![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F7e1dac43940801ba259535a302817ab2%2FScreen%20Shot%202020-03-16%20at%209.54.09%20PM.png?generation=1584421104555329&alt=media\")
I thought people would enjoy seeing some class activation maps. The areas in red are the areas that are the most important for my CNN to predict root, vowel, consonant respectively. These CAMs were generated from a 256x256 efficientNetB6 trained with cutmix, rotation, scale, and shift augmentation. I used 600,000 CAMs to perform CAM CutMix in my final model. Model described here. (CAM CutMix increased CV and LB by 0.001 over basic CutMix).
\n\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2Fedbb855597a2a4d5bc268aaf15fa6e00%2FScreen%20Shot%202020-03-16%20at%209.53.32%20PM.png?generation=1584421122146913&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F7ca5e40aba271742a48cc725d61eb902%2FScreen%20Shot%202020-03-16%20at%209.53.38%20PM.png?generation=1584421132642316&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F7fa325f243482315230a5075f89668e2%2FScreen%20Shot%202020-03-16%20at%209.53.44%20PM.png?generation=1584421146368106&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F05549915ad13709cf0f083a463763488%2FScreen%20Shot%202020-03-16%20at%209.53.50%20PM.png?generation=1584421158234302&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2Fc0b470e580c24fa8ee6861b8f8c1b45d%2FScreen%20Shot%202020-03-16%20at%209.53.57%20PM.png?generation=1584421167714596&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F79d4fba2e5f665a45b7a7280326dd307%2FScreen%20Shot%202020-03-16%20at%209.54.03%20PM.png?generation=1584421179121245&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F5eccf9346ec1f26e606806de42382587%2FScreen%20Shot%202020-03-16%20at%209.52.57%20PM.png?generation=1584421191704736&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2Fa574ba12b527478edc09fbee0c0d6174%2FScreen%20Shot%202020-03-16%20at%209.53.03%20PM.png?generation=1584421204427798&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F990091f7aca27806e14d5b7fa0f14a6b%2FScreen%20Shot%202020-03-16%20at%209.53.09%20PM.png?generation=1584421270190229&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F47709620093480df1672f7891d6d9d25%2FScreen%20Shot%202020-03-16%20at%209.53.14%20PM.png?generation=1584421279433880&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F1d73a8486342211d8021431bd59533f1%2FScreen%20Shot%202020-03-16%20at%209.53.21%20PM.png?generation=1584421291888922&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2Ffed6891e06804b84eb4e4e9c856c09f6%2FScreen%20Shot%202020-03-16%20at%209.53.26%20PM.png?generation=1584421301239290&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F487c00bc1f4a19a37d30dd9a0ded922a%2FScreen%20Shot%202020-03-16%20at%209.52.17%20PM.png?generation=1584421312822190&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F4794fa840b03352f6083cd96a7acf46b%2FScreen%20Shot%202020-03-16%20at%209.52.23%20PM.png?generation=1584421323946461&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F6b4a7994afb828aa180c4c122aeff0ed%2FScreen%20Shot%202020-03-16%20at%209.52.31%20PM.png?generation=1584421334478343&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F1997b5479064b36556175b7cf556d218%2FScreen%20Shot%202020-03-16%20at%209.52.37%20PM.png?generation=1584421346191713&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F5cc7f455e7c58320bb49ef7d7b5e4467%2FScreen%20Shot%202020-03-16%20at%209.52.44%20PM.png?generation=1584421360388187&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2Fb6076aa03c0e38d11b4e22f6c720d4b4%2FScreen%20Shot%202020-03-16%20at%209.52.51%20PM.png?generation=1584421369440086&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F199e96a69ea0f8bbbbf9825e1826ad66%2FScreen%20Shot%202020-03-16%20at%209.51.34%20PM.png?generation=1584421380882055&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F2f2ba6bfda8c7b79b100ed89d2263164%2FScreen%20Shot%202020-03-16%20at%209.51.43%20PM.png?generation=1584421390708646&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F8bef6493ab4668e5e895672940735835%2FScreen%20Shot%202020-03-16%20at%209.51.50%20PM.png?generation=1584421401219119&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F5bb87eb5ea732017f1a013ee7a278ccb%2FScreen%20Shot%202020-03-16%20at%209.51.57%20PM.png?generation=1584421411222165&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2Ff0446a0ae71692b6e9c7cba4fc6f2fd6%2FScreen%20Shot%202020-03-16%20at%209.52.03%20PM.png?generation=1584421420169763&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1723677%2F2ac553b64b3ef3da69ed1e5fab15971a%2FScreen%20Shot%202020-03-16%20at%209.52.10%20PM.png?generation=1584421431362471&alt=media\")
I thought people would enjoy seeing some class activation maps. The areas in red are the areas that are the most important for my CNN to predict root, vowel, consonant respectively. These CAMs were generated from a 256x256 efficientNetB6 trained with cutmix, rotation, scale, and shift augmentation. I used 600,000 CAMs to perform CAM CutMix in my final model. Model described here. (CAM CutMix increased CV and LB by 0.001 over basic CutMix).
\n\n
Thank you for sharing very interesting idea!\nHow much did your score increase by using cam cutmix?
", "rawMarkdown": "Thank you for sharing very interesting idea!\nHow much did your score increase by using cam cutmix?", "votes": null }, { "id": "776068", "postDate": "03/17/2020 05:24:30", "content": "it is interesting to see activation of vowel /consonant 0 (null).
", "rawMarkdown": "it is interesting to see activation of vowel /consonant 0 (null).", "votes": null }, { "id": "776071", "postDate": "03/17/2020 05:26:38", "content": "CAM CutMix increased CV and LB 0.0010 over my CutMix. (My CutMix was as follows. Instead of choosing lam = np.random.beta(1.0,1.0); cut_rat = np.sqrt(1. - lam), I just did cut_rat = np.random.uniform(0.19,0.31) that worked better than normal CutMix for me.)
Thank you! It's nice. Cam seems to be usable for other augmentations.
", "rawMarkdown": "Thank you! It's nice. Cam seems to be usable for other augmentations.", "votes": null }, { "id": "776295", "postDate": "03/17/2020 09:20:51", "content": "Try consonant 3 ... it has only 4 grapheme combinations so wrong CAM is very likely.
", "rawMarkdown": "Try consonant 3 ... it has only 4 grapheme combinations so wrong CAM is very likely.", "votes": null }, { "id": "776310", "postDate": "03/17/2020 09:49:42", "content": "Very interesting idea, we could also do CAM FMix with this right?
", "rawMarkdown": "Very interesting idea, we could also do CAM FMix with this right?", "votes": null } ], "comments": [ { "id": 776064, "author_name": "statsu", "author_url": "", "post_date": "03/17/2020 05:19:27", "content": "Thank you for sharing very interesting idea!\nHow much did your score increase by using cam cutmix?
", "votes": null, "replies": [ { "id": 776071, "author_name": "cdeotte", "author_url": "", "post_date": "03/17/2020 05:26:38", "content": "CAM CutMix increased CV and LB 0.0010 over my CutMix. (My CutMix was as follows. Instead of choosing lam = np.random.beta(1.0,1.0); cut_rat = np.sqrt(1. - lam), I just did cut_rat = np.random.uniform(0.19,0.31) that worked better than normal CutMix for me.)
Thank you! It's nice. Cam seems to be usable for other augmentations.
", "votes": null, "replies": [] } ] }, { "id": 776068, "author_name": "hengck23", "author_url": "", "post_date": "03/17/2020 05:24:30", "content": "it is interesting to see activation of vowel /consonant 0 (null).
", "votes": null, "replies": [] }, { "id": 776295, "author_name": "dipamc77", "author_url": "", "post_date": "03/17/2020 09:20:51", "content": "Try consonant 3 ... it has only 4 grapheme combinations so wrong CAM is very likely.
", "votes": null, "replies": [] }, { "id": 776310, "author_name": "dipamc77", "author_url": "", "post_date": "03/17/2020 09:49:42", "content": "Very interesting idea, we could also do CAM FMix with this right?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "776042": "I thought people would enjoy seeing some class activation maps. The areas in red are the areas that are the most important for my CNN to predict root, vowel, consonant respectively. These CAMs were generated from a 256x256 efficientNetB6 trained with cutmix, rotation, scale, and shift augmentation. I used 600,000 CAMs to perform CAM CutMix in my final model. Model described [here][1]. (CAM CutMix increased CV and LB by 0.001 over basic CutMix).\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n[1]: https://www.kaggle.com/c/bengaliai-cv19/discussion/136021", "776064": "Thank you for sharing very interesting idea!\nHow much did your score increase by using cam cutmix?", "776068": "it is interesting to see activation of vowel /consonant 0 (null).", "776071": "CAM CutMix increased CV and LB 0.0010 over my CutMix. (My CutMix was as follows. Instead of choosing `lam = np.random.beta(1.0,1.0); cut_rat = np.sqrt(1. - lam)`, I just did `cut_rat = np.random.uniform(0.19,0.31)` that worked better than normal CutMix for me.)", "776293": "Thank you! It's nice. Cam seems to be usable for other augmentations.", "776295": "Try consonant 3 ... it has only 4 grapheme combinations so wrong CAM is very likely.", "776310": "Very interesting idea, we could also do CAM FMix with this right?" }, "source": "meta" }