Hello,
\nI am trying to incorporate a VisionTransformer and got the following error. The same model works with CPU.
All the layers of TFViTModel were initialized the model checkpoint at google/vit-base-patch16--in21k.\n your is similar to the the model of the checkpoint was trained on, you can already use TFViTModel predictions without further training.\n---------------------------------------------------------------------------\nInvalidArgumentError Traceback (most recent last)\nCell In[], line \n loss_value = loss_fn(labels, logits)\n # Backward pass and optimization\n--> grads = tape.gradient(loss_value, model.trainable_variables)\n optimizer.apply_gradients(zip(grads, model.trainable_variables))\n # training loss epoch\n\n condapython3.tensorfloweager/backprop.py:, in GradientTape.gradient(self, target, sources, output_gradients, unconnected_gradients)\n output_gradients = (\n composite_tensor_gradient.get_flat_tensors_for_gradients(\n output_gradients))\n output_gradients = [None x is None ops.convert_to_tensor(x)\n x in output_gradients]\n-> flat_grad = imperative_grad.imperative_grad(\n self._tape,\n flat_targets,\n flat_sources,\n output_gradients=output_gradients,\n sources_raw=flat_sources_raw,\n unconnected_gradients=unconnected_gradients)\n not self._persistent:\n # Keep track of watched variables before setting tape to None\n self._watched_variables = self._tape.watched_variables()\n\n condapython3.tensorfloweager/imperative_grad.py:, in imperative_grad(tape, target, sources, output_gradients, sources_raw, unconnected_gradients)\n except ValueError:\n raise ValueError(\n % unconnected_gradients)\n---> pywrap_tfe.TFE_Py_TapeGradient(\n tape._tape, # pylint: disable=-access\n target,\n sources,\n output_gradients,\n sources_raw,\n compat.as_str(unconnected_gradients.value))\n\n condapython3.tensorfloweager/backprop.py:, in _gradient_function(op_name, attr_tuple, num_inputs, inputs, outputs, out_grads, skip_input_indices, forward_pass_name_scope)\n gradient_name_scope += forward_pass_name_scope + \n with ops.name_scope(gradient_name_scope):\n--> grad_fn(mock_op, *out_grads)\n :\n grad_fn(mock_op, *out_grads)\n\n condapython3.tensorflowops/array_grad.py:, in _TransposeGrad(op, grad)\n \n p = op.inputs[]\n--> [array_ops.transpose(grad, array_ops.invert_permutation(p)), None]\n\n condapython3.tensorflowops/gen_array_ops.py:, in invert_permutation(x, name)\n _result\n except _core._NotOkStatusException as e:\n-> _ops.raise_from_not_ok_status(e, name)\n except _core._FallbackException:\n pass\n\n condapython3.tensorflowframework/ops.py:, in raise_from_not_ok_status(e, name)\n raise_from_not_ok_status(e, name):\n e.message += ( + name name is not None )\n-> raise core._status_to_exception(e) None\n\nInvalidArgumentError: {{function_node __wrapped__InvertPermutation_device_replica:device:GPU:}} invert_permutation expects a D vector. [Op:InvertPermutation]\nAny ideas? Thank you.
", "messages": [ { "id": 2334193, "postDate": "2023-07-07T13:58:06.690Z", "content": "Hello,
\nI am trying to incorporate a VisionTransformer and got the following error. The same model works with CPU.
All the layers of TFViTModel were initialized the model checkpoint at google/vit-base-patch16--in21k.\n your is similar to the the model of the checkpoint was trained on, you can already use TFViTModel predictions without further training.\n---------------------------------------------------------------------------\nInvalidArgumentError Traceback (most recent last)\nCell In[], line \n loss_value = loss_fn(labels, logits)\n # Backward pass and optimization\n--> grads = tape.gradient(loss_value, model.trainable_variables)\n optimizer.apply_gradients(zip(grads, model.trainable_variables))\n # training loss epoch\n\n condapython3.tensorfloweager/backprop.py:, in GradientTape.gradient(self, target, sources, output_gradients, unconnected_gradients)\n output_gradients = (\n composite_tensor_gradient.get_flat_tensors_for_gradients(\n output_gradients))\n output_gradients = [None x is None ops.convert_to_tensor(x)\n x in output_gradients]\n-> flat_grad = imperative_grad.imperative_grad(\n self._tape,\n flat_targets,\n flat_sources,\n output_gradients=output_gradients,\n sources_raw=flat_sources_raw,\n unconnected_gradients=unconnected_gradients)\n not self._persistent:\n # Keep track of watched variables before setting tape to None\n self._watched_variables = self._tape.watched_variables()\n\n condapython3.tensorfloweager/imperative_grad.py:, in imperative_grad(tape, target, sources, output_gradients, sources_raw, unconnected_gradients)\n except ValueError:\n raise ValueError(\n % unconnected_gradients)\n---> pywrap_tfe.TFE_Py_TapeGradient(\n tape._tape, # pylint: disable=-access\n target,\n sources,\n output_gradients,\n sources_raw,\n compat.as_str(unconnected_gradients.value))\n\n condapython3.tensorfloweager/backprop.py:, in _gradient_function(op_name, attr_tuple, num_inputs, inputs, outputs, out_grads, skip_input_indices, forward_pass_name_scope)\n gradient_name_scope += forward_pass_name_scope + \n with ops.name_scope(gradient_name_scope):\n--> grad_fn(mock_op, *out_grads)\n :\n grad_fn(mock_op, *out_grads)\n\n condapython3.tensorflowops/array_grad.py:, in _TransposeGrad(op, grad)\n \n p = op.inputs[]\n--> [array_ops.transpose(grad, array_ops.invert_permutation(p)), None]\n\n condapython3.tensorflowops/gen_array_ops.py:, in invert_permutation(x, name)\n _result\n except _core._NotOkStatusException as e:\n-> _ops.raise_from_not_ok_status(e, name)\n except _core._FallbackException:\n pass\n\n condapython3.tensorflowframework/ops.py:, in raise_from_not_ok_status(e, name)\n raise_from_not_ok_status(e, name):\n e.message += ( + name name is not None )\n-> raise core._status_to_exception(e) None\n\nInvalidArgumentError: {{function_node __wrapped__InvertPermutation_device_replica:device:GPU:}} invert_permutation expects a D vector. [Op:InvertPermutation]\nAny ideas? Thank you.
", "rawMarkdown": "Hello,\nI am trying to incorporate a VisionTransformer and got the following error. The same model works with CPU.\n\n```\nAll the layers of TFViTModel were initialized from the model checkpoint at google/vit-base-patch16-224-in21k.\nIf your task is similar to the task the model of the checkpoint was trained on, you can already use TFViTModel for predictions without further training.\n---------------------------------------------------------------------------\nInvalidArgumentError Traceback (most recent call last)\nCell In[1], line 146\n 143 loss_value = loss_fn(labels, logits)\n 145 # Backward pass and optimization\n--> 146 grads = tape.gradient(loss_value, model.trainable_variables)\n 147 optimizer.apply_gradients(zip(grads, model.trainable_variables))\n 149 # Print training loss for each epoch\n\nFile /opt/conda/lib/python3.10/site-packages/tensorflow/python/eager/backprop.py:1063, in GradientTape.gradient(self, target, sources, output_gradients, unconnected_gradients)\n 1057 output_gradients = (\n 1058 composite_tensor_gradient.get_flat_tensors_for_gradients(\n 1059 output_gradients))\n 1060 output_gradients = [None if x is None else ops.convert_to_tensor(x)\n 1061 for x in output_gradients]\n-> 1063 flat_grad = imperative_grad.imperative_grad(\n 1064 self._tape,\n 1065 flat_targets,\n 1066 flat_sources,\n 1067 output_gradients=output_gradients,\n 1068 sources_raw=flat_sources_raw,\n 1069 unconnected_gradients=unconnected_gradients)\n 1071 if not self._persistent:\n 1072 # Keep track of watched variables before setting tape to None\n 1073 self._watched_variables = self._tape.watched_variables()\n\nFile /opt/conda/lib/python3.10/site-packages/tensorflow/python/eager/imperative_grad.py:67, in imperative_grad(tape, target, sources, output_gradients, sources_raw, unconnected_gradients)\n 63 except ValueError:\n 64 raise ValueError(\n 65 \"Unknown value for unconnected_gradients: %r\" % unconnected_gradients)\n---> 67 return pywrap_tfe.TFE_Py_TapeGradient(\n 68 tape._tape, # pylint: disable=protected-access\n 69 target,\n 70 sources,\n 71 output_gradients,\n 72 sources_raw,\n 73 compat.as_str(unconnected_gradients.value))\n\nFile /opt/conda/lib/python3.10/site-packages/tensorflow/python/eager/backprop.py:146, in _gradient_function(op_name, attr_tuple, num_inputs, inputs, outputs, out_grads, skip_input_indices, forward_pass_name_scope)\n 144 gradient_name_scope += forward_pass_name_scope + \"/\"\n 145 with ops.name_scope(gradient_name_scope):\n--> 146 return grad_fn(mock_op, *out_grads)\n 147 else:\n 148 return grad_fn(mock_op, *out_grads)\n\nFile /opt/conda/lib/python3.10/site-packages/tensorflow/python/ops/array_grad.py:832, in _TransposeGrad(op, grad)\n 830 \"\"\"Returns unshuffle(grad).\"\"\"\n 831 p = op.inputs[1]\n--> 832 return [array_ops.transpose(grad, array_ops.invert_permutation(p)), None]\n\nFile /opt/conda/lib/python3.10/site-packages/tensorflow/python/ops/gen_array_ops.py:4527, in invert_permutation(x, name)\n 4525 return _result\n 4526 except _core._NotOkStatusException as e:\n-> 4527 _ops.raise_from_not_ok_status(e, name)\n 4528 except _core._FallbackException:\n 4529 pass\n\nFile /opt/conda/lib/python3.10/site-packages/tensorflow/python/framework/ops.py:7262, in raise_from_not_ok_status(e, name)\n 7260 def raise_from_not_ok_status(e, name):\n 7261 e.message += (\" name: \" + name if name is not None else \"\")\n-> 7262 raise core._status_to_exception(e) from None\n\nInvalidArgumentError: {{function_node __wrapped__InvertPermutation_device_/job:localhost/replica:0/task:0/device:GPU:0}} invert_permutation expects a 1D vector. [Op:InvertPermutation]\n```\n\nAny ideas? Thank you." } ], "comments": [], "raw_markdown_by_id": { "2334193": "Hello,\nI am trying to incorporate a VisionTransformer and got the following error. The same model works with CPU.\n\n```\nAll the layers of TFViTModel were initialized from the model checkpoint at google/vit-base-patch16-224-in21k.\nIf your task is similar to the task the model of the checkpoint was trained on, you can already use TFViTModel for predictions without further training.\n---------------------------------------------------------------------------\nInvalidArgumentError Traceback (most recent call last)\nCell In[1], line 146\n 143 loss_value = loss_fn(labels, logits)\n 145 # Backward pass and optimization\n--> 146 grads = tape.gradient(loss_value, model.trainable_variables)\n 147 optimizer.apply_gradients(zip(grads, model.trainable_variables))\n 149 # Print training loss for each epoch\n\nFile /opt/conda/lib/python3.10/site-packages/tensorflow/python/eager/backprop.py:1063, in GradientTape.gradient(self, target, sources, output_gradients, unconnected_gradients)\n 1057 output_gradients = (\n 1058 composite_tensor_gradient.get_flat_tensors_for_gradients(\n 1059 output_gradients))\n 1060 output_gradients = [None if x is None else ops.convert_to_tensor(x)\n 1061 for x in output_gradients]\n-> 1063 flat_grad = imperative_grad.imperative_grad(\n 1064 self._tape,\n 1065 flat_targets,\n 1066 flat_sources,\n 1067 output_gradients=output_gradients,\n 1068 sources_raw=flat_sources_raw,\n 1069 unconnected_gradients=unconnected_gradients)\n 1071 if not self._persistent:\n 1072 # Keep track of watched variables before setting tape to None\n 1073 self._watched_variables = self._tape.watched_variables()\n\nFile /opt/conda/lib/python3.10/site-packages/tensorflow/python/eager/imperative_grad.py:67, in imperative_grad(tape, target, sources, output_gradients, sources_raw, unconnected_gradients)\n 63 except ValueError:\n 64 raise ValueError(\n 65 \"Unknown value for unconnected_gradients: %r\" % unconnected_gradients)\n---> 67 return pywrap_tfe.TFE_Py_TapeGradient(\n 68 tape._tape, # pylint: disable=protected-access\n 69 target,\n 70 sources,\n 71 output_gradients,\n 72 sources_raw,\n 73 compat.as_str(unconnected_gradients.value))\n\nFile /opt/conda/lib/python3.10/site-packages/tensorflow/python/eager/backprop.py:146, in _gradient_function(op_name, attr_tuple, num_inputs, inputs, outputs, out_grads, skip_input_indices, forward_pass_name_scope)\n 144 gradient_name_scope += forward_pass_name_scope + \"/\"\n 145 with ops.name_scope(gradient_name_scope):\n--> 146 return grad_fn(mock_op, *out_grads)\n 147 else:\n 148 return grad_fn(mock_op, *out_grads)\n\nFile /opt/conda/lib/python3.10/site-packages/tensorflow/python/ops/array_grad.py:832, in _TransposeGrad(op, grad)\n 830 \"\"\"Returns unshuffle(grad).\"\"\"\n 831 p = op.inputs[1]\n--> 832 return [array_ops.transpose(grad, array_ops.invert_permutation(p)), None]\n\nFile /opt/conda/lib/python3.10/site-packages/tensorflow/python/ops/gen_array_ops.py:4527, in invert_permutation(x, name)\n 4525 return _result\n 4526 except _core._NotOkStatusException as e:\n-> 4527 _ops.raise_from_not_ok_status(e, name)\n 4528 except _core._FallbackException:\n 4529 pass\n\nFile /opt/conda/lib/python3.10/site-packages/tensorflow/python/framework/ops.py:7262, in raise_from_not_ok_status(e, name)\n 7260 def raise_from_not_ok_status(e, name):\n 7261 e.message += (\" name: \" + name if name is not None else \"\")\n-> 7262 raise core._status_to_exception(e) from None\n\nInvalidArgumentError: {{function_node __wrapped__InvertPermutation_device_/job:localhost/replica:0/task:0/device:GPU:0}} invert_permutation expects a 1D vector. [Op:InvertPermutation]\n```\n\nAny ideas? Thank you." } }