ohatakenji / sketchrnn-pytorch

SketchRNN : a seq2seq VAE model which draws pictures

A Software Engineer

Portfolio: https://ohatakenji.github.io/portfolio/
AtCoder: https://atcoder.jp/users/Ohata_Kenji

sketchrnn-pytorch's Issues

plot reconstruction

Issue with different dataset i.e. cat.npz

Hi I was trying to run your code exactly while just changing the dataset to cat.npz that I downloaded from google cloud (https://console.cloud.google.com/storage/browser/quickdraw_dataset/full/numpy_bitmap;tab=objects?pli=1&prefix=&forceOnObjectsSortingFiltering=false).

import torch
# from dataset import To5vStrokes, V5Dataset
# from model import SketchRNN
# from trainer import Trainer
from pathlib import Path
from torch.utils.tensorboard import SummaryWriter

data_path = Path.home() / 'Desktop/Jupyter_Notebook/cat.npz'
dataset = V5Dataset(str(data_path), To5vStrokes(max_len=200), pre_scaling=True)
dataloader = torch.utils.data.DataLoader(
    dataset, batch_size=100, shuffle=True)

log_dir = Path.home() / 'Desktop/Jupyter_Notebook'
tb_writer = SummaryWriter(log_dir)

checkpoint_dir = Path.home() / 'Desktop/Jupyter_Notebook'
model = SketchRNN(enc_hidden_size=256, dec_hidden_size=512,
                  Nz=128, M=20, dropout=0.1)
trainer = Trainer(model, dataloader, tb_writer,
                  checkpoint_dir, learning_rate=0.0001)

trainer.train(epoch=300000)

tb_writer.close()

So I basically changed data_path, log_dir, and checkpoint_dir.

Then I have this error,

ValueError                                Traceback (most recent call last)
<ipython-input-14-b4a8b96ebdc7> in <module>
      7 
      8 data_path = Path.home() / 'Desktop/Jupyter_Notebook/cat.npz'
----> 9 dataset = V5Dataset(str(data_path), To5vStrokes(max_len=200), pre_scaling=True)
     10 dataloader = torch.utils.data.DataLoader(
     11     dataset, batch_size=100, shuffle=True)

<ipython-input-6-fbe5ea2c671c> in __init__(self, data_path, transform, pre_scaling)
     29         self.transform = transform
     30         self.data = np.load(data_path, encoding='bytes', allow_pickle=True)
---> 31         scale = self.scaling_factor()
     32         self.data = np.array(list(
     33             map(lambda x: self.scale_stroke(x, scale), self.data)))

<ipython-input-6-fbe5ea2c671c> in scaling_factor(self)
     42 
     43     def scaling_factor(self):
---> 44         data = np.concatenate([S for S in self.data])
     45         return np.std(data[:, 0:2])
     46 

<__array_function__ internals> in concatenate(*args, **kwargs)

ValueError: zero-dimensional arrays cannot be concatenated

Would you mind if you help me out with this please?

Save model

Normalize Input

Number of epochs to train

Nice repository, thank you.

Could you please indicate how many epochs of training did it take to generate that apple from readme?
In the main branch test.py I see 300,000 epochs, exp1.py uses 3,000, test.py in experimental branch uses 300.