multitasking #173, which cases a lot of rebase conflict. (Note that this PR for the M4C model is built to be compatible with #173, which I believe will eventually appear in master.)

Do you have suggestions on how to proceed here?

apsdehal

I am landing this internally, but it would be great if you can work on this on separate PR.

apsdehal · 2020-03-19T20:02:57Z

requirements.txt

 requests==2.21.0
-fasttext==0.9.1
+fastText


Any particular reason that we are not using a specific version here?

apsdehal · 2020-03-19T20:03:03Z

requirements.txt

 nltk==3.4.1
+pytorch-transformers==1.2.0
+editdistance


apsdehal · 2020-03-19T20:03:45Z

pythia/utils/phoc/src/cphoc.c

@@ -0,0 +1,146 @@
+// C implementation of the PHOC respresentation. Converts a string into a PHOC feature vector


We should move M4C specific utils to a folder utils/m4c_utils.

apsdehal · 2020-03-19T20:04:20Z

pythia/utils/objects_to_byte_tensor.py

@@ -0,0 +1,50 @@
+


This should go inside distributed utils.

apsdehal · 2020-03-19T20:05:32Z

pythia/common/defaults/configs/datasets/vqa/m4c_stvqa.yml

@@ -0,0 +1,22 @@
+dataset_attributes:


Configs currently have a lot of replication and not fully utilizing the power of inheritance in our configuration system.

vedanuj

In general looks good. I think lot of code/config duplication can be avoided.

vedanuj · 2020-03-19T20:15:55Z

projects/M4C/scripts/extract_ocr_frcn_feature.py

+# install `vqa-maskrcnn-benchmark` from
+# https://github.com/ronghanghu/vqa-maskrcnn-benchmark-m4c
+import sys; sys.path.append('/private/home/ronghanghu/workspace/vqa-maskrcnn-benchmark')  # NoQA


Is there some specific change done for m4c? If not can the https://gitlab.com/meetshah1995/vqa-maskrcnn-benchmark repo be used here? Because it is already being used in feature extraction scripts pythia provides here pythia/scripts/features/extract_features_vmb.py

Yeah, this needs a separate look.

Yes, there are specific changes for OCR feature extraction. The major change is to allow RoI-Pooling from an externally-specified bounding box (OCR boxes in our use case) instead of from the Faster R-CNN's own RPN proposal. The new branch (https://github.com/ronghanghu/vqa-maskrcnn-benchmark-m4c) is compatible with pythia/scripts/features/extract_features_vmb.py, so can also land it into https://gitlab.com/meetshah1995/vqa-maskrcnn-benchmark.

vedanuj · 2020-03-19T20:16:11Z

.gitignore

@@ -24,3 +24,5 @@ eggs/
 *.egg
 .DS_Store
 .vscode/*
+*.so
+*-checkpoint.ipynb


Are .ipynb files generated by any of the scripts added here? If not please remove it.

Thanks, I'll remove this change.

The *-checkpoint.ipynb files are generated by Jupyter Notebook servers automatically (similar to how vim generates .swp files -- but doesn't delete them after Jupyter Notebook is shut down). But these Jupyter Notebooks were added primarily during our internal analyses and I did not add them here.

vedanuj · 2020-03-19T20:16:23Z

configs/captioning/m4c_textcaps/m4c_captioner_coco_textcaps_joint.yml

+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images
+      - open_images/detectron_fix_100/fc6/train,m4c_textvqa_ocr_en_frcn_features/train_images


This behaviour should be configurable from code in future. This looks very messy now.

vedanuj · 2020-03-19T20:17:26Z

projects/M4C/scripts/extract_ocr_frcn_feature.py

+def _image_transform(image_path):
+    img = Image.open(image_path)
+    im = np.array(img).astype(np.float32)
+    # handle a few corner cases
+    if im.ndim == 2:  # gray => RGB
+        im = np.tile(im[:, :, None], (1, 1, 3))
+    if im.shape[2] > 3:  # RGBA => RGB
+        im = im[:, :, :3]
+
+    im = im[:, :, ::-1]  # RGB => BGR
+    im -= np.array([102.9801, 115.9465, 122.7717])
+    im_shape = im.shape
+    im_size_min = np.min(im_shape[0:2])
+    im_size_max = np.max(im_shape[0:2])
+    im_scale = float(800) / float(im_size_min)
+    # Prevent the biggest axis from being more than max_size
+    if np.round(im_scale * im_size_max) > 1333:
+        im_scale = float(1333) / float(im_size_max)
+    im = cv2.resize(
+        im,
+        None,
+        None,
+        fx=im_scale,
+        fy=im_scale,
+        interpolation=cv2.INTER_LINEAR
+    )
+    img = torch.from_numpy(im).permute(2, 0, 1)
+    return img, im_scale


Please check if code can be reused between these and pythia/scripts/features/extract_features_vmb.py

vedanuj · 2020-03-19T20:20:30Z

pythia/datasets/processors.py

+
+
+@registry.register_processor("bert_tokenizer")
+class BertTokenizerProcessor(BaseProcessor):


@apsdehal We should check if this is same or different from the BertTokenizer we have internally when merging.

Ours is inside processors/bert.

vedanuj · 2020-03-19T20:22:39Z

pythia/modules/losses.py

@@ -160,7 +160,7 @@ def forward(self, sample_list, model_output, *args, **kwargs):
        if loss.dim() == 0:
            loss = loss.view(1)

-            key = "{}/{}/{}".format(


Was this a bug earlier?

Yeah, I noticed this and fixed it in the rebase. It was already fixed in dev branch,

vedanuj · 2020-03-19T20:24:52Z

pythia/utils/checkpoint.py

@@ -42,7 +42,7 @@ def __init__(self, trainer):

        self.models_foldername = os.path.join(self.ckpt_foldername, "models")
        if not os.path.exists(self.models_foldername):
-            os.makedirs(self.models_foldername)
+            os.makedirs(self.models_foldername, exist_ok=True)


Can we not change this? This might have dangerous consequences where we overwrite trained models by mistake.

This is fine. This is an issue in case of distributed training which is already fixed in our dev branch. If you see the if statement above you will understand. This happens when there is a race condition and one of the jobs have already made the folder, then the whole thing fails. That's why exist_ok=True needs. exist_ok doesn't overwrite, it just doesn't throw an error if the folder is already there.

Hi, From Oleksii's experience, without this change, the program frequently crashes (over 50% time) in distributed training due to a race condition, since multiple processes are trying to make directories, and there's no lock on these lines.

vedanuj · 2020-03-19T20:25:56Z

requirements.txt

-torch==1.2.0
-torchvision==0.2.2
-tensorboardX==1.2
+torch>=1.2
+torchvision>0.2
+tensorboardX>=1.2
 numpy>=1.14
-tqdm==4.19.9
+tqdm>=4.19
 demjson>=2.2
 torchtext>=0.2
 GitPython>=2.1
 PyYAML>=3.11
-pytest==3.3.2
+pytest==5.2.0


Are these changes necessary for this PR? If not please remove it from this.

I removed the changes with >= as they can be breaking in nature. Rest I accepted and then updated with our changes as of now.

apsdehal · 2020-03-19T20:40:17Z

@ronghanghu This has landed internally on dev. Please make further PRs on that branch. This should be automatically closed once it lands in master.

ronghanghu · 2020-03-19T22:44:16Z

@apsdehal and @vedanuj Thanks a lot for your review and landing!

(I'll make PRs to dev branch for future changes)

apsdehal · 2020-03-20T01:22:23Z

Also, I would suggest you to move to automatic download api instead of making users download everything as per the instructions in your readme.

apsdehal · 2020-05-07T18:36:26Z

Closing as landed internally.

Closes #213 Merge the M4C model (https://arxiv.org/pdf/1911.06258.pdf) for TextVQA into Pythia. Summary of changes: * Adding `README.md` under `projects/M4C` * Adding new models: M4C under `pythia/models/m4c.py` * Adding new dataset classes: `m4c_textvqa`, `m4c_stvqa`, and `m4c_ocrvqa` under `pythia/datasets/vqa/` * Adding new config files under `configs/vqa` * Adding new processors, metrics and losses for M4C training and evaluation. * Adding other utilities (such as PHOC feature extraction). Introducing new dependencies (added to `requirements.txt`): * `pytorch-transformers` * `editdistance` * R. Hu, A. Singh, T. Darrell, M. Rohrbach, *Iterative Answer Prediction with Pointer-Augmented Multimodal Transformers for TextVQA*. arXiv preprint arXiv:1911.06258, 2019 ([PDF](https://arxiv.org/pdf/1911.06258.pdf)) ``` @Article{hu2019iterative, title={Iterative Answer Prediction with Pointer-Augmented Multimodal Transformers for TextVQA}, author={Hu, Ronghang and Singh, Amanpreet and Darrell, Trevor and Rohrbach, Marcus}, journal={arXiv preprint arXiv:1911.06258}, year={2019} } ``` | Datasets | M4C Vocabs | M4C ImDBs | Object Faster R-CNN Features | OCR Faster R-CNN Features | |--------------|-----|-----|-----------------------------------------------------------------------------------|---------------------------------------------------------------------------------| | TextVQA | [All Vocabs](https://dl.fbaipublicfiles.com/pythia/m4c/data/m4c_vocabs.tar.gz) | [TextVQA ImDB](https://dl.fbaipublicfiles.com/pythia/m4c/data/imdb/m4c_textvqa.tar.gz) | [OpenImages](https://dl.fbaipublicfiles.com/pythia/features/open_images.tar.gz) | [TextVQA Rosetta-en OCRs](https://dl.fbaipublicfiles.com/pythia/m4c/data/m4c_textvqa_ocr_en_frcn_features.tar.gz), [TextVQA Rosetta-ml OCRs](https://dl.fbaipublicfiles.com/pythia/m4c/data/m4c_textvqa_ocr_ml_frcn_features.tar.gz) | | ST-VQA | [All Vocabs](https://dl.fbaipublicfiles.com/pythia/m4c/data/m4c_vocabs.tar.gz) | [ST-VQA ImDB](https://dl.fbaipublicfiles.com/pythia/m4c/data/imdb/m4c_stvqa.tar.gz) | [ST-VQA Objects](https://dl.fbaipublicfiles.com/pythia/m4c/data/m4c_stvqa_obj_frcn_features.tar.gz) | [ST-VQA Rosetta-en OCRs](https://dl.fbaipublicfiles.com/pythia/m4c/data/m4c_stvqa_ocr_en_frcn_features.tar.gz) | | OCR-VQA | [All Vocabs](https://dl.fbaipublicfiles.com/pythia/m4c/data/m4c_vocabs.tar.gz) | [OCR-VQA ImDB](https://dl.fbaipublicfiles.com/pythia/m4c/data/imdb/m4c_ocrvqa.tar.gz) | [OCR-VQA Objects](https://dl.fbaipublicfiles.com/pythia/m4c/data/m4c_ocrvqa_obj_frcn_features.tar.gz) | [OCR-VQA Rosetta-en OCRs](https://dl.fbaipublicfiles.com/pythia/m4c/data/m4c_ocrvqa_ocr_en_frcn_features.tar.gz) | | Datasets | Configs (under `configs/vqa/`) | Pretrained Models | Metrics | Notes | |--------|------------------|----------------------------|-------------------------------|-------------------------------| | TextVQA (`m4c_textvqa`) | `m4c_textvqa/m4c_with_stvqa.yml` | [`download`](https://dl.fbaipublicfiles.com/pythia/m4c/m4c_release_models/m4c_textvqa/m4c_textvqa_m4c_with_stvqa.ckpt) | val accuracy - 40.55%; test accuracy - 40.46% | Rosetta-en OCRs; ST-VQA as additional data | | TextVQA (`m4c_textvqa`) | `m4c_textvqa/m4c.yml` | [`download`](https://dl.fbaipublicfiles.com/pythia/m4c/m4c_release_models/m4c_textvqa/m4c_textvqa_m4c.ckpt) | val accuracy - 39.40%; test accuracy - 39.01% | Rosetta-en OCRs | | TextVQA (`m4c_textvqa`) | `m4c_textvqa/m4c_ocr_ml.yml` | [`download`](https://dl.fbaipublicfiles.com/pythia/m4c/m4c_release_models/m4c_textvqa/m4c_textvqa_m4c_ocr_ml.ckpt) | val accuracy - 37.06% | Rosetta-ml OCRs | | ST-VQA (`m4c_stvqa`) | `m4c_stvqa/m4c.yml` | [`download`](https://dl.fbaipublicfiles.com/pythia/m4c/m4c_release_models/m4c_stvqa/m4c_stvqa_m4c.ckpt) | val ANLS - 0.472 (accuracy - 38.05%); test ANLS - 0.462 | Rosetta-en OCRs | | OCR-VQA (`m4c_ocrvqa`) | `m4c_ocrvqa/m4c.yml` | [`download`](https://dl.fbaipublicfiles.com/pythia/m4c/m4c_release_models/m4c_ocrvqa/m4c_ocrvqa_m4c.ckpt) | val accuracy - 63.52%; test accuracy - 63.87% | Rosetta-en OCRs |

facebook-github-bot added the CLA Signed Do not delete this pull request or issue due to inactivity. label Jan 6, 2020

ronghanghu force-pushed the project/m4c branch 6 times, most recently from 8423d60 to e5b58fd Compare January 6, 2020 17:14

[feat] add M4C model

d5c59f1

ronghanghu force-pushed the project/m4c branch from e5b58fd to d5c59f1 Compare January 6, 2020 17:15

vedanuj reviewed Jan 13, 2020

View reviewed changes

ronghanghu force-pushed the project/m4c branch 3 times, most recently from ec1b1dc to e0e99c9 Compare January 17, 2020 01:54

ronghanghu force-pushed the project/m4c branch from e0e99c9 to 5a15b40 Compare January 17, 2020 02:20

[feat] add M4C-Captioner

58732c7

ronghanghu force-pushed the project/m4c branch from e5a2962 to 58732c7 Compare March 15, 2020 06:10

[fix] upgrade pytest to fix CI build error

8f0dba5

apsdehal reviewed Mar 19, 2020

View reviewed changes

vedanuj reviewed Mar 19, 2020

View reviewed changes

update TextCaps bibtex

28e48ce

apsdehal closed this May 7, 2020

hackgoofer mentioned this pull request Apr 30, 2021

Working with pretrained UNITER and Hateful Memes dataset #912

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[feat] add M4C model for TextVQA #213

[feat] add M4C model for TextVQA #213

ronghanghu commented Jan 6, 2020 •

edited

Loading

apsdehal commented Jan 12, 2020

ronghanghu commented Jan 12, 2020 •

edited

Loading

vedanuj left a comment

vedanuj Jan 13, 2020

ronghanghu Jan 14, 2020 •

edited

Loading

ronghanghu commented Jan 17, 2020 •

edited

Loading

apsdehal left a comment

apsdehal Mar 19, 2020

apsdehal Mar 19, 2020

apsdehal Mar 19, 2020

apsdehal Mar 19, 2020

apsdehal Mar 19, 2020

vedanuj left a comment

vedanuj Mar 19, 2020 •

edited

Loading

apsdehal Mar 19, 2020

ronghanghu Mar 19, 2020

vedanuj Mar 19, 2020

ronghanghu Mar 19, 2020

vedanuj Mar 19, 2020

vedanuj Mar 19, 2020

vedanuj Mar 19, 2020

apsdehal Mar 19, 2020

vedanuj Mar 19, 2020

apsdehal Mar 19, 2020

vedanuj Mar 19, 2020

apsdehal Mar 19, 2020

ronghanghu Mar 19, 2020

vedanuj Mar 19, 2020

apsdehal Mar 19, 2020

apsdehal commented Mar 19, 2020

ronghanghu commented Mar 19, 2020 •

edited

Loading

apsdehal commented Mar 20, 2020

apsdehal commented May 7, 2020

		@@ -0,0 +1,146 @@
		// C implementation of the PHOC respresentation. Converts a string into a PHOC feature vector



		@registry.register_processor("bert_tokenizer")
		class BertTokenizerProcessor(BaseProcessor):

[feat] add M4C model for TextVQA #213

[feat] add M4C model for TextVQA #213

Conversation

ronghanghu commented Jan 6, 2020 • edited Loading

M4C for the TextVQA Task

Vocabs, ImDBs and Features:

Pretrained models:

apsdehal commented Jan 12, 2020

ronghanghu commented Jan 12, 2020 • edited Loading

vedanuj left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ronghanghu Jan 14, 2020 • edited Loading

Choose a reason for hiding this comment

ronghanghu commented Jan 17, 2020 • edited Loading

apsdehal left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vedanuj left a comment

Choose a reason for hiding this comment

vedanuj Mar 19, 2020 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

apsdehal commented Mar 19, 2020

ronghanghu commented Mar 19, 2020 • edited Loading

apsdehal commented Mar 20, 2020

apsdehal commented May 7, 2020

ronghanghu commented Jan 6, 2020 •

edited

Loading

ronghanghu commented Jan 12, 2020 •

edited

Loading

ronghanghu Jan 14, 2020 •

edited

Loading

ronghanghu commented Jan 17, 2020 •

edited

Loading

vedanuj Mar 19, 2020 •

edited

Loading

ronghanghu commented Mar 19, 2020 •

edited

Loading