python - 如何修复 ValueError：数据基数不明确，同时拟合模型

Question

所以，我试图重新创建sketch-rnn 项目，为你完成一个涂鸦。但我仍然是 tensorflow 和 keras 的新手，不太了解事情是如何工作的。

import tensorflow as tf
from tensorflow import keras
import numpy as np
import pickle, copy


with open("cake_100.pickle", "rb") as f:
    drawings = pickle.load(f)
    # each drawing will be a list of vectors, i.e [x,y,p1,p2,p3]
    # p1 means pendown, p2 is penup, and p3 is done.
    for i in range(len(drawings)):
        drawings[i] = np.expand_dims(drawings[i], 2) # expand dims, cuz thats what the model wants :/


def build_model():
    visible = keras.layers.Input(
        shape=(5, 1)
    )  # take one drawing at a time with shape (None,5,1)
    # how to handle all these states??? - done
    encoder = keras.layers.Bidirectional(
        keras.layers.LSTM(
            128, activation="tanh", return_state=True, return_sequences=True
        )
    )  # encode sequences into fixed length vectors of shape (5,256), i.e for each sequence in a drawing, for each scalar in that sequence - convert it to a 256 length vector
    (
        encoder_out,
        forward_state_h,
        forward_state_c,
        backward_state_h,
        backward_state_c,
    ) = encoder(visible)
    state_h, state_c = (
        forward_state_h + backward_state_h,
        forward_state_c + backward_state_c,
    )
    decoder1 = keras.layers.LSTM(128, activation="tanh", return_sequences=True,)
    decoder1_out = decoder1(encoder_out, initial_state=[state_h, state_c],)
    decoder2 = keras.layers.Dense(1, activation="relu")
    decoder_out = decoder2(decoder1_out)
    model = keras.Model(inputs=visible, outputs=decoder_out)
    return model


def get_x_and_y(drawings):
    X, y = [], []
    for drawing in drawings:
        # y is X shifted to the right, cuz the model predicts one step ahead in the future
        X.append(drawing[:-1]) # get all the values except the last one
        y.append(drawing[1:]) # get all the value except the first one
    return X, y


model = build_model()
model.summary(100)

model.compile(
    optimizer="adam", loss=keras.losses.CategoricalCrossentropy(), metrics=["accuracy"]
)

X_train, y_train = get_x_and_y(drawings)
# X_train[0].shape = y_train[0].shape = (49, 5, 1)
# X_train[1].shape = y_train[1].shape = (30, 5, 1)
# ...

model.fit(X_train, y_train, batch_size=1, epochs=3)

但这提出了一个ValueError: Data cardinality is ambiguous：

ValueError: Data cardinality is ambiguous:
  x sizes: 49, 30, 40, 53, 73, 45, 35, 58, 28, 59, 40, 34, 36, 26, 49, 35, 38, 37, 33, 47, 61, 38, 56, 44, 25, 38, 41, 32, 67, 25, 28, 20, 35, 35, 48, 34, 42, 43, 34, 88, 38, 41, 32, 55, 26, 47, 39, 50, 108, 27, 49, 35, 18, 74, 43, 51, 36, 49, 41, 42, 44, 76, 40, 30, 41, 31, 43, 18, 24, 89, 25, 67, 43, 44, 25, 58, 25, 47, 83, 39, 56, 56, 48, 25, 25, 43, 46, 35, 75, 49, 51, 33, 24, 37, 42, 35, 25, 29, 63, 36
  y sizes: 49, 30, 40, 53, 73, 45, 35, 58, 28, 59, 40, 34, 36, 26, 49, 35, 38, 37, 33, 47, 61, 38, 56, 44, 25, 38, 41, 32, 67, 25, 28, 20, 35, 35, 48, 34, 42, 43, 34, 88, 38, 41, 32, 55, 26, 47, 39, 50, 108, 27, 49, 35, 18, 74, 43, 51, 36, 49, 41, 42, 44, 76, 40, 30, 41, 31, 43, 18, 24, 89, 25, 67, 43, 44, 25, 58, 25, 47, 83, 39, 56, 56, 48, 25, 25, 43, 46, 35, 75, 49, 51, 33, 24, 37, 42, 35, 25, 29, 63, 36
Please provide data which shares the same first dimension.

对实际代码本身的任何帮助或建议，非常感谢。提前致谢。

标签： pythonnumpytensorflowkeras