AVhldZddlmZddlmZddlmZddlmZddlmZddl m Z ddl m Z d d Z d d Zy ) z Jacobian ops.)ops)tensor) array_ops) check_ops)gradients_impl)control_flow_ops)nestNc: tj| j}tj}tjdg fd} t jd}|rtj|||}n0tj|jgt z||}t|D]\} } t| tjs!tj |tj| ddgd} tj| | } | j#|j% | j| || <tj&||S#t $rtjd}Y*wxYw)aComputes jacobian of `output` w.r.t. `inputs`. Args: output: A tensor. inputs: A tensor or a nested structure of tensor objects. use_pfor: If true, uses pfor for computing the jacobian. Else uses tf.while_loop. parallel_iterations: A knob to control how many iterations and dispatched in parallel. This knob can be used to control the total memory usage. Returns: A tensor or a nested structure of tensors with the same structure as `inputs`. Each entry is the jacobian of `output` w.r.t. to the corresponding value in `inputs`. If output has shape [y_1, ..., y_n] and inputs_i has shape [x_1, ..., x_m], the corresponding jacobian has shape [y_1, ..., y_n, x_1, ..., x_m]. Note that in cases where the gradient is sparse (IndexedSlices), jacobian function currently makes it dense and returns a Tensor instead. This may change in the future. c\tj|}tj|S)Nrgather gradient_ops gradients)iy flat_inputsoutputs \/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/python/ops/parallel_for/gradients.pyloop_fnzjacobian..loop_fn2s(#A  ! !![ 11rparallel_iterationsNaxis)r flattenshaperreshapeint TypeErrorrpforfor_loopdtypelen enumerate isinstancerTensorconcat set_shape concatenatepack_sequence_as) rinputsuse_pforroutput_tensor_shape output_shaper output_size pfor_outputsrout new_shapers ` @rjacobianr5so( V$+ (,   VbT *&2-fll1o&K#((2EGL$,, [))/ 1L ,'fa#v}}%"" -ab1 2 [[[2, 0], [0, 4]], [[6, 0], [0, 8]]] Args: output: A tensor with shape [b, y1, ..., y_n]. `output[i,...]` should only depend on `inp[i,...]`. inp: A tensor with shape [b, x1, ..., x_m] use_pfor: If true, uses pfor for computing the Jacobian. Else uses a tf.while_loop. parallel_iterations: A knob to control how many iterations are vectorized and dispatched in parallel. The default value of None, when use_pfor is true, corresponds to vectorizing all the iterations. When use_pfor is false, the default value of None corresponds to parallel_iterations=10. This knob can be used to control the total memory usage. Returns: A tensor `t` with shape [b, y_1, ..., y_n, x1, ..., x_m] where `t[i, ...]` is the jacobian of `output[i, ...]` w.r.t. `inp[i, ...]`, i.e. stacked per-example jacobians. Raises: ValueError: if first dimension of `output` and `inp` do not match. rz(Need first dimension of `output` shape (z) and `inp` shape (z ) to match.Ncftj|d}tj|dS)Nrrrr )rrinprs rrzbatch_jacobian..loop_fn~s/+A  ! !!S )! ,,rrr )rrrr)ris_compatible_with ValueErroris_fully_definedr num_elementsrsizercontrol_dependenciesr assert_equalrrr"r#r$ transposer)) rr8r.rr0 batch_sizeoutput_row_size inp_shaper pfor_outputr4s `` rbatch_jacobianrFPs:, a + +CIIaL 9 ? ~N)), ;@ AA""$\!_%J"//1Z?O??6*LaJnnV, :Oooc") j)A,78:F   v O'D EFF-"''rMs++.++@?'45n@.r