can you share a trained model?

[sooyeonshin]

can you share a trained model?

[wbhu]

Hi @sooyeonshin,

There are bugs in the source code updated 2 weeks ago. And I am training the fixed model right now, you could also have a try of the updated code.

thanks.

[sooyeonshin]

oh last updated version seem's good.

thank you for answer my question.

and it's just my advice, It will be better, you modify save-load model part.

save routine coded like this
# save the model
if np.mod(iter_num, self.save_every_epoch) == 0:

I think change iter_num to epoch (if you want save model by epock size) or change self.save_every_epoch to number of iteration.

also in your code, exist save model method but not exist load pretrained model before train.

so I try to coding load part like this (add self.ckpt_train at initial part)

   if self.ckpt_train : # if your want load pretrained model, ckpt_train=true 
        model_dir = "%s_%s_%s" % (self.trainset, self.batch_size, self.patch_sioze)
        checkpoint_dir2 = os.path.join(self.ckpt_dir, model_dir)
        full_path = tf.train.latest_checkpoint(checkpoint_dir2)
        self.saver.restore(self.sess, full_path)
        print("[*] loaded check point model = %s" % full_path)
        global_step = int(full_path.split('/')[-1].split('-')[-1])
        start_epoch = global_step // numBatch
        start_step = global_step % numBatch
        print("[*] Model restore finished, current globle step: %d" % global_step)
        print("[*] previous final epoch: %d" % start_epoch)
        print("[*] start step: %d" % start_step)
        iter_num = global_step

        for epoch in xrange(**start_epoch,** self.epoch): #start at pretrained epoch
            for batch_id in xrange(start_step, numBatch):
                batch_images = data[batch_id * self.batch_size:(batch_id + 1) * self.batch_size, :, :, :]
                batch_images = np.array(batch_images / 255.0, dtype=np.float32)     #normalize the data to 0-1
                train_images = add_noise(batch_images, self.sigma, self.sess)
                _, loss, summary = self.sess.run([self.train_step, self.loss, merged], \
                                                 feed_dict={self.X: train_images, self.X_: batch_images})
                print("Epoch: [%2d] [%4d/%4d] time: %4.4f, loss: %.6f" \
                      % (epoch + 1, batch_id + 1, numBatch,
                         time.time() - start_time, loss))
                iter_num += 1
                writer.add_summary(summary, iter_num)
            if np.mod(epoch, self.eval_every_epoch) == 0:
                self.evaluate(epoch, iter_num, test_data)  # test_data value range is 0-255
            # save the model
            if np.mod(epoch, self.save_every_epoch) == 0:
                self.save(iter_num)
        print("[*] Finish training.")

    else : #if you want train at first, ckpt_train is false and this routine is not changed
        for epoch in xrange(self.epoch):
            for batch_id in xrange(numBatch):
                batch_images = data[batch_id * self.batch_size:(batch_id + 1) * self.batch_size, :, :, :]
                batch_images = np.array(batch_images / 255.0, dtype=np.float32)     #normalize the data to 0-1
                train_images = add_noise(batch_images, self.sigma, self.sess)
                _, loss, summary = self.sess.run([self.train_step, self.loss, merged], \
                                                 feed_dict={self.X: train_images, self.X_: batch_images})
                print("Epoch: [%2d] [%4d/%4d] time: %4.4f, loss: %.6f" \
                      % (epoch + 1, batch_id + 1, numBatch,
                         time.time() - start_time, loss))
                iter_num += 1
                writer.add_summary(summary, iter_num)
            if np.mod(epoch, self.eval_every_epoch) == 0:
                self.evaluate(epoch, iter_num, test_data)  # test_data value range is 0-255
            # save the model
            if np.mod(epoch, self.save_every_epoch) == 0:
                self.save(iter_num)
        print("[*] Finish training.")

It may be unnecessary interference.

anyway I got alot of help from your work. thanks alot.

[wbhu]

Hi @sooyeonshin ,

Thanks for your advice. You could pull a request for the improved code, it will be greatly appreciated.