Help to support tensorflow slim "Flatten" pattern to Tensorrt. about forward HOT 10 CLOSED

@ys0232 In the latest commit eb0eb94, Forward can support 'Flatten' Op. You could try it. And we'll support tf.depth_to_space soon.

from forward.

Another unsupport operation is tf.depth_to_space, it in torch is "PixelShuffle". Thanks for helping!

from forward.

Nice work! @yuanzexi

from forward.

There are another "Reshape" pattern like "Flatten". However, it doesn't work in Forward-1.1.0. Please help to support "Reshape" to tensorrt, thanks! Create model code look like this.@yuanzexi

def create_tf_reshape(model_file):
    import tensorflow as tf
    import tf_slim as slim
    with tf.Session() as sess:
        x1 = tf.placeholder(shape=(1,None,None,27),dtype=tf.float32, name='x')

        #op = slim.flatten(x1)
        gs = tf.shape(x1)
        op = tf.reshape(x1,shape=(1,gs[1],gs[2],3,9),name='reshape')

        sess.run(tf.global_variables_initializer())
        feed_dict = {x1: np.ones((1,12,12,27))}
        print(sess.run(op, feed_dict))

        graphdef = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['reshape'])
        tf.train.write_graph(graphdef, './', model_file, as_text=False)
        return feed_dict

from forward.

There are another "Reshape" pattern like "Flatten". However, it doesn't work in Forward-1.1.0. Please help to support "Reshape" to tensorrt, thanks! Create model code look like this.@yuanzexi

def create_tf_reshape(model_file):
    import tensorflow as tf
    import tf_slim as slim
    with tf.Session() as sess:
        x1 = tf.placeholder(shape=(1,None,None,27),dtype=tf.float32, name='x')

        #op = slim.flatten(x1)
        gs = tf.shape(x1)
        op = tf.reshape(x1,shape=(1,gs[1],gs[2],3,9),name='reshape')

        sess.run(tf.global_variables_initializer())
        feed_dict = {x1: np.ones((1,12,12,27))}
        print(sess.run(op, feed_dict))

        graphdef = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['reshape'])
        tf.train.write_graph(graphdef, './', model_file, as_text=False)
        return feed_dict

@ys0232 So far, Forward only support DYNAMIC BATCH dim, so using shape=(1, None, None, 27) as input is invalid.

from forward.

Another unsupport operation is tf.depth_to_space, it in torch is "PixelShuffle". Thanks for helping!

@yuanzexi Could help to support tf.depth_to_space and pytorch.PixelShuffle ?

from forward.

This issue is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.

from forward.

This issue was closed because it has been stalled for 5 days with no activity.

from forward.

@ys0232 Sorry for late reply. tf.depth_to_space and pytorch.PixelShuffle are now supported in Release v1.2.2.

from forward.

This issue was closed because it has been stalled for 5 days with no activity.

from forward.