Ошибка ValueError: Cannot reshape a tensor with 26912 elements to shape [1,30,30,32] (28800 elements) for '{{node Reshape_6}} = Reshape[T=DT_FLOAT,
Пытаюсь запустить сеть. Получил сигнатурные матрицы:
series_number is 30
series_length is 20000
signature_matrices_number is 2000
the shape of signature_matrices is (3, 2000, 30, 30)
train dataset shape is (790, 5, 30, 30, 3)
test dataset shape is (1200, 5, 30, 30, 3)
Дальше в коде:
import utils as util
import numpy as np
import os
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
def cnn_encoder_layer(data, filter_layer, strides):
result = tf.nn.conv2d(
input=data,
filters=filter_layer,
strides=strides,
padding="SAME")
return tf.nn.selu(result)
def tensor_variable(shape, name):
variable = tf.Variable(tf.zeros(shape), name=name)
variable = tf.compat.v1.get_variable(name, shape=shape, initializer=tf.compat.v1.keras.initializers.glorot_uniform())
return variable
def cnn_encoder(data):
filter1 = tensor_variable([3, 3, 3, 32], "filter1")
strides1 = (1, 1, 1, 1)
cnn1_out = cnn_encoder_layer(data, filter1, strides1)
filter2 = tensor_variable([3, 3, 32, 64], "filter2")
strides2 = (1, 2, 2, 1)
cnn2_out = cnn_encoder_layer(cnn1_out, filter2, strides2)
filter3 = tensor_variable([2, 2, 64, 128], "filter3")
strides3 = (1, 2, 2, 1)
cnn3_out = cnn_encoder_layer(cnn2_out, filter3, strides3)
filter4 = tensor_variable([2, 2, 128, 256], "filter4")
strides4 = (1, 2, 2, 1)
cnn4_out = cnn_encoder_layer(cnn3_out, filter4, strides4)
return cnn1_out, cnn2_out, cnn3_out, cnn4_out
def cnn_lstm_attention_layer(input_data, layer_number):
convlstm_layer = tf.compat.v1.keras.layers.ConvLSTM2D(
filters=input_data.shape[-1],
kernel_size=(2, 2),
use_bias=True,
return_sequences=True,
name="conv_lstm_cell" + str(layer_number))
outputs = convlstm_layer(input_data)
attention_w = []
for k in range(util.step_max):
attention_w.append(tf.reduce_sum(tf.multiply(outputs[:, k], outputs[:, -1])) / util.step_max)
attention_w = tf.reshape(tf.nn.softmax(tf.stack(attention_w)), [1, util.step_max])
outputs = tf.reshape(outputs, [util.step_max, -1])
outputs = tf.matmul(attention_w, outputs)
outputs = tf.reshape(outputs, [1, input_data.shape[2], input_data.shape[3], input_data.shape[4]])
return outputs, attention_w
def cnn_decoder_layer(conv_lstm_out_c, filter, output_shape, strides):
deconv = tf.nn.conv2d_transpose(
value=conv_lstm_out_c,
filters=filter,
output_shape=output_shape,
strides=strides,
padding="SAME")
deconv = tf.nn.selu(deconv)
return deconv
def cnn_decoder(lstm1_out, lstm2_out, lstm3_out, lstm4_out):
d_filter4 = tensor_variable([2, 2, 128, 256], "d_filter4")
dec4 = cnn_decoder_layer(lstm4_out, d_filter4, [1, 8, 8, 128], (1, 2, 2, 1))
dec4_concat = tf.concat([dec4, lstm3_out], axis=3)
d_filter3 = tensor_variable([2, 2, 64, 256], "d_filter3")
dec3 = cnn_decoder_layer(dec4_concat, d_filter3, [1, 15, 15, 64], (1, 2, 2, 1))
dec3_concat = tf.concat([dec3, lstm2_out], axis=3)
d_filter2 = tensor_variable([3, 3, 32, 128], "d_filter2")
dec2 = cnn_decoder_layer(dec3_concat, d_filter2, [1, 32, 29, 29], (1, 2, 2, 1))
dec2_concat = tf.concat([dec2, lstm1_out], axis=3)
d_filter1 = tensor_variable([3, 3, 3, 64], "d_filter1")
dec1 = cnn_decoder_layer(dec2_concat, d_filter1, [1, 30, 30, 3], (1, 1, 1, 1))
return dec1
def main():
matrix_data_path = util.train_data_path + "train.npy"
matrix_gt_1 = np.load(matrix_data_path)
data_input = tf.compat.v1.placeholder(tf.float32, [util.step_max, 30, 30, 3])
conv1_out, conv2_out, conv3_out, conv4_out = cnn_encoder(data_input)
conv1_out = tf.reshape(conv1_out, [-1, 5, 30, 30, 32])
conv2_out = tf.reshape(conv2_out, [-1, 5, 15, 15, 64])
conv3_out = tf.reshape(conv3_out, [-1, 5, 8, 8, 128])
conv4_out = tf.reshape(conv4_out, [-1, 5, 4, 4, 256])
conv1_lstm_attention_out, atten_weight_1 = cnn_lstm_attention_layer(conv1_out, 1)
conv2_lstm_attention_out, atten_weight_2 = cnn_lstm_attention_layer(conv2_out, 2)
conv3_lstm_attention_out, atten_weight_3 = cnn_lstm_attention_layer(conv3_out, 3)
conv4_lstm_attention_out, atten_weight_4 = cnn_lstm_attention_layer(conv4_out, 4)
deconv_out = cnn_decoder(conv1_lstm_attention_out, conv2_lstm_attention_out, conv3_lstm_attention_out,
conv4_lstm_attention_out)
loss = tf.reduce_mean(tf.square(data_input[-1] - deconv_out))
optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=util.learning_rate).minimize(loss)
init = tf.global_variables_initializer()
sess = tf.compat.v1.Session()
sess.run(init)
for idx in range(util.train_start_id, util.train_end_id):
matrix_gt = matrix_gt_1[idx - util.train_start_id]
feed_dict = {data_input: np.asarray(matrix_gt)}
a, loss_value = sess.run([optimizer, loss], feed_dict)
print("mse of last train data: " + str(loss_value))
matrix_data_path = util.test_data_path + "test.npy"
matrix_gt_1 = np.load(matrix_data_path)
result_all = []
for idx in range(util.test_start_id, util.test_end_id):
matrix_gt = matrix_gt_1[idx - util.test_start_id]
feed_dict = {data_input: np.asarray(matrix_gt)}
result, loss_value = sess.run([deconv_out, loss], feed_dict)
result_all.append(result)
print("mse of last test data: " + str(loss_value))
reconstructed_path = util.reconstructed_data_path
if not os.path.exists(reconstructed_path):
os.makedirs(reconstructed_path)
reconstructed_path = reconstructed_path + "test_reconstructed.npy"
result_all = np.asarray(result_all).reshape((-1, 30, 30, 3))
print(result_all.shape)
np.save(reconstructed_path, result_all)
if __name__ == '__main__':
main()
появляется ошибка:
ValueError: Cannot reshape a tensor with 26912 elements to shape
[1,30,30,32] (28800 elements) for '{{node Reshape_6}} =
Reshape[T=DT_FLOAT, Tshape=DT_INT32](MatMul, Reshape_6/shape)' with
input shapes: [1,26912], [4] and with input tensors computed as partial shapes: input[1] = [1,30,30,32]