ttgump
commented
5 years ago Hi,
Thanks for the email. Yes, ZINB loss need y_true and y_pred, when you call fit/train_on_batch function, keras will pass y_true and y_pred to ZINB loss, just like typical MSE and cross entropy losses. y is the parameter of train_on_batch function, look at the keras document: https://keras.io/models/model/ The input of the model is normalized data, but when fitting ZINB loss, the y_true is raw count data. It can work if passing the raw count data as input, but during optimization procedure, count data might cause problem for gradient calculation. So I used normalized data, since it is continuous. Best,
Tian
From: xuebaliang notifications@github.com Reply-To: ttgump/scDeepCluster reply@reply.github.com Date: Wednesday, June 26, 2019 at 7:59 AM To: ttgump/scDeepCluster scDeepCluster@noreply.github.com Cc: Subscribed subscribed@noreply.github.com Subject: [ttgump/scDeepCluster] some question about "code" (#3)
hello, I have some question about the code. the first question: ''' zinb = ZINB(pi, theta=disp, ridge_lambda=self.ridge, debug=self.debug) self.loss = zinb.loss ''' the model contains the above loss function but I find that the loss function of ZINB class needs two parameters(y_true and y_pred), you may not pass arguments to it. the second question: ''' loss = self.model.train_on_batch(x = [x_counts[index batch_size:(index + 1) batch_size], sf[index batch_size : (index + 1) batch_size]], y = [p[index batch_size:(index + 1) batch_size], raw_counts[index batch_size:(index + 1) batch_size]]) ''' what's the action of y? what can it calculate? the third question: the preprocessing progress transform the count data to normalized data, but you may pass the normalized data to the neural network, why not the count data based on your model? thank you!
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.
xuebaliang
hello, I have some question about the code. the first question: ''' zinb = ZINB(pi, theta=disp, ridge_lambda=self.ridge, debug=self.debug) self.loss = zinb.loss ''' the model contains the above loss function but I find that the loss function of ZINB class needs two parameters(y_true and y_pred), you may not pass arguments to it. the second question: ''' loss = self.model.train_on_batch(x = [x_counts[index batch_size:(index + 1) batch_size], sf[index batch_size : (index + 1) batch_size]], y = [p[index batch_size:(index + 1) batch_size], raw_counts[index batch_size:(index + 1) batch_size]]) ''' what's the action of y? what can it calculate? the third question: the preprocessing progress transform the count data to normalized data, but you may pass the normalized data to the neural network, why not the count data based on your model? thank you!