Image captioning aims at generating a natural language description of an image. Open domain captioning is a very challenging task, as it requires a fine-grained understanding of the global and the local entities in an image, as well as their attributes and relationships. The recently released MSCOCO challenge 
provides a new, larger scale platform for evaluating image captioning systems, complete with an evaluation server for benchmarking competing methods. Deep learning approaches to sequence modeling have yielded impressive results on the task, dominating the task leaderboard. Inspired by the recently introduced encoder/decoder paradigm for machine translation using recurrent neural networks (RNNs), , and 5] RNN decoder to generate the output caption. These systems are trained end-to-end using back-propagation, and have achieved state-of-the-art results on MSCOCO. More recently in , the use of spatial attention mechanisms on CNN layers to incorporate visual context—which implicitly conditions on the text generated so far—was incorporated into the generation process. It has been shown and we have qualitatively observed that captioning systems that utilize attention mechanisms lead to better generalization, as these models can compose novel text descriptions based on the recognition of the global and local entities that comprise images.
As discussed in , deep generative models for text are typically trained to maximize the likelihood of the next ground-truth word given the previous ground-truth word using back-propagation. This approach has been called “Teacher-Forcing” . However, this approach creates a mismatch between training and testing, since at test-time the model uses the previously generated words from the model distribution to predict the next word. This exposure bias , results in error accumulation during generation at test time, since the model has never been exposed to its own predictions.
Several approaches to overcoming the exposure bias problem described above have recently been proposed. In 
they show that feeding back the model’s own predictions and slowly increasing the feedback probabilityduring training leads to significantly better test-time performance. Another line of work proposes “Professor-Forcing” , a technique that uses adversarial training to encourage the dynamics of the recurrent network to be the same when training conditioned on ground truth previous words and when sampling freely from the network.
While sequence models are usually trained using the cross entropy loss, they are typically evaluated at test time using discrete and non-differentiable NLP metrics such as BLEU , ROUGE , METEOR  or CIDEr . Ideally sequence models for image captioning should be trained to avoid exposure bias and directly optimize metrics for the task at hand.
Recently it has been shown that both the exposure bias and non-differentiable task metric issues can be addressed by incorporating techniques from Reinforcement Learning (RL) . Specifically in , Ranzato et al. use the REINFORCE algorithm 
to directly optimize non-differentiable, sequence-based test metrics, and overcome both issues. REINFORCE as we will describe, allows one to optimize the gradient of the expected reward by sampling from the model during training, and treating those samples as ground-truth labels (that are re-weighted by the reward they deliver). The major limitation of the approach is that the expected gradient computed using mini-batches under REINFORCE typically exhibit high variance, and without proper context-dependent normalization, is typically unstable. The recent discovery that REINFORCE with proper bias correction using learned “baselines” is effective has led to a flurry of work in applying REINFORCE to problems in RL, supervised learning, and variational inference[16, 17, 18]. Actor-critic methods  , which instead train a second “critic” network to provide an estimate of the value of each generated word given the policy of an actor network, have also been investigated for sequence problems recently . These techniques overcome the need to sample from the policy’s (actor’s) action space, which can be enormous, at the expense of estimating future rewards, and training multiple networks based on one another’s outputs, which as  explore, can also be unstable.
In this paper we present a new approach to sequence training which we call self-critical sequence training (SCST), and demonstrate that SCST can improve the performance of image captioning systems dramatically. SCST is a REINFORCE algorithm that, rather than estimating the reward signal, or how the reward signal should be normalized, utilizes the output of its own test-time inference algorithm to normalize the rewards it experiences. As a result, only samples from the model that outperform the current test-time system are given positive weight, and inferior samples are suppressed. Using SCST, attempting to estimate the reward signal, as actor-critic methods must do, and estimating normalization, as REINFORCE algorithms must do, is avoided, while at the same time harmonizing the model with respect to its test-time inference procedure. Empirically we find that directly optimizing the CIDEr metric with SCST and greedy decoding at test-time is highly effective. Our results on the MSCOCO evaluation sever establish a new state-of-the-art on the task, improving the best result in terms of CIDEr from 104.9 to 114.7.
2 Captioning Models
In this section we describe the recurrent models that we use for caption generation.
. Words are represented with one hot vectors that are embedded with a linear embeddingthat has the same output dimension as . The beginning of each sentence is marked with a special BOS token, and the end with an EOS token. Under the model, words are generated and then fed back into the LSTM, with the image treated as the first word . The following updates for the hidden units and cells of an LSTM define the model :
where is a maxout non-linearity with units ( denotes the units) and
is the sigmoid function. We initializeand to zero. The LSTM outputs a distribution over the next word using the softmax function:
In our architecture, the hidden states and word and image embeddings have dimension . Let denote the parameters of the model. Traditionally the parameters are learned by maximizing the likelihood of the observed sequence. Specifically, given a target ground truth sequence , the objective is to minimize the cross entropy loss (XE):
is given by the parametric model in Equation (1).
Attention Model (Att2in).
Rather than utilizing a static, spatially pooled representation of the image, attention models dynamically re-weight the input spatial (CNN) features to focus on specific regions of the image at each time step. In this paper we consider a modification of the architecture of the attention model for captioning given in, and input the attention-derived image feature only to the cell node of the LSTM.
where is the attention-derived image feature. This feature is derived as in  as follows: given CNN features at locations , , where , and . In this work we set the dimension of to , and set and to zero. Let denote the parameters of the model. Then is again defined by (1). The parameters of attention models are also traditionally learned by optimizing the XE loss (2).
Attention Model (Att2all). The standard attention model presented in  also feeds then attention signal as an input into all gates of the LSTM, and the output posterior. In our experiments feeding to all gates in addition to the input did not boost performance, but feeding to both the gates and the outputs resulted in significant gains when ADAM  was used.
3 Reinforcement Learning
Sequence Generation as an RL problem. As described in the previous section, captioning systems are traditionally trained using the cross entropy loss. To directly optimize NLP metrics and address the exposure bias issue, we can cast our generative models in the Reinforcement Learning terminology as in . Our recurrent models (LSTMs) introduced above can be viewed as an “agent” that interacts with an external “environment” (words and image features). The parameters of the network, , define a policy , that results in an “action” that is the prediction of the next word. After each action, the agent (the LSTM) updates its internal “state” (cells and hidden states of the LSTM, attention weights etc). Upon generating the end-of-sequence (EOS) token, the agent observes a “reward” that is, for instance, the CIDEr score of the generated sentence—we denote this reward by
. The reward is computed by an evaluation metric by comparing the generated sequence to corresponding ground-truth sequences. The goal of training is to minimize the negative expected reward:
where and is the word sampled from the model at the time step . In practice is typically estimated with a single sample from :
Policy Gradient with REINFORCE. In order to compute the gradient , we use the REINFORCE algorithm (See also Chapter 13 in ). REINFORCE is based on the observation that the expected gradient of a non-differentiable reward function can be computed as follows:
In practice the expected gradient can be approximated using a single Monte-Carlo sample from , for each training example in the minibatch:
REINFORCE with a Baseline. The policy gradient given by REINFORCE can be generalized to compute the reward associated with an action value relative to a reference reward or baseline :
The baseline can be an arbitrary function, as long as it does not depend on the “action” , since in this case:
This shows that the baseline does not change the expected gradient, but importantly, it can reduce the variance of the gradient estimate. For each training case, we again approximate the expected gradient with a single sample :
4 Self-critical sequence training (SCST)
The central idea of the self-critical sequence training (SCST) approach is to baseline the REINFORCE algorithm with the reward obtained by the current model under the inference algorithm used at test time. The gradient of the negative reward of a sample from the model w.r.t. to the softmax activations at time-step then becomes:
where again is the reward obtained by the current model under the inference algorithm used at test time. Accordingly, samples from the model that return higher reward than will be “pushed up”, or increased in probability, while samples which result in lower reward will be suppressed. Like MIXER , SCST has all the advantages of REINFORCE algorithms, as it directly optimizes the true, sequence-level, evaluation metric, but avoids the usual scenario of having to learn a (context-dependent) estimate of expected future rewards as a baseline. In practice we have found that SCST has much lower variance, and can be more effectively trained on mini-batches of samples using SGD.
Since the SCST baseline is based on the test-time estimate under the current model, SCST is forced to improve the performance of the model under the inference algorithm used at test time. This encourages training/test time consistency like the maximum likelihood-based approaches “Data as Demonstrator” , “Professor Forcing” , and E2E , but importantly, it can directly optimize sequence metrics. Finally, SCST is self-critical, and so avoids all the inherent training difficulties associated with actor-critic methods, where a second “critic” network must be trained to estimate value functions, and the actor must be trained on estimated value functions rather than actual rewards.
In this paper we focus on scenario of greedy decoding, where:
This choice, depicted in Figure 1, minimizes the impact of baselining with the test-time inference algorithm on training time, since it requires only one additional forward pass, and trains the system to be optimized for fast, greedy decoding at test-time.
Generalizations. The basic SCST approach described above can be generalized in several ways.
One generalization is to condition the baseline on what has been generated (i.e. sampled) so far, which makes the baseline word-dependent, and further reduces the variance of the reward signal by making it dependent only on future rewards. This is achieved by baselining the reward for word at timestep with the reward obtained by the word sequence , which is generated by sampling tokens for timesteps , and then executing the inference algorithm to complete the sequence. The resulting reward signal, , is a baselined future reward (advantage) signal that conditions on both the input image and the sequence , and remains unbiased. We call this variant time-dependent SCST (TD-SCST).
Another important generalization is to utilize the inference algorithm as a critic to replace the learned critic of traditional actor-critic approaches. Like for traditional actor-critic methods, this biases the learning procedure, but can be used to trade off variance for bias. Specifically, the primary reward signal at time can be based on a sequence that samples only future tokens, and then executes the inference algorithm to complete the sequence. The primary reward is then based on , and can further be baselined in a time-dependent manner using TD-SCST. The resulting reward signal in this case is . We call this variant True SCST.
We have experimented with both TD-SCST and “True” SCST as described above on the MSCOCO task, but found that they did not lead to significant additional gain. We have also experimented with learning a control-variate for the SCST baseline on MSCOCO to no avail. Nevertheless, we anticipate that these generalizations will be important for other sequence modeling tasks, and policy-gradient-based RL more generally.
Dataset. We evaluate our proposed method on the MSCOCO dataset . For offline evaluation purposes we used the data splits from . The training set contains images, along with captions each. We use a set of image for validation and report results on a test set of
images as well, as given in . We report four widely used automatic evaluation metrics, BLEU-4, ROUGEL, METEOR, and CIDEr. We prune the vocabulary and drop any word that has count less then five, we end up with a vocabulary of size 10096 words.
Image Features 1) FC Models. We use two type of Features: a) (FC-2k) features, where we encode each image with Resnet-101 ( layers) . Note that we do not rescale or crop each image. Instead we encode the full image with the final convolutional layer of resnet, and apply average pooling, which results in a vector of dimension . b) (FC-15K) features where we stack average pooled layers of Resnet-101 ( and ). These
layers are the odd layers of conv4 and conv5, with the exception of the 23rd layer of conv4, which was omitted. This results in a feature vector of dimension 15360.
2) Spatial CNN features for Attention models: (Att2in) We encode each image using the residual convolutional neural network Resnet-101 
. Note that we do not rescale or crop the image. Instead we encode the full image with the final convolutional layer of Resnet-101, and apply spatially adaptive max-pooling so that the output has a fixed size of. At each time step the attention model produces an attention mask over the spatial locations. This mask is applied and then the result is spatially averaged to produce a dimension representation of the attended portion of the image.
Implementation Details. The LSTM hidden, image, word and attention embeddings dimension are fixed to 512 for all of the models discussed herein. All of our models are trained according to the following recipe, except where otherwise noted. We initialize all models by training the model under the XE objective using ADAM  optimizer with an initial learning rate of . We anneal the learning rate by a factor of
every three epochs, and increase the probability of feeding back a sample of the word posterior byevery 5 epochs until we reach a feedback probability . We evaluate at each epoch the model on the development set and select the model with best CIDEr score as an initialization for SCST training. We then run SCST training initialized with the XE model to optimize the CIDEr metric (specifically, the CIDEr-D metric) using ADAM with a learning rate 111In the case of the Att2all models, the XE model was trained for only 20 epochs, and the learning rate was also annealed during RL training.. Initially when experimenting with FC-2k and FC-15k models we utilized curriculum learning (CL) during training, as proposed in , by increasing the number of words that are sampled and trained under CIDEr by one each epoch (the prefix of the sentence remains under the XE criterion until eventually being subsumed). We have since realized that for the MSCOCO task CL is not required, and provides little to no boost in performance. The results reported here for the FC-2K and FC-15K models are trained with CL, while the attention models were trained directly on the entire sentence for all epochs after being initialized by the XE seed models.
Mean/std. performance of SCST versus REINFORCE and REINFORCE with learned baseline (MIXER less CL), for Att2all models (4 seeds) on the Karpathy test set (CIDEr optimized). A one-sample t-test on the gain of SCST over MIXER less CL rejects the null hypothesis on all metrics except ROUGE for(i.e. ).
5.1 Offline Evaluation
Evaluating different RL training strategies.
Table 1 compares the performance of SCST to MIXER  (test set, Karpathy splits). In this experiment, we utilize “curriculum learning” (CL) by optimizing the expected reward of the metric on the last words of each training sentence, optimizing XE on the remaining sentence prefix, and slowly increasing . The results reported were generated with the optimized CL schedule reported in . We found that CL was not necessary to train both SCST and REINFORCE with a learned baseline on MSCOCO—turning off CL sped up training and yielded equal or better results. The gain of SCST over learned baselines was consistent, regardless of the CL schedule and the initial seed. Figures 2 and 3 and table 2 compare the performance of SCST and MIXER less CL for Att2all models on the Karpathy validation and test splits, respectively. Figure 4 and 5 further compare their gradient variance and word posterior entropy on the Karpathy training set. While both techniques are unbiased, SCST in general has lower gradient variance, which translates to improved training performance. Interestingly, SCST has much higher gradient variance than MIXER less CL during the first epoch of training, as most sampled sentences initially score significantly lower than the sentence produced by the test-time inference algorithm.
Training on different metrics.
We experimented with training directly on the evaluation metrics of the MSCOCO challenge. Results for FC-2K models are depicted in table 3.
In general we can see that optimizing for a given metric during training leads to the best performance on that same metric at test time, an expected result.
We experimented with training on multiple test metrics, and found that we were unable to outperform the overall performance of the model trained only on the CIDEr metric, which lifts the performance of all other metrics considerably. For this reason most of our experimentation has since focused on optimizing CIDEr.
|Single Best Models (XE)|
|Single Best Models (SCST unless noted o.w.)|
Single FC-Models Versus Attention Models. We trained FC models (2K and 15 K), as well as attention models (Att2in and Att2all) using SCST with the CIDEr metric. We trained different models for each FC and attention type, starting the optimization from four different random seeds 222pls. consult the supp. material for further details on model training.. We report in Table 4, the system with best performance for each family of models on the test portion of Karpathy splits . We see that the FC-15K models outperform the FC-2K models. Both FC models are outperformed by the attention models, which establish a new state of the art for a single model performance on Karpathy splits. Note that this quantitative evaluation favors attention models is inline with our observation that attention models tend to generalize better and compose outside of the context of the training of MSCOCO, as we will see in Section 6.
|Ensembled Models (XE)|
|Ensembled Models (SCST unless o.w. noted)|
Model Ensembling. In this section we investigate the performance of ensembling over 4 random seeds of the XE and SCST-trained FC models and attention models. We see in Table 5 that ensembling improves performance and confirms the supremacy of attention modeling, and establishes yet another state of the art result on Karpathy splits . Note that in our case we ensemble only models and we don’t do any fine-tuning of the Resnet. NIC , in contrast, used an ensemble of models with fine-tuned CNNs.
5.2 Online Evaluation on MS-COCO Server
Table 6 reports the performance of two variants of 4 ensembled attention models trained with self-critical sequence training (SCST) on the official MSCOCO evaluation server. The previous best result on the leaderboard (as of April 10, 2017) is also depicted. We outperform the previous best system on all evaluation metrics.
|Ens. 4 (Att2all)||114.7||35.2||56.3||27.0|
|Ens. 4 (Att2in)||112.3||34.4||55.9||26.8|
6 Example of Generated Captions
Here we provide a qualitative example of the captions generated by our systems for the image in figure 6. This picture is taken from the objects out-of-context (OOOC) dataset of images . It depicts a boat situated in an unusual context, and tests the ability of our models to compose descriptions of images that differ from those seen during training. The top 5 captions returned by the XE and SCST-trained FC-2K, FC-15K, and attention model ensembles when deployed with a decoding “beam” of 5 are depicted in figure 7 333pls. consult the the supp. material for further details on beam search.. On this image the FC models fail completely, and the SCST-trained ensemble of attention models is the only system that is able to correctly describe the image. In general we found that the performance of all captioning systems on MSCOCO data is qualitatively similar, while on images containing objects situated in an uncommon context  (i.e. unlike the MSCOCO training set) our attention models perform much better, and SCST-trained attention models output yet more accurate and descriptive captions. In general we qualitatively found that SCST-trained attention models describe images more accurately, and with higher confidence, as reflected in Figure 7, where the average of the log-likelihoods of the words in each generated caption are also depicted. Additional examples can be found in the supplementary material. Note that we found that Att2in attention models actually performed better than our Att2all models when applied to images “from the wild”, so here we focus on demonstrating them.
7 Discussion and Future Work
In this paper we have presented a simple and efficient approach to more effectively baselining the REINFORCE algorithm for policy-gradient based RL, which allows us to more effectively train on non-differentiable metrics, and leads to significant improvements in captioning performance on MSCOCO—our results on the MSCOCO evaluation sever establish a new state-of-the-art on the task. The self-critical approach, which normalizes the reward obtained by sampled sentences with the reward obtained by the model under the test-time inference algorithm is intuitive, and avoids having to estimate both action-dependent and action-independent reward functions.
-  Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. EECV, 2014.
-  Kyunghyun Cho, Bart van Merrienboer, Çaglar Gülçehre, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder-decoder for statistical machine translation. EMNLP, 2014.
-  Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. CVPR, 2015.
-  Andrej Karpathy and Fei-Fei Li. Deep visual-semantic alignments for generating image descriptions. CVPR, 2015.
-  Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 1997.
-  Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In ICML, 2015.
-  Marc’Aurelio Ranzato, Sumit Chopra, Michael Auli, and Wojciech Zaremba. Sequence level training with recurrent neural networks. ICLR, 2015.
-  Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. In NIPS, 2015.
-  Alex Lamb, Anirudh Goyal, Ying Zhang, Saizheng Zhang, Aaron Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. Neural Information Processing Systems (NIPS) 2016, 2016.
-  Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: A method for automatic evaluation of machine translation. In ACL, 2002.
-  Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out: Proceedings of the ACL-04 workshop, volume 8. Barcelona, Spain, 2004.
-  Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, volume 29, pages 65–72, 2005.
-  Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, 2015.
-  Richard S. Sutton and Andrew G. Barto. Reinforcement learning: An introduction. MIT Press, 1998.
-  Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. In Machine Learning, pages 229–256, 1992.
-  John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438, 2015.
Wojciech Zaremba and Ilya Sutskever.
Reinforcement learning neural turing machines.Arxiv, 2015.
-  Andriy Mnih and Karol Gregor. Neural variational inference and learning in belief networks. arXiv preprint arXiv:1402.0030, 2014.
-  Dzmitry Bahdanau, Philemon Brakel, Kelvin Xu, Anirudh Goyal, Ryan Lowe, Joelle Pineau, Aaron C. Courville, and Yoshua Bengio. An actor-critic algorithm for sequence prediction. Arxiv, 2016.
-  Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ICLR, 2015.
-  Andrej Karpathy and Fei-Fei Li. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015.
-  Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016.
-  Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: Lessons learned from the 2015 MSCOCO image captioning challenge. PAMI, 2016.
-  Myung Jinchoi, Antonio Torralba, and Alan S. Willsky. Context models and out-of-context objects.
Appendix A Beam search
Throughout the paper and in this supplementary material we often refer to caption results and evaluation metric results obtained using “beam search”. This section briefly summarizes our beam search procedure. While decoding the image to generate captions that describe it, rather than greedily selecting the most probable word (), we can maintain a list of the
most probable sub-sequences generated so far, generate posterior probabilities for the next word of each of these sub-sequences, and then again prune down to the-best sub-sequences. This approach is widely referred to as a beam search, where is the width of the decoding “beam”. In our experiments we additionally prune away hypotheses within the -best list that have a log probability that is below that of the maximally probable partial sentence by more than . For all reported results, the value of is tuned on a per-model basis on the validation set (of the Karpathy splits). On MSCOCO data, is typically optimal for cross-entropy (XE) trained models and SCST-trained models, but in the latter case beam search provides only a very small boost in performance. For our captioning demonstrations we set for all models for illustrative purposes, and because we have qualitatively observed that for test images that are substantially different from those encountered during training, beam search is important.
Appendix B Performance of XE versus SCST trained models
In tables 4 and 5 of the main text we compared the performance of models trained to optimize the CIDEr metric with self-critical sequence training (SCST) with that of their corresponding bootstrap models, which were trained under the cross-entropy (XE) criterion using scheduled sampling . We provide some additional details about these experiments here. For all XE models, the probability of feeding forward the maximally probable word rather than the ground-truth word was increased by every 5 epochs until reaching a maximum value of . The XE model with the best performance on the validation set of the Karpathy splits was then selected as the bootstrap model for SCST (with the exception of the Att2all attention models, where CE training was intentionally terminated prematurely to encourage more exploration during early epochs of RL training).
For all models, the performance of greedily decoding each word at test time is reported, as is the performance of beam search as described in the previous section. As reported in , we found that beam search using RL-trained models resulted in very little performance gain. Figure 8 depicts the performance of our best Att2in model, which is trained to directly optimize the CIDEr metric, as a function of training epoch and evaluation metric, on the validation portion of the Karpathy splits. Optimizing CIDEr clearly improves all of the MSCOCO evaluation metrics substantially.
Appendix C Examples of Generated Captions
Figures 9-16 depict demonstrations of the captioning performance of all systems. In general we found that the performance of all captioning systems on MSCOCO data is qualitatively similar, while on images containing objects situated in an uncommon context  (i.e. unlike the MSCOCO training set) our attention models perform much better, and SCST-trained attention models output yet more accurate and descriptive captions.
Appendix D Further details and analysis of SCST training
One detail that was crucial to optimizing CIDEr to produce better models was to include the EOS tag as a word. When the EOS word was omitted, trivial sentence fragments such as “with a” and “and a” were dominating the metric gains, despite the “gaming” counter-measures (sentence length and precision clipping) that are included in CIDEr-D , which is what we optimized. Including the EOS tag substantially lowers the reward allocated to incomplete sentences, and completely resolved this issue. Another more obvious detail that is important is to associate the reward for the sentence with the first EOS encountered. Omitting the reward from the first EOS fails to reward sentence completion which leads to run-on, and rewarding any words that follow the first EOS token is inconsistent with the decoding procedure.
This work has focused on optimizing the CIDEr metric, since, as discussed in the paper, optimizing CIDER substantially improves all MSCOCO evaluation metrics, as was shown in tables 4 and 5 and is depicted in figure 8. Nevertheless, directly optimizing another metric does lead to higher evaluation scores on that same metric as shown, and so we have started to experiment with including models trained on Bleu, Rouge-L, and METEOR in our Att2in ensemble to attempt to improve it further. So far we have not been able to substantially improve performance w.r.t. the other metrics without more substantially degrading CIDEr.