Pushing ScienceQA version

model-00001-of-00003.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4214710d3c4b31d9a89527da2c196e344c9a41fbf6e4a7e942a8a626b9e911c5
+oid sha256:837c35fe2d467993abf7e629c20e2e85e280cd827d60f01f0b82a42c592eb3ca
 size 4917078632

model-00002-of-00003.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e74788d2d95dc3174fd671a7e987fb4fb0243e25b1c8803a1fef8e084117638e
+oid sha256:f4be49e2abc4ce683f7a190b95871b759deeab622aa488ee6e7051aaf2ae190a
 size 4983443424

model-00003-of-00003.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:24ca5c47d98179d7796291d5d4d6c2b8706f5173d2dcfc3fa57a6d394575f9fd
+oid sha256:51390a18b5727f7acebfc9ac9aab1d214882ad055fb8b165ea2b03136633ef49
 size 932581696

modeling_feynmodel.py

@@ -1,15 +1,19 @@
 # modeling_fynmodel : Imed MAGROUNE / 2024 - 09
+#
 # original code from modeling_FeynModel
+# Use of Gemma2 Layers
 # add DaVit Vision Tower
 #
-# update generate forward function
+# update generate and forward function
 #
 # add lora adapters
 #
 # train on coco OD and vision reasoning
+#
 # train on ScenceQA
 #
-# todo add mamaba layer
+# 
+# add mamaba layer
 #
 # todo train on Arc-AGI
 
@@ -50,7 +54,7 @@ from transformers.modeling_outputs import (
 from typing import List, Optional, Tuple, Union
 
 from transformers.models.gemma2.modeling_gemma2 import Gemma2Model, Gemma2ForCausalLM,Gemma2DecoderLayer,Gemma2RMSNorm
-from .configuration_feynmodel import FeynModelConfig,Florence2VisionConfig
+from configuration_feynmodel import FeynModelConfig,Florence2VisionConfig
 
 from transformers import AutoProcessor, AutoTokenizer, AutoModelForCausalLM
 import json
@@ -1330,7 +1334,7 @@ class FeynModelForCausalLM(Gemma2ForCausalLM):
                 inputs_embeds = self.get_input_embeddings()(input_ids)
             image_features = self._encode_image(pixel_values)
             inputs_embeds, causal_attention_mask = self._merge_input_ids_with_image_features(image_features, inputs_embeds )
-            causal_attention_mask = create_git_attention_mask(tgt=input_ids, memory=image_features,max_length=2048)
+            causal_attention_mask = create_git_attention_mask(tgt=input_ids, memory=image_features,max_length=8192)
             causal_attention_mask=causal_attention_mask.to(input_ids.device)
             self.__causal_attention_mask=causal_attention_mask
 
@@ -1495,7 +1499,7 @@ class FeynModelForCausalLM(Gemma2ForCausalLM):
             if input_ids is not None:
                 
                 inputs_embeds = self.get_input_embeddings()(input_ids)
-            print("pixels")
+            
             image_features = self._encode_image(pixel_values)
             inputs_embeds, causal_attention_mask = self._merge_input_ids_with_image_features(image_features, inputs_embeds )
             causal_attention_mask = create_git_attention_mask(tgt=input_ids, memory=image_features,max_length=max_length)