The game says the main objective of the competition is to develop machine learning-based models to accurately classify a given leaf image from the test dataset to a particular disease category, and to identify an individual disease from multiple disease symptoms on a single leaf image.
\nSome images have more than one type of disease.
\nWhen we make a prediction, do we take the first n maximum probabilities, or just the maximum probabilities. Can someone come and tell me
The game says the main objective of the competition is to develop machine learning-based models to accurately classify a given leaf image from the test dataset to a particular disease category, and to identify an individual disease from multiple disease symptoms on a single leaf image.
\nSome images have more than one type of disease.
\nWhen we make a prediction, do we take the first n maximum probabilities, or just the maximum probabilities. Can someone come and tell me
Hello Dlidli
\nIt is something that I ve been thinking for a while. I think the best solution is defining a threshold and then make the prediction any of the illness which surpases that limit. In case there is none I am gonna choose the highest one.
", "rawMarkdown": "Hello Dlidli\n\nIt is something that I ve been thinking for a while. I think the best solution is defining a threshold and then make the prediction any of the illness which surpases that limit. In case there is none I am gonna choose the highest one.", "votes": null }, { "id": "1278677", "postDate": "04/20/2021 06:57:55", "content": "Hello @ignaciomorenomarn thanks for your answer, this has helped me tremendously
", "rawMarkdown": "Hello @ignaciomorenomarn thanks for your answer, this has helped me tremendously", "votes": null }, { "id": "1279025", "postDate": "04/20/2021 14:13:42", "content": "hi, do you mean the code should may like this during the training step ? :
\ntrain_out = model(image)
\nloss_in = sigmoid(train_out)
\ntrain_loss = loss_fn(loss_in, label)
\npred_label = train_loss > thredshold
train_out = model(image)
\nloss_in = sigmoid(train_out)
\ntrain_loss = loss_fn(loss_in, label)
\nimagine this example
\npred_label =
\n[[0,2 0,6 0,2 0,1]
\n[0,85 0,1 0,05 0,9]]
With thredshold=0,7
\n[[0 1 0 0] #Because it is the largest (At least it have to be one)
\n[1 0 0 1]] #Because they are the larger the threshold
Another alternative is to do a model for healthy or not healthy and then a second model for the illness but I think it is much more complicated.
", "rawMarkdown": "train_out = model(image)\nloss_in = sigmoid(train_out)\ntrain_loss = loss_fn(loss_in, label)\nimagine this example\npred_label = \n[[0,2 0,6 0,2 0,1]\n[0,85 0,1 0,05 0,9]]\n\nWith thredshold=0,7\n[[0 1 0 0] #Because it is the largest (At least it have to be one)\n[1 0 0 1]] #Because they are the larger the threshold\n\nAnother alternative is to do a model for healthy or not healthy and then a second model for the illness but I think it is much more complicated.", "votes": null }, { "id": "1279518", "postDate": "04/21/2021 02:11:23", "content": "the code i find on github.
\nacc = []
\naccuracies = []
\nbest_threshold = np.zeros(out.shape[1])
\nfor i in range(out.shape[1]):
\n y_prob = np.array(out[:,i])
\n for j in threshold:
\n y_pred = [1 if prob>=j else 0 for prob in y_prob]
\n acc.append( matthews_corrcoef(y_test[:,i],y_pred))
\n acc = np.array(acc)
\n index = np.where(acc==acc.max())
\n accuracies.append(acc.max())
\n best_threshold[i] = threshold[index[0][0]]
\n acc = []
https://github.com/suraj-deshmukh/Keras-Multi-Label-Image-Classification/blob/master/model.py
", "rawMarkdown": "the code i find on github.\n\nacc = []\naccuracies = []\nbest_threshold = np.zeros(out.shape[1])\nfor i in range(out.shape[1]):\n y_prob = np.array(out[:,i])\n for j in threshold:\n y_pred = [1 if prob>=j else 0 for prob in y_prob]\n acc.append( matthews_corrcoef(y_test[:,i],y_pred))\n acc = np.array(acc)\n index = np.where(acc==acc.max()) \n accuracies.append(acc.max()) \n best_threshold[i] = threshold[index[0][0]]\n acc = []\n\nhttps://github.com/suraj-deshmukh/Keras-Multi-Label-Image-Classification/blob/master/model.py", "votes": null } ], "comments": [ { "id": 1278062, "author_name": "ignaciomorenomarn", "author_url": "", "post_date": "04/19/2021 14:32:42", "content": "Hello Dlidli
\nIt is something that I ve been thinking for a while. I think the best solution is defining a threshold and then make the prediction any of the illness which surpases that limit. In case there is none I am gonna choose the highest one.
", "votes": null, "replies": [ { "id": 1279025, "author_name": "gjzhongdf163com", "author_url": "", "post_date": "04/20/2021 14:13:42", "content": "hi, do you mean the code should may like this during the training step ? :
\ntrain_out = model(image)
\nloss_in = sigmoid(train_out)
\ntrain_loss = loss_fn(loss_in, label)
\npred_label = train_loss > thredshold
train_out = model(image)
\nloss_in = sigmoid(train_out)
\ntrain_loss = loss_fn(loss_in, label)
\nimagine this example
\npred_label =
\n[[0,2 0,6 0,2 0,1]
\n[0,85 0,1 0,05 0,9]]
With thredshold=0,7
\n[[0 1 0 0] #Because it is the largest (At least it have to be one)
\n[1 0 0 1]] #Because they are the larger the threshold
Another alternative is to do a model for healthy or not healthy and then a second model for the illness but I think it is much more complicated.
", "votes": null, "replies": [] }, { "id": 1279518, "author_name": "gjzhongdf163com", "author_url": "", "post_date": "04/21/2021 02:11:23", "content": "the code i find on github.
\nacc = []
\naccuracies = []
\nbest_threshold = np.zeros(out.shape[1])
\nfor i in range(out.shape[1]):
\n y_prob = np.array(out[:,i])
\n for j in threshold:
\n y_pred = [1 if prob>=j else 0 for prob in y_prob]
\n acc.append( matthews_corrcoef(y_test[:,i],y_pred))
\n acc = np.array(acc)
\n index = np.where(acc==acc.max())
\n accuracies.append(acc.max())
\n best_threshold[i] = threshold[index[0][0]]
\n acc = []
https://github.com/suraj-deshmukh/Keras-Multi-Label-Image-Classification/blob/master/model.py
", "votes": null, "replies": [] } ] }, { "id": 1278677, "author_name": "dlidli", "author_url": "", "post_date": "04/20/2021 06:57:55", "content": "Hello @ignaciomorenomarn thanks for your answer, this has helped me tremendously
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1276808": "The game says the main objective of the competition is to develop machine learning-based models to accurately classify a given leaf image from the test dataset to a particular disease category, and to identify an individual disease from multiple disease symptoms on a single leaf image.\n\nSome images have more than one type of disease.\nWhen we make a prediction, do we take the first n maximum probabilities, or just the maximum probabilities. Can someone come and tell me", "1278062": "Hello Dlidli\n\nIt is something that I ve been thinking for a while. I think the best solution is defining a threshold and then make the prediction any of the illness which surpases that limit. In case there is none I am gonna choose the highest one.", "1278677": "Hello @ignaciomorenomarn thanks for your answer, this has helped me tremendously", "1279025": "hi, do you mean the code should may like this during the training step ? : \ntrain_out = model(image)\nloss_in = sigmoid(train_out)\ntrain_loss = loss_fn(loss_in, label)\npred_label = train_loss > thredshold", "1279182": "train_out = model(image)\nloss_in = sigmoid(train_out)\ntrain_loss = loss_fn(loss_in, label)\nimagine this example\npred_label = \n[[0,2 0,6 0,2 0,1]\n[0,85 0,1 0,05 0,9]]\n\nWith thredshold=0,7\n[[0 1 0 0] #Because it is the largest (At least it have to be one)\n[1 0 0 1]] #Because they are the larger the threshold\n\nAnother alternative is to do a model for healthy or not healthy and then a second model for the illness but I think it is much more complicated.", "1279518": "the code i find on github.\n\nacc = []\naccuracies = []\nbest_threshold = np.zeros(out.shape[1])\nfor i in range(out.shape[1]):\n y_prob = np.array(out[:,i])\n for j in threshold:\n y_pred = [1 if prob>=j else 0 for prob in y_prob]\n acc.append( matthews_corrcoef(y_test[:,i],y_pred))\n acc = np.array(acc)\n index = np.where(acc==acc.max()) \n accuracies.append(acc.max()) \n best_threshold[i] = threshold[index[0][0]]\n acc = []\n\nhttps://github.com/suraj-deshmukh/Keras-Multi-Label-Image-Classification/blob/master/model.py" }, "source": "meta" }