From Numpy Array to TC Tensor and vice versa

[MuhamedEssam]

if i have X with type np.array how can i convert it to be Tinychain Tensor format
So i can use Tc.tensor methods ?

and for the opposite should i use .flatten.tolist() then cast it by np.array()

[MuhamedEssam]

Also do we have .shape Function over here in tinychain ?

[MuhamedEssam]

i tested this in numpy and it worked

        self.dbeta = 2/n_samples*(np.einsum("ij,ki->kj", self.beta,
            np.einsum("ij,ki->kj", self.X, self.X.T)) 
                                  - np.einsum("ij,ki->kj", self.y, self.X.T))  

can we say that

      d_weights = 2/len(inputs)*(tc.tensor.einsum("ij,ki->kj", weights,
                                                   tc.tensor.einsum("ij,ki->kj", inputs, inputs.transpose))-
                                                      tc.tensor.einsum("ij,ki->kj", labels, inputs.transpose)) 

[haydnv]

Hi Muhamed, there is a Tensor.shape function which returns a Tuple of U64s and I just committed a change (in #102) so that Tuple now supports len and iter when its contents are known at compile-time. So, the only issues I can possibly see with your code are len(inputs) (it might be better to call inputs.shape[0], for a Tensor, or inputs.len(), for a Tuple which may not be resolved at compile-time) and inputs.transpose() (note: transpose is a method and needs the parentheses to call it).

In general you can load a Tensor from an nparray x by calling Tensor.load(x.shape, tc.F32, x.flatten().tolist()) where you should replace F32 with whatever data type you're expecting. Eventually the Python client will be smart enough to support NumPy and Pandas types (almost) automatically but this is still a few months away.

[MuhamedEssam]

but when we flatten then we will lose information about width and hight and it will be replaced by size information only ? you got me ?

[haydnv]

The purpose of providing x.shape is to set the shape of the Tensor, so you won't lose that information. The reason you need to call .flatten().tolist() is that the TinyChain client doesn't know about NumPy (so that client users won't have to install NumPy in order to use TinyChain) and it encodes, stores, and transmits tensor elements as one long sequence for reasons of efficiency. You can check that the shape is treated correctly by using any of the Tensor methods that accept an axis argument, e.g. d_weights.reduce_sum(axis=-1).